lint

Signed-off-by: Duncan Moss <djm.moss@gmail.com>

.buildkite/lm-eval-harness/configs/Qwen2-1.5B-Instruct-W8A16-compressed-tensors.yaml

@@ -0,0 +1,12 @@
+# For vllm script, with -t option (tensor parallel size).
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Qwen2-1.5B-Instruct-W8A16-Channelwise -b "auto" -l 1000 -f 5 -t 1
+model_name: "nm-testing/Qwen2-1.5B-Instruct-W8A16-Channelwise"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.595
+  - name: "exact_match,flexible-extract"
+    value: 0.582
+limit: 1000
+num_fewshot: 5

.buildkite/release-pipeline.yaml

@@ -132,7 +132,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_AMXBF16=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest"
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:

.buildkite/scripts/annotate-release.sh

@@ -2,29 +2,22 @@
 
 set -ex
 
-# Get release version, default to 1.0.0.dev for nightly/per-commit builds
-RELEASE_VERSION=$(buildkite-agent meta-data get release-version 2>/dev/null | sed 's/^v//')
-if [ -z "${RELEASE_VERSION}" ]; then
-  RELEASE_VERSION="1.0.0.dev"
+# Get release version and strip leading 'v' if present
+RELEASE_VERSION=$(buildkite-agent meta-data get release-version | sed 's/^v//')
+
+if [ -z "$RELEASE_VERSION" ]; then
+  echo "Error: RELEASE_VERSION is empty. 'release-version' metadata might not be set or is invalid."
+  exit 1
 fi
 
 buildkite-agent annotate --style 'info' --context 'release-workflow' << EOF
-To download the wheel (by commit):
-\`\`\`
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux2014_aarch64.whl .
-
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${BUILDKITE_COMMIT}/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux1_x86_64.whl .
-\`\`\`
-
-To download the wheel (by version):
+To download the wheel:
 \`\`\`
 aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux1_x86_64.whl .
 aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux2014_aarch64.whl .
 
+aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}+cu126/vllm-${RELEASE_VERSION}+cu126-cp38-abi3-manylinux1_x86_64.whl .
 aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}+cu129/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}+cu130/vllm-${RELEASE_VERSION}+cu130-cp38-abi3-manylinux1_x86_64.whl .
 \`\`\`
 
 To download and upload the image:
@@ -45,10 +38,9 @@ docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
 docker push vllm/vllm-openai:latest-aarch64
 docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
 
-docker manifest rm vllm/vllm-openai:latest
-docker manifest create vllm/vllm-openai:latest vllm/vllm-openai:latest-x86_64 vllm/vllm-openai:latest-aarch64
-docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
+docker manifest create vllm/vllm-openai:latest vllm/vllm-openai:latest-x86_64 vllm/vllm-openai:latest-aarch64 --amend
+docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64 --amend
 docker manifest push vllm/vllm-openai:latest
 docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}
 \`\`\`
-EOF 
+EOF 

.buildkite/scripts/hardware_ci/run-amd-test.sh

@@ -59,7 +59,7 @@ while true; do
         fi
 done
 
-echo "--- Pulling container"
+echo "--- Pulling container" 
 image_name="rocm/vllm-ci:${BUILDKITE_COMMIT}"
 container_name="rocm_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
 docker pull "${image_name}"
@@ -78,13 +78,17 @@ HF_MOUNT="/root/.cache/huggingface"
 commands=$@
 echo "Commands:$commands"
 
-commands=${commands//"pytest -v -s basic_correctness/test_basic_correctness.py"/"pytest -v -s basic_correctness/test_basic_correctness.py"}
+if [[ $commands == *"pytest -v -s basic_correctness/test_basic_correctness.py"* ]]; then
+  commands=${commands//"pytest -v -s basic_correctness/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s basic_correctness/test_basic_correctness.py"}
+fi
 
 if [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
   commands=${commands//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
 fi
 
-commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"pytest -v -s compile/test_basic_correctness.py"}
+if [[ $commands == *"pytest -v -s compile/test_basic_correctness.py"* ]]; then
+  commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s compile/test_basic_correctness.py"}
+fi
 
 if [[ $commands == *"pytest -v -s lora"* ]]; then
   commands=${commands//"pytest -v -s lora"/"VLLM_ROCM_CUSTOM_PAGED_ATTN=0 pytest -v -s lora"}
@@ -169,28 +173,19 @@ fi
 PARALLEL_JOB_COUNT=8
 MYPYTHONPATH=".."
 
-# Test that we're launching on the machine that has
-# proper access to GPUs
-render_gid=$(getent group render | cut -d: -f3)
-if [[ -z "$render_gid" ]]; then
-  echo "Error: 'render' group not found. This is required for GPU access." >&2
-  exit 1
-fi
-
-# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs.
+# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs. 
 if [[ $commands == *"--shard-id="* ]]; then
-  # assign job count as the number of shards used
-  commands=$(echo "$commands" | sed -E "s/--num-shards[[:blank:]]*=[[:blank:]]*[0-9]*/--num-shards=${PARALLEL_JOB_COUNT} /g" | sed 's/ \\ / /g')
+  # assign job count as the number of shards used   
+  commands=${commands//"--num-shards= "/"--num-shards=${PARALLEL_JOB_COUNT} "}
   for GPU in $(seq 0 $(($PARALLEL_JOB_COUNT-1))); do
     # assign shard-id for each shard
-    commands_gpu=$(echo "$commands" | sed -E "s/--shard-id[[:blank:]]*=[[:blank:]]*[0-9]*/--shard-id=${GPU} /g" | sed 's/ \\ / /g')
+    commands_gpu=${commands//"--shard-id= "/"--shard-id=${GPU} "}
     echo "Shard ${GPU} commands:$commands_gpu"
     echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
     docker run \
         --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
         --network=host \
         --shm-size=16gb \
-        --group-add "$render_gid" \
         --rm \
         -e HIP_VISIBLE_DEVICES="${GPU}" \
         -e HF_TOKEN \
@@ -222,8 +217,8 @@ else
           --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
           --network=host \
           --shm-size=16gb \
-          --group-add "$render_gid" \
           --rm \
+          -e HIP_VISIBLE_DEVICES=0 \
           -e HF_TOKEN \
           -e AWS_ACCESS_KEY_ID \
           -e AWS_SECRET_ACCESS_KEY \

.buildkite/scripts/hardware_ci/run-cpu-test-arm.sh

@@ -1,62 +0,0 @@
-#!/bin/bash
-
-# This script build the CPU docker image and run the offline inference inside the container.
-# It serves a sanity check for compilation and basic model usage.
-set -ex
-
-# allow to bind to different cores
-CORE_RANGE=${CORE_RANGE:-0-16}
-OMP_CORE_RANGE=${OMP_CORE_RANGE:-0-16}
-
-export CMAKE_BUILD_PARALLEL_LEVEL=16
-
-# Setup cleanup
-remove_docker_container() {
-    set -e;
-    docker rm -f cpu-test || true;
-}
-trap remove_docker_container EXIT
-remove_docker_container
-
-# Try building the docker image
-docker build --tag cpu-test --target vllm-test -f docker/Dockerfile.cpu .
-
-# Run the image
-docker run -itd --cpuset-cpus="$CORE_RANGE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=16 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test cpu-test
-
-function cpu_tests() {
-  set -e
-
-  docker exec cpu-test bash -c "
-    set -e
-    pip list"
-
-  # offline inference
-  docker exec cpu-test bash -c "
-    set -e
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
-
-  # Run kernel tests
-  docker exec cpu-test bash -c "
-    set -e
-    pytest -x -v -s tests/kernels/test_onednn.py
-    pytest -x -v -s tests/kernels/attention/test_cpu_attn.py"
-
-  # basic online serving
-  docker exec cpu-test bash -c '
-    set -e
-    VLLM_CPU_OMP_THREADS_BIND=$E2E_OMP_THREADS vllm serve Qwen/Qwen3-0.6B --max-model-len 2048 &
-    server_pid=$!
-    timeout 600 bash -c "until curl localhost:8000/v1/models; do sleep 1; done" || exit 1
-    vllm bench serve \
-      --backend vllm \
-      --dataset-name random \
-      --model Qwen/Qwen3-0.6B \
-      --num-prompts 20 \
-      --endpoint /v1/completions
-    kill -s SIGTERM $server_pid &'
-}
-
-# All of CPU tests are expected to be finished less than 40 mins.
-export -f cpu_tests
-timeout 2h bash -c cpu_tests

.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh

@@ -25,22 +25,20 @@ function cpu_tests() {
 
   # offline inference
   podman exec -it "$container_id" bash -c "
-    export TORCH_COMPILE_DISABLE=1
     set -xve
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m" >> $HOME/test_basic.log
 
   # Run basic model test
   podman exec -it "$container_id" bash -c "
-    export TORCH_COMPILE_DISABLE=1
     set -evx
     pip install pytest pytest-asyncio einops peft Pillow soundfile transformers_stream_generator matplotlib
-    pip install sentence-transformers datamodel_code_generator tblib 
+    pip install sentence-transformers datamodel_code_generator
 
     # Note: disable Bart until supports V1
     # pytest -v -s tests/models/language/generation/test_bart.py -m cpu_model
-    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-False-5-32-openai-community/gpt2]
-    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-False-5-32-facebook/opt-125m]
-    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-False-5-32-google/gemma-1.1-2b-it]
+    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-openai-community/gpt2]
+    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-facebook/opt-125m]
+    pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-google/gemma-1.1-2b-it]
     pytest -v -s tests/models/language/pooling/test_classification.py::test_models[float-jason9693/Qwen2.5-1.5B-apeach]
     # TODO: Below test case tests/models/language/pooling/test_embedding.py::test_models[True-ssmits/Qwen2-7B-Instruct-embed-base] fails on ppc64le. Disabling it for time being.
     # pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model" >> $HOME/test_rest.log

.buildkite/scripts/hardware_ci/run-cpu-test.sh

@@ -21,8 +21,8 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --progress plain --tag cpu-test-"$NUMA_NODE" --target vllm-test -f docker/Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --progress plain --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$NUMA_NODE"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$NUMA_NODE" --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$NUMA_NODE"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
 
 # Run the image, setting --shm-size=4g for tensor parallel.
 docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=16 --env VLLM_CPU_CI_ENV=1 -e E2E_OMP_THREADS="$OMP_CORE_RANGE" --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
@@ -49,7 +49,6 @@ function cpu_tests() {
   # Run kernel tests
   docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
-    pytest -x -v -s tests/kernels/attention/test_cpu_attn.py
     pytest -x -v -s tests/kernels/test_onednn.py"
 
   # Run basic model test
@@ -73,11 +72,12 @@ function cpu_tests() {
     pytest -x -s -v \
     tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs"
 
-  # Run AWQ/GPTQ test
-  docker exec cpu-test-"$NUMA_NODE" bash -c "
-    set -e
-    pytest -x -s -v \
-    tests/quantization/test_cpu_wna16.py"
+  # Note: disable it until supports V1
+  # Run AWQ test
+  # docker exec cpu-test-"$NUMA_NODE" bash -c "
+  #   set -e
+  #   VLLM_USE_V1=0 pytest -x -s -v \
+  #   tests/quantization/test_ipex_quant.py"
 
   # Run multi-lora tests
   docker exec cpu-test-"$NUMA_NODE" bash -c "
@@ -116,4 +116,4 @@ function cpu_tests() {
 
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 2.5h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"
+timeout 2h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"

.buildkite/scripts/hardware_ci/run-xpu-test.sh

@@ -35,7 +35,7 @@ docker run \
     echo $ZE_AFFINITY_MASK
     pip install tblib==3.1.0
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -cc.cudagraph_mode=NONE
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -O.cudagraph_mode=NONE
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
     VLLM_ATTENTION_BACKEND=TRITON_ATTN python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
@@ -46,6 +46,6 @@ docker run \
     pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
     pytest -v -s v1/structured_output
     pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_tree_attention.py --ignore=v1/spec_decode/test_speculators_eagle3.py
-    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_shared_storage_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py
+    pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_nixl_connector.py --ignore=v1/kv_connector/unit/test_shared_storage_connector.py
     pytest -v -s v1/test_serial_utils.py
 '

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh

@@ -17,17 +17,7 @@ wait_for_server() {
 }
 
 MODEL="deepseek-ai/DeepSeek-V2-lite"
-
-# Set BACKENDS based on platform
-if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then
-  # ROCm platform
-  BACKENDS=("allgather_reducescatter")
-  # Disable MOE padding for ROCm since it is causing eplb to fail
-  export VLLM_ROCM_MOE_PADDING=0
-else
-  # Non-ROCm platform (CUDA/other)
-  BACKENDS=("deepep_high_throughput" "deepep_low_latency")
-fi
+BACKENDS=("deepep_high_throughput" "deepep_low_latency")
 
 cleanup() {
   if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then

.buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh

@@ -1,12 +1,10 @@
 #!/usr/bin/env bash
 set -euxo pipefail
 
-# args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT] [DATA_PARALLEL_SIZE] [TENSOR_PARALLEL_SIZE]
+# args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT]
 THRESHOLD=${1:-0.8}
 NUM_Q=${2:-1319}
 PORT=${3:-8020}
-DATA_PARALLEL_SIZE=${4:-2}
-TENSOR_PARALLEL_SIZE=${5:-2}
 OUT_DIR=${OUT_DIR:-/tmp/vllm-scheduled}
 mkdir -p "${OUT_DIR}"
 
@@ -19,16 +17,7 @@ wait_for_server() {
 }
 
 MODEL="QWen/Qwen3-30B-A3B-FP8"
-# Set BACKENDS based on platform
-if command -v rocm-smi &> /dev/null || [[ -d /opt/rocm ]] || [[ -n "${ROCM_PATH:-}" ]]; then
-  # ROCm platform
-  BACKENDS=("allgather_reducescatter")
-  # Disable MOE padding for ROCm since it is causing eplb to fail
-  export VLLM_ROCM_MOE_PADDING=0
-else
-  # Non-ROCm platform (CUDA/other)
-  BACKENDS=("deepep_high_throughput" "deepep_low_latency")
-fi
+BACKENDS=("deepep_high_throughput" "deepep_low_latency")
 
 cleanup() {
   if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
@@ -47,10 +36,8 @@ for BACK in "${BACKENDS[@]}"; do
   VLLM_ALL2ALL_BACKEND=$BACK \
   vllm serve "$MODEL" \
     --enforce-eager \
-    --enable-eplb \
-    --eplb-config '{"window_size":10, "step_interval":100, "num_redundant_experts":0, "log_balancedness":true}' \
-    --tensor-parallel-size ${TENSOR_PARALLEL_SIZE} \
-    --data-parallel-size ${DATA_PARALLEL_SIZE} \
+    --tensor-parallel-size 2 \
+    --data-parallel-size 2 \
     --enable-expert-parallel \
     --trust-remote-code \
     --max-model-len 2048 \

.buildkite/test-amd.yaml

@@ -48,8 +48,8 @@ steps:
   commands:
   - bash standalone_tests/pytorch_nightly_dependency.sh
 
-- label: Async Engine, Inputs, Utils, Worker Test # 10min
-  timeout_in_minutes: 15
+- label: Async Engine, Inputs, Utils, Worker Test # 36min
+  timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
@@ -61,8 +61,8 @@ steps:
   - pytest -v -s -m 'not cpu_test' multimodal
   - pytest -v -s utils_
 
-- label: Async Engine, Inputs, Utils, Worker, Config Test (CPU) # 15min
-  timeout_in_minutes: 20
+- label: Async Engine, Inputs, Utils, Worker Test (CPU) # 4 mins
+  timeout_in_minutes: 10
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
@@ -72,18 +72,14 @@ steps:
   - tests/test_outputs.py
   - tests/multimodal
   - tests/standalone_tests/lazy_imports.py
-  - tests/tokenizers_
   - tests/transformers_utils
-  - tests/config
   no_gpu: true
   commands:
   - python3 standalone_tests/lazy_imports.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s -m 'cpu_test' multimodal
-  - pytest -v -s tokenizers_
   - pytest -v -s transformers_utils
-  - pytest -v -s config
 
 - label: Python-only Installation Test # 10min
   timeout_in_minutes: 20
@@ -191,7 +187,7 @@ steps:
   - tests/distributed/test_utils
   - tests/distributed/test_pynccl
   - tests/distributed/test_events
-  - tests/compile/fullgraph/test_basic_correctness.py
+  - tests/compile/test_basic_correctness
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
@@ -219,7 +215,7 @@ steps:
   - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py
   - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
   - pytest -v -s distributed/test_utils.py
-  - pytest -v -s compile/fullgraph/test_basic_correctness.py
+  - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s distributed/test_events.py
   - pytest -v -s distributed/test_symm_mem_allreduce.py
@@ -230,27 +226,6 @@ steps:
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
   - popd
 
-- label: Distributed Tests (8 GPUs) # 4min
-  timeout_in_minutes: 10
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_8
-  # grade: Blocking
-  gpu: h100
-  num_gpus: 8
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - examples/offline_inference/torchrun_dp_example.py
-  - vllm/config/parallel.py
-  - vllm/distributed/
-  - vllm/v1/engine/llm_engine.py
-  - vllm/v1/executor/uniproc_executor.py
-  - vllm/v1/worker/gpu_worker.py
-  commands:
-  # https://github.com/NVIDIA/nccl/issues/1838
-  #- export NCCL_CUMEM_HOST_ENABLE=0
-  # test with torchrun tp=2 and dp=4 with ep
-  - torchrun --nproc-per-node=8 ../examples/offline_inference/torchrun_dp_example.py --tp-size=2 --pp-size=1 --dp-size=4 --enable-ep
-
 - label: EPLB Algorithm Test # 5min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
@@ -263,11 +238,11 @@ steps:
   commands:
   - pytest -v -s distributed/test_eplb_algo.py
 
-- label: EPLB Execution Test # 10min
+- label: EPLB Execution Test # 5min
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_4
   # grade: Blocking
-  timeout_in_minutes: 20
+  timeout_in_minutes: 15
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -275,7 +250,6 @@ steps:
   - tests/distributed/test_eplb_execute.py
   commands:
   - pytest -v -s distributed/test_eplb_execute.py
-  - pytest -v -s distributed/test_eplb_spec_decode.py
 
 - label: Metrics, Tracing Test # 12min
   timeout_in_minutes: 20
@@ -299,7 +273,7 @@ steps:
 
 - label: Regression Test # 7min
   timeout_in_minutes: 20
-  mirror_hardwares: [amdexperimental, amdproduction, amdtentative]
+  mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   grade: Blocking
   source_file_dependencies:
@@ -310,20 +284,23 @@ steps:
   - pytest -v -s test_regression.py
   working_dir: "/vllm-workspace/tests" # optional
 
-- label: Engine Test # 9min
-  timeout_in_minutes: 15
+- label: Engine Test # 25min
+  timeout_in_minutes: 40
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
-  # grade: Blocking
+  #grade: Blocking
   source_file_dependencies:
   - vllm/
   - tests/engine
+  - tests/tokenization
   - tests/test_sequence
   - tests/test_config
   - tests/test_logger
   - tests/test_vllm_port
   commands:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
+  # OOM in the CI unless we run this separately
+  - pytest -v -s tokenization
 
 - label: V1 Test e2e + engine # 30min
   timeout_in_minutes: 45
@@ -360,7 +337,6 @@ steps:
     - tests/v1
   commands:
     # split the test to avoid interference
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
     - pytest -v -s -m 'not cpu_test' v1/core
     - pytest -v -s v1/executor
     - pytest -v -s v1/kv_offload
@@ -372,34 +348,10 @@ steps:
     - pytest -v -s -m 'not cpu_test' v1/metrics
     - pytest -v -s v1/test_oracle.py
     - pytest -v -s v1/test_request.py
-    - pytest -v -s v1/test_outputs.py
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
-# TODO: Add the "V1 Test attetion (MI300)" test group
-
-- label: V1 Test attention (H100) # 10min
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_1
-  # grade: Blocking
-  timeout_in_minutes: 30
-  gpu: h100
-  source_file_dependencies:
-    - vllm/v1/attention
-    - tests/v1/attention
-  commands:
-    - pytest -v -s v1/attention
-
-- label: V1 Test attention (B200) # 10min
-  timeout_in_minutes: 30
-  gpu: b200
-  source_file_dependencies:
-    - vllm/v1/attention
-    - tests/v1/attention
-  commands:
-    - VLLM_DISABLE_FLASHINFER_PREFILL=1 pytest -v -s v1/attention # TODO: FI prefill is bugged and causes incorrectness, fix this
-
 - label: V1 Test others (CPU) # 5 mins
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
@@ -503,12 +455,17 @@ steps:
     - vllm/
     - tests/compile
   commands:
-  # Run unit tests defined directly under compile/,
-  # not including subdirectories, which are usually heavier
-  # tests covered elsewhere.
-  # Use `find` to launch multiple instances of pytest so that
-  # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
-  - "find compile/ -maxdepth 1 -name 'test_*.py' -exec pytest -s -v {} \\\\;"
+    - pytest -v -s compile/test_pass_manager.py
+    - pytest -v -s compile/test_fusion.py
+    - pytest -v -s compile/test_fusion_attn.py
+    - pytest -v -s compile/test_functionalization.py
+    - pytest -v -s compile/test_silu_mul_quant_fusion.py
+  #  - pytest -v -s compile/test_sequence_parallelism.py
+  #  - pytest -v -s compile/test_async_tp.py
+    - pytest -v -s compile/test_fusion_all_reduce.py
+    - pytest -v -s compile/test_decorator.py
+    - pytest -v -s compile/test_noop_elimination.py
+    - pytest -v -s compile/test_aot_compile.py
 
 - label: PyTorch Fullgraph Smoke Test # 15min
   timeout_in_minutes: 30
@@ -520,14 +477,11 @@ steps:
   - vllm/
   - tests/compile
   commands:
-  # Run smoke tests under fullgraph directory, except test_full_graph.py
-  # as it is a heavy test that is covered in other steps.
-  # Use `find` to launch multiple instances of pytest so that
-  # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
-  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -exec pytest -s -v {} \\\\;"
+  - pytest -v -s compile/test_basic_correctness.py
+  - pytest -v -s compile/piecewise/
 
-- label: PyTorch Fullgraph Test # 27min
-  timeout_in_minutes: 40
+- label: PyTorch Fullgraph Test # 22min
+  timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
@@ -536,23 +490,8 @@ steps:
   - vllm/
   - tests/compile
   commands:
-  - pytest -v -s compile/fullgraph/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
-    # Limit to no custom ops to reduce running time
-    # Wrap with quotes to escape yaml and avoid starting -k string with a -
-  - "pytest -v -s compile/distributed/test_fusions_e2e.py -k 'TRITON and not +quant_fp8 and not Llama-4'"
-
-- label: Cudagraph test
-  timeout_in_minutes: 20
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_1
-  source_file_dependencies:
-  - tests/v1/cudagraph
-  - vllm/v1/cudagraph_dispatcher.py
-  - vllm/config/compilation.py
-  - vllm/compilation
-  commands:
-    - pytest -v -s v1/cudagraph/test_cudagraph_dispatch.py
-    - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
+  - pytest -v -s compile/test_full_graph.py
+  - pytest -v -s compile/test_fusions_e2e.py
 
 - label: Kernels Core Operation Test # 48min
   timeout_in_minutes: 75
@@ -604,8 +543,6 @@ steps:
   - tests/kernels/moe
   - vllm/model_executor/layers/fused_moe/
   - vllm/distributed/device_communicators/
-  - vllm/envs.py
-  - vllm/config
   commands:
     - pytest -v -s kernels/moe --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 2
@@ -624,13 +561,10 @@ steps:
 
 - label: Model Executor Test # 23min
   timeout_in_minutes: 35
-  torch_nightly: true
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
-  - vllm/engine/arg_utils.py
-  - vllm/config/model.py
   - vllm/model_executor
   - tests/model_executor
   - tests/entrypoints/openai/test_tensorizer_entrypoint.py
@@ -682,9 +616,9 @@ steps:
   - uv pip install --system torchao==0.13.0
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
-- label: LM Eval Small Models # 15min
-  timeout_in_minutes: 20
-  mirror_hardwares: [amdexperimental, amdproduction]
+- label: LM Eval Small Models # 53min
+  timeout_in_minutes: 75
+  mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
   # grade: Blocking
   source_file_dependencies:
@@ -693,8 +627,8 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
 
-- label: OpenAI API correctness # 10min
-  timeout_in_minutes: 15
+- label: OpenAI API correctness # 22min
+  timeout_in_minutes: 30
   mirror_hardwares: [amdexperimental, amdproduction]
   agent_pool: mi325_1
   # grade: Blocking
@@ -704,7 +638,7 @@ steps:
   - vllm/model_executor/models/whisper.py
   commands: # LMEval
   # Transcription WER check is skipped because encoder-decoder models are not supported on ROCm, see https://github.com/vllm-project/vllm/issues/27442
-  - pytest -s entrypoints/openai/correctness/
+  - pytest -s entrypoints/openai/correctness/  --ignore entrypoints/openai/correctness/test_transcription_api_correctness.py
 
 - label: OpenAI-Compatible Tool Use # 23 min
   timeout_in_minutes: 35
@@ -753,7 +687,6 @@ steps:
   torch_nightly: true
   source_file_dependencies:
   - vllm/model_executor/models/
-  - vllm/transformers_utils/
   - tests/models/test_initialization.py
   commands:
     # Only when vLLM model source is modified - test initialization of a large
@@ -926,11 +859,10 @@ steps:
     - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
 
-- label: Multi-Modal Accuracy Eval (Small Models) # 10min
-  timeout_in_minutes: 70
-  mirror_hardwares: [amdexperimental, amdproduction]
+- label: Multi-Modal Accuracy Eval (Small Models) # 50min
+  mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
-  # grade: Blocking
+  timeout_in_minutes: 70
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
   - vllm/multimodal/
@@ -1001,17 +933,16 @@ steps:
 - label: Transformers Nightly Models Test
   mirror_hardwares: [amdexperimental]
   agent_pool: mi325_1
-  # grade: Blocking
   working_dir: "/vllm-workspace/"
   optional: true
   commands:
     - pip install --upgrade git+https://github.com/huggingface/transformers
-    - pytest -v -s tests/models/test_initialization.py -k 'not (Gemma3 or ModernBert or Qwen2_5_VL or Qwen2_5vl or Qwen2VL or TransformersMultiModalEmbeddingModel or TransformersMultiModalForSequenceClassification or Ultravox or Phi4Multimodal or LlavaNextVideo or MiniCPMO or Lfm2Moe or PaliGemma or RobertaForSequenceClassification or Ovis2_5 or Fuyu or DeepseekOCR or KimiVL)'
+    - pytest -v -s tests/models/test_initialization.py
     - pytest -v -s tests/models/test_transformers.py
-    # - pytest -v -s tests/models/multimodal/processing/
-    - pytest -v -s tests/models/multimodal/test_mapping.py -k 'not (Gemma3 or Qwen2VL or Qwen2_5_VL)'
+    - pytest -v -s tests/models/multimodal/processing/
+    - pytest -v -s tests/models/multimodal/test_mapping.py
     - python3 examples/offline_inference/basic/chat.py
-    # - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
+    - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
     # Whisper needs spawn method to avoid deadlock
     - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
 
@@ -1029,16 +960,11 @@ steps:
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
-  - vllm/v1/attention/backends/mla/cutlass_mla.py
-  - vllm/v1/attention/backends/mla/flashinfer_mla.py
-  - vllm/platforms/cuda.py
-  - vllm/attention/selector.py
   commands:
     - nvidia-smi
     - python3 examples/offline_inference/basic/chat.py
     # Attention
     # num_heads2 broken by https://github.com/flashinfer-ai/flashinfer/issues/1353
-    - pytest -v -s tests/kernels/attention/test_attention_selector.py
     - pytest -v -s tests/kernels/attention/test_flashinfer.py -k 'not num_heads2'
     - pytest -v -s tests/kernels/attention/test_flashinfer_trtllm_attention.py
     - pytest -v -s tests/kernels/attention/test_cutlass_mla_decode.py
@@ -1056,7 +982,7 @@ steps:
     - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
 
-- label: Blackwell Fusion and Compile Tests # 30 min
+- label: Blackwell Fusion Tests # 30 min
   timeout_in_minutes: 40
   working_dir: "/vllm-workspace/"
   gpu: b200
@@ -1074,40 +1000,13 @@ steps:
     - pytest -v -s tests/compile/test_fusion_attn.py
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
     # this runner has 2 GPUs available even though num_gpus=2 is not set
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
-    # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
-    # Wrap with quotes to escape yaml
-    - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
-    # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
-    - pytest -v -s tests/compile/distributed/test_full_graph.py::test_fp8_kv_scale_compile
-
-- label: Blackwell Fusion E2E Tests # 30 min
-  timeout_in_minutes: 40
-  working_dir: "/vllm-workspace/"
-  gpu: b200
-  optional: true
-  num_gpus: 2
-  source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
-  - vllm/compilation/
-  # can affect pattern matching
-  - vllm/model_executor/layers/layernorm.py
-  - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/distributed/test_fusions_e2e.py
-  - tests/compile/fullgraph/test_full_graph.py
-  commands:
-    - nvidia-smi
-    # Run all e2e fusion tests
+    - pytest -v -s tests/compile/test_fusion_all_reduce.py
     - pytest -v -s tests/compile/test_fusions_e2e.py
 
-- label: ROCm GPT-OSS Eval
+- label: Blackwell GPT-OSS Eval
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/"
-  agent_pool: mi325_1
-  mirror_hardwares: [amdexperimental, amdproduction]
+  gpu: b200
   optional: true # run on nightlies
   source_file_dependencies:
   - tests/evals/gpt_oss
@@ -1116,7 +1015,7 @@ steps:
   - vllm/v1/attention/backends/flashinfer.py
   commands:
     - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - VLLM_ROCM_USE_AITER_MHA=0 VLLM_ROCM_USE_AITER=1 VLLM_USE_AITER_UNIFIED_ATTENTION=1 pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
+    - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58
 
 - label: Blackwell Quantized MoE Test
   timeout_in_minutes: 60
@@ -1206,7 +1105,7 @@ steps:
   - vllm/worker/worker_base.py
   - vllm/v1/engine/
   - vllm/v1/worker/
-  - tests/compile/fullgraph/test_basic_correctness.py
+  - tests/compile/test_basic_correctness.py
   - tests/compile/test_wrapper.py
   - tests/distributed/
   - tests/entrypoints/llm/test_collective_rpc.py
@@ -1219,7 +1118,7 @@ steps:
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
   - DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py
   - pytest -v -s entrypoints/llm/test_collective_rpc.py
-  - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
+  - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
@@ -1319,10 +1218,7 @@ steps:
     - pytest -v -s -x lora/test_llama_tp.py
     - pytest -v -s -x lora/test_llm_with_multi_loras.py
     - pytest -v -s -x lora/test_olmoe_tp.py
-
-    # Disabled for now because MXFP4 backend on non-cuda platform 
-    # doesn't support LoRA yet
-    #- pytest -v -s -x lora/test_gptoss_tp.py
+    - pytest -v -s -x lora/test_gptoss_tp.py
 
 
 - label: Weight Loading Multiple GPU Test  # 33min
@@ -1337,7 +1233,7 @@ steps:
   - vllm/
   - tests/weight_loading
   commands:
-    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-amd.txt
+    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
 
 - label: Weight Loading Multiple GPU Test - Large Models # optional
   mirror_hardwares: [amdexperimental]
@@ -1345,17 +1241,17 @@ steps:
   # grade: Blocking
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
+  gpu: a100
   optional: true
   source_file_dependencies:
   - vllm/
   - tests/weight_loading
   commands:
-    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large-amd.txt
+    - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt
 
 - label: NixlConnector PD accuracy tests (Distributed) # 30min
   mirror_hardwares: [amdexperimental]
   agent_pool: mi325_4
-  # grade: Blocking
   timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
@@ -1370,9 +1266,6 @@ steps:
 ##### A100 test #####
 
 - label: Distributed Tests (A100) # optional
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_4
-  # grade: Blocking
   gpu: a100
   optional: true
   num_gpus: 4
@@ -1387,9 +1280,6 @@ steps:
   - pytest -v -s -x lora/test_mixtral.py
 
 - label: LM Eval Large Models # optional
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_4
-  # grade: Blocking
   gpu: a100
   optional: true
   num_gpus: 4
@@ -1401,41 +1291,19 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
-##### H100 test #####
-- label: LM Eval Large Models (H100) # optional
-  mirror_hardwares: [amdexperimental, amdproduction]
-  agent_pool: mi325_4
-  # grade: Blocking
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
-  source_file_dependencies:
-  - csrc/
-  - vllm/model_executor/layers/quantization
-  commands:
-    - export VLLM_USE_DEEP_GEMM=0  # We found Triton is faster than DeepGEMM for H100
-    - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
-
 ##### H200 test #####
 - label: Distributed Tests (H200) # optional
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_2
-  # grade: Blocking
   gpu: h200
   optional: true
   working_dir: "/vllm-workspace/"
   num_gpus: 2
   commands:
-    - pytest -v -s tests/compile/distributed/test_async_tp.py
-    - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
-    #- pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
-    - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
-    - pytest -v -s tests/compile/distributed/test_sequence_parallel.py
+    - pytest -v -s tests/compile/test_async_tp.py
+    - pytest -v -s tests/compile/test_sequence_parallelism.py
+    - pytest -v -s tests/compile/test_fusion_all_reduce.py
+    - pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
     - pytest -v -s tests/distributed/test_context_parallel.py
     - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
-    - pytest -v -s tests/v1/distributed/test_dbo.py
 
 ##### B200 test #####
 - label: Distributed Tests (B200) # optional
@@ -1446,7 +1314,6 @@ steps:
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
     - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
-    - pytest -v -s tests/v1/distributed/test_dbo.py
 
 ##### RL Integration Tests #####
 - label: Prime-RL Integration Test # 15min
@@ -1462,27 +1329,3 @@ steps:
   - .buildkite/scripts/run-prime-rl-test.sh
   commands:
     - bash .buildkite/scripts/run-prime-rl-test.sh
-
-- label: DeepSeek V2-Lite Accuracy
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_4
-  # grade: Blocking
-  timeout_in_minutes: 60
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
-
-- label: Qwen3-30B-A3B-FP8-block Accuracy
-  mirror_hardwares: [amdexperimental]
-  agent_pool: mi325_4
-  # grade: Blocking
-  timeout_in_minutes: 60
-  gpu: h100
-  optional: true
-  num_gpus: 4
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020

.buildkite/test-pipeline.yaml

@@ -25,7 +25,6 @@
 #     and $$BUILDKITE_PARALLEL_JOB_COUNT environment variables.
 # working_dir(str): specify the place where the command should execute, default to /vllm-workspace/tests
 # source_file_dependencies(list): the list of prefixes to opt-in the test for, if empty, the test will always run.
-# autorun_on_main (bool): default to false, if true, the test will run automatically when commit is pushed to main branch.
 
 # When adding a test
 # - If the test belongs to an existing group, add it there
@@ -57,26 +56,22 @@ steps:
   - pytest -v -s -m 'not cpu_test' multimodal
   - pytest -v -s utils_
 
-- label: Async Engine, Inputs, Utils, Worker, Config Test (CPU) # 15min
-  timeout_in_minutes: 20
+- label: Async Engine, Inputs, Utils, Worker Test (CPU) # 4 mins
+  timeout_in_minutes: 10
   source_file_dependencies:
   - vllm/
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/multimodal
   - tests/standalone_tests/lazy_imports.py
-  - tests/tokenizers_
   - tests/transformers_utils
-  - tests/config
   no_gpu: true
   commands:
   - python3 standalone_tests/lazy_imports.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s -m 'cpu_test' multimodal
-  - pytest -v -s tokenizers_
   - pytest -v -s transformers_utils
-  - pytest -v -s config
 
 - label: Python-only Installation Test # 10min
   timeout_in_minutes: 20
@@ -169,7 +164,7 @@ steps:
   - tests/distributed/test_utils
   - tests/distributed/test_pynccl
   - tests/distributed/test_events
-  - tests/compile/fullgraph/test_basic_correctness.py
+  - tests/compile/test_basic_correctness
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
@@ -194,13 +189,12 @@ steps:
   # test with internal dp
   - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
-  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
   - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_internal_lb_dp.py
   - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py
   - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
   - pytest -v -s distributed/test_utils.py
-  - pytest -v -s compile/fullgraph/test_basic_correctness.py
+  - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s distributed/test_events.py
   - pytest -v -s distributed/test_symm_mem_allreduce.py
@@ -215,7 +209,6 @@ steps:
   timeout_in_minutes: 10
   gpu: h100
   num_gpus: 8
-  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - examples/offline_inference/torchrun_dp_example.py
@@ -239,8 +232,8 @@ steps:
   commands:
   - pytest -v -s distributed/test_eplb_algo.py
 
-- label: EPLB Execution Test # 10min
-  timeout_in_minutes: 20
+- label: EPLB Execution Test # 5min
+  timeout_in_minutes: 15
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -248,7 +241,6 @@ steps:
   - tests/distributed/test_eplb_execute.py
   commands:
   - pytest -v -s distributed/test_eplb_execute.py
-  - pytest -v -s distributed/test_eplb_spec_decode.py
 
 - label: Metrics, Tracing Test # 12min
   timeout_in_minutes: 20
@@ -279,18 +271,21 @@ steps:
   - pytest -v -s test_regression.py
   working_dir: "/vllm-workspace/tests" # optional
 
-- label: Engine Test # 9min
-  timeout_in_minutes: 15
+- label: Engine Test # 25min
+  timeout_in_minutes: 40
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
   - tests/engine
+  - tests/tokenization
   - tests/test_sequence
   - tests/test_config
   - tests/test_logger
   - tests/test_vllm_port
   commands:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
+  # OOM in the CI unless we run this separately
+  - pytest -v -s tokenization
 
 - label: V1 Test e2e + engine # 30min
   timeout_in_minutes: 45
@@ -320,7 +315,6 @@ steps:
     - vllm/
     - tests/v1
   commands:
-    - uv pip install --system -r /vllm-workspace/requirements/kv_connectors.txt
     # split the test to avoid interference
     - pytest -v -s -m 'not cpu_test' v1/core
     - pytest -v -s v1/executor
@@ -333,7 +327,6 @@ steps:
     - pytest -v -s -m 'not cpu_test' v1/metrics
     - pytest -v -s v1/test_oracle.py
     - pytest -v -s v1/test_request.py
-    - pytest -v -s v1/test_outputs.py
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
@@ -347,18 +340,6 @@ steps:
   commands:
     - pytest -v -s v1/attention
 
-- label: Batch Invariance Tests (H100) # 10min
-  timeout_in_minutes: 25
-  gpu: h100
-  source_file_dependencies:
-    - vllm/
-    - tests/v1/determinism/
-  commands:
-    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    - pip install pytest-timeout pytest-forked
-    - pytest -v -s v1/determinism/test_batch_invariance.py
-    - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
-
 - label: V1 Test attention (B200) # 10min
   timeout_in_minutes: 30
   gpu: b200
@@ -458,12 +439,15 @@ steps:
     - vllm/
     - tests/compile
   commands:
-  # Run unit tests defined directly under compile/,
-  # not including subdirectories, which are usually heavier
-  # tests covered elsewhere.
-  # Use `find` to launch multiple instances of pytest so that
-  # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
-  - "find compile/ -maxdepth 1 -name 'test_*.py' -exec pytest -s -v {} \\\\;"
+    - pytest -v -s compile/test_pass_manager.py
+    - pytest -v -s compile/test_fusion.py
+    - pytest -v -s compile/test_fusion_attn.py
+    - pytest -v -s compile/test_functionalization.py
+    - pytest -v -s compile/test_silu_mul_quant_fusion.py
+    - pytest -v -s compile/test_fusion_all_reduce.py
+    - pytest -v -s compile/test_decorator.py
+    - pytest -v -s compile/test_noop_elimination.py
+    - pytest -v -s compile/test_aot_compile.py
 
 - label: PyTorch Fullgraph Smoke Test # 15min
   timeout_in_minutes: 30
@@ -473,25 +457,19 @@ steps:
   - vllm/
   - tests/compile
   commands:
-  # Run smoke tests under fullgraph directory, except test_full_graph.py
-  # as it is a heavy test that is covered in other steps.
-  # Use `find` to launch multiple instances of pytest so that
-  # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
-  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -exec pytest -s -v {} \\\\;"
+  - pytest -v -s compile/test_basic_correctness.py
+  - pytest -v -s compile/piecewise/
 
-- label: PyTorch Fullgraph Test # 27min
-  timeout_in_minutes: 40
+- label: PyTorch Fullgraph Test # 22min
+  timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
   - tests/compile
   commands:
-    # fp8 kv scales not supported on sm89, tested on Blackwell instead
-  - pytest -v -s compile/fullgraph/test_full_graph.py -k 'not test_fp8_kv_scale_compile'
-    # Limit to no custom ops to reduce running time
-    # Wrap with quotes to escape yaml and avoid starting -k string with a -
-  - "pytest -v -s compile/distributed/test_fusions_e2e.py -k 'TRITON and not +quant_fp8 and not Llama-4'"
+  - pytest -v -s compile/test_full_graph.py
+  - pytest -v -s compile/test_fusions_e2e.py
 
 - label: Cudagraph test
   timeout_in_minutes: 20
@@ -563,32 +541,10 @@ steps:
   commands:
     - pytest -v -s kernels/mamba
 
-- label: Kernels DeepGEMM Test (H100)
-  timeout_in_minutes: 45
-  gpu: h100
-  num_gpus: 1
-  source_file_dependencies:
-  - tools/install_deepgemm.sh
-  - vllm/utils/deep_gemm.py
-  - vllm/model_executor/layers/fused_moe
-  - vllm/model_executor/layers/quantization
-  - tests/kernels/quantization/test_block_fp8.py
-  - tests/kernels/moe/test_deepgemm.py
-  - tests/kernels/moe/test_batched_deepgemm.py
-  - tests/kernels/attention/test_deepgemm_attention.py
-  commands:
-    - pytest -v -s kernels/quantization/test_block_fp8.py -k deep_gemm
-    - pytest -v -s kernels/moe/test_deepgemm.py
-    - pytest -v -s kernels/moe/test_batched_deepgemm.py
-    - pytest -v -s kernels/attention/test_deepgemm_attention.py
-
 - label: Model Executor Test # 23min
   timeout_in_minutes: 35
-  torch_nightly: true
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
-  - vllm/engine/arg_utils.py
-  - vllm/config/model.py
   - vllm/model_executor
   - tests/model_executor
   - tests/entrypoints/openai/test_tensorizer_entrypoint.py
@@ -632,7 +588,6 @@ steps:
   # we can only upgrade after this is resolved
   # TODO(jerryzh168): resolve the above comment
   - uv pip install --system torchao==0.13.0 --index-url https://download.pytorch.org/whl/cu129
-  - uv pip install --system conch-triton-kernels
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
 - label: LM Eval Small Models # 53min
@@ -641,7 +596,6 @@ steps:
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
-  autorun_on_main: true
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt --tp-size=1
 
@@ -693,7 +647,6 @@ steps:
   torch_nightly: true
   source_file_dependencies:
   - vllm/model_executor/models/
-  - vllm/transformers_utils/
   - tests/models/test_initialization.py
   commands:
     # Only when vLLM model source is modified - test initialization of a large
@@ -820,24 +773,14 @@ steps:
   commands:
     - pytest -v -s models/language/pooling_mteb_test
 
-- label: Multi-Modal Processor Test (CPU)
-  timeout_in_minutes: 60
-  source_file_dependencies:
-  - vllm/
-  - tests/models/multimodal
-  no_gpu: true
-  commands:
-    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py
-
-- label: Multi-Modal Processor Test
+- label: Multi-Modal Processor Test # 44min
   timeout_in_minutes: 60
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/multimodal/processing/test_tensor_schema.py
+    - pytest -v -s models/multimodal/processing
 
 - label: Multi-Modal Models Test (Standard) # 60min
   timeout_in_minutes: 80
@@ -914,7 +857,6 @@ steps:
 - label: Transformers Nightly Models Test
   working_dir: "/vllm-workspace/"
   optional: true
-  soft_fail: true
   commands:
     - pip install --upgrade git+https://github.com/huggingface/transformers
     - pytest -v -s tests/models/test_initialization.py
@@ -940,16 +882,11 @@ steps:
   - vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
-  - vllm/v1/attention/backends/mla/cutlass_mla.py
-  - vllm/v1/attention/backends/mla/flashinfer_mla.py
-  - vllm/platforms/cuda.py
-  - vllm/attention/selector.py
   commands:
     - nvidia-smi
     - python3 examples/offline_inference/basic/chat.py
     # Attention
     # num_heads2 broken by https://github.com/flashinfer-ai/flashinfer/issues/1353
-    - pytest -v -s tests/kernels/attention/test_attention_selector.py
     - pytest -v -s tests/kernels/attention/test_flashinfer.py -k 'not num_heads2'
     - pytest -v -s tests/kernels/attention/test_flashinfer_trtllm_attention.py
     - pytest -v -s tests/kernels/attention/test_cutlass_mla_decode.py
@@ -966,9 +903,8 @@ steps:
     - pytest -v -s tests/kernels/moe/test_nvfp4_moe.py
     - pytest -v -s tests/kernels/moe/test_ocp_mx_moe.py
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
-    - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
 
-- label: Blackwell Fusion and Compile Tests # 30 min
+- label: Blackwell Fusion Tests # 30 min
   timeout_in_minutes: 40
   working_dir: "/vllm-workspace/"
   gpu: b200
@@ -976,50 +912,18 @@ steps:
   - csrc/quantization/fp4/
   - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
   - vllm/v1/attention/backends/flashinfer.py
-  - vllm/v1/worker/
-  - vllm/v1/cudagraph_dispatcher.py
   - vllm/compilation/
   # can affect pattern matching
   - vllm/model_executor/layers/layernorm.py
   - vllm/model_executor/layers/activation.py
   - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/test_fusion_attn.py
-  - tests/compile/test_silu_mul_quant_fusion.py
-  - tests/compile/distributed/test_fusion_all_reduce.py
-  - tests/compile/distributed/test_fusions_e2e.py
-  - tests/compile/fullgraph/test_full_graph.py
   commands:
     - nvidia-smi
     - pytest -v -s tests/compile/test_fusion_attn.py
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
     # this runner has 2 GPUs available even though num_gpus=2 is not set
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
-    # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
-    # Wrap with quotes to escape yaml
-    - "pytest -v -s tests/compile/distributed/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and not +quant_fp8 and not +rms_norm'"
-    # test_fp8_kv_scale_compile requires FlashAttention (not supported on default L4/L40)
-    - pytest -v -s tests/compile/fullgraph/test_full_graph.py::test_fp8_kv_scale_compile
-
-- label: Blackwell Fusion E2E Tests # 30 min
-  timeout_in_minutes: 40
-  working_dir: "/vllm-workspace/"
-  gpu: b200
-  optional: true
-  num_gpus: 2
-  source_file_dependencies:
-  - csrc/quantization/fp4/
-  - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
-  - vllm/v1/attention/backends/flashinfer.py
-  - vllm/compilation/
-  # can affect pattern matching
-  - vllm/model_executor/layers/layernorm.py
-  - vllm/model_executor/layers/activation.py
-  - vllm/model_executor/layers/quantization/input_quant_fp8.py
-  - tests/compile/distributed/test_fusions_e2e.py
-  commands:
-    - nvidia-smi
-    # Run all e2e fusion tests
-    - pytest -v -s tests/compile/distributed/test_fusions_e2e.py
+    - pytest -v -s tests/compile/test_fusion_all_reduce.py
+    - pytest -v -s tests/compile/test_fusions_e2e.py
 
 - label: Blackwell GPT-OSS Eval
   timeout_in_minutes: 60
@@ -1117,7 +1021,7 @@ steps:
   - vllm/worker/worker_base.py
   - vllm/v1/engine/
   - vllm/v1/worker/
-  - tests/compile/fullgraph/test_basic_correctness.py
+  - tests/compile/test_basic_correctness.py
   - tests/compile/test_wrapper.py
   - tests/distributed/
   - tests/entrypoints/llm/test_collective_rpc.py
@@ -1129,11 +1033,10 @@ steps:
   # https://github.com/NVIDIA/nccl/issues/1838
   - export NCCL_CUMEM_HOST_ENABLE=0
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
-  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
   - DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py
   - pytest -v -s entrypoints/llm/test_collective_rpc.py
-  - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
+  - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
   - VLLM_TEST_SAME_HOST=1 VLLM_TEST_WITH_DEFAULT_DEVICE_SET=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
@@ -1313,14 +1216,13 @@ steps:
   working_dir: "/vllm-workspace/"
   num_gpus: 2
   commands:
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_async_tp.py
-    - pytest -v -s tests/compile/distributed/test_sequence_parallelism.py
-    - pytest -v -s tests/compile/distributed/test_fusion_all_reduce.py
-    - "VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/compile/distributed/test_fusions_e2e.py -k 'not Llama-4'"
-    - VLLM_TEST_CLEAN_GPU_MEMORY=1 pytest -v -s tests/distributed/test_sequence_parallel.py
+    - pytest -v -s tests/compile/test_async_tp.py
+    - pytest -v -s tests/compile/test_sequence_parallelism.py
+    - pytest -v -s tests/compile/test_fusion_all_reduce.py
+    - pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm
     - pytest -v -s tests/distributed/test_context_parallel.py
     - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
-    - pytest -v -s tests/v1/distributed/test_dbo.py
+    - pytest -v -s tests/v1/distributed/test_dbo.py  
 
 ##### B200 test #####
 - label: Distributed Tests (B200) # optional
@@ -1354,20 +1256,11 @@ steps:
   commands:
   - bash .buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_ep_eplb.sh 0.25 200 8010
 
-- label: Qwen3-30B-A3B-FP8-block Accuracy (H100)
+- label: Qwen3-30B-A3B-FP8-block Accuracy
   timeout_in_minutes: 60
   gpu: h100
   optional: true
   num_gpus: 4
   working_dir: "/vllm-workspace"
   commands:
-  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020
-
-- label: Qwen3-30B-A3B-FP8-block Accuracy (B200)
-  timeout_in_minutes: 60
-  gpu: b200
-  optional: true
-  num_gpus: 2
-  working_dir: "/vllm-workspace"
-  commands:
-  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep_eplb.sh 0.8 200 8020 2 1
+  - bash .buildkite/scripts/scheduled_integration_test/qwen30b_a3b_fp8_block_ep.sh 0.8 200 8020

.github/CODEOWNERS

@@ -3,14 +3,13 @@
 
 # This lists cover the "core" components of vLLM that require careful review
 /vllm/attention @LucasWilkinson
-/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @njhill
-/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @njhill @22quinn
+/vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @22quinn
 /vllm/model_executor/layers/fused_moe @mgoin @pavanimajety
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256 @pavanimajety
 /vllm/model_executor/layers/mamba @tdoublep
 /vllm/model_executor/model_loader @22quinn
-/vllm/model_executor/layers/batch_invariant.py @yewentao256 
-/vllm/multimodal @DarkLight1337 @ywang96 @NickLucche @tjtanaa
+/vllm/multimodal @DarkLight1337 @ywang96 @NickLucche
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
 /vllm/reasoning @aarnphm @chaunceyjiang
@@ -21,30 +20,27 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
-/vllm/config @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
-/vllm/config/cache.py @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg @heheda12345
+/vllm/config @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
+/vllm/config/cache.py @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg @heheda12345
 
 # vLLM V1
 /vllm/v1/attention @LucasWilkinson
 /vllm/v1/attention/backends/mla @pavanimajety
 /vllm/v1/attention/backends/flashinfer.py @mgoin @pavanimajety
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
-/vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC
+/vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat @heheda12345 @ApostaC
 /vllm/v1/sample @22quinn @houseroad @njhill
 /vllm/v1/spec_decode @benchislett @luccafong
 /vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
 /vllm/v1/kv_cache_interface.py @heheda12345
 /vllm/v1/offloading @ApostaC
 
-# Model runner V2
-/vllm/v1/worker/gpu @WoosukKwon
-
 # Test ownership
-/.buildkite/lm-eval-harness @mgoin 
+/.buildkite/lm-eval-harness @mgoin @simon-mo
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
-/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @aarnphm @NickLucche
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm @NickLucche
 /tests/evals @mgoin
 /tests/kernels @mgoin @tlrmchlsmth @WoosukKwon @yewentao256
 /tests/models @DarkLight1337 @ywang96
@@ -53,29 +49,18 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm
-/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC
+/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat @heheda12345 @ApostaC
 /tests/weight_loading @mgoin @youkaichao @yewentao256
 /tests/lora @jeejeelee
 /tests/models/language/generation/test_hybrid.py @tdoublep
 /tests/v1/kv_connector/nixl_integration @NickLucche
 /tests/v1/kv_connector @ApostaC
 /tests/v1/offloading @ApostaC
-/tests/v1/determinism @yewentao256 
 
-# Transformers modeling backend
+# Transformers backend
 /vllm/model_executor/models/transformers @hmellor
 /tests/models/test_transformers.py @hmellor
 
-# Observability
-/vllm/config/observability.py @markmc
-/vllm/v1/metrics @markmc
-/tests/v1/metrics @markmc
-/vllm/tracing.py @markmc
-/tests/v1/tracing/test_tracing.py @markmc
-/vllm/config/kv_events.py @markmc
-/vllm/distributed/kv_events.py @markmc
-/tests/distributed/test_events.py @markmc
-
 # Docs
 /docs/mkdocs @hmellor
 /docs/**/*.yml @hmellor
@@ -120,21 +105,11 @@ mkdocs.yaml @hmellor
 /vllm/attention/ops/triton_unified_attention.py @tdoublep
 
 # ROCm related: specify owner with write access to notify AMD folks for careful code review
-/vllm/**/*rocm* @tjtanaa
-/docker/Dockerfile.rocm* @gshtras @tjtanaa
-/vllm/v1/attention/backends/rocm*.py @gshtras @tjtanaa
-/vllm/v1/attention/backends/mla/rocm*.py @gshtras @tjtanaa
-/vllm/attention/ops/rocm*.py @gshtras @tjtanaa
-/vllm/model_executor/layers/fused_moe/rocm*.py @gshtras @tjtanaa
-/csrc/rocm @gshtras @tjtanaa
-/requirements/*rocm* @tjtanaa
-/tests/**/*rocm* @tjtanaa
-/docs/**/*rocm* @tjtanaa
-/vllm/**/*quark* @tjtanaa
-/tests/**/*quark* @tjtanaa
-/docs/**/*quark* @tjtanaa
-/vllm/**/*aiter* @tjtanaa
-/tests/**/*aiter* @tjtanaa
+/docker/Dockerfile.rocm* @gshtras
+/vllm/v1/attention/backends/rocm*.py @gshtras
+/vllm/v1/attention/backends/mla/rocm*.py @gshtras
+/vllm/attention/ops/rocm*.py @gshtras
+/vllm/model_executor/layers/fused_moe/rocm*.py @gshtras
 
 # TPU
 /vllm/v1/worker/tpu* @NickLucche
@@ -149,12 +124,6 @@ mkdocs.yaml @hmellor
 /examples/*/pooling/ @noooop
 /tests/models/*/pooling* @noooop
 /tests/entrypoints/pooling @noooop
-/vllm/entrypoints/pooling @aarnphm @chaunceyjiang @noooop
 /vllm/config/pooler.py @noooop
 /vllm/pooling_params.py @noooop
 /vllm/model_executor/layers/pooler.py @noooop
-
-# Security guide and policies
-/docs/usage/security.md @russellb
-/SECURITY.md @russellb
-/docs/contributing/vulnerability_management.md @russellb

.github/mergify.yml

@@ -151,23 +151,6 @@ pull_request_rules:
       add:
         - gpt-oss
 
-- name: label-nvidia
-  description: Automatically apply nvidia label
-  conditions:
-    - label != stale
-    - or:
-      - files~=cuda
-      - files~=cutlass
-      - files~=flashinfer
-      - files~=trtllm
-      - title~=(?i)NVIDIA
-      - title~=(?i)CUDA
-      - title~=(?i)CUTLASS
-  actions:
-    label:
-      add:
-        - nvidia
-
 - name: label-rocm
   description: Automatically apply rocm label
   conditions:

.github/workflows/cleanup_pr_body.yml

@@ -13,7 +13,7 @@ jobs:
 
     steps:
       - name: Checkout repository
-        uses: actions/checkout@1af3b93b6815bc44a9784bd300feb67ff0d1eeb3 # v6.0.0
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
 
       - name: Set up Python
         uses: actions/setup-python@e797f83bcb11b83ae66e0230d6156d7c80228e7c # v6.0.0

.github/workflows/issue_autolabel.yml

@@ -105,31 +105,6 @@ jobs:
                   }
                 ],
               },
-              cpu: {
-                // Keyword search - matches whole words only (with word boundaries)
-                keywords: [
-                  {
-                    term: "CPU Backend",
-                    searchIn: "title"
-                  },
-                  {
-                    term: "x86",
-                    searchIn: "title"
-                  },
-                  {
-                    term: "ARM",
-                    searchIn: "title"
-                  },
-                  {
-                    term: "Apple Silicon",
-                    searchIn: "title"
-                  },
-                  {
-                    term: "IBM Z",
-                    searchIn: "title"
-                  },
-                ],
-              },
               // Add more label configurations here as needed
               // example: {
               //   keywords: [...],

.github/workflows/macos-smoke-test.yml

@@ -1,80 +0,0 @@
-name: macOS Apple Silicon Smoke Test
-
-on:
-  push:
-    branches:
-      - main
-  workflow_dispatch:  # Manual trigger
-
-jobs:
-  macos-m1-smoke-test:
-    runs-on: macos-latest
-    timeout-minutes: 30
-
-    steps:
-      - uses: actions/checkout@v6
-
-      - uses: astral-sh/setup-uv@v7
-        with:
-          enable-cache: true
-          cache-dependency-glob: |
-            requirements/**/*.txt
-            pyproject.toml
-          python-version: '3.12'
-
-      - name: Create virtual environment
-        run: |
-          uv venv
-          echo "$GITHUB_WORKSPACE/.venv/bin" >> "$GITHUB_PATH"
-
-      - name: Install dependencies and build vLLM
-        run: |
-          uv pip install -r requirements/cpu.txt --index-strategy unsafe-best-match
-          uv pip install -e .
-        env:
-          CMAKE_BUILD_PARALLEL_LEVEL: 4
-
-      - name: Verify installation
-        run: |
-          python -c "import vllm; print(f'vLLM version: {vllm.__version__}')"
-
-      - name: Smoke test vllm serve
-        run: |
-          # Start server in background
-          vllm serve Qwen/Qwen3-0.6B \
-            --max-model-len=2K \
-            --load-format=dummy \
-            --hf-overrides '{"num_hidden_layers": 2}' \
-            --enforce-eager \
-            --port 8000 &
-
-          SERVER_PID=$!
-
-          # Wait for server to start
-          for i in {1..30}; do
-            if curl -s http://localhost:8000/health > /dev/null; then
-              echo "Server started successfully"
-              break
-            fi
-            if [ "$i" -eq 30 ]; then
-              echo "Server failed to start"
-              kill "$SERVER_PID"
-              exit 1
-            fi
-            sleep 2
-          done
-
-          # Test health endpoint
-          curl -f http://localhost:8000/health
-
-          # Test completion
-          curl -f http://localhost:8000/v1/completions \
-            -H "Content-Type: application/json" \
-            -d '{
-              "model": "Qwen/Qwen3-0.6B",
-              "prompt": "Hello",
-              "max_tokens": 5
-            }'
-
-          # Cleanup
-          kill "$SERVER_PID"

.github/workflows/pre-commit.yml

@@ -16,7 +16,7 @@ jobs:
   pre-commit:
     runs-on: ubuntu-latest
     steps:
-    - uses: actions/checkout@1af3b93b6815bc44a9784bd300feb67ff0d1eeb3 # v6.0.0
+    - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
     - uses: actions/setup-python@e797f83bcb11b83ae66e0230d6156d7c80228e7c # v6.0.0
       with:
         python-version: "3.12"

.gitignore

@@ -4,9 +4,6 @@
 # vllm-flash-attn built from source
 vllm/vllm_flash_attn/*
 
-# OpenAI triton kernels copied from source
-vllm/third_party/triton_kernels/*
-
 # triton jit
 .triton
 

.markdownlint.yaml

@@ -3,9 +3,10 @@ MD007:
 MD013: false
 MD024:
   siblings_only: true
-MD031:
-  list_items: false
 MD033: false
+MD045: false
 MD046: false
+MD051: false
 MD052: false
+MD053: false
 MD059: false

.pre-commit-config.yaml

@@ -38,7 +38,7 @@ repos:
   rev: 0.9.1
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu129, --python-platform, x86_64-manylinux_2_28, --python-version, "3.12"]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu129, --python-platform, x86_64-manylinux_2_28]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:

CMakeLists.txt

@@ -39,13 +39,6 @@ set(PYTHON_SUPPORTED_VERSIONS "3.10" "3.11" "3.12" "3.13")
 # Supported AMD GPU architectures.
 set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151")
 
-# ROCm installation prefix. Default to /opt/rocm but allow override via
-# -DROCM_PATH=/your/rocm/path when invoking cmake.
-if(NOT DEFINED ROCM_PATH)
-  set(ROCM_PATH "/opt/rocm" CACHE PATH "ROCm installation prefix")
-else()
-  set(ROCM_PATH ${ROCM_PATH} CACHE PATH "ROCm installation prefix" FORCE)
-endif()
 #
 # Supported/expected torch versions for CUDA/ROCm.
 #
@@ -136,7 +129,7 @@ elseif(HIP_FOUND)
 
   # ROCm 5.X and 6.X
   if (ROCM_VERSION_DEV_MAJOR GREATER_EQUAL 5 AND
-      Torch_VERSION VERSION_LESS ${TORCH_SUPPORTED_VERSION_ROCM})
+      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM})
     message(WARNING "Pytorch version >= ${TORCH_SUPPORTED_VERSION_ROCM} "
       "expected for ROCm build, saw ${Torch_VERSION} instead.")
   endif()
@@ -244,28 +237,11 @@ set_gencode_flags_for_srcs(
   SRCS "${VLLM_CUMEM_EXT_SRC}"
   CUDA_ARCHS "${CUDA_ARCHS}")
 
-if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
+if(VLLM_GPU_LANG STREQUAL "CUDA")
   message(STATUS "Enabling cumem allocator extension.")
-  if(VLLM_GPU_LANG STREQUAL "CUDA")
-    # link against cuda driver library
-    list(APPEND CUMEM_LIBS CUDA::cuda_driver)
-  else()
-    # link against rocm driver library. Prefer an absolute path to
-    # libamdhip64.so inside ${ROCM_PATH}/lib if available, otherwise fall
-    # back to linking by name "amdhip64".
-    find_library(AMDHIP64_LIB
-      NAMES amdhip64 libamdhip64.so
-      PATHS ${ROCM_PATH}/lib
-      NO_DEFAULT_PATH)
-    if(AMDHIP64_LIB)
-      message(STATUS "Found libamdhip64 at ${AMDHIP64_LIB}")
-      list(APPEND CUMEM_LIBS ${AMDHIP64_LIB})
-    else()
-      message(WARNING "libamdhip64 not found in ${ROCM_PATH}/lib; falling back to linking 'amdhip64' by name")
-      list(APPEND CUMEM_LIBS amdhip64)
-    endif()
-  endif()
-  define_extension_target(
+  # link against cuda driver library
+  list(APPEND CUMEM_LIBS CUDA::cuda_driver)
+  define_gpu_extension_target(
     cumem_allocator
     DESTINATION vllm
     LANGUAGE CXX
@@ -289,7 +265,6 @@ set(VLLM_EXT_SRC
   "csrc/pos_encoding_kernels.cu"
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
-  "csrc/fused_qknorm_rope_kernel.cu"
   "csrc/layernorm_quant_kernels.cu"
   "csrc/sampler.cu"
   "csrc/cuda_view.cu"
@@ -307,7 +282,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   SET(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
 
   # Set CUTLASS_REVISION. Used for FetchContent. Also fixes some bogus messages when building.
-  set(CUTLASS_REVISION "v4.2.1")
+  set(CUTLASS_REVISION "v4.2.1" CACHE STRING "CUTLASS revision to use")
 
   # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
   if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
@@ -354,17 +329,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # Only build Marlin kernels if we are building for at least some compatible archs.
   # Keep building Marlin for 9.0 as there are some group sizes and shapes that
   # are not supported by Machete yet.
-
-  # marlin arches for fp16 output
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0+PTX" "${CUDA_ARCHS}")
-  # marlin arches for bf16 output (we need 9.0 for bf16 atomicAdd PTX)
-  cuda_archs_loose_intersection(MARLIN_BF16_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
-  # marlin arches for fp8 input
-  # - sm80 doesn't support fp8 computation
-  # - sm90 and sm100 don't support QMMA.16832.F32.E4M3.E4M3 SAAS instruction
-  # so we only enable fp8 computation for SM89 (e.g. RTX 40x0)  and 12.0 (e.g. RTX 50x0)
-  cuda_archs_loose_intersection(MARLIN_FP8_ARCHS "8.9;12.0" "${CUDA_ARCHS}")
-
+  # 9.0 for latest bf16 atomicAdd PTX
+  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
 
     #
@@ -374,18 +340,16 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set(MARLIN_GEN_SCRIPT
       ${CMAKE_CURRENT_SOURCE_DIR}/csrc/quantization/gptq_marlin/generate_kernels.py)
     file(MD5 ${MARLIN_GEN_SCRIPT} MARLIN_GEN_SCRIPT_HASH)
-    list(JOIN CUDA_ARCHS "," CUDA_ARCHS_STR)
-    set(MARLIN_GEN_SCRIPT_HASH_AND_ARCH "${MARLIN_GEN_SCRIPT_HASH}(ARCH:${CUDA_ARCHS_STR})")
 
-    message(STATUS "Marlin generation script hash: ${MARLIN_GEN_SCRIPT_HASH_AND_ARCH}")
-    message(STATUS "Last run Marlin generate script hash: $CACHE{MARLIN_GEN_SCRIPT_HASH_AND_ARCH}")
+    message(STATUS "Marlin generation script hash: ${MARLIN_GEN_SCRIPT_HASH}")
+    message(STATUS "Last run Marlin generate script hash: $CACHE{MARLIN_GEN_SCRIPT_HASH}")
 
-    if (NOT DEFINED CACHE{MARLIN_GEN_SCRIPT_HASH_AND_ARCH}
-        OR NOT $CACHE{MARLIN_GEN_SCRIPT_HASH_AND_ARCH} STREQUAL ${MARLIN_GEN_SCRIPT_HASH_AND_ARCH})
+    if (NOT DEFINED CACHE{MARLIN_GEN_SCRIPT_HASH}
+        OR NOT $CACHE{MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MARLIN_GEN_SCRIPT_HASH})
       execute_process(
         COMMAND ${CMAKE_COMMAND} -E env
         PYTHONPATH=$PYTHONPATH
-          ${Python_EXECUTABLE} ${MARLIN_GEN_SCRIPT} ${CUDA_ARCHS_STR}
+          ${Python_EXECUTABLE} ${MARLIN_GEN_SCRIPT}
         RESULT_VARIABLE marlin_generation_result
         OUTPUT_VARIABLE marlin_generation_result
         OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log
@@ -398,15 +362,15 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                             "\nCheck the log for details: "
                             "${CMAKE_CURRENT_BINARY_DIR}/marlin_generation.log")
       else()
-        set(MARLIN_GEN_SCRIPT_HASH_AND_ARCH ${MARLIN_GEN_SCRIPT_HASH_AND_ARCH}
-            CACHE STRING "Last run Marlin generate script hash and arch" FORCE)
+        set(MARLIN_GEN_SCRIPT_HASH ${MARLIN_GEN_SCRIPT_HASH}
+            CACHE STRING "Last run Marlin generate script hash" FORCE)
         message(STATUS "Marlin generation completed successfully.")
       endif()
     else()
       message(STATUS "Marlin generation script has not changed, skipping generation.")
     endif()
 
-    file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/gptq_marlin/sm80_kernel_*_float16.cu")
+    file(GLOB MARLIN_TEMPLATE_KERNEL_SRC "csrc/quantization/gptq_marlin/kernel_*.cu")
     set_gencode_flags_for_srcs(
       SRCS "${MARLIN_TEMPLATE_KERNEL_SRC}"
       CUDA_ARCHS "${MARLIN_ARCHS}")
@@ -414,34 +378,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       set_source_files_properties(${MARLIN_TEMPLATE_KERNEL_SRC}
         PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
     endif()
+
     list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_KERNEL_SRC})
 
-    file(GLOB MARLIN_TEMPLATE_BF16_KERNEL_SRC "csrc/quantization/gptq_marlin/sm80_kernel_*_bfloat16.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_TEMPLATE_BF16_KERNEL_SRC}"
-      CUDA_ARCHS "${MARLIN_BF16_ARCHS}")
-    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
-      set_source_files_properties(${MARLIN_TEMPLATE_BF16_KERNEL_SRC}
-        PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
-    endif()
-    list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_BF16_KERNEL_SRC})
-
-    if (MARLIN_FP8_ARCHS) 
-      file(GLOB MARLIN_TEMPLATE_FP8_KERNEL_SRC "csrc/quantization/gptq_marlin/sm89_kernel_*.cu")
-      set_gencode_flags_for_srcs(
-        SRCS "${MARLIN_TEMPLATE_FP8_KERNEL_SRC}"
-        CUDA_ARCHS "${MARLIN_FP8_ARCHS}")
-      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
-        set_source_files_properties(${MARLIN_TEMPLATE_FP8_KERNEL_SRC}
-          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
-      endif()
-      list(APPEND VLLM_EXT_SRC ${MARLIN_TEMPLATE_FP8_KERNEL_SRC})
-    endif()
-
     set(MARLIN_SRCS
        "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
        "csrc/quantization/gptq_marlin/gptq_marlin.cu"
-       "csrc/quantization/gptq_marlin/marlin_int4_fp8_preprocess.cu"
        "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
        "csrc/quantization/gptq_marlin/awq_marlin_repack.cu")
     set_gencode_flags_for_srcs(
@@ -545,9 +487,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
   # require CUDA 12.8 or later
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
   endif()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
@@ -637,15 +579,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
       "csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu"
-      "csrc/quantization/fp4/nvfp4_experts_quant.cu"
-      "csrc/quantization/fp4/nvfp4_scaled_mm_sm120_kernels.cu"
-      "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu")
+      "csrc/quantization/fp4/nvfp4_scaled_mm_sm120_kernels.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${FP4_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4_SM120=1")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM120=1")
     message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
   else()
     message(STATUS "Not building NVFP4 as no compatible archs were found.")
@@ -655,9 +594,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # FP4 Archs and flags
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(FP4_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(FP4_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(FP4_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(FP4_ARCHS "10.0a;10.1a;12.0a;12.1a" "${CUDA_ARCHS}")
   endif()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND FP4_ARCHS)
     set(SRCS
@@ -731,7 +670,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
     cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a" "${CUDA_ARCHS}")
   endif()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/quantization/w8a8/cutlass/moe/grouped_mm_c3x_sm100.cu")
@@ -777,9 +716,9 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
   endif()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/quantization/w8a8/cutlass/moe/blockwise_scaled_group_mm_sm100.cu")
@@ -897,7 +836,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   endif()
 
   # Hadacore kernels
-  cuda_archs_loose_intersection(HADACORE_ARCHS "8.0+PTX;9.0+PTX" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(HADACORE_ARCHS "8.0;8.9;9.0" "${CUDA_ARCHS}")
   if(HADACORE_ARCHS)
     set(SRCS "csrc/quantization/hadamard/hadacore/hadamard_transform_cuda.cu")
     set_gencode_flags_for_srcs(
@@ -919,7 +858,7 @@ if (VLLM_GPU_LANG STREQUAL "HIP")
 endif()
 
 message(STATUS "Enabling C extension.")
-define_extension_target(
+define_gpu_extension_target(
   _C
   DESTINATION vllm
   LANGUAGE ${VLLM_GPU_LANG}
@@ -974,15 +913,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     CUDA_ARCHS "${CUDA_ARCHS}")
 
   list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
-  # moe marlin arches
-  # note that we always set `use_atomic_add=False` for moe marlin now,
-  # so we don't need 9.0 for bf16 atomicAdd PTX
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0+PTX" "${CUDA_ARCHS}")
-  # moe marlin arches for fp8 input
-  # - sm80 doesn't support fp8 computation
-  # - sm90 and sm100 don't support QMMA.16832.F32.E4M3.E4M3 SAAS instruction
-  # so we only enable fp8 computation for SM89 (e.g. RTX 40x0)  and 12.0 (e.g. RTX 50x0)
-  cuda_archs_loose_intersection(MARLIN_MOE_FP8_ARCHS "8.9;12.0" "${CUDA_ARCHS}")
+  # 9.0 for latest bf16 atomicAdd PTX
+  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_MOE_ARCHS)
 
     #
@@ -992,18 +924,16 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set(MOE_MARLIN_GEN_SCRIPT
       ${CMAKE_CURRENT_SOURCE_DIR}/csrc/moe/marlin_moe_wna16/generate_kernels.py)
     file(MD5 ${MOE_MARLIN_GEN_SCRIPT} MOE_MARLIN_GEN_SCRIPT_HASH)
-    list(JOIN CUDA_ARCHS "," CUDA_ARCHS_STR)
-    set(MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH "${MOE_MARLIN_GEN_SCRIPT_HASH}(ARCH:${CUDA_ARCHS_STR})")
 
-    message(STATUS "Marlin MOE generation script hash with arch: ${MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH}")
-    message(STATUS "Last run Marlin MOE generate script hash with arch: $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH}")
+    message(STATUS "Marlin MOE generation script hash: ${MOE_MARLIN_GEN_SCRIPT_HASH}")
+    message(STATUS "Last run Marlin MOE generate script hash: $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}")
 
-    if (NOT DEFINED CACHE{MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH}
-        OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH})
+    if (NOT DEFINED CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}
+        OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH})
       execute_process(
         COMMAND ${CMAKE_COMMAND} -E env
         PYTHONPATH=$PYTHONPATH
-          ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT} ${CUDA_ARCHS_STR}
+          ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT}
         RESULT_VARIABLE moe_marlin_generation_result
         OUTPUT_VARIABLE moe_marlin_generation_output
         OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
@@ -1016,7 +946,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                             "\nCheck the log for details: "
                             "${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log")
       else()
-        set(MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH ${MOE_MARLIN_GEN_SCRIPT_HASH_AND_ARCH}
+        set(MOE_MARLIN_GEN_SCRIPT_HASH ${MOE_MARLIN_GEN_SCRIPT_HASH}
             CACHE STRING "Last run Marlin MOE generate script hash" FORCE)
         message(STATUS "Marlin MOE generation completed successfully.")
       endif()
@@ -1024,28 +954,16 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       message(STATUS "Marlin MOE generation script has not changed, skipping generation.")
     endif()
 
-    file(GLOB MARLIN_MOE_SRC "csrc/moe/marlin_moe_wna16/sm80_kernel_*.cu")
-    list(APPEND MARLIN_MOE_SRC "csrc/moe/marlin_moe_wna16/ops.cu")
+    file(GLOB MOE_WNAA16_MARLIN_SRC "csrc/moe/marlin_moe_wna16/*.cu")
     set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_MOE_SRC}"
+      SRCS "${MOE_WNAA16_MARLIN_SRC}"
       CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
     if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
-      set_source_files_properties(${MARLIN_MOE_SRC}
+      set_source_files_properties(${MOE_WNAA16_MARLIN_SRC}
         PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
     endif()
-    list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_SRC})
 
-    if (MARLIN_MOE_FP8_ARCHS)
-      file(GLOB MARLIN_MOE_FP8_SRC "csrc/moe/marlin_moe_wna16/sm89_kernel_*.cu")
-      set_gencode_flags_for_srcs(
-        SRCS "${MARLIN_MOE_FP8_SRC}"
-        CUDA_ARCHS "${MARLIN_MOE_FP8_ARCHS}")
-      if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
-        set_source_files_properties(${MARLIN_MOE_FP8_SRC}
-          PROPERTIES COMPILE_FLAGS "-static-global-template-stub=false")
-      endif()
-      list(APPEND VLLM_MOE_EXT_SRC ${MARLIN_MOE_FP8_SRC})
-    endif()
+    list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})
 
     message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
   else()
@@ -1055,7 +973,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 endif()
 
 message(STATUS "Enabling moe extension.")
-define_extension_target(
+define_gpu_extension_target(
   _moe_C
   DESTINATION vllm
   LANGUAGE ${VLLM_GPU_LANG}
@@ -1076,7 +994,7 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
     "csrc/rocm/skinny_gemms.cu"
     "csrc/rocm/attention.cu")
 
-  define_extension_target(
+  define_gpu_extension_target(
     _rocm_C
     DESTINATION vllm
     LANGUAGE ${VLLM_GPU_LANG}
@@ -1087,11 +1005,6 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
     WITH_SOABI)
 endif()
 
-# For CUDA and HIP builds also build the triton_kernels external package.
-if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
-    include(cmake/external_projects/triton_kernels.cmake)
-endif()
-
 # For CUDA we also build and ship some external projects.
 if (VLLM_GPU_LANG STREQUAL "CUDA")
     include(cmake/external_projects/flashmla.cmake)

README.md

@@ -21,9 +21,6 @@ Join us at the [PyTorch Conference, October 22-23](https://events.linuxfoundatio
 
 *Latest News* 🔥
 
-- [2025/11] We hosted [vLLM Bangkok Meetup](https://luma.com/v0f647nv). We explored vLLM and LMCache inference and low-resource language adaptation with speakers from Embedded LLM, AMD, and Red Hat. Please find the meetup slides [here](https://drive.google.com/drive/folders/1H0DS57F8HQ5q3kSOSoRmucPJWL3E0A_X?usp=sharing).
-- [2025/11] We hosted [the first vLLM Europe Meetup in Zurich](https://luma.com/0gls27kb) focused on quantization, distributed inference, and reinforcement learning at scale with speakers from Mistral, IBM, and Red Hat. Please find the meetup slides [here](https://docs.google.com/presentation/d/1UC9PTLCHYXQpOmJDSFg6Sljra3iVXzc09DeEI7dnxMc/edit?usp=sharing) and recording [here](https://www.youtube.com/watch?v=6m6ZE6yVEDI)
-- [2025/11] We hosted [vLLM Beijing Meetup](https://mp.weixin.qq.com/s/xSrYXjNgr1HbCP4ExYNG1w) focusing on distributed inference and diverse accelerator support with vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1nQJ8ZkLSjKxvu36sSHaceVXtttbLvvu-?usp=drive_link).
 - [2025/10] We hosted [vLLM Shanghai Meetup](https://mp.weixin.qq.com/s/__xb4OyOsImz-9eAVrdlcg) focused on hands-on vLLM inference optimization! Please find the meetup slides [here](https://drive.google.com/drive/folders/1KqwjsFJLfEsC8wlDugnrR61zsWHt94Q6).
 - [2025/09] We hosted [vLLM Toronto Meetup](https://luma.com/e80e0ymm) focused on tackling inference at scale and speculative decoding with speakers from NVIDIA and Red Hat! Please find the meetup slides [here](https://docs.google.com/presentation/d/1IYJYmJcu9fLpID5N5RbW_vO0XLo0CGOR14IXOjB61V8/edit?usp=sharing).
 - [2025/08] We hosted [vLLM Shenzhen Meetup](https://mp.weixin.qq.com/s/k8ZBO1u2_2odgiKWH_GVTQ) focusing on the ecosystem around vLLM! Please find the meetup slides [here](https://drive.google.com/drive/folders/1Ua2SVKVSu-wp5vou_6ElraDt2bnKhiEA).
@@ -86,7 +83,7 @@ vLLM is flexible and easy to use with:
 - Tensor, pipeline, data and expert parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support for NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, Arm CPUs, and TPU. Additionally, support for diverse hardware plugins such as Intel Gaudi, IBM Spyre and Huawei Ascend.
+- Support for NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, and TPU. Additionally, support for diverse hardware plugins such as Intel Gaudi, IBM Spyre and Huawei Ascend.
 - Prefix caching support
 - Multi-LoRA support
 

benchmarks/auto_tune/README.md

@@ -83,7 +83,7 @@ MIN_CACHE_HIT_PCT=0
 MAX_LATENCY_ALLOWED_MS=100000000000 # A very large number
 ```
 
-### 2. Maximize Throughput with a Latency Requirement
+#### 2. Maximize Throughput with a Latency Requirement
 
 - **Goal**: Find the best server parameters when P99 end-to-end latency must be below 500ms.
 - **Configuration**:
@@ -96,7 +96,7 @@ MIN_CACHE_HIT_PCT=0
 MAX_LATENCY_ALLOWED_MS=500
 ```
 
-### 3. Maximize Throughput with Prefix Caching and Latency Requirements
+#### 3. Maximize Throughput with Prefix Caching and Latency Requirements
 
 - **Goal**: Find the best server parameters assuming a 60% prefix cache hit rate and a latency requirement of 500ms.
 - **Configuration**:

benchmarks/backend_request_func.py

@@ -620,7 +620,7 @@ def get_tokenizer(
         kwargs["use_fast"] = False
     if tokenizer_mode == "mistral":
         try:
-            from vllm.tokenizers import MistralTokenizer
+            from vllm.transformers_utils.tokenizer import MistralTokenizer
         except ImportError as e:
             raise ImportError(
                 "MistralTokenizer requires vllm package.\n"

benchmarks/benchmark_batch_invariance.py

@@ -1,380 +0,0 @@
-#!/usr/bin/env python3
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""
-Benchmark to measure the performance overhead of VLLM_BATCH_INVARIANT mode.
-
-This benchmark runs the same workload twice:
-1. With VLLM_BATCH_INVARIANT=0 (baseline)
-2. With VLLM_BATCH_INVARIANT=1 (batch invariant mode)
-
-And reports the timing and throughput metrics for comparison.
-
-Environment variables:
-    VLLM_BENCH_MODEL: Model to benchmark (default: "Qwen/Qwen3-1.7B")
-    VLLM_BENCH_TP_SIZE: Tensor parallel size (default: 1, use 8 for deepseek)
-    VLLM_BENCH_BATCH_SIZE: Max batch size (default: 128)
-    VLLM_BENCH_NUM_TRIALS: Number of trials to run (default: 5)
-    VLLM_BENCH_MIN_PROMPT: Min prompt length in words (default: 1024)
-    VLLM_BENCH_MAX_PROMPT: Max prompt length in words (default: 2048)
-    VLLM_BENCH_MAX_TOKENS: Max tokens to generate (default: 128)
-    VLLM_BENCH_TEMPERATURE: Temperature for sampling (default: 0.0)
-    VLLM_BENCH_GPU_MEMORY_UTILIZATION: GPU memory utilization (default: 0.4)
-    VLLM_BENCH_MAX_MODEL_LEN: Max model length (default: 5120)
-    VLLM_BENCH_BACKEND: Attention backend (default: FLASH_ATTN)
-
-Example usage:
-    # Benchmark qwen3 (default)
-    python benchmarks/benchmark_batch_invariance.py
-
-    # Benchmark deepseek with 8 GPUs
-    VLLM_BENCH_MODEL="deepseek-ai/DeepSeek-V3" VLLM_BENCH_TP_SIZE=8 \\
-        python benchmarks/benchmark_batch_invariance.py
-
-    # Quick test with fewer trials
-    VLLM_BENCH_NUM_TRIALS=2 VLLM_BENCH_BATCH_SIZE=32 \\
-        python benchmarks/benchmark_batch_invariance.py
-"""
-
-import contextlib
-import os
-import random
-import time
-
-from vllm import LLM, SamplingParams
-from vllm.platforms import current_platform
-
-
-def _random_prompt(min_words: int = 1024, max_words: int = 1024 * 2) -> str:
-    """Generate a random prompt for benchmarking."""
-    prompt_templates = [
-        "Question: What is the capital of France?\nAnswer: The capital of France is",
-        "Q: How does photosynthesis work?\nA: Photosynthesis is the process by which",
-        "User: Can you explain quantum mechanics?\nAssistant: Quantum mechanics is",
-        "Once upon a time in a distant galaxy, there lived",
-        "The old man walked slowly down the street, remembering",
-        "In the year 2157, humanity finally discovered",
-        "To implement a binary search tree in Python, first we need to",
-        "The algorithm works by iterating through the array and",
-        "Here's how to optimize database queries using indexing:",
-        "The Renaissance was a period in European history that",
-        "Climate change is caused by several factors including",
-        "The human brain contains approximately 86 billion neurons which",
-        "I've been thinking about getting a new laptop because",
-        "Yesterday I went to the store and bought",
-        "My favorite thing about summer is definitely",
-    ]
-
-    base_prompt = random.choice(prompt_templates)
-
-    if max_words < min_words:
-        max_words = min_words
-    target_words = random.randint(min_words, max_words)
-
-    if target_words > 50:
-        padding_text = (
-            " This is an interesting topic that deserves more explanation. "
-            * (target_words // 50)
-        )
-        base_prompt = base_prompt + padding_text
-
-    return base_prompt
-
-
-def run_benchmark_with_batch_invariant(
-    model: str,
-    tp_size: int,
-    max_batch_size: int,
-    num_trials: int,
-    min_prompt: int,
-    max_prompt: int,
-    max_tokens: int,
-    temperature: float,
-    gpu_mem_util: float,
-    max_model_len: int,
-    backend: str,
-    batch_invariant: bool,
-    seed: int = 12345,
-) -> dict:
-    """
-    Run the benchmark with the specified configuration.
-
-    Returns a dict with timing and throughput metrics.
-    """
-    random.seed(seed)
-
-    # Set environment variables
-    os.environ["VLLM_ATTENTION_BACKEND"] = backend
-    if batch_invariant:
-        os.environ["VLLM_BATCH_INVARIANT"] = "1"
-    else:
-        os.environ["VLLM_BATCH_INVARIANT"] = "0"
-
-    print(f"\n{'=' * 80}")
-    print(f"BENCHMARK: VLLM_BATCH_INVARIANT={int(batch_invariant)}")
-    print(f"  Model: {model}")
-    print(f"  TP Size: {tp_size}")
-    print(f"  Backend: {backend}")
-    print(f"  Max Batch Size: {max_batch_size}")
-    print(f"  Trials: {num_trials}")
-    print(f"  Max Tokens: {max_tokens}")
-    print(f"{'=' * 80}\n")
-
-    sampling = SamplingParams(
-        temperature=temperature,
-        top_p=0.95,
-        max_tokens=max_tokens,
-        seed=20240919,
-    )
-
-    needle_prompt = "There once was a "
-
-    llm = None
-    try:
-        # Create LLM engine
-        start_init = time.perf_counter()
-        llm = LLM(
-            model=model,
-            max_num_seqs=max_batch_size,
-            gpu_memory_utilization=gpu_mem_util,
-            max_model_len=max_model_len,
-            dtype="bfloat16",
-            tensor_parallel_size=tp_size,
-            enable_prefix_caching=False,
-        )
-        init_time = time.perf_counter() - start_init
-        print(f"Engine initialization time: {init_time:.2f}s\n")
-
-        # Generate baseline
-        print("Generating baseline (warmup)...")
-        baseline_out = llm.generate([needle_prompt], sampling)
-        assert len(baseline_out) == 1
-        baseline_text = baseline_out[0].outputs[0].text
-        print(f"Baseline output: '{baseline_text[:50]}...'\n")
-
-        # Run trials and measure timing
-        trial_times: list[float] = []
-        total_tokens = 0
-        total_prompts = 0
-
-        for trial in range(num_trials):
-            # Create a batch
-            prompts: list[str] = []
-            batch_size = random.randint(max_batch_size // 2, max_batch_size)
-            needle_pos = random.randint(0, batch_size - 1)
-            for i in range(batch_size):
-                if i == needle_pos:
-                    prompts.append(needle_prompt)
-                else:
-                    prompts.append(_random_prompt(min_prompt, max_prompt))
-
-            # Measure time for this trial
-            start_time = time.perf_counter()
-            outputs = llm.generate(prompts, sampling)
-            trial_time = time.perf_counter() - start_time
-
-            trial_times.append(trial_time)
-            total_prompts += len(prompts)
-
-            # Count tokens
-            for output in outputs:
-                if output.outputs:
-                    total_tokens += len(output.outputs[0].token_ids)
-
-            print(
-                f"Trial {trial + 1}/{num_trials}: "
-                f"batch_size={batch_size}, "
-                f"time={trial_time:.2f}s"
-            )
-
-            # Verify needle output still matches
-            needle_output = outputs[needle_pos]
-            assert needle_output.prompt == needle_prompt
-
-        # Compute statistics
-        avg_time = sum(trial_times) / len(trial_times)
-        min_time = min(trial_times)
-        max_time = max(trial_times)
-        throughput = total_tokens / sum(trial_times)
-        prompts_per_sec = total_prompts / sum(trial_times)
-
-        print(f"\n{'=' * 80}")
-        print("RESULTS:")
-        print(f"  Average time per trial: {avg_time:.2f}s")
-        print(f"  Min time: {min_time:.2f}s")
-        print(f"  Max time: {max_time:.2f}s")
-        print(f"  Total tokens generated: {total_tokens}")
-        print(f"  Total prompts processed: {total_prompts}")
-        print(f"  Throughput: {throughput:.2f} tokens/s")
-        print(f"  Prompts/s: {prompts_per_sec:.2f}")
-        print(f"{'=' * 80}\n")
-
-        return {
-            "init_time": init_time,
-            "avg_time": avg_time,
-            "min_time": min_time,
-            "max_time": max_time,
-            "total_tokens": total_tokens,
-            "total_prompts": total_prompts,
-            "throughput": throughput,
-            "prompts_per_sec": prompts_per_sec,
-            "trial_times": trial_times,
-        }
-
-    finally:
-        # Cleanup
-        if llm is not None:
-            with contextlib.suppress(Exception):
-                llm.shutdown()
-
-
-def main():
-    # Check platform support
-    if not (current_platform.is_cuda() and current_platform.has_device_capability(90)):
-        print("ERROR: Requires CUDA and >= Hopper (SM90)")
-        print(f"Current platform: {current_platform.device_type}")
-        if current_platform.is_cuda():
-            print(f"Device capability: {current_platform.get_device_capability()}")
-        return 1
-
-    # Read configuration from environment
-    model = os.getenv("VLLM_BENCH_MODEL", "Qwen/Qwen3-1.7B")
-    tp_size = int(os.getenv("VLLM_BENCH_TP_SIZE", "1"))
-    max_batch_size = int(os.getenv("VLLM_BENCH_BATCH_SIZE", "128"))
-    num_trials = int(os.getenv("VLLM_BENCH_NUM_TRIALS", "5"))
-    min_prompt = int(os.getenv("VLLM_BENCH_MIN_PROMPT", "1024"))
-    max_prompt = int(os.getenv("VLLM_BENCH_MAX_PROMPT", "2048"))
-    max_tokens = int(os.getenv("VLLM_BENCH_MAX_TOKENS", "128"))
-    temperature = float(os.getenv("VLLM_BENCH_TEMPERATURE", "0.0"))
-    gpu_mem_util = float(os.getenv("VLLM_BENCH_GPU_MEMORY_UTILIZATION", "0.4"))
-    max_model_len = int(os.getenv("VLLM_BENCH_MAX_MODEL_LEN", "5120"))
-    backend = os.getenv("VLLM_BENCH_BACKEND", "FLASH_ATTN")
-
-    print("\n" + "=" * 80)
-    print("VLLM BATCH INVARIANCE BENCHMARK")
-    print("=" * 80)
-    print("\nConfiguration:")
-    print(f"  Model: {model}")
-    print(f"  Tensor Parallel Size: {tp_size}")
-    print(f"  Attention Backend: {backend}")
-    print(f"  Max Batch Size: {max_batch_size}")
-    print(f"  Number of Trials: {num_trials}")
-    print(f"  Prompt Length Range: {min_prompt}-{max_prompt} words")
-    print(f"  Max Tokens to Generate: {max_tokens}")
-    print(f"  Temperature: {temperature}")
-    print(f"  GPU Memory Utilization: {gpu_mem_util}")
-    print(f"  Max Model Length: {max_model_len}")
-    print("=" * 80)
-
-    # Run benchmark WITHOUT batch invariance (baseline)
-    print("\n" + "=" * 80)
-    print("PHASE 1: Running WITHOUT batch invariance (baseline)")
-    print("=" * 80)
-    baseline_results = run_benchmark_with_batch_invariant(
-        model=model,
-        tp_size=tp_size,
-        max_batch_size=max_batch_size,
-        num_trials=num_trials,
-        min_prompt=min_prompt,
-        max_prompt=max_prompt,
-        max_tokens=max_tokens,
-        temperature=temperature,
-        gpu_mem_util=gpu_mem_util,
-        max_model_len=max_model_len,
-        backend=backend,
-        batch_invariant=False,
-    )
-
-    # Run benchmark WITH batch invariance
-    print("\n" + "=" * 80)
-    print("PHASE 2: Running WITH batch invariance")
-    print("=" * 80)
-    batch_inv_results = run_benchmark_with_batch_invariant(
-        model=model,
-        tp_size=tp_size,
-        max_batch_size=max_batch_size,
-        num_trials=num_trials,
-        min_prompt=min_prompt,
-        max_prompt=max_prompt,
-        max_tokens=max_tokens,
-        temperature=temperature,
-        gpu_mem_util=gpu_mem_util,
-        max_model_len=max_model_len,
-        backend=backend,
-        batch_invariant=True,
-    )
-
-    # Compare results
-    print("\n" + "=" * 80)
-    print("COMPARISON: Batch Invariance vs Baseline")
-    print("=" * 80)
-
-    init_overhead_pct = (
-        (batch_inv_results["init_time"] - baseline_results["init_time"])
-        / baseline_results["init_time"]
-        * 100
-    )
-    time_overhead_pct = (
-        (batch_inv_results["avg_time"] - baseline_results["avg_time"])
-        / baseline_results["avg_time"]
-        * 100
-    )
-    throughput_change_pct = (
-        (batch_inv_results["throughput"] - baseline_results["throughput"])
-        / baseline_results["throughput"]
-        * 100
-    )
-
-    print("\nInitialization Time:")
-    print(f"  Baseline:         {baseline_results['init_time']:.2f}s")
-    print(f"  Batch Invariant:  {batch_inv_results['init_time']:.2f}s")
-    print(f"  Overhead:         {init_overhead_pct:+.2f}%")
-
-    print("\nAverage Trial Time:")
-    print(f"  Baseline:         {baseline_results['avg_time']:.2f}s")
-    print(f"  Batch Invariant:  {batch_inv_results['avg_time']:.2f}s")
-    print(f"  Overhead:         {time_overhead_pct:+.2f}%")
-
-    print("\nThroughput (tokens/s):")
-    print(f"  Baseline:         {baseline_results['throughput']:.2f}")
-    print(f"  Batch Invariant:  {batch_inv_results['throughput']:.2f}")
-    print(f"  Change:           {throughput_change_pct:+.2f}%")
-
-    print("\nPrompts/s:")
-    print(f"  Baseline:         {baseline_results['prompts_per_sec']:.2f}")
-    print(f"  Batch Invariant:  {batch_inv_results['prompts_per_sec']:.2f}")
-
-    print("\n" + "=" * 80)
-    print("SUMMARY")
-    print("=" * 80)
-    if time_overhead_pct > 0:
-        print(
-            f"Batch invariance mode adds approximately {time_overhead_pct:.1f}% "
-            "overhead"
-        )
-    else:
-        print(
-            f"Batch invariance mode is approximately {-time_overhead_pct:.1f}% "
-            "faster (unexpected!)"
-        )
-
-    if abs(throughput_change_pct) < 1.0:
-        print("Throughput difference is negligible (< 1%)")
-    elif throughput_change_pct < 0:
-        print(
-            f"Throughput decreased by {-throughput_change_pct:.1f}% "
-            "with batch invariance"
-        )
-    else:
-        print(
-            f"Throughput increased by {throughput_change_pct:.1f}% "
-            "with batch invariance (unexpected!)"
-        )
-
-    print("=" * 80 + "\n")
-
-    return 0
-
-
-if __name__ == "__main__":
-    exit(main())

benchmarks/benchmark_prefix_caching.py

@@ -69,7 +69,7 @@ def sample_tokens(tokenizer: PreTrainedTokenizerBase, length: int) -> list[int]:
 
     # Remove the special tokens.
     return random.choices(
-        [v for v in vocab.values() if v not in all_special_ids],
+        [v for k, v in vocab.items() if k not in all_special_ids],
         k=length,
     )
 

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

@@ -5,12 +5,11 @@ import argparse
 import asyncio
 import logging
 import os
-import time
-import uuid
-from urllib.parse import urlparse
 
 import aiohttp
 from quart import Quart, Response, make_response, request
+from rate_limiter import RateLimiter
+from request_queue import RequestQueue
 
 # Configure logging
 logging.basicConfig(level=logging.INFO)
@@ -25,8 +24,26 @@ def parse_args():
     parser.add_argument(
         "--timeout",
         type=float,
-        default=6 * 60 * 60,
-        help="Timeout for backend service requests in seconds (default: 21600)",
+        default=300,
+        help="Timeout for backend service requests in seconds (default: 300)",
+    )
+    parser.add_argument(
+        "--max-concurrent",
+        type=int,
+        default=100,
+        help="Maximum concurrent requests to backend services (default: 100)",
+    )
+    parser.add_argument(
+        "--queue-size",
+        type=int,
+        default=500,
+        help="Maximum number of requests in the queue (default: 500)",
+    )
+    parser.add_argument(
+        "--rate-limit",
+        type=int,
+        default=40,
+        help="Maximum requests per second (default: 40)",
     )
     parser.add_argument(
         "--port",
@@ -37,32 +54,14 @@ def parse_args():
     parser.add_argument(
         "--prefill-url",
         type=str,
-        default="http://localhost:8100",
-        help="Prefill service base URL (protocol + host[:port])",
+        default="http://localhost:8100/v1/completions",
+        help="Prefill service endpoint URL",
     )
     parser.add_argument(
         "--decode-url",
         type=str,
-        default="http://localhost:8200",
-        help="Decode service base URL (protocol + host[:port])",
-    )
-    parser.add_argument(
-        "--kv-host",
-        type=str,
-        default="localhost",
-        help="Hostname or IP used by KV transfer (default: localhost)",
-    )
-    parser.add_argument(
-        "--prefill-kv-port",
-        type=int,
-        default=14579,
-        help="Prefill KV port (default: 14579)",
-    )
-    parser.add_argument(
-        "--decode-kv-port",
-        type=int,
-        default=14580,
-        help="Decode KV port (default: 14580)",
+        default="http://localhost:8200/v1/completions",
+        help="Decode service endpoint URL",
     )
 
     return parser.parse_args()
@@ -74,129 +73,70 @@ def main():
 
     # Initialize configuration using command line parameters
     AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=args.timeout)
+    MAX_CONCURRENT_REQUESTS = args.max_concurrent
+    REQUEST_QUEUE_SIZE = args.queue_size
+    RATE_LIMIT = args.rate_limit
     PREFILL_SERVICE_URL = args.prefill_url
     DECODE_SERVICE_URL = args.decode_url
     PORT = args.port
 
-    PREFILL_KV_ADDR = f"{args.kv_host}:{args.prefill_kv_port}"
-    DECODE_KV_ADDR = f"{args.kv_host}:{args.decode_kv_port}"
-
-    logger.info(
-        "Proxy resolved KV addresses -> prefill: %s, decode: %s",
-        PREFILL_KV_ADDR,
-        DECODE_KV_ADDR,
-    )
-
     app = Quart(__name__)
 
-    # Attach the configuration object to the application instance so helper
-    # coroutines can read the resolved backend URLs and timeouts without using
-    # globals.
+    # Initialize the rate limiter and request queue
+    rate_limiter = RateLimiter(RATE_LIMIT)
+    request_queue = RequestQueue(MAX_CONCURRENT_REQUESTS, REQUEST_QUEUE_SIZE)
+
+    # Attach the configuration object to the application instance
     app.config.update(
         {
             "AIOHTTP_TIMEOUT": AIOHTTP_TIMEOUT,
+            "rate_limiter": rate_limiter,
+            "request_queue": request_queue,
             "PREFILL_SERVICE_URL": PREFILL_SERVICE_URL,
             "DECODE_SERVICE_URL": DECODE_SERVICE_URL,
-            "PREFILL_KV_ADDR": PREFILL_KV_ADDR,
-            "DECODE_KV_ADDR": DECODE_KV_ADDR,
         }
     )
 
-    def _normalize_base_url(url: str) -> str:
-        """Remove any trailing slash so path joins behave predictably."""
-        return url.rstrip("/")
+    # Start queue processing on app startup
+    @app.before_serving
+    async def startup():
+        """Start request processing task when app starts serving"""
+        asyncio.create_task(request_queue.process())
 
-    def _get_host_port(url: str) -> str:
-        """Return the hostname:port portion for logging and KV headers."""
-        parsed = urlparse(url)
-        host = parsed.hostname or "localhost"
-        port = parsed.port
-        if port is None:
-            port = 80 if parsed.scheme == "http" else 443
-        return f"{host}:{port}"
+    async def forward_request(url, data):
+        """Forward request to backend service with rate limiting and error handling"""
+        headers = {"Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}"}
 
-    PREFILL_BASE = _normalize_base_url(PREFILL_SERVICE_URL)
-    DECODE_BASE = _normalize_base_url(DECODE_SERVICE_URL)
-    KV_TARGET = _get_host_port(DECODE_SERVICE_URL)
-
-    def _build_headers(request_id: str) -> dict[str, str]:
-        """Construct the headers expected by vLLM's P2P disagg connector."""
-        headers: dict[str, str] = {"X-Request-Id": request_id, "X-KV-Target": KV_TARGET}
-        api_key = os.environ.get("OPENAI_API_KEY")
-        if api_key:
-            headers["Authorization"] = f"Bearer {api_key}"
-        return headers
-
-    async def _run_prefill(
-        request_path: str,
-        payload: dict,
-        headers: dict[str, str],
-        request_id: str,
-    ):
-        url = f"{PREFILL_BASE}{request_path}"
-        start_ts = time.perf_counter()
-        logger.info("[prefill] start request_id=%s url=%s", request_id, url)
-        try:
-            async with (
-                aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session,
-                session.post(url=url, json=payload, headers=headers) as resp,
-            ):
-                if resp.status != 200:
-                    error_text = await resp.text()
-                    raise RuntimeError(
-                        f"Prefill backend error {resp.status}: {error_text}"
-                    )
-                await resp.read()
-                logger.info(
-                    "[prefill] done request_id=%s status=%s elapsed=%.2fs",
-                    request_id,
-                    resp.status,
-                    time.perf_counter() - start_ts,
-                )
-        except asyncio.TimeoutError as exc:
-            raise RuntimeError(f"Prefill service timeout at {url}") from exc
-        except aiohttp.ClientError as exc:
-            raise RuntimeError(f"Prefill service unavailable at {url}") from exc
-
-    async def _stream_decode(
-        request_path: str,
-        payload: dict,
-        headers: dict[str, str],
-        request_id: str,
-    ):
-        url = f"{DECODE_BASE}{request_path}"
-        # Stream tokens from the decode service once the prefill stage has
-        # materialized KV caches on the target workers.
-        logger.info("[decode] start request_id=%s url=%s", request_id, url)
-        try:
-            async with (
-                aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session,
-                session.post(url=url, json=payload, headers=headers) as resp,
-            ):
-                if resp.status != 200:
-                    error_text = await resp.text()
-                    logger.error(
-                        "Decode backend error %s - %s", resp.status, error_text
-                    )
-                    err_msg = (
-                        '{"error": "Decode backend error ' + str(resp.status) + '"}'
-                    )
-                    yield err_msg.encode()
-                    return
-                logger.info(
-                    "[decode] streaming response request_id=%s status=%s",
-                    request_id,
-                    resp.status,
-                )
-                async for chunk_bytes in resp.content.iter_chunked(1024):
-                    yield chunk_bytes
-                logger.info("[decode] finished streaming request_id=%s", request_id)
-        except asyncio.TimeoutError:
-            logger.error("Decode service timeout at %s", url)
-            yield b'{"error": "Decode service timeout"}'
-        except aiohttp.ClientError as exc:
-            logger.error("Decode service error at %s: %s", url, exc)
-            yield b'{"error": "Decode service unavailable"}'
+        # Use rate limiter as context manager
+        async with (
+            rate_limiter,
+            aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session,
+        ):
+            try:
+                async with session.post(
+                    url=url, json=data, headers=headers
+                ) as response:
+                    if response.status == 200:
+                        # Stream response chunks
+                        async for chunk_bytes in response.content.iter_chunked(1024):
+                            yield chunk_bytes
+                    else:
+                        # Handle backend service errors
+                        error_text = await response.text()
+                        logger.error(
+                            "Backend service error: %s - %s",
+                            response.status,
+                            error_text,
+                        )
+                        yield b'{"error": "Backend service error"}'
+            except aiohttp.ClientError as e:
+                # Handle connection errors
+                logger.error("Connection error to %s: %s", url, str(e))
+                yield b'{"error": "Service unavailable"}'
+            except asyncio.TimeoutError:
+                # Handle timeout errors
+                logger.error("Timeout connecting to %s", url)
+                yield b'{"error": "Service timeout"}'
 
     async def process_request():
         """Process a single request through prefill and decode stages"""
@@ -206,27 +146,13 @@ def main():
             # Create prefill request (max_tokens=1)
             prefill_request = original_request_data.copy()
             prefill_request["max_tokens"] = 1
-            if "max_completion_tokens" in prefill_request:
-                prefill_request["max_completion_tokens"] = 1
 
             # Execute prefill stage
-            # The request id encodes both KV socket addresses so the backend can
-            # shuttle tensors directly via NCCL once the prefill response
-            # completes.
-            request_id = (
-                f"___prefill_addr_{PREFILL_KV_ADDR}___decode_addr_"
-                f"{DECODE_KV_ADDR}_{uuid.uuid4().hex}"
-            )
-
-            headers = _build_headers(request_id)
-            await _run_prefill(request.path, prefill_request, headers, request_id)
+            async for _ in forward_request(PREFILL_SERVICE_URL, prefill_request):
+                continue
 
             # Execute decode stage and stream response
-            # Pass the unmodified user request so the decode phase can continue
-            # sampling with the already-populated KV cache.
-            generator = _stream_decode(
-                request.path, original_request_data, headers, request_id
-            )
+            generator = forward_request(DECODE_SERVICE_URL, original_request_data)
             response = await make_response(generator)
             response.timeout = None  # Disable timeout for streaming response
             return response
@@ -242,10 +168,23 @@ def main():
     @app.route("/v1/completions", methods=["POST"])
     async def handle_request():
         """Handle incoming API requests with concurrency and rate limiting"""
+        # Create task for request processing
+        task = asyncio.create_task(process_request())
+
+        # Enqueue request or reject if queue is full
+        if not await request_queue.enqueue(task):
+            return Response(
+                response=b'{"error": "Server busy, try again later"}',
+                status=503,
+                content_type="application/json",
+            )
+
         try:
-            return await process_request()
+            # Return the response from the processing task
+            return await task
         except asyncio.CancelledError:
-            logger.warning("Request cancelled")
+            # Handle task cancellation (timeout or queue full)
+            logger.warning("Request cancelled due to timeout or queue full")
             return Response(
                 response=b'{"error": "Request cancelled"}',
                 status=503,

benchmarks/kernels/bench_block_fp8_gemm.py

@@ -1,18 +1,10 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
-import os
-
-# Disable DeepGEMM for this benchmark to use CUTLASS
-os.environ["VLLM_USE_DEEP_GEMM"] = "0"
-
 import torch
 
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    W8A8BlockFp8LinearOp,
-)
-from vllm.model_executor.layers.quantization.utils.quant_utils import (
-    GroupShape,
+    apply_w8a8_block_fp8_linear,
 )
 from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
     CUTLASS_BLOCK_FP8_SUPPORTED,
@@ -47,14 +39,13 @@ def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
     fp8_info = torch.finfo(torch.float8_e4m3fn)
     fp8_max, fp8_min = fp8_info.max, fp8_info.min
 
-    # Create random input tensor (bfloat16, will be quantized by W8A8BlockFp8LinearOp)
+    # Create random FP8 tensors
     A_ref = (torch.rand(M, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
 
-    # Create quantized weight tensor
     B_ref = (torch.rand(N, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
     B = B_ref.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
 
-    # Create weight scales
+    # Create scales
     block_n, block_k = block_size[0], block_size[1]
     n_tiles = (N + block_n - 1) // block_n
     k_tiles = (K + block_k - 1) // block_k
@@ -64,25 +55,19 @@ def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
         * factor_for_scale
     )
 
-    # Create W8A8BlockFp8LinearOp instance
-    weight_group_shape = GroupShape(block_n, block_k)
-    act_quant_group_shape = GroupShape(1, block_k)  # Per-token, per-group quantization
-
-    linear_op = W8A8BlockFp8LinearOp(
-        weight_group_shape=weight_group_shape,
-        act_quant_group_shape=act_quant_group_shape,
-        cutlass_block_fp8_supported=use_cutlass,
-        use_aiter_and_is_supported=False,
-    )
+    # SM90 CUTLASS requires row-major format for scales
+    if use_cutlass and current_platform.is_device_capability(90):
+        Bs = Bs.T.contiguous()
 
     def run():
-        return linear_op.apply(
-            input=A_ref,
-            weight=B,
-            weight_scale=Bs,
-            input_scale=None,
-            bias=None,
-        )
+        if use_cutlass:
+            return apply_w8a8_block_fp8_linear(
+                A_ref, B, block_size, Bs, cutlass_block_fp8_supported=True
+            )
+        else:
+            return apply_w8a8_block_fp8_linear(
+                A_ref, B, block_size, Bs, cutlass_block_fp8_supported=False
+            )
 
     return run
 

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -255,8 +255,8 @@ def bench_run(
         torch.cuda.synchronize()
 
         # Timing
-        start_event = torch.Event(enable_timing=True)
-        end_event = torch.Event(enable_timing=True)
+        start_event = torch.cuda.Event(enable_timing=True)
+        end_event = torch.cuda.Event(enable_timing=True)
 
         latencies = []
         for _ in range(num_iters):

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -16,8 +16,8 @@ from vllm.model_executor.layers.fused_moe.fused_moe import (
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = [
-    "mistralai/Mixtral-8x7B-Instruct-v0.1",
-    "deepseek-ai/DeepSeek-V2-Lite",
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1",
+    "nm-testing/deepseekv2-lite",
     "ibm-granite/granite-3.0-1b-a400m",
     "ibm-granite/granite-3.0-3b-a800m",
 ]

benchmarks/kernels/benchmark_machete.py

@@ -237,7 +237,6 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
             b_q_weight=w_q,
             b_bias=None,
             b_scales=w_s,
-            a_scales=None,
             global_scale=None,
             b_zeros=w_zp,
             g_idx=g_idx,

benchmarks/kernels/benchmark_marlin.py

@@ -263,7 +263,7 @@ def bench_run(
 
     results.append(
         benchmark.Timer(
-            stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, None, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False, False)",  # noqa: E501
+            stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False, False)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
@@ -273,7 +273,7 @@ def bench_run(
 
     results.append(
         benchmark.Timer(
-            stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, None, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True, False)",  # noqa: E501
+            stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True, False)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,

benchmarks/kernels/benchmark_moe.py

@@ -185,8 +185,8 @@ def benchmark_config(
         graph.replay()
     torch.cuda.synchronize()
 
-    start_event = torch.Event(enable_timing=True)
-    end_event = torch.Event(enable_timing=True)
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
 
     latencies: list[float] = []
     for i in range(num_iters):
@@ -211,7 +211,7 @@ def get_rocm_tuning_space(use_fp16):
     num_warps_range = [1, 2, 4, 8]
     group_m_range = [1, 4, 8, 16, 32]
     num_stage_range = [2]
-    waves_per_eu_range = [0, 1, 2, 4]
+    waves_per_eu_range = [0]
     matrix_instr_nonkdim_range = [16, 32] if use_fp16 else []
     kpack_range = [1, 2] if use_fp16 else []
 
@@ -616,11 +616,6 @@ def main(args: argparse.Namespace):
         topk = config.moe_topk[0]
         intermediate_size = config.moe_intermediate_size[0]
         hidden_size = config.hidden_size
-    elif config.architectures[0] in ["Qwen3OmniMoeForConditionalGeneration"]:
-        E = config.thinker_config.text_config.num_experts
-        topk = config.thinker_config.text_config.num_experts_per_tok
-        intermediate_size = config.thinker_config.text_config.moe_intermediate_size
-        hidden_size = config.thinker_config.text_config.hidden_size
     else:
         # Support for llama4
         config = config.get_text_config()

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -105,8 +105,8 @@ def benchmark_permute(
         graph.replay()
     torch.cuda.synchronize()
 
-    start_event = torch.Event(enable_timing=True)
-    end_event = torch.Event(enable_timing=True)
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
 
     latencies: list[float] = []
     for i in range(num_iters):
@@ -241,8 +241,8 @@ def benchmark_unpermute(
         graph.replay()
     torch.cuda.synchronize()
 
-    start_event = torch.Event(enable_timing=True)
-    end_event = torch.Event(enable_timing=True)
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
 
     latencies: list[float] = []
     for i in range(num_iters):

benchmarks/kernels/benchmark_mrope.py

@@ -6,7 +6,7 @@
 #
 # The CSV file (named with current date/time) contains these columns:
 # model_name, tp_size, num_tokens, num_heads, num_kv_heads, head_dim, max_position,
-# is_neox_style, rope_parameters, dtype, torch_mean, torch_median, torch_p99,
+# rope_theta, is_neox_style, rope_scaling, dtype, torch_mean, torch_median, torch_p99,
 # torch_min, torch_max, triton_mean, triton_median, triton_p99, triton_min, triton_max,
 # speedup
 #
@@ -86,8 +86,9 @@ def benchmark_mrope(
     num_heads: int,
     num_kv_heads: int,
     max_position: int = 8192,
+    rope_theta: float = 10000,
     is_neox_style: bool = True,
-    rope_parameters: dict[str, Any] | None = None,
+    rope_scaling: dict[str, Any] = None,
     dtype: torch.dtype = torch.bfloat16,
     seed: int = 0,
     warmup_iter: int = 10,
@@ -101,8 +102,9 @@ def benchmark_mrope(
         head_size=head_dim,
         rotary_dim=head_dim,
         max_position=max_position,
+        base=rope_theta,
         is_neox_style=is_neox_style,
-        rope_parameters=rope_parameters,
+        rope_scaling=rope_scaling,
         dtype=dtype,
     ).to(device=device)
 
@@ -201,8 +203,9 @@ def benchmark_mrope(
             num_kv_heads,
             head_dim,
             max_position,
+            rope_theta,
             is_neox_style,
-            str(rope_parameters),
+            str(rope_scaling),
             str(dtype).split(".")[-1],
             torch_stats["mean"],
             torch_stats["median"],
@@ -252,8 +255,9 @@ if __name__ == "__main__":
             "num_kv_heads",
             "head_dim",
             "max_position",
+            "rope_theta",
             "is_neox_style",
-            "rope_parameters",
+            "rope_scaling",
             "dtype",
             "torch_mean",
             "torch_median",
@@ -299,7 +303,7 @@ if __name__ == "__main__":
                 q_size = num_heads * head_dim
                 kv_size = num_kv_heads * head_dim
                 is_neox_style = True
-                rope_parameters = config.rope_parameters
+                rope_theta = config.rope_theta
                 max_position = config.max_position_embeddings
 
                 for num_tokens in num_tokens_list:
@@ -311,8 +315,9 @@ if __name__ == "__main__":
                         num_heads=num_heads,
                         num_kv_heads=num_kv_heads,
                         max_position=max_position,
+                        rope_theta=rope_theta,
                         is_neox_style=is_neox_style,
-                        rope_parameters=rope_parameters,
+                        rope_scaling=config.rope_scaling,
                         dtype=getattr(torch, args.dtype),
                         seed=args.seed,
                         warmup_iter=args.warmup_iter,

benchmarks/kernels/benchmark_per_token_group_quant.py

@@ -30,8 +30,8 @@ def _time_cuda(
         fn()
     torch.cuda.synchronize()
 
-    start = torch.Event(enable_timing=True)
-    end = torch.Event(enable_timing=True)
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
 
     start.record()
     for _ in range(bench_iters):

benchmarks/kernels/benchmark_rope.py

@@ -1,76 +1,97 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
-import itertools
+from itertools import accumulate
 
+import nvtx
 import torch
 
-from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.triton_utils import triton
+from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding, get_rope
+from vllm.platforms import current_platform
 from vllm.utils.argparse_utils import FlexibleArgumentParser
 
-batch_size_range = [2**i for i in range(0, 8, 2)]
-seq_len_range = [2**i for i in range(6, 10, 1)]
-num_heads_range = [32, 48]
-configs = list(itertools.product(batch_size_range, seq_len_range, num_heads_range))
 
+def benchmark_rope_kernels_multi_lora(
+    is_neox_style: bool,
+    batch_size: int,
+    seq_len: int,
+    num_heads: int,
+    head_size: int,
+    rotary_dim: int | None,
+    dtype: torch.dtype,
+    seed: int,
+    device: str,
+    max_position: int = 8192,
+    base: float = 10000,
+) -> None:
+    current_platform.seed_everything(seed)
+    torch.set_default_device(device)
+    if rotary_dim is None:
+        rotary_dim = head_size
+    # silulating serving 4 LoRAs
+    scaling_factors = [1, 2, 4, 8]
+    # batched RoPE can take multiple scaling factors
+    batched_rope = get_rope(
+        head_size,
+        rotary_dim,
+        max_position,
+        base,
+        is_neox_style,
+        {"rope_type": "linear", "factor": tuple(scaling_factors)},
+    )
+    # non-batched RoPE takes only one scaling factor, we create multiple
+    # instances to simulate the same behavior
+    non_batched_ropes: list[RotaryEmbedding] = []
+    for scaling_factor in scaling_factors:
+        non_batched_ropes.append(
+            get_rope(
+                head_size,
+                rotary_dim,
+                max_position,
+                base,
+                is_neox_style,
+                {"rope_type": "linear", "factor": (scaling_factor,)},
+            )
+        )
 
-def get_benchmark(head_size, rotary_dim, is_neox_style, device):
-    @triton.testing.perf_report(
-        triton.testing.Benchmark(
-            x_names=["batch_size", "seq_len", "num_heads"],
-            x_vals=[list(_) for _ in configs],
-            line_arg="provider",
-            line_vals=["torch", "flashinfer", "vllm"],
-            line_names=["PyTorch", "FlashInfer", "vLLM"],
-            styles=[("blue", "-"), ("green", "-"), ("red", "-")],
-            ylabel="us",
-            plot_name=f"rope-perf{'-neox-style' if is_neox_style else ''}",
-            args={},
+    positions = torch.randint(0, max_position, (batch_size, seq_len))
+    query = torch.randn(batch_size, seq_len, num_heads * head_size, dtype=dtype)
+    key = torch.randn_like(query)
+
+    # create query offsets for batched RoPE, we concat multiple kv cache
+    # together and each query needs to find the right kv cache of its type
+    offset_map = torch.tensor(
+        list(
+            accumulate(
+                [0]
+                + [
+                    max_position * scaling_factor * 2
+                    for scaling_factor in scaling_factors[:-1]
+                ]
+            )
         )
     )
-    def benchmark(batch_size, seq_len, num_heads, provider):
-        dtype = torch.bfloat16
-        max_position = 8192
-        base = 10000
-        rope = get_rope(head_size, rotary_dim, max_position, base, is_neox_style)
-        rope = rope.to(dtype=dtype, device=device)
-        cos_sin_cache = rope.cos_sin_cache.to(dtype=torch.float, device=device)
+    query_types = torch.randint(
+        0, len(scaling_factors), (batch_size, seq_len), device=device
+    )
+    # map query types to offsets
+    query_offsets = offset_map[query_types]
+    # the kernel takes flattened offsets
+    flatten_offsets = query_offsets.flatten()
 
-        positions = torch.randint(0, max_position, (batch_size, seq_len), device=device)
-        query = torch.randn(
-            (batch_size, seq_len, num_heads * head_size), dtype=dtype, device=device
-        )
-        key = torch.randn_like(query)
-
-        quantiles = [0.5, 0.2, 0.8]
-
-        if provider == "torch":
-            ms, min_ms, max_ms = triton.testing.do_bench(
-                lambda: rope.forward_native(positions, query.clone(), key.clone()),
-                quantiles=quantiles,
-            )
-        elif provider == "flashinfer":
-            ms, min_ms, max_ms = triton.testing.do_bench(
-                lambda: torch.ops.vllm.flashinfer_rotary_embedding(
-                    positions,
-                    query.clone(),
-                    key.clone(),
-                    head_size,
-                    cos_sin_cache,
-                    is_neox_style,
-                ),
-                quantiles=quantiles,
-            )
-        else:
-            ms, min_ms, max_ms = triton.testing.do_bench(
-                lambda: rope.forward_cuda(positions, query.clone(), key.clone()),
-                quantiles=quantiles,
-            )
-
-        return 1000 * ms, 1000 * max_ms, 1000 * min_ms
-
-    return benchmark
+    # batched queries of the same type together for non-batched RoPE
+    queries = [query[query_types == i] for i in range(len(scaling_factors))]
+    keys = [key[query_types == i] for i in range(len(scaling_factors))]
+    packed_qkr = zip(queries, keys, non_batched_ropes)
+    # synchronize before start timing
+    torch.cuda.synchronize()
+    with nvtx.annotate("non-batched", color="yellow"):
+        for q, k, r in packed_qkr:
+            r.forward(positions, q, k)
+    torch.cuda.synchronize()
+    with nvtx.annotate("batched", color="green"):
+        batched_rope.forward(positions, query, key, flatten_offsets)
+    torch.cuda.synchronize()
 
 
 if __name__ == "__main__":
@@ -95,12 +116,17 @@ if __name__ == "__main__":
     parser.add_argument(
         "--device", type=str, choices=["cuda:0", "cuda:1"], default="cuda:0"
     )
-    parser.add_argument("--save-path", type=str, default="./configs/rope/")
     args = parser.parse_args()
+    print(args)
 
-    # Get the benchmark function
-    benchmark = get_benchmark(
-        args.head_size, args.rotary_dim, args.is_neox_style, args.device
+    benchmark_rope_kernels_multi_lora(
+        is_neox_style=args.is_neox_style,
+        batch_size=args.batch_size,
+        seq_len=args.seq_len,
+        num_heads=args.num_heads,
+        head_size=args.head_size,
+        rotary_dim=args.rotary_dim,
+        dtype=getattr(torch, args.dtype),
+        seed=args.seed,
+        device=args.device,
     )
-    # Run performance benchmark
-    benchmark.run(print_data=True, save_path=args.save_path)

benchmarks/kernels/benchmark_shapes.py

@@ -78,11 +78,11 @@ WEIGHT_SHAPES = {
 }
 
 WEIGHT_SHAPES_MOE = {
-    "mistralai/Mixtral-8x7B-Instruct-v0.1": [
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1": [
         [8, 2, 4096, 28672],
         [8, 2, 14336, 4096],
     ],
-    "deepseek-ai/DeepSeek-V2-Lite": [
+    "nm-testing/deepseekv2-lite": [
         [64, 6, 2048, 1408],
     ],
     "ibm-granite/granite-3.0-1b-a400m": [

benchmarks/kernels/benchmark_silu_mul_fp8_quant.py

@@ -253,8 +253,8 @@ def benchmark(
         )
     torch.cuda.synchronize()
 
-    start_event = torch.Event(enable_timing=True)
-    end_event = torch.Event(enable_timing=True)
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
 
     # Benchmark
     latencies: list[float] = []

benchmarks/kernels/benchmark_trtllm_decode_attention.py

@@ -127,8 +127,8 @@ def benchmark_decode(
 
     def time_fn(fn, warmup=10, trials=20):
         torch.cuda.synchronize()
-        start = torch.Event(enable_timing=True)
-        end = torch.Event(enable_timing=True)
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
         times = []
         for i in range(warmup):
             fn()

benchmarks/kernels/benchmark_trtllm_prefill_attention.py

@@ -139,8 +139,8 @@ def benchmark_prefill(
 
     def time_fn(fn, warmup=10, trials=20):
         torch.cuda.synchronize()
-        start = torch.Event(enable_timing=True)
-        end = torch.Event(enable_timing=True)
+        start = torch.cuda.Event(enable_timing=True)
+        end = torch.cuda.Event(enable_timing=True)
         times = []
         for i in range(warmup):
             fn()

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -183,8 +183,8 @@ def benchmark_config(
         run()
     torch.cuda.synchronize()
 
-    start_event = torch.Event(enable_timing=True)
-    end_event = torch.Event(enable_timing=True)
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
 
     latencies: list[float] = []
     for i in range(num_iters):

benchmarks/kernels/deepgemm/README.md

@@ -2,7 +2,7 @@
 
 This directory includes benchmarks between DeepSeek's DeepGEMM block fp8 kernels against vLLM's existing triton and CUTLASS-based kernels.
 
-Currently, this just includes dense GEMMs and only works on Hopper GPUs.
+Currently this just includes dense GEMMs and only works on Hopper GPUs.
 
 ## Setup
 

benchmarks/multi_turn/README.md

@@ -55,10 +55,6 @@ output_num_chunks  166.0    99.01   11.80    79.00    90.00    98.00   108.75
 ----------------------------------------------------------------------------------------------------
 ```
 
-If you run with `--warmup-step`, the summary will also include `warmup_runtime_sec`
-and `total_runtime_incl_warmup_sec` (while `runtime_sec` continues to reflect the
-benchmark-only runtime so the reported throughput stays comparable).
-
 ### JSON configuration file for synthetic conversations generation
 
 The input flag `--input-file` is used to determine the input conversations for the benchmark.<br/>

benchmarks/multi_turn/bench_dataset.py

@@ -11,7 +11,6 @@ from bench_utils import (
     Color,
     logger,
 )
-from tqdm import tqdm
 from transformers import AutoTokenizer  # type: ignore
 
 # Conversation ID is a string (e.g: "UzTK34D")
@@ -418,10 +417,6 @@ def generate_conversations(
             data = file.read()
             tokens_in_file = tokenizer.encode(data, add_special_tokens=False)
             list_of_tokens.extend(tokens_in_file)
-        logger.info(
-            f"Loaded {len(tokens_in_file)} tokens from file {filename}, "
-            f"total tokens so far: {len(list_of_tokens)}"
-        )
 
     conversations: ConversationsMap = {}
     conv_id = 0
@@ -454,25 +449,18 @@ def generate_conversations(
         )
         base_offset += common_prefix_tokens
 
-    for conv_id in tqdm(
-        range(args.num_conversations),
-        total=args.num_conversations,
-        desc="Generating conversations",
-        unit="conv",
-    ):
+    for conv_id in range(args.num_conversations):
         # Generate a single conversation
         messages: MessagesList = []
 
         nturns = turn_count[conv_id]
 
         # User prompt token count per turn (with lower limit)
-        input_token_count: np.ndarray = args.input_num_tokens.sample(nturns).astype(int)
+        input_token_count: np.ndarray = args.input_num_tokens.sample(nturns)
         input_token_count = np.maximum(input_token_count, base_prompt_token_count)
 
         # Assistant answer token count per turn (with lower limit)
-        output_token_count: np.ndarray = args.output_num_tokens.sample(nturns).astype(
-            int
-        )
+        output_token_count: np.ndarray = args.output_num_tokens.sample(nturns)
         output_token_count = np.maximum(output_token_count, 1)
 
         user_turn = True

benchmarks/multi_turn/benchmark_serving_multi_turn.py

@@ -55,7 +55,6 @@ class ClientArgs(NamedTuple):
     verify_output: bool
     conversation_sampling: ConversationSampling
     request_rate: float
-    max_retries: int
 
 
 class RequestArgs(NamedTuple):
@@ -64,7 +63,6 @@ class RequestArgs(NamedTuple):
     stream: bool
     limit_min_tokens: int  # Use negative value for no limit
     limit_max_tokens: int  # Use negative value for no limit
-    timeout_sec: int
 
 
 class BenchmarkArgs(NamedTuple):
@@ -216,7 +214,6 @@ async def send_request(
     stream: bool = True,
     min_tokens: int | None = None,
     max_tokens: int | None = None,
-    timeout_sec: int = 120,
 ) -> ServerResponse:
     payload = {
         "model": model,
@@ -238,16 +235,10 @@ async def send_request(
     headers = {"Content-Type": "application/json"}
 
     # Calculate the timeout for the request
+    timeout_sec = 120
     if max_tokens is not None:
         # Assume TPOT of 200ms and use max_tokens to determine timeout
-        token_based_timeout = int(max_tokens * 0.2)
-        if token_based_timeout > timeout_sec:
-            timeout_sec = token_based_timeout
-            logger.info(
-                "Using timeout of %ds based on max_tokens %d",
-                timeout_sec,
-                max_tokens,
-            )
+        timeout_sec = max(timeout_sec, int(max_tokens * 0.2))
     timeout = aiohttp.ClientTimeout(total=timeout_sec)
 
     valid_response = True
@@ -418,7 +409,6 @@ async def send_turn(
         req_args.stream,
         min_tokens,
         max_tokens,
-        req_args.timeout_sec,
     )
 
     if response.valid is False:
@@ -528,25 +518,6 @@ async def poisson_sleep(request_rate: float, verbose: bool = False) -> None:
     await asyncio.sleep(interval)
 
 
-async def exponential_backoff_sleep(
-    attempt_cnt: int,
-    base_rate: float = 1.0,
-    backoff_factor: float = 2.0,
-    jitter_fraction: float = 0.10,
-    verbose: bool = False,
-) -> None:
-    # Sleep with exponential backoff and jitter after a failed request.
-    backoff_delay = base_rate * (backoff_factor**attempt_cnt)
-    jittered_delay = backoff_delay * (
-        1 + np.random.uniform(-jitter_fraction, jitter_fraction)
-    )
-
-    if verbose:
-        logger.info(f"Backoff for {jittered_delay:.3f} seconds...")
-
-    await asyncio.sleep(jittered_delay)
-
-
 async def client_main(
     args: ClientArgs,
     req_args: RequestArgs,
@@ -561,11 +532,8 @@ async def client_main(
         f"{Color.CYAN}Started client {client_id}: max_num_requests={args.max_num_requests}, max_active_conversations={args.max_active_conversations}{Color.RESET}"  # noqa: E501
     )
 
-    # Set unique seed per client (each client runs in its own process)
-    # Add 1 to ensure no client uses the same seed as the main process
-    client_seed = args.seed + client_id + 1
-    random.seed(client_seed)
-    np.random.seed(client_seed)
+    random.seed(args.seed)
+    np.random.seed(args.seed)
 
     # Active conversations
     active_convs: ConversationsMap = {}
@@ -678,62 +646,49 @@ async def client_main(
                 )
                 time_of_last_turn[conv_id] = curr_time_sec
 
-            success = False
-            for attempt_cnt in range(args.max_retries + 1):
-                try:
-                    exception = False
-                    result = await send_turn(
-                        session,
-                        client_id,
-                        conv_id,
-                        messages,
-                        current_turn,
-                        tokenizer,
-                        req_args,
-                        args.print_content,
-                        args.verify_output,
-                    )
-                    if result is not None:
-                        result_queue.put(result)
-                        success = True
-                        break
-                    else:
-                        logger.warning(
-                            f"{Color.YELLOW}Client {client_id} - Request rejected during conversation ID {conv_id} (turn: {current_turn}){Color.RESET}"  # noqa: E501
-                        )
-                except asyncio.exceptions.TimeoutError:
-                    exception = True
-                    logger.error(
-                        "%sClient %d - Timeout during conversation ID %s (turn: %d). "
-                        "Base timeout is %ss (set with --request-timeout-sec), but the "
-                        "effective timeout may be longer based on max_tokens. If this "
-                        "is unexpected, consider increasing the timeout or checking "
-                        "model performance.%s",
-                        Color.RED,
-                        client_id,
-                        conv_id,
-                        current_turn,
-                        req_args.timeout_sec,
-                        Color.RESET,
-                    )
-                except Exception:
-                    exception = True
-                    logger.exception(
-                        f"{Color.RED}Client {client_id} - Exception during conversation ID {conv_id} (turn: {current_turn}){Color.RESET}"  # noqa: E501
+            success = True
+            try:
+                result = await send_turn(
+                    session,
+                    client_id,
+                    conv_id,
+                    messages,
+                    current_turn,
+                    tokenizer,
+                    req_args,
+                    args.print_content,
+                    args.verify_output,
+                )
+                if result is not None:
+                    result_queue.put(result)
+                else:
+                    # None means that the request failed,
+                    # and should not be added to the statistics.
+                    success = False
+                    num_failures += 1
+
+                    logger.warning(
+                        f"{Color.YELLOW}Client {client_id} - Request rejected during conversation ID {conv_id} (turn: {current_turn}){Color.RESET}"  # noqa: E501
                     )
 
-                # Sleep before retry if not last attempt
-                if not success and attempt_cnt < args.max_retries:
-                    await exponential_backoff_sleep(attempt_cnt, verbose=args.verbose)
+                    # Remove the conversation (should not be used again)
+                    active_convs.pop(conv_id)
 
-            if not success:
+            except asyncio.exceptions.TimeoutError:
                 num_failures += 1
-                # Remove the conversation (should not be used again)
-                active_convs.pop(conv_id)
-                if exception:
-                    break  # Exit gracefully instead of raising an error
+                logger.exception(
+                    f"{Color.RED}Client {client_id} - Timeout during conversation ID {conv_id} (turn: {current_turn}){Color.RESET}"  # noqa: E501
+                )
+                break  # Exit gracefully instead of raising an error
 
-            else:
+            except Exception:
+                num_failures += 1
+                logger.exception(
+                    f"{Color.RED}Client {client_id} - Exception during conversation ID {conv_id} (turn: {current_turn}){Color.RESET}"  # noqa: E501
+                )
+                break  # Exit gracefully instead of raising an error
+
+            if success:
                 num_successes += 1
 
                 # Update the turns counter to include the LLM response
@@ -848,7 +803,6 @@ def get_client_config(
         verify_output=args.verify_output,
         conversation_sampling=args.conversation_sampling,
         request_rate=args.request_rate,
-        max_retries=args.max_retries,
     )
 
     if args.limit_min_tokens > 0 or args.limit_max_tokens > 0:
@@ -861,9 +815,6 @@ def get_client_config(
                 "Invalid min/max tokens limits (min should not be larger than max)"
             )
 
-    if args.request_timeout_sec <= 0:
-        raise ValueError("Request timeout must be a positive number")
-
     # Arguments for API requests
     chat_url = f"{args.url}/v1/chat/completions"
     model_name = args.served_model_name if args.served_model_name else args.model
@@ -874,7 +825,6 @@ def get_client_config(
         stream=not args.no_stream,
         limit_min_tokens=args.limit_min_tokens,
         limit_max_tokens=args.limit_max_tokens,
-        timeout_sec=args.request_timeout_sec,
     )
 
     return client_args, req_args
@@ -1018,7 +968,7 @@ async def main_mp(
             f"(is alive: {client.is_alive()}){Color.RESET}"
         )
 
-        client.join(timeout=req_args.timeout_sec + 1)
+        client.join(timeout=120)
 
         if client.is_alive():
             logger.warning(
@@ -1076,7 +1026,6 @@ def process_statistics(
     verbose: bool,
     gen_conv_args: GenConvArgs | None = None,
     excel_output: bool = False,
-    warmup_runtime_sec: float | None = None,
 ) -> None:
     if len(client_metrics) == 0:
         logger.info("No samples to process")
@@ -1170,13 +1119,8 @@ def process_statistics(
         # Convert milliseconds to seconds
         runtime_sec = runtime_sec / 1000.0
         requests_per_sec = float(len(df)) / runtime_sec
-        params = {
-            "runtime_sec": runtime_sec,
-            "requests_per_sec": requests_per_sec,
-        }
-        if warmup_runtime_sec is not None:
-            params["warmup_runtime_sec"] = warmup_runtime_sec
-            params["total_runtime_incl_warmup_sec"] = runtime_sec + warmup_runtime_sec
+
+        params = {"runtime_sec": runtime_sec, "requests_per_sec": requests_per_sec}
 
         # Generate a summary of relevant metrics (and drop irrelevant data)
         df = df.drop(columns=exclude).describe(percentiles=percentiles).transpose()
@@ -1390,16 +1334,6 @@ async def main() -> None:
         help="Expected request rate (Poisson process) per client in requests/sec."
         "Set to 0 for no delay between requests.",
     )
-    parser.add_argument(
-        "--max-retries",
-        type=int,
-        default=int(os.environ.get("MULTITURN_BENCH_MAX_RETRIES", "0")),
-        help="Maximum number of retry attempts for timed-out requests. "
-        "Default is 0 (no retries). "
-        "Set to higher values to retry failed requests and maintain "
-        "fair workload distribution. "
-        "Can also be set via MULTITURN_BENCH_MAX_RETRIES environment variable.",
-    )
     parser.add_argument(
         "--conversation-sampling",
         type=ConversationSampling,
@@ -1417,13 +1351,6 @@ async def main() -> None:
         action="store_true",
         help="Verify the LLM output (compare to the answers in the input JSON file)",
     )
-    parser.add_argument(
-        "--request-timeout-sec",
-        type=int,
-        default=120,
-        help="Timeout in seconds for each API request (default: 120). "
-        "Automatically increased if max tokens imply longer decoding.",
-    )
 
     parser.add_argument(
         "--no-stream",
@@ -1499,7 +1426,6 @@ async def main() -> None:
             f"Invalid --warmup-percentage={args.warmup_percentage}"
         ) from None
 
-    # Set global seeds for main process
     random.seed(args.seed)
     np.random.seed(args.seed)
 
@@ -1558,8 +1484,6 @@ async def main() -> None:
         url=args.url, num_clients=args.num_clients, early_stop=not args.no_early_stop
     )
 
-    warmup_runtime_sec: float | None = None
-
     # Warm-up step
     if args.warmup_step:
         # Only send a single user prompt from every conversation.
@@ -1574,56 +1498,26 @@ async def main() -> None:
         # all clients should finish their work before exiting
         warmup_bench_args = bench_args._replace(early_stop=False)
 
-        logger.info("%sWarmup start%s", Color.PURPLE, Color.RESET)
-        warmup_start_ns = time.perf_counter_ns()
+        logger.info(f"{Color.PURPLE}Warmup start{Color.RESET}")
         conversations, _ = await main_mp(
             warmup_client_args, req_args, warmup_bench_args, tokenizer, conversations
         )
-        warmup_runtime_sec = nanosec_to_sec(time.perf_counter_ns() - warmup_start_ns)
-        logger.info(
-            "%sWarmup runtime: %.3f sec (%.3f ms)%s",
-            Color.PURPLE,
-            warmup_runtime_sec,
-            warmup_runtime_sec * 1000,
-            Color.RESET,
-        )
-        logger.info("%sWarmup done%s", Color.PURPLE, Color.RESET)
+        logger.info(f"{Color.PURPLE}Warmup done{Color.RESET}")
 
     # Run the benchmark
-    benchmark_start_ns = time.perf_counter_ns()
+    start_time = time.perf_counter_ns()
     client_convs, client_metrics = await main_mp(
         client_args, req_args, bench_args, tokenizer, conversations
     )
-    benchmark_runtime_sec = nanosec_to_sec(time.perf_counter_ns() - benchmark_start_ns)
+    total_runtime_ms = nanosec_to_millisec(time.perf_counter_ns() - start_time)
 
     # Calculate requests per second
-    requests_per_sec = len(client_metrics) / benchmark_runtime_sec
-    benchmark_runtime_ms = benchmark_runtime_sec * 1000.0
+    total_runtime_sec = total_runtime_ms / 1000.0
+    rps = len(client_metrics) / total_runtime_sec
     logger.info(
-        "%sAll clients finished, benchmark runtime: %.3f sec (%.3f ms), "
-        "requests per second: %.3f%s",
-        Color.GREEN,
-        benchmark_runtime_sec,
-        benchmark_runtime_ms,
-        requests_per_sec,
-        Color.RESET,
+        f"{Color.GREEN}All clients finished, total runtime: {total_runtime_sec:.3f} sec"
+        f" ({total_runtime_ms:.3f} ms), requests per second: {rps:.3f}{Color.RESET}"
     )
-    if warmup_runtime_sec is not None:
-        total_runtime_sec = benchmark_runtime_sec + warmup_runtime_sec
-        logger.info(
-            "%sWarmup runtime: %.3f sec (%.3f ms)%s",
-            Color.GREEN,
-            warmup_runtime_sec,
-            warmup_runtime_sec * 1000,
-            Color.RESET,
-        )
-        logger.info(
-            "%sTotal runtime (including warmup): %.3f sec (%.3f ms)%s",
-            Color.GREEN,
-            total_runtime_sec,
-            total_runtime_sec * 1000,
-            Color.RESET,
-        )
 
     # Benchmark parameters
     params = {
@@ -1648,7 +1542,6 @@ async def main() -> None:
         verbose=args.verbose,
         gen_conv_args=gen_conv_args,
         excel_output=args.excel_output,
-        warmup_runtime_sec=warmup_runtime_sec,
     )
 
     if args.output_file is not None:

benchmarks/multi_turn/requirements.txt

@@ -2,5 +2,4 @@ numpy>=1.24
 pandas>=2.0.0
 aiohttp>=3.10
 transformers>=4.46
-xlsxwriter>=3.2.1
-tqdm>=4.66
+xlsxwriter>=3.2.1

cmake/cpu_extension.cmake

@@ -15,7 +15,6 @@ endif()
 #
 set(ENABLE_AVX512BF16 $ENV{VLLM_CPU_AVX512BF16})
 set(ENABLE_AVX512VNNI $ENV{VLLM_CPU_AVX512VNNI})
-set(ENABLE_AMXBF16 $ENV{VLLM_CPU_AMXBF16})
 
 include_directories("${CMAKE_SOURCE_DIR}/csrc")
 
@@ -141,22 +140,6 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         set(ENABLE_AVX512VNNI OFF)
         message(WARNING "Disable AVX512-VNNI ISA support, no avx512_vnni found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512VNNI=1.")
     endif()
-
-    find_isa(${CPUINFO} "amx_bf16" AMXBF16_FOUND)
-    if (AMXBF16_FOUND OR ENABLE_AMXBF16)
-        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
-            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
-            list(APPEND CXX_COMPILE_FLAGS "-mamx-bf16" "-mamx-tile")
-            set(ENABLE_AMXBF16 ON)
-            add_compile_definitions(-DCPU_CAPABILITY_AMXBF16)
-        else()
-            set(ENABLE_AMXBF16 OFF)
-            message(WARNING "Disable AMX_BF16 ISA support, requires gcc/g++ >= 12.3")
-        endif()
-    else()
-        set(ENABLE_AMXBF16 OFF)
-        message(WARNING "Disable AMX_BF16 ISA support, no amx_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AMXBF16=1.")
-    endif()
     
 elseif (AVX2_FOUND)
     list(APPEND CXX_COMPILE_FLAGS "-mavx2")
@@ -210,30 +193,7 @@ endif()
 if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
     # Fetch and build Arm Compute Library (ACL) as oneDNN's backend for AArch64
     # TODO [fadara01]: remove this once ACL can be fetched and built automatically as a dependency of oneDNN
-    set(ONEDNN_AARCH64_USE_ACL OFF CACHE BOOL "")
     if(ASIMD_FOUND)
-        # Set number of parallel build processes
-        include(ProcessorCount)
-        ProcessorCount(NPROC)
-        if(NOT NPROC)
-            set(NPROC 4)
-        endif()
-        # locate PyTorch's libgomp (e.g. site-packages/torch.libs/libgomp-947d5fa1.so.1.0.0)
-        # and create a local shim dir with it
-        vllm_prepare_torch_gomp_shim(VLLM_TORCH_GOMP_SHIM_DIR)
-
-        find_library(OPEN_MP
-            NAMES gomp
-            PATHS ${VLLM_TORCH_GOMP_SHIM_DIR}
-            NO_DEFAULT_PATH
-            REQUIRED
-        )
-        # Set LD_LIBRARY_PATH to include the shim dir at build time to use the same libgomp as PyTorch
-        if (OPEN_MP)
-            set(ENV{LD_LIBRARY_PATH} "${VLLM_TORCH_GOMP_SHIM_DIR}:$ENV{LD_LIBRARY_PATH}")
-        endif()
-
-        # Fetch and populate ACL
         if(DEFINED ENV{ACL_ROOT_DIR} AND IS_DIRECTORY "$ENV{ACL_ROOT_DIR}")
             message(STATUS "Using ACL from specified source directory: $ENV{ACL_ROOT_DIR}")
         else()
@@ -242,58 +202,43 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
                 SUBBUILD_DIR "${FETCHCONTENT_BASE_DIR}/arm_compute-subbuild"
                 SOURCE_DIR   "${FETCHCONTENT_BASE_DIR}/arm_compute-src"
                 GIT_REPOSITORY https://github.com/ARM-software/ComputeLibrary.git
-                GIT_TAG        v52.6.0
+                GIT_TAG        v52.2.0
                 GIT_SHALLOW    TRUE
                 GIT_PROGRESS   TRUE
             )
             set(ENV{ACL_ROOT_DIR} "${arm_compute_SOURCE_DIR}")
-            set(ACL_LIB_DIR "$ENV{ACL_ROOT_DIR}/build")
         endif()
 
-        # Build ACL with CMake
-        set(ARM_COMPUTE_BUILD_SHARED_LIB "OFF")
-        set(CMAKE_BUILD_TYPE "Release")
-        set(ARM_COMPUTE_ARCH "armv8.2-a")
-        set(ARM_COMPUTE_ENABLE_ASSERTS "OFF")
-        set(ARM_COMPUTE_ENABLE_CPPTHREADS "OFF")
-        set(ONEDNN_ENABLE_PRIMITIVE "MATMUL;REORDER")
-        set(ARM_COMPUTE_ENABLE_OPENMP "ON")
-        set(ARM_COMPUTE_ENABLE_WERROR "OFF")
-        set(ARM_COMPUTE_BUILD_EXAMPLES "OFF")
-        set(ARM_COMPUTE_BUILD_TESTING "OFF")
-
-        set(_cmake_config_cmd
-             ${CMAKE_COMMAND} -G Ninja -B build 
-            -DARM_COMPUTE_BUILD_SHARED_LIB=OFF 
-            -DCMAKE_BUILD_TYPE=Release 
-            -DARM_COMPUTE_ARCH=armv8.2-a 
-            -DARM_COMPUTE_ENABLE_ASSERTS=OFF 
-            -DARM_COMPUTE_ENABLE_CPPTHREADS=OFF 
-            -DARM_COMPUTE_ENABLE_OPENMP=ON 
-            -DARM_COMPUTE_ENABLE_WERROR=OFF 
-            -DARM_COMPUTE_BUILD_EXAMPLES=OFF 
-            -DARM_COMPUTE_BUILD_TESTING=OFF)
-        set(_cmake_build_cmd
-            ${CMAKE_COMMAND} --build build -- -j${NPROC}
+        # Build ACL with scons
+        include(ProcessorCount)
+        ProcessorCount(_NPROC)
+        set(_scons_cmd
+        scons -j${_NPROC}
+            Werror=0 debug=0 neon=1 examples=0 embed_kernels=0 os=linux
+            arch=armv8.2-a build=native benchmark_examples=0 fixed_format_kernels=1
+            multi_isa=1 openmp=1 cppthreads=0
         )
 
+        # locate PyTorch's libgomp (e.g. site-packages/torch.libs/libgomp-947d5fa1.so.1.0.0)
+        # and create a local shim dir with it
+        include("${CMAKE_CURRENT_LIST_DIR}/utils.cmake")
+        vllm_prepare_torch_gomp_shim(VLLM_TORCH_GOMP_SHIM_DIR)
+
+        if(NOT VLLM_TORCH_GOMP_SHIM_DIR STREQUAL "")
+            list(APPEND _scons_cmd extra_link_flags=-L${VLLM_TORCH_GOMP_SHIM_DIR})
+        endif()
+
         execute_process(
-            COMMAND ${_cmake_config_cmd}
-            WORKING_DIRECTORY "$ENV{ACL_ROOT_DIR}"
-        )
-        execute_process(
-            COMMAND ${_cmake_build_cmd}
+            COMMAND ${_scons_cmd}
             WORKING_DIRECTORY "$ENV{ACL_ROOT_DIR}"
             RESULT_VARIABLE _acl_rc
         )
-
         if(NOT _acl_rc EQUAL 0)
             message(FATAL_ERROR "ACL SCons build failed (exit ${_acl_rc}).")
         endif()
-        message(STATUS "Arm Compute Library (ACL) built successfully.")
 
-        # VLLM/oneDNN settings for ACL
-        set(ONEDNN_AARCH64_USE_ACL ON CACHE BOOL "" FORCE)
+        set(ONEDNN_AARCH64_USE_ACL "ON")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wl,-rpath,$ENV{ACL_ROOT_DIR}/build/")
         add_compile_definitions(VLLM_USE_ACL)
     endif()
 
@@ -310,7 +255,7 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
         FetchContent_Declare(
             oneDNN
             GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-            GIT_TAG v3.10
+            GIT_TAG v3.9
             GIT_PROGRESS TRUE
             GIT_SHALLOW TRUE
         )
@@ -330,10 +275,7 @@ if ((AVX512_FOUND AND NOT AVX512_DISABLED) OR (ASIMD_FOUND AND NOT APPLE_SILICON
     set(ONEDNN_VERBOSE "OFF")
     set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
 
-    set(VLLM_BUILD_TYPE ${CMAKE_BUILD_TYPE})
-    set(CMAKE_BUILD_TYPE "Release") # remove oneDNN debug symbols to reduce size
     FetchContent_MakeAvailable(oneDNN)
-    set(CMAKE_BUILD_TYPE ${VLLM_BUILD_TYPE})
     add_library(dnnl_ext OBJECT "csrc/cpu/dnnl_helper.cpp")
     target_include_directories(
         dnnl_ext
@@ -363,19 +305,18 @@ endif()
 #
 set(VLLM_EXT_SRC
     "csrc/cpu/activation.cpp"
+    "csrc/cpu/attention.cpp"
+    "csrc/cpu/cache.cpp"
     "csrc/cpu/utils.cpp"
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
-    "csrc/moe/dynamic_4bit_int_moe_cpu.cpp"
-    "csrc/cpu/cpu_attn.cpp"
-    "csrc/cpu/scratchpad_manager.cpp"
-    "csrc/cpu/torch_bindings.cpp")
+    "csrc/cpu/torch_bindings.cpp"
+    "csrc/moe/dynamic_4bit_int_moe_cpu.cpp")
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     set(VLLM_EXT_SRC
         "csrc/cpu/shm.cpp"
-        "csrc/cpu/cpu_wna16.cpp"
         ${VLLM_EXT_SRC})
     if (ENABLE_AVX512BF16 AND ENABLE_AVX512VNNI)
         set(VLLM_EXT_SRC
@@ -402,7 +343,7 @@ message(STATUS "CPU extension source files: ${VLLM_EXT_SRC}")
 # Define extension targets
 #
 
-define_extension_target(
+define_gpu_extension_target(
     _C
     DESTINATION vllm
     LANGUAGE CXX
@@ -413,4 +354,4 @@ define_extension_target(
     WITH_SOABI
 )
 
-message(STATUS "Enabling C extension.")
+message(STATUS "Enabling C extension.")

cmake/external_projects/flashmla.cmake

@@ -92,7 +92,7 @@ if(FLASH_MLA_ARCHS)
         SRCS "${FlashMLA_Extension_SOURCES}"
         CUDA_ARCHS "${FLASH_MLA_ARCHS}")
 
-    define_extension_target(
+    define_gpu_extension_target(
         _flashmla_C
         DESTINATION vllm
         LANGUAGE ${VLLM_GPU_LANG}
@@ -109,7 +109,7 @@ if(FLASH_MLA_ARCHS)
         $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
         $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
 
-    define_extension_target(
+    define_gpu_extension_target(
         _flashmla_extension_C
         DESTINATION vllm
         LANGUAGE ${VLLM_GPU_LANG}

cmake/external_projects/triton_kernels.cmake

@@ -1,53 +0,0 @@
-# Install OpenAI triton_kernels from https://github.com/triton-lang/triton/tree/main/python/triton_kernels
-
-set(DEFAULT_TRITON_KERNELS_TAG "v3.5.0")
-
-# Set TRITON_KERNELS_SRC_DIR for use with local development with vLLM. We expect TRITON_KERNELS_SRC_DIR to
-# be directly set to the triton_kernels python directory. 
-if (DEFINED ENV{TRITON_KERNELS_SRC_DIR})
-  message(STATUS "[triton_kernels] Fetch from $ENV{TRITON_KERNELS_SRC_DIR}")
-  FetchContent_Declare(
-          triton_kernels
-          SOURCE_DIR $ENV{TRITON_KERNELS_SRC_DIR}
-  )
-
-else()
-  set(TRITON_GIT "https://github.com/triton-lang/triton.git")
-  message (STATUS "[triton_kernels] Fetch from ${TRITON_GIT}:${DEFAULT_TRITON_KERNELS_TAG}")
-  FetchContent_Declare(
-          triton_kernels
-          # TODO (varun) : Fetch just the triton_kernels directory from Triton
-          GIT_REPOSITORY https://github.com/triton-lang/triton.git
-          GIT_TAG ${DEFAULT_TRITON_KERNELS_TAG}
-          GIT_PROGRESS TRUE
-          SOURCE_SUBDIR python/triton_kernels/triton_kernels
-  )
-endif()
-
-# Fetch content 
-FetchContent_MakeAvailable(triton_kernels)
-
-if (NOT triton_kernels_SOURCE_DIR)
-  message (FATAL_ERROR "[triton_kernels] Cannot resolve triton_kernels_SOURCE_DIR")
-endif()
-
-if (DEFINED ENV{TRITON_KERNELS_SRC_DIR})
-  set(TRITON_KERNELS_PYTHON_DIR "${triton_kernels_SOURCE_DIR}/")
-else()
-  set(TRITON_KERNELS_PYTHON_DIR "${triton_kernels_SOURCE_DIR}/python/triton_kernels/triton_kernels/")
-endif()
-
-message (STATUS "[triton_kernels] triton_kernels is available at ${TRITON_KERNELS_PYTHON_DIR}")
-
-add_custom_target(triton_kernels)
-
-# Ensure the vllm/third_party directory exists before installation
-install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/third_party/triton_kernels\")")
-
-## Copy .py files to install directory.
-install(DIRECTORY
-        ${TRITON_KERNELS_PYTHON_DIR}
-        DESTINATION 
-        vllm/third_party/triton_kernels/
-        COMPONENT triton_kernels
-        FILES_MATCHING PATTERN "*.py")

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 86f8f157cf82aa2342743752b97788922dd7de43
+          GIT_TAG a893712401d70362fbb299cd9c4b3476e8e9ed54
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

@@ -453,20 +453,21 @@ macro(override_gpu_arches GPU_ARCHES GPU_LANG GPU_SUPPORTED_ARCHES)
 endmacro()
 
 #
-# Define a target named `MOD_NAME` for a single extension. The
+# Define a target named `GPU_MOD_NAME` for a single extension. The
 # arguments are:
 #
 # DESTINATION <dest>         - Module destination directory.
-# LANGUAGE <lang>            - The language for this module, e.g. CUDA, HIP,
-#                              CXX, etc.
+# LANGUAGE <lang>            - The GPU language for this module, e.g CUDA, HIP,
+#                              etc.
 # SOURCES <sources>          - List of source files relative to CMakeLists.txt
 #                              directory.
 #
 # Optional arguments:
 #
-# ARCHITECTURES <arches>     - A list of target architectures in cmake format.
-#                              For GPU, refer to CMAKE_CUDA_ARCHITECTURES and
-#                              CMAKE_HIP_ARCHITECTURES for more info.
+# ARCHITECTURES <arches>     - A list of target GPU architectures in cmake
+#                              format.
+#                              Refer `CMAKE_CUDA_ARCHITECTURES` documentation
+#                              and `CMAKE_HIP_ARCHITECTURES` for more info.
 #                              ARCHITECTURES will use cmake's defaults if
 #                              not provided.
 # COMPILE_FLAGS <flags>      - Extra compiler flags passed to NVCC/hip.
@@ -477,67 +478,63 @@ endmacro()
 #
 # Note: optimization level/debug info is set via cmake build type.
 #
-function (define_extension_target MOD_NAME)
+function (define_gpu_extension_target GPU_MOD_NAME)
   cmake_parse_arguments(PARSE_ARGV 1
-    ARG
+    GPU
     "WITH_SOABI"
     "DESTINATION;LANGUAGE;USE_SABI"
     "SOURCES;ARCHITECTURES;COMPILE_FLAGS;INCLUDE_DIRECTORIES;LIBRARIES")
 
   # Add hipify preprocessing step when building with HIP/ROCm.
-  if (ARG_LANGUAGE STREQUAL "HIP")
-    hipify_sources_target(ARG_SOURCES ${MOD_NAME} "${ARG_SOURCES}")
+  if (GPU_LANGUAGE STREQUAL "HIP")
+    hipify_sources_target(GPU_SOURCES ${GPU_MOD_NAME} "${GPU_SOURCES}")
   endif()
 
-  if (ARG_WITH_SOABI)
-    set(SOABI_KEYWORD WITH_SOABI)
+  if (GPU_WITH_SOABI)
+    set(GPU_WITH_SOABI WITH_SOABI)
   else()
-    set(SOABI_KEYWORD "")
+    set(GPU_WITH_SOABI)
   endif()
 
-  run_python(IS_FREETHREADED_PYTHON
-    "import sysconfig; print(1 if sysconfig.get_config_var(\"Py_GIL_DISABLED\") else 0)"
-    "Failed to determine whether interpreter is free-threaded")
-
-  # Free-threaded Python doesn't yet support the stable ABI (see PEP 803/809),
-  # so avoid using the stable ABI under free-threading only.
-  if (ARG_USE_SABI AND NOT IS_FREETHREADED_PYTHON)
-    Python_add_library(${MOD_NAME} MODULE USE_SABI ${ARG_USE_SABI} ${SOABI_KEYWORD} "${ARG_SOURCES}")
+  if (GPU_USE_SABI)
+    Python_add_library(${GPU_MOD_NAME} MODULE USE_SABI ${GPU_USE_SABI} ${GPU_WITH_SOABI} "${GPU_SOURCES}")
   else()
-    Python_add_library(${MOD_NAME} MODULE ${SOABI_KEYWORD} "${ARG_SOURCES}")
+    Python_add_library(${GPU_MOD_NAME} MODULE ${GPU_WITH_SOABI} "${GPU_SOURCES}")
   endif()
 
-  if (ARG_LANGUAGE STREQUAL "HIP")
+  if (GPU_LANGUAGE STREQUAL "HIP")
     # Make this target dependent on the hipify preprocessor step.
-    add_dependencies(${MOD_NAME} hipify${MOD_NAME})
+    add_dependencies(${GPU_MOD_NAME} hipify${GPU_MOD_NAME})
     # Make sure we include the hipified versions of the headers, and avoid conflicts with the ones in the original source folder
-    target_include_directories(${MOD_NAME} PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/csrc
-      ${ARG_INCLUDE_DIRECTORIES})
+    target_include_directories(${GPU_MOD_NAME} PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/csrc
+      ${GPU_INCLUDE_DIRECTORIES})
   else()
-    target_include_directories(${MOD_NAME} PRIVATE csrc
-      ${ARG_INCLUDE_DIRECTORIES})
+    target_include_directories(${GPU_MOD_NAME} PRIVATE csrc
+      ${GPU_INCLUDE_DIRECTORIES})
   endif()
 
-  if (ARG_ARCHITECTURES)
-    set_target_properties(${MOD_NAME} PROPERTIES
-      ${ARG_LANGUAGE}_ARCHITECTURES "${ARG_ARCHITECTURES}")
+  if (GPU_ARCHITECTURES)
+    set_target_properties(${GPU_MOD_NAME} PROPERTIES
+      ${GPU_LANGUAGE}_ARCHITECTURES "${GPU_ARCHITECTURES}")
   endif()
 
-  target_compile_options(${MOD_NAME} PRIVATE
-    $<$<COMPILE_LANGUAGE:${ARG_LANGUAGE}>:${ARG_COMPILE_FLAGS}>)
 
-  target_compile_definitions(${MOD_NAME} PRIVATE
-    "-DTORCH_EXTENSION_NAME=${MOD_NAME}")
+  target_compile_options(${GPU_MOD_NAME} PRIVATE
+    $<$<COMPILE_LANGUAGE:${GPU_LANGUAGE}>:${GPU_COMPILE_FLAGS}>)
 
-  target_link_libraries(${MOD_NAME} PRIVATE torch ${ARG_LIBRARIES})
+  target_compile_definitions(${GPU_MOD_NAME} PRIVATE
+    "-DTORCH_EXTENSION_NAME=${GPU_MOD_NAME}")
+
+
+  target_link_libraries(${GPU_MOD_NAME} PRIVATE torch ${GPU_LIBRARIES})
 
   # Don't use `TORCH_LIBRARIES` for CUDA since it pulls in a bunch of
   # dependencies that are not necessary and may not be installed.
-  if (ARG_LANGUAGE STREQUAL "CUDA")
-    target_link_libraries(${MOD_NAME} PRIVATE torch CUDA::cudart CUDA::cuda_driver ${ARG_LIBRARIES})
+  if (GPU_LANGUAGE STREQUAL "CUDA")
+    target_link_libraries(${GPU_MOD_NAME} PRIVATE CUDA::cudart CUDA::cuda_driver)
   else()
-    target_link_libraries(${MOD_NAME} PRIVATE torch ${TORCH_LIBRARIES} ${ARG_LIBRARIES})
+    target_link_libraries(${GPU_MOD_NAME} PRIVATE ${TORCH_LIBRARIES})
   endif()
 
-  install(TARGETS ${MOD_NAME} LIBRARY DESTINATION ${ARG_DESTINATION} COMPONENT ${MOD_NAME})
+  install(TARGETS ${GPU_MOD_NAME} LIBRARY DESTINATION ${GPU_DESTINATION} COMPONENT ${GPU_MOD_NAME})
 endfunction()

csrc/attention/merge_attn_states.cu

@@ -16,8 +16,7 @@ __global__ void merge_attn_states_kernel(
     scalar_t* output, float* output_lse, const scalar_t* prefix_output,
     const float* prefix_lse, const scalar_t* suffix_output,
     const float* suffix_lse, const uint num_tokens, const uint num_heads,
-    const uint head_size, const uint prefix_head_stride,
-    const uint output_head_stride) {
+    const uint head_size) {
   using pack_128b_t = uint4;
   const uint pack_size = 16 / sizeof(scalar_t);
   const uint threads_per_head = head_size / pack_size;
@@ -35,13 +34,11 @@ __global__ void merge_attn_states_kernel(
   const uint head_idx = token_head_idx % num_heads;
 
   const uint pack_offset = pack_idx * pack_size;  // (0~15)*8, etc.
-  const uint src_head_offset = token_idx * num_heads * prefix_head_stride +
-                               head_idx * prefix_head_stride;
-  const uint dst_head_offset = token_idx * num_heads * output_head_stride +
-                               head_idx * output_head_stride;
-  const scalar_t* prefix_head_ptr = prefix_output + src_head_offset;
-  const scalar_t* suffix_head_ptr = suffix_output + src_head_offset;
-  scalar_t* output_head_ptr = output + dst_head_offset;
+  const uint head_offset =
+      token_idx * num_heads * head_size + head_idx * head_size;
+  const scalar_t* prefix_head_ptr = prefix_output + head_offset;
+  const scalar_t* suffix_head_ptr = suffix_output + head_offset;
+  scalar_t* output_head_ptr = output + head_offset;
 
   float p_lse = prefix_lse[head_idx * num_tokens + token_idx];
   float s_lse = suffix_lse[head_idx * num_tokens + token_idx];
@@ -49,32 +46,6 @@ __global__ void merge_attn_states_kernel(
   s_lse = std::isinf(s_lse) ? -std::numeric_limits<float>::infinity() : s_lse;
 
   const float max_lse = fmaxf(p_lse, s_lse);
-
-  /* In certain edge cases, MLA can produce p_lse = s_lse = -inf;
-     continuing the pipeline then yields NaN. Root cause: with chunked prefill
-     a batch may be split into two chunks; if a request in that batch has no
-     prefix hit, every LSE entry for that request’s position is -inf, and at
-     this moment we merge cross-attention at first. For now we simply emit
-     prefix_output (expected to be all zeros) and prefix_lse (-inf) to fix
-     this problem.
-  */
-  if (std::isinf(max_lse)) {
-    if (pack_offset < head_size) {
-      // Pack 128b load
-      pack_128b_t p_out_pack = reinterpret_cast<const pack_128b_t*>(
-          prefix_head_ptr)[pack_offset / pack_size];
-
-      // Pack 128b storage
-      reinterpret_cast<pack_128b_t*>(output_head_ptr)[pack_offset / pack_size] =
-          p_out_pack;
-    }
-    // We only need to write to output_lse once per head.
-    if (output_lse != nullptr && pack_idx == 0) {
-      output_lse[head_idx * num_tokens + token_idx] = max_lse;
-    }
-    return;
-  }
-
   p_lse = p_lse - max_lse;
   s_lse = s_lse - max_lse;
   const float p_se = expf(p_lse);
@@ -143,7 +114,7 @@ __global__ void merge_attn_states_kernel(
             reinterpret_cast<float*>(prefix_lse.data_ptr()),                \
             reinterpret_cast<scalar_t*>(suffix_output.data_ptr()),          \
             reinterpret_cast<float*>(suffix_lse.data_ptr()), num_tokens,    \
-            num_heads, head_size, prefix_head_stride, output_head_stride);  \
+            num_heads, head_size);                                          \
   }
 
 /*@brief Merges the attention states from prefix and suffix
@@ -169,11 +140,17 @@ void merge_attn_states_launcher(torch::Tensor& output,
   const uint num_tokens = output.size(0);
   const uint num_heads = output.size(1);
   const uint head_size = output.size(2);
-  const uint prefix_head_stride = prefix_output.stride(1);
-  const uint output_head_stride = output.stride(1);
   const uint pack_size = 16 / sizeof(scalar_t);
   TORCH_CHECK(head_size % pack_size == 0,
               "headsize must be multiple of pack_size:", pack_size);
+  TORCH_CHECK(output.stride(-2) == head_size && output.stride(-1) == 1,
+              "output heads must be contiguous in memory");
+  TORCH_CHECK(
+      prefix_output.stride(-2) == head_size && prefix_output.stride(-1) == 1,
+      "prefix_output heads must be contiguous in memory");
+  TORCH_CHECK(
+      suffix_output.stride(-2) == head_size && suffix_output.stride(-1) == 1,
+      "suffix_output heads must be contiguous in memory");
   float* output_lse_ptr = nullptr;
   if (output_lse.has_value()) {
     output_lse_ptr = output_lse.value().data_ptr<float>();

csrc/cache.h

@@ -41,12 +41,11 @@ void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
                  const double scale, const std::string& kv_cache_dtype);
 
 void gather_and_maybe_dequant_cache(
-    torch::Tensor const& src_cache,     // [NUM_BLOCKS, BLOCK_SIZE, ENTRIES...]
-    torch::Tensor const& dst,           // [TOT_TOKENS, ENTRIES...]
-    torch::Tensor const& block_table,   // [BATCH, BLOCK_INDICES]
-    torch::Tensor const& cu_seq_lens,   // [BATCH+1]
-    torch::Tensor const& token_to_seq,  // [MAX_TOKEN_ACROSS_CHUNKS]
-    int64_t num_tokens, const std::string& kv_cache_dtype,
+    torch::Tensor const& src_cache,    // [NUM_BLOCKS, BLOCK_SIZE, ENTRIES...]
+    torch::Tensor const& dst,          // [TOT_TOKENS, ENTRIES...]
+    torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
+    torch::Tensor const& cu_seq_lens,  // [BATCH+1]
+    int64_t batch_size, const std::string& kv_cache_dtype,
     torch::Tensor const& scale,
     std::optional<torch::Tensor> seq_starts = std::nullopt);
 

csrc/cache_kernels.cu

@@ -552,11 +552,7 @@ __global__ void indexer_k_quant_and_cache_kernel(
 #ifndef USE_ROCM
   __syncwarp();
 #endif
-#if defined(__gfx942__)
-  float scale = fmaxf(amax, 1e-4) / 224.0f;
-#else
   float scale = fmaxf(amax, 1e-4) / 448.0f;
-#endif
   if (use_ue8m0) {
     scale = exp2f(ceilf(log2f(scale)));
   }
@@ -905,80 +901,87 @@ void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
 namespace vllm {
 
 // grid is launched with dimensions (batch, num_splits)
-template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt,
-          int ENTRY_SIZE, int CTA_SIZE>
+template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
 __global__ void gather_and_maybe_dequant_cache(
-    const cache_t* __restrict__ src_cache,     // [NUM_BLOCKS, BLOCK_SIZE,
-                                               // ENTRIES...]
-    scalar_t* __restrict__ dst,                // [TOT_TOKENS, ENTRIES...]
-    const int32_t* __restrict__ block_table,   // [BATCH, BLOCK_INDICES]
-    const int32_t* __restrict__ cu_seq_lens,   // [BATCH+1]
-    const int32_t* __restrict__ token_to_seq,  // [MAX_TOKEN_ACROSS_CHUNK]
-    const int32_t num_tokens, const int32_t block_size,
+    const cache_t* __restrict__ src_cache,    // [NUM_BLOCKS, BLOCK_SIZE,
+                                              // ENTRIES...]
+    scalar_t* __restrict__ dst,               // [TOT_TOKENS, ENTRIES...]
+    const int32_t* __restrict__ block_table,  // [BATCH, BLOCK_INDICES]
+    const int32_t* __restrict__ cu_seq_lens,  // [BATCH+1]
+    const int32_t block_size, const int32_t entry_size,
     const int64_t block_table_stride, const int64_t cache_block_stride,
     const int64_t cache_entry_stride, const int64_t dst_entry_stride,
     const float* __restrict__ scale,
     const int32_t* __restrict__ seq_starts) {  // Optional: starting offsets per
                                                // batch
-  constexpr int vec_size = sizeof(float4) / sizeof(scalar_t);
-  using ltype = vllm::vec_n_t<cache_t, vec_size>;
-  using stype = vllm::vec_n_t<scalar_t, vec_size>;
-  // We are adding this for code readability which will be optimized out when
-  // build in release.
-  assert(CTA_SIZE == blockDim.x);
 
-#pragma unroll
-  for (int token_id = blockIdx.x; token_id < num_tokens;
-       token_id += gridDim.x) {
-    int64_t batch_id = token_to_seq[token_id];
-    int64_t batch_start = cu_seq_lens[batch_id];
-    int64_t batch_end = cu_seq_lens[batch_id + 1];
-    int32_t batch_offset = token_id - batch_start;
+  const int64_t bid = blockIdx.x;  // Batch ID
+  const int32_t num_splits = gridDim.y;
+  const int32_t split = blockIdx.y;
+  const int32_t seq_start = cu_seq_lens[bid];
+  const int32_t seq_end = cu_seq_lens[bid + 1];
+  const int32_t seq_len = seq_end - seq_start;
+  const int32_t tot_blocks = cuda_utils::ceil_div(seq_len, block_size);
+  const int32_t split_blocks = cuda_utils::ceil_div(tot_blocks, num_splits);
 
-    if (token_id >= batch_end) return;
-    int32_t offset = 0;
-    if (seq_starts != nullptr) {
-      offset = seq_starts[batch_id];
-    }
-    batch_offset += offset;
-    int32_t block_table_id = batch_offset / block_size;
-    int32_t slot_id = batch_offset % block_size;
-    int32_t block_table_offset = batch_id * block_table_stride + block_table_id;
-    int32_t block_id = block_table[block_table_offset];
-    int64_t cache_offset =
-        block_id * cache_block_stride + slot_id * cache_entry_stride;
-    constexpr int32_t vec_iter_cnt = ENTRY_SIZE / vec_size;
-    scalar_t* dst_ = dst + token_id * dst_entry_stride;
-    cache_t* src_ = const_cast<cache_t*>(src_cache) + cache_offset;
+  const int32_t split_start = split * split_blocks;
+  const int32_t split_end = min((split + 1) * split_blocks, tot_blocks);
 
-#pragma unroll
-    for (int idx = threadIdx.x; idx < vec_iter_cnt; idx += CTA_SIZE) {
+  const bool is_active_split = (split_start < tot_blocks);
+  const bool is_last_split = (split_end == tot_blocks);
+
+  if (!is_active_split) return;
+
+  int32_t full_blocks_end = split_end;
+  int32_t partial_block_size = 0;
+
+  // Adjust the pointer for the block_table for this batch.
+  // If seq_starts is provided, compute an offset based on (seq_starts[bid] /
+  // page_size)
+  const int32_t batch_offset = bid * block_table_stride;
+  int32_t offset = 0;
+  if (seq_starts != nullptr) {
+    offset = seq_starts[bid] / block_size;
+  }
+  const int32_t* batch_block_table = block_table + batch_offset + offset;
+
+  // Adjust dst pointer based on the cumulative sequence lengths.
+  dst += seq_start * dst_entry_stride;
+
+  if (is_last_split) {
+    partial_block_size = seq_len % block_size;
+    if (partial_block_size) full_blocks_end -= 1;
+  }
+
+  auto copy_entry = [&](const cache_t* __restrict__ _src,
+                        scalar_t* __restrict__ _dst) {
+    for (int i = threadIdx.x; i < entry_size; i += blockDim.x) {
       if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
-        reinterpret_cast<stype*>(dst_)[idx] =
-            static_cast<stype>(reinterpret_cast<ltype*>(src_)[idx]);
+        _dst[i] = static_cast<scalar_t>(_src[i]);
       } else {
-        ltype loaded_val = reinterpret_cast<ltype*>(src_)[idx];
-        stype store_val;
-#pragma unroll
-        for (int j = 0; j < vec_size; ++j) {
-          store_val.val[j] = fp8::scaled_convert<scalar_t, cache_t, kv_dt>(
-              loaded_val.val[j], *scale);
-        }
-        reinterpret_cast<stype*>(dst_)[idx] = store_val;
+        _dst[i] =
+            fp8::scaled_convert<scalar_t, cache_t, kv_dt>(_src[i], *scale);
       }
     }
-    // process tail
-    constexpr int32_t tail_cnt = ENTRY_SIZE % vec_size;
-    dst_ = dst_ + ENTRY_SIZE - tail_cnt;
-    src_ = src_ + ENTRY_SIZE - tail_cnt;
-#pragma unroll
-    for (int idx = threadIdx.x; idx < tail_cnt; idx += CTA_SIZE) {
-      if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
-        dst_[idx] = static_cast<scalar_t>(src_[idx]);
-      } else {
-        dst_[idx] =
-            fp8::scaled_convert<scalar_t, cache_t, kv_dt>(src_[idx], *scale);
-      }
+  };
+
+  for (int pid = split_start; pid < full_blocks_end; ++pid) {
+    auto block_id = batch_block_table[pid];
+    auto block_start_ptr = src_cache + block_id * cache_block_stride;
+    auto block_dst_ptr = dst + pid * block_size * dst_entry_stride;
+    for (int eid = 0; eid < block_size; ++eid) {
+      copy_entry(block_start_ptr + eid * cache_entry_stride,
+                 block_dst_ptr + eid * dst_entry_stride);
+    }
+  }
+
+  if (partial_block_size) {
+    auto block_id = batch_block_table[full_blocks_end];
+    auto block_start_ptr = src_cache + block_id * cache_block_stride;
+    auto block_dst_ptr = dst + full_blocks_end * block_size * dst_entry_stride;
+    for (int eid = 0; eid < partial_block_size; ++eid) {
+      copy_entry(block_start_ptr + eid * cache_entry_stride,
+                 block_dst_ptr + eid * dst_entry_stride);
     }
   }
 }
@@ -989,38 +992,34 @@ __global__ void gather_and_maybe_dequant_cache(
 // SCALAR_T is the data type of the destination tensor.
 // CACHE_T is the stored data type of kv-cache.
 // KV_DTYPE is the real data type of kv-cache.
-#define CALL_GATHER_CACHE(SCALAR_T, CACHE_T, KV_DTYPE)                        \
-  vllm::gather_and_maybe_dequant_cache<SCALAR_T, CACHE_T, KV_DTYPE, 576,      \
-                                       thread_block_size>                     \
-      <<<grid, block, 0, stream>>>(                                           \
-          reinterpret_cast<CACHE_T*>(src_cache.data_ptr()),                   \
-          reinterpret_cast<SCALAR_T*>(dst.data_ptr()),                        \
-          block_table.data_ptr<int32_t>(), cu_seq_lens.data_ptr<int32_t>(),   \
-          token_to_seq.data_ptr<int32_t>(), num_tokens, block_size,           \
-          block_table_stride, cache_block_stride, cache_entry_stride,         \
-          dst_entry_stride, reinterpret_cast<const float*>(scale.data_ptr()), \
-          seq_starts_ptr);
+#define CALL_GATHER_CACHE(SCALAR_T, CACHE_T, KV_DTYPE)                      \
+  vllm::gather_and_maybe_dequant_cache<SCALAR_T, CACHE_T, KV_DTYPE>         \
+      <<<grid, block, 0, stream>>>(                                         \
+          reinterpret_cast<CACHE_T*>(src_cache.data_ptr()),                 \
+          reinterpret_cast<SCALAR_T*>(dst.data_ptr()),                      \
+          block_table.data_ptr<int32_t>(), cu_seq_lens.data_ptr<int32_t>(), \
+          block_size, entry_size, block_table_stride, cache_block_stride,   \
+          cache_entry_stride, dst_entry_stride,                             \
+          reinterpret_cast<const float*>(scale.data_ptr()), seq_starts_ptr);
 
 // Gather sequences from the cache into the destination tensor.
 //  - cu_seq_lens contains the cumulative sequence lengths for each batch
 //  - block_table contains the cache block indices for each sequence
-//  - token_to_seq contains the back mapping from token_id to batch_id
 //  - Optionally, seq_starts (if provided) offsets the starting block index by
 //  (seq_starts[bid] / page_size)
 void gather_and_maybe_dequant_cache(
-    torch::Tensor const& src_cache,     // [NUM_BLOCKS, BLOCK_SIZE, ENTRIES...]
-    torch::Tensor const& dst,           // [TOT_TOKENS, ENTRIES...]
-    torch::Tensor const& block_table,   // [BATCH, BLOCK_INDICES]
-    torch::Tensor const& cu_seq_lens,   // [BATCH+1]
-    torch::Tensor const& token_to_seq,  // [MAX_TOKEN_ACROSS_CHUNKS]
-    int64_t num_tokens, const std::string& kv_cache_dtype,
+    torch::Tensor const& src_cache,    // [NUM_BLOCKS, BLOCK_SIZE, ENTRIES...]
+    torch::Tensor const& dst,          // [TOT_TOKENS, ENTRIES...]
+    torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
+    torch::Tensor const& cu_seq_lens,  // [BATCH+1]
+    int64_t batch_size, const std::string& kv_cache_dtype,
     torch::Tensor const& scale,
     std::optional<torch::Tensor> seq_starts = std::nullopt) {
   at::cuda::OptionalCUDAGuard device_guard(src_cache.device());
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
   int32_t block_size = src_cache.size(1);
-  int32_t head_dim = dst.size(-1);
+  int32_t entry_size = src_cache.flatten(2, -1).size(2);
 
   TORCH_CHECK(block_table.dtype() == torch::kInt32,
               "block_table must be int32");
@@ -1030,9 +1029,6 @@ void gather_and_maybe_dequant_cache(
     TORCH_CHECK(seq_starts.value().dtype() == torch::kInt32,
                 "seq_starts must be int32");
   }
-  TORCH_CHECK(head_dim == 576,
-              "gather_and_maybe_dequant_cache only support the head_dim to 576 "
-              "for better performance")
 
   TORCH_CHECK(src_cache.device() == dst.device(),
               "src_cache and dst must be on the same device");
@@ -1050,9 +1046,10 @@ void gather_and_maybe_dequant_cache(
   int64_t cache_entry_stride = src_cache.stride(1);
   int64_t dst_entry_stride = dst.stride(0);
 
-  constexpr int32_t thread_block_size = 64;
-  dim3 grid(num_tokens);
-  dim3 block(thread_block_size);
+  // Decide on the number of splits based on the batch size.
+  int num_splits = batch_size > 128 ? 2 : batch_size > 64 ? 4 : 16;
+  dim3 grid(batch_size, num_splits);
+  dim3 block(1024);
 
   const int32_t* seq_starts_ptr =
       seq_starts.has_value() ? seq_starts.value().data_ptr<int32_t>() : nullptr;

csrc/cpu/attention.cpp

@@ -0,0 +1,798 @@
+#include "cpu_types.hpp"
+
+namespace {
+
+template <typename scalar_t>
+struct KernelVecType {
+  using q_load_vec_type = void;
+  using q_vec_type = void;
+  using k_load_vec_type = void;
+  using k_vec_type = void;
+  using qk_acc_vec_type = void;
+  using v_load_vec_type = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using q_load_vec_type = vec_op::FP32Vec4;
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_load_vec_type = vec_op::FP32Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+};
+
+template <>
+struct KernelVecType<c10::Half> {
+#if defined(__powerpc64__) || defined(__s390x__)
+  // Power and s390x architecture-specific vector types
+  using q_load_vec_type = vec_op::FP32Vec8;
+  using k_load_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+#else
+  // Fallback for other architectures, including x86
+  using q_load_vec_type = vec_op::FP16Vec8;
+  using k_load_vec_type = vec_op::FP16Vec16;
+  using v_load_vec_type = vec_op::FP16Vec16;
+#endif
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+};
+
+#ifdef __AVX512BF16__
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using q_load_vec_type = vec_op::BF16Vec8;
+  using q_vec_type = vec_op::BF16Vec32;
+  using k_load_vec_type = vec_op::BF16Vec32;
+  using k_vec_type = vec_op::BF16Vec32;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+#else
+  #ifdef __aarch64__
+    #ifndef ARM_BF16_SUPPORT
+    // pass
+    #else
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using q_load_vec_type = vec_op::BF16Vec8;
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_load_vec_type = vec_op::BF16Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+    #endif
+  #else
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using q_load_vec_type = vec_op::BF16Vec8;
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_load_vec_type = vec_op::BF16Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+  #endif
+#endif
+
+template <typename T>
+FORCE_INLINE std::pair<T, T> reduceSoftmax(T* data, const int size,
+                                           const int capacity) {
+  T max = data[0];
+  for (int i = 1; i < size; ++i) {
+    max = max >= data[i] ? max : data[i];
+  }
+
+  T sum = 0;
+  for (int i = 0; i < size; ++i) {
+    data[i] = std::exp(data[i] - max);
+    sum += data[i];
+  }
+
+  int i = 0;
+  for (; i < size; ++i) {
+    data[i] /= sum;
+  }
+
+  for (; i < capacity; ++i) {
+    data[i] = 0;
+  }
+
+  return {max, sum};
+}
+
+template <typename T>
+FORCE_INLINE std::pair<T, T> reduceSoftmaxAlibi(T* data, const int size,
+                                                const int capacity,
+                                                const float alibi_slope,
+                                                const int start_index,
+                                                const int seq_len) {
+  data[0] += alibi_slope * (start_index - seq_len + 1);
+  T max = data[0];
+  for (int i = 1; i < size; ++i) {
+    T qk = data[i] + alibi_slope * (start_index + i - seq_len + 1);
+    data[i] = qk;
+    max = max >= qk ? max : qk;
+  }
+
+  T sum = 0;
+  for (int i = 0; i < size; ++i) {
+    data[i] = std::exp(data[i] - max);
+    sum += data[i];
+  }
+
+  int i = 0;
+  for (; i < size; ++i) {
+    data[i] /= sum;
+  }
+
+  for (; i < capacity; ++i) {
+    data[i] = 0;
+  }
+
+  return {max, sum};
+}
+
+template <typename T>
+FORCE_INLINE void reducePartitionSoftmax(const T* max_data, T* sum_data,
+                                         const int size) {
+  T max = max_data[0];
+  for (int i = 1; i < size; ++i) {
+    max = max >= max_data[i] ? max : max_data[i];
+  }
+
+  T rescaled_sum = 0;
+  for (int i = 0; i < size; ++i) {
+    T rescale_factor = std::exp(max_data[i] - max);
+    rescaled_sum += rescale_factor * sum_data[i];
+    sum_data[i] *= rescale_factor;
+  }
+  for (int i = 0; i < size; ++i) {
+    sum_data[i] /= rescaled_sum + 1e-8;
+  }
+}
+
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE, int x>
+struct reduceQKBlockKernel {
+  using q_load_vec_type = typename KernelVecType<scalar_t>::q_load_vec_type;
+  using q_vec_type = typename KernelVecType<scalar_t>::q_vec_type;
+  using k_load_vec_type = typename KernelVecType<scalar_t>::k_load_vec_type;
+  using k_vec_type = typename KernelVecType<scalar_t>::k_vec_type;
+  using qk_acc_vec_type = typename KernelVecType<scalar_t>::qk_acc_vec_type;
+
+  constexpr static int TOKEN_PER_GROUP = k_load_vec_type::get_elem_num() / x;
+  constexpr static int MAX_GROUP_NUM = 16 / TOKEN_PER_GROUP;
+  constexpr static int UNROLL_GROUP_NUM = MAX_GROUP_NUM / 4;
+
+  static_assert(MAX_GROUP_NUM == 8 || MAX_GROUP_NUM == 4);
+  static_assert(k_load_vec_type::get_elem_num() % x == 0);
+  static_assert(q_load_vec_type::get_elem_num() * sizeof(scalar_t) == 16);
+
+  FORCE_INLINE static void call(const scalar_t* __restrict__ q,
+                                const scalar_t* __restrict__ k_block,
+                                float* __restrict__ logits, float scale,
+                                const int token_num) {
+    const int group_num = (token_num + TOKEN_PER_GROUP - 1) / TOKEN_PER_GROUP;
+
+    qk_acc_vec_type group_accums[MAX_GROUP_NUM];
+    if (token_num == BLOCK_SIZE) {
+      for (int q_offset = 0; q_offset < HEAD_SIZE;
+           q_offset += x, k_block += x * BLOCK_SIZE) {
+        q_load_vec_type q_load_group_vec(q + q_offset);
+        q_vec_type q_group_vec(q_load_group_vec);
+
+        vec_op::unroll_loop<int, MAX_GROUP_NUM>(
+            [k_block, &q_group_vec, &group_accums](int token_group_idx) {
+              k_load_vec_type k_load_group_vec(k_block + token_group_idx * x *
+                                                             TOKEN_PER_GROUP);
+              k_vec_type k_group_vec(k_load_group_vec);
+              vec_op::fma(group_accums[token_group_idx], q_group_vec,
+                          k_group_vec);
+              vec_op::prefetch(k_block + x * BLOCK_SIZE +
+                               token_group_idx * x * TOKEN_PER_GROUP);
+            });
+      }
+    } else {
+      for (int q_offset = 0; q_offset < HEAD_SIZE;
+           q_offset += x, k_block += x * BLOCK_SIZE) {
+        q_load_vec_type q_load_group_vec(q + q_offset);
+        q_vec_type q_group_vec(q_load_group_vec);
+        for (int token_group_start = 0; token_group_start < group_num;
+             token_group_start += UNROLL_GROUP_NUM) {
+          vec_op::unroll_loop<int, UNROLL_GROUP_NUM>(
+              [token_group_start, k_block, &q_group_vec,
+               &group_accums](int token_group_idx) {
+                token_group_idx += token_group_start;
+                k_load_vec_type k_load_group_vec(k_block + token_group_idx * x *
+                                                               TOKEN_PER_GROUP);
+                k_vec_type k_group_vec(k_load_group_vec);
+                vec_op::fma(group_accums[token_group_idx], q_group_vec,
+                            k_group_vec);
+                vec_op::prefetch(k_block + x * BLOCK_SIZE +
+                                 token_group_idx * x * TOKEN_PER_GROUP);
+              });
+        }
+      }
+    }
+
+    for (int token_group_idx = 0; token_group_idx < group_num;
+         ++token_group_idx) {
+      vec_op::unroll_loop<int, TOKEN_PER_GROUP>(
+          [&group_accums, logits, scale, token_group_idx](int token_idx) {
+            float dot_v =
+                group_accums[token_group_idx]
+                    .template reduce_sub_sum<qk_acc_vec_type::get_elem_num() /
+                                             TOKEN_PER_GROUP>(token_idx);
+            logits[token_group_idx * TOKEN_PER_GROUP + token_idx] =
+                dot_v * scale;
+          });
+    }
+  }
+};
+
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE,
+          int HEAD_PARTITION_SIZE, typename acc_t>
+FORCE_INLINE void reduceValueBlock(const float* prob, const scalar_t* v_block,
+                                   acc_t&& acc) {
+  using v_load_vec_type = typename KernelVecType<scalar_t>::v_load_vec_type;
+  constexpr int ELEM_NUM = v_load_vec_type::get_elem_num();
+  static_assert(BLOCK_SIZE == ELEM_NUM);
+  vec_op::FP32Vec16 prob_vec(prob);
+
+  vec_op::unroll_loop<int, HEAD_PARTITION_SIZE>([&](int head_elem_idx) {
+    v_load_vec_type v_vec(v_block + BLOCK_SIZE * head_elem_idx);
+    vec_op::FP32Vec16 fp32_v_vec(v_vec);
+    acc[head_elem_idx] = acc[head_elem_idx] + prob_vec * fp32_v_vec;
+  });
+}
+};  // namespace
+
+// Paged attention v1
+namespace {
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE>
+struct paged_attention_v1_impl {
+  static void call(
+      scalar_t* __restrict__ out,            // [num_seqs, num_heads, head_size]
+      const scalar_t* __restrict__ q,        // [num_seqs, num_heads, head_size]
+      const scalar_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size/x, block_size, x]
+      const scalar_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size, block_size]
+      const int num_kv_heads, const float scale,
+      const int* __restrict__ block_tables,  // [num_seqs,
+                                             // max_num_blocks_per_seq]
+      const int* __restrict__ seq_lens,      // [num_seqs]
+      const int max_num_blocks_per_seq,
+      const float* __restrict__ alibi_slopes,  // [num_heads]
+      const int q_stride, const int kv_block_stride, const int kv_head_stride,
+      const int num_seqs, const int num_heads) {
+    constexpr int x = 16 / sizeof(scalar_t);
+    const int num_queries_per_kv = num_heads / num_kv_heads;
+
+    static_assert(BLOCK_SIZE == 16);
+
+    int max_seq_len = max_num_blocks_per_seq * BLOCK_SIZE;
+    int max_seq_len_padded = (max_seq_len + 15) & 0xFFFFFFF0;
+    TORCH_CHECK((max_seq_len_padded * sizeof(float)) % 64 == 0);
+
+    const int parallel_work_item_num = omp_get_max_threads();
+
+    size_t logits_bytes =
+        parallel_work_item_num * max_seq_len_padded * sizeof(float);
+    float* logits = (float*)std::aligned_alloc(
+        64, logits_bytes);  // Cacheline alignment for each context token.
+                            // [parallel_work_item_num, max_seq_len_padded]
+
+#pragma omp parallel for collapse(2) schedule(dynamic, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        int seq_len = seq_lens[seq_idx];
+        const int* seq_block_table =
+            block_tables + max_num_blocks_per_seq * seq_idx;
+        const int block_num = (seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+        const int64_t kv_head_idx = head_idx / num_queries_per_kv;
+        const scalar_t* __restrict__ q_vec_ptr =
+            q + seq_idx * q_stride + head_idx * HEAD_SIZE;
+        const int last_block_token_num = seq_len - (block_num - 1) * BLOCK_SIZE;
+        float* __restrict__ thread_block_logits =
+            logits + omp_get_thread_num() * max_seq_len_padded;
+
+        // Compute logits
+        for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+          const int64_t physical_block_idx = seq_block_table[block_idx];
+          const scalar_t* __restrict__ k_block_cache_ptr =
+              k_cache + physical_block_idx * kv_block_stride +
+              kv_head_idx * kv_head_stride;
+          float* __restrict__ head_block_logits =
+              thread_block_logits + block_idx * BLOCK_SIZE;
+
+          reduceQKBlockKernel<scalar_t, HEAD_SIZE, BLOCK_SIZE, x>::call(
+              q_vec_ptr, k_block_cache_ptr, head_block_logits, scale,
+              block_idx == block_num - 1 ? last_block_token_num : BLOCK_SIZE);
+        }
+
+        // Compute softmax
+        if (alibi_slopes) {
+          reduceSoftmaxAlibi(thread_block_logits, seq_len,
+                             block_num * BLOCK_SIZE, alibi_slopes[head_idx], 0,
+                             seq_len);
+        } else {
+          reduceSoftmax(thread_block_logits, seq_len, block_num * BLOCK_SIZE);
+        }
+
+        // Compute value
+        constexpr int head_elem_num_per_partition = 16;
+        constexpr int head_partition_num =
+            HEAD_SIZE / head_elem_num_per_partition;
+        for (int head_part_idx = 0; head_part_idx < head_partition_num;
+             ++head_part_idx) {
+          vec_op::FP32Vec16 accums[head_elem_num_per_partition];
+          scalar_t* __restrict__ out_ptr =
+              out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE +
+              head_part_idx * head_elem_num_per_partition;
+          for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+            const int64_t physical_block_idx = seq_block_table[block_idx];
+            const float* __restrict__ prob_vec_ptr =
+                thread_block_logits + block_idx * BLOCK_SIZE;
+            const scalar_t* __restrict__ v_block_cache_ptr =
+                v_cache + physical_block_idx * kv_block_stride +
+                kv_head_idx * kv_head_stride +
+                BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+            reduceValueBlock<scalar_t, HEAD_SIZE, BLOCK_SIZE,
+                             head_elem_num_per_partition>(
+                prob_vec_ptr, v_block_cache_ptr, accums);
+
+            if (block_idx != block_num - 1) {
+              const int64_t next_physical_block_idx =
+                  seq_block_table[block_idx + 1];
+              const scalar_t* __restrict__ next_v_block_cache_ptr =
+                  v_cache + next_physical_block_idx * kv_block_stride +
+                  kv_head_idx * kv_head_stride +
+                  BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+              vec_op::unroll_loop<int, head_elem_num_per_partition>(
+                  [&](int head_elem_idx) {
+                    if (head_elem_idx % 2 == 0) {
+                      vec_op::prefetch(next_v_block_cache_ptr +
+                                       BLOCK_SIZE * head_elem_idx);
+                    }
+                  });
+            }
+          }
+
+          vec_op::unroll_loop<int, head_elem_num_per_partition>(
+              [&](int head_elem_idx) {
+                float value = accums[head_elem_idx].reduce_sum();
+                vec_op::storeFP32(value, out_ptr + head_elem_idx);
+              });
+        }
+      }
+    }
+    std::free(logits);
+  }
+};
+
+#define LAUNCH_V1_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE)                   \
+  paged_attention_v1_impl<T, HEAD_SIZE, BLOCK_SIZE>::call(                     \
+      out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale, \
+      block_tables_ptr, seq_lens_ptr, max_num_blocks_per_seq,                  \
+      alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, num_seqs,   \
+      num_heads);
+
+template <typename T, int BLOCK_SIZE>
+void paged_attention_v1_impl_launcher(
+    torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
+    const std::optional<torch::Tensor>& alibi_slopes) {
+  int num_seqs = query.size(0);
+  int num_heads = query.size(1);
+  int head_size = query.size(2);
+  int max_num_blocks_per_seq = block_tables.size(1);
+  int q_stride = query.stride(0);
+  int kv_block_stride = key_cache.stride(0);
+  int kv_head_stride = key_cache.stride(1);
+
+  // NOTE: alibi_slopes is optional.
+  const float* alibi_slopes_ptr =
+      alibi_slopes
+          ? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
+          : nullptr;
+
+  T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
+  T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
+  T* key_cache_ptr = reinterpret_cast<T*>(key_cache.data_ptr());
+  T* value_cache_ptr = reinterpret_cast<T*>(value_cache.data_ptr());
+  int* block_tables_ptr = block_tables.data_ptr<int>();
+  int* seq_lens_ptr = seq_lens.data_ptr<int>();
+
+  switch (head_size) {
+    case 32:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 32, BLOCK_SIZE);
+      break;
+    case 64:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 64, BLOCK_SIZE);
+      break;
+    case 80:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 80, BLOCK_SIZE);
+      break;
+    case 96:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 96, BLOCK_SIZE);
+      break;
+    case 112:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 112, BLOCK_SIZE);
+      break;
+    case 128:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 128, BLOCK_SIZE);
+      break;
+    case 192:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 192, BLOCK_SIZE);
+      break;
+    case 256:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 256, BLOCK_SIZE);
+      break;
+    default:
+      TORCH_CHECK(false, "Unsupported head size: ", head_size);
+      break;
+  }
+}
+
+#define CALL_V1_KERNEL_LAUNCHER(T, BLOCK_SIZE)                               \
+  paged_attention_v1_impl_launcher<T, BLOCK_SIZE>(                           \
+      out, query, key_cache, value_cache, num_kv_heads, scale, block_tables, \
+      seq_lens, max_seq_len, alibi_slopes);
+
+#define CALL_V1_KERNEL_LAUNCHER_BLOCK_SIZE(T)                     \
+  switch (block_size) {                                           \
+    case 16:                                                      \
+      CALL_V1_KERNEL_LAUNCHER(T, 16);                             \
+      break;                                                      \
+    default:                                                      \
+      TORCH_CHECK(false, "Unsupported block size: ", block_size); \
+      break;                                                      \
+  }
+}  // namespace
+
+void paged_attention_v1(
+    torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
+    int64_t max_seq_len, const std::optional<torch::Tensor>& alibi_slopes,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
+    const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
+    const int64_t blocksparse_head_sliding_step) {
+  TORCH_CHECK(blocksparse_vert_stride <= 1,
+              "CPU backend does not support blocksparse attention yet.");
+  VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v1_impl",
+                               [&] {
+                                 CPU_KERNEL_GUARD_IN(paged_attention_v1_impl)
+                                 CALL_V1_KERNEL_LAUNCHER_BLOCK_SIZE(scalar_t);
+                                 CPU_KERNEL_GUARD_OUT(paged_attention_v1_impl)
+                               });
+}
+
+// Paged attention v2
+namespace {
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE, int PARTITION_SIZE>
+struct paged_attention_v2_impl {
+  static void call(
+      scalar_t* __restrict__ out,            // [num_seqs, num_heads, head_size]
+      float* __restrict__ exp_sums,          // [num_seqs, num_heads,
+                                             // max_num_partitions]
+      float* __restrict__ max_logits,        // [num_seqs, num_heads,
+                                             // max_num_partitions]
+      scalar_t* __restrict__ tmp_out,        // [num_seqs, num_heads,
+                                             // max_num_partitions, head_size]
+      const scalar_t* __restrict__ q,        // [num_seqs, num_heads, head_size]
+      const scalar_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size/x, block_size, x]
+      const scalar_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size, block_size]
+      const int num_kv_heads, const float scale,
+      const int* __restrict__ block_tables,  // [num_seqs,
+                                             // max_num_blocks_per_seq]
+      const int* __restrict__ seq_lens,      // [num_seqs]
+      const int max_num_blocks_per_seq,
+      const float* __restrict__ alibi_slopes,  // [num_heads]
+      const int q_stride, const int kv_block_stride, const int kv_head_stride,
+      const int num_seqs, const int num_heads, const int max_num_partitions) {
+    constexpr int x = 16 / sizeof(scalar_t);
+    const int num_queries_per_kv = num_heads / num_kv_heads;
+
+    static_assert(BLOCK_SIZE == 16);
+    static_assert(PARTITION_SIZE * sizeof(float) % 64 == 0);
+    static_assert(PARTITION_SIZE % BLOCK_SIZE == 0);
+
+#pragma omp parallel for collapse(3) schedule(static, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int partition_idx = 0; partition_idx < max_num_partitions;
+           ++partition_idx) {
+        for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+          const int seq_len = seq_lens[seq_idx];
+          const int start_token_idx = partition_idx * PARTITION_SIZE;
+
+          if (start_token_idx >= seq_len) continue;
+
+          const int partition_num =
+              (seq_len + PARTITION_SIZE - 1) / PARTITION_SIZE;
+          const bool no_reduce = (partition_num == 1);
+          const int token_num =
+              (std::min(seq_len, start_token_idx + PARTITION_SIZE) -
+               start_token_idx);
+          const int block_num = (token_num + BLOCK_SIZE - 1) / BLOCK_SIZE;
+          const int last_block_token_num =
+              token_num - (block_num - 1) * BLOCK_SIZE;
+          const int* seq_block_table = block_tables +
+                                       max_num_blocks_per_seq * seq_idx +
+                                       start_token_idx / BLOCK_SIZE;
+          const int64_t kv_head_idx = head_idx / num_queries_per_kv;
+          const scalar_t* __restrict__ q_vec_ptr =
+              q + seq_idx * q_stride + head_idx * HEAD_SIZE;
+
+          float logits[PARTITION_SIZE] __attribute__((aligned(64))) = {0};
+
+          // Compute logits
+          for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+            const int64_t physical_block_idx = seq_block_table[block_idx];
+            const scalar_t* __restrict__ k_block_cache_ptr =
+                k_cache + physical_block_idx * kv_block_stride +
+                kv_head_idx * kv_head_stride;
+            float* __restrict__ head_block_logits =
+                logits + block_idx * BLOCK_SIZE;
+
+            reduceQKBlockKernel<scalar_t, HEAD_SIZE, BLOCK_SIZE, x>::call(
+                q_vec_ptr, k_block_cache_ptr, head_block_logits, scale,
+                block_idx == block_num - 1 ? last_block_token_num : BLOCK_SIZE);
+          }
+
+          std::pair<float, float> max_and_sum;
+          if (alibi_slopes) {
+            max_and_sum = reduceSoftmaxAlibi(
+                logits, token_num, block_num * BLOCK_SIZE,
+                alibi_slopes[head_idx], start_token_idx, seq_len);
+          } else {
+            max_and_sum =
+                reduceSoftmax(logits, token_num, block_num * BLOCK_SIZE);
+          }
+
+          auto&& [max_logit, exp_sum] = max_and_sum;
+
+          scalar_t* __restrict__ output_buffer = nullptr;
+          if (!no_reduce) {
+            auto idx = seq_idx * num_heads * max_num_partitions +
+                       head_idx * max_num_partitions + partition_idx;
+            max_logits[idx] = max_logit;
+            exp_sums[idx] = exp_sum;
+            output_buffer =
+                tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
+                head_idx * max_num_partitions * HEAD_SIZE +
+                partition_idx * HEAD_SIZE;
+          } else {
+            output_buffer =
+                out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
+          }
+
+          // Compute value
+          constexpr int head_elem_num_per_partition = 16;
+          constexpr int head_partition_num =
+              HEAD_SIZE / head_elem_num_per_partition;
+          for (int head_part_idx = 0; head_part_idx < head_partition_num;
+               ++head_part_idx) {
+            vec_op::FP32Vec16 accums[head_elem_num_per_partition];
+            scalar_t* __restrict__ out_ptr =
+                output_buffer + head_part_idx * head_elem_num_per_partition;
+            for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+              const int64_t physical_block_idx = seq_block_table[block_idx];
+              const float* __restrict__ prob_vec_ptr =
+                  logits + block_idx * BLOCK_SIZE;
+              const scalar_t* __restrict__ v_block_cache_ptr =
+                  v_cache + physical_block_idx * kv_block_stride +
+                  kv_head_idx * kv_head_stride +
+                  BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+              reduceValueBlock<scalar_t, HEAD_SIZE, BLOCK_SIZE,
+                               head_elem_num_per_partition>(
+                  prob_vec_ptr, v_block_cache_ptr, accums);
+
+              if (block_idx != block_num - 1) {
+                const int64_t next_physical_block_idx =
+                    seq_block_table[block_idx + 1];
+                const scalar_t* __restrict__ next_v_block_cache_ptr =
+                    v_cache + next_physical_block_idx * kv_block_stride +
+                    kv_head_idx * kv_head_stride +
+                    BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+                vec_op::unroll_loop<int, head_elem_num_per_partition>(
+                    [&](int head_elem_idx) {
+                      if (head_elem_idx % 2 == 0) {
+                        vec_op::prefetch(next_v_block_cache_ptr +
+                                         BLOCK_SIZE * head_elem_idx);
+                      }
+                    });
+              }
+            }
+
+            vec_op::unroll_loop<int, head_elem_num_per_partition>(
+                [&](int head_elem_idx) {
+                  float value = accums[head_elem_idx].reduce_sum();
+                  vec_op::storeFP32(value, out_ptr + head_elem_idx);
+                });
+          }
+        }
+      }
+    }
+
+    // Rescale partition softmax and store the factors to exp_sums
+#pragma omp parallel for collapse(2) schedule(static, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        const int seq_len = seq_lens[seq_idx];
+        const int partition_num =
+            (seq_len + PARTITION_SIZE - 1) / PARTITION_SIZE;
+
+        if (partition_num == 1) continue;
+
+        reducePartitionSoftmax(
+            max_logits + seq_idx * num_heads * max_num_partitions +
+                head_idx * max_num_partitions,
+            exp_sums + seq_idx * num_heads * max_num_partitions +
+                head_idx * max_num_partitions,
+            partition_num);
+      }
+    }
+
+    // Reduce values
+    using v_load_vec_type = typename KernelVecType<scalar_t>::v_load_vec_type;
+    static_assert(v_load_vec_type::get_elem_num() == BLOCK_SIZE);
+    constexpr int head_elem_num_per_group =
+        16;  // Note: didn't align with the cacheline size, due to some
+             // HEAD_SIZE didn't align with 64 bytes
+    static_assert(HEAD_SIZE % head_elem_num_per_group == 0);
+    constexpr int head_group_num = HEAD_SIZE / head_elem_num_per_group;
+    const float* __restrict__ rescale_factors = exp_sums;
+#pragma omp parallel for collapse(3) schedule(static, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        for (int group_idx = 0; group_idx < head_group_num; ++group_idx) {
+          const int seq_len = seq_lens[seq_idx];
+          const int partition_num =
+              (seq_len + PARTITION_SIZE - 1) / PARTITION_SIZE;
+
+          if (partition_num == 1) continue;
+
+          const float* __restrict__ seq_head_rescale_factors =
+              rescale_factors + seq_idx * num_heads * max_num_partitions +
+              head_idx * max_num_partitions;
+          const scalar_t* __restrict__ seq_head_tmp_out =
+              tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
+              head_idx * max_num_partitions * HEAD_SIZE +
+              group_idx * head_elem_num_per_group;
+          scalar_t* __restrict__ seq_head_output =
+              out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE +
+              group_idx * head_elem_num_per_group;
+
+          vec_op::FP32Vec16 acc;
+          for (int i = 0; i < partition_num; ++i) {
+            vec_op::FP32Vec16 rescale_factor(seq_head_rescale_factors[i]);
+            v_load_vec_type value(seq_head_tmp_out + i * HEAD_SIZE);
+            vec_op::FP32Vec16 fp32_value(value);
+            acc = acc + fp32_value * rescale_factor;
+          }
+          v_load_vec_type cast_acc(acc);
+          cast_acc.save(seq_head_output);
+        }
+      }
+    }
+  }
+};
+
+#define LAUNCH_V2_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE)                 \
+  paged_attention_v2_impl<T, HEAD_SIZE, BLOCK_SIZE, PARTITION_SIZE>::call(   \
+      out_ptr, exp_sums_ptr, max_logits_ptr, tmp_out_ptr, query_ptr,         \
+      key_cache_ptr, value_cache_ptr, num_kv_heads, scale, block_tables_ptr, \
+      seq_lens_ptr, max_num_blocks_per_seq, alibi_slopes_ptr, q_stride,      \
+      kv_block_stride, kv_head_stride, num_seqs, num_heads,                  \
+      max_num_partitions);
+
+template <typename T, int BLOCK_SIZE, int PARTITION_SIZE = 512>
+void paged_attention_v2_impl_launcher(
+    torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
+    torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int block_size,
+    int max_seq_len, const std::optional<torch::Tensor>& alibi_slopes) {
+  int num_seqs = query.size(0);
+  int num_heads = query.size(1);
+  int head_size = query.size(2);
+  int max_num_blocks_per_seq = block_tables.size(1);
+  int q_stride = query.stride(0);
+  int kv_block_stride = key_cache.stride(0);
+  int kv_head_stride = key_cache.stride(1);
+  int max_num_partitions = exp_sums.size(-1);
+
+  // NOTE: alibi_slopes is optional.
+  const float* alibi_slopes_ptr =
+      alibi_slopes
+          ? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
+          : nullptr;
+
+  T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
+  float* exp_sums_ptr = reinterpret_cast<float*>(exp_sums.data_ptr());
+  float* max_logits_ptr = reinterpret_cast<float*>(max_logits.data_ptr());
+  T* tmp_out_ptr = reinterpret_cast<T*>(tmp_out.data_ptr());
+  T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
+  T* key_cache_ptr = reinterpret_cast<T*>(key_cache.data_ptr());
+  T* value_cache_ptr = reinterpret_cast<T*>(value_cache.data_ptr());
+  int* block_tables_ptr = block_tables.data_ptr<int>();
+  int* seq_lens_ptr = seq_lens.data_ptr<int>();
+
+  switch (head_size) {
+    case 32:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 32, BLOCK_SIZE);
+      break;
+    case 64:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 64, BLOCK_SIZE);
+      break;
+    case 80:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 80, BLOCK_SIZE);
+      break;
+    case 96:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 96, BLOCK_SIZE);
+      break;
+    case 112:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 112, BLOCK_SIZE);
+      break;
+    case 128:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 128, BLOCK_SIZE);
+      break;
+    case 192:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 192, BLOCK_SIZE);
+      break;
+    case 256:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 256, BLOCK_SIZE);
+      break;
+    default:
+      TORCH_CHECK(false, "Unsupported head size: ", head_size);
+      break;
+  }
+}
+
+#define CALL_V2_KERNEL_LAUNCHER(T, BLOCK_SIZE)                              \
+  paged_attention_v2_impl_launcher<T, BLOCK_SIZE>(                          \
+      out, exp_sums, max_logits, tmp_out, query, key_cache, value_cache,    \
+      num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len, \
+      alibi_slopes);
+
+#define CALL_V2_KERNEL_LAUNCHER_BLOCK_SIZE(T)                     \
+  switch (block_size) {                                           \
+    case 16:                                                      \
+      CALL_V2_KERNEL_LAUNCHER(T, 16);                             \
+      break;                                                      \
+    default:                                                      \
+      TORCH_CHECK(false, "Unsupported block size: ", block_size); \
+      break;                                                      \
+  }
+}  // namespace
+
+void paged_attention_v2(
+    torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
+    torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int64_t num_kv_heads, double scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int64_t block_size,
+    int64_t max_seq_len, const std::optional<torch::Tensor>& alibi_slopes,
+    const std::string& kv_cache_dtype, torch::Tensor& k_scale,
+    torch::Tensor& v_scale, const int64_t tp_rank,
+    const int64_t blocksparse_local_blocks,
+    const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
+    const int64_t blocksparse_head_sliding_step) {
+  TORCH_CHECK(blocksparse_vert_stride <= 1,
+              "CPU backend does not support blocksparse attention yet.");
+  VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v2_impl",
+                               [&] {
+                                 CPU_KERNEL_GUARD_IN(paged_attention_v2_impl)
+                                 CALL_V2_KERNEL_LAUNCHER_BLOCK_SIZE(scalar_t);
+                                 CPU_KERNEL_GUARD_OUT(paged_attention_v2_impl)
+                               });
+}

csrc/cpu/cache.cpp

@@ -0,0 +1,214 @@
+#include <map>
+#include <vector>
+
+#include "cpu_types.hpp"
+
+#if defined(__x86_64__)
+  #define DISPATCH_MACRO VLLM_DISPATCH_FLOATING_TYPES_WITH_E5M2
+#else
+  #define DISPATCH_MACRO VLLM_DISPATCH_FLOATING_TYPES
+#endif
+
+namespace {
+template <typename scalar_t>
+void copy_blocks_cpu_impl(std::vector<torch::Tensor> const& key_caches,
+                          std::vector<torch::Tensor> const& value_caches,
+                          const torch::Tensor& mapping_pairs,
+                          const int element_num_per_block,
+                          const int layer_num) {
+  const size_t pair_num = mapping_pairs.size(0);
+  const size_t block_bytes = sizeof(scalar_t) * element_num_per_block;
+#pragma omp parallel for collapse(2)
+  for (int layer = 0; layer < layer_num; ++layer) {
+    for (size_t pair = 0; pair < pair_num; ++pair) {
+      int64_t source_offset =
+          element_num_per_block * mapping_pairs[pair][0].item<int64_t>();
+      int64_t target_offset =
+          element_num_per_block * mapping_pairs[pair][1].item<int64_t>();
+      scalar_t* key_cache_ptr = key_caches[layer].data_ptr<scalar_t>();
+      scalar_t* source_ptr = key_cache_ptr + source_offset;
+      scalar_t* target_ptr = key_cache_ptr + target_offset;
+      std::memcpy(target_ptr, source_ptr, block_bytes);
+
+      scalar_t* value_cache_ptr = value_caches[layer].data_ptr<scalar_t>();
+      source_ptr = value_cache_ptr + source_offset;
+      target_ptr = value_cache_ptr + target_offset;
+      std::memcpy(target_ptr, source_ptr, block_bytes);
+    }
+  }
+}
+
+template <typename scalar_t>
+void reshape_and_cache_cpu_impl(
+    const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
+    scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
+    const int64_t* __restrict__ slot_mapping, const int num_tokens,
+    const int key_stride, const int value_stride, const int num_heads,
+    const int head_size, const int block_size, const int x) {
+  const int block_elem_num = num_heads * head_size * block_size;
+
+#pragma omp parallel for collapse(2)
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+      const int64_t slot_idx = slot_mapping[token_idx];
+      if (slot_idx >= 0) {
+        int src_key_head_idx = token_idx * key_stride + head_idx * head_size;
+        int src_value_head_idx =
+            token_idx * value_stride + head_idx * head_size;
+        const scalar_t* src_key_head_ptr = key + src_key_head_idx;
+        const scalar_t* src_value_head_ptr = value + src_value_head_idx;
+        const int64_t block_index = slot_idx / block_size;
+        const int64_t block_offset = slot_idx % block_size;
+        scalar_t* target_key_head_ptr = key_cache +
+                                        block_elem_num * block_index +
+                                        head_idx * block_size * head_size;
+        scalar_t* target_value_head_ptr = value_cache +
+                                          block_elem_num * block_index +
+                                          head_idx * block_size * head_size;
+
+        for (int src_key_idx = 0; src_key_idx < head_size; src_key_idx += x) {
+          const int64_t target_offset =
+              src_key_idx * block_size + block_offset * x;
+          for (int i = 0; i < x; ++i) {
+            target_key_head_ptr[target_offset + i] =
+                src_key_head_ptr[src_key_idx + i];
+          }
+        }
+
+        for (int src_value_idx = 0; src_value_idx < head_size;
+             ++src_value_idx) {
+          const int64_t target_offset =
+              src_value_idx * block_size + block_offset;
+          target_value_head_ptr[target_offset] =
+              src_value_head_ptr[src_value_idx];
+        }
+      }
+    }
+  }
+}
+};  // namespace
+
+template <typename scalar_t>
+void concat_and_cache_mla_cpu_impl(
+    const scalar_t* __restrict__ kv_c,  // [num_tokens, kv_lora_rank]
+    const scalar_t* __restrict__ k_pe,  // [num_tokens, pe_dim]
+    scalar_t* __restrict__ kv_cache,  // [num_blocks, block_size, (kv_lora_rank
+                                      // + pe_dim)]
+    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
+    const int num_tokens,                      //
+    const int block_stride,                    //
+    const int entry_stride,                    //
+    const int kv_c_stride,                     //
+    const int k_pe_stride,                     //
+    const int kv_lora_rank,                    //
+    const int pe_dim,                          //
+    const int block_size                       //
+) {
+#pragma omp parallel for
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    const int64_t slot_idx = slot_mapping[token_idx];
+    // NOTE: slot_idx can be -1 if the token is padded
+    if (slot_idx < 0) {
+      continue;
+    }
+    const int64_t block_idx = slot_idx / block_size;
+    const int64_t block_offset = slot_idx % block_size;
+
+    auto copy = [&](const scalar_t* __restrict__ src,
+                    scalar_t* __restrict__ dst, int src_stride, int dst_stride,
+                    int size, int offset) {
+      for (int i = 0; i < size; i++) {
+        const int64_t src_idx = token_idx * src_stride + i;
+        const int64_t dst_idx =
+            block_idx * block_stride + block_offset * entry_stride + i + offset;
+        dst[dst_idx] = src[src_idx];
+      }
+    };
+
+    copy(kv_c, kv_cache, kv_c_stride, block_stride, kv_lora_rank, 0);
+    copy(k_pe, kv_cache, k_pe_stride, block_stride, pe_dim, kv_lora_rank);
+  }
+}
+
+// Note: the key_caches and value_caches vectors are constant but
+// not the Tensors they contain. The vectors need to be const refs
+// in order to satisfy pytorch's C++ operator registration code.
+void copy_blocks(std::vector<torch::Tensor> const& key_caches,
+                 std::vector<torch::Tensor> const& value_caches,
+                 const torch::Tensor& block_mapping) {
+  unsigned num_layers = key_caches.size();
+  TORCH_CHECK(num_layers == value_caches.size());
+  if (num_layers == 0) {
+    return;
+  }
+
+  const int element_num_per_block = key_caches[0][0].numel();
+  DISPATCH_MACRO(key_caches[0].scalar_type(), "copy_blocks_cpu_impl", [&] {
+    CPU_KERNEL_GUARD_IN(copy_blocks_cpu_impl)
+    copy_blocks_cpu_impl<scalar_t>(key_caches, value_caches, block_mapping,
+                                   element_num_per_block, num_layers);
+    CPU_KERNEL_GUARD_OUT(copy_blocks_cpu_impl)
+  });
+}
+
+void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
+                       torch::Tensor& key_cache, torch::Tensor& value_cache,
+                       torch::Tensor& slot_mapping,
+                       const std::string& kv_cache_dtype,
+                       torch::Tensor& k_scale, torch::Tensor& v_scale) {
+  int num_tokens = key.size(0);
+  int num_heads = key.size(1);
+  int head_size = key.size(2);
+  int block_size = key_cache.size(3);
+  int x = key_cache.size(4);
+
+  int key_stride = key.stride(0);
+  int value_stride = value.stride(0);
+
+  DISPATCH_MACRO(key.scalar_type(), "reshape_and_cache_cpu_impl", [&] {
+    CPU_KERNEL_GUARD_IN(reshape_and_cache_cpu_impl)
+    reshape_and_cache_cpu_impl<scalar_t>(
+        key.data_ptr<scalar_t>(), value.data_ptr<scalar_t>(),
+        key_cache.data_ptr<scalar_t>(), value_cache.data_ptr<scalar_t>(),
+        slot_mapping.data_ptr<int64_t>(), num_tokens, key_stride, value_stride,
+        num_heads, head_size, block_size, x);
+    CPU_KERNEL_GUARD_OUT(reshape_and_cache_cpu_impl)
+  });
+}
+
+void concat_and_cache_mla(
+    torch::Tensor& kv_c,          // [num_tokens, kv_lora_rank]
+    torch::Tensor& k_pe,          // [num_tokens, pe_dim]
+    torch::Tensor& kv_cache,      // [num_blocks, block_size, (kv_lora_rank +
+                                  // pe_dim)]
+    torch::Tensor& slot_mapping,  // [num_tokens] or [num_actual_tokens]
+    const std::string& kv_cache_dtype, torch::Tensor& scale) {
+  int num_tokens = slot_mapping.size(0);
+  int kv_lora_rank = kv_c.size(1);
+  int pe_dim = k_pe.size(1);
+  int block_size = kv_cache.size(1);
+
+  TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
+  TORCH_CHECK(kv_cache_dtype != "fp8");
+
+  int kv_c_stride = kv_c.stride(0);
+  int k_pe_stride = k_pe.stride(0);
+  int block_stride = kv_cache.stride(0);
+  int entry_stride = kv_cache.stride(1);
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      kv_c.scalar_type(), "concat_and_cache_mla_cpu_impl", [&] {
+        CPU_KERNEL_GUARD_IN(concat_and_cache_mla_cpu_impl)
+        concat_and_cache_mla_cpu_impl<scalar_t>(
+            kv_c.data_ptr<scalar_t>(), k_pe.data_ptr<scalar_t>(),
+            kv_cache.data_ptr<scalar_t>(), slot_mapping.data_ptr<int64_t>(),
+            num_tokens, block_stride, entry_stride, kv_c_stride, k_pe_stride,
+            kv_lora_rank, pe_dim, block_size);
+        CPU_KERNEL_GUARD_OUT(concat_and_cache_mla_cpu_impl)
+      });
+}
+
+void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
+                 const torch::Tensor& block_mapping) {
+  TORCH_CHECK(false, "swap_blocks is unsupported on CPU.")
+}

csrc/cpu/cpu_attn.cpp

@@ -1,266 +0,0 @@
-#include "cpu_attn_vec.hpp"
-#include "cpu_attn_vec16.hpp"
-
-#ifdef CPU_CAPABILITY_AMXBF16
-  #include "cpu_attn_amx.hpp"
-  #define AMX_DISPATCH(...)                                                   \
-    case cpu_attention::ISA::AMX: {                                           \
-      using attn_impl = cpu_attention::AttentionImpl<cpu_attention::ISA::AMX, \
-                                                     scalar_t, head_dim>;     \
-      return __VA_ARGS__();                                                   \
-    }
-#else
-  #define AMX_DISPATCH(...) case cpu_attention::ISA::AMX:
-#endif
-
-#ifdef __aarch64__
-  #include "cpu_attn_neon.hpp"
-  #define NEON_DISPATCH(...)                                                   \
-    case cpu_attention::ISA::NEON: {                                           \
-      using attn_impl = cpu_attention::AttentionImpl<cpu_attention::ISA::NEON, \
-                                                     scalar_t, head_dim>;      \
-      return __VA_ARGS__();                                                    \
-    }
-#else
-  #define NEON_DISPATCH(...) case cpu_attention::ISA::NEON:
-#endif  // #ifdef __aarch64__
-
-#define CPU_ATTN_DISPATCH_CASE(HEAD_DIM, ...) \
-  case HEAD_DIM: {                            \
-    constexpr size_t head_dim = HEAD_DIM;     \
-    return __VA_ARGS__();                     \
-  }
-
-#define CPU_ATTN_DISPATCH_CASE_HEADDIM(HEAD_DIM, ...)           \
-  [&] {                                                         \
-    switch (HEAD_DIM) {                                         \
-      CPU_ATTN_DISPATCH_CASE(32, __VA_ARGS__)                   \
-      CPU_ATTN_DISPATCH_CASE(64, __VA_ARGS__)                   \
-      CPU_ATTN_DISPATCH_CASE(96, __VA_ARGS__)                   \
-      CPU_ATTN_DISPATCH_CASE(128, __VA_ARGS__)                  \
-      CPU_ATTN_DISPATCH_CASE(160, __VA_ARGS__)                  \
-      CPU_ATTN_DISPATCH_CASE(192, __VA_ARGS__)                  \
-      CPU_ATTN_DISPATCH_CASE(224, __VA_ARGS__)                  \
-      CPU_ATTN_DISPATCH_CASE(256, __VA_ARGS__)                  \
-      default: {                                                \
-        TORCH_CHECK(false, "Invalid CPU attention head_dim: " + \
-                               std::to_string(HEAD_DIM));       \
-      }                                                         \
-    }                                                           \
-  }()
-
-#define CPU_ATTN_DISPATCH_IMPL(ISA_TYPE, ...)                                 \
-  [&] {                                                                       \
-    switch (ISA_TYPE) {                                                       \
-      AMX_DISPATCH(__VA_ARGS__)                                               \
-      NEON_DISPATCH(__VA_ARGS__)                                              \
-      case cpu_attention::ISA::VEC: {                                         \
-        using attn_impl =                                                     \
-            cpu_attention::AttentionImpl<cpu_attention::ISA::VEC, scalar_t,   \
-                                         head_dim>;                           \
-        return __VA_ARGS__();                                                 \
-      }                                                                       \
-      case cpu_attention::ISA::VEC16: {                                       \
-        using attn_impl =                                                     \
-            cpu_attention::AttentionImpl<cpu_attention::ISA::VEC16, scalar_t, \
-                                         head_dim>;                           \
-        return __VA_ARGS__();                                                 \
-      }                                                                       \
-      default: {                                                              \
-        TORCH_CHECK(false, "Invalid CPU attention ISA type.");                \
-      }                                                                       \
-    }                                                                         \
-  }()
-
-torch::Tensor get_scheduler_metadata(
-    const int64_t num_req, const int64_t num_heads_q,
-    const int64_t num_heads_kv, const int64_t head_dim,
-    const torch::Tensor& seq_lens, at::ScalarType dtype,
-    const torch::Tensor& query_start_loc, const bool casual,
-    const int64_t window_size, const std::string& isa_hint,
-    const bool enable_kv_split) {
-  cpu_attention::ISA isa;
-  if (isa_hint == "amx") {
-    isa = cpu_attention::ISA::AMX;
-  } else if (isa_hint == "vec") {
-    isa = cpu_attention::ISA::VEC;
-  } else if (isa_hint == "vec16") {
-    isa = cpu_attention::ISA::VEC16;
-  } else if (isa_hint == "neon") {
-    isa = cpu_attention::ISA::NEON;
-  } else {
-    TORCH_CHECK(false, "Unsupported CPU attention ISA hint: " + isa_hint);
-  }
-
-  cpu_attention::AttentionScheduler::ScheduleInput input;
-  input.num_reqs = num_req;
-  input.num_heads_q = num_heads_q;
-  input.num_heads_kv = num_heads_kv;
-  input.head_dim = head_dim;
-  input.query_start_loc = query_start_loc.data_ptr<int32_t>();
-  input.seq_lens = seq_lens.data_ptr<int32_t>();
-  if (window_size != -1) {
-    input.left_sliding_window_size = window_size - 1;
-    if (casual) {
-      input.right_sliding_window_size = 0;
-    } else {
-      input.right_sliding_window_size = window_size - 1;
-    }
-  } else {
-    input.left_sliding_window_size = -1;
-    if (casual) {
-      input.right_sliding_window_size = 0;
-    } else {
-      input.right_sliding_window_size = -1;
-    }
-  }
-  input.casual = casual;
-  input.isa = isa;
-  input.enable_kv_split = enable_kv_split;
-  TORCH_CHECK(casual, "Only supports casual mask for now.");
-
-  VLLM_DISPATCH_FLOATING_TYPES(dtype, "get_scheduler_metadata", [&]() {
-    CPU_ATTN_DISPATCH_CASE_HEADDIM(head_dim, [&] {
-      CPU_ATTN_DISPATCH_IMPL(isa, [&]() {
-        input.elem_size = sizeof(scalar_t);
-        input.q_buffer_elem_size = sizeof(attn_impl::q_buffer_t);
-        input.logits_buffer_elem_size = sizeof(attn_impl::logits_buffer_t);
-        input.output_buffer_elem_size =
-            sizeof(attn_impl::partial_output_buffer_t);
-        input.max_num_q_per_iter = attn_impl::MaxQHeadNumPerIteration;
-        input.kv_block_alignment = attn_impl::BlockSizeAlignment;
-      });
-    });
-  });
-
-  cpu_attention::AttentionScheduler scheduler;
-  torch::Tensor metadata = scheduler.schedule(input);
-  return metadata;
-}
-
-void cpu_attn_reshape_and_cache(
-    const torch::Tensor& key,    // [token_num, head_num, head_size]
-    const torch::Tensor& value,  // [token_num, head_num, head_size]
-    torch::Tensor&
-        key_cache,  // [num_blocks, num_kv_heads, block_size, head_size]
-    torch::Tensor&
-        value_cache,  // [num_blocks, num_kv_heads, block_size, head_size]
-    const torch::Tensor& slot_mapping, const std::string& isa) {
-  TORCH_CHECK_EQ(key.dim(), 3);
-  TORCH_CHECK_EQ(value.dim(), 3);
-  TORCH_CHECK_EQ(key_cache.dim(), 4);
-  TORCH_CHECK_EQ(value_cache.dim(), 4);
-  TORCH_CHECK_EQ(key.stride(2), 1);
-  TORCH_CHECK_EQ(value.stride(2), 1);
-
-  const int64_t token_num = key.size(0);
-  const int64_t key_token_num_stride = key.stride(0);
-  const int64_t value_token_num_stride = value.stride(0);
-  const int64_t head_num = value.size(1);
-  const int64_t key_head_num_stride = key.stride(1);
-  const int64_t value_head_num_stride = value.stride(1);
-  const int64_t num_blocks = key_cache.size(0);
-  const int64_t num_blocks_stride = key_cache.stride(0);
-  const int64_t cache_head_num_stride = key_cache.stride(1);
-  const int64_t block_size = key_cache.size(2);
-  const int64_t block_size_stride = key_cache.stride(2);
-  const int64_t head_dim = key.size(-1);
-
-  cpu_attention::ISA isa_tag = [&]() {
-    if (isa == "amx") {
-      return cpu_attention::ISA::AMX;
-    } else if (isa == "vec") {
-      return cpu_attention::ISA::VEC;
-    } else if (isa == "vec16") {
-      return cpu_attention::ISA::VEC16;
-    } else if (isa == "neon") {
-      return cpu_attention::ISA::NEON;
-    } else {
-      TORCH_CHECK(false, "Invalid ISA type: " + isa);
-    }
-  }();
-
-  VLLM_DISPATCH_FLOATING_TYPES(
-      key.scalar_type(), "cpu_attn_reshape_and_cache", [&]() {
-        CPU_ATTN_DISPATCH_CASE_HEADDIM(head_dim, [&] {
-          CPU_ATTN_DISPATCH_IMPL(isa_tag, [&]() {
-            attn_impl::reshape_and_cache(
-                key.data_ptr<scalar_t>(), value.data_ptr<scalar_t>(),
-                key_cache.data_ptr<scalar_t>(),
-                value_cache.data_ptr<scalar_t>(),
-                slot_mapping.data_ptr<int64_t>(), token_num,
-                key_token_num_stride, value_token_num_stride, head_num,
-                key_head_num_stride, value_head_num_stride, num_blocks,
-                num_blocks_stride, cache_head_num_stride, block_size,
-                block_size_stride);
-          });
-        });
-      });
-}
-
-void cpu_attention_with_kv_cache(
-    const torch::Tensor& query,  // [num_tokens, num_heads, head_size]
-    const torch::Tensor&
-        key_cache,  // [num_blocks, num_kv_heads, block_size, head_size]
-    const torch::Tensor&
-        value_cache,        // [num_blocks, num_kv_heads, block_size, head_size]
-    torch::Tensor& output,  // [num_tokens, num_heads, head_size]
-    const torch::Tensor& query_start_loc,  // [num_tokens + 1]
-    const torch::Tensor& seq_lens,         // [num_tokens]
-    const double scale, const bool causal,
-    const std::optional<torch::Tensor>& alibi_slopes,  // [num_heads]
-    const int64_t sliding_window_left, const int64_t sliding_window_right,
-    const torch::Tensor& block_table,  // [num_tokens, max_block_num]
-    const double softcap, const torch::Tensor& scheduler_metadata,
-    const std::optional<torch::Tensor>& s_aux  // [num_heads]
-) {
-  TORCH_CHECK_EQ(query.dim(), 3);
-  TORCH_CHECK_EQ(query.stride(2), 1);
-  TORCH_CHECK_EQ(key_cache.dim(), 4);
-  TORCH_CHECK_EQ(value_cache.dim(), 4);
-
-  cpu_attention::AttentionInput input;
-  input.metadata = reinterpret_cast<cpu_attention::AttentionMetadata*>(
-      scheduler_metadata.data_ptr());
-  input.num_tokens = query.size(0);
-  input.num_heads = query.size(1);
-  input.num_kv_heads = key_cache.size(1);
-  input.block_size = key_cache.size(2);
-  input.query = query.data_ptr();
-  input.query_num_tokens_stride = query.stride(0);
-  input.query_num_heads_stride = query.stride(1);
-  input.cache_num_blocks_stride = key_cache.stride(0);
-  input.cache_num_kv_heads_stride = key_cache.stride(1);
-  input.blt_num_tokens_stride = block_table.stride(0);
-  input.key_cache = key_cache.data_ptr();
-  input.value_cache = value_cache.data_ptr();
-  input.output = output.data_ptr();
-  input.query_start_loc = query_start_loc.data_ptr<int32_t>();
-  input.seq_lens = seq_lens.data_ptr<int32_t>();
-  input.block_table = block_table.data_ptr<int32_t>();
-  input.alibi_slopes =
-      alibi_slopes.has_value() ? alibi_slopes->data_ptr<float>() : nullptr;
-  // For now sink must be bf16
-  input.s_aux = s_aux.has_value() ? s_aux->data_ptr<c10::BFloat16>() : nullptr;
-  input.scale = scale;
-  input.causal = causal;
-  input.sliding_window_left = sliding_window_left;
-  input.sliding_window_right = sliding_window_right;
-  if (input.causal) {
-    // to make boundary calculation easier
-    input.sliding_window_right = 0;
-  }
-  float softcap_fp32 = softcap;
-  input.softcap = softcap_fp32;
-
-  VLLM_DISPATCH_FLOATING_TYPES(
-      query.scalar_type(), "cpu_attention_with_kv_cache", [&]() {
-        CPU_ATTN_DISPATCH_CASE_HEADDIM(query.size(2), [&] {
-          CPU_ATTN_DISPATCH_IMPL(input.metadata->isa, [&]() {
-            TORCH_CHECK_EQ(input.block_size % attn_impl::BlockSizeAlignment, 0);
-            cpu_attention::AttentionMainLoop<attn_impl> mainloop;
-            mainloop(&input);
-          });
-        });
-      });
-}

csrc/cpu/cpu_attn_amx.hpp

@@ -1,511 +0,0 @@
-#ifndef CPU_ATTN_AMX_HPP
-#define CPU_ATTN_AMX_HPP
-
-#include "cpu_attn_impl.hpp"
-
-namespace cpu_attention {
-namespace {
-// AMX specific
-constexpr static int64_t AMX_TILE_ROW_BYTES = 64;
-constexpr static int64_t AMX_TILE_ROW_NUM = 16;
-constexpr static int64_t AMX_TILE_BYTES = AMX_TILE_ROW_BYTES * AMX_TILE_ROW_NUM;
-
-typedef struct __tile_config {
-  uint8_t palette_id = 1;
-  uint8_t start_row = 0;
-  uint8_t reserved_0[14] = {0};
-  uint16_t colsb[16] = {0};
-  uint8_t rows[16] = {0};
-} __tilecfg;
-
-// 2-2-4 pattern, for 16 < m <= 32
-// TILE 0, 1: load A matrix, row num should be 16, m - 16
-// TILE 2, 3: load B matrix, row num should be 16
-// TILE 4, 5, 6, 7: store results C matrix, row num should be 16, 16, m - 16, m
-// - 16
-template <typename kv_cache_t>
-class TileGemm224 {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size, void* __restrict__ a_tile,
-                                void* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    TORCH_CHECK(false, "Unsupported kv cache type for TileGemm224");
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    TORCH_CHECK(false, "Unsupported kv cache type for TileGemm224");
-  }
-};
-
-template <>
-class TileGemm224<c10::BFloat16> {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size,
-                                c10::BFloat16* __restrict__ a_tile,
-                                c10::BFloat16* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    const int32_t k_times =
-        dynamic_k_size / (AMX_TILE_ROW_NUM * 4 / sizeof(c10::BFloat16));
-    c10::BFloat16* __restrict__ a_tile_0 = a_tile;
-    c10::BFloat16* __restrict__ a_tile_1 = a_tile + lda * AMX_TILE_ROW_NUM;
-    const int64_t a_tile_stride = [&]() {
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // q_buffer is prepacked
-        return AMX_TILE_ROW_BYTES;
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // logits_buffer is row-major
-        return lda * sizeof(c10::BFloat16);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-    }();
-
-    c10::BFloat16* __restrict__ b_tile_2 = b_tile;
-    c10::BFloat16* __restrict__ b_tile_3 = [&]() {
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // k_cache is prepacked
-        return b_tile + (k_size * AMX_TILE_ROW_BYTES / 4);
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // v_cache is prepacked
-        return b_tile + (block_size * AMX_TILE_ROW_BYTES / 4);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-    }();
-    // k_cache, v_cache are prepacked
-    const int32_t b_tile_stride = AMX_TILE_ROW_BYTES;
-
-    // logits_buffer, output_buffer are not prepacked
-    float* __restrict__ c_tile_4 = c_tile;
-    float* __restrict__ c_tile_5 =
-        c_tile_4 + AMX_TILE_ROW_BYTES / sizeof(float);
-    float* __restrict__ c_tile_6 = c_tile + AMX_TILE_ROW_NUM * ldc;
-    float* __restrict__ c_tile_7 =
-        c_tile_6 + AMX_TILE_ROW_BYTES / sizeof(float);
-    const int32_t c_tile_stride = ldc * sizeof(float);
-
-    if (accum_c) {
-      _tile_loadd(4, c_tile_4, c_tile_stride);
-      _tile_loadd(5, c_tile_5, c_tile_stride);
-      _tile_loadd(6, c_tile_6, c_tile_stride);
-      _tile_loadd(7, c_tile_7, c_tile_stride);
-    } else {
-      _tile_zero(4);
-      _tile_zero(5);
-      _tile_zero(6);
-      _tile_zero(7);
-    }
-
-    for (int32_t k = 0; k < k_times; ++k) {
-      _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_tile_stride);
-      _tile_dpbf16ps(4, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_tile_stride);
-      _tile_dpbf16ps(5, 0, 3);
-      _tile_loadd(1, a_tile_1, a_tile_stride);
-      _tile_dpbf16ps(6, 1, 2);
-      _tile_dpbf16ps(7, 1, 3);
-
-      // update ptrs
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // Q buffer is prepacked
-        a_tile_0 += AMX_TILE_BYTES / sizeof(c10::BFloat16);
-        a_tile_1 += AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // P buffer is not prepacked
-        a_tile_0 += AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-        a_tile_1 += AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-      b_tile_2 += AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_3 += AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    }
-
-    _tile_stored(4, c_tile_4, c_tile_stride);
-    _tile_stored(5, c_tile_5, c_tile_stride);
-    _tile_stored(6, c_tile_6, c_tile_stride);
-    _tile_stored(7, c_tile_7, c_tile_stride);
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    const int32_t m_0 = AMX_TILE_ROW_NUM;
-    const int32_t m_1 = m - AMX_TILE_ROW_NUM;
-    config.rows[0] = m_0;
-    config.rows[1] = m_1;
-    config.rows[2] = AMX_TILE_ROW_NUM;
-    config.rows[3] = AMX_TILE_ROW_NUM;
-    config.rows[4] = m_0;
-    config.rows[5] = m_0;
-    config.rows[6] = m_1;
-    config.rows[7] = m_1;
-    _tile_loadconfig(&config);
-  }
-};
-
-// 1-2-2 pattern, for 0 < m <= 16
-// TILE 0, (1): load A matrix, use extra 1 tile for prefetch, row num should be
-// m, m
-// TILE 2, 3, (4, 5): load B matrix, use extra 2 tiles for prefetch, row
-// num should be 16
-// TILE 6, 7, (6, 7): store results C matrix, row num should be
-// m
-template <typename kv_cache_t>
-class TileGemm122 {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size, void* __restrict__ a_tile,
-                                void* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    TORCH_CHECK(false, "Unsupported kv cache type for TileGemm122");
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    TORCH_CHECK(false, "Unsupported kv cache type for TileGemm122");
-  }
-};
-
-template <>
-class TileGemm122<c10::BFloat16> {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size,
-                                c10::BFloat16* __restrict__ a_tile,
-                                c10::BFloat16* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    c10::BFloat16* __restrict__ a_tile_0 = a_tile;
-    c10::BFloat16* __restrict__ a_tile_1 = [&]() {
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // q_buffer is prepacked
-        return a_tile + AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // logits_buffer is row-major
-        return a_tile + AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-    }();
-    const int64_t a_tile_stride = [&]() {
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // q_buffer is prepacked
-        return AMX_TILE_ROW_BYTES;
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // logits_buffer is row-major
-        return lda * sizeof(c10::BFloat16);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-    }();
-
-    c10::BFloat16* __restrict__ b_tile_2 = b_tile;
-    c10::BFloat16* __restrict__ b_tile_3 = [&]() {
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // k_cache is prepacked
-        return b_tile + (k_size * AMX_TILE_ROW_BYTES / 4);
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // v_cache is prepacked
-        return b_tile + (block_size * AMX_TILE_ROW_BYTES / 4);
-      } else {
-        TORCH_CHECK(false, "Unreachable");
-      }
-    }();
-    c10::BFloat16* __restrict__ b_tile_4 =
-        b_tile_2 + AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    c10::BFloat16* __restrict__ b_tile_5 =
-        b_tile_3 + AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    int64_t b_stride = AMX_TILE_ROW_BYTES;
-
-    float* __restrict__ c_tile_6 = c_tile;
-    float* __restrict__ c_tile_7 = c_tile + AMX_TILE_ROW_BYTES / sizeof(float);
-    int64_t c_stride = ldc * sizeof(float);
-
-    const int32_t k_times =
-        dynamic_k_size / (AMX_TILE_ROW_NUM * 4 / sizeof(c10::BFloat16));
-    const int32_t k_group_times = k_times / 2;
-    const bool has_tail = (k_times % 2 == 1);
-
-    if (accum_c) {
-      _tile_loadd(6, c_tile_6, c_stride);
-      _tile_loadd(7, c_tile_7, c_stride);
-    } else {
-      _tile_zero(6);
-      _tile_zero(7);
-    }
-
-    for (int32_t k = 0; k < k_group_times; ++k) {
-      _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_stride);
-      _tile_dpbf16ps(6, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_stride);
-      _tile_dpbf16ps(7, 0, 3);
-      _tile_loadd(1, a_tile_1, a_tile_stride);
-      _tile_stream_loadd(4, b_tile_4, b_stride);
-      _tile_dpbf16ps(6, 1, 4);
-      _tile_stream_loadd(5, b_tile_5, b_stride);
-      _tile_dpbf16ps(7, 1, 5);
-
-      // update ptrs
-      if constexpr (phase == AttentionGemmPhase::QK) {
-        // Q buffer is prepacked
-        a_tile_0 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-        a_tile_1 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      } else if constexpr (phase == AttentionGemmPhase::PV) {
-        // P buffer is not prepacked
-        a_tile_0 += 2 * AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-        a_tile_1 += 2 * AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-      }
-      b_tile_2 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_3 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_4 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_5 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    }
-
-    if (has_tail) {
-      _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_stride);
-      _tile_dpbf16ps(6, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_stride);
-      _tile_dpbf16ps(7, 0, 3);
-    }
-
-    _tile_stored(6, c_tile_6, c_stride);
-    _tile_stored(7, c_tile_7, c_stride);
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    config.rows[0] = m;
-    config.rows[1] = m;
-    config.rows[2] = AMX_TILE_ROW_NUM;
-    config.rows[3] = AMX_TILE_ROW_NUM;
-    config.rows[4] = AMX_TILE_ROW_NUM;
-    config.rows[5] = AMX_TILE_ROW_NUM;
-    config.rows[6] = m;
-    config.rows[7] = m;
-    _tile_loadconfig(&config);
-  }
-};
-}  // namespace
-
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::AMX, scalar_t, head_dim> {
- public:
-  using query_t = scalar_t;
-  using q_buffer_t = scalar_t;
-  using kv_cache_t = scalar_t;
-  using logits_buffer_t = float;
-  using partial_output_buffer_t = float;
-  using prob_buffer_t = scalar_t;
-
-  constexpr static int64_t BlockSizeAlignment =
-      AMX_TILE_ROW_BYTES /
-      sizeof(kv_cache_t);  // KV token num unit of QK and PV phases
-  constexpr static int64_t HeadDimAlignment =
-      2 * (AMX_TILE_ROW_BYTES / 4);  // headdim num unit of PV phase
-  constexpr static int64_t MaxQHeadNumPerIteration = 32;
-  constexpr static int64_t HeadDim = head_dim;
-  constexpr static ISA ISAType = ISA::AMX;
-  constexpr static bool scale_on_logits = true;
-
- public:
-  AttentionImpl() : current_q_head_num_(0) {
-    // Use all columns in AMX tiles
-    vec_op::unroll_loop<int, 8>([&](int i) { amx_tile_config_.colsb[i] = 64; });
-  }
-
-  ~AttentionImpl() { _tile_release(); }
-
-  template <template <typename tile_gemm_t> typename attention>
-  FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
-    if (q_head_num > AMX_TILE_ROW_NUM) {
-      if (q_head_num != current_q_head_num_) {
-        current_q_head_num_ = q_head_num;
-        TileGemm224<kv_cache_t>::init_tile_config(q_head_num, amx_tile_config_);
-      }
-      attention<TileGemm224<kv_cache_t>> attention_iteration;
-      attention_iteration(CPU_ATTENTION_PARAMS);
-    } else {
-      if (q_head_num != current_q_head_num_) {
-        current_q_head_num_ = q_head_num;
-        TileGemm122<kv_cache_t>::init_tile_config(q_head_num, amx_tile_config_);
-      }
-      attention<TileGemm122<kv_cache_t>> attention_iteration;
-      attention_iteration(CPU_ATTENTION_PARAMS);
-    }
-  }
-
-  // k_cache_token_group_stride: stride of K cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t k_cache_token_group_stride(
-      const int32_t block_size) {
-    return BlockSizeAlignment * head_dim;
-  }
-
-  // v_cache_token_group_stride: stride of V cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t v_cache_token_group_stride(
-      const int32_t block_size) {
-    return BlockSizeAlignment * (AMX_TILE_ROW_BYTES / 4);
-  }
-
-  // v_cache_head_group_stride: stride of V cache when move to next
-  // HeadDimAlignment head dims in a block
-  constexpr static int64_t v_cache_head_group_stride(const int32_t block_size) {
-    return block_size * HeadDimAlignment;
-  }
-
-  static void copy_q_heads_tile(
-      scalar_t* __restrict__ src,  // [q_num, q_heads_per_kv, head_size]
-      scalar_t* __restrict__ q_buffer, const int32_t q_num,
-      const int32_t q_heads_per_kv, const int64_t q_num_stride,
-      const int64_t q_head_stride, const float scale) {
-    constexpr int64_t bytes_per_head = head_dim * sizeof(scalar_t);
-    static_assert(bytes_per_head % AMX_TILE_ROW_BYTES == 0);
-    constexpr int64_t head_size_block_num = bytes_per_head / AMX_TILE_ROW_BYTES;
-    constexpr int64_t head_elem_num_pre_block =
-        AMX_TILE_ROW_BYTES / sizeof(scalar_t);
-
-    int32_t idx = 0;
-    int8_t* __restrict__ q_buffer_iter = reinterpret_cast<int8_t*>(q_buffer);
-    for (int32_t q_num_idx = 0; q_num_idx < q_num;
-         ++q_num_idx, src += q_num_stride) {
-      scalar_t* __restrict__ src_iter = src;
-      for (int32_t q_head_idx = 0; q_head_idx < q_heads_per_kv;
-           ++q_head_idx, src_iter += q_head_stride) {
-        vec_op::unroll_loop<int32_t, head_size_block_num>(
-            [&](int32_t head_size_block_idx) {
-              // Use INT8Vec64 for 64 bytes block
-              vec_op::INT8Vec64 vec(src_iter + head_size_block_idx *
-                                                   head_elem_num_pre_block);
-              vec.save(q_buffer_iter + head_size_block_idx * AMX_TILE_BYTES);
-            });
-
-        ++idx;
-        q_buffer_iter += AMX_TILE_ROW_BYTES;
-        if ((idx & (AMX_TILE_ROW_NUM - 1)) == 0) {
-          // head is in another amx tile
-          q_buffer_iter -= AMX_TILE_ROW_NUM * AMX_TILE_ROW_BYTES;
-          q_buffer_iter += head_size_block_num * AMX_TILE_BYTES;
-        }
-      }
-    }
-  }
-
-  // reshape KV to AMX friendly layout
-  static void reshape_and_cache(
-      const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
-      scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
-      const int64_t* __restrict__ slot_mapping, const int64_t token_num,
-      const int64_t key_token_num_stride, const int64_t value_token_num_stride,
-      const int64_t head_num, const int64_t key_head_num_stride,
-      const int64_t value_head_num_stride, const int64_t num_blocks,
-      const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
-      const int64_t block_size, const int64_t block_size_stride) {
-    // For AMX 2D tiles, size of each line is 64 bytes
-    constexpr int64_t amx_tile_row_size = AMX_TILE_ROW_BYTES;
-    // For AMX B martix, N always is 16
-    constexpr int64_t amx_b_tile_n_size = AMX_TILE_ROW_BYTES / 4;
-    constexpr int64_t amx_b_tile_k_size = amx_tile_row_size / sizeof(scalar_t);
-    // For now suppose block_size is divisible by amx_tile_column_num
-    TORCH_CHECK_EQ(block_size % amx_b_tile_k_size, 0);
-
-#pragma omp parallel for collapse(2)
-    for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
-      for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
-        const int64_t pos = slot_mapping[token_idx];
-        if (pos < 0) {
-          // skip
-          continue;
-        }
-
-        const int64_t block_idx = pos / block_size;
-        const int64_t block_offset = pos % block_size;
-        {
-          // Write Key
-          // Head elements should be packed as quand-words and stored in token
-          // groups with (quadword_stride/4) tokens
-          constexpr int64_t token_num_per_group = amx_tile_row_size / 4;
-          static_assert(head_dim % (4 / sizeof(scalar_t)) == 0);
-          constexpr int64_t quadword_num = head_dim / (4 / sizeof(scalar_t));
-          const int32_t* key_start_quadword_ptr =
-              reinterpret_cast<const int32_t*>(
-                  key + token_idx * key_token_num_stride +
-                  head_idx * key_head_num_stride);
-          const int64_t group_idx = block_offset / token_num_per_group;
-          const int64_t group_offset = block_offset % token_num_per_group;
-          constexpr int64_t quadword_num_per_group =
-              token_num_per_group * quadword_num;
-          int32_t* key_cache_start_ptr =
-              reinterpret_cast<int32_t*>(key_cache +
-                                         block_idx * num_blocks_stride +
-                                         head_idx * cache_head_num_stride) +
-              group_idx * quadword_num_per_group + group_offset;
-
-#pragma GCC unroll 8
-          for (int64_t i = 0, j = 0; j < quadword_num;
-               i += token_num_per_group, ++j) {
-            key_cache_start_ptr[i] = key_start_quadword_ptr[j];
-          }
-        }
-        {
-          // Write Value
-          // Different from Key, block_size dimension is packed rather than
-          // head_size dimension block_size dimension is packed as quand-words;
-          constexpr int64_t token_num_per_sub_group = 4 / sizeof(scalar_t);
-          const int64_t token_num_per_group = block_size;
-          constexpr int64_t head_elems_per_group = amx_b_tile_n_size;
-          const int64_t group_size = token_num_per_group * head_elems_per_group;
-          // For now suppose head_dim is divisible by amx_b_tile_n_size
-          static_assert(head_dim % head_elems_per_group == 0);
-          constexpr int64_t group_num = head_dim / head_elems_per_group;
-          const int64_t sub_group_idx = block_offset / token_num_per_sub_group;
-          const int64_t sub_group_offset =
-              block_offset % token_num_per_sub_group;
-
-          const scalar_t* value_start_ptr = value +
-                                            token_idx * value_token_num_stride +
-                                            head_idx * value_head_num_stride;
-          scalar_t* value_cache_start_ptr =
-              value_cache + block_idx * num_blocks_stride +
-              head_idx * cache_head_num_stride +
-              sub_group_idx * token_num_per_sub_group * amx_b_tile_n_size +
-              sub_group_offset;
-
-          for (int64_t i = 0; i < group_num; ++i) {
-#pragma GCC unroll head_elems_per_group
-            for (int64_t j = 0, k = 0; j < head_elems_per_group;
-                 ++j, k += token_num_per_sub_group) {
-              value_cache_start_ptr[k] = value_start_ptr[j];
-            }
-            value_start_ptr += head_elems_per_group;
-            value_cache_start_ptr += group_size;
-          }
-        }
-      }
-    }
-  }
-
- private:
-  alignas(64) __tilecfg amx_tile_config_;
-  int32_t current_q_head_num_;
-};
-}  // namespace cpu_attention
-
-#endif

csrc/cpu/cpu_attn_macros.h

@@ -1,63 +0,0 @@
-#ifndef CPU_ATTN_MACROS_H
-#define CPU_ATTN_MACROS_H
-
-// x86_64
-#ifdef __x86_64__
-  #define FAST_SPINNING _mm_pause();
-
-  #ifdef __AVX512F__
-    #define DEFINE_FAST_EXP                                                    \
-      const __m512 vec_factorial_1 = _mm512_set1_ps(0.999999701f);             \
-      const __m512 vec_factorial_2 = _mm512_set1_ps(0.499991506f);             \
-      const __m512 vec_factorial_3 = _mm512_set1_ps(0.166676521f);             \
-      const __m512 vec_factorial_4 = _mm512_set1_ps(0.0418978221f);            \
-      const __m512 vec_factorial_5 = _mm512_set1_ps(0.00828929059f);           \
-      const __m512 vec_exp_log2ef =                                            \
-          _mm512_castsi512_ps(_mm512_set1_epi32(0x3fb8aa3b));                  \
-      const __m512 vec_half = _mm512_set1_ps(0.5f);                            \
-      const __m512 vec_one = _mm512_set1_ps(1.f);                              \
-      const __m512 vec_zero = _mm512_set1_ps(0.f);                             \
-      const __m512 vec_two = _mm512_set1_ps(2.f);                              \
-      const __m512 vec_ln2f =                                                  \
-          _mm512_castsi512_ps(_mm512_set1_epi32(0x3f317218));                  \
-      const __m512 vec_ln_flt_min =                                            \
-          _mm512_castsi512_ps(_mm512_set1_epi32(0xc2aeac50));                  \
-      const __m512 vec_ln_flt_max =                                            \
-          _mm512_castsi512_ps(_mm512_set1_epi32(0x42b17218));                  \
-      const __m512i vec_127 = _mm512_set1_epi32(0x0000007f);                   \
-      const int n_mantissa_bits = 23;                                          \
-      auto fast_exp = [&](vec_op::FP32Vec16& vec) __attribute__((              \
-                          always_inline)) {                                    \
-        __m512 values = vec.reg;                                               \
-        auto less_ln_flt_min_mask =                                            \
-            _mm512_cmp_ps_mask(values, vec_ln_flt_min, 1 /*_CMP_LT_OS*/);      \
-        auto vec_src = _mm512_min_ps(values, vec_ln_flt_max);                  \
-        vec_src = _mm512_max_ps(vec_src, vec_ln_flt_min);                      \
-        auto vec_fx = _mm512_fmadd_ps(vec_src, vec_exp_log2ef, vec_half);      \
-        auto vec_fx_i = _mm512_cvt_roundps_epi32(                              \
-            vec_fx, _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC);                \
-        vec_fx = _mm512_cvtepi32_ps(vec_fx_i);                                 \
-        auto vec_exp_poly = _mm512_fnmadd_ps(vec_fx, vec_ln2f, vec_src);       \
-        auto vec_res =                                                         \
-            _mm512_fmadd_ps(vec_exp_poly, vec_factorial_5, vec_factorial_4);   \
-        vec_res = _mm512_fmadd_ps(vec_exp_poly, vec_res, vec_factorial_3);     \
-        vec_res = _mm512_fmadd_ps(vec_exp_poly, vec_res, vec_factorial_2);     \
-        vec_res = _mm512_fmadd_ps(vec_exp_poly, vec_res, vec_factorial_1);     \
-        vec_res = _mm512_fmadd_ps(vec_exp_poly, vec_res, vec_one);             \
-        auto vec_exp_number = _mm512_sub_ps(vec_fx, vec_one);                  \
-        auto vec_exp_number_i = _mm512_cvtps_epi32(vec_exp_number);            \
-        auto vec_two_pow_n_i = _mm512_add_epi32(vec_exp_number_i, vec_127);    \
-        vec_two_pow_n_i = _mm512_slli_epi32(vec_two_pow_n_i, n_mantissa_bits); \
-        auto vec_two_pow_n = _mm512_castsi512_ps(vec_two_pow_n_i);             \
-        vec_two_pow_n = _mm512_mask_blend_ps(less_ln_flt_min_mask,             \
-                                             vec_two_pow_n, vec_zero);         \
-        vec_res = _mm512_mul_ps(vec_res, vec_two_pow_n);                       \
-        vec_res = _mm512_mul_ps(vec_res, vec_two);                             \
-        vec_op::FP32Vec16 res(vec_res);                                        \
-        return res;                                                            \
-      };
-  #endif
-
-#endif
-
-#endif

csrc/cpu/cpu_attn_neon.hpp

@@ -1,386 +0,0 @@
-#ifndef CPU_ATTN_NEON_HPP
-#define CPU_ATTN_NEON_HPP
-
-#include "cpu_attn_impl.hpp"
-#include <arm_neon.h>
-#include <type_traits>
-namespace cpu_attention {
-
-namespace {
-
-#define BLOCK_SIZE_ALIGNMENT 32
-#define HEAD_SIZE_ALIGNMENT 32
-#define MAX_Q_HEAD_NUM_PER_ITER 16
-
-// These do not use vectorized class for loading / converting
-// because csrc/cpu/cpu_types_arm.hpp does not have fallback options
-// for vec_op::BF16Vec* / vec_op::BF16Vec* on Arm HW that
-// doesn't support BF16.
-// We don't use vec_op::FP32Vec* or vec_op::FP16Vec* for consistency.
-template <typename kv_cache_t>
-FORCE_INLINE void load_row8_B_as_f32(const kv_cache_t* p, float32x4_t& b0,
-                                     float32x4_t& b1);
-
-template <>
-FORCE_INLINE void load_row8_B_as_f32<float>(const float* p, float32x4_t& b0,
-                                            float32x4_t& b1) {
-  b0 = vld1q_f32(p + 0);
-  b1 = vld1q_f32(p + 4);
-}
-
-template <>
-FORCE_INLINE void load_row8_B_as_f32<c10::Half>(const c10::Half* p,
-                                                float32x4_t& b0,
-                                                float32x4_t& b1) {
-  const float16_t* h = reinterpret_cast<const float16_t*>(p);
-  float16x8_t v = vld1q_f16(h);
-  b0 = vcvt_f32_f16(vget_low_f16(v));
-  b1 = vcvt_f32_f16(vget_high_f16(v));
-}
-
-template <>
-FORCE_INLINE void load_row8_B_as_f32<c10::BFloat16>(const c10::BFloat16* p,
-                                                    float32x4_t& b0,
-                                                    float32x4_t& b1) {
-  const uint16_t* u = reinterpret_cast<const uint16_t*>(p);
-#ifdef ARM_BF16_SUPPORT
-  uint16x8_t u0 = vld1q_u16(u);
-  bfloat16x8_t bf0 = vreinterpretq_bf16_u16(u0);
-  b0 = vcvtq_low_f32_bf16(bf0);
-  b1 = vcvtq_high_f32_bf16(bf0);
-#else
-  uint16x8_t x0 = vld1q_u16(u);
-  uint32x4_t lo = vshlq_n_u32(vmovl_u16(vget_low_u16(x0)), 16);
-  uint32x4_t hi = vshlq_n_u32(vmovl_u16(vget_high_u16(x0)), 16);
-  b0 = vreinterpretq_f32_u32(lo);
-  b1 = vreinterpretq_f32_u32(hi);
-#endif
-}
-
-// Mx8, with 1 <= M <= 8 , K streamed, unroll-by-4 with NEON FMLAs
-// #Loads = (K // 4) * (M + 4 * sizeof(kv_cache_t) / 2)
-// #FMLAs = (K // 4) * (4 * 2 * M)
-// We have (4 * 2 * M) FMLAs for (M + 4 * sizeof(kv_cache_t) / 2) loads
-template <int32_t M, typename kv_cache_t>
-FORCE_INLINE void gemm_micro_neon_fmla_Mx8_Ku4(
-    const float* __restrict A,       // [M x K],
-    const kv_cache_t* __restrict B,  // [K x 8],
-    float* __restrict C,             // [M x 8],
-    int64_t lda, int64_t ldb, int64_t ldc, int32_t K, bool accumulate) {
-  // kernel supports max M of 8, as it'd spill for larger M
-  static_assert(1 <= M && M <= 8, "M must be in [1,8]");
-
-// helpers for per-M codegen
-#define ROWS_APPLY(OP) OP(0) OP(1) OP(2) OP(3) OP(4) OP(5) OP(6) OP(7)
-#define IF_M(i) if constexpr (M > (i))
-
-  // A row base pointers
-#define DECL_A(i) const float* a##i = A + (i) * lda;
-  ROWS_APPLY(DECL_A)
-#undef DECL_A
-
-  // declare 2 accumulators per row of M
-#define DECL_ACC(i) float32x4_t acc##i##_0, acc##i##_1;
-  ROWS_APPLY(DECL_ACC)
-#undef DECL_ACC
-
-  // initialize accumulators
-#define INIT_ACC(i)                              \
-  IF_M(i) {                                      \
-    if (accumulate) {                            \
-      acc##i##_0 = vld1q_f32(C + (i) * ldc + 0); \
-      acc##i##_1 = vld1q_f32(C + (i) * ldc + 4); \
-    } else {                                     \
-      acc##i##_0 = vdupq_n_f32(0.f);             \
-      acc##i##_1 = vdupq_n_f32(0.f);             \
-    }                                            \
-  }
-  ROWS_APPLY(INIT_ACC)
-#undef INIT_ACC
-
-  int32_t k = 0;
-
-  // K unrolled by 4
-  for (; k + 3 < K; k += 4) {
-    // load A[k..k+3] for each active row (M)
-#define LOAD_A4(i)     \
-  float32x4_t a##i##v; \
-  IF_M(i) a##i##v = vld1q_f32(a##i + k);
-    ROWS_APPLY(LOAD_A4)
-#undef LOAD_A4
-
-    // helper: FMA lane L from aiv
-#define FMAS_LANE(i, aiv, L)                              \
-  IF_M(i) {                                               \
-    acc##i##_0 = vfmaq_laneq_f32(acc##i##_0, b0, aiv, L); \
-    acc##i##_1 = vfmaq_laneq_f32(acc##i##_1, b1, aiv, L); \
-  }
-
-    // k + 0
-    {
-      float32x4_t b0, b1;
-      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 0) * ldb, b0, b1);
-#define STEP_K0(i) FMAS_LANE(i, a##i##v, 0)
-      ROWS_APPLY(STEP_K0)
-#undef STEP_K0
-    }
-    // k + 1
-    {
-      float32x4_t b0, b1;
-      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 1) * ldb, b0, b1);
-#define STEP_K1(i) FMAS_LANE(i, a##i##v, 1)
-      ROWS_APPLY(STEP_K1)
-#undef STEP_K1
-    }
-    // k + 2
-    {
-      float32x4_t b0, b1;
-      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 2) * ldb, b0, b1);
-#define STEP_K2(i) FMAS_LANE(i, a##i##v, 2)
-      ROWS_APPLY(STEP_K2)
-#undef STEP_K2
-    }
-    // k + 3
-    {
-      float32x4_t b0, b1;
-      load_row8_B_as_f32<kv_cache_t>(B + (int64_t)(k + 3) * ldb, b0, b1);
-#define STEP_K3(i) FMAS_LANE(i, a##i##v, 3)
-      ROWS_APPLY(STEP_K3)
-#undef STEP_K3
-    }
-#undef FMAS_LANE
-  }
-
-  // K tail
-  for (; k < K; ++k) {
-    float32x4_t b0, b1;
-    load_row8_B_as_f32<kv_cache_t>(B + (int64_t)k * ldb, b0, b1);
-#define TAIL_ROW(i)                             \
-  IF_M(i) {                                     \
-    float32x4_t ai = vdupq_n_f32(*(a##i + k));  \
-    acc##i##_0 = vfmaq_f32(acc##i##_0, b0, ai); \
-    acc##i##_1 = vfmaq_f32(acc##i##_1, b1, ai); \
-  }
-    ROWS_APPLY(TAIL_ROW)
-#undef TAIL_ROW
-  }
-
-  // store accumulators to C
-#define STORE_ROW(i)                          \
-  IF_M(i) {                                   \
-    vst1q_f32(C + (i) * ldc + 0, acc##i##_0); \
-    vst1q_f32(C + (i) * ldc + 4, acc##i##_1); \
-  }
-  ROWS_APPLY(STORE_ROW)
-#undef STORE_ROW
-
-#undef ROWS_APPLY
-#undef IF_M
-}
-
-template <int32_t N, typename kv_cache_t>
-FORCE_INLINE void gemm_macro_neon_fmla_Mx8_Ku4(const float* __restrict A,
-                                               const kv_cache_t* __restrict B,
-                                               float* __restrict C, int32_t M,
-                                               int32_t K, int64_t lda,
-                                               int64_t ldb, int64_t ldc,
-                                               bool accumulate) {
-  // micro kernel is Mx8
-  static_assert(N % 8 == 0, "N must be a multiple of 8");
-  for (int32_t m = 0; m < M;) {
-    int32_t mb = (M - m >= 8) ? 8 : (M - m >= 4) ? 4 : (M - m >= 2) ? 2 : 1;
-    const float* Ab = A + m * lda;
-    float* Cb = C + m * ldc;
-
-    for (int32_t n = 0; n < N; n += 8) {
-      const kv_cache_t* Bn = B + n;
-      float* Cn = Cb + n;
-      switch (mb) {
-        case 8:
-          gemm_micro_neon_fmla_Mx8_Ku4<8, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc,
-                                                      K, accumulate);
-          break;
-        case 4:
-          gemm_micro_neon_fmla_Mx8_Ku4<4, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc,
-                                                      K, accumulate);
-          break;
-        case 2:
-          gemm_micro_neon_fmla_Mx8_Ku4<2, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc,
-                                                      K, accumulate);
-          break;
-        default:
-          gemm_micro_neon_fmla_Mx8_Ku4<1, kv_cache_t>(Ab, Bn, Cn, lda, ldb, ldc,
-                                                      K, accumulate);
-          break;
-      }
-    }
-    // no tail loop for N as it's guaranteed to be a multiple of 8
-    m += mb;
-  }
-}
-
-template <typename kv_cache_t>
-class TileGemmNeonFMLA {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size,
-                                float* __restrict__ a_tile,
-                                kv_cache_t* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    if constexpr (phase == AttentionGemmPhase::QK) {
-      gemm_macro_neon_fmla_Mx8_Ku4<BLOCK_SIZE_ALIGNMENT, kv_cache_t>(
-          a_tile, b_tile, c_tile, m_size, k_size, lda, ldb, ldc, accum_c);
-    } else {
-      gemm_macro_neon_fmla_Mx8_Ku4<HEAD_SIZE_ALIGNMENT, kv_cache_t>(
-          a_tile, b_tile, c_tile, m_size, dynamic_k_size, lda, ldb, ldc,
-          accum_c);
-    }
-  }
-};
-
-}  // namespace
-
-// this is similar to "ISA::VEC" at the moment
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::NEON, scalar_t, head_dim> {
- public:
-  using query_t = scalar_t;
-  using q_buffer_t = float;
-  using kv_cache_t = scalar_t;
-  using logits_buffer_t = float;
-  using partial_output_buffer_t = float;
-  using prob_buffer_t = float;
-
-  constexpr static int64_t BlockSizeAlignment =
-      BLOCK_SIZE_ALIGNMENT;  // KV token num unit of QK and PV phases
-  constexpr static int64_t HeadDimAlignment =
-      HEAD_SIZE_ALIGNMENT;  // headdim num unit of PV phase
-  constexpr static int64_t MaxQHeadNumPerIteration = MAX_Q_HEAD_NUM_PER_ITER;
-  constexpr static int64_t HeadDim = head_dim;
-  constexpr static ISA ISAType = ISA::NEON;
-  constexpr static bool scale_on_logits = false;  // apply scale on q_buffer
-
-  static_assert(HeadDim % HeadDimAlignment == 0);
-  // the gemm micro kernel is Mx8
-  static_assert(HeadDimAlignment % 8 == 0);
-  static_assert(BlockSizeAlignment % 8 == 0);
-
- public:
-  template <template <typename tile_gemm_t> typename attention>
-  FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
-    attention<TileGemmNeonFMLA<kv_cache_t>> attention_iteration;
-    attention_iteration(CPU_ATTENTION_PARAMS);
-  }
-
-  // k_cache_token_group_stride: stride of K cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t k_cache_token_group_stride(
-      const int32_t block_size) {
-    return BlockSizeAlignment;  // layout of k_cache block is [head_dim,
-                                // block_size], row-major
-  }
-
-  // v_cache_token_group_stride: stride of V cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t v_cache_token_group_stride(
-      const int32_t block_size) {
-    return head_dim * BlockSizeAlignment;  // layout of v_cache is [block_size,
-                                           // head_dim], row-major
-  }
-
-  // v_cache_head_group_stride: stride of V cache when move to next
-  // HeadDimAlignment head dims in a block
-  constexpr static int64_t v_cache_head_group_stride(const int32_t block_size) {
-    return HeadDimAlignment;  // layout of v_cache is [block_size, head_dim],
-                              // row-major
-  }
-
-  // Copy q to q_buffer and cast it to fp32
-  static void copy_q_heads_tile(
-      scalar_t* __restrict__ src,  // [q_num, q_heads_per_kv, head_size]
-      float* __restrict__ q_buffer, const int32_t q_num,
-      const int32_t q_heads_per_kv, const int64_t q_num_stride,
-      const int64_t q_head_stride, float scale) {
-    static_assert(head_dim % 16 == 0);
-    constexpr int32_t unroll_size = head_dim / 16;
-    using load_vec_t = typename VecTypeTrait<scalar_t>::vec_t;
-
-    vec_op::FP32Vec16 scale_vec(scale);
-    for (int32_t q_num_idx = 0; q_num_idx < q_num; ++q_num_idx) {
-      for (int32_t q_head_idx = 0; q_head_idx < q_heads_per_kv; ++q_head_idx) {
-        scalar_t* __restrict__ curr_q =
-            src + q_num_idx * q_num_stride + q_head_idx * q_head_stride;
-        float* __restrict__ curr_q_buffer =
-            q_buffer + q_num_idx * q_heads_per_kv * head_dim +
-            q_head_idx * head_dim;
-
-        vec_op::unroll_loop<int32_t, unroll_size>([&](int32_t i) {
-          load_vec_t vec(curr_q);
-          vec_op::FP32Vec16 fp32_vec(vec);
-          fp32_vec = fp32_vec * scale_vec;
-          fp32_vec.save(curr_q_buffer);
-
-          curr_q += 16;
-          curr_q_buffer += 16;
-        });
-      }
-    }
-  }
-
-  // reshape K as column-major and V as row-major
-  static void reshape_and_cache(
-      const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
-      scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
-      const int64_t* __restrict__ slot_mapping, const int64_t token_num,
-      const int64_t key_token_num_stride, const int64_t value_token_num_stride,
-      const int64_t head_num, const int64_t key_head_num_stride,
-      const int64_t value_head_num_stride, const int64_t num_blocks,
-      const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
-      const int64_t block_size, const int64_t block_size_stride) {
-#pragma omp parallel for collapse(2)
-    for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
-      for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
-        const int64_t pos = slot_mapping[token_idx];
-        if (pos < 0) {
-          // skip
-          continue;
-        }
-
-        const int64_t block_idx = pos / block_size;
-        const int64_t block_offset = pos % block_size;
-        {
-          // Write Key
-          const scalar_t* key_start_ptr = key +
-                                          token_idx * key_token_num_stride +
-                                          head_idx * key_head_num_stride;
-          scalar_t* key_cache_start_ptr =
-              key_cache + block_idx * num_blocks_stride +
-              head_idx * cache_head_num_stride + block_offset;
-
-#pragma GCC unroll 8
-          for (int64_t i = 0, j = 0; i < head_dim; ++i, j += block_size) {
-            key_cache_start_ptr[j] = key_start_ptr[i];
-          }
-        }
-        {
-          // Write Value
-          const scalar_t* value_start_ptr = value +
-                                            token_idx * value_token_num_stride +
-                                            head_idx * value_head_num_stride;
-          scalar_t* value_cache_start_ptr =
-              value_cache + block_idx * num_blocks_stride +
-              head_idx * cache_head_num_stride + block_offset * head_dim;
-          std::memcpy(value_cache_start_ptr, value_start_ptr,
-                      sizeof(scalar_t) * head_dim);
-        }
-      }
-    }
-  }
-};
-}  // namespace cpu_attention
-
-#endif  // #ifndef CPU_ATTN_NEON_HPP

csrc/cpu/cpu_attn_vec.hpp

@@ -1,248 +0,0 @@
-#ifndef CPU_ATTN_VEC_HPP
-#define CPU_ATTN_VEC_HPP
-
-#include "cpu_attn_impl.hpp"
-
-namespace cpu_attention {
-
-namespace {
-// 8-2-16 pattern, 8 regs for A, 2 regs for B, 16 regs for C, [8, K] @ [k, 32]
-template <typename kv_cache_t>
-class TileGemm82 {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size,
-                                float* __restrict__ a_tile,
-                                kv_cache_t* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    switch (m_size) {
-      case 1:
-        gemm_micro<1>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 2:
-        gemm_micro<2>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 3:
-      case 4:
-        gemm_micro<4>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 5:
-      case 6:
-        gemm_micro<6>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 7:
-      case 8:
-        gemm_micro<8>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-    }
-  }
-
-  template <int32_t M>
-  static void gemm_micro(float* __restrict__ a_tile,
-                         kv_cache_t* __restrict__ b_tile,
-                         float* __restrict__ c_tile, const int64_t lda,
-                         const int64_t ldb, const int64_t ldc,
-                         const int32_t block_size, const int32_t dynamic_k_size,
-                         const bool accum_c) {
-    static_assert(0 < M <= 8);
-    using load_vec_t = typename VecTypeTrait<kv_cache_t>::vec_t;
-
-    kv_cache_t* __restrict__ curr_b_0 = b_tile;
-    kv_cache_t* __restrict__ curr_b_1 = b_tile + 16;
-    float* __restrict__ curr_c_0 = c_tile;
-    float* __restrict__ curr_c_1 = c_tile + 16;
-
-    vec_op::FP32Vec16 c_regs[M * 2];
-    if (accum_c) {
-      float* __restrict__ curr_m_c_0 = curr_c_0;
-      float* __restrict__ curr_m_c_1 = curr_c_1;
-      vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        c_regs[i * 2] = vec_op::FP32Vec16(curr_m_c_0);
-        c_regs[i * 2 + 1] = vec_op::FP32Vec16(curr_m_c_1);
-
-        // update
-        curr_m_c_0 += ldc;
-        curr_m_c_1 += ldc;
-      });
-    }
-
-    float* __restrict__ curr_a = a_tile;
-    for (int32_t k = 0; k < dynamic_k_size; ++k) {
-      load_vec_t b_0_reg(curr_b_0);
-      vec_op::FP32Vec16 fp32_b_0_reg(b_0_reg);
-      load_vec_t b_1_reg(curr_b_1);
-      vec_op::FP32Vec16 fp32_b_1_reg(b_1_reg);
-
-      float* __restrict__ curr_m_a = curr_a;
-      vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        float v = *curr_m_a;
-        vec_op::FP32Vec16 a_reg(v);
-        c_regs[i * 2] = c_regs[i * 2] + a_reg * fp32_b_0_reg;
-        c_regs[i * 2 + 1] = c_regs[i * 2 + 1] + a_reg * fp32_b_1_reg;
-
-        // update
-        curr_m_a += lda;
-      });
-
-      // update
-      curr_a += 1;
-      curr_b_0 += ldb;
-      curr_b_1 += ldb;
-    }
-
-    vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-      c_regs[i * 2].save(curr_c_0);
-      c_regs[i * 2 + 1].save(curr_c_1);
-
-      // update
-      curr_c_0 += ldc;
-      curr_c_1 += ldc;
-    });
-  }
-};
-}  // namespace
-
-// This is a general but naive implementation based on vector instructions
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::VEC, scalar_t, head_dim> {
- public:
-  using query_t = scalar_t;
-  using q_buffer_t = float;
-  using kv_cache_t = scalar_t;
-  using logits_buffer_t = float;
-  using partial_output_buffer_t = float;
-  using prob_buffer_t = float;
-
-  constexpr static int64_t BlockSizeAlignment =
-      32;  // KV token num unit of QK and PV phases
-  constexpr static int64_t HeadDimAlignment =
-      32;  // headdim num unit of PV phase
-  constexpr static int64_t MaxQHeadNumPerIteration = 8;
-  constexpr static int64_t HeadDim = head_dim;
-  constexpr static ISA ISAType = ISA::VEC;
-  constexpr static bool scale_on_logits = false;  // apply scale on q_buffer
-
- public:
-  template <template <typename tile_gemm_t> typename attention>
-  FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
-    attention<TileGemm82<kv_cache_t>> attention_iteration;
-    attention_iteration(CPU_ATTENTION_PARAMS);
-  }
-
-  // k_cache_token_group_stride: stride of K cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t k_cache_token_group_stride(
-      const int32_t block_size) {
-    return BlockSizeAlignment;  // layout of k_cache block is [head_dim,
-                                // block_size], row-major
-  }
-
-  // v_cache_token_group_stride: stride of V cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t v_cache_token_group_stride(
-      const int32_t block_size) {
-    return head_dim * BlockSizeAlignment;  // layout of v_cache is [block_size,
-                                           // head_dim], row-major
-  }
-
-  // v_cache_head_group_stride: stride of V cache when move to next
-  // HeadDimAlignment head dims in a block
-  constexpr static int64_t v_cache_head_group_stride(const int32_t block_size) {
-    return HeadDimAlignment;  // layout of v_cache is [block_size, head_dim],
-                              // row-major
-  }
-
-  // Copy q to q_buffer and cast it to fp32
-  static void copy_q_heads_tile(
-      scalar_t* __restrict__ src,  // [q_num, q_heads_per_kv, head_size]
-      float* __restrict__ q_buffer, const int32_t q_num,
-      const int32_t q_heads_per_kv, const int64_t q_num_stride,
-      const int64_t q_head_stride, float scale) {
-    static_assert(head_dim % 16 == 0);
-    constexpr int32_t unroll_size = head_dim / 16;
-    using load_vec_t = typename VecTypeTrait<scalar_t>::vec_t;
-
-    vec_op::FP32Vec16 scale_vec(scale);
-    for (int32_t q_num_idx = 0; q_num_idx < q_num; ++q_num_idx) {
-      for (int32_t q_head_idx = 0; q_head_idx < q_heads_per_kv; ++q_head_idx) {
-        scalar_t* __restrict__ curr_q =
-            src + q_num_idx * q_num_stride + q_head_idx * q_head_stride;
-        float* __restrict__ curr_q_buffer =
-            q_buffer + q_num_idx * q_heads_per_kv * head_dim +
-            q_head_idx * head_dim;
-
-        vec_op::unroll_loop<int32_t, unroll_size>([&](int32_t i) {
-          load_vec_t vec(curr_q);
-          vec_op::FP32Vec16 fp32_vec(vec);
-          fp32_vec = fp32_vec * scale_vec;
-          fp32_vec.save(curr_q_buffer);
-
-          curr_q += 16;
-          curr_q_buffer += 16;
-        });
-      }
-    }
-  }
-
-  // reshape K as column-major and V as row-major
-  static void reshape_and_cache(
-      const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
-      scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
-      const int64_t* __restrict__ slot_mapping, const int64_t token_num,
-      const int64_t key_token_num_stride, const int64_t value_token_num_stride,
-      const int64_t head_num, const int64_t key_head_num_stride,
-      const int64_t value_head_num_stride, const int64_t num_blocks,
-      const int64_t num_blocks_stride, const int64_t cache_head_num_stride,
-      const int64_t block_size, const int64_t block_size_stride) {
-#pragma omp parallel for collapse(2)
-    for (int64_t token_idx = 0; token_idx < token_num; ++token_idx) {
-      for (int64_t head_idx = 0; head_idx < head_num; ++head_idx) {
-        const int64_t pos = slot_mapping[token_idx];
-        if (pos < 0) {
-          // skip
-          continue;
-        }
-
-        const int64_t block_idx = pos / block_size;
-        const int64_t block_offset = pos % block_size;
-        {
-          // Write Key as column-major
-          const scalar_t* key_start_ptr = key +
-                                          token_idx * key_token_num_stride +
-                                          head_idx * key_head_num_stride;
-          scalar_t* key_cache_start_ptr =
-              key_cache + block_idx * num_blocks_stride +
-              head_idx * cache_head_num_stride + block_offset;
-
-#pragma GCC unroll 8
-          for (int64_t i = 0, j = 0; i < head_dim; ++i, j += block_size) {
-            key_cache_start_ptr[j] = key_start_ptr[i];
-          }
-        }
-        {
-          // Write Value as row-major
-          const scalar_t* value_start_ptr = value +
-                                            token_idx * value_token_num_stride +
-                                            head_idx * value_head_num_stride;
-          scalar_t* value_cache_start_ptr =
-              value_cache + block_idx * num_blocks_stride +
-              head_idx * cache_head_num_stride + block_offset * head_dim;
-          std::memcpy(value_cache_start_ptr, value_start_ptr,
-                      sizeof(scalar_t) * head_dim);
-        }
-      }
-    }
-  }
-};
-}  // namespace cpu_attention
-
-#endif

csrc/cpu/cpu_attn_vec16.hpp

@@ -1,171 +0,0 @@
-#ifndef CPU_ATTN_VEC16_HPP
-#define CPU_ATTN_VEC16_HPP
-
-#include "cpu_attn_vec.hpp"
-
-namespace cpu_attention {
-
-namespace {
-// 16-1-16 pattern, 16 regs for A, 1 regs for B, 16 regs for C, [16, K] @ [k,
-// 16]
-template <typename kv_cache_t>
-class TileGemm161 {
- public:
-  template <AttentionGemmPhase phase, int32_t k_size>
-  FORCE_INLINE static void gemm(const int32_t m_size,
-                                float* __restrict__ a_tile,
-                                kv_cache_t* __restrict__ b_tile,
-                                float* __restrict__ c_tile, const int64_t lda,
-                                const int64_t ldb, const int64_t ldc,
-                                const int32_t block_size,
-                                const int32_t dynamic_k_size,
-                                const bool accum_c) {
-    switch (m_size) {
-      case 1:
-        gemm_micro<1>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 2:
-        gemm_micro<2>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 3:
-      case 4:
-        gemm_micro<4>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 5:
-      case 6:
-        gemm_micro<6>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 7:
-      case 8:
-        gemm_micro<8>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                      dynamic_k_size, accum_c);
-        break;
-      case 9:
-      case 10:
-      case 11:
-      case 12:
-        gemm_micro<12>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                       dynamic_k_size, accum_c);
-        break;
-      case 13:
-      case 14:
-      case 15:
-      case 16:
-        gemm_micro<16>(a_tile, b_tile, c_tile, lda, ldb, ldc, block_size,
-                       dynamic_k_size, accum_c);
-        break;
-    }
-  }
-
-  template <int32_t M>
-  static void gemm_micro(float* __restrict__ a_tile,
-                         kv_cache_t* __restrict__ b_tile,
-                         float* __restrict__ c_tile, const int64_t lda,
-                         const int64_t ldb, const int64_t ldc,
-                         const int32_t block_size, const int32_t dynamic_k_size,
-                         const bool accum_c) {
-    static_assert(0 < M <= 16);
-    using load_vec_t = typename VecTypeTrait<kv_cache_t>::vec_t;
-
-    kv_cache_t* __restrict__ curr_b_0 = b_tile;
-    float* __restrict__ curr_c_0 = c_tile;
-
-    vec_op::FP32Vec16 c_regs[M];
-    if (accum_c) {
-      float* __restrict__ curr_m_c_0 = curr_c_0;
-      vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        c_regs[i] = vec_op::FP32Vec16(curr_m_c_0);
-
-        // update
-        curr_m_c_0 += ldc;
-      });
-    }
-
-    float* __restrict__ curr_a = a_tile;
-    for (int32_t k = 0; k < dynamic_k_size; ++k) {
-      load_vec_t b_0_reg(curr_b_0);
-      vec_op::FP32Vec16 fp32_b_0_reg(b_0_reg);
-
-      float* __restrict__ curr_m_a = curr_a;
-      vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        float v = *curr_m_a;
-        vec_op::FP32Vec16 a_reg(v);
-        c_regs[i] = c_regs[i] + a_reg * fp32_b_0_reg;
-
-        // update
-        curr_m_a += lda;
-      });
-
-      // update
-      curr_a += 1;
-      curr_b_0 += ldb;
-    }
-
-    vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-      c_regs[i].save(curr_c_0);
-
-      // update
-      curr_c_0 += ldc;
-    });
-  }
-};
-}  // namespace
-
-// This is a general but naive implementation based on vector instructions
-template <typename scalar_t, int64_t head_dim>
-class AttentionImpl<ISA::VEC16, scalar_t, head_dim>
-    : public AttentionImpl<ISA::VEC, scalar_t, head_dim> {
- public:
-  using query_t = scalar_t;
-  using q_buffer_t = float;
-  using kv_cache_t = scalar_t;
-  using logits_buffer_t = float;
-  using partial_output_buffer_t = float;
-  using prob_buffer_t = float;
-
-  constexpr static int64_t BlockSizeAlignment =
-      16;  // KV token num unit of QK and PV phases
-  constexpr static int64_t HeadDimAlignment =
-      16;  // headdim num unit of PV phase
-  constexpr static int64_t MaxQHeadNumPerIteration = 16;
-  constexpr static int64_t HeadDim = head_dim;
-  constexpr static ISA ISAType = ISA::VEC16;
-  constexpr static bool scale_on_logits = false;  // apply scale on q_buffer
-
- public:
-  template <template <typename tile_gemm_t> typename attention>
-  FORCE_INLINE void execute_attention(DEFINE_CPU_ATTENTION_PARAMS) {
-    attention<TileGemm161<kv_cache_t>> attention_iteration;
-    attention_iteration(CPU_ATTENTION_PARAMS);
-  }
-
-  // k_cache_token_group_stride: stride of K cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t k_cache_token_group_stride(
-      const int32_t block_size) {
-    return BlockSizeAlignment;  // layout of k_cache block is [head_dim,
-                                // block_size], row-major
-  }
-
-  // v_cache_token_group_stride: stride of V cache when move to next
-  // BlockSizeAlignment tokens in a block
-  constexpr static int64_t v_cache_token_group_stride(
-      const int32_t block_size) {
-    return head_dim * BlockSizeAlignment;  // layout of v_cache is [block_size,
-                                           // head_dim], row-major
-  }
-
-  // v_cache_head_group_stride: stride of V cache when move to next
-  // HeadDimAlignment head dims in a block
-  constexpr static int64_t v_cache_head_group_stride(const int32_t block_size) {
-    return HeadDimAlignment;  // layout of v_cache is [block_size, head_dim],
-                              // row-major
-  }
-};
-}  // namespace cpu_attention
-
-#endif

csrc/cpu/cpu_types_scalar.hpp

@@ -26,6 +26,10 @@ namespace vec_op {
 
 #define FORCE_INLINE __attribute__((always_inline)) inline
 
+#define __max(a, b) ((a) > (b) ? (a) : (b))
+#define __min(a, b) ((a) < (b) ? (a) : (b))
+#define __abs(a) ((a) < (0) ? (0 - a) : (a))
+
 typedef struct f16x8_t {
   uint16_t val[8];
 } f16x8_t;
@@ -95,7 +99,7 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
   void save(void* ptr) const { *reinterpret_cast<f16x16_t*>(ptr) = reg; }
 
   void save(void* ptr, const int elem_num) const {
-    int num = std::min(elem_num, VEC_ELEM_NUM);
+    int num = __min(elem_num, VEC_ELEM_NUM);
     std::memcpy(ptr, &(reg.val[0]), num * sizeof(uint16_t));
   }
 };
@@ -124,7 +128,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
   void save(void* ptr) const { *reinterpret_cast<f16x16_t*>(ptr) = reg; }
 
   void save(void* ptr, const int elem_num) const {
-    int num = std::min(elem_num, VEC_ELEM_NUM);
+    int num = __min(elem_num, VEC_ELEM_NUM);
     std::memcpy(ptr, &(reg.val[0]), num * sizeof(uint16_t));
   }
 };
@@ -139,9 +143,9 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
   explicit BF16Vec32(f16x32_t data) : reg(data) {};
 
   explicit BF16Vec32(BF16Vec8& vec8_data) {
-    unroll_loop<int, VEC_ELEM_NUM>([&vec8_data, this](int i) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
       reg.val[i] = vec8_data.reg.val[i % BF16Vec8::VEC_ELEM_NUM];
-    });
+    }
   }
 
   void save(void* ptr) const { *reinterpret_cast<f16x32_t*>(ptr) = reg; }
@@ -153,11 +157,15 @@ struct FP32Vec4 : public Vec<FP32Vec4> {
   f32x4_t reg;
 
   explicit FP32Vec4(float v) {
-    unroll_loop<int, VEC_ELEM_NUM>([&v, this](int i) { reg.val[i] = v; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = v;
+    }
   }
 
   explicit FP32Vec4() {
-    unroll_loop<int, VEC_ELEM_NUM>([this](int i) { reg.val[i] = 0.0f; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = 0.0f;
+    }
   }
 
   explicit FP32Vec4(const float* ptr)
@@ -174,11 +182,15 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
   f32x8_t reg;
 
   explicit FP32Vec8(float v) {
-    unroll_loop<int, VEC_ELEM_NUM>([&v, this](int i) { reg.val[i] = v; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = v;
+    }
   }
 
   explicit FP32Vec8() {
-    unroll_loop<int, VEC_ELEM_NUM>([this](int i) { reg.val[i] = 0.0f; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = 0.0f;
+    }
   }
 
   explicit FP32Vec8(const float* ptr)
@@ -189,68 +201,78 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
   explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {};
 
   explicit FP32Vec8(const FP16Vec8& v) {
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&v, this](int i) { reg.val[i] = fp16_to_float(v.reg.val[i]); });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = fp16_to_float(v.reg.val[i]);
+    }
   }
 
   FP32Vec8(const BF16Vec8& v) {
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&v, this](int i) { reg.val[i] = bf16_to_float(v.reg.val[i]); });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = bf16_to_float(v.reg.val[i]);
+    }
   }
 
   float reduce_sum() const {
     float result = 0;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&result, this](int i) { result += reg.val[i]; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result += reg.val[i];
+    }
     return result;
   }
 
   FP32Vec8 exp() const {
     f32x8_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, this](int i) { ret.val[i] = expf(reg.val[i]); });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = expf(reg.val[i]);
+    }
     return FP32Vec8(ret);
   }
 
   FP32Vec8 tanh() const {
     f32x8_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, this](int i) { ret.val[i] = tanhf(reg.val[i]); });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = tanhf(reg.val[i]);
+    }
     return FP32Vec8(ret);
   }
 
   FP32Vec8 er() const {
     f32x8_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, this](int i) { ret.val[i] = erf(reg.val[i]); });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = erf(reg.val[i]);
+    }
     return FP32Vec8(ret);
   }
 
   FP32Vec8 operator*(const FP32Vec8& b) const {
     f32x8_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, &b, this](int i) { ret.val[i] = reg.val[i] * b.reg.val[i]; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = reg.val[i] * b.reg.val[i];
+    }
     return FP32Vec8(ret);
   }
 
   FP32Vec8 operator+(const FP32Vec8& b) const {
     f32x8_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, &b, this](int i) { ret.val[i] = reg.val[i] + b.reg.val[i]; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = reg.val[i] + b.reg.val[i];
+    }
     return FP32Vec8(ret);
   }
 
   FP32Vec8 operator-(const FP32Vec8& b) const {
     f32x8_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, &b, this](int i) { ret.val[i] = reg.val[i] - b.reg.val[i]; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = reg.val[i] - b.reg.val[i];
+    }
     return FP32Vec8(ret);
   }
 
   FP32Vec8 operator/(const FP32Vec8& b) const {
     f32x8_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, &b, this](int i) { ret.val[i] = reg.val[i] / b.reg.val[i]; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      ret.val[i] = reg.val[i] / b.reg.val[i];
+    }
     return FP32Vec8(ret);
   }
 
@@ -262,11 +284,15 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   f32x16_t reg;
 
   explicit FP32Vec16(float v) {
-    unroll_loop<int, VEC_ELEM_NUM>([&v, this](int i) { reg.val[i] = v; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = v;
+    }
   }
 
   explicit FP32Vec16() {
-    unroll_loop<int, VEC_ELEM_NUM>([this](int i) { reg.val[i] = 0.0f; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = 0.0f;
+    }
   }
 
   explicit FP32Vec16(const float* ptr)
@@ -275,27 +301,29 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   explicit FP32Vec16(f32x16_t data) : reg(data) {};
 
   FP32Vec16(const FP32Vec4& data) {
-    unroll_loop<int, VEC_ELEM_NUM>([&data, this](int i) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
       reg.val[i] = data.reg.val[i % FP32Vec4::VEC_ELEM_NUM];
-    });
+    }
   }
 
   FP32Vec16(const FP32Vec8& data) {
-    unroll_loop<int, VEC_ELEM_NUM>([&data, this](int i) {
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
       reg.val[i] = data.reg.val[i % FP32Vec8::VEC_ELEM_NUM];
-    });
+    }
   }
 
   FP32Vec16(const FP32Vec16& data) : reg(data.reg) {};
 
   explicit FP32Vec16(const FP16Vec16& v) {
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&v, this](int i) { reg.val[i] = fp16_to_float(v.reg.val[i]); });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = fp16_to_float(v.reg.val[i]);
+    }
   }
 
   explicit FP32Vec16(const BF16Vec16& v) {
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&v, this](int i) { reg.val[i] = bf16_to_float(v.reg.val[i]); });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      reg.val[i] = bf16_to_float(v.reg.val[i]);
+    }
   }
 
   explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
@@ -303,74 +331,82 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
 
   FP32Vec16 operator*(const FP32Vec16& b) const {
-    f32x16_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, &b, this](int i) { ret.val[i] = reg.val[i] * b.reg.val[i]; });
-    return FP32Vec16(ret);
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = reg.val[i] * b.reg.val[i];
+    }
+    return result;
   }
 
   FP32Vec16 operator+(const FP32Vec16& b) const {
-    f32x16_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, &b, this](int i) { ret.val[i] = reg.val[i] + b.reg.val[i]; });
-    return FP32Vec16(ret);
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = reg.val[i] + b.reg.val[i];
+    }
+    return result;
   }
 
   FP32Vec16 operator-(const FP32Vec16& b) const {
-    f32x16_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, &b, this](int i) { ret.val[i] = reg.val[i] - b.reg.val[i]; });
-    return FP32Vec16(ret);
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = reg.val[i] - b.reg.val[i];
+    }
+    return result;
   }
 
   FP32Vec16 operator/(const FP32Vec16& b) const {
-    f32x16_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, &b, this](int i) { ret.val[i] = reg.val[i] / b.reg.val[i]; });
-    return FP32Vec16(ret);
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = reg.val[i] / b.reg.val[i];
+    }
+    return result;
   }
 
   FP32Vec16 max(const FP32Vec16& b) const {
-    f32x16_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>([&ret, &b, this](int i) {
-      ret.val[i] = std::max(reg.val[i], b.reg.val[i]);
-    });
-    return FP32Vec16(ret);
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = __max(reg.val[i], b.reg.val[i]);
+    }
+    return result;
   }
 
   FP32Vec16 min(const FP32Vec16& b) const {
-    f32x16_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>([&ret, &b, this](int i) {
-      ret.val[i] = std::min(reg.val[i], b.reg.val[i]);
-    });
-    return FP32Vec16(ret);
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = __min(reg.val[i], b.reg.val[i]);
+    }
+    return result;
   }
 
   FP32Vec16 abs() const {
-    f32x16_t ret;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&ret, this](int i) { ret.val[i] = std::abs(reg.val[i]); });
-    return FP32Vec16(ret);
+    FP32Vec16 result(0.0f);
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result.reg.val[i] = __abs(reg.val[i]);
+    }
+    return result;
   }
 
   float reduce_sum() const {
     float result = 0.0f;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&result, this](int i) { result += reg.val[i]; });
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result += reg.val[i];
+    }
     return result;
   }
 
   float reduce_max() const {
-    float result = std::numeric_limits<float>::lowest();
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&result, this](int i) { result = std::max(reg.val[i], result); });
+    float result = reg.val[0];
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result = __max(reg.val[i], result);
+    }
     return result;
   }
 
   float reduce_min() const {
-    float result = std::numeric_limits<float>::max();
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&result, this](int i) { result = std::min(reg.val[i], result); });
+    float result = reg.val[0];
+    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
+      result = __min(reg.val[i], result);
+    }
     return result;
   }
 
@@ -378,9 +414,13 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   float reduce_sub_sum(int idx) {
     static_assert(VEC_ELEM_NUM % group_size == 0);
     float sum = 0.0;
-    const int start = idx * group_size;
-    unroll_loop<int, group_size>(
-        [&sum, &start, this](int i) { sum += reg.val[start + i]; });
+    int start = idx * group_size;
+    int end = (idx + 1) * group_size;
+
+    for (; (start < VEC_ELEM_NUM) && (start < end); ++start) {
+      sum += reg.val[start];
+    }
+
     return sum;
   }
 
@@ -437,13 +477,17 @@ inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
 }
 
 inline FP16Vec16::FP16Vec16(const FP32Vec16& v) {
-  unroll_loop<int, FP16Vec16::VEC_ELEM_NUM>(
-      [&v, this](int i) { reg.val[i] = float_to_fp16(v.reg.val[i]); });
+  int i = 0;
+  for (i = 0; i < FP16Vec16::VEC_ELEM_NUM; ++i) {
+    reg.val[i] = float_to_fp16(v.reg.val[i]);
+  }
 }
 
 inline FP16Vec8 ::FP16Vec8(const FP32Vec8& v) {
-  unroll_loop<int, FP16Vec8::VEC_ELEM_NUM>(
-      [&v, this](int i) { reg.val[i] = float_to_fp16(v.reg.val[i]); });
+  int i = 0;
+  for (i = 0; i < FP16Vec8::VEC_ELEM_NUM; ++i) {
+    reg.val[i] = float_to_fp16(v.reg.val[i]);
+  }
 }
 
 inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) {
@@ -451,13 +495,17 @@ inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) {
 }
 
 inline BF16Vec8::BF16Vec8(const FP32Vec8& v) {
-  unroll_loop<int, BF16Vec8::VEC_ELEM_NUM>(
-      [&v, this](int i) { reg.val[i] = float_to_bf16(v.reg.val[i]); });
+  int i = 0;
+  for (i = 0; i < BF16Vec8::VEC_ELEM_NUM; ++i) {
+    reg.val[i] = float_to_bf16(v.reg.val[i]);
+  }
 }
 
 inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
-  unroll_loop<int, BF16Vec16::VEC_ELEM_NUM>(
-      [&v, this](int i) { reg.val[i] = float_to_bf16(v.reg.val[i]); });
+  int i = 0;
+  for (i = 0; i < BF16Vec16::VEC_ELEM_NUM; ++i) {
+    reg.val[i] = float_to_bf16(v.reg.val[i]);
+  }
 }
 
 inline void prefetch(const void* addr) { __builtin_prefetch(addr, 0, 3); }

csrc/cpu/cpu_types_vxe.hpp

@@ -4,7 +4,6 @@
 
 #include <vecintrin.h>
 #include <cmath>
-#include <limits>
 #include <torch/all.h>
 namespace vec_op {
 
@@ -175,9 +174,8 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
   }
 
   explicit FP32Vec8(const BF16Vec8& v) {
-    // On big-endian s390x, place BF16 first to get correct byte order
-    reg.val[0] = (__vector float)vec_mergeh(v.reg, zero);
-    reg.val[1] = (__vector float)vec_mergel(v.reg, zero);
+    reg.val[0] = (__vector float)vec_mergeh(zero, v.reg);
+    reg.val[1] = (__vector float)vec_mergel(zero, v.reg);
   }
 
   float reduce_sum() const {
@@ -191,257 +189,51 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
   }
 
   FP32Vec8 exp() const {
-    f32x4x2_t out;
-
-    const __vector float log2e = vec_splats(1.44269504088896341f);
-    const __vector float one = vec_splats(1.0f);
-    const __vector float min_x = vec_splats(-87.3f);
-    const __vector float max_x = vec_splats(88.7f);
-
-    // 5th-degree minimax polynomial for 2^r (r in [0,1))
-    const __vector float c1 = vec_splats(0.6931471805599453f);
-    const __vector float c2 = vec_splats(0.240226506959101f);
-    const __vector float c3 = vec_splats(0.05550410866482158f);
-    const __vector float c4 = vec_splats(0.009618129107628477f);
-    const __vector float c5 = vec_splats(0.0013333558146428443f);
-
-    for (int i = 0; i < 2; i++) {
-      __vector float x = reg.val[i];
-
-      x = vec_max(x, min_x);
-      x = vec_min(x, max_x);
-
-      __vector float y = vec_mul(x, log2e);
-
-      __vector float kf = vec_floor(y);
-      __vector float r = vec_sub(y, kf);
-
-      __vector signed int k = vec_signed(kf);
-      const __vector signed int min_k = vec_splats((signed int)-126);
-      const __vector signed int max_k = vec_splats((signed int)127);
-      k = vec_min(vec_max(k, min_k), max_k);
-
-      // Build 2^k from exponent bits
-      __vector signed int exp_int = vec_add(k, vec_splats((signed int)127));
-      __vector unsigned int bits = (__vector unsigned int)exp_int;
-      bits = vec_sl(bits, vec_splats((unsigned int)23));
-      __vector float pow2k = (__vector float)bits;
-
-      // Improved minimax polynomial
-      __vector float poly = vec_madd(c5, r, c4);
-      poly = vec_madd(poly, r, c3);
-      poly = vec_madd(poly, r, c2);
-      poly = vec_madd(poly, r, c1);
-      poly = vec_madd(poly, r, one);
-
-      out.val[i] = vec_mul(pow2k, poly);
-    }
-
-    return FP32Vec8(out);
+    // TODO: Vectorize this
+    AliasReg ar;
+    ar.reg = reg;
+    f32x4x4_t ret;
+    ret.val[0][0] = std::exp(ar.values[0]);
+    ret.val[0][1] = std::exp(ar.values[1]);
+    ret.val[0][2] = std::exp(ar.values[2]);
+    ret.val[0][3] = std::exp(ar.values[3]);
+    ret.val[1][0] = std::exp(ar.values[4]);
+    ret.val[1][1] = std::exp(ar.values[5]);
+    ret.val[1][2] = std::exp(ar.values[6]);
+    ret.val[1][3] = std::exp(ar.values[7]);
+    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
   }
 
   FP32Vec8 tanh() const {
-    // tanh(x) = (exp(2x) - 1) / (exp(2x) + 1)
-    const __vector float one = vec_splats(1.0f);
-    const __vector float two = vec_splats(2.0f);
-    const __vector float zero = vec_splats(0.0f);
-    const __vector float sat =
-        vec_splats(9.0f);  // beyond this, tanh(x) ~ sign(x)
-
-    f32x4x2_t out;
-
-    for (int i = 0; i < 2; i++) {
-      __vector float x = reg.val[i];
-      __vector float ax = vec_abs(x);
-
-      // sign(x): +1 or -1
-      __vector float sign = vec_sel(vec_splats(-1.0f), one, vec_cmpgt(x, zero));
-
-      // saturation mask: |x| > sat
-      __vector __bool int saturated = vec_cmpgt(ax, sat);
-
-      // 2x
-      __vector float two_x = vec_mul(x, two);
-
-      // Build a temporary FP32Vec8 with both lanes = 2x, reuse exp()
-      f32x4x2_t tmp;
-      tmp.val[0] = two_x;
-      tmp.val[1] = two_x;
-      FP32Vec8 exp_2x_vec(tmp);
-
-      FP32Vec8 e2x = exp_2x_vec.exp();
-      __vector float e = e2x.reg.val[i];
-
-      // tanh(x) = (e - 1) / (e + 1)
-      __vector float num = vec_sub(e, one);
-      __vector float den = vec_add(e, one);
-
-      __vector float t = vec_div(num, den);
-
-      // For large |x|, clamp to sign(x)
-      out.val[i] = vec_sel(t, sign, saturated);
-    }
-
-    return FP32Vec8(out);
+    // TODO: Vectorize this
+    AliasReg ar;
+    ar.reg = reg;
+    f32x4x4_t ret;
+    ret.val[0][0] = std::tanh(ar.values[0]);
+    ret.val[0][1] = std::tanh(ar.values[1]);
+    ret.val[0][2] = std::tanh(ar.values[2]);
+    ret.val[0][3] = std::tanh(ar.values[3]);
+    ret.val[1][0] = std::tanh(ar.values[4]);
+    ret.val[1][1] = std::tanh(ar.values[5]);
+    ret.val[1][2] = std::tanh(ar.values[6]);
+    ret.val[1][3] = std::tanh(ar.values[7]);
+    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
   }
 
   FP32Vec8 er() const {
-    // A&S 7.1.26 approximation:
-    // erf(x) = sign(x) * (1 - ((((a5*t + a4)*t + a3)*t + a2)*t + a1) * t *
-    // exp(-x^2)) t = 1 / (1 + p*|x|),  p = 0.3275911
-
-    const __vector float one = vec_splats(1.0f);
-    const __vector float zero = vec_splats(0.0f);
-    const __vector float p = vec_splats(0.3275911f);
-
-    // Polynomial coeffs
-    const __vector float a1 = vec_splats(0.254829592f);
-    const __vector float a2 = vec_splats(-0.284496736f);
-    const __vector float a3 = vec_splats(1.421413741f);
-    const __vector float a4 = vec_splats(-1.453152027f);
-    const __vector float a5 = vec_splats(1.061405429f);
-
-    // Threshold where erf(x) ~ sign(x)
-    const __vector float sat = vec_splats(6.0f);
-
-    f32x4x2_t out;
-
-    for (int lane = 0; lane < 2; lane++) {
-      __vector float x = reg.val[lane];
-      __vector float ax = vec_abs(x);
-
-      // sign(x)
-      __vector float sign = vec_sel(vec_splats(-1.0f), one, vec_cmpgt(x, zero));
-
-      // |x| > 6 → erf(x) = ±1
-      __vector __bool int saturated = vec_cmpgt(ax, sat);
-
-      // t = 1 / (1 + p * |x|)
-      __vector float t = vec_madd(p, ax, one);
-      t = vec_div(one, t);
-
-      // poly = a5
-      __vector float poly = a5;
-      poly = vec_madd(poly, t, a4);
-      poly = vec_madd(poly, t, a3);
-      poly = vec_madd(poly, t, a2);
-      poly = vec_madd(poly, t, a1);
-
-      // full polynomial: poly = poly * t
-      poly = vec_mul(poly, t);
-
-      // Compute exp(-x^2)
-      __vector float x2 = vec_mul(x, x);
-      __vector float neg_x2 = vec_neg(x2);
-
-      f32x4x2_t tmp;
-      tmp.val[0] = neg_x2;
-      tmp.val[1] = neg_x2;
-      FP32Vec8 exp_neg_x2(tmp);
-
-      FP32Vec8 e = exp_neg_x2.exp();
-      __vector float ex = e.reg.val[lane];
-
-      // erf(x) = sign * (1 - poly * exp(-x^2))
-      __vector float term = vec_mul(poly, ex);
-      __vector float y = vec_sub(one, term);
-      y = vec_mul(y, sign);
-
-      // saturated → ±1
-      __vector float sat_val = vec_mul(sign, one);
-      out.val[lane] = vec_sel(y, sat_val, saturated);
-    }
-
-    return FP32Vec8(out);
-  }
-  // Elementwise sigmoid(x) = 1 / (1 + exp(-x))
-  FP32Vec8 sigmoid() const {
-    const __vector float one = vec_splats(1.0f);
-
-    f32x4x2_t neg;
-    for (int i = 0; i < 2; ++i) {
-      neg.val[i] = vec_neg(reg.val[i]);
-    }
-
-    FP32Vec8 neg_x(neg);
-    FP32Vec8 e = neg_x.exp();  // exp(-x)
-
-    f32x4x2_t denom;
-    for (int i = 0; i < 2; ++i) {
-      denom.val[i] = vec_add(one, e.reg.val[i]);
-    }
-
-    FP32Vec8 denom_vec(denom);
-    FP32Vec8 one_vec(1.0f);
-
-    return one_vec / denom_vec;
-  }
-
-  // Tanh-based GELU:
-  // gelu(x) = 0.5 * x * (1 + tanh(√(2/π) * (x + 0.044715 * x^3)))
-  FP32Vec8 gelu_tanh() const {
-    const __vector float k_s2pi = vec_splats(0.7978845608028654f);  // √(2/π)
-    const __vector float k_0_0447 = vec_splats(0.044715f);
-
-    f32x4x2_t x2, x3, inner;
-    for (int i = 0; i < 2; ++i) {
-      __vector float x = reg.val[i];
-      x2.val[i] = vec_mul(x, x);                            // x^2
-      x3.val[i] = vec_mul(x2.val[i], x);                    // x^3
-      __vector float t = vec_madd(k_0_0447, x3.val[i], x);  // x + 0.044715*x^3
-      inner.val[i] = vec_mul(k_s2pi, t);                    // √(2/π)*(...)
-    }
-
-    FP32Vec8 inner_vec(inner);
-    FP32Vec8 t = inner_vec.tanh();  // tanh part
-
-    FP32Vec8 one_vec(1.0f);
-    FP32Vec8 half_vec(0.5f);
-
-    FP32Vec8 x_vec(*this);
-    return x_vec * half_vec * (one_vec + t);
-  }
-
-  // Erf-based GELU:
-  // gelu(x) = 0.5 * x * (1 + erf(x / √2))
-  FP32Vec8 gelu_erf() const {
-    const __vector float inv_sqrt2 = vec_splats(0.7071067811865476f);  // 1/√2
-    FP32Vec8 x_vec(*this);
-
-    f32x4x2_t scaled;
-    for (int i = 0; i < 2; ++i) {
-      scaled.val[i] = vec_mul(reg.val[i], inv_sqrt2);
-    }
-    FP32Vec8 x_scaled(scaled);
-
-    FP32Vec8 erf_x = x_scaled.er();
-
-    FP32Vec8 one_vec(1.0f);
-    FP32Vec8 half_vec(0.5f);
-
-    return x_vec * half_vec * (one_vec + erf_x);
-  }
-
-  // Elementwise reciprocal: 1/x (scalar per lane, for correctness)
-  FP32Vec8 rcp() const {
-    AliasReg in, out;
-    in.reg = reg;
-
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      out.values[i] = 1.0f / in.values[i];
-    }
-    return FP32Vec8(out.reg);
-  }
-
-  // Elementwise rsqrt(x) = 1 / sqrt(x) (scalar per lane, for correctness)
-  FP32Vec8 rsqrt() const {
-    AliasReg in, out;
-    in.reg = reg;
-
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      out.values[i] = 1.0f / std::sqrt(in.values[i]);
-    }
-    return FP32Vec8(out.reg);
+    // TODO: Vectorize this
+    AliasReg ar;
+    ar.reg = reg;
+    f32x4x4_t ret;
+    ret.val[0][0] = std::erf(ar.values[0]);
+    ret.val[0][1] = std::erf(ar.values[1]);
+    ret.val[0][2] = std::erf(ar.values[2]);
+    ret.val[0][3] = std::erf(ar.values[3]);
+    ret.val[1][0] = std::erf(ar.values[4]);
+    ret.val[1][1] = std::erf(ar.values[5]);
+    ret.val[1][2] = std::erf(ar.values[6]);
+    ret.val[1][3] = std::erf(ar.values[7]);
+    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
   }
 
   FP32Vec8 operator*(const FP32Vec8& b) const {
@@ -524,11 +316,10 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   }
 
   explicit FP32Vec16(const BF16Vec16& v) {
-    // On big-endian s390x, place BF16 first to get correct byte order
-    reg.val[0] = (__vector float)vec_mergeh(v.reg.val[0], zero);
-    reg.val[1] = (__vector float)vec_mergel(v.reg.val[0], zero);
-    reg.val[2] = (__vector float)vec_mergeh(v.reg.val[1], zero);
-    reg.val[3] = (__vector float)vec_mergel(v.reg.val[1], zero);
+    reg.val[0] = (__vector float)vec_mergeh(zero, v.reg.val[0]);
+    reg.val[1] = (__vector float)vec_mergel(zero, v.reg.val[0]);
+    reg.val[2] = (__vector float)vec_mergeh(zero, v.reg.val[1]);
+    reg.val[3] = (__vector float)vec_mergel(zero, v.reg.val[1]);
   }
 
   explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
@@ -585,23 +376,6 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return result;
   }
 
-  FP32Vec16 max(const FP32Vec16& b) const {
-    return FP32Vec16(f32x4x4_t({vec_max(reg.val[0], b.reg.val[0]),
-                                vec_max(reg.val[1], b.reg.val[1]),
-                                vec_max(reg.val[2], b.reg.val[2]),
-                                vec_max(reg.val[3], b.reg.val[3])}));
-  }
-
-  float reduce_max() const {
-    AliasReg ar;
-    ar.reg = reg;
-    float result = ar.values[0];
-    unroll_loop<int, VEC_ELEM_NUM>([&result, &ar](int i) {
-      if (ar.values[i] > result) result = ar.values[i];
-    });
-    return result;
-  }
-
   void save(float* ptr) const {
     vec_xst(reg.val[0], 0, ptr);
     vec_xst(reg.val[1], 16, ptr);
@@ -628,14 +402,15 @@ struct VecType<c10::BFloat16> {
   using vec_type = BF16Vec8;
 };
 
-// On s390x, FP16 (Half) is not natively supported, use FP32 vectors instead
-using FP16Vec16 = FP32Vec16;
-
 template <typename T>
 void storeFP32(float v, T* ptr) {
   *ptr = v;
 }
 
+inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) {
+  acc = acc + a * b;
+}
+
 namespace c10 {
 struct BFloat16 {
   uint16_t value;  // Assume BFloat16 is defined as a struct containing a 16-bit
@@ -654,79 +429,6 @@ inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
   #define __VEC_CLASS_FP_NAN (1 << 6)
 #endif
 
-// Optimized FMA (Fused Multiply-Add) implementations using IBM Z vector
-// intrinsics
-
-// FP32Vec4 FMA: acc = acc + (a * b) or equivalently acc = fma(a, b, acc)
-FORCE_INLINE void fma(FP32Vec4& acc, const FP32Vec4& a, const FP32Vec4& b) {
-  acc.reg = vec_madd(a.reg, b.reg, acc.reg);
-}
-
-// FP32Vec8 FMA: acc = acc + (a * b)
-FORCE_INLINE void fma(FP32Vec8& acc, const FP32Vec8& a, const FP32Vec8& b) {
-  acc.reg.val[0] = vec_madd(a.reg.val[0], b.reg.val[0], acc.reg.val[0]);
-  acc.reg.val[1] = vec_madd(a.reg.val[1], b.reg.val[1], acc.reg.val[1]);
-}
-
-// FP32Vec16 FMA: acc = acc + (a * b)
-FORCE_INLINE void fma(FP32Vec16& acc, const FP32Vec16& a, const FP32Vec16& b) {
-  acc.reg.val[0] = vec_madd(a.reg.val[0], b.reg.val[0], acc.reg.val[0]);
-  acc.reg.val[1] = vec_madd(a.reg.val[1], b.reg.val[1], acc.reg.val[1]);
-  acc.reg.val[2] = vec_madd(a.reg.val[2], b.reg.val[2], acc.reg.val[2]);
-  acc.reg.val[3] = vec_madd(a.reg.val[3], b.reg.val[3], acc.reg.val[3]);
-}
-
-// Multiply-Subtract: acc = acc - (a * b)
-FORCE_INLINE void fms(FP32Vec4& acc, const FP32Vec4& a, const FP32Vec4& b) {
-  acc.reg = vec_msub(a.reg, b.reg, acc.reg);
-}
-
-FORCE_INLINE void fms(FP32Vec8& acc, const FP32Vec8& a, const FP32Vec8& b) {
-  acc.reg.val[0] = vec_msub(a.reg.val[0], b.reg.val[0], acc.reg.val[0]);
-  acc.reg.val[1] = vec_msub(a.reg.val[1], b.reg.val[1], acc.reg.val[1]);
-}
-
-FORCE_INLINE void fms(FP32Vec16& acc, const FP32Vec16& a, const FP32Vec16& b) {
-  acc.reg.val[0] = vec_msub(a.reg.val[0], b.reg.val[0], acc.reg.val[0]);
-  acc.reg.val[1] = vec_msub(a.reg.val[1], b.reg.val[1], acc.reg.val[1]);
-  acc.reg.val[2] = vec_msub(a.reg.val[2], b.reg.val[2], acc.reg.val[2]);
-  acc.reg.val[3] = vec_msub(a.reg.val[3], b.reg.val[3], acc.reg.val[3]);
-}
-
-// Negative Multiply-Add: acc = -(a * b) + acc
-FORCE_INLINE void nfma(FP32Vec4& acc, const FP32Vec4& a, const FP32Vec4& b) {
-  acc.reg = vec_nmadd(a.reg, b.reg, acc.reg);
-}
-
-FORCE_INLINE void nfma(FP32Vec8& acc, const FP32Vec8& a, const FP32Vec8& b) {
-  acc.reg.val[0] = vec_nmadd(a.reg.val[0], b.reg.val[0], acc.reg.val[0]);
-  acc.reg.val[1] = vec_nmadd(a.reg.val[1], b.reg.val[1], acc.reg.val[1]);
-}
-
-FORCE_INLINE void nfma(FP32Vec16& acc, const FP32Vec16& a, const FP32Vec16& b) {
-  acc.reg.val[0] = vec_nmadd(a.reg.val[0], b.reg.val[0], acc.reg.val[0]);
-  acc.reg.val[1] = vec_nmadd(a.reg.val[1], b.reg.val[1], acc.reg.val[1]);
-  acc.reg.val[2] = vec_nmadd(a.reg.val[2], b.reg.val[2], acc.reg.val[2]);
-  acc.reg.val[3] = vec_nmadd(a.reg.val[3], b.reg.val[3], acc.reg.val[3]);
-}
-
-// Negative Multiply-Subtract: acc = -(a * b) - acc
-FORCE_INLINE void nfms(FP32Vec4& acc, const FP32Vec4& a, const FP32Vec4& b) {
-  acc.reg = vec_nmsub(a.reg, b.reg, acc.reg);
-}
-
-FORCE_INLINE void nfms(FP32Vec8& acc, const FP32Vec8& a, const FP32Vec8& b) {
-  acc.reg.val[0] = vec_nmsub(a.reg.val[0], b.reg.val[0], acc.reg.val[0]);
-  acc.reg.val[1] = vec_nmsub(a.reg.val[1], b.reg.val[1], acc.reg.val[1]);
-}
-
-FORCE_INLINE void nfms(FP32Vec16& acc, const FP32Vec16& a, const FP32Vec16& b) {
-  acc.reg.val[0] = vec_nmsub(a.reg.val[0], b.reg.val[0], acc.reg.val[0]);
-  acc.reg.val[1] = vec_nmsub(a.reg.val[1], b.reg.val[1], acc.reg.val[1]);
-  acc.reg.val[2] = vec_nmsub(a.reg.val[2], b.reg.val[2], acc.reg.val[2]);
-  acc.reg.val[3] = vec_nmsub(a.reg.val[3], b.reg.val[3], acc.reg.val[3]);
-}
-
 const static __vector unsigned char omask = {2,  3,  6,  7,  10, 11, 14, 15,
                                              18, 19, 22, 23, 26, 27, 30, 31};
 const static __vector unsigned int bias = {0x00007fff, 0x00007fff, 0x00007fff,
@@ -739,24 +441,13 @@ const static __vector unsigned int one = {1, 1, 1, 1};
 inline BF16Vec8::BF16Vec8(const FP32Vec8& v) {
   __vector unsigned int inp0 = (__vector unsigned int)(v.reg.val[0]);
   __vector unsigned int inp1 = (__vector unsigned int)(v.reg.val[1]);
-  __vector unsigned int lsb0 = inp0 >> sh16;
-  __vector unsigned int lsb1 = inp1 >> sh16;
-  lsb0 = lsb0 & one;
-  lsb1 = lsb1 & one;
-  __vector unsigned int rnd0 = lsb0 + bias;
-  __vector unsigned int rnd1 = lsb1 + bias;
-  inp0 = inp0 + rnd0;
-  inp1 = inp1 + rnd1;
   int cc;
   __vector __bool int sel0 =
       vec_fp_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN, &cc);
   __vector __bool int sel1 =
       vec_fp_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN, &cc);
-  inp0 = vec_sel(inp0, nan, sel0);
-  inp1 = vec_sel(inp1, nan, sel1);
-  inp0 = inp0 >> sh16;
-  inp1 = inp1 >> sh16;
-
+  inp0 = vec_sel(inp0, nan, sel0) >> sh16;
+  inp1 = vec_sel(inp1, nan, sel1) >> sh16;
   reg = (__vector signed short)vec_perm(inp0, inp1, omask);
 }
 
@@ -765,22 +456,6 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
   __vector unsigned int inp1 = (__vector unsigned int)(v.reg.val[1]);
   __vector unsigned int inp2 = (__vector unsigned int)(v.reg.val[2]);
   __vector unsigned int inp3 = (__vector unsigned int)(v.reg.val[3]);
-  __vector unsigned int lsb0 = inp0 >> sh16;
-  __vector unsigned int lsb1 = inp1 >> sh16;
-  __vector unsigned int lsb2 = inp2 >> sh16;
-  __vector unsigned int lsb3 = inp3 >> sh16;
-  lsb0 = lsb0 & one;
-  lsb1 = lsb1 & one;
-  lsb2 = lsb2 & one;
-  lsb3 = lsb3 & one;
-  __vector unsigned int rnd0 = lsb0 + bias;
-  __vector unsigned int rnd1 = lsb1 + bias;
-  __vector unsigned int rnd2 = lsb2 + bias;
-  __vector unsigned int rnd3 = lsb3 + bias;
-  inp0 = inp0 + rnd0;
-  inp1 = inp1 + rnd1;
-  inp2 = inp2 + rnd2;
-  inp3 = inp3 + rnd3;
   int cc;
   __vector __bool int sel0 =
       vec_fp_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN, &cc);
@@ -790,164 +465,15 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
       vec_fp_test_data_class(v.reg.val[2], __VEC_CLASS_FP_NAN, &cc);
   __vector __bool int sel3 =
       vec_fp_test_data_class(v.reg.val[3], __VEC_CLASS_FP_NAN, &cc);
-  inp0 = vec_sel(inp0, nan, sel0);
-  inp1 = vec_sel(inp1, nan, sel1);
-  inp2 = vec_sel(inp2, nan, sel2);
-  inp3 = vec_sel(inp3, nan, sel3);
-  inp0 = inp0 >> sh16;
-  inp1 = inp1 >> sh16;
-  inp2 = inp2 >> sh16;
-  inp3 = inp3 >> sh16;
-
+  inp0 = vec_sel(inp0, nan, sel0) >> sh16;
+  inp1 = vec_sel(inp1, nan, sel1) >> sh16;
+  inp2 = vec_sel(inp2, nan, sel2) >> sh16;
+  inp3 = vec_sel(inp3, nan, sel3) >> sh16;
   reg.val[0] = (__vector signed short)vec_perm(inp0, inp1, omask);
   reg.val[1] = (__vector signed short)vec_perm(inp2, inp3, omask);
 }
 
-// 1D softmax over `n` elements in `input`, writes result to `output`.
-// Uses FP32Vec8 for main body, scalar tail handling.
-// Requirement: n > 0
-FORCE_INLINE void softmax_fp32vec8(float* output, const float* input, int n) {
-  if (n <= 0) return;
-
-  // ---------- Pass 1: find max ----------
-  float max_val = -std::numeric_limits<float>::infinity();
-  int i = 0;
-
-  for (; i + FP32Vec8::VEC_ELEM_NUM <= n; i += FP32Vec8::VEC_ELEM_NUM) {
-    FP32Vec8 v(input + i);
-    FP32Vec8::AliasReg ar;
-    ar.reg = v.reg;
-    for (int j = 0; j < FP32Vec8::VEC_ELEM_NUM; ++j) {
-      if (ar.values[j] > max_val) max_val = ar.values[j];
-    }
-  }
-  for (; i < n; ++i) {
-    if (input[i] > max_val) max_val = input[i];
-  }
-
-  // ---------- Pass 2: compute exp(x - max) and sum ----------
-  float sum = 0.0f;
-  i = 0;
-
-  for (; i + FP32Vec8::VEC_ELEM_NUM <= n; i += FP32Vec8::VEC_ELEM_NUM) {
-    float tmp[FP32Vec8::VEC_ELEM_NUM];
-    for (int j = 0; j < FP32Vec8::VEC_ELEM_NUM; ++j) {
-      tmp[j] = input[i + j] - max_val;
-    }
-
-    FP32Vec8 v(tmp);
-    FP32Vec8 e = v.exp();
-
-    FP32Vec8::AliasReg ar;
-    ar.reg = e.reg;
-    for (int j = 0; j < FP32Vec8::VEC_ELEM_NUM; ++j) {
-      output[i + j] = ar.values[j];
-      sum += ar.values[j];
-    }
-  }
-
-  // Tail
-  for (; i < n; ++i) {
-    float x = input[i] - max_val;
-    float ex = std::exp(x);  // scalar tail
-    output[i] = ex;
-    sum += ex;
-  }
-
-  // ---------- Pass 3: normalize ----------
-  float inv_sum = 1.0f / sum;
-  i = 0;
-
-  for (; i + FP32Vec8::VEC_ELEM_NUM <= n; i += FP32Vec8::VEC_ELEM_NUM) {
-    float tmp[FP32Vec8::VEC_ELEM_NUM];
-    for (int j = 0; j < FP32Vec8::VEC_ELEM_NUM; ++j) {
-      tmp[j] = output[i + j] * inv_sum;
-    }
-    FP32Vec8 v(tmp);
-    v.save(output + i);
-  }
-
-  for (; i < n; ++i) {
-    output[i] *= inv_sum;
-  }
-}
-
-// 1D RMSNorm kernel:
-//   input:  x[0..n-1]
-//   weight: w[0..n-1] (gamma), may be nullptr
-//   output: y[i] = x[i] * inv_rms * (weight[i] if weight != nullptr else 1)
-//   eps: small epsilon for numerical stability
-FORCE_INLINE void rmsnorm_fp32vec8(float* output, const float* input,
-                                   const float* weight, int n, float eps) {
-  if (n <= 0) return;
-
-  // ---------- Pass 1: compute sum of squares ----------
-  float sum_sq = 0.0f;
-  int i = 0;
-
-  for (; i + FP32Vec8::VEC_ELEM_NUM <= n; i += FP32Vec8::VEC_ELEM_NUM) {
-    FP32Vec8 x_vec(input + i);
-
-    FP32Vec8 sq = x_vec * x_vec;
-
-    FP32Vec8::AliasReg ar;
-    ar.reg = sq.reg;
-    for (int j = 0; j < FP32Vec8::VEC_ELEM_NUM; ++j) {
-      sum_sq += ar.values[j];
-    }
-  }
-
-  // Tail
-  for (; i < n; ++i) {
-    float v = input[i];
-    sum_sq += v * v;
-  }
-
-  float mean_sq = sum_sq / static_cast<float>(n);
-  float inv_rms = 1.0f / std::sqrt(mean_sq + eps);
-
-  // ---------- Pass 2: scale (and apply weight if given) ----------
-  const float inv_rms_f = inv_rms;
-  i = 0;
-
-  if (weight) {
-    // with gamma
-    for (; i + FP32Vec8::VEC_ELEM_NUM <= n; i += FP32Vec8::VEC_ELEM_NUM) {
-      FP32Vec8 x_vec(input + i);
-
-      float wtmp[FP32Vec8::VEC_ELEM_NUM];
-      for (int j = 0; j < FP32Vec8::VEC_ELEM_NUM; ++j) {
-        wtmp[j] = weight[i + j];
-      }
-      FP32Vec8 w_vec(wtmp);
-
-      FP32Vec8 scale_vec(inv_rms_f);
-      FP32Vec8 y = x_vec * scale_vec * w_vec;
-      y.save(output + i);
-    }
-
-    for (; i < n; ++i) {
-      output[i] = input[i] * inv_rms_f * weight[i];
-    }
-  } else {
-    // without gamma
-    for (; i + FP32Vec8::VEC_ELEM_NUM <= n; i += FP32Vec8::VEC_ELEM_NUM) {
-      FP32Vec8 x_vec(input + i);
-      FP32Vec8 scale_vec(inv_rms_f);
-      FP32Vec8 y = x_vec * scale_vec;
-      y.save(output + i);
-    }
-
-    for (; i < n; ++i) {
-      output[i] = input[i] * inv_rms_f;
-    }
-  }
-}
-
-// Prefetch data to cache for better memory access performance
-FORCE_INLINE void prefetch(const void* addr) {
-  __builtin_prefetch(addr, 0, 3);  // 0=read, 3=high temporal locality
-}
+inline void prefetch(const void* addr) { void __dcbt(const void* addr); }
 
 };  // namespace vec_op
 

csrc/cpu/cpu_types_x86.hpp

@@ -40,23 +40,6 @@ namespace vec_op {
 
 #define FORCE_INLINE __attribute__((always_inline)) inline
 
-// Function to get the timestamp using RDTSCP
-FORCE_INLINE uint64_t bench_timestamp() {
-  unsigned int cycles_low, cycles_high;
-  asm volatile(
-      ".intel_syntax noprefix\n\t"
-      "CPUID\n\t"        // Serialize instruction stream to ensure previous
-                         // instructions complete
-      "RDTSCP\n\t"       // Read TSC and core ID
-      "mov %0, edx\n\t"  // Store high 32 bits of TSC
-      "mov %1, eax\n\t"  // Store low 32 bits of TSC
-      ".att_syntax"
-      : "=r"(cycles_high), "=r"(cycles_low)::"rax", "rbx", "rcx",
-        "rdx"  // Clobbered registers
-  );
-  return (uint64_t)cycles_high << 32 | cycles_low;
-}
-
 namespace {
 template <typename T, T... indexes, typename F>
 constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
@@ -104,8 +87,6 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
   explicit FP16Vec16(bool, void* ptr)
       : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
 
-  explicit FP16Vec16(const c10::Half v) : reg(_mm256_set1_epi16(v.x)) {}
-
   explicit FP16Vec16(const FP32Vec16&);
 
   void save(void* ptr) const { _mm256_storeu_si256((__m256i*)ptr, reg); }
@@ -143,8 +124,6 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
   explicit BF16Vec16(bool, void* ptr)
       : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
 
-  explicit BF16Vec16(const c10::BFloat16 v) : reg(_mm256_set1_epi16(v.x)) {}
-
   explicit BF16Vec16(const FP32Vec16&);
 
   void save(void* ptr) const { _mm256_storeu_si256((__m256i*)ptr, reg); }
@@ -354,22 +333,6 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   explicit FP32Vec16(__m512 data) : reg(data) {}
 
-  // de-pack 4 bit values
-  explicit FP32Vec16(int64_t value, const FP32Vec16& lut) {
-    int64_t mask_0 = 0x0F0F0F0F0F0F0F0F;
-    int64_t mask_1 = 0xF0F0F0F0F0F0F0F0;
-    int64_t value_0 = value & mask_0;
-    int64_t value_1 = value & mask_1;
-    __m128i vec_0 = _mm_movpi64_epi64((__m64)value_0);
-    __m128i vec_1 = _mm_movpi64_epi64((__m64)value_1);
-    vec_0 = _mm_cvtepu8_epi16(vec_0);
-    vec_1 = _mm_cvtepu8_epi16(vec_1);
-    vec_1 = _mm_slli_epi16(vec_1, 4);
-    __m128i vec = _mm_or_si128(vec_0, vec_1);
-    __m512i vec_i32 = _mm512_cvtepu8_epi32(vec);
-    reg = _mm512_permutexvar_ps(vec_i32, lut.reg);
-  }
-
   explicit FP32Vec16(const FP32Vec4& data)
       : reg((__m512)_mm512_inserti32x4(
             _mm512_inserti32x4(
@@ -444,7 +407,13 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   float reduce_min() const { return _mm512_reduce_min_ps(reg); }
 
-  float get_last_elem() const { return _mm512_cvtss_f32(reg); }
+  template <int group_size>
+  float reduce_sub_sum(int idx) {
+    static_assert(VEC_ELEM_NUM % group_size == 0);
+    constexpr uint32_t base_mask = (0xFFFF >> (16 - group_size));
+    __mmask16 mask = _cvtu32_mask16(base_mask << (idx * group_size));
+    return _mm512_mask_reduce_add_ps(mask, reg);
+  }
 
   void save(float* ptr) const { _mm512_storeu_ps(ptr, reg); }
 
@@ -477,6 +446,9 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   explicit FP32Vec16(__m256 low, __m256 high) : reg_low(low), reg_high(high) {}
 
+  explicit FP32Vec16(const FP32Vec16& data)
+      : reg_low(data.reg_low), reg_high(data.reg_high) {}
+
   explicit FP32Vec16(const FP32Vec4& data)
       : reg_low((__m256)_mm256_inserti128_si256(
             _mm256_castsi128_si256((__m128i)data.reg), (__m128i)data.reg, 1)),
@@ -532,32 +504,6 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
                      _mm256_div_ps(reg_high, b.reg_high));
   }
 
-  FP32Vec16 max(const FP32Vec16& b) const {
-    return FP32Vec16(_mm256_max_ps(reg_low, b.reg_low),
-                     _mm256_max_ps(reg_high, b.reg_high));
-  }
-
-  float reduce_max() const {
-    __m256 v = _mm256_max_ps(reg_low, reg_high);
-    // Permute to compare elements within 128-bit lanes
-    __m256 v_shuffled = _mm256_permute_ps(
-        v, 0b00001011);  // Swap halves within each 128-bit lane
-    __m256 v_max = _mm256_max_ps(v, v_shuffled);
-
-    v_shuffled = _mm256_permute_ps(
-        v_max, 0b00000001);  // Shuffle elements within each 128-bit lane
-    v_max = _mm256_max_ps(v_max, v_shuffled);
-
-    // Permute to compare elements between 128-bit lanes
-    v_shuffled =
-        _mm256_permute2f128_ps(v_max, v_max, 0b00000001);  // Swap 128-bit lanes
-    v_max = _mm256_max_ps(v_max, v_shuffled);
-
-    // At this point, the maximum value is present in all elements of v_max.
-    // Extract the first element for the scalar result.
-    return _mm256_cvtss_f32(v_max);  // Extract the lowest 32-bit float
-  }
-
   float reduce_sum() const {
     FP32Vec8 low = FP32Vec8(reg_low);
     FP32Vec8 high = FP32Vec8(reg_high);
@@ -696,7 +642,7 @@ inline FP16Vec16::FP16Vec16(const FP32Vec16& v)
 inline FP16Vec16::FP16Vec16(const FP32Vec16& v)
     : reg(_mm256_insertf128_si256(
           _mm256_castsi128_si256(FP16Vec8(FP32Vec8(v.reg_low)).reg),
-          FP16Vec8(FP32Vec8(v.reg_high)).reg, 1)) {}
+          FP16Vec8(FP32Vec8(v.reg_low)).reg, 1)) {}
 #endif
 
 #ifdef __AVX512BF16__
@@ -767,25 +713,6 @@ inline void non_temporal_save(BF16Vec16& vec, void* ptr) {
 inline void non_temporal_save(FP32Vec16& vec, void* ptr) {
   _mm512_stream_ps((float*)ptr, vec.reg);
 }
-
-static void interleave_save(const BF16Vec16& vec0, const BF16Vec16& vec1,
-                            void* ptr) {
-  __m512i vec_0 = _mm512_cvtepu16_epi32(vec0.reg);
-  __m512i vec_1 = _mm512_cvtepu16_epi32(vec1.reg);
-  vec_1 = _mm512_slli_epi32(vec_1, 16);
-  vec_0 = _mm512_or_si512(vec_0, vec_1);
-  _mm512_storeu_epi32(ptr, vec_0);
-}
-
-static void interleave_save(const FP16Vec16& vec0, const FP16Vec16& vec1,
-                            void* ptr) {
-  __m512i vec_0 = _mm512_cvtepu16_epi32(vec0.reg);
-  __m512i vec_1 = _mm512_cvtepu16_epi32(vec1.reg);
-  vec_1 = _mm512_slli_epi32(vec_1, 16);
-  vec_0 = _mm512_or_si512(vec_0, vec_1);
-  _mm512_storeu_epi32(ptr, vec_0);
-}
-
 #endif
 
 inline void mem_barrier() { _mm_mfence(); }

csrc/cpu/cpu_wna16.cpp

@@ -1,402 +0,0 @@
-#include "cpu_types.hpp"
-#include "scratchpad_manager.h"
-#include "utils.hpp"
-
-#ifdef CPU_CAPABILITY_AMXBF16
-  #include "cpu/micro_gemm/cpu_micro_gemm_amx.hpp"
-#endif
-#include "cpu/micro_gemm/cpu_micro_gemm_vec.hpp"
-
-#define VLLM_DISPATCH_CASE_16B_TYPES(...)                 \
-  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
-  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
-
-#define VLLM_DISPATCH_16B_TYPES(TYPE, NAME, ...) \
-  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_16B_TYPES(__VA_ARGS__))
-
-template <typename T>
-void print_logits(const char* name, T* ptr, int32_t row, int32_t col,
-                  int32_t stride) {
-  std::stringstream ss;
-  ss << std::fixed << std::setprecision(5) << name << ": [\n";
-  auto* curr_logits_buffer = ptr;
-  for (int32_t m = 0; m < row; ++m) {
-    for (int32_t n = 0; n < col; ++n) {
-      ss << curr_logits_buffer[n] << ", ";
-    }
-    ss << "\n";
-    curr_logits_buffer += stride;
-  }
-  ss << "]\n";
-  std::printf("%s", ss.str().c_str());
-}
-
-namespace {
-using cpu_utils::ISA;
-using cpu_utils::VecTypeTrait;
-
-template <typename scalar_t, ISA isa, bool has_zp, bool use_desc_act>
-class Dequantizer4b {
- public:
-  constexpr static int32_t pack_num = 32 / 4;
-  using scalar_vec_t = typename VecTypeTrait<scalar_t>::vec_t;
-
- public:
-  static void dequant(int32_t* __restrict__ q_weight,
-                      scalar_t* __restrict__ weight,
-                      scalar_t* __restrict__ scales,
-                      int32_t* __restrict__ zeros, int32_t* __restrict__ g_idx,
-                      const int64_t scales_stride, const int64_t zeros_stride,
-                      const int32_t k_size, const int32_t group_size) {
-    vec_op::FP32Vec16 lut;
-    if constexpr (has_zp) {
-      // AWQ
-      alignas(64) static const float LUT[16] = {
-          0.0f, 1.0f, 2.0f,  3.0f,  4.0f,  5.0f,  6.0f,  7.0f,
-          8.0f, 9.0f, 10.0f, 11.0f, 12.0f, 13.0f, 14.0f, 15.0f};
-      lut = vec_op::FP32Vec16(LUT);
-    } else {
-      // GPTQ
-      alignas(64) static const float LUT[16] = {
-          -8.0f, -7.0f, -6.0f, -5.0f, -4.0f, -3.0f, -2.0f, -1.0f,
-          0.0f,  1.0f,  2.0f,  3.0f,  4.0f,  5.0f,  6.0f,  7.0f};
-      lut = vec_op::FP32Vec16(LUT);
-    }
-
-    // per 64-bits elem contains 16 output channels
-    int64_t* __restrict__ curr_q_weight = reinterpret_cast<int64_t*>(q_weight);
-    int64_t* __restrict__ curr_zeros = reinterpret_cast<int64_t*>(zeros);
-    scalar_t* __restrict__ curr_weight = weight;
-    scalar_t* __restrict__ curr_scale = scales;
-    vec_op::FP32Vec16 scale_0;
-    vec_op::FP32Vec16 scale_1;
-    vec_op::FP32Vec16 zero_0;
-    vec_op::FP32Vec16 zero_1;
-    int32_t group_counter = 0;
-    for (int32_t k_idx = 0; k_idx < k_size; k_idx += 2) {
-      int64_t qwb_0 = *curr_q_weight;
-      int64_t qwb_1 = *(curr_q_weight + 1);
-      vec_op::FP32Vec16 wb_0(qwb_0, lut);
-      vec_op::FP32Vec16 wb_1(qwb_1, lut);
-
-      if constexpr (!use_desc_act) {
-        if (group_counter == 0) {
-          scale_0 = vec_op::FP32Vec16(scalar_vec_t(curr_scale));
-          scale_1 = vec_op::FP32Vec16(scale_0);
-          curr_scale += scales_stride;
-
-          if constexpr (has_zp) {
-            zero_0 = vec_op::FP32Vec16(*curr_zeros, lut);
-            zero_1 = vec_op::FP32Vec16(zero_0);
-            curr_zeros += zeros_stride / 2;
-          }
-        }
-      } else {
-        int32_t g_idx_0 = g_idx[k_idx];
-        int32_t g_idx_1 = g_idx[k_idx + 1];
-        scale_0 = vec_op::FP32Vec16(
-            scalar_vec_t(curr_scale + g_idx_0 * scales_stride));
-        scale_1 = vec_op::FP32Vec16(
-            scalar_vec_t(curr_scale + g_idx_1 * scales_stride));
-        if constexpr (has_zp) {
-          zero_0 = vec_op::FP32Vec16(*(curr_zeros + g_idx_0 * zeros_stride / 2),
-                                     lut);
-          zero_1 = vec_op::FP32Vec16(*(curr_zeros + g_idx_1 * zeros_stride / 2),
-                                     lut);
-        }
-      }
-
-      if constexpr (has_zp) {
-        wb_0 = wb_0 - zero_0;
-        wb_1 = wb_1 - zero_1;
-      }
-
-      wb_0 = wb_0 * scale_0;
-      wb_1 = wb_1 * scale_1;
-
-      scalar_vec_t output_vec_0(wb_0);
-      scalar_vec_t output_vec_1(wb_1);
-
-      // AMX needs to interlave K elements to pack as 32 bits
-      if constexpr (isa == ISA::AMX) {
-        vec_op::interleave_save(output_vec_0, output_vec_1, curr_weight);
-      } else {
-        output_vec_0.save(curr_weight);
-        output_vec_1.save(curr_weight + 16);
-      }
-
-      // update
-      curr_q_weight += 2;
-      curr_weight += 32;
-      if constexpr (!use_desc_act) {
-        group_counter += 2;
-        if (group_counter == group_size) {
-          group_counter = 0;
-        }
-      }
-    }
-  }
-};
-};  // namespace
-
-template <typename scalar_t, typename dequantizer_t, typename gemm_t>
-void cpu_gemm_wna16_impl(
-    scalar_t* __restrict__ input, int32_t* __restrict__ q_weight,
-    scalar_t* __restrict__ output, scalar_t* __restrict__ scales,
-    int32_t* __restrict__ zeros, int32_t* __restrict__ g_idx,
-    scalar_t* __restrict__ bias, const int32_t m_size, const int32_t n_size,
-    const int32_t k_size, const int64_t input_stride,
-    const int64_t output_stride, const int64_t scales_group_stride,
-    const int64_t zeros_group_stride, const int32_t group_num,
-    const int32_t group_size, const int64_t pack_factor) {
-  constexpr int32_t gemm_n_tile_size = gemm_t::NSize;
-  constexpr int32_t gemm_m_tile_size = gemm_t::MaxMSize;
-  constexpr int32_t n_block_size = 16;
-  static_assert(gemm_n_tile_size % n_block_size == 0);
-  const int32_t thread_num = omp_get_max_threads();
-
-  // a simple schedule policy, just to hold more B tiles in L2 and make sure
-  // each thread has tasks
-  const int32_t n_partition_size = [&]() {
-    const int64_t cache_size = cpu_utils::get_l2_size();
-    int64_t ps_cache_limit = cache_size / (k_size * sizeof(scalar_t));
-    int64_t ps_thread_limit = n_size / thread_num;
-    ps_cache_limit =
-        std::max((ps_cache_limit / gemm_n_tile_size) * gemm_n_tile_size,
-                 (int64_t)gemm_n_tile_size);
-    ps_thread_limit =
-        std::max((ps_thread_limit / gemm_n_tile_size) * gemm_n_tile_size,
-                 (int64_t)gemm_n_tile_size);
-    return std::min(ps_cache_limit, ps_thread_limit);
-  }();
-  const int32_t task_num = (n_size + n_partition_size - 1) / n_partition_size;
-
-  // get buffer size
-  const int64_t b_buffer_size =
-      (((n_partition_size * k_size * sizeof(scalar_t) + 63) / 64) * 64);
-  const int64_t c_buffer_size =
-      (((gemm_m_tile_size * gemm_n_tile_size * sizeof(float) + 63) / 64) * 64);
-  const int64_t b_buffer_offset = 0;
-  const int64_t c_buffer_offset = b_buffer_size;
-  const int64_t buffer_size = b_buffer_size + c_buffer_size;
-  DNNLScratchPadManager::get_dnnl_scratchpad_manager()->realloc(buffer_size *
-                                                                thread_num);
-
-  alignas(64) cpu_utils::Counter counter;
-  cpu_utils::Counter* counter_ptr = &counter;
-
-#pragma omp parallel for schedule(static, 1)
-  for (int32_t thread_id = 0; thread_id < thread_num; ++thread_id) {
-    scalar_t* __restrict__ b_buffer = nullptr;
-    float* __restrict__ c_buffer = nullptr;
-    {
-      uint8_t* buffer_ptr = DNNLScratchPadManager::get_dnnl_scratchpad_manager()
-                                ->get_data<uint8_t>() +
-                            thread_id * buffer_size;
-      b_buffer = reinterpret_cast<scalar_t*>(buffer_ptr + b_buffer_offset);
-      c_buffer = reinterpret_cast<float*>(buffer_ptr + c_buffer_offset);
-    }
-
-    const int64_t q_weight_block_stride = n_block_size / pack_factor * k_size;
-    const int64_t b_buffer_block_stride = n_block_size * k_size;
-    const int32_t zeros_block_stride = n_block_size / pack_factor;
-
-    gemm_t gemm;
-
-    for (;;) {
-      int32_t task_id = counter_ptr->acquire_counter();
-
-      if (task_id >= task_num) {
-        break;
-      }
-
-      const int32_t n_start_idx = task_id * n_partition_size;
-      const int32_t n_block_start_idx = n_start_idx / n_block_size;
-      const int32_t n_num = std::min(n_partition_size, n_size - n_start_idx);
-      const int32_t n_block_num = n_num / n_block_size;
-      // std::printf("thread_id: %d, task_id: %d, n_start_idx: %d, n_num: %d\n",
-      // thread_id, task_id, n_start_idx, n_num);
-
-      // dequant weight
-      {
-        int32_t* __restrict__ curr_q_weight =
-            q_weight + n_block_start_idx * q_weight_block_stride;
-        scalar_t* __restrict__ curr_b_buffer = b_buffer;
-        scalar_t* __restrict__ curr_scales = scales + n_start_idx;
-        int32_t* __restrict__ curr_zeros = zeros + n_start_idx / pack_factor;
-        for (int32_t block_idx = 0; block_idx < n_block_num; ++block_idx) {
-          dequantizer_t::dequant(curr_q_weight, curr_b_buffer, curr_scales,
-                                 curr_zeros, g_idx, scales_group_stride,
-                                 zeros_group_stride, k_size, group_size);
-
-          // if (block_idx == 0 && n_start_idx == 0) {
-          //     print_logits("depacked weight", curr_b_buffer, k_size,
-          //     n_block_size, n_block_size);
-          // }
-
-          // update
-          curr_q_weight += q_weight_block_stride;
-          curr_b_buffer += b_buffer_block_stride;
-          curr_scales += n_block_size;
-          curr_zeros += zeros_block_stride;
-        }
-      }
-
-      // compute loop
-      {
-        const int32_t n_tile_num = n_num / gemm_n_tile_size;
-        scalar_t* __restrict__ curr_input = input;
-        scalar_t* __restrict__ init_bias = bias;
-        if (bias != nullptr) {
-          init_bias += n_start_idx;
-        }
-        scalar_t* __restrict__ init_output = output + n_start_idx;
-        for (int32_t m_idx = 0; m_idx < m_size; m_idx += gemm_m_tile_size) {
-          const int32_t curr_m_size =
-              std::min(gemm_m_tile_size, m_size - m_idx);
-          scalar_t* __restrict__ curr_b_buffer = b_buffer;
-          scalar_t* __restrict__ curr_bias = init_bias;
-          scalar_t* __restrict__ curr_output = init_output;
-          for (int32_t n_tile_idx = 0; n_tile_idx < n_tile_num; ++n_tile_idx) {
-            gemm.gemm(curr_input, curr_b_buffer, c_buffer, curr_m_size, k_size,
-                      input_stride, b_buffer_block_stride, gemm_n_tile_size,
-                      false);
-
-            if (bias != nullptr) {
-              cpu_micro_gemm::bias_epilogue<gemm_n_tile_size>(
-                  c_buffer, curr_output, curr_bias, curr_m_size,
-                  gemm_n_tile_size, output_stride);
-              curr_bias += gemm_n_tile_size;
-            } else {
-              cpu_micro_gemm::default_epilogue<gemm_n_tile_size>(
-                  c_buffer, curr_output, curr_m_size, gemm_n_tile_size,
-                  output_stride);
-            }
-
-            curr_b_buffer +=
-                b_buffer_block_stride * (gemm_n_tile_size / n_block_size);
-            curr_output += gemm_n_tile_size;
-          }
-          curr_input += gemm_m_tile_size * input_stride;
-          init_output += gemm_m_tile_size * output_stride;
-        }
-      }
-    }
-  }
-}
-
-void cpu_gemm_wna16(
-    const torch::Tensor& input,  // [M, K]
-    const torch::Tensor&
-        q_weight,           // [N / 16, K * 16 / pack_factor], packed as int32
-    torch::Tensor& output,  // [M, N]
-    const torch::Tensor& scales,  // [group_num, N]
-    const std::optional<torch::Tensor>&
-        zeros,  // [group_num, N / pack_factor], packed as int32
-    const std::optional<torch::Tensor>& g_idx,  // [K]
-    const std::optional<torch::Tensor>& bias,   // [N]
-    const int64_t pack_factor, const std::string& isa_hint) {
-  using cpu_utils::ISA;
-  TORCH_CHECK_EQ(pack_factor, 8);  // only supports 4bits
-  const int32_t a_m_size = input.size(0);
-  const int32_t a_k_size = input.size(1);
-  const int64_t a_m_stride = input.stride(0);
-  const int32_t b_n_size = q_weight.size(0) * 16;
-  TORCH_CHECK_EQ(a_k_size % 32, 0);
-  TORCH_CHECK_EQ(b_n_size % 32, 0);
-  const int32_t group_num = scales.size(0);
-  const int32_t group_size = a_k_size / group_num;
-  TORCH_CHECK_EQ(group_size % 2, 0);
-  const int64_t scales_group_stride = scales.stride(0);
-  const int64_t output_m_stride = output.stride(0);
-
-  bool has_zp = zeros.has_value();
-  bool use_desc_act = g_idx.has_value();
-  TORCH_CHECK(!(has_zp && use_desc_act));
-
-  ISA isa = [&]() {
-    if (isa_hint == "amx") {
-      return ISA::AMX;
-    } else if (isa_hint == "vec") {
-      return ISA::VEC;
-    } else {
-      TORCH_CHECK(false, "unsupported isa hint: " + isa_hint);
-    }
-  }();
-
-  int32_t* zeros_ptr = has_zp ? zeros->data_ptr<int32_t>() : nullptr;
-  const int64_t zeros_group_stride = has_zp ? zeros->stride(0) : 0;
-  int32_t* g_idx_ptr = use_desc_act ? g_idx->data_ptr<int32_t>() : nullptr;
-
-  VLLM_DISPATCH_16B_TYPES(input.scalar_type(), "cpu_gemm_wna16", [&]() {
-    if (isa == ISA::AMX) {
-      using gemm_t = cpu_micro_gemm::MicroGemm<ISA::AMX, scalar_t>;
-      if (has_zp) {
-        using dequantizer_t = Dequantizer4b<scalar_t, ISA::AMX, true, false>;
-        cpu_gemm_wna16_impl<scalar_t, dequantizer_t, gemm_t>(
-            input.data_ptr<scalar_t>(), q_weight.data_ptr<int32_t>(),
-            output.data_ptr<scalar_t>(), scales.data_ptr<scalar_t>(), zeros_ptr,
-            g_idx_ptr, bias.has_value() ? bias->data_ptr<scalar_t>() : nullptr,
-            a_m_size, b_n_size, a_k_size, a_m_stride, output_m_stride,
-            scales_group_stride, zeros_group_stride, group_num, group_size,
-            pack_factor);
-        return;
-      }
-      if (use_desc_act) {
-        using dequantizer_t = Dequantizer4b<scalar_t, ISA::AMX, false, true>;
-        cpu_gemm_wna16_impl<scalar_t, dequantizer_t, gemm_t>(
-            input.data_ptr<scalar_t>(), q_weight.data_ptr<int32_t>(),
-            output.data_ptr<scalar_t>(), scales.data_ptr<scalar_t>(), zeros_ptr,
-            g_idx_ptr, bias.has_value() ? bias->data_ptr<scalar_t>() : nullptr,
-            a_m_size, b_n_size, a_k_size, a_m_stride, output_m_stride,
-            scales_group_stride, zeros_group_stride, group_num, group_size,
-            pack_factor);
-        return;
-      } else {
-        using dequantizer_t = Dequantizer4b<scalar_t, ISA::AMX, false, false>;
-        cpu_gemm_wna16_impl<scalar_t, dequantizer_t, gemm_t>(
-            input.data_ptr<scalar_t>(), q_weight.data_ptr<int32_t>(),
-            output.data_ptr<scalar_t>(), scales.data_ptr<scalar_t>(), zeros_ptr,
-            g_idx_ptr, bias.has_value() ? bias->data_ptr<scalar_t>() : nullptr,
-            a_m_size, b_n_size, a_k_size, a_m_stride, output_m_stride,
-            scales_group_stride, zeros_group_stride, group_num, group_size,
-            pack_factor);
-        return;
-      }
-    } else if (isa == ISA::VEC) {
-      using gemm_t = cpu_micro_gemm::MicroGemm<ISA::VEC, scalar_t>;
-      if (has_zp) {
-        using dequantizer_t = Dequantizer4b<scalar_t, ISA::VEC, true, false>;
-        cpu_gemm_wna16_impl<scalar_t, dequantizer_t, gemm_t>(
-            input.data_ptr<scalar_t>(), q_weight.data_ptr<int32_t>(),
-            output.data_ptr<scalar_t>(), scales.data_ptr<scalar_t>(), zeros_ptr,
-            g_idx_ptr, bias.has_value() ? bias->data_ptr<scalar_t>() : nullptr,
-            a_m_size, b_n_size, a_k_size, a_m_stride, output_m_stride,
-            scales_group_stride, zeros_group_stride, group_num, group_size,
-            pack_factor);
-        return;
-      }
-      if (use_desc_act) {
-        using dequantizer_t = Dequantizer4b<scalar_t, ISA::VEC, false, true>;
-        cpu_gemm_wna16_impl<scalar_t, dequantizer_t, gemm_t>(
-            input.data_ptr<scalar_t>(), q_weight.data_ptr<int32_t>(),
-            output.data_ptr<scalar_t>(), scales.data_ptr<scalar_t>(), zeros_ptr,
-            g_idx_ptr, bias.has_value() ? bias->data_ptr<scalar_t>() : nullptr,
-            a_m_size, b_n_size, a_k_size, a_m_stride, output_m_stride,
-            scales_group_stride, zeros_group_stride, group_num, group_size,
-            pack_factor);
-        return;
-      } else {
-        using dequantizer_t = Dequantizer4b<scalar_t, ISA::VEC, false, false>;
-        cpu_gemm_wna16_impl<scalar_t, dequantizer_t, gemm_t>(
-            input.data_ptr<scalar_t>(), q_weight.data_ptr<int32_t>(),
-            output.data_ptr<scalar_t>(), scales.data_ptr<scalar_t>(), zeros_ptr,
-            g_idx_ptr, bias.has_value() ? bias->data_ptr<scalar_t>() : nullptr,
-            a_m_size, b_n_size, a_k_size, a_m_stride, output_m_stride,
-            scales_group_stride, zeros_group_stride, group_num, group_size,
-            pack_factor);
-        return;
-      }
-    }
-  });
-}

csrc/cpu/dnnl_helper.cpp

@@ -5,7 +5,6 @@
 #include "common/memory.hpp"
 
 #include "dnnl_helper.h"
-#include "scratchpad_manager.h"
 
 static dnnl::engine& default_engine() {
   static dnnl::engine engine(dnnl::engine::kind::cpu, 0);
@@ -23,6 +22,23 @@ void release_dnnl_matmul_handler(int64_t handler) {
   delete ptr;
 }
 
+DNNLScratchPadManager::DNNLScratchPadManager() : size_(0), ptr_(nullptr) {
+  this->realloc(allocation_unit * 128);
+}
+
+void DNNLScratchPadManager::realloc(size_t new_size) {
+  new_size = round(new_size);
+  if (new_size > size_) {
+    ptr_ = std::aligned_alloc(64, new_size);
+    size_ = new_size;
+  }
+}
+
+DNNLScratchPadManager* DNNLScratchPadManager::get_dnnl_scratchpad_manager() {
+  static DNNLScratchPadManager manager;
+  return &manager;
+}
+
 template <typename KT, typename VT>
 class DNNLPrimitiveCache {
  public:
@@ -396,9 +412,9 @@ MatMulPrimitiveHandler::MatMulPrimitiveHandler(const Args& args)
     : DNNLMatMulPrimitiveHandler(
           static_cast<DNNLMatMulPrimitiveHandler::Args>(args), args.ab_type),
       m_size_cache_(nullptr) {
-  assert(b_type_ == dnnl::memory::data_type::f32 ||
-         b_type_ == dnnl::memory::data_type::bf16 ||
-         b_type_ == dnnl::memory::data_type::f16);
+  assert(ab_type_ == dnnl::memory::data_type::f32 ||
+         ab_type_ == dnnl::memory::data_type::bf16 ||
+         ab_type_ == dnnl::memory::data_type::f16);
 
   dnnl::memory::desc original_b_md({b_k_size_, b_n_size_}, b_type_,
                                    {b_k_stride_, b_n_stride_});

csrc/cpu/dnnl_helper.h

@@ -59,6 +59,30 @@ constexpr inline dnnl::memory::data_type get_dnnl_type() {
   return DNNLType<std::decay_t<T>>::type;
 }
 
+class DNNLScratchPadManager {
+ public:
+  static constexpr size_t allocation_unit = 4 * 1024 * 1024;  // 4KB
+
+  static DNNLScratchPadManager* get_dnnl_scratchpad_manager();
+
+  DNNLScratchPadManager();
+
+  template <typename T>
+  T* get_data() {
+    return reinterpret_cast<T*>(ptr_);
+  }
+
+  static size_t round(size_t size) {
+    return ((size + allocation_unit - 1) / allocation_unit) * allocation_unit;
+  }
+
+  void realloc(size_t new_size);
+
+ private:
+  size_t size_;
+  void* ptr_;
+};
+
 class DNNLMatMulPrimitiveHandler {
  public:
   virtual ~DNNLMatMulPrimitiveHandler() = default;

csrc/cpu/micro_gemm/cpu_micro_gemm_amx.hpp

@@ -1,245 +0,0 @@
-#ifndef CPU_MICRO_GEMM_AMX_HPP
-#define CPU_MICRO_GEMM_AMX_HPP
-#include "cpu/micro_gemm/cpu_micro_gemm_impl.hpp"
-
-namespace cpu_micro_gemm {
-namespace {
-// AMX specific
-constexpr static int64_t AMX_TILE_ROW_BYTES = 64;
-constexpr static int64_t AMX_TILE_ROW_NUM = 16;
-constexpr static int64_t AMX_TILE_BYTES = AMX_TILE_ROW_BYTES * AMX_TILE_ROW_NUM;
-
-typedef struct __tile_config {
-  uint8_t palette_id = 1;
-  uint8_t start_row = 0;
-  uint8_t reserved_0[14] = {0};
-  uint16_t colsb[16] = {0};
-  uint8_t rows[16] = {0};
-} __tilecfg;
-
-// 2-2-4 pattern, for 16 < m <= 32
-// TILE 0, 1: load A matrix, row num should be 16, m - 16
-// TILE 2, 3: load B matrix, row num should be 16
-// TILE 4, 5, 6, 7: store results C matrix, row num should be 16, 16, m - 16, m
-// - 16
-template <typename scalar_t>
-class TileGemm224 {
- public:
-  FORCE_INLINE static void gemm(DEFINE_CPU_MICRO_GEMM_PARAMS) {
-    TORCH_CHECK(false, "Unsupported data type for TileGemm224");
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    TORCH_CHECK(false, "Unsupported data type for TileGemm224");
-  }
-};
-
-template <>
-class TileGemm224<c10::BFloat16> {
- public:
-  using scalar_t = c10::BFloat16;
-  FORCE_INLINE static void gemm(DEFINE_CPU_MICRO_GEMM_PARAMS) {
-    const int32_t k_times = k / (AMX_TILE_ROW_NUM * 4 / sizeof(c10::BFloat16));
-    c10::BFloat16* __restrict__ a_tile_0 = a_ptr;
-    c10::BFloat16* __restrict__ a_tile_1 = a_ptr + lda * AMX_TILE_ROW_NUM;
-    const int64_t a_tile_stride = lda * sizeof(c10::BFloat16);
-
-    // B is always packed as 16 output channels block
-    c10::BFloat16* __restrict__ b_tile_2 = b_ptr;
-    c10::BFloat16* __restrict__ b_tile_3 = b_ptr + b_n_group_stride;
-    const int32_t b_tile_stride = AMX_TILE_ROW_BYTES;
-
-    float* __restrict__ c_tile_4 = c_ptr;
-    float* __restrict__ c_tile_5 =
-        c_tile_4 + AMX_TILE_ROW_BYTES / sizeof(float);
-    float* __restrict__ c_tile_6 = c_ptr + AMX_TILE_ROW_NUM * ldc;
-    float* __restrict__ c_tile_7 =
-        c_tile_6 + AMX_TILE_ROW_BYTES / sizeof(float);
-    const int32_t c_tile_stride = ldc * sizeof(float);
-
-    if (accum_c) {
-      _tile_loadd(4, c_tile_4, c_tile_stride);
-      _tile_loadd(5, c_tile_5, c_tile_stride);
-      _tile_loadd(6, c_tile_6, c_tile_stride);
-      _tile_loadd(7, c_tile_7, c_tile_stride);
-    } else {
-      _tile_zero(4);
-      _tile_zero(5);
-      _tile_zero(6);
-      _tile_zero(7);
-    }
-
-    for (int32_t k = 0; k < k_times; ++k) {
-      _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_tile_stride);
-      _tile_dpbf16ps(4, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_tile_stride);
-      _tile_dpbf16ps(5, 0, 3);
-      _tile_loadd(1, a_tile_1, a_tile_stride);
-      _tile_dpbf16ps(6, 1, 2);
-      _tile_dpbf16ps(7, 1, 3);
-
-      // update ptrs
-      a_tile_0 += AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-      a_tile_1 += AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-      b_tile_2 += AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_3 += AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    }
-
-    _tile_stored(4, c_tile_4, c_tile_stride);
-    _tile_stored(5, c_tile_5, c_tile_stride);
-    _tile_stored(6, c_tile_6, c_tile_stride);
-    _tile_stored(7, c_tile_7, c_tile_stride);
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    const int32_t m_0 = AMX_TILE_ROW_NUM;
-    const int32_t m_1 = m - AMX_TILE_ROW_NUM;
-    config.rows[0] = m_0;
-    config.rows[1] = m_1;
-    config.rows[2] = AMX_TILE_ROW_NUM;
-    config.rows[3] = AMX_TILE_ROW_NUM;
-    config.rows[4] = m_0;
-    config.rows[5] = m_0;
-    config.rows[6] = m_1;
-    config.rows[7] = m_1;
-    _tile_loadconfig(&config);
-  }
-};
-
-// 1-2-2 pattern, for 0 < m <= 16
-// TILE 0, (1): load A matrix, use extra 1 tile for prefetch, row num should be
-// m, m
-// TILE 2, 3, (4, 5): load B matrix, use extra 2 tiles for prefetch, row
-// num should be 16
-// TILE 6, 7, (6, 7): store results C matrix, row num should be
-// m
-template <typename scalar_t>
-class TileGemm122 {
- public:
-  FORCE_INLINE static void gemm(DEFINE_CPU_MICRO_GEMM_PARAMS) {
-    TORCH_CHECK(false, "Unsupported data type for TileGemm122");
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    TORCH_CHECK(false, "Unsupported data type for TileGemm122");
-  }
-};
-
-template <>
-class TileGemm122<c10::BFloat16> {
- public:
-  using scalar_t = c10::BFloat16;
-  FORCE_INLINE static void gemm(DEFINE_CPU_MICRO_GEMM_PARAMS) {
-    c10::BFloat16* __restrict__ a_tile_0 = a_ptr;
-    c10::BFloat16* __restrict__ a_tile_1 =
-        a_ptr + AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-    const int64_t a_tile_stride = lda * sizeof(c10::BFloat16);
-
-    c10::BFloat16* __restrict__ b_tile_2 = b_ptr;
-    c10::BFloat16* __restrict__ b_tile_3 = b_ptr + b_n_group_stride;
-    c10::BFloat16* __restrict__ b_tile_4 =
-        b_tile_2 + AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    c10::BFloat16* __restrict__ b_tile_5 =
-        b_tile_3 + AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    int64_t b_stride = AMX_TILE_ROW_BYTES;
-
-    float* __restrict__ c_tile_6 = c_ptr;
-    float* __restrict__ c_tile_7 = c_ptr + AMX_TILE_ROW_BYTES / sizeof(float);
-    int64_t c_stride = ldc * sizeof(float);
-
-    const int32_t k_times = k / (AMX_TILE_ROW_NUM * 4 / sizeof(c10::BFloat16));
-    const int32_t k_group_times = k_times / 2;
-    const bool has_tail = (k_times % 2 == 1);
-
-    if (accum_c) {
-      _tile_loadd(6, c_tile_6, c_stride);
-      _tile_loadd(7, c_tile_7, c_stride);
-    } else {
-      _tile_zero(6);
-      _tile_zero(7);
-    }
-
-    for (int32_t k = 0; k < k_group_times; ++k) {
-      _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_stride);
-      _tile_dpbf16ps(6, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_stride);
-      _tile_dpbf16ps(7, 0, 3);
-      _tile_loadd(1, a_tile_1, a_tile_stride);
-      _tile_stream_loadd(4, b_tile_4, b_stride);
-      _tile_dpbf16ps(6, 1, 4);
-      _tile_stream_loadd(5, b_tile_5, b_stride);
-      _tile_dpbf16ps(7, 1, 5);
-
-      // update ptrs
-      a_tile_0 += 2 * AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-      a_tile_1 += 2 * AMX_TILE_ROW_BYTES / sizeof(c10::BFloat16);
-      b_tile_2 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_3 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_4 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-      b_tile_5 += 2 * AMX_TILE_BYTES / sizeof(c10::BFloat16);
-    }
-
-    if (has_tail) {
-      _tile_loadd(0, a_tile_0, a_tile_stride);
-      _tile_stream_loadd(2, b_tile_2, b_stride);
-      _tile_dpbf16ps(6, 0, 2);
-      _tile_stream_loadd(3, b_tile_3, b_stride);
-      _tile_dpbf16ps(7, 0, 3);
-    }
-
-    _tile_stored(6, c_tile_6, c_stride);
-    _tile_stored(7, c_tile_7, c_stride);
-  }
-
-  FORCE_INLINE static void init_tile_config(int32_t m, __tilecfg& config) {
-    config.rows[0] = m;
-    config.rows[1] = m;
-    config.rows[2] = AMX_TILE_ROW_NUM;
-    config.rows[3] = AMX_TILE_ROW_NUM;
-    config.rows[4] = AMX_TILE_ROW_NUM;
-    config.rows[5] = AMX_TILE_ROW_NUM;
-    config.rows[6] = m;
-    config.rows[7] = m;
-    _tile_loadconfig(&config);
-  }
-};
-}  // namespace
-
-// Gemm kernel uses AMX, requires B matrix to be packed
-template <typename scalar_t>
-class MicroGemm<cpu_utils::ISA::AMX, scalar_t> {
- public:
-  static constexpr int32_t MaxMSize = 32;
-  static constexpr int32_t NSize = 32;
-
- public:
-  MicroGemm() : curr_m_(-1) {
-    vec_op::unroll_loop<int, 8>([&](int i) { amx_tile_config_.colsb[i] = 64; });
-  }
-
-  void gemm(DEFINE_CPU_MICRO_GEMM_PARAMS) {
-    if (m > AMX_TILE_ROW_NUM) {
-      if (m != curr_m_) {
-        curr_m_ = m;
-        TileGemm224<scalar_t>::init_tile_config(m, amx_tile_config_);
-      }
-      TileGemm224<scalar_t>::gemm(CPU_MICRO_GEMM_PARAMS);
-    } else {
-      if (m != curr_m_) {
-        curr_m_ = m;
-        TileGemm122<scalar_t>::init_tile_config(m, amx_tile_config_);
-      }
-      TileGemm122<scalar_t>::gemm(CPU_MICRO_GEMM_PARAMS);
-    }
-  }
-
- private:
-  alignas(64) __tilecfg amx_tile_config_;
-  int32_t curr_m_;
-};
-
-}  // namespace cpu_micro_gemm
-
-#endif

csrc/cpu/micro_gemm/cpu_micro_gemm_impl.hpp

@@ -1,91 +0,0 @@
-#ifndef CPU_MICRO_GEMM_IMPL_HPP
-#define CPU_MICRO_GEMM_IMPL_HPP
-#include "cpu/utils.hpp"
-#include "cpu/cpu_types.hpp"
-
-namespace cpu_micro_gemm {
-#define DEFINE_CPU_MICRO_GEMM_PARAMS                                        \
-  scalar_t *__restrict__ a_ptr, scalar_t *__restrict__ b_ptr,               \
-      float *__restrict__ c_ptr, const int32_t m, const int32_t k,          \
-      const int64_t lda, const int64_t b_n_group_stride, const int64_t ldc, \
-      const bool accum_c
-
-#define CPU_MICRO_GEMM_PARAMS \
-  a_ptr, b_ptr, c_ptr, m, k, lda, b_n_group_stride, ldc, accum_c
-
-template <cpu_utils::ISA isa, typename scalar_t>
-class MicroGemm {
- public:
-  static constexpr int32_t MaxMSize = 16;
-  static constexpr int32_t NSize = 16;
-
- public:
-  void gemm(DEFINE_CPU_MICRO_GEMM_PARAMS) {
-    TORCH_CHECK(false, "Unimplemented MicroGemm.");
-  }
-};
-
-template <int32_t n_size, typename scalar_t>
-FORCE_INLINE void default_epilogue(float* __restrict__ c_ptr,
-                                   scalar_t* __restrict__ d_ptr,
-                                   const int32_t m, const int64_t ldc,
-                                   const int64_t ldd) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  static_assert(n_size % 16 == 0);
-
-  float* __restrict__ curr_c = c_ptr;
-  scalar_t* __restrict__ curr_d = d_ptr;
-  for (int32_t i = 0; i < m; ++i) {
-    float* __restrict__ curr_c_iter = curr_c;
-    scalar_t* __restrict__ curr_d_iter = curr_d;
-    vec_op::unroll_loop<int32_t, n_size / 16>([&](int32_t n_g_idx) {
-      vec_op::FP32Vec16 c_vec_fp32(curr_c_iter);
-      scalar_vec_t c_vec(c_vec_fp32);
-      c_vec.save(curr_d_iter);
-      curr_c_iter += 16;
-      curr_d_iter += 16;
-    });
-    curr_c += ldc;
-    curr_d += ldd;
-  }
-}
-
-template <int32_t n_size, typename scalar_t>
-FORCE_INLINE void bias_epilogue(float* __restrict__ c_ptr,
-                                scalar_t* __restrict__ d_ptr,
-                                scalar_t* __restrict__ bias_ptr,
-                                const int32_t m, const int64_t ldc,
-                                const int64_t ldd) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  static_assert(n_size % 16 == 0);
-  constexpr int32_t n_group_num = n_size / 16;
-  static_assert(n_group_num <= 16);
-
-  vec_op::FP32Vec16 bias_vecs[n_group_num];
-  scalar_t* __restrict__ curr_bias = bias_ptr;
-  vec_op::unroll_loop<int32_t, n_group_num>([&](int32_t i) {
-    scalar_vec_t vec(curr_bias);
-    bias_vecs[i] = vec_op::FP32Vec16(vec);
-    curr_bias += 16;
-  });
-
-  float* __restrict__ curr_c = c_ptr;
-  scalar_t* __restrict__ curr_d = d_ptr;
-  for (int32_t i = 0; i < m; ++i) {
-    float* __restrict__ curr_c_iter = curr_c;
-    scalar_t* __restrict__ curr_d_iter = curr_d;
-    vec_op::unroll_loop<int32_t, n_group_num>([&](int32_t n_g_idx) {
-      vec_op::FP32Vec16 c_vec_fp32(curr_c_iter);
-      c_vec_fp32 = c_vec_fp32 + bias_vecs[n_g_idx];
-      scalar_vec_t c_vec(c_vec_fp32);
-      c_vec.save(curr_d_iter);
-      curr_c_iter += 16;
-      curr_d_iter += 16;
-    });
-    curr_c += ldc;
-    curr_d += ldd;
-  }
-}
-}  // namespace cpu_micro_gemm
-
-#endif

csrc/cpu/micro_gemm/cpu_micro_gemm_vec.hpp

@@ -1,115 +0,0 @@
-#ifndef CPU_MICRO_GEMM_VEC_HPP
-#define CPU_MICRO_GEMM_VEC_HPP
-#include "cpu/micro_gemm/cpu_micro_gemm_impl.hpp"
-
-namespace cpu_micro_gemm {
-namespace {
-// 8-2-16 pattern, 8 regs for A, 2 regs for B, 16 regs for C, [8, K] @ [k, 32]
-template <typename scalar_t>
-class TileGemm82 {
- public:
-  FORCE_INLINE static void gemm(DEFINE_CPU_MICRO_GEMM_PARAMS) {
-    switch (m) {
-      case 1:
-        gemm_micro<1>(CPU_MICRO_GEMM_PARAMS);
-        break;
-      case 2:
-        gemm_micro<2>(CPU_MICRO_GEMM_PARAMS);
-        break;
-      case 3:
-        gemm_micro<3>(CPU_MICRO_GEMM_PARAMS);
-        break;
-      case 4:
-        gemm_micro<4>(CPU_MICRO_GEMM_PARAMS);
-        break;
-      case 5:
-        gemm_micro<5>(CPU_MICRO_GEMM_PARAMS);
-        break;
-      case 6:
-        gemm_micro<6>(CPU_MICRO_GEMM_PARAMS);
-        break;
-      case 7:
-        gemm_micro<7>(CPU_MICRO_GEMM_PARAMS);
-        break;
-      case 8:
-        gemm_micro<8>(CPU_MICRO_GEMM_PARAMS);
-        break;
-    }
-  }
-
-  template <int32_t M>
-  static void gemm_micro(DEFINE_CPU_MICRO_GEMM_PARAMS) {
-    static_assert(0 < M <= 8);
-    using load_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-
-    scalar_t* __restrict__ curr_b_0 = b_ptr;
-    scalar_t* __restrict__ curr_b_1 = b_ptr + b_n_group_stride;
-    float* __restrict__ curr_c_0 = c_ptr;
-    float* __restrict__ curr_c_1 = c_ptr + 16;
-
-    vec_op::FP32Vec16 c_regs[M * 2];
-    if (accum_c) {
-      float* __restrict__ curr_m_c_0 = curr_c_0;
-      float* __restrict__ curr_m_c_1 = curr_c_1;
-      vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        c_regs[i * 2] = vec_op::FP32Vec16(curr_m_c_0);
-        c_regs[i * 2 + 1] = vec_op::FP32Vec16(curr_m_c_1);
-
-        // update
-        curr_m_c_0 += ldc;
-        curr_m_c_1 += ldc;
-      });
-    }
-
-    scalar_t* __restrict__ curr_a = a_ptr;
-    for (int32_t k_idx = 0; k_idx < k; ++k_idx) {
-      load_vec_t b_0_reg(curr_b_0);
-      vec_op::FP32Vec16 fp32_b_0_reg(b_0_reg);
-      load_vec_t b_1_reg(curr_b_1);
-      vec_op::FP32Vec16 fp32_b_1_reg(b_1_reg);
-
-      scalar_t* __restrict__ curr_m_a = curr_a;
-      vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-        scalar_t v = *curr_m_a;
-        load_vec_t a_reg_original(v);
-        vec_op::FP32Vec16 a_reg(a_reg_original);
-        c_regs[i * 2] = c_regs[i * 2] + a_reg * fp32_b_0_reg;
-        c_regs[i * 2 + 1] = c_regs[i * 2 + 1] + a_reg * fp32_b_1_reg;
-
-        // update
-        curr_m_a += lda;
-      });
-
-      // update
-      curr_a += 1;
-      curr_b_0 += 16;
-      curr_b_1 += 16;
-    }
-
-    vec_op::unroll_loop<int32_t, M>([&](int32_t i) {
-      c_regs[i * 2].save(curr_c_0);
-      c_regs[i * 2 + 1].save(curr_c_1);
-
-      // update
-      curr_c_0 += ldc;
-      curr_c_1 += ldc;
-    });
-  }
-};
-}  // namespace
-
-// Gemm kernel uses vector instructions, requires B matrix to be packed
-template <typename scalar_t>
-class MicroGemm<cpu_utils::ISA::VEC, scalar_t> {
- public:
-  static constexpr int32_t MaxMSize = 8;
-  static constexpr int32_t NSize = 32;
-
- public:
-  void gemm(DEFINE_CPU_MICRO_GEMM_PARAMS) {
-    TileGemm82<scalar_t>::gemm(CPU_MICRO_GEMM_PARAMS);
-  }
-};
-}  // namespace cpu_micro_gemm
-
-#endif

csrc/cpu/scratchpad_manager.cpp

@@ -1,23 +0,0 @@
-#include <cstdlib>
-
-#include "scratchpad_manager.h"
-
-DNNLScratchPadManager::DNNLScratchPadManager() : size_(0), ptr_(nullptr) {
-  this->realloc(allocation_unit * 128);
-}
-
-void DNNLScratchPadManager::realloc(size_t new_size) {
-  new_size = round(new_size);
-  if (new_size > size_) {
-    if (ptr_ != nullptr) {
-      std::free(ptr_);
-    }
-    ptr_ = std::aligned_alloc(64, new_size);
-    size_ = new_size;
-  }
-}
-
-DNNLScratchPadManager* DNNLScratchPadManager::get_dnnl_scratchpad_manager() {
-  static DNNLScratchPadManager manager;
-  return &manager;
-}

csrc/cpu/scratchpad_manager.h

@@ -1,31 +0,0 @@
-#ifndef SCRATCHPAD_MANAGER_H
-#define SCRATCHPAD_MANAGER_H
-
-#include <cstddef>
-#include <cstdio>
-
-class DNNLScratchPadManager {
- public:
-  static constexpr size_t allocation_unit = 4 * 1024;  // 4KB
-
-  static DNNLScratchPadManager* get_dnnl_scratchpad_manager();
-
-  DNNLScratchPadManager();
-
-  template <typename T>
-  T* get_data() {
-    return reinterpret_cast<T*>(ptr_);
-  }
-
-  static size_t round(size_t size) {
-    return ((size + allocation_unit - 1) / allocation_unit) * allocation_unit;
-  }
-
-  void realloc(size_t new_size);
-
- private:
-  size_t size_;
-  void* ptr_;
-};
-
-#endif

csrc/cpu/shm.cpp

@@ -192,7 +192,7 @@ class SHMManager {
                       const int group_size)
       : _rank(rank),
         _group_size(group_size),
-        _thread_num(omp_get_max_threads()),
+        _thread_num(torch::get_num_threads()),
         _shm_names({""}),
         _shared_mem_ptrs({nullptr}),
         _shm_ctx(nullptr) {

csrc/cpu/torch_bindings.cpp

@@ -74,45 +74,25 @@ at::Tensor int8_scaled_mm_with_quant(at::Tensor& mat1, at::Tensor& mat2,
                                      const std::optional<at::Tensor>& bias,
                                      at::ScalarType out_dtype, bool is_vnni);
 
-torch::Tensor get_scheduler_metadata(
-    const int64_t num_req, const int64_t num_heads_q,
-    const int64_t num_heads_kv, const int64_t head_dim,
-    const torch::Tensor& seq_lens, at::ScalarType dtype,
-    const torch::Tensor& query_start_loc, const bool casual,
-    const int64_t window_size, const std::string& isa_hint,
-    const bool enable_kv_split);
-
-void cpu_attn_reshape_and_cache(const torch::Tensor& key,
-                                const torch::Tensor& value,
-                                torch::Tensor& key_cache,
-                                torch::Tensor& value_cache,
-                                const torch::Tensor& slot_mapping,
-                                const std::string& isa);
-
-void cpu_attention_with_kv_cache(
-    const torch::Tensor& query, const torch::Tensor& key_cache,
-    const torch::Tensor& value_cache, torch::Tensor& output,
-    const torch::Tensor& query_start_loc, const torch::Tensor& seq_lens,
-    const double scale, const bool causal,
-    const std::optional<torch::Tensor>& alibi_slopes,
-    const int64_t sliding_window_left, const int64_t sliding_window_right,
-    const torch::Tensor& block_table, const double softcap,
-    const torch::Tensor& scheduler_metadata,
-    const std::optional<torch::Tensor>& s_aux);
-
-// Note: just for avoiding importing errors
-void placeholder_op() { TORCH_CHECK(false, "Unimplemented"); }
-
-void cpu_gemm_wna16(const torch::Tensor& input, const torch::Tensor& q_weight,
-                    torch::Tensor& output, const torch::Tensor& scales,
-                    const std::optional<torch::Tensor>& zeros,
-                    const std::optional<torch::Tensor>& g_idx,
-                    const std::optional<torch::Tensor>& bias,
-                    const int64_t pack_factor, const std::string& isa_hint);
-
 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // vLLM custom ops
 
+  // Attention ops
+  // Compute the attention between an input query and the cached keys/values
+  // using PagedAttention.
+  ops.def(
+      "paged_attention_v1("
+      "    Tensor! out, Tensor query, Tensor key_cache,"
+      "    Tensor value_cache, int num_kv_heads, float scale,"
+      "    Tensor block_tables, Tensor seq_lens, int block_size,"
+      "    int max_seq_len, Tensor? alibi_slopes,"
+      "    str kv_cache_dtype, Tensor k_scale, Tensor v_scale,"
+      "    int tp_rank, int blocksparse_local_blocks,"
+      "    int blocksparse_vert_stride, int blocksparse_block_size,"
+      "    int blocksparse_head_sliding_step) -> ()");
+
+  ops.impl("paged_attention_v1", torch::kCPU, &paged_attention_v1);
+
   ops.def(
       "dynamic_4bit_int_moe("
       "Tensor x, Tensor topk_ids, Tensor topk_weights,"
@@ -122,6 +102,20 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 
   ops.impl("dynamic_4bit_int_moe", torch::kCPU, &dynamic_4bit_int_moe_cpu);
 
+  // PagedAttention V2.
+  ops.def(
+      "paged_attention_v2("
+      "    Tensor! out, Tensor! exp_sums, Tensor! max_logits,"
+      "    Tensor! tmp_out, Tensor query, Tensor key_cache,"
+      "    Tensor value_cache, int num_kv_heads, float scale,"
+      "    Tensor block_tables, Tensor seq_lens, int block_size,"
+      "    int max_seq_len, Tensor? alibi_slopes,"
+      "    str kv_cache_dtype, Tensor k_scale, Tensor v_scale,"
+      "    int tp_rank, int blocksparse_local_blocks,"
+      "    int blocksparse_vert_stride, int blocksparse_block_size,"
+      "    int blocksparse_head_sliding_step) -> ()");
+  ops.impl("paged_attention_v2", torch::kCPU, &paged_attention_v2);
+
   // Activation ops
 
   // Activation function used in SwiGLU.
@@ -172,6 +166,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // Quantization
 #if defined(__AVX512F__) || (defined(__aarch64__) && !defined(__APPLE__)) || \
     defined(__powerpc64__)
+  at::Tag stride_tag = at::Tag::needs_fixed_stride_order;
   // Helper function to release oneDNN handlers
   ops.def("release_dnnl_matmul_handler(int handler) -> ()",
           &release_dnnl_matmul_handler);
@@ -207,13 +202,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // Compute int8 quantized tensor for given scaling factor.
   ops.def(
       "static_scaled_int8_quant(Tensor! out, Tensor input, Tensor scale,"
-      "Tensor? azp) -> ()");
+      "Tensor? azp) -> ()",
+      {stride_tag});
   ops.impl("static_scaled_int8_quant", torch::kCPU, &static_scaled_int8_quant);
 
   // Compute int8 quantized tensor and scaling factor
   ops.def(
       "dynamic_scaled_int8_quant(Tensor! out, Tensor input, Tensor! scale, "
-      "Tensor!? azp) -> ()");
+      "Tensor!? azp) -> ()",
+      {stride_tag});
   ops.impl("dynamic_scaled_int8_quant", torch::kCPU,
            &dynamic_scaled_int8_quant);
 #endif
@@ -262,40 +259,37 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("int8_scaled_mm_with_quant", torch::kCPU,
            &int8_scaled_mm_with_quant);
 #endif
+}
 
-  // CPU attention kernels
-  ops.def(
-      "get_scheduler_metadata(int num_req, int num_heads_q, int num_heads_kv, "
-      "int head_dim, Tensor seq_lens, ScalarType dtype, Tensor "
-      "query_start_loc, bool casual, int window_size, str isa_hint, bool "
-      "enable_kv_split) -> Tensor",
-      &get_scheduler_metadata);
-  ops.def(
-      "cpu_attn_reshape_and_cache(Tensor key, Tensor value, Tensor(a2!) "
-      "key_cache, Tensor(a3!) value_cache, Tensor slot_mapping, str "
-      "isa) -> ()",
-      &cpu_attn_reshape_and_cache);
-  ops.def(
-      "cpu_attention_with_kv_cache(Tensor query, Tensor key_cache, Tensor "
-      "value_cache, Tensor(a3!) output, Tensor query_start_loc, Tensor "
-      "seq_lens, float scale, bool causal, Tensor? alibi_slopes, SymInt "
-      "sliding_window_left, SymInt sliding_window_right, Tensor block_table, "
-      "float softcap, Tensor sheduler_metadata, Tensor? s_aux) -> ()",
-      &cpu_attention_with_kv_cache);
+TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
+  // Cache ops
+  // Swap in (out) the cache blocks from src to dst.
+  cache_ops.def(
+      "swap_blocks(Tensor src, Tensor! dst, Tensor block_mapping) -> ()");
+  cache_ops.impl("swap_blocks", torch::kCPU, &swap_blocks);
 
-  // placeholders
-  ops.def("static_scaled_fp8_quant() -> ()", placeholder_op);
-  ops.def("dynamic_scaled_fp8_quant() -> ()", placeholder_op);
-  ops.def("dynamic_per_token_scaled_fp8_quant() -> ()", placeholder_op);
+  // Copy the cache blocks from src to dst.
+  cache_ops.def(
+      "copy_blocks(Tensor(a!)[] key_caches, Tensor[](b!) value_caches, "
+      "Tensor block_mapping) -> ()");
+  cache_ops.impl("copy_blocks", torch::kCPU, &copy_blocks);
 
-  // WNA16
-#if defined(__AVX512F__)
-  ops.def(
-      "cpu_gemm_wna16(Tensor input, Tensor q_weight, Tensor(a2!) output, "
-      "Tensor scales, Tensor? zeros, Tensor? g_idx, Tensor? bias, SymInt "
-      "pack_factor, str isa_hint) -> ()");
-  ops.impl("cpu_gemm_wna16", torch::kCPU, &cpu_gemm_wna16);
-#endif
+  // Reshape the key and value tensors and cache them.
+  cache_ops.def(
+      "reshape_and_cache(Tensor key, Tensor value,"
+      "                  Tensor! key_cache, Tensor! value_cache,"
+      "                  Tensor slot_mapping,"
+      "                  str kv_cache_dtype,"
+      "                  Tensor k_scale, Tensor v_scale) -> ()");
+  cache_ops.impl("reshape_and_cache", torch::kCPU, &reshape_and_cache);
+
+  cache_ops.def(
+      "concat_and_cache_mla(Tensor kv_c, Tensor k_pe,"
+      "                     Tensor! kv_cache,"
+      "                     Tensor slot_mapping,"
+      "                     str kv_cache_dtype,"
+      "                     Tensor scale) -> ()");
+  cache_ops.impl("concat_and_cache_mla", torch::kCPU, &concat_and_cache_mla);
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _utils), utils) {

csrc/cpu/utils.cpp

@@ -45,55 +45,21 @@ std::string init_cpu_threads_env(const std::string& cpu_ids) {
   // Memory node binding
   if (numa_available() != -1) {
     int mem_node_id = numa_node_of_cpu(omp_cpu_ids.front());
-    std::set<int> node_ids;
-    for (const auto& cpu_id : omp_cpu_ids) {
-      int node_id = numa_node_of_cpu(cpu_id);
-      if (node_id != -1) {
-        node_ids.insert(node_id);
-      }
-      if (node_id != mem_node_id) {
-        TORCH_WARN("CPU ", cpu_id, " is on NUMA node ", node_id, ", but CPU ",
-                   omp_cpu_ids.front(), " is on NUMA node ", mem_node_id,
-                   ". All CPUs should be on the same NUMA node for optimal "
-                   "performance. Memory will be bound to NUMA node ",
-                   mem_node_id, ".");
-      }
+    bitmask* mask = numa_parse_nodestring(std::to_string(mem_node_id).c_str());
+    bitmask* src_mask = numa_get_membind();
+
+    int pid = getpid();
+
+    // move all existing pages to the specified numa node.
+    *(src_mask->maskp) = *(src_mask->maskp) ^ *(mask->maskp);
+    int page_num = numa_migrate_pages(pid, src_mask, mask);
+    if (page_num == -1) {
+      TORCH_WARN("numa_migrate_pages failed. errno: " + std::to_string(errno));
     }
-    // Concatenate all node_ids into a single comma-separated string
-    if (!node_ids.empty()) {
-      std::string node_ids_str;
-      for (const int node_id : node_ids) {
-        if (!node_ids_str.empty()) {
-          node_ids_str += ",";
-        }
-        node_ids_str += std::to_string(node_id);
-      }
 
-      bitmask* mask = numa_parse_nodestring(node_ids_str.c_str());
-      bitmask* src_mask = numa_get_membind();
-
-      int pid = getpid();
-
-      if (mask && src_mask) {
-        // move all existing pages to the specified numa node.
-        *(src_mask->maskp) = *(src_mask->maskp) ^ *(mask->maskp);
-        int page_num = numa_migrate_pages(pid, src_mask, mask);
-        if (page_num == -1) {
-          TORCH_WARN("numa_migrate_pages failed. errno: " +
-                     std::to_string(errno));
-        }
-
-        // restrict memory allocation node.
-        numa_set_membind(mask);
-        numa_set_strict(1);
-
-        numa_free_nodemask(mask);
-        numa_free_nodemask(src_mask);
-      } else {
-        TORCH_WARN("numa_parse_nodestring or numa_get_membind failed. errno: " +
-                   std::to_string(errno));
-      }
-    }
+    // restrict memory allocation node.
+    numa_set_membind(mask);
+    numa_set_strict(1);
   }
 
   // OMP threads binding

csrc/cpu/utils.hpp

@@ -1,73 +0,0 @@
-#ifndef UTILS_HPP
-#define UTILS_HPP
-
-#include <atomic>
-#include <cassert>
-#include <cstdint>
-#include <unistd.h>
-
-#if defined(__APPLE__)
-  #include <sys/sysctl.h>
-#endif
-
-#include "cpu_types.hpp"
-
-namespace cpu_utils {
-enum class ISA { AMX, VEC };
-
-template <typename T>
-struct VecTypeTrait {
-  using vec_t = void;
-};
-
-template <>
-struct VecTypeTrait<float> {
-  using vec_t = vec_op::FP32Vec16;
-};
-
-#if !defined(__aarch64__) || defined(ARM_BF16_SUPPORT)
-template <>
-struct VecTypeTrait<c10::BFloat16> {
-  using vec_t = vec_op::BF16Vec16;
-};
-#endif
-
-template <>
-struct VecTypeTrait<c10::Half> {
-  using vec_t = vec_op::FP16Vec16;
-};
-
-struct Counter {
-  std::atomic<int64_t> counter;
-  char _padding[56];
-
-  Counter() : counter(0) {}
-
-  void reset_counter() { counter.store(0); }
-
-  int64_t acquire_counter() { return counter++; }
-};
-
-inline int64_t get_l2_size() {
-  static int64_t size = []() {
-#if defined(__APPLE__)
-    // macOS doesn't have _SC_LEVEL2_CACHE_SIZE. Use sysctlbyname.
-    int64_t l2_cache_size = 0;
-    size_t len = sizeof(l2_cache_size);
-    if (sysctlbyname("hw.l2cachesize", &l2_cache_size, &len, NULL, 0) == 0 &&
-        l2_cache_size > 0) {
-      return l2_cache_size >> 1;  // use 50% of L2 cache
-    }
-    // Fallback if sysctlbyname fails
-    return 128LL * 1024 >> 1;  // use 50% of 128KB
-#else
-    long l2_cache_size = sysconf(_SC_LEVEL2_CACHE_SIZE);
-    assert(l2_cache_size != -1);
-    return l2_cache_size >> 1;  // use 50% of L2 cache
-#endif
-  }();
-  return size;
-}
-}  // namespace cpu_utils
-
-#endif

csrc/cuda_view.cu

@@ -22,10 +22,15 @@ torch::Tensor get_cuda_view_from_cpu_tensor(torch::Tensor& cpu_tensor) {
   auto strides = cpu_tensor.strides();
   auto options = cpu_tensor.options().device(torch::kCUDA);
 
-  // use default no-op deleter, since the memory is owned by the original CPU
-  // tensor
+  // from_blob signature: from_blob(void *data, IntArrayRef sizes, ..., Deleter,
+  // const TensorOptions &) Provide a no-op deleter. The CPU tensor holds the
+  // memory, so we don't free it here.
+  auto deleter = [](void*) {
+    // no-op, since the memory is owned by the original CPU tensor
+  };
+
   torch::Tensor cuda_tensor =
-      torch::from_blob(device_ptr, sizes, strides, options);
+      torch::from_blob(device_ptr, sizes, strides, deleter, options);
 
   TORCH_CHECK(cuda_tensor.device().is_cuda(),
               "Resulting tensor is not on CUDA device");

csrc/cumem_allocator.cpp

@@ -3,58 +3,14 @@
 // need to be unsigned long long
 #include <iostream>
 
-#include "cumem_allocator_compat.h"
-
-#ifndef USE_ROCM
-static const char* PYARGS_PARSE = "KKKK";
-#else
-  #include <cstdlib>
-  #include <cerrno>
-  #include <climits>
-
-// Default chunk size 256MB for ROCm. Can be overridden at runtime by the
-// environment variable VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE, specified in megabytes
-// (MB). The env value is parsed with strtoull as an integer number of MB
-// (decimal or 0x hex). The parsed MB value is converted to bytes. If
-// parsing fails, the value is 0, or the multiplication would overflow,
-// the default (256MB) is used.
-static const unsigned long long DEFAULT_MEMCREATE_CHUNK_SIZE =
-    (256ULL * 1024ULL * 1024ULL);
-
-static unsigned long long get_memcreate_chunk_size() {
-  const char* env = getenv("VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE");
-  if (!env) return DEFAULT_MEMCREATE_CHUNK_SIZE;
-  char* endptr = nullptr;
-  errno = 0;
-  unsigned long long val_mb = strtoull(env, &endptr, 0);
-  if (endptr == env || errno != 0) {
-    // parsing failed, fallback to default
-    return DEFAULT_MEMCREATE_CHUNK_SIZE;
-  }
-  if (val_mb == 0) return DEFAULT_MEMCREATE_CHUNK_SIZE;
-
-  const unsigned long long MB = 1024ULL * 1024ULL;
-  // guard against overflow when converting MB -> bytes
-  if (val_mb > (ULLONG_MAX / MB)) {
-    return DEFAULT_MEMCREATE_CHUNK_SIZE;
-  }
-  return val_mb * MB;
-}
-
-static inline unsigned long long my_min(unsigned long long a,
-                                        unsigned long long b) {
-  return a < b ? a : b;
-}
-
-static const char* PYARGS_PARSE = "KKKO";
-#endif
-
 extern "C" {
 
 #define PY_SSIZE_T_CLEAN
 #include <Python.h>
 
 #include <sys/types.h>
+#include <cuda_runtime_api.h>
+#include <cuda.h>
 
 char error_msg[10240];  // 10KB buffer to store error messages
 CUresult no_error = CUresult(0);
@@ -93,12 +49,7 @@ void ensure_context(unsigned long long device) {
 }
 
 void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
-#ifndef USE_ROCM
                     CUmemGenericAllocationHandle* p_memHandle) {
-#else
-                    CUmemGenericAllocationHandle** p_memHandle,
-                    unsigned long long* chunk_sizes, size_t num_chunks) {
-#endif
   ensure_context(device);
   // Define memory allocation properties
   CUmemAllocationProp prop = {};
@@ -107,7 +58,6 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
   prop.location.id = device;
   prop.allocFlags.compressionType = CU_MEM_ALLOCATION_COMP_NONE;
 
-#ifndef USE_ROCM
   // Allocate memory using cuMemCreate
   CUDA_CHECK(cuMemCreate(p_memHandle, size, &prop, 0));
   if (error_code != 0) {
@@ -117,39 +67,6 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
   if (error_code != 0) {
     return;
   }
-#else
-  for (auto i = 0; i < num_chunks; ++i) {
-    CUDA_CHECK(cuMemCreate(p_memHandle[i], chunk_sizes[i], &prop, 0));
-    if (error_code != 0) {
-      // Clean up previously created handles
-      for (auto j = 0; j < i; ++j) {
-        cuMemRelease(*(p_memHandle[j]));
-      }
-      return;
-    }
-  }
-  unsigned long long allocated_size = 0;
-  for (auto i = 0; i < num_chunks; ++i) {
-    void* map_addr = (void*)((uintptr_t)d_mem + allocated_size);
-    CUDA_CHECK(cuMemMap(map_addr, chunk_sizes[i], 0, *(p_memHandle[i]), 0));
-    if (error_code != 0) {
-      // unmap previously mapped chunks
-      unsigned long long unmapped_size = 0;
-      for (auto j = 0; j < i; ++j) {
-        void* unmap_addr = (void*)((uintptr_t)d_mem + unmapped_size);
-        cuMemUnmap(unmap_addr, chunk_sizes[j]);
-        unmapped_size += chunk_sizes[j];
-      }
-      // release all created handles
-      for (auto j = 0; j < num_chunks; ++j) {
-        cuMemRelease(*(p_memHandle[j]));
-      }
-      return;
-    }
-    allocated_size += chunk_sizes[i];
-  }
-#endif
-
   CUmemAccessDesc accessDesc = {};
   accessDesc.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
   accessDesc.location.id = device;
@@ -165,16 +82,10 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
 
 void unmap_and_release(unsigned long long device, ssize_t size,
                        CUdeviceptr d_mem,
-#ifndef USE_ROCM
                        CUmemGenericAllocationHandle* p_memHandle) {
-#else
-                       CUmemGenericAllocationHandle** p_memHandle,
-                       unsigned long long* chunk_sizes, size_t num_chunks) {
-#endif
   // std::cout << "unmap_and_release: device=" << device << ", size=" << size <<
   // ", d_mem=" << d_mem << ", p_memHandle=" << p_memHandle << std::endl;
   ensure_context(device);
-#ifndef USE_ROCM
   CUDA_CHECK(cuMemUnmap(d_mem, size));
   if (error_code != 0) {
     return;
@@ -183,30 +94,6 @@ void unmap_and_release(unsigned long long device, ssize_t size,
   if (error_code != 0) {
     return;
   }
-#else
-  unsigned long long allocated_size = 0;
-  CUresult first_error = no_error;
-
-  for (auto i = 0; i < num_chunks; ++i) {
-    void* map_addr = (void*)((uintptr_t)d_mem + allocated_size);
-    CUresult status = cuMemUnmap(map_addr, chunk_sizes[i]);
-    if (status != no_error && first_error == no_error) {
-      first_error = status;
-    }
-    allocated_size += chunk_sizes[i];
-  }
-
-  for (auto i = 0; i < num_chunks; ++i) {
-    CUresult status = cuMemRelease(*(p_memHandle[i]));
-    if (status != no_error && first_error == no_error) {
-      first_error = status;
-    }
-  }
-
-  if (first_error != no_error) {
-    CUDA_CHECK(first_error);
-  }
-#endif
 }
 
 PyObject* create_tuple_from_c_integers(unsigned long long a,
@@ -233,36 +120,6 @@ PyObject* create_tuple_from_c_integers(unsigned long long a,
   return tuple;  // Return the created tuple
 }
 
-PyObject* create_tuple_from_c_mixed(unsigned long long a, unsigned long long b,
-                                    unsigned long long c,
-                                    CUmemGenericAllocationHandle** vec,
-                                    unsigned long long* chunk_sizes,
-                                    size_t num_chunks) {
-  PyObject* tuple = PyTuple_New(4);
-  if (!tuple) {
-    return NULL;
-  }
-
-  // PyObject* list = PyList_New(vec.size());
-  PyObject* list = PyList_New(num_chunks);
-  for (auto i = 0; i < num_chunks; ++i) {
-    PyObject* addr_size_pair = PyTuple_New(2);
-    PyObject* addr = PyLong_FromUnsignedLongLong((unsigned long long)(vec[i]));
-    PyObject* size =
-        PyLong_FromUnsignedLongLong((unsigned long long)(chunk_sizes[i]));
-    PyTuple_SetItem(addr_size_pair, 0, addr);
-    PyTuple_SetItem(addr_size_pair, 1, size);
-    PyList_SetItem(list, i, addr_size_pair);
-  }
-
-  PyTuple_SetItem(tuple, 0, PyLong_FromUnsignedLongLong(a));
-  PyTuple_SetItem(tuple, 1, PyLong_FromUnsignedLongLong(b));
-  PyTuple_SetItem(tuple, 2, PyLong_FromUnsignedLongLong(c));
-  PyTuple_SetItem(tuple, 3, list);
-
-  return tuple;
-}
-
 // ---------------------------------------------------------------------------
 // Our exported C functions that call Python:
 
@@ -290,55 +147,14 @@ void* my_malloc(ssize_t size, int device, CUstream stream) {
   size_t alignedSize = ((size + granularity - 1) / granularity) * granularity;
 
   CUdeviceptr d_mem;
-#ifndef USE_ROCM
   CUDA_CHECK(cuMemAddressReserve(&d_mem, alignedSize, 0, 0, 0));
   if (error_code != 0) {
     return nullptr;
   }
-#else
-  CUDA_CHECK(cuMemAddressReserve(&d_mem, alignedSize, granularity, 0, 0));
-  if (error_code != 0) {
-    return nullptr;
-  }
-#endif
-
-#ifndef USE_ROCM
   // allocate the CUmemGenericAllocationHandle
   CUmemGenericAllocationHandle* p_memHandle =
       (CUmemGenericAllocationHandle*)malloc(
           sizeof(CUmemGenericAllocationHandle));
-#else
-  // Make sure chunk size is aligned with hardware granularity. The base
-  // chunk size can be configured via environment variable
-  // ``VLLM_ROCM_SLEEP_MEM_CHUNK_SIZE``; otherwise
-  // DEFAULT_MEMCREATE_CHUNK_SIZE is used.
-  size_t base_chunk = (size_t)get_memcreate_chunk_size();
-  size_t aligned_chunk_size =
-      ((base_chunk + granularity - 1) / granularity) * granularity;
-  size_t num_chunks =
-      (alignedSize + aligned_chunk_size - 1) / aligned_chunk_size;
-  CUmemGenericAllocationHandle** p_memHandle =
-      (CUmemGenericAllocationHandle**)malloc(
-          num_chunks * sizeof(CUmemGenericAllocationHandle*));
-  unsigned long long* chunk_sizes =
-      (unsigned long long*)malloc(num_chunks * sizeof(unsigned long long));
-  for (auto i = 0; i < num_chunks; ++i) {
-    p_memHandle[i] = (CUmemGenericAllocationHandle*)malloc(
-        sizeof(CUmemGenericAllocationHandle));
-    if (p_memHandle[i] == nullptr) {
-      std::cerr << "ERROR: malloc failed for p_memHandle[" << i << "].\n";
-      for (auto j = 0; j < i; ++j) {
-        free(p_memHandle[j]);
-      }
-      free(p_memHandle);
-      free(chunk_sizes);
-      return nullptr;
-    }
-    chunk_sizes[i] = (unsigned long long)my_min(
-        (unsigned long long)(alignedSize - i * aligned_chunk_size),
-        (unsigned long long)aligned_chunk_size);
-  }
-#endif
 
   if (!g_python_malloc_callback) {
     std::cerr << "ERROR: g_python_malloc_callback not set.\n";
@@ -348,15 +164,9 @@ void* my_malloc(ssize_t size, int device, CUstream stream) {
   // Acquire GIL (not in stable ABI officially, but often works)
   PyGILState_STATE gstate = PyGILState_Ensure();
 
-#ifndef USE_ROCM
   PyObject* arg_tuple = create_tuple_from_c_integers(
       (unsigned long long)device, (unsigned long long)alignedSize,
       (unsigned long long)d_mem, (unsigned long long)p_memHandle);
-#else
-  PyObject* arg_tuple = create_tuple_from_c_mixed(
-      (unsigned long long)device, (unsigned long long)alignedSize,
-      (unsigned long long)d_mem, p_memHandle, chunk_sizes, num_chunks);
-#endif
 
   // Call g_python_malloc_callback
   PyObject* py_result =
@@ -372,27 +182,7 @@ void* my_malloc(ssize_t size, int device, CUstream stream) {
   PyGILState_Release(gstate);
 
   // do the final mapping
-#ifndef USE_ROCM
   create_and_map(device, alignedSize, d_mem, p_memHandle);
-#else
-  create_and_map(device, alignedSize, d_mem, p_memHandle, chunk_sizes,
-                 num_chunks);
-  free(chunk_sizes);
-#endif
-
-  if (error_code != 0) {
-    // free address and the handle
-    CUDA_CHECK(cuMemAddressFree(d_mem, alignedSize));
-#ifndef USE_ROCM
-    free(p_memHandle);
-#else
-    for (size_t i = 0; i < num_chunks; ++i) {
-      free(p_memHandle[i]);
-    }
-    free(p_memHandle);
-#endif
-    return nullptr;
-  }
 
   return (void*)d_mem;
 }
@@ -416,96 +206,36 @@ void my_free(void* ptr, ssize_t size, int device, CUstream stream) {
 
   if (!py_result || !PyTuple_Check(py_result) || PyTuple_Size(py_result) != 4) {
     PyErr_SetString(PyExc_TypeError, "Expected a tuple of size 4");
-    Py_XDECREF(py_result);
-    Py_XDECREF(py_ptr);
     return;
   }
 
   unsigned long long recv_device, recv_size;
-  unsigned long long recv_d_mem;
-#ifndef USE_ROCM
-  unsigned long long recv_p_memHandle;
-#else
-  PyObject* recv_p_memHandle;
-#endif
+  unsigned long long recv_d_mem, recv_p_memHandle;
   // Unpack the tuple into four C integers
-  if (!PyArg_ParseTuple(py_result, PYARGS_PARSE, &recv_device, &recv_size,
+  if (!PyArg_ParseTuple(py_result, "KKKK", &recv_device, &recv_size,
                         &recv_d_mem, &recv_p_memHandle)) {
     // PyArg_ParseTuple sets an error if it fails
-    Py_XDECREF(py_result);
-    Py_XDECREF(py_ptr);
     return;
   }
 
-  // For ROCm, copy the Python list of (addr,size) pairs into C arrays while
-  // holding the GIL. Then release the GIL and call the unmap/release helper
-  // using the copied arrays. This avoids calling PyList_* APIs without the
-  // GIL (which is undefined behavior and can crash when called from other
-  // threads).
+  PyGILState_Release(gstate);
+
+  // recv_size == size
+  // recv_device == device
+
+  // Free memory
+
   CUdeviceptr d_mem = (CUdeviceptr)recv_d_mem;
-#ifdef USE_ROCM
-  Py_ssize_t num_chunks = PyList_Size(recv_p_memHandle);
-  CUmemGenericAllocationHandle** p_memHandle =
-      (CUmemGenericAllocationHandle**)malloc(
-          num_chunks * sizeof(CUmemGenericAllocationHandle*));
-  if (p_memHandle == nullptr) {
-    Py_DECREF(py_ptr);
-    Py_DECREF(py_result);
-    PyGILState_Release(gstate);
-    std::cerr << "ERROR: malloc failed for p_memHandle in my_free."
-              << std::endl;
-    return;
-  }
-  unsigned long long* chunk_sizes =
-      (unsigned long long*)malloc(num_chunks * sizeof(unsigned long long));
-  if (chunk_sizes == nullptr) {
-    free(p_memHandle);
-    Py_DECREF(py_ptr);
-    Py_DECREF(py_result);
-    PyGILState_Release(gstate);
-    std::cerr << "ERROR: malloc failed for chunk_sizes in my_free."
-              << std::endl;
-    return;
-  }
-  for (Py_ssize_t i = 0; i < num_chunks; ++i) {
-    PyObject* item = PyList_GetItem(recv_p_memHandle, i);
-    PyObject* addr_py = PyTuple_GetItem(item, 0);
-    PyObject* size_py = PyTuple_GetItem(item, 1);
-    p_memHandle[i] =
-        (CUmemGenericAllocationHandle*)PyLong_AsUnsignedLongLong(addr_py);
-    chunk_sizes[i] = (unsigned long long)PyLong_AsUnsignedLongLong(size_py);
-  }
-
-  // Drop temporary Python refs, then release the GIL before calling into
-  // non-Python APIs.
-  Py_DECREF(py_ptr);
-  Py_DECREF(py_result);
-  PyGILState_Release(gstate);
-
-  unmap_and_release(device, size, d_mem, p_memHandle, chunk_sizes, num_chunks);
-#else
-  // Non-ROCm path: simple integer handle already extracted; drop temporary
-  // Python refs while still holding the GIL, then release it.
-  Py_DECREF(py_ptr);
-  Py_DECREF(py_result);
-  PyGILState_Release(gstate);
-
   CUmemGenericAllocationHandle* p_memHandle =
       (CUmemGenericAllocationHandle*)recv_p_memHandle;
   unmap_and_release(device, size, d_mem, p_memHandle);
-#endif
 
   // free address and the handle
   CUDA_CHECK(cuMemAddressFree(d_mem, size));
-#ifndef USE_ROCM
-  free(p_memHandle);
-#else
-  for (auto i = 0; i < num_chunks; ++i) {
-    free(p_memHandle[i]);
+  if (error_code != 0) {
+    return;
   }
   free(p_memHandle);
-  free(chunk_sizes);
-#endif
 }
 
 // ---------------------------------------------------------------------------
@@ -541,87 +271,19 @@ static PyObject* python_unmap_and_release(PyObject* self, PyObject* args) {
   }
 
   unsigned long long recv_device, recv_size;
-  unsigned long long recv_d_mem;
-#ifndef USE_ROCM
-  unsigned long long recv_p_memHandle;
-#else
-  PyObject* recv_p_memHandle;
-#endif
+  unsigned long long recv_d_mem, recv_p_memHandle;
   // Unpack the tuple into four C integers
-  if (!PyArg_ParseTuple(args, PYARGS_PARSE, &recv_device, &recv_size,
-                        &recv_d_mem, &recv_p_memHandle)) {
+  if (!PyArg_ParseTuple(args, "KKKK", &recv_device, &recv_size, &recv_d_mem,
+                        &recv_p_memHandle)) {
     // PyArg_ParseTuple sets an error if it fails
     return nullptr;
   }
 
   CUdeviceptr d_mem_ptr = (CUdeviceptr)recv_d_mem;
-#ifndef USE_ROCM
   CUmemGenericAllocationHandle* p_memHandle =
       (CUmemGenericAllocationHandle*)recv_p_memHandle;
 
   unmap_and_release(recv_device, recv_size, d_mem_ptr, p_memHandle);
-#else
-  if (!PyList_Check(recv_p_memHandle)) {
-    PyErr_SetString(PyExc_TypeError,
-                    "Expected a list for the 4th argument on ROCm");
-    return nullptr;
-  }
-  Py_ssize_t num_chunks = PyList_Size(recv_p_memHandle);
-  if (num_chunks < 0) {
-    return nullptr;  // PyList_Size sets an exception on error.
-  }
-  CUmemGenericAllocationHandle** p_memHandle =
-      (CUmemGenericAllocationHandle**)malloc(
-          num_chunks * sizeof(CUmemGenericAllocationHandle*));
-  if (p_memHandle == nullptr) {
-    PyErr_SetString(PyExc_MemoryError, "malloc failed for p_memHandle");
-    return nullptr;
-  }
-  unsigned long long* chunk_sizes =
-      (unsigned long long*)malloc(num_chunks * sizeof(unsigned long long));
-  if (chunk_sizes == nullptr) {
-    free(p_memHandle);
-    PyErr_SetString(PyExc_MemoryError, "malloc failed for chunk_sizes");
-    return nullptr;
-  }
-  for (Py_ssize_t i = 0; i < num_chunks; ++i) {
-    PyObject* item = PyList_GetItem(recv_p_memHandle, i);
-    if (item == nullptr || !PyTuple_Check(item) || PyTuple_Size(item) != 2) {
-      free(p_memHandle);
-      free(chunk_sizes);
-      PyErr_SetString(
-          PyExc_TypeError,
-          "List items must be tuples of size 2 (handle_addr, size)");
-      return nullptr;
-    }
-    PyObject* addr_py = PyTuple_GetItem(item, 0);
-    PyObject* size_py = PyTuple_GetItem(item, 1);
-    if (addr_py == nullptr || size_py == nullptr) {
-      free(p_memHandle);
-      free(chunk_sizes);
-      return nullptr;  // PyTuple_GetItem sets an exception
-    }
-    p_memHandle[i] =
-        (CUmemGenericAllocationHandle*)PyLong_AsUnsignedLongLong(addr_py);
-    if (PyErr_Occurred()) {
-      free(p_memHandle);
-      free(chunk_sizes);
-      return nullptr;
-    }
-    chunk_sizes[i] = (unsigned long long)PyLong_AsUnsignedLongLong(size_py);
-    if (PyErr_Occurred()) {
-      free(p_memHandle);
-      free(chunk_sizes);
-      return nullptr;
-    }
-  }
-
-  unmap_and_release(recv_device, recv_size, d_mem_ptr, p_memHandle, chunk_sizes,
-                    num_chunks);
-
-  free(p_memHandle);
-  free(chunk_sizes);
-#endif
 
   if (error_code != 0) {
     error_code = no_error;
@@ -639,56 +301,19 @@ static PyObject* python_create_and_map(PyObject* self, PyObject* args) {
   }
 
   unsigned long long recv_device, recv_size;
-  unsigned long long recv_d_mem;
-#ifndef USE_ROCM
-  unsigned long long recv_p_memHandle;
-#else
-  PyObject* recv_p_memHandle;
-#endif
+  unsigned long long recv_d_mem, recv_p_memHandle;
   // Unpack the tuple into four C integers
-  if (!PyArg_ParseTuple(args, PYARGS_PARSE, &recv_device, &recv_size,
-                        &recv_d_mem, &recv_p_memHandle)) {
+  if (!PyArg_ParseTuple(args, "KKKK", &recv_device, &recv_size, &recv_d_mem,
+                        &recv_p_memHandle)) {
     // PyArg_ParseTuple sets an error if it fails
     return nullptr;
   }
 
   CUdeviceptr d_mem_ptr = (CUdeviceptr)recv_d_mem;
-#ifndef USE_ROCM
   CUmemGenericAllocationHandle* p_memHandle =
       (CUmemGenericAllocationHandle*)recv_p_memHandle;
 
   create_and_map(recv_device, recv_size, d_mem_ptr, p_memHandle);
-#else
-  Py_ssize_t num_chunks = PyList_Size(recv_p_memHandle);
-  CUmemGenericAllocationHandle** p_memHandle =
-      (CUmemGenericAllocationHandle**)malloc(
-          num_chunks * sizeof(CUmemGenericAllocationHandle*));
-  if (p_memHandle == nullptr) {
-    PyErr_SetString(PyExc_MemoryError, "malloc failed for p_memHandle");
-    return nullptr;
-  }
-  unsigned long long* chunk_sizes =
-      (unsigned long long*)malloc(num_chunks * sizeof(unsigned long long));
-  if (chunk_sizes == nullptr) {
-    free(p_memHandle);
-    PyErr_SetString(PyExc_MemoryError, "malloc failed for chunk_sizes");
-    return nullptr;
-  }
-  for (auto i = 0; i < num_chunks; ++i) {
-    PyObject* item = PyList_GetItem(recv_p_memHandle, i);
-    PyObject* addr_py = PyTuple_GetItem(item, 0);
-    PyObject* size_py = PyTuple_GetItem(item, 1);
-    p_memHandle[i] =
-        (CUmemGenericAllocationHandle*)PyLong_AsUnsignedLongLong(addr_py);
-    chunk_sizes[i] = PyLong_AsUnsignedLongLong(size_py);
-  }
-
-  create_and_map(recv_device, recv_size, d_mem_ptr, p_memHandle, chunk_sizes,
-                 num_chunks);
-
-  free(p_memHandle);
-  free(chunk_sizes);
-#endif
 
   if (error_code != 0) {
     error_code = no_error;

csrc/cumem_allocator_compat.h

@@ -1,109 +0,0 @@
-#pragma once
-
-#ifdef USE_ROCM
-////////////////////////////////////////
-// For compatibility with CUDA and ROCm
-////////////////////////////////////////
-  #include <hip/hip_runtime_api.h>
-
-extern "C" {
-  #ifndef CUDA_SUCCESS
-    #define CUDA_SUCCESS hipSuccess
-  #endif  // CUDA_SUCCESS
-
-// https://rocm.docs.amd.com/projects/HIPIFY/en/latest/tables/CUDA_Driver_API_functions_supported_by_HIP.html
-typedef unsigned long long CUdevice;
-typedef hipDeviceptr_t CUdeviceptr;
-typedef hipError_t CUresult;
-typedef hipCtx_t CUcontext;
-typedef hipStream_t CUstream;
-typedef hipMemGenericAllocationHandle_t CUmemGenericAllocationHandle;
-typedef hipMemAllocationGranularity_flags CUmemAllocationGranularity_flags;
-typedef hipMemAllocationProp CUmemAllocationProp;
-typedef hipMemAccessDesc CUmemAccessDesc;
-
-  #define CU_MEM_ALLOCATION_TYPE_PINNED hipMemAllocationTypePinned
-  #define CU_MEM_LOCATION_TYPE_DEVICE hipMemLocationTypeDevice
-  #define CU_MEM_ACCESS_FLAGS_PROT_READWRITE hipMemAccessFlagsProtReadWrite
-  #define CU_MEM_ALLOC_GRANULARITY_MINIMUM hipMemAllocationGranularityMinimum
-
-  // https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__TYPES.html
-  #define CU_MEM_ALLOCATION_COMP_NONE 0x0
-
-// Error Handling
-// https://docs.nvidia.com/cuda/archive/11.4.4/cuda-driver-api/group__CUDA__ERROR.html
-CUresult cuGetErrorString(CUresult hipError, const char** pStr) {
-  *pStr = hipGetErrorString(hipError);
-  return CUDA_SUCCESS;
-}
-
-// Context Management
-// https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__CTX.html
-CUresult cuCtxGetCurrent(CUcontext* ctx) {
-  // This API is deprecated on the AMD platform, only for equivalent cuCtx
-  // driver API on the NVIDIA platform.
-  return hipCtxGetCurrent(ctx);
-}
-
-CUresult cuCtxSetCurrent(CUcontext ctx) {
-  // This API is deprecated on the AMD platform, only for equivalent cuCtx
-  // driver API on the NVIDIA platform.
-  return hipCtxSetCurrent(ctx);
-}
-
-// Primary Context Management
-// https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__PRIMARY__CTX.html
-CUresult cuDevicePrimaryCtxRetain(CUcontext* ctx, CUdevice dev) {
-  return hipDevicePrimaryCtxRetain(ctx, dev);
-}
-
-// Virtual Memory Management
-// https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__VA.html
-CUresult cuMemAddressFree(CUdeviceptr ptr, size_t size) {
-  return hipMemAddressFree(ptr, size);
-}
-
-CUresult cuMemAddressReserve(CUdeviceptr* ptr, size_t size, size_t alignment,
-                             CUdeviceptr addr, unsigned long long flags) {
-  return hipMemAddressReserve(ptr, size, alignment, addr, flags);
-}
-
-CUresult cuMemCreate(CUmemGenericAllocationHandle* handle, size_t size,
-                     const CUmemAllocationProp* prop,
-                     unsigned long long flags) {
-  return hipMemCreate(handle, size, prop, flags);
-}
-
-CUresult cuMemGetAllocationGranularity(
-    size_t* granularity, const CUmemAllocationProp* prop,
-    CUmemAllocationGranularity_flags option) {
-  return hipMemGetAllocationGranularity(granularity, prop, option);
-}
-
-CUresult cuMemMap(CUdeviceptr dptr, size_t size, size_t offset,
-                  CUmemGenericAllocationHandle handle,
-                  unsigned long long flags) {
-  return hipMemMap(dptr, size, offset, handle, flags);
-}
-
-CUresult cuMemRelease(CUmemGenericAllocationHandle handle) {
-  return hipMemRelease(handle);
-}
-
-CUresult cuMemSetAccess(CUdeviceptr ptr, size_t size,
-                        const CUmemAccessDesc* desc, size_t count) {
-  return hipMemSetAccess(ptr, size, desc, count);
-}
-
-CUresult cuMemUnmap(CUdeviceptr ptr, size_t size) {
-  return hipMemUnmap(ptr, size);
-}
-}  // extern "C"
-
-#else
-////////////////////////////////////////
-// Import CUDA headers for NVIDIA GPUs
-////////////////////////////////////////
-  #include <cuda_runtime_api.h>
-  #include <cuda.h>
-#endif

csrc/dispatch_utils.h

@@ -88,53 +88,3 @@
 #define VLLM_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(TYPE, NAME, ...) \
   AT_DISPATCH_SWITCH(                                              \
       TYPE, NAME, VLLM_DISPATCH_CASE_INTEGRAL_AND_UNSIGNED_TYPES(__VA_ARGS__))
-
-#define VLLM_DISPATCH_VEC_SIZE(VEC_SIZE, ...) \
-  switch (VEC_SIZE) {                         \
-    case 16: {                                \
-      constexpr int vec_size = 16;            \
-      __VA_ARGS__();                          \
-      break;                                  \
-    }                                         \
-    case 8: {                                 \
-      constexpr int vec_size = 8;             \
-      __VA_ARGS__();                          \
-      break;                                  \
-    }                                         \
-    case 4: {                                 \
-      constexpr int vec_size = 4;             \
-      __VA_ARGS__();                          \
-      break;                                  \
-    }                                         \
-    case 2: {                                 \
-      constexpr int vec_size = 2;             \
-      __VA_ARGS__();                          \
-      break;                                  \
-    }                                         \
-    default: {                                \
-      constexpr int vec_size = 1;             \
-      __VA_ARGS__();                          \
-      break;                                  \
-    }                                         \
-  }
-
-#define VLLM_DISPATCH_RANK234(NUM_DIMS, ...)                                   \
-  switch (NUM_DIMS) {                                                          \
-    case 2: {                                                                  \
-      constexpr int tensor_rank = 2;                                           \
-      __VA_ARGS__();                                                           \
-      break;                                                                   \
-    }                                                                          \
-    case 3: {                                                                  \
-      constexpr int tensor_rank = 3;                                           \
-      __VA_ARGS__();                                                           \
-      break;                                                                   \
-    }                                                                          \
-    case 4: {                                                                  \
-      constexpr int tensor_rank = 4;                                           \
-      __VA_ARGS__();                                                           \
-      break;                                                                   \
-    }                                                                          \
-    default:                                                                   \
-      TORCH_CHECK(false, "Expects rank 2, 3 or 4 tensors but got ", NUM_DIMS); \
-  }

csrc/fused_qknorm_rope_kernel.cu

@@ -1,428 +0,0 @@
-/*
- * Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
- *
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- *     http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- */
-
-#include <cmath>
-#include <cuda_runtime.h>
-#include <type_traits>
-
-#include <torch/cuda.h>
-#include <c10/cuda/CUDAGuard.h>
-
-#include "cuda_compat.h"
-#include "dispatch_utils.h"
-#include "type_convert.cuh"
-
-#define CHECK_TYPE(x, st)                                              \
-  TORCH_CHECK(x.scalar_type() == st, #x " dtype is ", x.scalar_type(), \
-              ", while ", st, " is expected")
-#define CHECK_TH_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")
-#define CHECK_CONTIGUOUS(x) \
-  TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
-#define CHECK_INPUT(x) \
-  CHECK_TH_CUDA(x);    \
-  CHECK_CONTIGUOUS(x)
-
-#ifdef USE_ROCM
-  #define FINAL_MASK 0xffffffffffffffffULL
-
-  #if defined(HIP_VERSION) && HIP_VERSION < 70000000
-// On ROCm versions before 7.0, __syncwarp isn't defined. The below
-// implementation is copy/pasted from the implementation in ROCm 7.0
-__device__ inline void __syncwarp() {
-  __builtin_amdgcn_fence(__ATOMIC_RELEASE, "wavefront");
-  __builtin_amdgcn_wave_barrier();
-  __builtin_amdgcn_fence(__ATOMIC_ACQUIRE, "wavefront");
-}
-  #endif
-#else
-  #define FINAL_MASK 0xffffffff
-#endif
-
-namespace tensorrt_llm::common {
-template <typename T, int num>
-struct packed_as;
-// Specialization for packed_as used in this kernel.
-template <>
-struct packed_as<uint, 1> {
-  using type = uint;
-};
-
-template <>
-struct packed_as<uint, 2> {
-  using type = uint2;
-};
-
-template <>
-struct packed_as<uint, 4> {
-  using type = uint4;
-};
-
-template <typename T>
-__inline__ __device__ T warpReduceSum(T val) {
-#pragma unroll
-  for (int mask = 16; mask > 0; mask >>= 1)
-    val += __shfl_xor_sync(FINAL_MASK, val, mask, 32);
-  return val;
-}
-
-template <typename T>
-inline __device__ __host__ T divUp(T m, T n) {
-  return (m + n - 1) / n;
-}
-
-}  // namespace tensorrt_llm::common
-
-namespace tensorrt_llm::kernels {
-// NOTE(zhuhaoran): This kernel is adapted from TensorRT-LLM implementation,
-// with added support for passing the cos_sin_cache as an input.
-// https://github.com/NVIDIA/TensorRT-LLM/blob/main/cpp/tensorrt_llm/kernels/fusedQKNormRopeKernel.cu
-
-// Perform per-head QK Norm and RoPE in a single kernel.
-// scalar_t_in: data type of QKV and RMSNorm weights
-// scalar_t_cache: data type of cos/sin cache
-// head_dim: the dimension of each head
-// interleave: interleave=!is_neox.
-template <typename scalar_t_in, typename scalar_t_cache, int head_dim,
-          bool interleave>
-__global__ void fusedQKNormRopeKernel(
-    void* qkv_void,                  // Combined QKV tensor
-    int const num_heads_q,           // Number of query heads
-    int const num_heads_k,           // Number of key heads
-    int const num_heads_v,           // Number of value heads
-    float const eps,                 // Epsilon for RMS normalization
-    void const* q_weight_void,       // RMSNorm weights for query
-    void const* k_weight_void,       // RMSNorm weights for key
-    void const* cos_sin_cache_void,  // Pre-computed cos/sin cache
-    int64_t const* position_ids,     // Position IDs for RoPE
-    int const num_tokens             // Number of tokens
-) {
-#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
-  if constexpr ((std::is_same_v<scalar_t_in, c10::BFloat16>) ||
-                std::is_same_v<scalar_t_cache, c10::BFloat16>) {
-    return;
-  } else {
-#endif
-
-    using Converter = vllm::_typeConvert<scalar_t_in>;
-    static_assert(Converter::exists,
-                  "Input QKV data type is not supported for this CUDA "
-                  "architecture or toolkit version.");
-    using T_in = typename Converter::hip_type;
-    using T2_in = typename Converter::packed_hip_type;
-
-    using CacheConverter = vllm::_typeConvert<scalar_t_cache>;
-    static_assert(CacheConverter::exists,
-                  "Cache data type is not supported for this CUDA architecture "
-                  "or toolkit version.");
-    using T_cache = typename CacheConverter::hip_type;
-
-    T_in* qkv = reinterpret_cast<T_in*>(qkv_void);
-    T_in const* q_weight = reinterpret_cast<T_in const*>(q_weight_void);
-    T_in const* k_weight = reinterpret_cast<T_in const*>(k_weight_void);
-    T_cache const* cos_sin_cache =
-        reinterpret_cast<T_cache const*>(cos_sin_cache_void);
-
-    int const warpsPerBlock = blockDim.x / 32;
-    int const warpId = threadIdx.x / 32;
-    int const laneId = threadIdx.x % 32;
-
-    // Calculate global warp index to determine which head/token this warp
-    // processes
-    int const globalWarpIdx = blockIdx.x * warpsPerBlock + warpId;
-
-    // Total number of attention heads (Q and K)
-    int const total_qk_heads = num_heads_q + num_heads_k;
-
-    // Determine which token and head type (Q or K) this warp processes
-    int const tokenIdx = globalWarpIdx / total_qk_heads;
-    int const localHeadIdx = globalWarpIdx % total_qk_heads;
-
-    // Skip if this warp is assigned beyond the number of tokens
-    if (tokenIdx >= num_tokens) return;
-
-    bool const isQ = localHeadIdx < num_heads_q;
-    int const headIdx = isQ ? localHeadIdx : localHeadIdx - num_heads_q;
-
-    int const num_heads = num_heads_q + num_heads_k + num_heads_v;
-
-    static_assert(head_dim % (32 * 2) == 0,
-                  "head_dim must be divisible by 64 (each warp processes one "
-                  "head, and each thread gets even number of "
-                  "elements)");
-    constexpr int numElemsPerThread = head_dim / 32;
-    float elements[numElemsPerThread];
-    constexpr int elemSizeBytes = numElemsPerThread * sizeof(__nv_bfloat16);
-    static_assert(elemSizeBytes % 4 == 0,
-                  "numSizeBytes must be a multiple of 4");
-    constexpr int vecSize =
-        elemSizeBytes /
-        4;  // Use packed_as<uint, vecSize> to perform loading/saving.
-    using vec_T = typename tensorrt_llm::common::packed_as<uint, vecSize>::type;
-
-    int offsetWarp;  // Offset for the warp
-    if (isQ) {
-      // Q segment: token offset + head offset within Q segment
-      offsetWarp = tokenIdx * num_heads * head_dim + headIdx * head_dim;
-    } else {
-      // K segment: token offset + entire Q segment + head offset within K
-      // segment
-      offsetWarp = tokenIdx * num_heads * head_dim + num_heads_q * head_dim +
-                   headIdx * head_dim;
-    }
-    int offsetThread = offsetWarp + laneId * numElemsPerThread;
-
-    // Sum of squares for RMSNorm
-    float sumOfSquares = 0.0f;
-
-    // Load.
-    {
-      vec_T vec = *reinterpret_cast<vec_T const*>(&qkv[offsetThread]);
-      constexpr int num_packed_elems = elemSizeBytes / sizeof(T2_in);
-#pragma unroll
-      for (int i = 0; i < num_packed_elems; i++) {
-        // Interpret the generic vector chunk as the specific packed type
-        T2_in packed_val = *(reinterpret_cast<T2_in*>(&vec) + i);
-        // Convert to float2 for computation
-        float2 vals = Converter::convert(packed_val);
-        sumOfSquares += vals.x * vals.x;
-        sumOfSquares += vals.y * vals.y;
-
-        elements[2 * i] = vals.x;
-        elements[2 * i + 1] = vals.y;
-      }
-    }
-
-    // Reduce sum across warp using the utility function
-    sumOfSquares = tensorrt_llm::common::warpReduceSum(sumOfSquares);
-
-    // Compute RMS normalization factor
-    float rms_rcp = rsqrtf(sumOfSquares / static_cast<float>(head_dim) + eps);
-
-    // Normalize elements
-#pragma unroll
-    for (int i = 0; i < numElemsPerThread; i++) {
-      int dim = laneId * numElemsPerThread + i;
-      float weight = isQ ? Converter::convert(q_weight[dim])
-                         : Converter::convert(k_weight[dim]);
-      elements[i] *= rms_rcp * weight;
-    }
-
-    // Apply RoPE to normalized elements
-    float elements2[numElemsPerThread];  // Additional buffer required for RoPE.
-
-    int64_t pos_id = position_ids[tokenIdx];
-
-    // Calculate cache pointer for this position - similar to
-    // pos_encoding_kernels.cu
-    T_cache const* cache_ptr = cos_sin_cache + pos_id * head_dim;
-    int const embed_dim = head_dim / 2;
-    T_cache const* cos_ptr = cache_ptr;
-    T_cache const* sin_ptr = cache_ptr + embed_dim;
-
-    if constexpr (interleave) {
-      // Perform interleaving. Use pre-computed cos/sin values.
-#pragma unroll
-      for (int i = 0; i < numElemsPerThread / 2; ++i) {
-        int const idx0 = 2 * i;
-        int const idx1 = 2 * i + 1;
-
-        float const val0 = elements[idx0];
-        float const val1 = elements[idx1];
-
-        int const dim_idx = laneId * numElemsPerThread + idx0;
-        int const half_dim = dim_idx / 2;
-        float const cos_val =
-            CacheConverter::convert(VLLM_LDG(cos_ptr + half_dim));
-        float const sin_val =
-            CacheConverter::convert(VLLM_LDG(sin_ptr + half_dim));
-
-        elements[idx0] = val0 * cos_val - val1 * sin_val;
-        elements[idx1] = val0 * sin_val + val1 * cos_val;
-      }
-    } else {
-      // Before data exchange with in warp, we need to sync.
-      __syncwarp();
-      // Get the data from the other half of the warp. Use pre-computed cos/sin
-      // values.
-#pragma unroll
-      for (int i = 0; i < numElemsPerThread; i++) {
-        elements2[i] = __shfl_xor_sync(FINAL_MASK, elements[i], 16);
-        if (laneId < 16) {
-          elements2[i] = -elements2[i];
-        }
-
-        int dim_idx = laneId * numElemsPerThread + i;
-        dim_idx = (dim_idx * 2) % head_dim;
-        int half_dim = dim_idx / 2;
-        // Use pre-computed cos/sin from cache
-        float cos_val = CacheConverter::convert(VLLM_LDG(cos_ptr + half_dim));
-        float sin_val = CacheConverter::convert(VLLM_LDG(sin_ptr + half_dim));
-
-        elements[i] = elements[i] * cos_val + elements2[i] * sin_val;
-      }
-      // __shfl_xor_sync does not provide memfence. Need to sync again.
-      __syncwarp();
-    }
-
-    // Store.
-    {
-      vec_T vec;
-      constexpr int num_packed_elems = elemSizeBytes / sizeof(T2_in);
-#pragma unroll
-      for (int i = 0; i < num_packed_elems; i++) {
-        // Convert from float2 back to the specific packed type
-        T2_in packed_val = Converter::convert(
-            make_float2(elements[2 * i], elements[2 * i + 1]));
-        // Place it into the generic vector
-        *(reinterpret_cast<T2_in*>(&vec) + i) = packed_val;
-      }
-      *reinterpret_cast<vec_T*>(&qkv[offsetThread]) = vec;
-    }
-
-#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
-  }
-#endif
-}
-
-// Borrowed from
-// https://github.com/flashinfer-ai/flashinfer/blob/8125d079a43e9a0ba463a4ed1b639cefd084cec9/include/flashinfer/pos_enc.cuh#L568
-#define DISPATCH_INTERLEAVE(interleave, INTERLEAVE, ...) \
-  if (interleave) {                                      \
-    const bool INTERLEAVE = true;                        \
-    __VA_ARGS__                                          \
-  } else {                                               \
-    const bool INTERLEAVE = false;                       \
-    __VA_ARGS__                                          \
-  }
-
-template <typename scalar_t_in, typename scalar_t_cache>
-void launchFusedQKNormRope(void* qkv, int const num_tokens,
-                           int const num_heads_q, int const num_heads_k,
-                           int const num_heads_v, int const head_dim,
-                           float const eps, void const* q_weight,
-                           void const* k_weight, void const* cos_sin_cache,
-                           bool const interleave, int64_t const* position_ids,
-                           cudaStream_t stream) {
-  constexpr int blockSize = 256;
-
-  int const warpsPerBlock = blockSize / 32;
-  int const totalQKHeads = num_heads_q + num_heads_k;
-  int const totalWarps = num_tokens * totalQKHeads;
-
-  int const gridSize = common::divUp(totalWarps, warpsPerBlock);
-  dim3 gridDim(gridSize);
-  dim3 blockDim(blockSize);
-
-  switch (head_dim) {
-    case 64:
-      DISPATCH_INTERLEAVE(interleave, INTERLEAVE, {
-        fusedQKNormRopeKernel<scalar_t_in, scalar_t_cache, 64, INTERLEAVE>
-            <<<gridDim, blockDim, 0, stream>>>(
-                qkv, num_heads_q, num_heads_k, num_heads_v, eps, q_weight,
-                k_weight, cos_sin_cache, position_ids, num_tokens);
-      });
-      break;
-    case 128:
-      DISPATCH_INTERLEAVE(interleave, INTERLEAVE, {
-        fusedQKNormRopeKernel<scalar_t_in, scalar_t_cache, 128, INTERLEAVE>
-            <<<gridDim, blockDim, 0, stream>>>(
-                qkv, num_heads_q, num_heads_k, num_heads_v, eps, q_weight,
-                k_weight, cos_sin_cache, position_ids, num_tokens);
-      });
-      break;
-    case 256:
-      DISPATCH_INTERLEAVE(interleave, INTERLEAVE, {
-        fusedQKNormRopeKernel<scalar_t_in, scalar_t_cache, 256, INTERLEAVE>
-            <<<gridDim, blockDim, 0, stream>>>(
-                qkv, num_heads_q, num_heads_k, num_heads_v, eps, q_weight,
-                k_weight, cos_sin_cache, position_ids, num_tokens);
-      });
-      break;
-    default:
-      TORCH_CHECK(false,
-                  "Unsupported head dimension for fusedQKNormRope: ", head_dim);
-  }
-}
-}  // namespace tensorrt_llm::kernels
-
-void fused_qk_norm_rope(
-    torch::Tensor& qkv,       // Combined QKV tensor [num_tokens,
-                              // (num_heads_q+num_heads_k+num_heads_v)*head_dim]
-    int64_t num_heads_q,      // Number of query heads
-    int64_t num_heads_k,      // Number of key heads
-    int64_t num_heads_v,      // Number of value heads
-    int64_t head_dim,         // Dimension per head
-    double eps,               // Epsilon for RMS normalization
-    torch::Tensor& q_weight,  // RMSNorm weights for query [head_dim]
-    torch::Tensor& k_weight,  // RMSNorm weights for key [head_dim]
-    torch::Tensor& cos_sin_cache,  // Cos/sin cache [max_position, head_dim]
-    bool is_neox,                  // Whether RoPE is applied in Neox style
-    torch::Tensor& position_ids    // Position IDs for RoPE [num_tokens]
-) {
-  // Input validation
-  CHECK_INPUT(qkv);
-  CHECK_INPUT(position_ids);
-  CHECK_INPUT(q_weight);
-  CHECK_INPUT(k_weight);
-  CHECK_INPUT(cos_sin_cache);
-  CHECK_TYPE(position_ids, torch::kInt64);
-
-  TORCH_CHECK(qkv.dim() == 2,
-              "QKV tensor must be 2D: [num_tokens, "
-              "(num_heads_q+num_heads_k+num_heads_v)*head_dim]");
-  TORCH_CHECK(position_ids.dim() == 1, "Position IDs must be 1D: [num_tokens]");
-  TORCH_CHECK(q_weight.dim() == 1, "Query weights must be 1D: [head_dim]");
-  TORCH_CHECK(k_weight.dim() == 1, "Key weights must be 1D: [head_dim]");
-  TORCH_CHECK(cos_sin_cache.dim() == 2,
-              "Cos/sin cache must be 2D: [max_position, head_dim]");
-  TORCH_CHECK(q_weight.size(0) == head_dim,
-              "Query weights size must match head dimension");
-  TORCH_CHECK(k_weight.size(0) == head_dim,
-              "Key weights size must match head dimension");
-  TORCH_CHECK(cos_sin_cache.size(1) == head_dim,
-              "Cos/sin cache dimension must match head_dim");
-  TORCH_CHECK(qkv.scalar_type() == q_weight.scalar_type() &&
-                  qkv.scalar_type() == k_weight.scalar_type(),
-              "qkv, q_weight and k_weight must have the same dtype");
-
-  int64_t num_tokens = qkv.size(0);
-  TORCH_CHECK(position_ids.size(0) == num_tokens,
-              "Number of tokens in position_ids must match QKV");
-
-  int64_t total_heads = num_heads_q + num_heads_k + num_heads_v;
-  TORCH_CHECK(
-      qkv.size(1) == total_heads * head_dim,
-      "QKV tensor size must match total number of heads and head dimension");
-
-  auto stream = at::cuda::getCurrentCUDAStream(qkv.get_device());
-
-  VLLM_DISPATCH_HALF_TYPES(qkv.scalar_type(), "fused_qk_norm_rope_kernel", [&] {
-    using qkv_scalar_t = scalar_t;
-    VLLM_DISPATCH_FLOATING_TYPES(
-        cos_sin_cache.scalar_type(), "fused_qk_norm_rope_kernel", [&] {
-          using cache_scalar_t = scalar_t;
-          tensorrt_llm::kernels::launchFusedQKNormRope<qkv_scalar_t,
-                                                       cache_scalar_t>(
-              qkv.data_ptr(), static_cast<int>(num_tokens),
-              static_cast<int>(num_heads_q), static_cast<int>(num_heads_k),
-              static_cast<int>(num_heads_v), static_cast<int>(head_dim),
-              static_cast<float>(eps), q_weight.data_ptr(), k_weight.data_ptr(),
-              cos_sin_cache.data_ptr(), !is_neox,
-              reinterpret_cast<int64_t const*>(position_ids.data_ptr()),
-              stream);
-        });
-  });
-}

csrc/layernorm_kernels.cu

@@ -10,39 +10,18 @@
 namespace vllm {
 
 // TODO(woosuk): Further optimize this kernel.
-template <typename scalar_t, int VEC_SIZE, int NUM_DIMS>
+template <typename scalar_t>
 __global__ void rms_norm_kernel(
-    scalar_t* __restrict__ out,           // [..., hidden_size]
-    const scalar_t* __restrict__ input,   // [..., hidden_size]
-    const int64_t input_stride_d2,        // input.stride(-2)
-    const int64_t input_stride_d3,        // input.stride(-3)
-    const int64_t input_stride_d4,        // input.stride(-4)
-    const int64_t input_shape_d2,         // input.size(-2)
-    const int64_t input_shape_d3,         // input.size(-3)
+    scalar_t* __restrict__ out,          // [..., hidden_size]
+    const scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int64_t input_stride,
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
   __shared__ float s_variance;
   float variance = 0.0f;
-  const scalar_t* input_row;
-  if constexpr (NUM_DIMS == 2) {
-    // 2D for layernorm normal case [batch_size, hidden]
-    input_row = input + blockIdx.x * input_stride_d2;
-  } else if constexpr (NUM_DIMS == 3) {
-    // 3D for q/k norm [batch_size, num_heads, head_size]
-    int batch_idx = blockIdx.x / input_shape_d2;
-    int head_idx = blockIdx.x % input_shape_d2;
-    input_row =
-        input + batch_idx * input_stride_d3 + head_idx * input_stride_d2;
-  } else if constexpr (NUM_DIMS == 4) {
-    // 4D for transformers model_impl qk norm [batch, seq, head, head_dim]
-    int batch_idx = blockIdx.x / (input_shape_d3 * input_shape_d2);
-    int remaining = blockIdx.x % (input_shape_d3 * input_shape_d2);
-    int seq_idx = remaining / input_shape_d2;
-    int head_idx = remaining % input_shape_d2;
-    input_row = input + batch_idx * input_stride_d4 +
-                seq_idx * input_stride_d3 + head_idx * input_stride_d2;
-  }
+  const scalar_t* input_row = input + blockIdx.x * input_stride;
 
+  constexpr int VEC_SIZE = 8;
   auto vec_op = [&variance](const vec_n_t<scalar_t, VEC_SIZE>& vec) {
 #pragma unroll
     for (int i = 0; i < VEC_SIZE; ++i) {
@@ -66,20 +45,10 @@ __global__ void rms_norm_kernel(
   }
   __syncthreads();
 
-  scalar_t* out_row = out + blockIdx.x * hidden_size;
-  auto* v_in = reinterpret_cast<const vec_n_t<scalar_t, VEC_SIZE>*>(input_row);
-  auto* v_w = reinterpret_cast<const vec_n_t<scalar_t, VEC_SIZE>*>(weight);
-  auto* v_out = reinterpret_cast<vec_n_t<scalar_t, VEC_SIZE>*>(out_row);
-  for (int i = threadIdx.x; i < hidden_size / VEC_SIZE; i += blockDim.x) {
-    vec_n_t<scalar_t, VEC_SIZE> dst;
-    vec_n_t<scalar_t, VEC_SIZE> src1 = v_in[i];
-    vec_n_t<scalar_t, VEC_SIZE> src2 = v_w[i];
-#pragma unroll
-    for (int j = 0; j < VEC_SIZE; j++) {
-      float x = static_cast<float>(src1.val[j]);
-      dst.val[j] = ((scalar_t)(x * s_variance)) * src2.val[j];
-    }
-    v_out[i] = dst;
+  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
+    float x = (float)input[blockIdx.x * input_stride + idx];
+    out[blockIdx.x * hidden_size + idx] =
+        ((scalar_t)(x * s_variance)) * weight[idx];
   }
 }
 
@@ -186,44 +155,30 @@ void rms_norm(torch::Tensor& out,     // [..., hidden_size]
               torch::Tensor& weight,  // [hidden_size]
               double epsilon) {
   TORCH_CHECK(out.is_contiguous());
-  if (input.stride(-1) != 1) {
-    input = input.contiguous();
-  }
   TORCH_CHECK(input.stride(-1) == 1);
   TORCH_CHECK(weight.is_contiguous());
 
   int hidden_size = input.size(-1);
 
-  int num_tokens = input.numel() / hidden_size;
-  int num_dims = input.dim();
-  int64_t input_stride_d2 = input.stride(-2);
-  int64_t input_stride_d3 = (num_dims >= 3) ? input.stride(-3) : 0;
-  int64_t input_stride_d4 = (num_dims >= 4) ? input.stride(-4) : 0;
-  int64_t input_shape_d2 = (num_dims >= 3) ? input.size(-2) : 0;
-  int64_t input_shape_d3 = (num_dims >= 4) ? input.size(-3) : 0;
+  // We cannot just use `input.stride(-2)` if the tensor is not row-major.
+  // Instead, we use a 2d view to get the second-innermost stride.
+  // That way the dimensions (except the last one) can be arbitrarily permuted.
+  torch::Tensor input_view = input.view({-1, hidden_size});
+
+  int num_tokens = input_view.numel() / hidden_size;
+  int64_t input_stride = input_view.stride(-2);
 
-  // For large num_tokens, use smaller blocks to increase SM concurrency.
-  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
   dim3 grid(num_tokens);
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
+  dim3 block(std::min(hidden_size, 1024));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input_view));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  VLLM_DISPATCH_RANK234(num_dims, [&] {
-    VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "rms_norm_kernel", [&] {
-      const int calculated_vec_size =
-          std::gcd(16 / sizeof(scalar_t), hidden_size);
-      const int block_size =
-          std::min(hidden_size / calculated_vec_size, max_block_size);
-      dim3 block(block_size);
-      VLLM_DISPATCH_VEC_SIZE(calculated_vec_size, [&] {
-        vllm::rms_norm_kernel<scalar_t, vec_size, tensor_rank>
-            <<<grid, block, 0, stream>>>(
-                out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),
-                input_stride_d2, input_stride_d3, input_stride_d4,
-                input_shape_d2, input_shape_d3, weight.data_ptr<scalar_t>(),
-                epsilon, num_tokens, hidden_size);
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input_view.scalar_type(), "rms_norm_kernel", [&] {
+        vllm::rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
+            out.data_ptr<scalar_t>(), input_view.data_ptr<scalar_t>(),
+            input_stride, weight.data_ptr<scalar_t>(), epsilon, num_tokens,
+            hidden_size);
       });
-    });
-  });
 }
 
 #define LAUNCH_FUSED_ADD_RMS_NORM(width)                                    \

csrc/layernorm_quant_kernels.cu

@@ -18,7 +18,7 @@
 namespace vllm {
 
 // TODO(woosuk): Further optimize this kernel.
-template <typename scalar_t, typename fp8_type, int VEC_SIZE>
+template <typename scalar_t, typename fp8_type>
 __global__ void rms_norm_static_fp8_quant_kernel(
     fp8_type* __restrict__ out,          // [..., hidden_size]
     const scalar_t* __restrict__ input,  // [..., hidden_size]
@@ -31,6 +31,7 @@ __global__ void rms_norm_static_fp8_quant_kernel(
 
   const scalar_t* input_row = input + blockIdx.x * input_stride;
 
+  constexpr int VEC_SIZE = 8;
   auto vec_op = [&variance](const vec_n_t<scalar_t, VEC_SIZE>& vec) {
 #pragma unroll
     for (int i = 0; i < VEC_SIZE; ++i) {
@@ -57,18 +58,11 @@ __global__ void rms_norm_static_fp8_quant_kernel(
   // invert scale to avoid division
   float const scale_inv = 1.0f / *scale;
 
-  auto* v_in = reinterpret_cast<const vec_n_t<scalar_t, VEC_SIZE>*>(input_row);
-  auto* v_w = reinterpret_cast<const vec_n_t<scalar_t, VEC_SIZE>*>(weight);
-  for (int idx = threadIdx.x; idx < hidden_size / VEC_SIZE; idx += blockDim.x) {
-    vec_n_t<scalar_t, VEC_SIZE> src1 = v_in[idx];
-    vec_n_t<scalar_t, VEC_SIZE> src2 = v_w[idx];
-#pragma unroll
-    for (int j = 0; j < VEC_SIZE; j++) {
-      float x = static_cast<float>(src1.val[j]);
-      float const out_norm = ((scalar_t)(x * s_variance)) * src2.val[j];
-      out[blockIdx.x * hidden_size + idx * VEC_SIZE + j] =
-          scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
-    }
+  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
+    float x = (float)input[blockIdx.x * input_stride + idx];
+    float const out_norm = ((scalar_t)(x * s_variance)) * weight[idx];
+    out[blockIdx.x * hidden_size + idx] =
+        scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
   }
 }
 
@@ -194,29 +188,20 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
   int input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
-  // For large num_tokens, use smaller blocks to increase SM concurrency.
-  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
   dim3 grid(num_tokens);
+  dim3 block(std::min(hidden_size, 1024));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   VLLM_DISPATCH_FLOATING_TYPES(
       input.scalar_type(), "rms_norm_kernel_scalar_type", [&] {
         VLLM_DISPATCH_FP8_TYPES(
             out.scalar_type(), "rms_norm_kernel_fp8_type", [&] {
-              const int calculated_vec_size =
-                  std::gcd(16 / sizeof(scalar_t), hidden_size);
-              const int block_size =
-                  std::min(hidden_size / calculated_vec_size, max_block_size);
-              dim3 block(block_size);
-              VLLM_DISPATCH_VEC_SIZE(calculated_vec_size, [&] {
-                vllm::rms_norm_static_fp8_quant_kernel<scalar_t, fp8_t,
-                                                       vec_size>
-                    <<<grid, block, 0, stream>>>(
-                        out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),
-                        input_stride, weight.data_ptr<scalar_t>(),
-                        scale.data_ptr<float>(), epsilon, num_tokens,
-                        hidden_size);
-              });
+              vllm::rms_norm_static_fp8_quant_kernel<scalar_t, fp8_t>
+                  <<<grid, block, 0, stream>>>(
+                      out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),
+                      input_stride, weight.data_ptr<scalar_t>(),
+                      scale.data_ptr<float>(), epsilon, num_tokens,
+                      hidden_size);
             });
       });
 }

csrc/moe/grouped_topk_kernels.cu

@@ -427,29 +427,11 @@ __device__ inline bool is_finite(const T val) {
 #endif
 }
 
-// Scoring function enums
-enum ScoringFunc {
-  SCORING_NONE = 0,    // no activation function
-  SCORING_SIGMOID = 1  // apply sigmoid
-};
-
-// Efficient sigmoid approximation from TensorRT-LLM
-__device__ inline float sigmoid_accurate(float x) {
-  return 0.5f * tanhf(0.5f * x) + 0.5f;
-}
-
 template <typename T>
-__device__ inline T apply_sigmoid(T val) {
-  float f = cuda_cast<float, T>(val);
-  return cuda_cast<T, float>(sigmoid_accurate(f));
-}
-
-template <typename T>
-__device__ void topk_with_k2(T* output, T const* input, T const* bias,
+__device__ void topk_with_k2(T* output, T const* input,
                              cg::thread_block_tile<32> const& tile,
                              int32_t const lane_id,
-                             int const num_experts_per_group,
-                             int const scoring_func) {
+                             int const num_experts_per_group) {
   // Get the top2 per thread
   T largest = neg_inf<T>();
   T second_largest = neg_inf<T>();
@@ -457,12 +439,6 @@ __device__ void topk_with_k2(T* output, T const* input, T const* bias,
   if (num_experts_per_group > WARP_SIZE) {
     for (int i = lane_id; i < num_experts_per_group; i += WARP_SIZE) {
       T value = input[i];
-      // Apply scoring function if needed
-      if (scoring_func == SCORING_SIGMOID) {
-        value = apply_sigmoid(value);
-      }
-      value = value + bias[i];
-
       if (value > largest) {
         second_largest = largest;
         largest = value;
@@ -472,13 +448,7 @@ __device__ void topk_with_k2(T* output, T const* input, T const* bias,
     }
   } else {
     for (int i = lane_id; i < num_experts_per_group; i += WARP_SIZE) {
-      T value = input[i];
-      // Apply scoring function if needed
-      if (scoring_func == SCORING_SIGMOID) {
-        value = apply_sigmoid(value);
-      }
-      value = value + bias[i];
-      largest = value;
+      largest = input[i];
     }
   }
 
@@ -502,21 +472,17 @@ __device__ void topk_with_k2(T* output, T const* input, T const* bias,
 }
 
 template <typename T>
-__global__ void topk_with_k2_kernel(T* output, T* input, T const* bias,
+__global__ void topk_with_k2_kernel(T* output, T* input,
                                     int64_t const num_tokens,
                                     int64_t const num_cases,
                                     int64_t const n_group,
-                                    int64_t const num_experts_per_group,
-                                    int const scoring_func) {
+                                    int64_t const num_experts_per_group) {
   int32_t warp_id = threadIdx.x / WARP_SIZE;
   int32_t lane_id = threadIdx.x % WARP_SIZE;
 
   int32_t case_id = blockIdx.x * NUM_WARPS_PER_BLOCK + warp_id;
   if (case_id < num_cases) {
     input += case_id * num_experts_per_group;
-    // bias is per expert group, offset to current group
-    int32_t group_id = case_id % n_group;
-    T const* group_bias = bias + group_id * num_experts_per_group;
     output += case_id;
 
     cg::thread_block block = cg::this_thread_block();
@@ -525,8 +491,7 @@ __global__ void topk_with_k2_kernel(T* output, T* input, T const* bias,
 #if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
     asm volatile("griddepcontrol.wait;");
 #endif
-    topk_with_k2(output, input, group_bias, tile, lane_id,
-                 num_experts_per_group, scoring_func);
+    topk_with_k2(output, input, tile, lane_id, num_experts_per_group);
   }
 #if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
   asm volatile("griddepcontrol.launch_dependents;");
@@ -535,15 +500,16 @@ __global__ void topk_with_k2_kernel(T* output, T* input, T const* bias,
 
 template <typename T, typename IdxT>
 __global__ void group_idx_and_topk_idx_kernel(
-    T* scores, T const* group_scores, float* topk_values, IdxT* topk_indices,
-    T const* bias, int64_t const num_tokens, int64_t const n_group,
+    T* scores, T const* group_scores, T* topk_values, IdxT* topk_indices,
+    T* scores_with_bias, int64_t const num_tokens, int64_t const n_group,
     int64_t const topk_group, int64_t const topk, int64_t const num_experts,
     int64_t const num_experts_per_group, bool renormalize,
-    double routed_scaling_factor, int scoring_func) {
+    double routed_scaling_factor) {
   int32_t warp_id = threadIdx.x / WARP_SIZE;
   int32_t lane_id = threadIdx.x % WARP_SIZE;
   int32_t case_id =
       blockIdx.x * NUM_WARPS_PER_BLOCK + warp_id;  // one per token
+  scores_with_bias += case_id * num_experts;
   scores += case_id * num_experts;
   group_scores += case_id * n_group;
   topk_values += case_id * topk;
@@ -611,16 +577,10 @@ __global__ void group_idx_and_topk_idx_kernel(
         int32_t offset = i_group * num_experts_per_group;
         for (int32_t i = lane_id; i < align_num_experts_per_group;
              i += WARP_SIZE) {
-          T candidates = neg_inf<T>();
-          if (i < num_experts_per_group) {
-            // Apply scoring function (if any) and add bias
-            T input = scores[offset + i];
-            if (is_finite(input)) {
-              T score = (scoring_func == SCORING_SIGMOID) ? apply_sigmoid(input)
-                                                          : input;
-              candidates = score + bias[offset + i];
-            }
-          }
+          T candidates = (i < num_experts_per_group) &&
+                                 is_finite(scores_with_bias[offset + i])
+                             ? scores_with_bias[offset + i]
+                             : neg_inf<T>();
           queue.add(candidates, offset + i);
         }
         if (group_scores[i_group] == topk_group_value) {
@@ -642,12 +602,11 @@ __global__ void group_idx_and_topk_idx_kernel(
     for (int i = lane_id;
          i < warp_topk::round_up_to_multiple_of<WARP_SIZE>(topk);
          i += WARP_SIZE) {
-      T value = cuda_cast<T, float>(0.0f);
+      T value =
+          i < topk
+              ? scores[s_topk_idx[i]]
+              : cuda_cast<T, float>(0.0f);  // Load the valid value of expert
       if (i < topk) {
-        // Load the score value (without bias) for normalization
-        T input = scores[s_topk_idx[i]];
-        value =
-            (scoring_func == SCORING_SIGMOID) ? apply_sigmoid(input) : input;
         s_topk_value[i] = value;
       }
       topk_sum +=
@@ -668,12 +627,12 @@ __global__ void group_idx_and_topk_idx_kernel(
           value = cuda_cast<float, T>(s_topk_value[i]) * routed_scaling_factor;
         }
         topk_indices[i] = s_topk_idx[i];
-        topk_values[i] = value;
+        topk_values[i] = cuda_cast<T, float>(value);
       }
     } else {
       for (int i = lane_id; i < topk; i += WARP_SIZE) {
         topk_indices[i] = i;
-        topk_values[i] = 1.0f / topk;
+        topk_values[i] = cuda_cast<T, float>(1.0f / topk);
       }
     }
     // Note: when if_proceed_next_topk==false, choose the first 8 experts as the
@@ -685,12 +644,12 @@ __global__ void group_idx_and_topk_idx_kernel(
 }
 
 template <typename T, typename IdxT>
-void invokeNoAuxTc(T* scores, T* group_scores, float* topk_values,
-                   IdxT* topk_indices, T const* bias, int64_t const num_tokens,
-                   int64_t const num_experts, int64_t const n_group,
-                   int64_t const topk_group, int64_t const topk,
-                   bool const renormalize, double const routed_scaling_factor,
-                   int const scoring_func, bool enable_pdl = false,
+void invokeNoAuxTc(T* scores, T* group_scores, T* topk_values,
+                   IdxT* topk_indices, T* scores_with_bias,
+                   int64_t const num_tokens, int64_t const num_experts,
+                   int64_t const n_group, int64_t const topk_group,
+                   int64_t const topk, bool const renormalize,
+                   double const routed_scaling_factor, bool enable_pdl = false,
                    cudaStream_t const stream = 0) {
   int64_t num_cases = num_tokens * n_group;
   int64_t topk_with_k2_num_blocks = (num_cases - 1) / NUM_WARPS_PER_BLOCK + 1;
@@ -705,9 +664,8 @@ void invokeNoAuxTc(T* scores, T* group_scores, float* topk_values,
   attrs[0].val.programmaticStreamSerializationAllowed = enable_pdl;
   config.numAttrs = 1;
   config.attrs = attrs;
-  cudaLaunchKernelEx(&config, kernel_instance1, group_scores, scores, bias,
-                     num_tokens, num_cases, n_group, num_experts / n_group,
-                     scoring_func);
+  cudaLaunchKernelEx(&config, kernel_instance1, group_scores, scores_with_bias,
+                     num_tokens, num_cases, n_group, num_experts / n_group);
 
   int64_t topk_with_k_group_num_blocks =
       (num_tokens - 1) / NUM_WARPS_PER_BLOCK + 1;
@@ -724,18 +682,19 @@ void invokeNoAuxTc(T* scores, T* group_scores, float* topk_values,
   config.numAttrs = 1;
   config.attrs = attrs;
   cudaLaunchKernelEx(&config, kernel_instance2, scores, group_scores,
-                     topk_values, topk_indices, bias, num_tokens, n_group,
-                     topk_group, topk, num_experts, num_experts / n_group,
-                     renormalize, routed_scaling_factor, scoring_func);
+                     topk_values, topk_indices, scores_with_bias, num_tokens,
+                     n_group, topk_group, topk, num_experts,
+                     num_experts / n_group, renormalize, routed_scaling_factor);
 }
 
 #define INSTANTIATE_NOAUX_TC(T, IdxT)                                       \
   template void invokeNoAuxTc<T, IdxT>(                                     \
-      T * scores, T * group_scores, float* topk_values, IdxT* topk_indices, \
-      T const* bias, int64_t const num_tokens, int64_t const num_experts,   \
-      int64_t const n_group, int64_t const topk_group, int64_t const topk,  \
-      bool const renormalize, double const routed_scaling_factor,           \
-      int const scoring_func, bool enable_pdl, cudaStream_t const stream);
+      T * scores, T * group_scores, T * topk_values, IdxT * topk_indices,   \
+      T * scores_with_bias, int64_t const num_tokens,                       \
+      int64_t const num_experts, int64_t const n_group,                     \
+      int64_t const topk_group, int64_t const topk, bool const renormalize, \
+      double const routed_scaling_factor, bool enable_pdl,                  \
+      cudaStream_t const stream);
 
 INSTANTIATE_NOAUX_TC(float, int32_t);
 INSTANTIATE_NOAUX_TC(half, int32_t);
@@ -744,32 +703,28 @@ INSTANTIATE_NOAUX_TC(__nv_bfloat16, int32_t);
 }  // namespace vllm
 
 std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
-    torch::Tensor const& scores, int64_t n_group, int64_t topk_group,
-    int64_t topk, bool renormalize, double routed_scaling_factor,
-    torch::Tensor const& bias, int64_t scoring_func = 0) {
-  auto data_type = scores.scalar_type();
-  auto input_size = scores.sizes();
+    torch::Tensor const& scores, torch::Tensor const& scores_with_bias,
+    int64_t n_group, int64_t topk_group, int64_t topk, bool renormalize,
+    double routed_scaling_factor) {
+  auto data_type = scores_with_bias.scalar_type();
+  auto input_size = scores_with_bias.sizes();
   int64_t num_tokens = input_size[0];
   int64_t num_experts = input_size[1];
-  TORCH_CHECK(input_size.size() == 2, "scores must be a 2D Tensor");
+  TORCH_CHECK(input_size.size() == 2, "scores_with_bias must be a 2D Tensor");
   TORCH_CHECK(num_experts % n_group == 0,
               "num_experts should be divisible by n_group");
   TORCH_CHECK(n_group <= 32,
               "n_group should be smaller than or equal to 32 for now");
   TORCH_CHECK(topk <= 32, "topk should be smaller than or equal to 32 for now");
-  TORCH_CHECK(scoring_func == vllm::moe::SCORING_NONE ||
-                  scoring_func == vllm::moe::SCORING_SIGMOID,
-              "scoring_func must be SCORING_NONE (0) or SCORING_SIGMOID (1)");
 
   torch::Tensor group_scores = torch::empty(
       {num_tokens, n_group}, torch::dtype(data_type).device(torch::kCUDA));
-  // Always output float32 for topk_values (eliminates Python-side conversion)
   torch::Tensor topk_values = torch::empty(
-      {num_tokens, topk}, torch::dtype(torch::kFloat32).device(torch::kCUDA));
+      {num_tokens, topk}, torch::dtype(data_type).device(torch::kCUDA));
   torch::Tensor topk_indices = torch::empty(
       {num_tokens, topk}, torch::dtype(torch::kInt32).device(torch::kCUDA));
 
-  auto stream = c10::cuda::getCurrentCUDAStream(scores.get_device());
+  auto stream = c10::cuda::getCurrentCUDAStream(scores_with_bias.get_device());
 
   switch (data_type) {
     case torch::kFloat16:
@@ -777,11 +732,11 @@ std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
       vllm::moe::invokeNoAuxTc<half, int32_t>(
           reinterpret_cast<half*>(scores.mutable_data_ptr()),
           reinterpret_cast<half*>(group_scores.mutable_data_ptr()),
-          reinterpret_cast<float*>(topk_values.mutable_data_ptr()),
+          reinterpret_cast<half*>(topk_values.mutable_data_ptr()),
           reinterpret_cast<int32_t*>(topk_indices.mutable_data_ptr()),
-          reinterpret_cast<half const*>(bias.data_ptr()), num_tokens,
+          reinterpret_cast<half*>(scores_with_bias.data_ptr()), num_tokens,
           num_experts, n_group, topk_group, topk, renormalize,
-          routed_scaling_factor, static_cast<int>(scoring_func), false, stream);
+          routed_scaling_factor, false, stream);
       break;
     case torch::kFloat32:
       // Handle Float32
@@ -790,20 +745,20 @@ std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
           reinterpret_cast<float*>(group_scores.mutable_data_ptr()),
           reinterpret_cast<float*>(topk_values.mutable_data_ptr()),
           reinterpret_cast<int32_t*>(topk_indices.mutable_data_ptr()),
-          reinterpret_cast<float const*>(bias.data_ptr()), num_tokens,
+          reinterpret_cast<float*>(scores_with_bias.data_ptr()), num_tokens,
           num_experts, n_group, topk_group, topk, renormalize,
-          routed_scaling_factor, static_cast<int>(scoring_func), false, stream);
+          routed_scaling_factor, false, stream);
       break;
     case torch::kBFloat16:
       // Handle BFloat16
       vllm::moe::invokeNoAuxTc<__nv_bfloat16, int32_t>(
           reinterpret_cast<__nv_bfloat16*>(scores.mutable_data_ptr()),
           reinterpret_cast<__nv_bfloat16*>(group_scores.mutable_data_ptr()),
-          reinterpret_cast<float*>(topk_values.mutable_data_ptr()),
+          reinterpret_cast<__nv_bfloat16*>(topk_values.mutable_data_ptr()),
           reinterpret_cast<int32_t*>(topk_indices.mutable_data_ptr()),
-          reinterpret_cast<__nv_bfloat16 const*>(bias.data_ptr()), num_tokens,
-          num_experts, n_group, topk_group, topk, renormalize,
-          routed_scaling_factor, static_cast<int>(scoring_func), false, stream);
+          reinterpret_cast<__nv_bfloat16*>(scores_with_bias.data_ptr()),
+          num_tokens, num_experts, n_group, topk_group, topk, renormalize,
+          routed_scaling_factor, false, stream);
       break;
     default:
       // Handle other data types

csrc/moe/marlin_moe_wna16/.gitignore

@@ -1,2 +1 @@
-sm*_kernel_*.cu
-kernel_selector.h
+kernel_*.cu

csrc/moe/marlin_moe_wna16/generate_kernels.py

@@ -4,282 +4,134 @@ import glob
 import itertools
 import os
 import subprocess
-import sys
 
 import jinja2
 
-ARCHS = []
-SUPPORT_FP8 = False
-for arch in sys.argv[1].split(","):
-    arch = arch[: arch.index(".") + 2].replace(".", "")
-    arch = int(arch)
-    # only SM89 and SM120 fully support
-    # mma.sync.aligned.m16n8k32.row.col.f32.e4m3.e4m3.f32.
-    # SM90 and SM100 can use this PTX, but it’s simulated
-    # with FP16 MMA, so it cannot achieve any acceleration.
-    if arch in [89, 120]:
-        SUPPORT_FP8 = True
-
-FILE_HEAD_COMMENT = """
-// auto generated by generate_kernels.py
+FILE_HEAD = """
+// auto generated by generate.py
 // clang-format off
-""".lstrip()
 
-FILE_HEAD = (
-    FILE_HEAD_COMMENT
-    + """
 #include "kernel.h"
 #include "marlin_template.h"
 
 namespace MARLIN_NAMESPACE_NAME {
-"""
-)
+""".strip()
 
 TEMPLATE = (
     "template __global__ void Marlin<"
-    "{{a_type_id}}, "
-    "{{b_type_id}}, "
-    "{{c_type_id}}, "
+    "{{scalar_t}}, "
+    "{{w_type_id}}, "
     "{{s_type_id}}, "
     "{{threads}}, "
     "{{thread_m_blocks}}, "
     "{{thread_n_blocks}}, "
     "{{thread_k_blocks}}, "
-    "{{m_block_size_8}}, "
+    "{{'true' if m_block_size_8 else 'false'}}, "
     "{{stages}}, "
     "{{group_blocks}}, "
-    "{{is_zp_float}}>"
+    "{{'true' if is_zp_float else 'false'}}>"
     "( MARLIN_KERNEL_PARAMS );"
 )
 
+# int8 with zero point case (vllm::kU8) is also supported,
+# we don't add it to reduce wheel size.
+SCALAR_TYPES = [
+    "vllm::kU4",
+    "vllm::kU4B8",
+    "vllm::kU8B128",
+    "vllm::kFE4M3fn",
+    "vllm::kFE2M1f",
+]
 THREAD_CONFIGS = [(128, 128, 256), (64, 256, 256), (64, 128, 128)]
 
 THREAD_M_BLOCKS = [0.5, 1, 2, 3, 4]
-
-QUANT_CONFIGS = [
-    # AWQ-INT4
-    {
-        "b_type": "kU4",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": THREAD_M_BLOCKS,
-        "group_blocks": [-1, 2, 4, 8],
-    },
-    # GPTQ-INT4
-    {
-        "b_type": "kU4B8",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": THREAD_M_BLOCKS,
-        "group_blocks": [-1, 0, 2, 4, 8],
-    },
-    # AWQ-INT8
-    {
-        "b_type": "kU8B128",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": THREAD_M_BLOCKS,
-        "group_blocks": [-1, 0, 2, 4, 8],
-    },
-    # FP8
-    {
-        "b_type": "kFE4M3fn",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": THREAD_M_BLOCKS,
-        "group_blocks": [-1, 8],
-    },
-    # NVFP4
-    {
-        "b_type": "kFE2M1f",
-        "s_type": "kFE4M3fn",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": THREAD_M_BLOCKS,
-        "group_blocks": [1],
-    },
-    # MXFP4
-    {
-        "a_type": ["kBFloat16"],
-        "b_type": "kFE2M1f",
-        "s_type": "kFE8M0fnu",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": THREAD_M_BLOCKS,
-        "group_blocks": [2],
-    },
-    # AWQ-INT4 with INT8 activation
-    {
-        "a_type": ["kS8"],
-        "b_type": "kU4",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": [1, 2, 3, 4],
-        "group_blocks": [-1, 2, 4, 8],
-    },
-    # GPTQ-INT4 with INT8 activation
-    {
-        "a_type": ["kS8"],
-        "b_type": "kU4B8",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": [1, 2, 3, 4],
-        "group_blocks": [-1, 2, 4, 8],
-    },
-    # GPTQ-INT4 with FP8 activation
-    {
-        "a_type": ["kFE4M3fn"],
-        "b_type": "kU4B8",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": [1, 2, 3, 4],
-        "group_blocks": [-1, 2, 4, 8],
-    },
-    # AWQ-INT4 with FP8 activation
-    {
-        "a_type": ["kFE4M3fn"],
-        "b_type": "kU4",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": [1, 2, 3, 4],
-        "group_blocks": [-1, 2, 4, 8],
-    },
-    # MXFP4 with FP8 activation
-    {
-        "a_type": ["kFE4M3fn"],
-        "b_type": "kFE2M1f",
-        "c_type": ["kBFloat16"],
-        "s_type": "kFE8M0fnu",
-        "thread_configs": THREAD_CONFIGS,
-        "thread_m_blocks": [1, 2, 3, 4],
-        "group_blocks": [2],
-    },
-]
+# group_blocks:
+#   = 0 : act order case
+#   = -1 : channelwise quantization
+#   > 0 : group_size=16*group_blocks
+GROUP_BLOCKS = [0, -1, 1, 2, 4, 8]
+DTYPES = ["fp16", "bf16"]
 
 
 def remove_old_kernels():
-    for filename in glob.glob(os.path.dirname(__file__) + "/*kernel_*.cu"):
+    for filename in glob.glob(os.path.dirname(__file__) + "/kernel_*.cu"):
         subprocess.call(["rm", "-f", filename])
 
-    filename = os.path.dirname(__file__) + "/kernel_selector.h"
-    subprocess.call(["rm", "-f", filename])
-
 
 def generate_new_kernels():
-    result_dict = {}
-
-    for quant_config in QUANT_CONFIGS:
-        c_types = quant_config.get("c_type", ["kFloat16", "kBFloat16"])
-        a_types = quant_config.get("a_type", ["kFloat16", "kBFloat16"])
-        b_type = quant_config["b_type"]
-        all_group_blocks = quant_config["group_blocks"]
-        all_m_blocks = quant_config["thread_m_blocks"]
-        all_thread_configs = quant_config["thread_configs"]
-
-        for a_type, c_type in itertools.product(a_types, c_types):
-            if not SUPPORT_FP8 and a_type == "kFE4M3fn":
-                continue
-            if "16" in a_type and "16" in c_type and a_type != c_type:
-                continue
-            s_type = quant_config.get("s_type", c_type)
-            if (a_type, b_type, c_type) not in result_dict:
-                result_dict[(a_type, b_type, c_type)] = []
-
-            for group_blocks, m_blocks, thread_configs in itertools.product(
-                all_group_blocks, all_m_blocks, all_thread_configs
-            ):
-                thread_k, thread_n, threads = thread_configs
-
-                if threads == 256:
-                    # for small batch (m_blocks == 1),
-                    #     we only need (128, 128, 256)
-                    # for large batch (m_blocks > 1),
-                    #     we only need (64, 256, 256)
-                    if m_blocks <= 1 and (thread_k, thread_n) != (128, 128):
-                        continue
-                    if m_blocks > 1 and (thread_k, thread_n) != (64, 256):
-                        continue
-
-                config = {
-                    "threads": threads,
-                    "s_type": s_type,
-                    "thread_m_blocks": max(m_blocks, 1),
-                    "thread_k_blocks": thread_k // 16,
-                    "thread_n_blocks": thread_n // 16,
-                    "m_block_size_8": "true" if m_blocks == 0.5 else "false",
-                    "stages": "pipe_stages",
-                    "group_blocks": group_blocks,
-                    "is_zp_float": "false",
-                }
-
-                result_dict[(a_type, b_type, c_type)].append(config)
-
-    kernel_selector_str = FILE_HEAD_COMMENT
-
-    for (a_type, b_type, c_type), config_list in result_dict.items():
+    for scalar_type, dtype in itertools.product(SCALAR_TYPES, DTYPES):
         all_template_str_list = []
-        for config in config_list:
-            s_type = config["s_type"]
+
+        for group_blocks, m_blocks, thread_configs in itertools.product(
+            GROUP_BLOCKS, THREAD_M_BLOCKS, THREAD_CONFIGS
+        ):
+            # act order case only support gptq-int4 and gptq-int8
+            if group_blocks == 0 and scalar_type not in [
+                "vllm::kU4B8",
+                "vllm::kU8B128",
+            ]:
+                continue
+            if thread_configs[2] == 256:
+                # for small batch (m_blocks == 1), we only need (128, 128, 256)
+                # for large batch (m_blocks > 1), we only need (64, 256, 256)
+                if m_blocks <= 1 and thread_configs[0] != 128:
+                    continue
+                if m_blocks > 1 and thread_configs[0] != 64:
+                    continue
+
+            # we only support channelwise quantization and group_size == 128
+            # for fp8
+            if scalar_type == "vllm::kFE4M3fn" and group_blocks not in [-1, 8]:
+                continue
+            # nvfp4 only supports group_size == 16
+            # mxfp4 only supports group_size == 32
+            if scalar_type == "vllm::kFE2M1f" and group_blocks not in [1, 2]:
+                continue
+            # other quantization methods don't support group_size = 16
+            if scalar_type != "vllm::kFE2M1f" and group_blocks == 1:
+                continue
+
+            k_blocks = thread_configs[0] // 16
+            n_blocks = thread_configs[1] // 16
+            threads = thread_configs[2]
+
+            c_dtype = "half" if dtype == "fp16" else "nv_bfloat16"
+
+            if scalar_type == "vllm::kFE2M1f" and group_blocks == 1:
+                s_type = "vllm::kFE4M3fn"
+            elif scalar_type == "vllm::kFE2M1f" and group_blocks == 2:
+                s_type = "vllm::kFE8M0fnu"
+                if dtype == "fp16":
+                    # we cannot safely dequantize e8m0 to fp16, so skip this
+                    continue
+            elif dtype == "fp16":
+                s_type = "vllm::kFloat16"
+            elif dtype == "bf16":
+                s_type = "vllm::kBFloat16"
+
             template_str = jinja2.Template(TEMPLATE).render(
-                a_type_id=f"vllm::{a_type}.id()",
-                b_type_id=f"vllm::{b_type}.id()",
-                c_type_id=f"vllm::{c_type}.id()",
-                s_type_id=f"vllm::{s_type}.id()",
-                **config,
+                scalar_t=c_dtype,
+                w_type_id=scalar_type + ".id()",
+                s_type_id=s_type + ".id()",
+                threads=threads,
+                thread_m_blocks=max(m_blocks, 1),
+                thread_n_blocks=n_blocks,
+                thread_k_blocks=k_blocks,
+                m_block_size_8=m_blocks == 0.5,
+                stages="pipe_stages",
+                group_blocks=group_blocks,
+                is_zp_float=False,
             )
+
             all_template_str_list.append(template_str)
 
-            conditions = [
-                f"a_type == vllm::{a_type}",
-                f"b_type == vllm::{b_type}",
-                f"c_type == vllm::{c_type}",
-                f"s_type == vllm::{s_type}",
-                f"threads == {config['threads']}",
-                f"thread_m_blocks == {config['thread_m_blocks']}",
-                f"thread_n_blocks == {config['thread_n_blocks']}",
-                f"thread_k_blocks == {config['thread_k_blocks']}",
-                f"m_block_size_8 == {config['m_block_size_8']}",
-                f"group_blocks == {config['group_blocks']}",
-                f"is_zp_float == {config['is_zp_float']}",
-            ]
-            conditions = " && ".join(conditions)
-
-            if kernel_selector_str == FILE_HEAD_COMMENT:
-                kernel_selector_str += f"if ({conditions})\n  kernel = "
-            else:
-                kernel_selector_str += f"else if ({conditions})\n  kernel = "
-
-            kernel_template2 = (
-                "Marlin<{{a_type_id}}, {{b_type_id}}, {{c_type_id}}, "
-                "{{s_type_id}}, {{threads}}, {{thread_m_blocks}}, "
-                "{{thread_n_blocks}}, {{thread_k_blocks}}, "
-                "{{m_block_size_8}}, {{stages}}, {{group_blocks}}, "
-                "{{is_zp_float}}>;"
-            )
-
-            kernel_selector_str += (
-                jinja2.Template(kernel_template2).render(
-                    a_type_id=f"vllm::{a_type}.id()",
-                    b_type_id=f"vllm::{b_type}.id()",
-                    c_type_id=f"vllm::{c_type}.id()",
-                    s_type_id=f"vllm::{s_type}.id()",
-                    **config,
-                )
-                + "\n"
-            )
-
         file_content = FILE_HEAD + "\n\n"
         file_content += "\n\n".join(all_template_str_list) + "\n\n}\n"
-        if a_type == "kFE4M3fn":
-            filename = f"sm89_kernel_{a_type[1:]}_{b_type[1:]}_{c_type[1:]}.cu"
-        else:
-            filename = f"sm80_kernel_{a_type[1:]}_{b_type[1:]}_{c_type[1:]}.cu"
-
-        filename = filename.lower()
+        filename = f"kernel_{dtype}_{scalar_type[6:].lower()}.cu"
 
         with open(os.path.join(os.path.dirname(__file__), filename), "w") as f:
             f.write(file_content)
 
-    if not SUPPORT_FP8 and kernel_selector_str != FILE_HEAD_COMMENT:
-        kernel_selector_str += (
-            "else if (a_type == vllm::kFE4M3fn)\n"
-            "  TORCH_CHECK(false, "
-            '"marlin kernel with fp8 activation is not built.");'
-        )
-
-    with open(os.path.join(os.path.dirname(__file__), "kernel_selector.h"), "w") as f:
-        f.write(kernel_selector_str)
-
 
 if __name__ == "__main__":
     remove_old_kernels()

csrc/moe/marlin_moe_wna16/kernel.h

@@ -11,9 +11,8 @@
   const int4 *__restrict__ A, const int4 *__restrict__ B,                     \
       int4 *__restrict__ C, int4 *__restrict__ C_tmp,                         \
       const int4 *__restrict__ b_bias_ptr,                                    \
-      const float *__restrict__ a_scales_ptr,                                 \
       const int4 *__restrict__ scales_ptr,                                    \
-      const uint16_t *__restrict__ global_scale_ptr,                          \
+      const uint16_t *__restrict__ scale2_ptr,                                \
       const int4 *__restrict__ zp_ptr, const int *__restrict__ g_idx,         \
       const int32_t *__restrict__ sorted_token_ids_ptr,                       \
       const int32_t *__restrict__ expert_ids_ptr,                             \
@@ -21,13 +20,12 @@
       const float *__restrict__ topk_weights_ptr, int top_k,                  \
       bool mul_topk_weights, bool is_ep, int num_groups, int prob_m,          \
       int prob_n, int prob_k, int *locks, bool has_bias, bool use_atomic_add, \
-      bool use_fp32_reduce
+      bool use_fp32_reduce, int max_shared_mem
 
 namespace MARLIN_NAMESPACE_NAME {
-template <const vllm::ScalarTypeId a_type_id,  // A ScalarType id
-          const vllm::ScalarTypeId b_type_id,  // B ScalarType id
-          const vllm::ScalarTypeId c_type_id,  // C ScalarType id
-          const vllm::ScalarTypeId s_type_id,  // B_SCALE ScalarType id
+template <typename scalar_t,  // compute dtype, half or nv_float16
+          const vllm::ScalarTypeId w_type_id,  // weight ScalarType id
+          const vllm::ScalarTypeId s_type_id,  // weight scale ScalarType id
           const int threads,          // number of threads in a threadblock
           const int thread_m_blocks,  // number of 16x16 blocks in the m
                                       // dimension (batchsize) of the

csrc/moe/marlin_moe_wna16/ops.cu

@@ -37,6 +37,39 @@ __global__ void MarlinDefault(MARLIN_KERNEL_PARAMS){};
 
 using MarlinFuncPtr = void (*)(MARLIN_KERNEL_PARAMS);
 
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
+
+template <int moe_block_size>
+__global__ void permute_cols_kernel(
+    int4 const* __restrict__ a_int4_ptr, int const* __restrict__ perm_int_ptr,
+    int4* __restrict__ out_int4_ptr,
+    const int32_t* __restrict__ sorted_token_ids_ptr,
+    const int32_t* __restrict__ expert_ids_ptr,
+    const int32_t* __restrict__ num_tokens_past_padded_ptr, int size_m,
+    int size_k, int top_k) {};
+
+}  // namespace marlin
+
+torch::Tensor moe_wna16_marlin_gemm(
+    torch::Tensor& a, std::optional<torch::Tensor> c_or_none,
+    torch::Tensor& b_q_weight,
+    std::optional<torch::Tensor> const& b_bias_or_none, torch::Tensor& b_scales,
+    std::optional<torch::Tensor> const& b_zeros_or_none,
+    std::optional<torch::Tensor> const& g_idx_or_none,
+    std::optional<torch::Tensor> const& perm_or_none, torch::Tensor& workspace,
+    torch::Tensor& sorted_token_ids, torch::Tensor& expert_ids,
+    torch::Tensor& num_tokens_past_padded, torch::Tensor& topk_weights,
+    int64_t moe_block_size, int64_t top_k, bool mul_topk_weights, bool is_ep,
+    vllm::ScalarTypeId const& b_q_type_id, int64_t size_m, int64_t size_n,
+    int64_t size_k, bool is_k_full, bool use_atomic_add, bool use_fp32_reduce,
+    bool is_zp_float) {
+  TORCH_CHECK_NOT_IMPLEMENTED(false,
+                              "marlin_gemm(..) requires CUDA_ARCH >= 8.0");
+  return torch::empty({1, 1});
+}
+
+#else
+
 // For a given "a" of size [M,K] performs a permutation of the K columns based
 // on the given "perm" indices.
 template <int moe_block_size>
@@ -174,7 +207,7 @@ int get_kernel_cache_size(thread_config_t const& th_config, bool m_block_size_8,
                           int thread_m_blocks, int prob_m, int prob_n,
                           int prob_k, int num_bits, int group_size,
                           bool has_act_order, bool is_k_full, int has_zp,
-                          int is_zp_float, bool is_a_8bit) {
+                          int is_zp_float) {
   int pack_factor = 32 / num_bits;
 
   // Get B size
@@ -184,8 +217,8 @@ int get_kernel_cache_size(thread_config_t const& th_config, bool m_block_size_8,
 
   // shm size for block_sorted_ids/rd_block_sorted_ids/block_topk_weights
   // both of them requires tb_m * 4 bytes (tb_m * int32 or tb_m * float32)
-  int sh_block_meta_size = tb_m * 16;
-  int sh_a_size = pipe_stages * (tb_m * tb_k) * (is_a_8bit ? 1 : 2);
+  int sh_block_meta_size = tb_m * 4;
+  int sh_a_size = pipe_stages * (tb_m * tb_k) * 2;
   int sh_b_size = pipe_stages * (tb_k * tb_n / pack_factor) * 4;
   int sh_red_size = tb_m * (tb_n + 8) * 2;
   int sh_bias_size = tb_n * 2;
@@ -217,7 +250,7 @@ bool is_valid_config(thread_config_t const& th_config, bool m_block_size_8,
                      int thread_m_blocks, int prob_m, int prob_n, int prob_k,
                      int num_bits, int group_size, bool has_act_order,
                      bool is_k_full, int has_zp, int is_zp_float,
-                     int max_shared_mem, bool is_a_8bit) {
+                     int max_shared_mem) {
   // Sanity
   if (th_config.thread_k == -1 || th_config.thread_n == -1 ||
       th_config.num_threads == -1) {
@@ -240,34 +273,188 @@ bool is_valid_config(thread_config_t const& th_config, bool m_block_size_8,
   }
 
   // Check that pipeline fits into cache
-  int cache_size =
-      get_kernel_cache_size(th_config, m_block_size_8, thread_m_blocks, prob_m,
-                            prob_n, prob_k, num_bits, group_size, has_act_order,
-                            is_k_full, has_zp, is_zp_float, is_a_8bit);
-  return cache_size <= max_shared_mem;
+  int cache_size = get_kernel_cache_size(
+      th_config, m_block_size_8, thread_m_blocks, prob_m, prob_n, prob_k,
+      num_bits, group_size, has_act_order, is_k_full, has_zp, is_zp_float);
+  return cache_size + 512 <= max_shared_mem;
 }
 
-MarlinFuncPtr get_marlin_kernel(
-    const vllm::ScalarType a_type, const vllm::ScalarType b_type,
-    const vllm::ScalarType c_type, const vllm::ScalarType s_type,
-    int thread_m_blocks, int thread_n_blocks, int thread_k_blocks,
-    bool m_block_size_8, bool has_act_order, bool has_zp, int group_blocks,
-    int threads, bool is_zp_float) {
-  int num_bits = b_type.size_bits();
-  auto kernel = MarlinDefault;
+  #define _GET_IF(W_TYPE, THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS,   \
+                  M_BLOCK_SIZE_8, GROUP_BLOCKS, NUM_THREADS, IS_ZP_FLOAT)      \
+    else if (q_type == W_TYPE && thread_m_blocks == THREAD_M_BLOCKS &&         \
+             thread_n_blocks == THREAD_N_BLOCKS &&                             \
+             thread_k_blocks == THREAD_K_BLOCKS &&                             \
+             m_block_size_8 == M_BLOCK_SIZE_8 &&                               \
+             group_blocks == GROUP_BLOCKS && num_threads == NUM_THREADS &&     \
+             is_zp_float == IS_ZP_FLOAT) {                                     \
+      constexpr auto S_TYPE =                                                  \
+          W_TYPE == vllm::kFE2M1f                                              \
+              ? (GROUP_BLOCKS == 1 ? vllm::kFE4M3fn : vllm::kFE8M0fnu)         \
+              : (std::is_same<scalar_t, half>::value ? vllm::kFloat16          \
+                                                     : vllm::kBFloat16);       \
+      kernel = Marlin<scalar_t, W_TYPE.id(), S_TYPE.id(), NUM_THREADS,         \
+                      THREAD_M_BLOCKS, THREAD_N_BLOCKS, THREAD_K_BLOCKS,       \
+                      M_BLOCK_SIZE_8, pipe_stages, GROUP_BLOCKS, IS_ZP_FLOAT>; \
+    }
 
-#include "kernel_selector.h"
+  // COMMON: cases for (group_blocks in [-1, 2, 4, 8] and is_zp_float == false)
+  //         this is the most common cases
+  // BIGGROUP: cases for big group size (group_blocks in [-1, 8])
+  // FZP: cases for float-zero-point (is_zp_float = true)
+  // ACT: cases for act order case (group_blocks == 0)
+  // FP4: cases for nvfp4(e2m1) (group_blocks == 1)
+  #define COMMON_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)       \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, -1, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 2, NUM_THREADS, false)   \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 4, NUM_THREADS, false)   \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 8, NUM_THREADS, false)   \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)
+
+  #define COMMON_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)     \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)  \
+                                                                          \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)  \
+                                                                          \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)
+
+  #define COMMON_GET_IF(W_TYPE)            \
+    COMMON_GET_IF_M1(W_TYPE, 8, 8, 256)    \
+    COMMON_GET_IF_M1(W_TYPE, 8, 4, 128)    \
+    COMMON_GET_IF_M234(W_TYPE, 16, 4, 256) \
+    COMMON_GET_IF_M234(W_TYPE, 8, 4, 128)
+
+  #define BIGGROUP_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)     \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, -1, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 8, NUM_THREADS, false)   \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)
+
+  #define BIGGROUP_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)   \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)  \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, -1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 8, NUM_THREADS, false)
+
+  #define BIGGROUP_GET_IF(W_TYPE)            \
+    BIGGROUP_GET_IF_M1(W_TYPE, 8, 8, 256)    \
+    BIGGROUP_GET_IF_M1(W_TYPE, 8, 4, 128)    \
+    BIGGROUP_GET_IF_M234(W_TYPE, 16, 4, 256) \
+    BIGGROUP_GET_IF_M234(W_TYPE, 8, 4, 128)
+
+  #define NVFP4_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)      \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false)
+
+  #define NVFP4_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)     \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 1, NUM_THREADS, false)
+
+  #define NVFP4_GET_IF(W_TYPE)            \
+    NVFP4_GET_IF_M1(W_TYPE, 8, 8, 256)    \
+    NVFP4_GET_IF_M1(W_TYPE, 8, 4, 128)    \
+    NVFP4_GET_IF_M234(W_TYPE, 16, 4, 256) \
+    NVFP4_GET_IF_M234(W_TYPE, 8, 4, 128)
+
+  #define MXFP4_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)      \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 2, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false)
+
+  #define MXFP4_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)     \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 2, NUM_THREADS, false)
+
+  #define MXFP4_GET_IF(W_TYPE)            \
+    MXFP4_GET_IF_M1(W_TYPE, 8, 8, 256)    \
+    MXFP4_GET_IF_M1(W_TYPE, 8, 4, 128)    \
+    MXFP4_GET_IF_M234(W_TYPE, 16, 4, 256) \
+    MXFP4_GET_IF_M234(W_TYPE, 8, 4, 128)
+
+  // We currently have 4-bit models only with group_blocks == 4
+  #define FZP_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)       \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 4, NUM_THREADS, true) \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, true)
+
+  #define FZP_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)      \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, true) \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, true) \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 4, NUM_THREADS, true)
+
+  #define FZP_GET_IF(W_TYPE)            \
+    FZP_GET_IF_M1(W_TYPE, 8, 8, 256)    \
+    FZP_GET_IF_M1(W_TYPE, 8, 4, 128)    \
+    FZP_GET_IF_M234(W_TYPE, 16, 4, 256) \
+    FZP_GET_IF_M234(W_TYPE, 8, 4, 128)
+
+  // We currently have 4-bit models only with group_blocks == 4
+  #define ACT_GET_IF_M1(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)        \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, true, 0, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 1, N_BLOCKS, K_BLOCKS, false, 0, NUM_THREADS, false)
+
+  #define ACT_GET_IF_M234(W_TYPE, N_BLOCKS, K_BLOCKS, NUM_THREADS)       \
+    _GET_IF(W_TYPE, 2, N_BLOCKS, K_BLOCKS, false, 0, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 3, N_BLOCKS, K_BLOCKS, false, 0, NUM_THREADS, false) \
+    _GET_IF(W_TYPE, 4, N_BLOCKS, K_BLOCKS, false, 0, NUM_THREADS, false)
+
+  #define ACT_GET_IF(W_TYPE)            \
+    ACT_GET_IF_M1(W_TYPE, 8, 8, 256)    \
+    ACT_GET_IF_M1(W_TYPE, 8, 4, 128)    \
+    ACT_GET_IF_M234(W_TYPE, 16, 4, 256) \
+    ACT_GET_IF_M234(W_TYPE, 8, 4, 128)
+
+template <typename scalar_t>
+MarlinFuncPtr get_marlin_kernel(const vllm::ScalarType q_type,
+                                int thread_m_blocks, int thread_n_blocks,
+                                int thread_k_blocks, bool m_block_size_8,
+                                bool has_act_order, bool has_zp,
+                                int group_blocks, int num_threads,
+                                bool is_zp_float) {
+  int num_bits = q_type.size_bits();
+  auto kernel = MarlinDefault;
+  if (false) {
+  }
+
+  COMMON_GET_IF(vllm::kU4)
+  COMMON_GET_IF(vllm::kU4B8)
+  COMMON_GET_IF(vllm::kU8B128)
+
+  NVFP4_GET_IF(vllm::kFE2M1f)
+
+  BIGGROUP_GET_IF(vllm::kFE4M3fn)
+
+  ACT_GET_IF(vllm::kU4B8)
+  ACT_GET_IF(vllm::kU8B128)
+  if (std::is_same<scalar_t, nv_bfloat16>::value) {
+    if (false) {
+    }
+    MXFP4_GET_IF(vllm::kFE2M1f)
+  }
 
   return kernel;
 }
 
-exec_config_t determine_exec_config(
-    const vllm::ScalarType& a_type, const vllm::ScalarType& b_type,
-    const vllm::ScalarType& c_type, const vllm::ScalarType& s_type, int prob_m,
-    int prob_n, int prob_k, int num_experts, int top_k, int thread_m_blocks,
-    bool m_block_size_8, int num_bits, int group_size, bool has_act_order,
-    bool is_k_full, bool has_zp, bool is_zp_float, int max_shared_mem, int sms,
-    bool is_a_8bit) {
+template <typename scalar_t>
+exec_config_t determine_exec_config(const vllm::ScalarType& q_type, int prob_m,
+                                    int prob_n, int prob_k, int thread_m_blocks,
+                                    bool m_block_size_8, int num_bits,
+                                    int group_size, bool has_act_order,
+                                    bool is_k_full, bool has_zp,
+                                    bool is_zp_float, int max_shared_mem) {
   exec_config_t exec_cfg = exec_config_t{1, thread_config_t{-1, -1, -1}};
   thread_config_t* thread_configs = thread_m_blocks > 1
                                         ? large_batch_thread_configs
@@ -284,69 +471,73 @@ exec_config_t determine_exec_config(
 
     if (!is_valid_config(th_config, m_block_size_8, thread_m_blocks, prob_m,
                          prob_n, prob_k, num_bits, group_size, has_act_order,
-                         is_k_full, has_zp, is_zp_float, max_shared_mem - 512,
-                         is_a_8bit)) {
+                         is_k_full, has_zp, is_zp_float, max_shared_mem)) {
       continue;
     }
 
     int cache_size = get_kernel_cache_size(
         th_config, m_block_size_8, thread_m_blocks, prob_m, prob_n, prob_k,
-        num_bits, group_size, has_act_order, is_k_full, has_zp, is_zp_float,
-        is_a_8bit);
+        num_bits, group_size, has_act_order, is_k_full, has_zp, is_zp_float);
 
     int group_blocks = 0;
     if (!has_act_order) {
       group_blocks = group_size == -1 ? -1 : (group_size / 16);
     }
 
-    auto kernel =
-        get_marlin_kernel(a_type, b_type, c_type, s_type, thread_m_blocks,
-                          th_config.thread_n / 16, th_config.thread_k / 16,
-                          m_block_size_8, has_act_order, has_zp, group_blocks,
-                          th_config.num_threads, is_zp_float);
+    auto kernel = get_marlin_kernel<scalar_t>(
+        q_type, thread_m_blocks, th_config.thread_n / 16,
+        th_config.thread_k / 16, m_block_size_8, has_act_order, has_zp,
+        group_blocks, th_config.num_threads, is_zp_float);
 
     if (kernel == MarlinDefault) continue;
 
-    cudaFuncAttributes attr;
-    cudaFuncGetAttributes(&attr, kernel);
-    int reg_size = max(attr.numRegs, 1) * th_config.num_threads * 4;
-    int allow_count = min(device_max_reg_size / reg_size,
-                          max_shared_mem / (cache_size + 1536));
-    if (thread_m_blocks == 1)
+    if (thread_m_blocks > 1) {
+      exec_cfg = {1, th_config};
+      break;
+    } else {
+      cudaFuncAttributes attr;
+      cudaFuncGetAttributes(&attr, kernel);
+      int reg_size = max(attr.numRegs, 1) * th_config.num_threads * 4;
+      int allow_count = min(device_max_reg_size / reg_size,
+                            max_shared_mem / (cache_size + 1024));
       allow_count = max(min(allow_count, 4), 1);
-    else
-      allow_count = max(min(allow_count, 2), 1);
-
-    if (prob_n / th_config.thread_n * prob_m * top_k * 4 < sms * allow_count) {
-      allow_count =
-          max(prob_n / th_config.thread_n * prob_m * top_k * 4 / sms, 1);
+      if (allow_count > count) {
+        count = allow_count;
+        exec_cfg = {count, th_config};
+      };
     }
-
-    if (allow_count > count) {
-      count = allow_count;
-      exec_cfg = {count, th_config};
-    };
   }
 
   return exec_cfg;
 }
 
+template <typename scalar_t>
 void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
-               void* a_s, void* b_s, void* g_s, void* zp, void* g_idx,
-               void* perm, void* a_tmp, void* sorted_token_ids,
-               void* expert_ids, void* num_tokens_past_padded,
-               void* topk_weights, int moe_block_size, int num_experts,
-               int top_k, bool mul_topk_weights, bool is_ep, int prob_m,
-               int prob_n, int prob_k, void* workspace,
-               vllm::ScalarType const& a_type, vllm::ScalarType const& b_type,
-               vllm::ScalarType const& c_type, vllm::ScalarType const& s_type,
-               bool has_bias, bool has_act_order, bool is_k_full, bool has_zp,
-               int num_groups, int group_size, int dev, cudaStream_t stream,
-               int thread_k, int thread_n, int sms, int blocks_per_sm,
-               bool use_atomic_add, bool use_fp32_reduce, bool is_zp_float) {
+               void* s, void* s2, void* zp, void* g_idx, void* perm,
+               void* a_tmp, void* sorted_token_ids, void* expert_ids,
+               void* num_tokens_past_padded, void* topk_weights,
+               int moe_block_size, int top_k, bool mul_topk_weights, bool is_ep,
+               int prob_m, int prob_n, int prob_k, void* workspace,
+               vllm::ScalarType const& q_type, bool has_bias,
+               bool has_act_order, bool is_k_full, bool has_zp, int num_groups,
+               int group_size, int dev, cudaStream_t stream, int thread_k,
+               int thread_n, int sms, bool use_atomic_add, bool use_fp32_reduce,
+               bool is_zp_float) {
   int thread_m_blocks = div_ceil(moe_block_size, 16);
   bool m_block_size_8 = moe_block_size == 8;
-  bool is_a_8bit = a_type.size_bits() == 8;
+
+  if (has_zp) {
+    TORCH_CHECK(
+        q_type == vllm::kU4 || q_type == vllm::kU8,
+        "q_type must be u4 or u8 when has_zp = True. Got = ", q_type.str());
+  } else {
+    TORCH_CHECK(
+        q_type == vllm::kU4B8 || q_type == vllm::kU8B128 ||
+            q_type == vllm::kFE4M3fn || q_type == vllm::kFE2M1f,
+        "q_type must be uint4b8, uint8b128, float8_e4m3fn or float4_e2m1f when "
+        "has_zp = False. Got = ",
+        q_type.str());
+  }
 
   TORCH_CHECK(prob_m > 0 && prob_n > 0 && prob_k > 0, "Invalid MNK = [", prob_m,
               ", ", prob_n, ", ", prob_k, "]");
@@ -372,15 +563,14 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
     }
   }
 
-  int num_bits = b_type.size_bits();
+  int num_bits = q_type.size_bits();
   const int4* A_ptr = (const int4*)A;
   const int4* B_ptr = (const int4*)B;
   int4* C_ptr = (int4*)C;
   int4* C_tmp_ptr = (int4*)C_tmp;
   const int4* bias_ptr = (const int4*)b_bias;
-  const float* a_s_ptr = (const float*)a_s;
-  const int4* b_s_ptr = (const int4*)b_s;
-  const uint16_t* g_s_ptr = (const uint16_t*)g_s;
+  const int4* s_ptr = (const int4*)s;
+  const uint16_t* s2_ptr = (const uint16_t*)s2;
   const int4* zp_ptr = (const int4*)zp;
   const int* g_idx_ptr = (const int*)g_idx;
   const int* perm_ptr = (const int*)perm;
@@ -428,41 +618,22 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
                          cudaDevAttrMaxSharedMemoryPerBlockOptin, dev);
   TORCH_CHECK(max_shared_mem > 0);
 
-  int major_capability, minor_capability;
-  cudaDeviceGetAttribute(&major_capability, cudaDevAttrComputeCapabilityMajor,
-                         dev);
-  cudaDeviceGetAttribute(&minor_capability, cudaDevAttrComputeCapabilityMinor,
-                         dev);
-  TORCH_CHECK(major_capability * 10 + minor_capability >= 80,
-              "marlin kernel only support Ampere or newer GPUs.");
-  if (a_type == vllm::kFE4M3fn) {
-    TORCH_CHECK(major_capability * 10 + minor_capability >= 89,
-                "FP8 only support Ada Lovelace or newer GPUs.");
-    TORCH_CHECK(
-        major_capability * 10 + minor_capability == 89 ||
-            major_capability * 10 + minor_capability == 120,
-        "Marlin W4A8-FP8 only support SM89 or SM120 device (It is slower than "
-        "Marlin W4A16 on other devices).");
-  }
-
   // Set thread config
   exec_config_t exec_cfg;
   thread_config_t thread_tfg;
   if (thread_k != -1 && thread_n != -1) {
-    thread_tfg = thread_config_t{thread_k, thread_n, thread_k * thread_n / 64};
-    if (blocks_per_sm == -1) blocks_per_sm = 1;
-    exec_cfg = exec_config_t{blocks_per_sm, thread_tfg};
+    thread_tfg = thread_config_t{thread_k, thread_n, default_threads};
+    exec_cfg = exec_config_t{1, thread_tfg};
     TORCH_CHECK(prob_n % thread_n == 0, "prob_n = ", prob_n,
                 " is not divisible by thread_n = ", thread_n);
     TORCH_CHECK(prob_k % thread_k == 0, "prob_k = ", prob_k,
                 " is not divisible by thread_k = ", thread_k);
   } else {
     // Auto config
-    exec_cfg = determine_exec_config(
-        a_type, b_type, c_type, s_type, prob_m, prob_n, prob_k, num_experts,
-        top_k, thread_m_blocks, m_block_size_8, num_bits, group_size,
-        has_act_order, is_k_full, has_zp, is_zp_float, max_shared_mem, sms,
-        is_a_8bit);
+    exec_cfg = determine_exec_config<scalar_t>(
+        q_type, prob_m, prob_n, prob_k, thread_m_blocks, m_block_size_8,
+        num_bits, group_size, has_act_order, is_k_full, has_zp, is_zp_float,
+        max_shared_mem);
     thread_tfg = exec_cfg.tb_cfg;
   }
 
@@ -476,29 +647,22 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
   int thread_k_blocks = thread_k / 16;
   int thread_n_blocks = thread_n / 16;
 
-  TORCH_CHECK(is_valid_config(thread_tfg, m_block_size_8, thread_m_blocks,
-                              prob_m, prob_n, prob_k, num_bits, group_size,
-                              has_act_order, is_k_full, has_zp, is_zp_float,
-                              max_shared_mem, is_a_8bit),
-              "Invalid thread config: thread_m_blocks = ", thread_m_blocks,
-              ", thread_k = ", thread_tfg.thread_k,
-              ", thread_n = ", thread_tfg.thread_n,
-              ", num_threads = ", thread_tfg.num_threads, " for MKN = [",
-              prob_m, ", ", prob_k, ", ", prob_n, "] and num_bits = ", num_bits,
-              ", group_size = ", group_size,
-              ", has_act_order = ", has_act_order, ", is_k_full = ", is_k_full,
-              ", has_zp = ", has_zp, ", is_zp_float = ", is_zp_float,
-              ", max_shared_mem = ", max_shared_mem);
+  TORCH_CHECK(
+      is_valid_config(thread_tfg, m_block_size_8, thread_m_blocks, prob_m,
+                      prob_n, prob_k, num_bits, group_size, has_act_order,
+                      is_k_full, has_zp, is_zp_float, max_shared_mem),
+      "Invalid thread config: thread_m_blocks = ", thread_m_blocks,
+      ", thread_k = ", thread_tfg.thread_k,
+      ", thread_n = ", thread_tfg.thread_n,
+      ", num_threads = ", thread_tfg.num_threads, " for MKN = [", prob_m, ", ",
+      prob_k, ", ", prob_n, "] and num_bits = ", num_bits,
+      ", group_size = ", group_size, ", has_act_order = ", has_act_order,
+      ", is_k_full = ", is_k_full, ", has_zp = ", has_zp,
+      ", is_zp_float = ", is_zp_float, ", max_shared_mem = ", max_shared_mem);
 
-  int sh_cache_size =
-      get_kernel_cache_size(thread_tfg, m_block_size_8, thread_m_blocks, prob_m,
-                            prob_n, prob_k, num_bits, group_size, has_act_order,
-                            is_k_full, has_zp, is_zp_float, is_a_8bit);
-
-  auto kernel = get_marlin_kernel(
-      a_type, b_type, c_type, s_type, thread_m_blocks, thread_n_blocks,
-      thread_k_blocks, m_block_size_8, has_act_order, has_zp, group_blocks,
-      num_threads, is_zp_float);
+  auto kernel = get_marlin_kernel<scalar_t>(
+      q_type, thread_m_blocks, thread_n_blocks, thread_k_blocks, m_block_size_8,
+      has_act_order, has_zp, group_blocks, num_threads, is_zp_float);
 
   if (kernel == MarlinDefault) {
     TORCH_CHECK(false, "Unsupported shapes: MNK = [", prob_m, ", ", prob_n,
@@ -515,20 +679,19 @@ void marlin_mm(const void* A, const void* B, void* C, void* C_tmp, void* b_bias,
   // avoid ">>>" being formatted to "> > >"
   // clang-format off
   kernel<<<blocks, num_threads, max_shared_mem, stream>>>(
-      A_ptr, B_ptr, C_ptr, C_tmp_ptr, bias_ptr, a_s_ptr, b_s_ptr, g_s_ptr, zp_ptr, g_idx_ptr,
+      A_ptr, B_ptr, C_ptr, C_tmp_ptr, bias_ptr, s_ptr, s2_ptr, zp_ptr, g_idx_ptr,
       sorted_token_ids_ptr, expert_ids_ptr, num_tokens_past_padded_ptr,
       topk_weights_ptr, top_k, mul_topk_weights, is_ep, num_groups, prob_m,
-      prob_n, prob_k, locks, has_bias, use_atomic_add, use_fp32_reduce);
+      prob_n, prob_k, locks, has_bias, use_atomic_add, use_fp32_reduce, max_shared_mem);
   // clang-format on
 }
 
 }  // namespace MARLIN_NAMESPACE_NAME
 
 torch::Tensor moe_wna16_marlin_gemm(
-    torch::Tensor& a, std::optional<torch::Tensor> c_or_none,
+    torch::Tensor& a, std::optional<torch::Tensor> const& c_or_none,
     torch::Tensor& b_q_weight,
     std::optional<torch::Tensor> const& b_bias_or_none, torch::Tensor& b_scales,
-    std::optional<torch::Tensor> const& a_scales_or_none,
     std::optional<torch::Tensor> const& global_scale_or_none,
     std::optional<torch::Tensor> const& b_zeros_or_none,
     std::optional<torch::Tensor> const& g_idx_or_none,
@@ -536,70 +699,11 @@ torch::Tensor moe_wna16_marlin_gemm(
     torch::Tensor& sorted_token_ids, torch::Tensor& expert_ids,
     torch::Tensor& num_tokens_past_padded, torch::Tensor& topk_weights,
     int64_t moe_block_size, int64_t top_k, bool mul_topk_weights, bool is_ep,
-    vllm::ScalarTypeId const& b_type_id, int64_t size_m, int64_t size_n,
+    vllm::ScalarTypeId const& b_q_type_id, int64_t size_m, int64_t size_n,
     int64_t size_k, bool is_k_full, bool use_atomic_add, bool use_fp32_reduce,
-    bool is_zp_float, int64_t thread_k, int64_t thread_n,
-    int64_t blocks_per_sm) {
-  vllm::ScalarTypeId a_type_id, c_type_id, s_type_id;
-
-  auto c_dtype = a.dtype();
-  if (a.scalar_type() == at::ScalarType::Half) {
-    a_type_id = vllm::kFloat16.id();
-    c_type_id = vllm::kFloat16.id();
-  } else if (a.scalar_type() == at::ScalarType::BFloat16) {
-    a_type_id = vllm::kBFloat16.id();
-    c_type_id = vllm::kBFloat16.id();
-  } else {
-    c_dtype = b_scales.dtype();
-    if (b_scales.scalar_type() == at::ScalarType::Half) {
-      c_type_id = vllm::kFloat16.id();
-    } else if (b_scales.scalar_type() == at::ScalarType::BFloat16) {
-      c_type_id = vllm::kBFloat16.id();
-    } else {
-      c_type_id = vllm::kBFloat16.id();
-
-      TORCH_CHECK(c_or_none.has_value(), "c must be passed for W4A8-FP4");
-      torch::Tensor c = c_or_none.value();
-      c_dtype = c.dtype();
-
-      if (c.scalar_type() == at::ScalarType::Half) {
-        c_type_id = vllm::kFloat16.id();
-      } else if (c.scalar_type() == at::ScalarType::BFloat16) {
-        c_type_id = vllm::kBFloat16.id();
-      } else {
-        TORCH_CHECK(false, "unsupported c dtype");
-      }
-    }
-
-    if (a.scalar_type() == at::ScalarType::Float8_e4m3fn) {
-      a_type_id = vllm::kFE4M3fn.id();
-    } else if (a.scalar_type() == at::ScalarType::Char) {
-      a_type_id = vllm::kS8.id();
-    } else {
-      TORCH_CHECK(false, "unsupported `a` scalar_type");
-    }
-  }
-
-  s_type_id = c_type_id;
-  if (b_type_id == vllm::kFE2M1f.id()) {
-    if (b_scales.scalar_type() == at::ScalarType::Float8_e4m3fn) {
-      s_type_id = vllm::kFE4M3fn.id();
-    } else if (b_scales.scalar_type() == at::ScalarType::Float8_e8m0fnu) {
-      s_type_id = vllm::kFE8M0fnu.id();
-    } else {
-      TORCH_CHECK(false,
-                  "When b_type = float4_e2m1f, b_scale scalar type must be",
-                  "float8_e4m3fn (for NVFP4) or float8_e8m0fnu (for MXFP4).");
-    }
-  }
-
-  vllm::ScalarType a_type = vllm::ScalarType::from_id(a_type_id);
-  vllm::ScalarType b_type = vllm::ScalarType::from_id(b_type_id);
-  vllm::ScalarType c_type = vllm::ScalarType::from_id(c_type_id);
-  vllm::ScalarType s_type = vllm::ScalarType::from_id(s_type_id);
-
-  int pack_factor = 32 / b_type.size_bits();
-  int num_experts = b_q_weight.size(0);
+    bool is_zp_float) {
+  vllm::ScalarType const b_q_type = vllm::ScalarType::from_id(b_q_type_id);
+  int pack_factor = 32 / b_q_type.size_bits();
 
   if (moe_block_size != 8) {
     TORCH_CHECK(moe_block_size % 16 == 0,
@@ -641,27 +745,19 @@ torch::Tensor moe_wna16_marlin_gemm(
   TORCH_CHECK(b_scales.device().is_cuda(), "b_scales is not on GPU");
   TORCH_CHECK(b_scales.is_contiguous(), "b_scales is not contiguous");
 
-  torch::Tensor a_scales;
-  auto options = torch::TensorOptions().dtype(c_dtype).device(a.device());
-  auto options_fp32 =
-      torch::TensorOptions().dtype(at::kFloat).device(a.device());
-
-  if (a_scales_or_none.has_value()) {
-    a_scales = a_scales_or_none.value();
-    TORCH_CHECK(a_type.size_bits() == 8,
-                "a_scales can only be used for 8bit activation.");
-  } else {
-    a_scales = torch::empty({0}, options_fp32);
-    TORCH_CHECK(a_type.size_bits() != 8,
-                "the a_scales parameter must be passed for 8bit activation.");
-  }
-
+  // thread_k: `k` size of a thread_tile in `weights` (can usually be left as
+  // auto -1)
+  int thread_k = -1;
+  // thread_n: `n` size of a thread_tile in `weights` (can usually be left as
+  // auto -1)
+  int thread_n = -1;
   // sms: number of SMs to use for the kernel
   int sms = -1;
   cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, a.get_device());
 
   // Alloc buffers
   const at::cuda::OptionalCUDAGuard device_guard(device_of(a));
+  auto options = torch::TensorOptions().dtype(a.dtype()).device(a.device());
   torch::Tensor c;
   if (c_or_none.has_value()) {
     c = c_or_none.value();
@@ -678,6 +774,8 @@ torch::Tensor moe_wna16_marlin_gemm(
 
   // Alloc C tmp buffer that is going to be used for the global reduce
   torch::Tensor c_tmp;
+  auto options_fp32 =
+      torch::TensorOptions().dtype(at::kFloat).device(a.device());
   if (use_fp32_reduce && !use_atomic_add) {
     // max num of threadblocks is sms * 4
     long max_c_tmp_size = min(
@@ -748,11 +846,11 @@ torch::Tensor moe_wna16_marlin_gemm(
   torch::Tensor global_scale;
   if (global_scale_or_none.has_value()) {
     global_scale = global_scale_or_none.value();
-    TORCH_CHECK(b_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn,
+    TORCH_CHECK(b_q_type == vllm::kFE2M1f && group_size == 16,
                 "global_scale can only be used for nvfp4 format.");
   } else {
     global_scale = torch::empty({0}, options);
-    TORCH_CHECK(!(b_type == vllm::kFE2M1f && s_type == vllm::kFE4M3fn),
+    TORCH_CHECK(!(b_q_type == vllm::kFE2M1f && group_size == 16),
                 "the global_scale parameter must be passed for nvfp4 format.");
   }
 
@@ -779,15 +877,15 @@ torch::Tensor moe_wna16_marlin_gemm(
   bool has_zp = b_zeros.size(-1) > 0;
   if (has_zp) {
     TORCH_CHECK(
-        b_type == vllm::kU4 || b_type == vllm::kU8,
-        "b_type must be u4 or u8 when has_zp = True. Got = ", b_type.str());
+        b_q_type == vllm::kU4 || b_q_type == vllm::kU8,
+        "b_q_type must be u4 or u8 when has_zp = True. Got = ", b_q_type.str());
   } else {
-    TORCH_CHECK(b_type == vllm::kU4B8 || b_type == vllm::kU8B128 ||
-                    b_type == vllm::kS4 || b_type == vllm::kS8 ||
-                    b_type == vllm::kFE4M3fn || b_type == vllm::kFE2M1f,
-                "b_type must be uint4b8, uint8b128, int4, int8, "
-                "float8_e4m3fn or float4_e2m1f when has_zp = False. Got = ",
-                b_type.str());
+    TORCH_CHECK(b_q_type == vllm::kU4B8 || b_q_type == vllm::kU8B128 ||
+                    b_q_type == vllm::kFE4M3fn || b_q_type == vllm::kFE2M1f,
+                "b_q_type must be uint4b8, uint8b128, float8_e4m3fn or "
+                "float4_e2m1f when "
+                "has_zp = False. Got = ",
+                b_q_type.str());
   }
 
   if (has_zp && is_zp_float) {
@@ -831,33 +929,71 @@ torch::Tensor moe_wna16_marlin_gemm(
               " is below min_workspace_size = ", min_workspace_size);
 
   int dev = a.get_device();
+  if (a.scalar_type() == at::ScalarType::Half) {
+    void* scales_ptr;
+    if (b_q_type == vllm::kFE2M1f) {
+      if (group_size == 16)
+        scales_ptr = b_scales.data_ptr<at::Float8_e4m3fn>();
+      else if (group_size == 32)
+        scales_ptr = b_scales.data_ptr<at::Float8_e8m0fnu>();
+      else
+        TORCH_CHECK(false,
+                    "float4_e2m1f only supports group_size == 16 (NVFP4) ",
+                    "and group_size == 32 (MXFP4)");
+    } else {
+      scales_ptr = b_scales.data_ptr<at::Half>();
+    }
 
-  TORCH_CHECK(a_scales.scalar_type() == at::ScalarType::Float,
-              "scalar type of a_scales must be float");
-  TORCH_CHECK(global_scale.scalar_type() == c.scalar_type(),
-              "scalar type of global_scale must be the same with c");
-  if (a_type.size_bits() == 16) {
-    TORCH_CHECK(
-        a.scalar_type() == c.scalar_type(),
-        "scalar type of a must be the same with c for 16 bit activation");
+    MARLIN_NAMESPACE_NAME::marlin_mm<half>(
+        a.data_ptr<at::Half>(), b_q_weight.data_ptr(), c.data_ptr<at::Half>(),
+        c_tmp.data_ptr<float>(), b_bias.data_ptr<at::Half>(), scales_ptr,
+        global_scale.data_ptr<at::Half>(), b_zeros.data_ptr(), g_idx.data_ptr(),
+        perm.data_ptr(), a_tmp.data_ptr<at::Half>(),
+        sorted_token_ids.data_ptr(), expert_ids.data_ptr(),
+        num_tokens_past_padded.data_ptr(), topk_weights.data_ptr(),
+        moe_block_size, top_k, mul_topk_weights, is_ep, size_m, size_n, size_k,
+        workspace.data_ptr(), b_q_type, has_bias, has_act_order, is_k_full,
+        has_zp, num_groups, group_size, dev,
+        at::cuda::getCurrentCUDAStream(dev), thread_k, thread_n, sms,
+        use_atomic_add, use_fp32_reduce, is_zp_float);
+  } else if (a.scalar_type() == at::ScalarType::BFloat16) {
+    void* scales_ptr;
+    if (b_q_type == vllm::kFE2M1f) {
+      if (group_size == 16)
+        scales_ptr = b_scales.data_ptr<at::Float8_e4m3fn>();
+      else if (group_size == 32)
+        scales_ptr = b_scales.data_ptr<at::Float8_e8m0fnu>();
+      else
+        TORCH_CHECK(false,
+                    "float4_e2m1f only supports group_size == 16 (NVFP4) ",
+                    "and group_size == 32 (MXFP4)");
+    } else {
+      scales_ptr = b_scales.data_ptr<at::BFloat16>();
+    }
+
+    MARLIN_NAMESPACE_NAME::marlin_mm<nv_bfloat16>(
+        a.data_ptr<at::BFloat16>(), b_q_weight.data_ptr(),
+        c.data_ptr<at::BFloat16>(), c_tmp.data_ptr<float>(),
+        b_bias.data_ptr<at::BFloat16>(), scales_ptr,
+        global_scale.data_ptr<at::BFloat16>(), b_zeros.data_ptr(),
+        g_idx.data_ptr(), perm.data_ptr(), a_tmp.data_ptr<at::BFloat16>(),
+        sorted_token_ids.data_ptr(), expert_ids.data_ptr(),
+        num_tokens_past_padded.data_ptr(), topk_weights.data_ptr(),
+        moe_block_size, top_k, mul_topk_weights, is_ep, size_m, size_n, size_k,
+        workspace.data_ptr(), b_q_type, has_bias, has_act_order, is_k_full,
+        has_zp, num_groups, group_size, dev,
+        at::cuda::getCurrentCUDAStream(dev), thread_k, thread_n, sms,
+        use_atomic_add, use_fp32_reduce, is_zp_float);
+  } else {
+    TORCH_CHECK(false,
+                "moe_wna16_marlin_gemm only supports bfloat16 and float16");
   }
 
-  MARLIN_NAMESPACE_NAME::marlin_mm(
-      a.data_ptr(), b_q_weight.data_ptr(), c.data_ptr(), c_tmp.data_ptr(),
-      b_bias.data_ptr(), a_scales.data_ptr(), b_scales.data_ptr(),
-      global_scale.data_ptr(), b_zeros.data_ptr(), g_idx.data_ptr(),
-      perm.data_ptr(), a_tmp.data_ptr(), sorted_token_ids.data_ptr(),
-      expert_ids.data_ptr(), num_tokens_past_padded.data_ptr(),
-      topk_weights.data_ptr(), moe_block_size, num_experts, top_k,
-      mul_topk_weights, is_ep, size_m, size_n, size_k, workspace.data_ptr(),
-      a_type, b_type, c_type, s_type, has_bias, has_act_order, is_k_full,
-      has_zp, num_groups, group_size, dev, at::cuda::getCurrentCUDAStream(dev),
-      thread_k, thread_n, sms, blocks_per_sm, use_atomic_add, use_fp32_reduce,
-      is_zp_float);
-
   return c;
 }
 
+#endif
+
 TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
   m.impl("moe_wna16_marlin_gemm", &moe_wna16_marlin_gemm);
 }

csrc/moe/moe_ops.h

@@ -39,9 +39,9 @@ torch::Tensor moe_wna16_gemm(torch::Tensor input, torch::Tensor output,
                              int64_t BLOCK_SIZE_K, int64_t bit);
 
 std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
-    torch::Tensor const& scores, int64_t n_group, int64_t topk_group,
-    int64_t topk, bool renormalize, double routed_scaling_factor,
-    torch::Tensor const& bias, int64_t scoring_func);
+    torch::Tensor const& scores, torch::Tensor const& scores_with_bias,
+    int64_t n_group, int64_t topk_group, int64_t topk, bool renormalize,
+    double routed_scaling_factor);
 #endif
 
 bool moe_permute_unpermute_supported();

csrc/moe/torch_bindings.cpp

@@ -63,18 +63,16 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
   m.def(
       "moe_wna16_marlin_gemm(Tensor! a, Tensor? c_or_none,"
       "Tensor! b_q_weight, Tensor? b_bias_or_none,"
-      "Tensor! b_scales, Tensor? a_scales, Tensor? global_scale, Tensor? "
+      "Tensor! b_scales, Tensor? global_scale, Tensor? "
       "b_zeros_or_none,"
       "Tensor? g_idx_or_none, Tensor? perm_or_none, Tensor! workspace,"
       "Tensor sorted_token_ids,"
       "Tensor! expert_ids, Tensor! num_tokens_past_padded,"
       "Tensor! topk_weights, int moe_block_size, int top_k, "
-      "bool mul_topk_weights, bool is_ep, int b_type_id,"
+      "bool mul_topk_weights, bool is_ep, int b_q_type_id,"
       "int size_m, int size_n, int size_k,"
       "bool is_full_k, bool use_atomic_add,"
-      "bool use_fp32_reduce, bool is_zp_float,"
-      "int thread_k, int thread_n, int blocks_per_sm) -> Tensor");
-
+      "bool use_fp32_reduce, bool is_zp_float) -> Tensor");
   m.def(
       "marlin_gemm_moe(Tensor! a, Tensor! b_q_weights, Tensor! sorted_ids, "
       "Tensor! topk_weights, Tensor! topk_ids, Tensor! b_scales, Tensor! "
@@ -109,10 +107,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
 
   // Apply grouped topk routing to select experts.
   m.def(
-      "grouped_topk(Tensor scores, int n_group, int "
+      "grouped_topk(Tensor scores, Tensor scores_with_bias, int n_group, int "
       "topk_group, int topk, bool renormalize, float "
-      "routed_scaling_factor, Tensor bias, int scoring_func) -> (Tensor, "
-      "Tensor)");
+      "routed_scaling_factor) -> (Tensor, Tensor)");
   m.impl("grouped_topk", torch::kCUDA, &grouped_topk);
 #endif
 }

csrc/ops.h

@@ -52,13 +52,14 @@ void paged_attention_v2(
     const int64_t blocksparse_vert_stride, const int64_t blocksparse_block_size,
     const int64_t blocksparse_head_sliding_step);
 
+#ifndef USE_ROCM
 void merge_attn_states(torch::Tensor& output,
                        std::optional<torch::Tensor> output_lse,
                        const torch::Tensor& prefix_output,
                        const torch::Tensor& prefix_lse,
                        const torch::Tensor& suffix_output,
                        const torch::Tensor& suffix_lse);
-#ifndef USE_ROCM
+
 void convert_vertical_slash_indexes(
     torch::Tensor& block_count,      // [BATCH, N_HEADS, NUM_ROWS]
     torch::Tensor& block_offset,     // [BATCH, N_HEADS, NUM_ROWS, NNZ_S]
@@ -91,12 +92,6 @@ void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
 void fused_add_rms_norm(torch::Tensor& input, torch::Tensor& residual,
                         torch::Tensor& weight, double epsilon);
 
-void fused_qk_norm_rope(torch::Tensor& qkv, int64_t num_heads_q,
-                        int64_t num_heads_k, int64_t num_heads_v,
-                        int64_t head_dim, double eps, torch::Tensor& q_weight,
-                        torch::Tensor& k_weight, torch::Tensor& cos_sin_cache,
-                        bool is_neox, torch::Tensor& position_ids);
-
 void apply_repetition_penalties_(torch::Tensor& logits,
                                  const torch::Tensor& prompt_mask,
                                  const torch::Tensor& output_mask,

csrc/quantization/activation_kernels.cu

@@ -279,17 +279,17 @@ __device__ __forceinline__ void token_bounds(int32_t n_tokens,
 }
 
 template <int BLOCK_COUNT, int SMEM_SIZE_BYTES_Y, typename fp8_type,
-          typename scale_t, int THREADS, typename Idx_t, bool CEIL_UE8M0,
-          int GROUP_SIZE = 128, int NUM_STAGES = 3>
+          int THREADS, typename Idx_t, bool USE_UE8M0, int GROUP_SIZE = 128,
+          int NUM_STAGES = 3>
 __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
     const __nv_bfloat16* __restrict__ _input, fp8_type* __restrict__ _y_q,
-    scale_t* __restrict__ _y_s, const int32_t* __restrict__ tokens_per_expert,
+    float* __restrict__ _y_s, const int32_t* __restrict__ tokens_per_expert,
     // sizes
     Idx_t E, Idx_t T, Idx_t H,
     // strides (in elements)
     Idx_t stride_i_e, Idx_t stride_i_t, Idx_t stride_i_h, Idx_t stride_yq_e,
     Idx_t stride_yq_t, Idx_t stride_yq_h, Idx_t stride_ys_e, Idx_t stride_ys_t,
-    Idx_t stride_ys_g, Idx_t stride_ys_p, Idx_t stride_counts_e) {
+    Idx_t stride_ys_g, Idx_t stride_counts_e) {
 #ifndef USE_ROCM
   static constexpr int NUM_WARPS = THREADS / WARP_SIZE;
 
@@ -466,22 +466,9 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
 
   __nv_fp8x4_e4m3* y_q_base_ptr =
       reinterpret_cast<__nv_fp8x4_e4m3*>(_y_q) + lane_id;
-
-  Idx_t scale_group_offset = 0;
-  if constexpr (std::is_same<scale_t, uint8_t>::value) {
-    // packed int32_t format
-    int pack_id = warp_position_scales / 4;
-    int scale_in_pack = warp_position_scales % 4;
-    scale_group_offset = pack_id * stride_ys_p + scale_in_pack * stride_ys_g;
-  } else {
-    scale_group_offset = warp_position_scales * stride_ys_g;
-  }
-
-  scale_t* const y_scale_base_ptr = _y_s + scale_group_offset;
+  auto y_scale_base_ptr = _y_s + warp_position_scales * stride_ys_g;
 
   for (auto j = tokens_lower; j < tokens_upper; j++) {
-    int current_group_id = warp_position_scales;  // Running count of which
-                                                  // group is being processed
     const Idx_t base_ys = expert_id * stride_ys_e;
     auto y_s_ptr = y_scale_base_ptr + base_ys + token_offset * stride_ys_t;
     __nv_fp8x4_e4m3* y_q_ptr =
@@ -522,7 +509,7 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
 
       __nv_bfloat16 y_s = __hmul(warp_max(_y_max2.x), fp8_inv);
 
-      if constexpr (CEIL_UE8M0) {
+      if constexpr (USE_UE8M0) {
         y_s = hexp2(hceil(hlog2(y_s)));
       }
 
@@ -540,24 +527,8 @@ __global__ void silu_mul_fp8_quant_deep_gemm_kernel(
       y_q_ptr += WARP_SIZE * stride_yq_h;
 
       if (!lane_id) {
-        // Store scales.
-        if constexpr (std::is_same<scale_t, uint8_t>::value) {
-          // Packed UE8MO format. Remove Mantissa.
-          *y_s_ptr = reinterpret_cast<int16_t&>(y_s) >> 7;
-
-          bool const jump_pack = (current_group_id + 1) % 4 == 0;
-          // Minus 3 because we need to get to the first group in the
-          // next pack.
-          y_s_ptr += jump_pack ? (stride_ys_p - 3) : stride_ys_g;
-
-        } else {
-          // float32 format
-          static_assert(std::is_same<scale_t, float>::value);
-          *y_s_ptr = y_s;
-          y_s_ptr += stride_ys_g;
-        }
-
-        current_group_id += 1;
+        *y_s_ptr = y_s;
+        y_s_ptr += stride_ys_g;
       }
     }
   }
@@ -602,21 +573,16 @@ void persistent_masked_m_silu_mul_quant(
     const at::Tensor& tokens_per_expert,  // (E)
     at::Tensor& y_q,                      // (E, T, H) [OUT]
     at::Tensor& y_s,                      // (E, T, H//group_size) [OUT]
-    bool cast_scale_ue8m0) {
+    bool use_ue8m0) {
 #ifndef USE_ROCM
 
   // This kernel currently only supports H % 128 == 0 and assumes a
   // fixed GROUP_SIZE of 128.
-  static constexpr int GROUP_SIZE = 128;
-
   TORCH_CHECK(input.dtype() == torch::kBFloat16);
   TORCH_CHECK(y_q.dtype() == torch::kFloat8_e4m3fn ||
               y_q.dtype() == torch::kFloat8_e4m3fnuz);
-  TORCH_CHECK(input.size(-1) % (GROUP_SIZE * 2) == 0);
-
-  bool const is_packed_ue8m0 =
-      (y_s.dtype() == torch::kInt32 && cast_scale_ue8m0);
-  TORCH_CHECK(y_s.dtype() == torch::kFloat32 || is_packed_ue8m0);
+  TORCH_CHECK(y_s.dtype() == torch::kFloat32);
+  TORCH_CHECK(input.size(-1) % 256 == 0);
 
   using Idx_t = int64_t;
 
@@ -629,18 +595,17 @@ void persistent_masked_m_silu_mul_quant(
   Idx_t stride_yq_e = y_q.stride(0);
   Idx_t stride_yq_t = y_q.stride(1);
   Idx_t stride_yq_h = y_q.stride(2);
+  Idx_t stride_ys_e = y_s.stride(0);
+  Idx_t stride_ys_t = y_s.stride(1);
+  Idx_t stride_ys_g = y_s.stride(2);
 
   Idx_t stride_counts_e = tokens_per_expert.stride(0);
 
-  int const NUM_GROUPS = H / GROUP_SIZE;
+  static constexpr int GROUP_SIZE = 128;
 
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  // TODO: Get this from cuda_arch ?
-  static constexpr int SILU_V2_BLOCK_COUNT = 132 * 32;
-
-  #define KERNEL(BLOCK_COUNT, scale_t, STRIDE_YS_E, STRIDE_YS_T, STRIDE_YS_G,  \
-                 STRIDE_YS_P, CEIL_UE8M0, THREAD_COUNT, STAGES)                \
+  #define KERNEL(BLOCK_COUNT, USE_UE8M0, THREAD_COUNT, STAGES)                 \
     static constexpr int NUM_WARPS = THREAD_COUNT / WARP_SIZE;                 \
     int sms = SILU_V2_BLOCK_COUNT;                                             \
     static constexpr int max_shared_mem_bytes =                                \
@@ -650,86 +615,38 @@ void persistent_masked_m_silu_mul_quant(
     VLLM_DISPATCH_FP8_TYPES(                                                   \
         y_q.scalar_type(), "silu_mul_fp8_quant_deep_gemm_kernel", [&] {        \
           vllm::silu_mul_fp8_quant_deep_gemm_kernel<                           \
-              BLOCK_COUNT, max_shared_mem_bytes, fp8_t, scale_t, THREAD_COUNT, \
-              Idx_t, CEIL_UE8M0, GROUP_SIZE, STAGES>                           \
+              BLOCK_COUNT, max_shared_mem_bytes, fp8_t, THREAD_COUNT, Idx_t,   \
+              USE_UE8M0, GROUP_SIZE, STAGES>                                   \
               <<<grid, block, max_shared_mem_bytes + (E + 1) * 16, stream>>>(  \
                   reinterpret_cast<__nv_bfloat16*>(input.data_ptr()),          \
-                  (fp8_t*)y_q.data_ptr(),                                      \
-                  reinterpret_cast<scale_t*>(y_s.data_ptr()),                  \
+                  (fp8_t*)y_q.data_ptr(), y_s.data_ptr<float>(),               \
                   reinterpret_cast<int32_t*>(tokens_per_expert.data_ptr()), E, \
                   T, H, stride_i_e, stride_i_t, stride_i_h, stride_yq_e,       \
-                  stride_yq_t, stride_yq_h, STRIDE_YS_E, STRIDE_YS_T,          \
-                  STRIDE_YS_G, STRIDE_YS_P, stride_counts_e);                  \
+                  stride_yq_t, stride_yq_h, stride_ys_e, stride_ys_t,          \
+                  stride_ys_g, stride_counts_e);                               \
         });
 
-  #define LAUNCH_ON_H(scale_t, STRIDE_YS_E, STRIDE_YS_T, STRIDE_YS_G,         \
-                      STRIDE_YS_P, CEIL_UE8M0)                                \
-    if (H >= 4096 && (NUM_GROUPS % 8) == 0) {                                 \
-      /* 8 warp config */                                                     \
-      static constexpr int NUM_STAGES = 4;                                    \
-      static constexpr int THREAD_COUNT = 256;                                \
-      KERNEL(SILU_V2_BLOCK_COUNT, scale_t, STRIDE_YS_E, STRIDE_YS_T,          \
-             STRIDE_YS_G, STRIDE_YS_P, CEIL_UE8M0, THREAD_COUNT, NUM_STAGES); \
-    } else {                                                                  \
-      /* 1 warp config */                                                     \
-      static constexpr int THREAD_COUNT = 32;                                 \
-      KERNEL(SILU_V2_BLOCK_COUNT, scale_t, STRIDE_YS_E, STRIDE_YS_T,          \
-             STRIDE_YS_G, STRIDE_YS_P, CEIL_UE8M0, THREAD_COUNT, 2);          \
+  static constexpr int SILU_V2_BLOCK_COUNT = 132 * 32;
+
+  if (!use_ue8m0) {
+    if (H >= 4096) {
+      static constexpr int NUM_STAGES = 4;
+      static constexpr int THREAD_COUNT = 256;
+      KERNEL(SILU_V2_BLOCK_COUNT, false, THREAD_COUNT, NUM_STAGES);
+    } else {
+      static constexpr int THREAD_COUNT = 32;
+      KERNEL(SILU_V2_BLOCK_COUNT, false, THREAD_COUNT, 2);
+    }
+  } else {
+    if (H >= 4096) {
+      static constexpr int NUM_STAGES = 4;
+      static constexpr int THREAD_COUNT = 256;
+      KERNEL(SILU_V2_BLOCK_COUNT, true, THREAD_COUNT, NUM_STAGES);
+    } else {
+      static constexpr int THREAD_COUNT = 32;
+      KERNEL(SILU_V2_BLOCK_COUNT, true, THREAD_COUNT, 2);
     }
-
-  Idx_t stride_ys_e = y_s.stride(0);
-  Idx_t stride_ys_t = y_s.stride(1);
-  Idx_t stride_ys_g = y_s.stride(2);
-  Idx_t stride_ys_p = 0;
-  if (!cast_scale_ue8m0) {
-    TORCH_CHECK(!is_packed_ue8m0);
-    LAUNCH_ON_H(float, stride_ys_e, stride_ys_t, stride_ys_g, stride_ys_p,
-                false);
-    return;
   }
 
-  if (!is_packed_ue8m0) {
-    // UE8M0 but not packed
-    LAUNCH_ON_H(float, stride_ys_e, stride_ys_t, stride_ys_g, stride_ys_p,
-                true);
-    return;
-  }
-
-  TORCH_CHECK(cast_scale_ue8m0 && is_packed_ue8m0);
-  TORCH_CHECK(y_s.dtype() == torch::kInt32);
-
-  // Int32 packed ue8m0 scales tensor.
-  // Let E, T, G be the number to experts, number of tokens and number of groups
-  // respectively. Let, E = 2, T = 4, G = 6, in this case the int32 scales
-  // tensor are of shape [1, 4, 2] and stride [8, 1, 4]. The scales are expected
-  // to be arranged as follows,
-  // [[T0G0-T0G1-T0G2-T0G3, T0G4-T0G5-X-X,],
-  //  [T1G0-T1G1-T1G2-T1G3, T1G4-T1G5-X-X,]
-  //  [T2G0-T2G1-T2G2-T2G3, T2G4-T2G5-X-X,]
-  //  [T3G0-T3G1-T3G2-T3G3, T3G4-T3G5-X-X,]]
-  // where, TxGy is the scale ue8m0 scale value of Token x, Group y.
-  //
-  // In memory (in bytes) the scale values are arranged as,
-  //  [T0G0, T0G1, T0G2, T0G3, T1G0, T1G2, T1G3, T1G4, T2G0, T2G1, T2G3, T2G4,
-  //   T3G0, T3G1, T3G2, T3G3, T0G4, T0G5, X, X, T1G4, T1G5, X, X, T2G4, T2G5,
-  //   X, X, T3G4, T3G5, X, X]
-  //
-  // An Int32 tensor of size [1, 4, 2] and stride [8, 1, 4] can be represented
-  // as an uint8 tensor of shape [1, 2, 4, 4] and stride [32, 16, 4, 1]. In
-  // english, ignoring the Experts dimension, the original int32 tensor is
-  // simply treated as two packed [4, 4] uint8 tensor (or two [4, 1] int32
-  // tensor). The following strides setting reflects this change. Caveat: This
-  // means that the G dimension is no longer contiguous. i.e. Note that to move
-  // from G3 to G4, we need to jump along the packing dimension. The kernel
-  // handles this case.
-
-  stride_ys_e *= sizeof(int32_t);
-  stride_ys_p = T * sizeof(int32_t);  // Packing dimension
-  stride_ys_t = sizeof(int32_t);
-  stride_ys_g = 1;
-
-  LAUNCH_ON_H(uint8_t, stride_ys_e, stride_ys_t, stride_ys_g, stride_ys_p,
-              true);
-
 #endif
 }

csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu

@@ -22,7 +22,6 @@
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
 #include <c10/cuda/CUDAStream.h>
-#include "cutlass_extensions/common.hpp"
 
 #include "cute/tensor.hpp"
 #include "cutlass/tensor_ref.h"
@@ -174,7 +173,7 @@ void run_get_group_gemm_starts(
 }
 
 template <typename OutType>
-void run_fp4_blockwise_scaled_group_mm_sm100(
+void run_fp4_blockwise_scaled_group_mm(
     torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
     const torch::Tensor& a_blockscale, const torch::Tensor& b_blockscales,
     const torch::Tensor& alphas, const torch::Tensor& problem_sizes,
@@ -344,225 +343,17 @@ void run_fp4_blockwise_scaled_group_mm_sm100(
 
   auto can_implement_status = gemm_op.can_implement(args);
   TORCH_CHECK(can_implement_status == cutlass::Status::kSuccess,
-              "Failed to implement GEMM: status=", (int)can_implement_status);
+              "Failed to implement GEMM");
 
   // Run the GEMM
   auto status = gemm_op.initialize(args, workspace.data_ptr());
-  TORCH_CHECK(status == cutlass::Status::kSuccess,
-              "Failed to initialize GEMM: status=", (int)status,
-              " workspace_size=", workspace_size, " num_experts=", num_experts,
-              " M=", M, " N=", N, " K=", K);
+  TORCH_CHECK(status == cutlass::Status::kSuccess, "Failed to initialize GEMM");
 
   status = gemm_op.run(args, workspace.data_ptr(), stream);
   TORCH_CHECK(status == cutlass::Status::kSuccess, "Failed to run GEMM");
 }
 
-void run_fp4_blockwise_scaled_group_mm_sm120(
-    torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
-    const torch::Tensor& a_blockscale, const torch::Tensor& b_blockscales,
-    const torch::Tensor& alphas, const torch::Tensor& problem_sizes,
-    const torch::Tensor& expert_offsets, const torch::Tensor& sf_offsets, int M,
-    int N, int K) {
-  using ProblemShape =
-      cutlass::gemm::GroupProblemShape<Shape<int32_t, int32_t, int32_t>>;
-  using ElementType = cutlass::float_e2m1_t;
-  using ElementSFType = cutlass::float_ue4m3_t;
-  using ElementA = cutlass::nv_float4_t<cutlass::float_e2m1_t>;
-  using ElementB = cutlass::nv_float4_t<cutlass::float_e2m1_t>;
-
-  // NOTE: For SM120 it seems templating the output type is not supported and
-  // we need to hardcode the output type to bfloat16
-  using ElementC = cutlass::bfloat16_t;
-  using ElementD = ElementC;
-  using ElementAccumulator = float;
-  // Layout definitions
-  using LayoutA = cutlass::layout::RowMajor;
-  using LayoutB = cutlass::layout::ColumnMajor;
-  using LayoutC = cutlass::layout::RowMajor;
-  using LayoutD = LayoutC;
-
-  // Alignment constraints
-  static constexpr int AlignmentA = 32;
-  static constexpr int AlignmentB = 32;
-  static constexpr int AlignmentC = 128 / cutlass::sizeof_bits<ElementC>::value;
-  static constexpr int AlignmentD = 128 / cutlass::sizeof_bits<ElementD>::value;
-
-  // Architecture definitions
-  using ArchTag = cutlass::arch::Sm120;
-  using OperatorClass = cutlass::arch::OpClassBlockScaledTensorOp;
-
-  using ClusterShape = Shape<_1, _1, _1>;
-  using MmaTileShape = Shape<_128, _128, _128>;
-
-  using FusionOperation = cutlass::epilogue::fusion::LinearCombination<
-      ElementD, ElementAccumulator, ElementC, ElementAccumulator>;
-
-  using CollectiveEpilogue =
-      typename cutlass::epilogue::collective::CollectiveBuilder<
-          ArchTag, OperatorClass, MmaTileShape, ClusterShape,
-          cutlass::epilogue::collective::EpilogueTileAuto, ElementAccumulator,
-          ElementAccumulator, ElementC, LayoutC*, AlignmentC, ElementD,
-          LayoutD*, AlignmentD,
-          cutlass::epilogue::collective::EpilogueScheduleAuto,
-          FusionOperation>::CollectiveOp;
-
-  using CollectiveMainloop =
-      typename cutlass::gemm::collective::CollectiveBuilder<
-          ArchTag, OperatorClass, ElementA, LayoutA*, AlignmentA, ElementB,
-          LayoutB*, AlignmentB, ElementAccumulator, MmaTileShape, ClusterShape,
-          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
-              sizeof(typename CollectiveEpilogue::SharedStorage))>,
-          cutlass::gemm::collective::KernelScheduleAuto>::CollectiveOp;
-
-  using GemmKernel =
-      cutlass::gemm::kernel::GemmUniversal<ProblemShape, CollectiveMainloop,
-                                           CollectiveEpilogue>;
-
-  using Gemm = cutlass::gemm::device::GemmUniversalAdapter<GemmKernel>;
-  using StrideA = typename Gemm::GemmKernel::InternalStrideA;
-  using StrideB = typename Gemm::GemmKernel::InternalStrideB;
-  using StrideC = typename Gemm::GemmKernel::InternalStrideC;
-  using StrideD = typename Gemm::GemmKernel::InternalStrideD;
-
-  using LayoutSFA =
-      typename Gemm::GemmKernel::CollectiveMainloop::InternalLayoutSFA;
-  using LayoutSFB =
-      typename Gemm::GemmKernel::CollectiveMainloop::InternalLayoutSFB;
-  using ScaleConfig =
-      typename Gemm::GemmKernel::CollectiveMainloop::Sm1xxBlkScaledConfig;
-
-  using UnderlyingProblemShape = ProblemShape::UnderlyingProblemShape;
-  int num_experts = static_cast<int>(expert_offsets.size(0));
-  auto options_int =
-      torch::TensorOptions().dtype(torch::kInt64).device(a.device());
-
-  torch::Tensor a_ptrs = torch::empty(num_experts, options_int);
-  torch::Tensor b_ptrs = torch::empty(num_experts, options_int);
-  torch::Tensor out_ptrs = torch::empty(num_experts, options_int);
-  torch::Tensor a_scales_ptrs = torch::empty(num_experts, options_int);
-  torch::Tensor b_scales_ptrs = torch::empty(num_experts, options_int);
-  torch::Tensor alpha_ptrs = torch::empty(num_experts, options_int);
-  torch::Tensor layout_sfa = torch::empty({num_experts, 5}, options_int);
-  torch::Tensor layout_sfb = torch::empty({num_experts, 5}, options_int);
-  torch::Tensor c_strides1 =
-      torch::full({num_experts}, output.stride(0), options_int);
-  torch::Tensor a_strides1 =
-      torch::full({num_experts}, a.stride(0) * 2, options_int);
-  torch::Tensor b_strides1 =
-      torch::full({num_experts}, b.stride(1) * 2, options_int);
-
-  run_get_group_gemm_starts<LayoutSFA, LayoutSFB, ScaleConfig>(
-      a_ptrs, b_ptrs, out_ptrs, a_scales_ptrs, b_scales_ptrs, alpha_ptrs,
-      layout_sfa, layout_sfb, a, b, output, a_blockscale, b_blockscales, alphas,
-      expert_offsets, sf_offsets, problem_sizes, M, N, K);
-
-  // Create an instance of the GEMM
-  Gemm gemm_op;
-
-  // Initialize problem_sizes_as_shapes correctly
-  UnderlyingProblemShape* problem_sizes_as_shapes =
-      static_cast<UnderlyingProblemShape*>(problem_sizes.data_ptr());
-
-  // Set the Scheduler info
-  cutlass::KernelHardwareInfo hw_info;
-  using RasterOrderOptions = cutlass::gemm::kernel::detail::RasterOrderOptions;
-  typename Gemm::GemmKernel::TileSchedulerArguments scheduler;
-  scheduler.raster_order = RasterOrderOptions::AlongM;
-  hw_info.device_id = a.get_device();
-  static std::unordered_map<int, int> cached_sm_counts;
-  if (cached_sm_counts.find(hw_info.device_id) == cached_sm_counts.end()) {
-    cached_sm_counts[hw_info.device_id] =
-        cutlass::KernelHardwareInfo::query_device_multiprocessor_count(
-            hw_info.device_id);
-  }
-  hw_info.sm_count = min(cached_sm_counts[hw_info.device_id], INT_MAX);
-
-  // Mainloop Arguments
-  typename GemmKernel::MainloopArguments mainloop_args{
-      static_cast<const ElementType**>(a_ptrs.data_ptr()),
-      static_cast<StrideA*>(a_strides1.data_ptr()),
-      static_cast<const ElementType**>(b_ptrs.data_ptr()),
-      static_cast<StrideB*>(b_strides1.data_ptr()),
-      static_cast<const ElementSFType**>(a_scales_ptrs.data_ptr()),
-      reinterpret_cast<LayoutSFA*>(layout_sfa.data_ptr()),
-      static_cast<const ElementSFType**>(b_scales_ptrs.data_ptr()),
-      reinterpret_cast<LayoutSFB*>(layout_sfb.data_ptr())};
-
-  // Epilogue Arguments
-  typename GemmKernel::EpilogueArguments epilogue_args{
-      {},  // epilogue.thread
-      nullptr,
-      static_cast<StrideC*>(c_strides1.data_ptr()),
-      static_cast<ElementD**>(out_ptrs.data_ptr()),
-      static_cast<StrideC*>(c_strides1.data_ptr())};
-  auto& fusion_args = epilogue_args.thread;
-  fusion_args.alpha_ptr_array =
-      reinterpret_cast<float**>(alpha_ptrs.data_ptr());
-  fusion_args.dAlpha = {_0{}, _0{}, 1};
-  fusion_args.beta = 0.0f;
-
-  // Gemm Arguments
-  typename GemmKernel::Arguments args{
-      cutlass::gemm::GemmUniversalMode::kGrouped,
-      {num_experts, problem_sizes_as_shapes, nullptr},
-      mainloop_args,
-      epilogue_args,
-      hw_info,
-      scheduler};
-
-  size_t workspace_size = Gemm::get_workspace_size(args);
-  auto const workspace_options =
-      torch::TensorOptions().dtype(torch::kUInt8).device(a.device());
-  auto workspace = torch::empty(workspace_size, workspace_options);
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream(a.get_device());
-
-  auto can_implement_status = gemm_op.can_implement(args);
-  TORCH_CHECK(can_implement_status == cutlass::Status::kSuccess,
-              "Failed to implement GEMM: status=", (int)can_implement_status);
-
-  // Run the GEMM
-  auto status = gemm_op.initialize(args, workspace.data_ptr());
-  TORCH_CHECK(status == cutlass::Status::kSuccess,
-              "Failed to initialize GEMM: status=", (int)status,
-              " workspace_size=", workspace_size, " num_experts=", num_experts,
-              " M=", M, " N=", N, " K=", K);
-
-  status = gemm_op.run(args, workspace.data_ptr(), stream);
-  TORCH_CHECK(status == cutlass::Status::kSuccess, "Failed to run GEMM");
-}
-
-template <typename OutType>
-void run_fp4_blockwise_scaled_group_mm(
-    torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
-    const torch::Tensor& a_blockscale, const torch::Tensor& b_blockscales,
-    const torch::Tensor& alphas, const torch::Tensor& problem_sizes,
-    const torch::Tensor& expert_offsets, const torch::Tensor& sf_offsets, int M,
-    int N, int K) {
-  int32_t version_num = get_sm_version_num();
-#if defined ENABLE_NVFP4_SM120 && ENABLE_NVFP4_SM120
-  if (version_num >= 120 && version_num < 130) {
-    run_fp4_blockwise_scaled_group_mm_sm120(
-        output, a, b, a_blockscale, b_blockscales, alphas, problem_sizes,
-        expert_offsets, sf_offsets, M, N, K);
-    return;
-  }
-#endif
 #if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100
-  if (version_num >= 100 && version_num < 120) {
-    run_fp4_blockwise_scaled_group_mm_sm100<OutType>(
-        output, a, b, a_blockscale, b_blockscales, alphas, problem_sizes,
-        expert_offsets, sf_offsets, M, N, K);
-    return;
-  }
-#endif
-  TORCH_CHECK_NOT_IMPLEMENTED(
-      false,
-      "No compiled cutlass_fp4_group_mm kernel for CUDA device capability: ",
-      version_num, ". Required capability: 100 or 120");
-}
-
-#if (defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100) || \
-    (defined ENABLE_NVFP4_SM120 && ENABLE_NVFP4_SM120)
 constexpr auto FLOAT4_E2M1X2 = at::ScalarType::Byte;
 constexpr auto SF_DTYPE = at::ScalarType::Float8_e4m3fn;
 #endif
@@ -583,8 +374,7 @@ void cutlass_fp4_group_mm(
     const torch::Tensor& a_blockscale, const torch::Tensor& b_blockscales,
     const torch::Tensor& alphas, const torch::Tensor& problem_sizes,
     const torch::Tensor& expert_offsets, const torch::Tensor& sf_offsets) {
-#if (defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100) || \
-    (defined ENABLE_NVFP4_SM120 && ENABLE_NVFP4_SM120)
+#if defined ENABLE_NVFP4_SM100 && ENABLE_NVFP4_SM100
   // Input validation
   CHECK_INPUT(a, FLOAT4_E2M1X2, "a");
   CHECK_INPUT(b, FLOAT4_E2M1X2, "b");
@@ -618,14 +408,6 @@ void cutlass_fp4_group_mm(
         output, a, b, a_blockscale, b_blockscales, alphas, problem_sizes,
         expert_offsets, sf_offsets, M, N, K);
   } else {
-  #if defined ENABLE_NVFP4_SM120 && ENABLE_NVFP4_SM120
-    int32_t version_num = get_sm_version_num();
-    if (version_num >= 120 && version_num < 130) {
-      TORCH_CHECK_NOT_IMPLEMENTED(
-          false, "SM120 NVFP4 MOE only supports bfloat16 output, got: ",
-          output.scalar_type());
-    }
-  #endif
     run_fp4_blockwise_scaled_group_mm<cutlass::half_t>(
         output, a, b, a_blockscale, b_blockscales, alphas, problem_sizes,
         expert_offsets, sf_offsets, M, N, K);
@@ -634,8 +416,8 @@ void cutlass_fp4_group_mm(
   TORCH_CHECK_NOT_IMPLEMENTED(
       false,
       "No compiled cutlass_fp4_group_mm kernel, vLLM must "
-      "be compiled with ENABLE_NVFP4_SM100 or ENABLE_NVFP4_SM120 for SM100/120 "
-      "and CUDA 12.8 or above.");
+      "be compiled with ENABLE_NVFP4_SM100 for SM100+ and CUDA "
+      "12.8 or above.");
 #endif
 }
 

csrc/quantization/fp4/nvfp4_experts_quant.cu

@@ -307,7 +307,7 @@ constexpr auto FLOAT = at::ScalarType::Float;
 constexpr auto INT = at::ScalarType::Int;
 constexpr auto UINT8 = at::ScalarType::Byte;
 
-void scaled_fp4_experts_quant_sm1xxa(
+void scaled_fp4_experts_quant_sm100a(
     torch::Tensor& output, torch::Tensor& output_scale,
     torch::Tensor const& input, torch::Tensor const& input_global_scale,
     torch::Tensor const& input_offset_by_experts,