[Hybrid] calling get_mamba_groups() once at MambaCopyBuffers.create() (#37318)

Signed-off-by: Francesco Fusco <ffu@zurich.ibm.com>

.buildkite/hardware_tests/amd.yaml

@@ -10,7 +10,7 @@ steps:
       docker build
       --build-arg max_jobs=16
       --build-arg REMOTE_VLLM=1
-      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx942;gfx950'
+      --build-arg ARG_PYTORCH_ROCM_ARCH='gfx90a;gfx942;gfx950'
       --build-arg VLLM_BRANCH=$BUILDKITE_COMMIT
       --tag "rocm/vllm-ci:${BUILDKITE_COMMIT}"
       -f docker/Dockerfile.rocm

.buildkite/hardware_tests/cpu.yaml

@@ -21,6 +21,20 @@ steps:
       pytest -x -v -s tests/kernels/moe/test_cpu_fused_moe.py
       pytest -x -v -s tests/kernels/test_onednn.py"
 
+- label: CPU-Compatibility Tests
+  depends_on: []
+  soft_fail: true
+  device: intel_cpu
+  no_plugin: true
+  source_file_dependencies:
+  - cmake/cpu_extension.cmake
+  - setup.py
+  - vllm/platforms/cpu.py
+  commands:
+    - |
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 20m "
+      bash .buildkite/scripts/hardware_ci/run-cpu-compatibility-test.sh"
+
 - label: CPU-Language Generation and Pooling Model Tests
   depends_on: []
   soft_fail: true

.buildkite/image_build/image_build_cpu.sh

@@ -25,9 +25,7 @@ fi
 docker build --file docker/Dockerfile.cpu \
   --build-arg max_jobs=16 \
   --build-arg buildkite_commit="$BUILDKITE_COMMIT" \
-  --build-arg VLLM_CPU_AVX512BF16=true \
-  --build-arg VLLM_CPU_AVX512VNNI=true \
-  --build-arg VLLM_CPU_AMXBF16=true \
+  --build-arg VLLM_CPU_X86=true \
   --tag "$REGISTRY"/"$REPO":"$BUILDKITE_COMMIT"-cpu \
   --target vllm-test \
   --progress plain .

.buildkite/lm-eval-harness/configs/models-large-rocm-fp8.txt

@@ -0,0 +1 @@
+Qwen3-235B-A22B-Instruct-2507-FP8.yaml

.buildkite/release-pipeline.yaml

@@ -83,7 +83,7 @@ steps:
         agents:
           queue: cpu_queue_postmerge
         commands:
-          - "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 vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
+          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_X86=true --tag vllm-ci:build-image --target vllm-build --progress plain -f docker/Dockerfile.cpu ."
           - "mkdir artifacts"
           - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
           - "bash .buildkite/scripts/upload-nightly-wheels.sh manylinux_2_35"
@@ -152,7 +152,7 @@ steps:
           queue: cpu_queue_postmerge
         commands:
           - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-          - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --build-arg VLLM_CPU_AVX512BF16=true --build-arg VLLM_CPU_AVX512VNNI=true --build-arg VLLM_CPU_AMXBF16=true --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(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_X86=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/check-ray-compatibility.sh

@@ -16,6 +16,23 @@ RAY_BASE_URL="https://raw.githubusercontent.com/ray-project/ray/master/python"
 WORK_DIR=$(mktemp -d)
 trap 'rm -rf "$WORK_DIR"' EXIT
 
+# ── Detect PyTorch index URL ─────────────────────────────────────────────
+
+if python3 -c "import torch; assert torch.version.hip" 2>/dev/null; then
+    ROCM_VER=$(python3 -c "import torch; print(torch.version.hip.rsplit('.', 1)[0])")
+    CANDIDATE_URL="https://download.pytorch.org/whl/rocm${ROCM_VER}"
+    if curl -fsSL --head "${CANDIDATE_URL}/" >/dev/null 2>&1; then
+        TORCH_INDEX_URL="${CANDIDATE_URL}"
+    else
+        echo ">>> WARNING: ROCm ${ROCM_VER} wheel index not found at ${CANDIDATE_URL}"
+        echo ">>>          Falling back to default PyPI (resolution may be incomplete)"
+        TORCH_INDEX_URL=""
+    fi
+else
+    TORCH_INDEX_URL="https://download.pytorch.org/whl/cu129"
+fi
+echo ">>> Using PyTorch index: ${TORCH_INDEX_URL:-PyPI default}"
+
 # Fetch all Ray requirement files used in the LLM depset pipeline
 echo ">>> Fetching Ray requirement files"
 RAY_FILES=(
@@ -116,6 +133,11 @@ echo "============================================================"
 echo ">>> Resolving: Can Ray generate compatible lock files?"
 echo "============================================================"
 
+EXTRA_INDEX_ARGS=()
+if [[ -n "${TORCH_INDEX_URL}" ]]; then
+    EXTRA_INDEX_ARGS+=(--extra-index-url "${TORCH_INDEX_URL}")
+fi
+
 set +e
 uv pip compile \
     "${WORK_DIR}/requirements.txt" \
@@ -126,7 +148,7 @@ uv pip compile \
     -c "${WORK_DIR}/vllm-constraints.txt" \
     --python-version 3.12 \
     --python-platform x86_64-manylinux_2_31 \
-    --extra-index-url https://download.pytorch.org/whl/cu129 \
+    "${EXTRA_INDEX_ARGS[@]}" \
     --index-strategy unsafe-best-match \
     --unsafe-package setuptools \
     --unsafe-package ray \

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

@@ -205,6 +205,13 @@ re_quote_pytest_markers() {
       esac
 
       if $is_boundary; then
+        # Strip surrounding double quotes if present (from upstream
+        # single-to-double conversion); without this, wrapping below
+        # would produce '"expr"' with literal double-quote characters.
+        if [[ "$marker_buf" == '"'*'"' ]]; then
+          marker_buf="${marker_buf#\"}"
+          marker_buf="${marker_buf%\"}"
+        fi
         # Flush the collected marker expression
         if [[ "$marker_buf" == *" "* || "$marker_buf" == *"("* ]]; then
           output+="'${marker_buf}' "
@@ -242,6 +249,11 @@ re_quote_pytest_markers() {
 
   # Flush any trailing marker expression (marker at end of command)
   if $collecting && [[ -n "$marker_buf" ]]; then
+    # Strip surrounding double quotes (see mid-stream flush comment)
+    if [[ "$marker_buf" == '"'*'"' ]]; then
+      marker_buf="${marker_buf#\"}"
+      marker_buf="${marker_buf%\"}"
+    fi
     if [[ "$marker_buf" == *" "* || "$marker_buf" == *"("* ]]; then
       output+="'${marker_buf}'"
     else
@@ -321,15 +333,18 @@ apply_rocm_test_overrides() {
   # --- Entrypoint ignores ---
   if [[ $cmds == *" entrypoints/openai "* ]]; then
     cmds=${cmds//" entrypoints/openai "/" entrypoints/openai \
-    --ignore=entrypoints/openai/test_audio.py \
-    --ignore=entrypoints/openai/test_shutdown.py \
+    --ignore=entrypoints/openai/chat_completion/test_audio.py \
+    --ignore=entrypoints/openai/completion/test_shutdown.py \
     --ignore=entrypoints/openai/test_completion.py \
-    --ignore=entrypoints/openai/test_models.py \
-    --ignore=entrypoints/openai/test_lora_adapters.py \
+    --ignore=entrypoints/openai/models/test_models.py \
     --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
-    --ignore=entrypoints/openai/test_root_path.py \
-    --ignore=entrypoints/openai/test_tokenization.py \
-    --ignore=entrypoints/openai/test_prompt_validation.py "}
+    --ignore=entrypoints/openai/chat_completion/test_root_path.py \
+    --ignore=entrypoints/openai/completion/test_prompt_validation.py "}
+  fi
+
+  if [[ $cmds == *" entrypoints/serve"* ]]; then
+    cmds="${cmds} \
+    --ignore=entrypoints/serve/lora/test_lora_adapters.py"
   fi
 
   if [[ $cmds == *" entrypoints/llm "* ]]; then
@@ -492,6 +507,8 @@ else
     -e HF_TOKEN \
     -e AWS_ACCESS_KEY_ID \
     -e AWS_SECRET_ACCESS_KEY \
+    -e BUILDKITE_PARALLEL_JOB \
+    -e BUILDKITE_PARALLEL_JOB_COUNT \
     -v "${HF_CACHE}:${HF_MOUNT}" \
     -e "HF_HOME=${HF_MOUNT}" \
     -e "PYTHONPATH=${MYPYTHONPATH}" \

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

@@ -0,0 +1,65 @@
+#!/bin/bash
+set -euox pipefail
+
+export VLLM_CPU_KVCACHE_SPACE=1 
+export VLLM_CPU_CI_ENV=1
+# Reduce sub-processes for acceleration
+export TORCH_COMPILE_DISABLE=1 
+export VLLM_ENABLE_V1_MULTIPROCESSING=0
+
+SDE_ARCHIVE="sde-external-10.7.0-2026-02-18-lin.tar.xz"
+SDE_CHECKSUM="CA3D4086DE4ACB3FAEDF9F57B541C6936B7D5E19AE2BF763B6EA933573A0A217"
+wget "https://downloadmirror.intel.com/913594/${SDE_ARCHIVE}"
+echo "${SDE_CHECKSUM}  ${SDE_ARCHIVE}" | sha256sum --check
+mkdir -p sde
+tar -xvf "./${SDE_ARCHIVE}" --strip-components=1 -C ./sde/
+
+wait_for_pid_and_check_log() {
+    local pid="$1"
+    local log_file="$2"
+    local exit_status
+
+    if [ -z "$pid" ] || [ -z "$log_file" ]; then
+        echo "Usage: wait_for_pid_and_check_log <PID> <LOG_FILE>"
+        return 1
+    fi
+
+    echo "Waiting for process $pid to finish..."
+    
+    # Use the 'wait' command to pause the script until the specific PID exits.
+    # The 'wait' command's own exit status will be that of the waited-for process.
+    if wait "$pid"; then
+        exit_status=$?
+        echo "Process $pid finished with exit status $exit_status (Success)."
+    else
+        exit_status=$?
+        echo "Process $pid finished with exit status $exit_status (Failure)."
+    fi
+
+    if [ "$exit_status" -ne 0 ]; then
+        echo "Process exited with a non-zero status."
+        echo "--- Last few lines of log file: $log_file ---"
+        tail -n 50 "$log_file"
+        echo "---------------------------------------------"
+        return 1 # Indicate failure based on exit status
+    fi
+
+    echo "No errors detected in log file and process exited successfully."
+    return 0
+}
+
+# Test Sky Lake (AVX512F)
+./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_0.log 2>&1 &
+PID_TEST_0=$!
+
+# Test Cascade Lake (AVX512F + VNNI)
+./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_1.log 2>&1 &
+PID_TEST_1=$!
+
+# Test Cooper Lake (AVX512F + VNNI + BF16)
+./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_2.log 2>&1 &
+PID_TEST_2=$!
+
+wait_for_pid_and_check_log $PID_TEST_0 test_0.log
+wait_for_pid_and_check_log $PID_TEST_1 test_1.log
+wait_for_pid_and_check_log $PID_TEST_2 test_2.log

.buildkite/scripts/hardware_ci/run-tpu-v1-test-part2.sh

@@ -127,7 +127,7 @@ run_and_track_test() {
 
 # --- Actual Test Execution ---
 run_and_track_test 1 "test_struct_output_generate.py" \
-    "python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
+    "python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
 run_and_track_test 2 "test_moe_pallas.py" \
     "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
 run_and_track_test 3 "test_lora.py" \

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

@@ -33,23 +33,22 @@ docker run \
     bash -c '
     set -e
     echo $ZE_AFFINITY_MASK
-    pip install tblib==3.1.0
     python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
     python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 -O3 -cc.cudagraph_mode=NONE
     python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
     python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
     python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --attention-backend=TRITON_ATTN
     python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager --quantization fp8
-    python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4  --block-size 64 --enforce-eager
+    python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4  --block-size 64 --enforce-eager --max-model-len 8192
     python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2
     python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b  --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel
     cd tests
     pytest -v -s v1/core --ignore=v1/core/test_reset_prefix_cache_e2e.py --ignore=v1/core/test_scheduler_e2e.py
     pytest -v -s v1/engine
     pytest -v -s v1/sample --ignore=v1/sample/test_logprobs.py --ignore=v1/sample/test_logprobs_e2e.py
-    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py
+    pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.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 --ignore=v1/spec_decode/test_acceptance_length.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_example_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_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py -k "not (test_register_kv_caches and FLASH_ATTN and True)"
     pytest -v -s v1/test_serial_utils.py
 '

.buildkite/scripts/scheduled_integration_test/deepseek_v2_lite_prefetch_offload.sh

@@ -1,11 +1,14 @@
 #!/usr/bin/env bash
 set -euxo pipefail
-
 # Nightly e2e test for prefetch offloading with a MoE model.
 # Runs DeepSeek-V2-Lite with prefetch offloading of MoE expert weights
 # and validates GSM8K accuracy matches baseline (no offloading).
 #
 # args: [THRESHOLD] [NUM_QUESTIONS] [START_PORT]
+#
+# Environment variables:
+#   ATTENTION_BACKEND   - attention backend to use (e.g., FLASH_ATTN,
+#                         ROCM_ATTN, FLASHINFER). If unset, uses vllm default.
 THRESHOLD=${1:-0.25}
 NUM_Q=${2:-1319}
 PORT=${3:-8030}
@@ -22,6 +25,14 @@ wait_for_server() {
 
 MODEL="deepseek-ai/DeepSeek-V2-Lite"
 
+# ── Build optional vllm serve flags ─────────────────────────────────────
+
+EXTRA_ARGS=()
+if [[ -n "${ATTENTION_BACKEND:-}" ]]; then
+  echo "Using attention backend: ${ATTENTION_BACKEND}"
+  EXTRA_ARGS+=(--attention-backend "${ATTENTION_BACKEND}")
+fi
+
 cleanup() {
   if [[ -n "${SERVER_PID:-}" ]] && kill -0 "${SERVER_PID}" 2>/dev/null; then
     kill "${SERVER_PID}" 2>/dev/null || true
@@ -40,7 +51,8 @@ vllm serve "$MODEL" \
   --offload-num-in-group 2 \
   --offload-prefetch-step 1 \
   --offload-params w13_weight w2_weight \
-  --port "$PORT" &
+  --port "$PORT" \
+  ${EXTRA_ARGS+"${EXTRA_ARGS[@]}"} &
 SERVER_PID=$!
 wait_for_server "$PORT"
 

.buildkite/test_areas/compile.yaml

@@ -59,7 +59,7 @@ steps:
   - export VLLM_TEST_CLEAN_GPU_MEMORY=1
   - pytest -s -v tests/compile/passes/distributed
 
-- label: Fusion and Compile Unit Tests (B200)
+- label: Fusion and Compile Unit Tests (2xB200)
   timeout_in_minutes: 20
   working_dir: "/vllm-workspace/"
   device: b200

.buildkite/test_areas/distributed.yaml

@@ -15,8 +15,29 @@ steps:
   - pytest -v -s distributed/test_shm_buffer.py
   - pytest -v -s distributed/test_shm_storage.py
 
-- label: Distributed (2 GPUs)
-  timeout_in_minutes: 60
+- label: Distributed DP Tests (2 GPUs)
+  timeout_in_minutes: 20
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/executor/
+  - vllm/worker/worker_base.py
+  - vllm/v1/engine/
+  - vllm/v1/worker/
+  - tests/v1/distributed
+  - tests/entrypoints/openai/test_multi_api_servers.py
+  commands:
+  # 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 entrypoints/openai/test_multi_api_servers.py
+
+- label: Distributed Compile + RPC Tests (2 GPUs)
+  timeout_in_minutes: 20
   working_dir: "/vllm-workspace/tests"
   num_devices: 2
   source_file_dependencies:
@@ -29,62 +50,80 @@ steps:
   - vllm/v1/worker/
   - tests/compile/fullgraph/test_basic_correctness.py
   - tests/compile/test_wrapper.py
-  - tests/distributed/
   - tests/entrypoints/llm/test_collective_rpc.py
-  - tests/v1/distributed
-  - tests/v1/entrypoints/openai/test_multi_api_servers.py
+  commands:
+  # https://github.com/NVIDIA/nccl/issues/1838
+  - export NCCL_CUMEM_HOST_ENABLE=0
+  - pytest -v -s entrypoints/llm/test_collective_rpc.py
+  - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
+  - pytest -v -s ./compile/test_wrapper.py
+
+- label: Distributed Torchrun + Shutdown Tests (2 GPUs)
+  timeout_in_minutes: 20
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/executor/
+  - vllm/worker/worker_base.py
+  - vllm/v1/engine/
+  - vllm/v1/worker/
+  - tests/distributed/
   - tests/v1/shutdown
   - tests/v1/worker/test_worker_memory_snapshot.py
   commands:
   # 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_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'
   - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
   - pytest -v -s v1/worker/test_worker_memory_snapshot.py
 
-- label: Distributed Tests (4 GPUs)
-  timeout_in_minutes: 50
-  working_dir: "/vllm-workspace/tests"
+- label: Distributed Torchrun + Examples (4 GPUs)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace"
   num_devices: 4
   source_file_dependencies:
   - vllm/distributed/
-  - tests/distributed/test_utils
-  - tests/distributed/test_pynccl
-  - tests/distributed/test_events
-  - tests/compile/fullgraph/test_basic_correctness.py
-  - examples/offline_inference/rlhf.py
+  - tests/distributed/test_torchrun_example.py
+  - tests/distributed/test_torchrun_example_moe.py
   - examples/offline_inference/rlhf_colocate.py
-  - examples/offline_inference/new_weight_syncing/
+  - examples/rl/
   - tests/examples/offline_inference/data_parallel.py
-  - tests/v1/distributed
-  - tests/v1/engine/test_engine_core_client.py
-  - tests/distributed/test_symm_mem_allreduce.py
-  - tests/distributed/test_multiproc_executor.py
   commands:
   # https://github.com/NVIDIA/nccl/issues/1838
   - export NCCL_CUMEM_HOST_ENABLE=0
   # test with torchrun tp=2 and external_dp=2
-  - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  - torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example.py
   # test with torchrun tp=2 and pp=2
-  - PP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  - PP_SIZE=2 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example.py
   # test with torchrun tp=4 and dp=1
-  - TP_SIZE=4 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
+  - TP_SIZE=4 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example_moe.py
   # test with torchrun tp=2, pp=2 and dp=1
-  - PP_SIZE=2 TP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
+  - PP_SIZE=2 TP_SIZE=2 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example_moe.py
   # test with torchrun tp=1 and dp=4 with ep
-  - DP_SIZE=4 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
+  - DP_SIZE=4 ENABLE_EP=1 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example_moe.py
   # test with torchrun tp=2 and dp=2 with ep
-  - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
+  - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 tests/distributed/test_torchrun_example_moe.py
   # test with internal dp
-  - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
+  - python3 examples/offline_inference/data_parallel.py --enforce-eager
+  # rlhf examples
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_nccl.py
+  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_ipc.py
+
+- label: Distributed DP Tests (4 GPUs)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+  - vllm/distributed/
+  - tests/v1/distributed
+  - tests/v1/engine/test_engine_core_client.py
+  - tests/distributed/test_utils
+  commands:
+  # https://github.com/NVIDIA/nccl/issues/1838
+  - export NCCL_CUMEM_HOST_ENABLE=0
   - 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
@@ -92,22 +131,27 @@ 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
+
+- label: Distributed Compile + Comm (4 GPUs)
+  timeout_in_minutes: 30
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 4
+  source_file_dependencies:
+  - vllm/distributed/
+  - tests/distributed/test_pynccl
+  - tests/distributed/test_events
+  - tests/compile/fullgraph/test_basic_correctness.py
+  - tests/distributed/test_symm_mem_allreduce.py
+  - tests/distributed/test_multiproc_executor.py
+  commands:
+  # https://github.com/NVIDIA/nccl/issues/1838
+  - export NCCL_CUMEM_HOST_ENABLE=0
   - pytest -v -s compile/fullgraph/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
   # test multi-node TP with multiproc executor (simulated on single node)
   - pytest -v -s distributed/test_multiproc_executor.py::test_multiproc_executor_multi_node
-  # TODO: create a dedicated test section for multi-GPU example tests
-  # when we have multiple distributed example tests
-  # OLD rlhf examples
-  - cd ../examples/offline_inference
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf.py
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
-  # NEW rlhf examples
-  - cd new_weight_syncing
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_nccl.py
-  - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_ipc.py
 
 - label: Distributed Tests (8 GPUs)(H100)
   timeout_in_minutes: 10
@@ -149,7 +193,7 @@ steps:
   num_devices: 2
   commands:
     - pytest -v -s tests/distributed/test_context_parallel.py
-    - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/offline_inference/new_weight_syncing/rlhf_async_new_apis.py
+    - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 examples/rl/rlhf_async_new_apis.py
     - VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model=Qwen/Qwen1.5-MoE-A2.7B -tp=1 -dp=2 --max-model-len=2048 --all2all-backend=deepep_high_throughput
     - pytest -v -s tests/v1/distributed/test_dbo.py
 

.buildkite/test_areas/engine.yaml

@@ -1,5 +1,5 @@
 group: Engine
-depends_on: 
+depends_on:
   - image-build
 steps:
 - label: Engine
@@ -14,28 +14,30 @@ steps:
   commands:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
 
-- label: V1 e2e + engine (1 GPU)
-  timeout_in_minutes: 45
+- label: Engine (1 GPU)
+  timeout_in_minutes: 30
   source_file_dependencies:
-    - vllm/
-    - tests/v1
+    - vllm/v1/engine/
+    - tests/v1/engine/
   commands:
-    # TODO: accuracy does not match, whether setting
-    # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
-    - pytest -v -s v1/e2e
-    # Run this test standalone for now;
-    # need to untangle use (implicit) use of spawn/fork across the tests.
     - pytest -v -s v1/engine/test_preprocess_error_handling.py
-    # Run the rest of v1/engine tests
     - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
-      commands:
-      - pytest -v -s v1/e2e
-      - pytest -v -s v1/engine
+
+- label: e2e Scheduling (1 GPU)
+  timeout_in_minutes: 30
+  source_file_dependencies:
+    - vllm/v1/
+    - tests/v1/e2e/general/
+  commands:
+    - pytest -v -s v1/e2e/general/test_async_scheduling.py
+
+- label: e2e Core (1 GPU)
+  timeout_in_minutes: 30
+  source_file_dependencies:
+    - vllm/v1/
+    - tests/v1/e2e/general/
+  commands:
+    - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
 
 - label: V1 e2e (2 GPUs)
   timeout_in_minutes: 60 # TODO: Fix timeout after we have more confidence in the test stability
@@ -46,7 +48,7 @@ steps:
     - tests/v1/e2e
   commands:
     # Only run tests that need exactly 2 GPUs
-    - pytest -v -s v1/e2e/test_spec_decode.py -k "tensor_parallelism"
+    - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
   mirror:
     amd:
       device: mi325_2
@@ -62,9 +64,21 @@ steps:
     - tests/v1/e2e
   commands:
     # Only run tests that need 4 GPUs
-    - pytest -v -s v1/e2e/test_spec_decode.py -k "eagle_correctness_heavy"
+    - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "eagle_correctness_heavy"
   mirror:
     amd:
       device: mi325_4
       depends_on:
       - image-build-amd
+
+- label: V1 e2e (4xH100)
+  timeout_in_minutes: 60
+  device: h100
+  num_devices: 4
+  optional: true
+  source_file_dependencies:
+    - vllm/v1/attention/backends/utils.py
+    - vllm/v1/worker/gpu_model_runner.py
+    - tests/v1/e2e/test_hybrid_chunked_prefill.py
+  commands:
+    - pytest -v -s v1/e2e/test_hybrid_chunked_prefill.py

.buildkite/test_areas/entrypoints.yaml

@@ -10,7 +10,7 @@ steps:
   - tests/entrypoints/
   commands:
   - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
+  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/rpc --ignore=entrypoints/sleep --ignore=entrypoints/serve/instrumentator --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
 
 - label: Entrypoints Integration (LLM)
   timeout_in_minutes: 40
@@ -34,7 +34,7 @@ steps:
   - tests/entrypoints/test_chat_utils
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/  --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/chat_completion/test_oot_registration.py --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/  --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/test_multi_api_servers.py
   - pytest -v -s entrypoints/test_chat_utils.py
   mirror:
     amd:
@@ -48,11 +48,11 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/rpc
-  - tests/entrypoints/instrumentator
+  - tests/entrypoints/serve/instrumentator
   - tests/tool_use
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/instrumentator
+  - pytest -v -s entrypoints/serve/instrumentator
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/rpc
   - pytest -v -s tool_use
 
@@ -75,19 +75,6 @@ steps:
   commands:
   - pytest -v -s entrypoints/openai/responses
 
-- label: Entrypoints V1
-  timeout_in_minutes: 50
-  source_file_dependencies:
-    - vllm/
-    - tests/v1
-  commands:
-    - pytest -v -s v1/entrypoints
-  mirror:
-    amd:
-      device: mi325_1
-      depends_on:
-      - image-build-amd
-
 - label: OpenAI API Correctness
   timeout_in_minutes: 30
   source_file_dependencies:

.buildkite/test_areas/expert_parallelism.yaml

@@ -24,8 +24,7 @@ steps:
 
 - label: Elastic EP Scaling Test
   timeout_in_minutes: 20
-  device: b200
-  optional: true
+  device: h100
   working_dir: "/vllm-workspace/tests"
   num_devices: 4
   source_file_dependencies:

.buildkite/test_areas/kernels.yaml

@@ -35,7 +35,7 @@ steps:
   parallelism: 2
 
 - label: Kernels MoE Test %N
-  timeout_in_minutes: 60
+  timeout_in_minutes: 25
   source_file_dependencies:
   - csrc/quantization/cutlass_w8a8/moe/
   - csrc/moe/
@@ -47,7 +47,7 @@ steps:
   commands:
     - pytest -v -s kernels/moe --ignore=kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
     - pytest -v -s kernels/moe/test_modular_oai_triton_moe.py --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
-  parallelism: 2
+  parallelism: 5
 
 - label: Kernels Mamba Test
   timeout_in_minutes: 45

.buildkite/test_areas/lm_eval.yaml

@@ -45,6 +45,22 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-blackwell.txt
 
+- label: LM Eval Qwen3.5 Models (B200)
+  timeout_in_minutes: 120
+  device: b200
+  optional: true
+  num_devices: 2
+  source_file_dependencies:
+  - vllm/model_executor/models/qwen3_5.py
+  - vllm/model_executor/models/qwen3_5_mtp.py
+  - vllm/transformers_utils/configs/qwen3_5.py
+  - vllm/transformers_utils/configs/qwen3_5_moe.py
+  - vllm/model_executor/models/qwen3_next.py
+  - vllm/model_executor/models/qwen3_next_mtp.py
+  - vllm/model_executor/layers/fla/ops/
+  commands:
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-blackwell.txt
+
 - label: LM Eval Large Models (H200)
   timeout_in_minutes: 60
   device: h200

.buildkite/test_areas/lora.yaml

@@ -8,7 +8,7 @@ steps:
   - vllm/lora
   - tests/lora
   commands:
-    - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_llm_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py
+    - pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_llm_with_multi_loras.py --ignore=lora/test_olmoe_tp.py --ignore=lora/test_deepseekv2_tp.py --ignore=lora/test_gptoss_tp.py --ignore=lora/test_qwen3moe_tp.py --ignore=lora/test_qwen35_densemoel_lora.py 
   parallelism: 4
 
 
@@ -30,4 +30,5 @@ 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
-    - pytest -v -s -x lora/test_gptoss_tp.py
+    - pytest -v -s -x lora/test_gptoss_tp.py
+    - pytest -v -s -x lora/test_qwen35_densemoel_lora.py

.buildkite/test_areas/model_executor.yaml

@@ -9,9 +9,9 @@ steps:
   - vllm/config/model.py
   - vllm/model_executor
   - tests/model_executor
-  - tests/entrypoints/openai/test_tensorizer_entrypoint.py
+  - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
   commands:
     - apt-get update && apt-get install -y curl libsodium23
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pytest -v -s model_executor
-    - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py
+    - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py

.buildkite/test_areas/model_runner_v2.yaml

@@ -11,18 +11,18 @@ steps:
   - vllm/v1/attention/
   - tests/v1/engine/test_llm_engine.py
   - tests/v1/e2e/
-  - tests/v1/entrypoints/llm/test_struct_output_generate.py
+  - tests/entrypoints/llm/test_struct_output_generate.py
   commands:
   - set -x
   - export VLLM_USE_V2_MODEL_RUNNER=1
   - pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics"
   # This requires eager until we sort out CG correctness issues.
   # TODO: remove ENFORCE_EAGER here after https://github.com/vllm-project/vllm/pull/32936 is merged.
-  - ENFORCE_EAGER=1 pytest -v -s v1/e2e/test_async_scheduling.py -k "not ngram"
-  - pytest -v -s v1/e2e/test_context_length.py
-  - pytest -v -s v1/e2e/test_min_tokens.py
+  - ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram"
+  - pytest -v -s v1/e2e/general/test_context_length.py
+  - pytest -v -s v1/e2e/general/test_min_tokens.py
   # Temporary hack filter to exclude ngram spec decoding based tests.
-  - pytest -v -s v1/entrypoints/llm/test_struct_output_generate.py -k "xgrammar and not speculative_config6 and not speculative_config7 and not speculative_config8 and not speculative_config0"
+  - pytest -v -s entrypoints/llm/test_struct_output_generate.py -k "xgrammar and not speculative_config6 and not speculative_config7 and not speculative_config8 and not speculative_config0"
 
 - label: Model Runner V2 Examples
   timeout_in_minutes: 45
@@ -102,9 +102,9 @@ steps:
   - vllm/v1/worker/gpu/
   - vllm/v1/worker/gpu_worker.py
   - tests/v1/spec_decode/test_max_len.py
-  - tests/v1/e2e/test_spec_decode.py
+  - tests/v1/e2e/spec_decode/test_spec_decode.py
   commands:
   - set -x
   - export VLLM_USE_V2_MODEL_RUNNER=1
   - pytest -v -s v1/spec_decode/test_max_len.py -k "eagle or mtp"
-  - pytest -v -s v1/e2e/test_spec_decode.py -k "eagle or mtp"
+  - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "eagle or mtp"

.buildkite/test_areas/models_multimodal.yaml

@@ -2,15 +2,59 @@ group: Models - Multimodal
 depends_on: 
   - image-build
 steps:
-- label: Multi-Modal Models (Standard) # 60min
-  timeout_in_minutes: 80
+- label: "Multi-Modal Models (Standard) 1: qwen2"
+  timeout_in_minutes: 45
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pip freeze | grep -E 'torch'
-    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
+    - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen2"
+    - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+
+- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
+  timeout_in_minutes: 45
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "qwen3 or gemma"
+    - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+
+- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
+  timeout_in_minutes: 45
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s models/multimodal/generation/test_common.py -m core_model -k "not qwen2 and not qwen3 and not gemma"
+    - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+
+- label: "Multi-Modal Models (Standard) 4: other + whisper"
+  timeout_in_minutes: 45
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/generation/test_ultravox.py --ignore models/multimodal/generation/test_qwen2_5_vl.py --ignore models/multimodal/generation/test_qwen2_vl.py --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
   mirror:
     amd:
@@ -18,7 +62,7 @@ steps:
       depends_on:
       - image-build-amd
 
-- label: Multi-Modal Processor Test (CPU)
+- label: Multi-Modal Processor (CPU)
   depends_on: 
   - image-build-cpu
   timeout_in_minutes: 60
@@ -51,34 +95,44 @@ steps:
   commands:
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
 
-- label: Multi-Modal Models (Extended) 1
+- label: Multi-Modal Models (Extended Generation 1)
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/multimodal
+  - tests/models/multimodal/generation
+  - tests/models/multimodal/test_mapping.py
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
-    - pytest -v -s models/multimodal -m 'not core_model' --ignore models/multimodal/generation/test_common.py --ignore models/multimodal/processing
+    - pytest -v -s models/multimodal/generation -m 'not core_model' --ignore models/multimodal/generation/test_common.py
+    - pytest -v -s models/multimodal/test_mapping.py
   mirror:
     amd:
       device: mi325_1
       depends_on:
       - image-build-amd
 
-- label: Multi-Modal Models (Extended) 2
+- label: Multi-Modal Models (Extended Generation 2)
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/multimodal
+  - tests/models/multimodal/generation
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=0) and not core_model'
 
-- label: Multi-Modal Models (Extended) 3
+- label: Multi-Modal Models (Extended Generation 3)
   optional: true
   source_file_dependencies:
   - vllm/
-  - tests/models/multimodal
+  - tests/models/multimodal/generation
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal/generation/test_common.py -m 'split(group=1) and not core_model'
+
+- label: Multi-Modal Models (Extended Pooling)
+  optional: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal/pooling
+  commands:
+    - pytest -v -s models/multimodal/pooling -m 'not core_model'

.buildkite/test_areas/plugins.yaml

@@ -36,6 +36,6 @@ steps:
   - pytest -v -s plugins_tests/test_scheduler_plugins.py
   - pip install -e ./plugins/vllm_add_dummy_model
   - pytest -v -s distributed/test_distributed_oot.py
-  - pytest -v -s entrypoints/openai/test_oot_registration.py # it needs a clean process
+  - pytest -v -s entrypoints/openai/chat_completion/test_oot_registration.py # it needs a clean process
   - pytest -v -s models/test_oot_registration.py # it needs a clean process
   - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins

.buildkite/test_areas/pytorch.yaml

@@ -35,7 +35,7 @@ steps:
   # 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 {} \\;"
+  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
 
 - label: PyTorch Fullgraph
   timeout_in_minutes: 30

.buildkite/test_areas/spec_decode.yaml

@@ -0,0 +1,40 @@
+group: Spec Decode
+depends_on:
+  - image-build
+steps:
+- label: Spec Decode Eagle
+  timeout_in_minutes: 30
+  source_file_dependencies:
+    - vllm/v1/spec_decode/
+    - vllm/v1/worker/gpu/spec_decode/
+    - tests/v1/e2e/spec_decode/
+  commands:
+    - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
+
+- label: Spec Decode Speculators + MTP
+  timeout_in_minutes: 30
+  source_file_dependencies:
+    - vllm/v1/spec_decode/
+    - vllm/v1/worker/gpu/spec_decode/
+    - vllm/transformers_utils/configs/speculators/
+    - tests/v1/e2e/spec_decode/
+  commands:
+    - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
+
+- label: Spec Decode Ngram + Suffix
+  timeout_in_minutes: 30
+  source_file_dependencies:
+    - vllm/v1/spec_decode/
+    - vllm/v1/worker/gpu/spec_decode/
+    - tests/v1/e2e/spec_decode/
+  commands:
+    - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
+
+- label: Spec Decode Draft Model
+  timeout_in_minutes: 30
+  source_file_dependencies:
+    - vllm/v1/spec_decode/
+    - vllm/v1/worker/gpu/spec_decode/
+    - tests/v1/e2e/spec_decode/
+  commands:
+    - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"

.github/CODEOWNERS

@@ -75,7 +75,7 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /tests/multimodal @DarkLight1337 @ywang96 @NickLucche
 /tests/quantization @mgoin @robertgshaw2-redhat @yewentao256 @pavanimajety
 /tests/test_inputs.py @DarkLight1337 @ywang96
-/tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
+/tests/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 @orozery
 /tests/weight_loading @mgoin @youkaichao @yewentao256
@@ -171,6 +171,7 @@ mkdocs.yaml @hmellor
 
 # Pooling models
 /examples/pooling @noooop
+/docs/models/pooling_models @noooop
 /tests/models/*/pooling* @noooop
 /tests/entrypoints/pooling @noooop
 /vllm/config/pooler.py @noooop

.github/mergify.yml

@@ -27,7 +27,7 @@ pull_request_rules:
         Hi @{{author}}, the pre-commit checks have failed. Please run:
 
         ```bash 
-        uv pip install pre-commit
+        uv pip install pre-commit>=4.5.1
         pre-commit install
         pre-commit run --all-files
         ```
@@ -260,7 +260,7 @@ pull_request_rules:
       - files=examples/offline_inference/structured_outputs.py
       - files=examples/online_serving/structured_outputs/structured_outputs.py
       - files~=^tests/v1/structured_output/
-      - files=tests/v1/entrypoints/llm/test_struct_output_generate.py
+      - files=tests/entrypoints/llm/test_struct_output_generate.py
       - files~=^vllm/v1/structured_output/
   actions:
     label:
@@ -333,9 +333,10 @@ pull_request_rules:
     - label != stale
     - or:
       - files~=^tests/tool_use/
-      - files~=^tests/entrypoints/openai/tool_parsers/
-      - files=tests/entrypoints/openai/test_chat_with_tool_reasoning.py
-      - files~=^vllm/entrypoints/openai/tool_parsers/
+      - files~=^tests/tool_parsers/
+      - files~=^tests/entrypoints/openai/.*tool.*
+      - files~=^tests/entrypoints/anthropic/.*tool.*
+      - files~=^vllm/tool_parsers/
       - files=docs/features/tool_calling.md
       - files~=^examples/tool_chat_*
       - files=examples/offline_inference/chat_with_tools.py
@@ -381,7 +382,7 @@ pull_request_rules:
     - or:
       - files~=^vllm/model_executor/model_loader/tensorizer.py
       - files~=^vllm/model_executor/model_loader/tensorizer_loader.py
-      - files~=^tests/entrypoints/openai/test_tensorizer_entrypoint.py
+      - files~=^tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
       - files~=^tests/model_executor/model_loader/tensorizer_loader/
   actions:
     assign:

.github/scripts/cleanup_pr_body.sh

@@ -1,50 +0,0 @@
-#!/bin/bash
-
-set -eu
-
-# ensure 1 argument is passed
-if [ "$#" -ne 1 ]; then
-    echo "Usage: $0 <pr_number>"
-    exit 1
-fi
-
-PR_NUMBER=$1
-OLD=/tmp/orig_pr_body.txt
-NEW=/tmp/new_pr_body.txt
-
-gh pr view --json body --template "{{.body}}" "${PR_NUMBER}" > "${OLD}"
-cp "${OLD}" "${NEW}"
-
-# Remove markdown comments (like the <!-- markdownlint-disable --> at the start)
-sed -i '/<!--.*-->$/d' "${NEW}"
-
-# Remove "PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED."
-sed -i '/PLEASE FILL IN THE PR DESCRIPTION HERE.*$/d' "${NEW}"
-
-# Remove all lines after and including "**BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE**"
-sed -i '/\*\*BEFORE SUBMITTING, PLEASE READ.*\*\*/,$d' "${NEW}"
-
-# Remove HTML <details> section that includes <summary> text of "PR Checklist (Click to Expand)"
-python3 - <<EOF
-import regex as re
-
-with open("${NEW}", "r") as file:
-    content = file.read()
-
-pattern = re.compile(r'(---\n\n)?<details>.*?<summary>.*?PR Checklist \(Click to Expand\).*?</summary>.*?</details>', re.DOTALL)
-content = re.sub(pattern, '', content)
-
-with open("${NEW}", "w") as file:
-    file.write(content)
-EOF
-
-# Run this only if ${NEW} is different than ${OLD}
-if ! cmp -s "${OLD}" "${NEW}"; then
-    gh pr edit --body-file "${NEW}" "${PR_NUMBER}"
-    echo
-    echo "Updated PR body:"
-    echo
-    cat "${NEW}"
-else
-    echo "No changes needed"
-fi

.github/workflows/cleanup_pr_body.yml

@@ -1,32 +0,0 @@
-name: Cleanup PR Body
-
-on:
-  pull_request_target:
-    types: [opened, reopened, edited]
-
-permissions:
-  pull-requests: write
-
-jobs:
-  update-description:
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
-
-      - name: Set up Python
-        uses: actions/setup-python@83679a892e2d95755f2dac6acb0bfd1e9ac5d548 # v6.1.0
-        with:
-          python-version: '3.12'
-          cache: 'pip'
-
-      - name: Install Python dependencies
-        run: |
-          python3 -m pip install --upgrade pip
-          python3 -m pip install regex
-
-      - name: Update PR description
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: bash .github/scripts/cleanup_pr_body.sh "${{ github.event.number }}"

.github/workflows/issue_autolabel.yml

@@ -383,4 +383,107 @@ jobs:
                   core.notice(`All users for label "${label}" already mentioned, skipping comment`);
                 }
               }
-            }
+            }
+
+      - name: Request missing ROCm info from issue author
+        if: contains(steps.label-step.outputs.labels_added, 'rocm') && contains(toJSON(github.event.issue.labels.*.name), 'bug')
+        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd  # v8.0.0
+        with:
+          script: |
+            const body = (context.payload.issue.body || '').toLowerCase();
+
+            // Check for existing bot comments to avoid duplicate requests
+            const comments = await github.rest.issues.listComments({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              issue_number: context.issue.number,
+            });
+            const botAlreadyAsked = comments.data.some(
+              c => c.user.type === 'Bot' && c.body.includes('<!-- rocm-info-request -->')
+            );
+            if (botAlreadyAsked) {
+              core.notice('ROCm info request already posted, skipping');
+              return;
+            }
+
+            // Define required information and detection patterns
+            const requiredInfo = [
+              {
+                name: 'Reproducer',
+                patterns: [
+                  /reproduc/i, /minimal.?example/i, /repro\b/i, /steps to reproduce/i,
+                  /code.?snippet/i, /sample.?code/i,
+                  /```python[\s\S]*?```/, /```bash[\s\S]*?```/, /```sh[\s\S]*?```/,
+                ],
+                ask: 'A minimal reproducer (code snippet or script that triggers the issue)',
+              },
+              {
+                name: 'Error message',
+                patterns: [
+                  /error/i, /traceback/i, /exception/i, /fault/i, /crash/i,
+                  /failed/i, /abort/i, /panic/i,
+                ],
+                ask: 'The full error message or traceback',
+              },
+              {
+                name: 'Installation method',
+                patterns: [
+                  /docker/i, /rocm\/pytorch/i, /dockerfile/i, /from source/i,
+                  /pip install/i, /build.?from/i, /container/i, /image/i,
+                  /wheel/i, /\.whl/i, /nightly/i,
+                ],
+                ask: 'How you installed vLLM (Docker image name, pip install, or build from source steps)',
+              },
+              {
+                name: 'Command',
+                patterns: [
+                  /vllm serve/i, /python\s+\S+\.py/i, /```bash[\s\S]*?```/,
+                  /```sh[\s\S]*?```/, /command/i, /launch/i, /run\s/i,
+                  /--model/i, /--tensor-parallel/i, /--gpu-memory/i,
+                ],
+                ask: 'The command you used to launch vLLM (e.g., `vllm serve ...` or the Python script)',
+              },
+              {
+                name: 'GFX architecture',
+                patterns: [
+                  /gfx\d{3,4}/i, /mi\d{3}/i, /mi\d{2}\b/i, /radeon/i,
+                  /gpu.?arch/i, /rocm-smi/i, /rocminfo/i, /navi/i,
+                  /instinct/i,
+                ],
+                ask: 'Your GPU model and GFX architecture (e.g., MI300X / gfx942) — run `rocminfo | grep gfx`',
+              },
+            ];
+
+            const issueBody = context.payload.issue.body || '';
+            const missing = requiredInfo.filter(info =>
+              !info.patterns.some(p => p.test(issueBody))
+            );
+
+            if (missing.length === 0) {
+              core.notice('All required ROCm info appears to be present');
+              return;
+            }
+
+            const author = context.payload.issue.user.login;
+            const checklist = requiredInfo.map(info => {
+              const found = !missing.includes(info);
+              return `- [${found ? 'x' : ' '}] ${info.ask}`;
+            }).join('\n');
+            const message = [
+              '<!-- rocm-info-request -->',
+              `Hi @${author}, thanks for reporting this ROCm issue!`,
+              '',
+              'To help us investigate, please make sure the following information is included:',
+              '',
+              checklist,
+              '',
+              'Please provide any unchecked items above. This will help us reproduce and resolve the issue faster. Thank you!',
+            ].join('\n');
+
+            await github.rest.issues.createComment({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              issue_number: context.issue.number,
+              body: message,
+            });
+            core.notice(`Requested missing ROCm info from @${author}: ${missing.map(m => m.name).join(', ')}`);

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

@@ -1,9 +1,9 @@
 name: macOS Apple Silicon Smoke Test
 
 on:
-  push:
-    branches:
-      - main
+  schedule:
+    # Daily at 2:30 AM UTC
+    - cron: '30 2 * * *'
   workflow_dispatch:  # Manual trigger
 
 permissions:

.github/workflows/new_pr_bot.yml

@@ -0,0 +1,96 @@
+name: New PR Bot
+
+on:
+  pull_request_target:
+    types: [opened]
+
+permissions:
+  pull-requests: write
+
+jobs:
+  update-description:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Update PR description
+        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+        with:
+          script: |
+            const { owner, repo } = context.repo;
+            const pr_number = context.issue.number;
+
+            const { data: pr } = await github.rest.pulls.get({
+              owner,
+              repo,
+              pull_number: pr_number,
+            });
+
+            let body = pr.body || '';
+            const original = body;
+
+            // Remove markdown comments (<!-- ... -->)
+            body = body.replace(/^<!--.*-->$/gm, '');
+
+            // Remove "PLEASE FILL IN THE PR DESCRIPTION HERE ..."
+            body = body.replace(/^PLEASE FILL IN THE PR DESCRIPTION HERE.*$/gm, '');
+
+            // Remove all lines after and including "**BEFORE SUBMITTING, PLEASE READ ..."
+            body = body.replace(/\*\*BEFORE SUBMITTING, PLEASE READ.*\*\*[\s\S]*$/, '');
+
+            // Remove <details> section containing "PR Checklist (Click to Expand)"
+            body = body.replace(/(---\n\n)?<details>[\s\S]*?<summary>[\s\S]*?PR Checklist \(Click to Expand\)[\s\S]*?<\/summary>[\s\S]*?<\/details>/g, '');
+
+            if (body !== original) {
+              await github.rest.pulls.update({
+                owner,
+                repo,
+                pull_number: pr_number,
+                body,
+              });
+              console.log('Updated PR body');
+            } else {
+              console.log('No changes needed');
+            }
+
+  reminder-comment:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Post welcome comment for first-time contributors
+        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+        with:
+          script: |
+            const { owner, repo } = context.repo;
+            const prAuthor = context.payload.pull_request.user.login;
+
+            const { data: searchResults } = await github.rest.search.issuesAndPullRequests({
+              q: `repo:${owner}/${repo} type:pr author:${prAuthor}`,
+              per_page: 1,
+            });
+
+            const authorPRCount = searchResults.total_count;
+            console.log(`Found ${authorPRCount} PRs by ${prAuthor}`);
+
+            if (authorPRCount === 1) {
+              console.log(`Posting welcome comment for first-time contributor: ${prAuthor}`);
+              await github.rest.issues.createComment({
+                owner,
+                repo,
+                issue_number: context.issue.number,
+                body: [
+                  '\u{1f44b} Hi! Thank you for contributing to the vLLM project.',
+                  '',
+                  '\u{1f4ac} Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.',
+                  '',
+                  'Just a reminder: PRs would not trigger full CI run by default.',
+                  '',
+                  'Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.',
+                  '',
+                  'To run CI, PR reviewers can either: Add `ready` label to the PR or enable auto-merge.',
+                  '',
+                  'If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.',
+                  '',
+                  '\u{1f680}',
+                ].join('\n'),
+              });
+            } else {
+              console.log(`Skipping comment for ${prAuthor} - not their first PR (${authorPRCount} PRs found)`);
+            }

.github/workflows/pre-commit.yml

@@ -11,9 +11,39 @@ concurrency:
 
 permissions:
   contents: read
+  pull-requests: read
 
 jobs:
+  pre-run-check:
+    if: github.event_name == 'pull_request'
+    runs-on: ubuntu-latest
+    steps:
+    - name: Check PR label and author merge count
+      uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+      with:
+        script: |
+          const { data: pr } = await github.rest.pulls.get({
+            ...context.repo,
+            pull_number: context.payload.pull_request.number,
+          });
+
+          const hasReadyLabel = pr.labels.some(l => l.name === 'ready');
+
+          const { data: mergedPRs } = await github.rest.search.issuesAndPullRequests({
+            q: `repo:${context.repo.owner}/${context.repo.repo} is:pr is:merged author:${pr.user.login}`,
+            per_page: 4,
+          });
+          const mergedCount = mergedPRs.total_count;
+
+          if (hasReadyLabel || mergedCount >= 4) {
+            core.info(`Check passed: ready label=${hasReadyLabel}, 4+ merged PRs=${mergedCount >= 4}`);
+          } else {
+            core.setFailed(`PR must have the 'ready' label or the author must have at least 4 merged PRs (found ${mergedCount}).`);
+          }
+
   pre-commit:
+    needs: pre-run-check
+    if: always() && (needs.pre-run-check.result == 'success' || needs.pre-run-check.result == 'skipped')
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1

.github/workflows/reminder_comment.yml

@@ -1,54 +0,0 @@
-name: PR Reminder Comment Bot
-permissions:
-  pull-requests: write
-on:
-  pull_request_target:
-    types: [opened]
-jobs:
-  pr_reminder:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Remind to run full CI on PR
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
-        with:
-          script: |
-            try {
-              // Get the PR author
-              const prAuthor = context.payload.pull_request.user.login;
-              
-              // Check if this is the author's first PR in this repository
-              // Use GitHub's search API to find all PRs by this author
-              const { data: searchResults } = await github.rest.search.issuesAndPullRequests({
-                q: `repo:${context.repo.owner}/${context.repo.repo} type:pr author:${prAuthor}`,
-                per_page: 100  
-              });
-              
-              const authorPRCount = searchResults.total_count;
-              
-              console.log(`Found ${authorPRCount} PRs by ${prAuthor}`);
-              
-              // Only post comment if this is the first PR (only one PR by this author)
-              if (authorPRCount === 1) {
-                console.log(`Posting welcome comment for first-time contributor: ${prAuthor}`);
-                await github.rest.issues.createComment({
-                owner: context.repo.owner,
-                repo: context.repo.repo,
-                issue_number: context.issue.number,
-                body: '👋 Hi! Thank you for contributing to the vLLM project.\n\n' +
-                  '💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.\n\n' +
-                  'Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run `fastcheck` CI which starts running only a small and essential subset of CI tests to quickly catch errors. \n\n' +
-                  'You ask your reviewers to trigger select CI tests on top of `fastcheck` CI. \n\n' +
-                  'Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.\n\n' +
-                  'To run CI, PR reviewers can either: Add `ready` label to the PR or enable auto-merge.\n\n' +
-                  'If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.\n\n' +
-                  '🚀'
-                });
-              } else {
-                console.log(`Skipping comment for ${prAuthor} - not their first PR (${authorPRCount} PRs found)`);
-              }
-            } catch (error) {
-              console.error('Error checking PR history or posting comment:', error);
-              // Don't fail the workflow, just log the error
-            }
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.gitignore

@@ -189,11 +189,9 @@ cython_debug/
 .vscode/
 
 # Claude
-CLAUDE.md
 .claude/
 
 # Codex
-AGENTS.md
 .codex/
 
 # Cursor

.pre-commit-config.yaml

@@ -30,6 +30,7 @@ repos:
   - id: markdownlint-cli2
     language_version: lts
     args: [--fix]
+    exclude: ^CLAUDE\.md$
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:

AGENTS.md

@@ -0,0 +1,113 @@
+# Agent Instructions for vLLM
+
+> These instructions apply to **all** AI-assisted contributions to `vllm-project/vllm`.
+> Breaching these guidelines can result in automatic banning.
+
+## 1. Contribution Policy (Mandatory)
+
+### Duplicate-work checks
+
+Before proposing a PR, run these checks:
+
+```bash
+gh issue view <issue_number> --repo vllm-project/vllm --comments
+gh pr list --repo vllm-project/vllm --state open --search "<issue_number> in:body"
+gh pr list --repo vllm-project/vllm --state open --search "<short area keywords>"
+```
+
+- If an open PR already addresses the same fix, do not open another.
+- If your approach is materially different, explain the difference in the issue.
+
+### No low-value busywork PRs
+
+Do not open one-off PRs for tiny edits (single typo, isolated style change, one mutable default, etc.). Mechanical cleanups are acceptable only when bundled with substantive work.
+
+### Accountability
+
+- Pure code-agent PRs are **not allowed**. A human submitter must understand and defend the change end-to-end.
+- The submitting human must review every changed line and run relevant tests.
+- PR descriptions for AI-assisted work **must** include:
+    - Why this is not duplicating an existing PR.
+    - Test commands run and results.
+    - Clear statement that AI assistance was used.
+
+### Fail-closed behavior
+
+If work is duplicate/trivial busywork, **do not proceed**. Return a short explanation of what is missing.
+
+---
+
+## 2. Development Workflow
+
+### Environment setup
+
+```bash
+# Install `uv` if you don't have it already:
+curl -LsSf https://astral.sh/uv/install.sh | sh
+
+# Always use `uv` for Python environment management:
+uv venv --python 3.12
+source .venv/bin/activate
+
+# Always make sure `pre-commit` and its hooks are installed:
+uv pip install -r requirements/lint.txt
+pre-commit install
+```
+
+### Installing dependencies
+
+```bash
+# If you are only making Python changes:
+VLLM_USE_PRECOMPILED=1 uv pip install -e .
+
+# If you are also making C/C++ changes:
+uv pip install -e .
+```
+
+### Running tests
+
+Tests require extra dependencies.
+All versions for test dependencies should be read from `requirements/test.txt`
+
+```bash
+# Install bare minimum test dependencies:
+uv pip install pytest pytest-asyncio tblib
+
+# Install additional test dependencies as needed, or install them all as follows:
+uv pip install -r requirements/test.txt
+
+# Run specific test from specific test file
+pytest tests/path/to/test.py -v -s -k test_name
+
+# Run all tests in directory
+pytest tests/path/to/dir -v -s
+```
+
+### Running linters
+
+```bash
+# Run all pre-commit hooks on staged files:
+pre-commit run
+
+# Run on all files:
+pre-commit run --all-files
+
+# Run a specific hook:
+pre-commit run ruff-check --all-files
+
+# Run mypy as it is in CI:
+pre-commit run mypy-3.10 --all-files --hook-stage manual
+```
+
+### Commit messages
+
+Add attribution using commit trailers such as `Co-authored-by:` (other projects use `Assisted-by:` or `Generated-by:`). For example:
+
+```text
+Your commit message here
+
+Co-authored-by: GitHub Copilot
+Co-authored-by: Claude
+Co-authored-by: gemini-code-assist
+Signed-off-by: Your Name <your.email@example.com>
+```

CLAUDE.md

@@ -0,0 +1 @@
+@AGENTS.md

CMakeLists.txt

@@ -340,7 +340,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   list(APPEND VLLM_EXT_SRC
     "csrc/quantization/awq/gemm_kernels.cu"
-    "csrc/permute_cols.cu"
     "csrc/quantization/w8a8/cutlass/scaled_mm_entry.cu"
     "csrc/quantization/fp4/nvfp4_quant_entry.cu"
     "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
@@ -986,6 +985,48 @@ define_extension_target(
 # Setting this variable sidesteps the issue by calling the driver directly.
 target_compile_definitions(_C PRIVATE CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
 
+# add OR VLLM_GPU_LANG STREQUAL "HIP" here once
+# https://github.com/vllm-project/vllm/issues/35163 is resolved
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  #
+  # _C_stable_libtorch extension (ops registered via STABLE_TORCH_LIBRARY)
+  #
+  set(VLLM_STABLE_EXT_SRC
+    "csrc/libtorch_stable/torch_bindings.cpp")
+
+  if(VLLM_GPU_LANG STREQUAL "CUDA")
+    list(APPEND VLLM_STABLE_EXT_SRC "csrc/libtorch_stable/permute_cols.cu")
+  endif()
+
+  if(VLLM_GPU_LANG STREQUAL "CUDA")
+    set_gencode_flags_for_srcs(
+      SRCS "${VLLM_STABLE_EXT_SRC}"
+      CUDA_ARCHS "${CUDA_ARCHS}")
+  endif()
+
+  message(STATUS "Enabling C_stable extension.")
+  define_extension_target(
+    _C_stable_libtorch
+    DESTINATION vllm
+    LANGUAGE ${VLLM_GPU_LANG}
+    SOURCES ${VLLM_STABLE_EXT_SRC}
+    COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+    ARCHITECTURES ${VLLM_GPU_ARCHES}
+    USE_SABI 3
+    WITH_SOABI)
+
+  # Set TORCH_TARGET_VERSION for stable ABI compatibility.
+  # This ensures we only use C-shim APIs available in PyTorch 2.10.
+  # _C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
+  # which is currently set to 2.10.
+  target_compile_definitions(_C_stable_libtorch PRIVATE
+    TORCH_TARGET_VERSION=0x020A000000000000ULL)
+
+  # Needed to use cuda APIs from C-shim
+  target_compile_definitions(_C_stable_libtorch PRIVATE
+    USE_CUDA)
+endif()
+
 #
 # _moe_C extension
 #
@@ -999,6 +1040,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_MOE_EXT_SRC
     "csrc/moe/moe_wna16.cu"
     "csrc/moe/grouped_topk_kernels.cu"
+    "csrc/moe/gpt_oss_router_gemm.cu"
     "csrc/moe/router_gemm.cu")
 endif()
 

benchmarks/attention_benchmarks/benchmark.py

@@ -47,6 +47,8 @@ from common import (
     is_mla_backend,
 )
 
+from vllm.v1.worker.workspace import init_workspace_manager
+
 
 def run_standard_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
     """Run standard attention benchmark (Flash/Triton/FlashInfer)."""
@@ -59,7 +61,9 @@ def run_mla_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
     """Run MLA benchmark with appropriate backend."""
     from mla_runner import run_mla_benchmark as run_mla
 
-    return run_mla(config.backend, config, **kwargs)
+    return run_mla(
+        config.backend, config, prefill_backend=config.prefill_backend, **kwargs
+    )
 
 
 def run_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
@@ -440,20 +444,27 @@ def main():
     # Backend selection
     parser.add_argument(
         "--backends",
+        "--decode-backends",
         nargs="+",
-        help="Backends to benchmark (flash, triton, flashinfer, cutlass_mla, "
+        help="Decode backends to benchmark (flash, triton, flashinfer, cutlass_mla, "
         "flashinfer_mla, flashattn_mla, flashmla)",
     )
     parser.add_argument(
         "--backend",
         help="Single backend (alternative to --backends)",
     )
+    parser.add_argument(
+        "--prefill-backends",
+        nargs="+",
+        help="Prefill backends to compare (fa2, fa3, fa4). "
+        "Uses the first decode backend for impl construction.",
+    )
 
     # Batch specifications
     parser.add_argument(
         "--batch-specs",
         nargs="+",
-        default=["q2k", "8q1s1k"],
+        default=None,
         help="Batch specifications using extended grammar",
     )
 
@@ -469,6 +480,21 @@ def main():
     parser.add_argument("--repeats", type=int, default=1, help="Repetitions")
     parser.add_argument("--warmup-iters", type=int, default=3, help="Warmup iterations")
     parser.add_argument("--profile-memory", action="store_true", help="Profile memory")
+    parser.add_argument(
+        "--kv-cache-dtype",
+        default="auto",
+        choices=["auto", "fp8"],
+        help="KV cache dtype: auto or fp8",
+    )
+    parser.add_argument(
+        "--cuda-graphs",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+        help=(
+            "Launch kernels with CUDA graphs to eliminate CPU overhead"
+            "in measurements (default: True)"
+        ),
+    )
 
     # Parameter sweep (use YAML config for advanced sweeps)
     parser.add_argument(
@@ -502,7 +528,7 @@ def main():
 
         # Override args with YAML values, but CLI args take precedence
         # Check if CLI provided backends (they would be non-None and not default)
-        cli_backends_provided = args.backends is not None or args.backend is not None
+        cli_backends_provided = args.backend is not None or args.backends is not None
 
         # Backend(s) - only use YAML if CLI didn't specify
         if not cli_backends_provided:
@@ -512,6 +538,12 @@ def main():
             elif "backends" in yaml_config:
                 args.backends = yaml_config["backends"]
                 args.backend = None
+            elif "decode_backends" in yaml_config:
+                args.backends = yaml_config["decode_backends"]
+                args.backend = None
+
+        # Prefill backends (e.g., ["fa3", "fa4"])
+        args.prefill_backends = yaml_config.get("prefill_backends", None)
 
         # Check for special modes
         if "mode" in yaml_config:
@@ -521,21 +553,24 @@ def main():
 
         # Batch specs and sizes
         # Support both explicit batch_specs and generated batch_spec_ranges
-        if "batch_spec_ranges" in yaml_config:
-            # Generate batch specs from ranges
-            generated_specs = generate_batch_specs_from_ranges(
-                yaml_config["batch_spec_ranges"]
-            )
-            # Combine with any explicit batch_specs
-            if "batch_specs" in yaml_config:
-                args.batch_specs = yaml_config["batch_specs"] + generated_specs
-            else:
-                args.batch_specs = generated_specs
-            console.print(
-                f"[dim]Generated {len(generated_specs)} batch specs from ranges[/]"
-            )
-        elif "batch_specs" in yaml_config:
-            args.batch_specs = yaml_config["batch_specs"]
+        # CLI --batch-specs takes precedence over YAML when provided.
+        cli_batch_specs_provided = args.batch_specs is not None
+        if not cli_batch_specs_provided:
+            if "batch_spec_ranges" in yaml_config:
+                # Generate batch specs from ranges
+                generated_specs = generate_batch_specs_from_ranges(
+                    yaml_config["batch_spec_ranges"]
+                )
+                # Combine with any explicit batch_specs
+                if "batch_specs" in yaml_config:
+                    args.batch_specs = yaml_config["batch_specs"] + generated_specs
+                else:
+                    args.batch_specs = generated_specs
+                console.print(
+                    f"[dim]Generated {len(generated_specs)} batch specs from ranges[/]"
+                )
+            elif "batch_specs" in yaml_config:
+                args.batch_specs = yaml_config["batch_specs"]
 
         if "batch_sizes" in yaml_config:
             args.batch_sizes = yaml_config["batch_sizes"]
@@ -560,6 +595,10 @@ def main():
             args.warmup_iters = yaml_config["warmup_iters"]
         if "profile_memory" in yaml_config:
             args.profile_memory = yaml_config["profile_memory"]
+        if "kv_cache_dtype" in yaml_config:
+            args.kv_cache_dtype = yaml_config["kv_cache_dtype"]
+        if "cuda_graphs" in yaml_config:
+            args.cuda_graphs = yaml_config["cuda_graphs"]
 
         # Parameter sweep configuration
         if "parameter_sweep" in yaml_config:
@@ -613,10 +652,19 @@ def main():
 
     # Determine backends
     backends = args.backends or ([args.backend] if args.backend else ["flash"])
+    prefill_backends = getattr(args, "prefill_backends", None)
+    if not args.batch_specs:
+        args.batch_specs = ["q2k", "8q1s1k"]
     console.print(f"Backends: {', '.join(backends)}")
+    if prefill_backends:
+        console.print(f"Prefill backends: {', '.join(prefill_backends)}")
     console.print(f"Batch specs: {', '.join(args.batch_specs)}")
+    console.print(f"KV cache dtype: {args.kv_cache_dtype}")
+    console.print(f"CUDA graphs: {args.cuda_graphs}")
     console.print()
 
+    init_workspace_manager(args.device)
+
     # Run benchmarks
     all_results = []
 
@@ -669,6 +717,8 @@ def main():
                         repeats=args.repeats,
                         warmup_iters=args.warmup_iters,
                         profile_memory=args.profile_memory,
+                        kv_cache_dtype=args.kv_cache_dtype,
+                        use_cuda_graphs=args.cuda_graphs,
                     )
 
                     # Add decode pipeline config
@@ -821,6 +871,8 @@ def main():
             "repeats": args.repeats,
             "warmup_iters": args.warmup_iters,
             "profile_memory": args.profile_memory,
+            "kv_cache_dtype": args.kv_cache_dtype,
+            "use_cuda_graphs": args.cuda_graphs,
         }
         all_results = run_model_parameter_sweep(
             backends,
@@ -843,6 +895,8 @@ def main():
             "repeats": args.repeats,
             "warmup_iters": args.warmup_iters,
             "profile_memory": args.profile_memory,
+            "kv_cache_dtype": args.kv_cache_dtype,
+            "use_cuda_graphs": args.cuda_graphs,
         }
         all_results = run_parameter_sweep(
             backends, args.batch_specs, base_config_args, args.parameter_sweep, console
@@ -850,37 +904,95 @@ def main():
 
     else:
         # Normal mode: compare backends
-        total = len(backends) * len(args.batch_specs)
+        decode_results = []
+        prefill_results = []
 
-        with tqdm(total=total, desc="Benchmarking") as pbar:
-            for spec in args.batch_specs:
-                for backend in backends:
-                    config = BenchmarkConfig(
-                        backend=backend,
-                        batch_spec=spec,
-                        num_layers=args.num_layers,
-                        head_dim=args.head_dim,
-                        num_q_heads=args.num_q_heads,
-                        num_kv_heads=args.num_kv_heads,
-                        block_size=args.block_size,
-                        device=args.device,
-                        repeats=args.repeats,
-                        warmup_iters=args.warmup_iters,
-                        profile_memory=args.profile_memory,
-                    )
+        # Run decode backend comparison
+        if not prefill_backends:
+            # No prefill backends specified: compare decode backends as before
+            total = len(backends) * len(args.batch_specs)
 
-                    result = run_benchmark(config)
-                    all_results.append(result)
+            with tqdm(total=total, desc="Benchmarking") as pbar:
+                for spec in args.batch_specs:
+                    for backend in backends:
+                        config = BenchmarkConfig(
+                            backend=backend,
+                            batch_spec=spec,
+                            num_layers=args.num_layers,
+                            head_dim=args.head_dim,
+                            num_q_heads=args.num_q_heads,
+                            num_kv_heads=args.num_kv_heads,
+                            block_size=args.block_size,
+                            device=args.device,
+                            repeats=args.repeats,
+                            warmup_iters=args.warmup_iters,
+                            profile_memory=args.profile_memory,
+                            kv_cache_dtype=args.kv_cache_dtype,
+                            use_cuda_graphs=args.cuda_graphs,
+                        )
 
-                    if not result.success:
-                        console.print(f"[red]Error {backend} {spec}: {result.error}[/]")
+                        result = run_benchmark(config)
+                        decode_results.append(result)
 
-                    pbar.update(1)
+                        if not result.success:
+                            console.print(
+                                f"[red]Error {backend} {spec}: {result.error}[/]"
+                            )
 
-        # Display results
-        console.print("\n[bold green]Results:[/]")
-        formatter = ResultsFormatter(console)
-        formatter.print_table(all_results, backends)
+                        pbar.update(1)
+
+            console.print("\n[bold green]Results:[/]")
+            formatter = ResultsFormatter(console)
+            formatter.print_table(decode_results, backends)
+
+        # Run prefill backend comparison
+        if prefill_backends:
+            # Use first decode backend for impl construction
+            decode_backend = backends[0]
+            total = len(prefill_backends) * len(args.batch_specs)
+
+            console.print(
+                f"[yellow]Prefill comparison mode: "
+                f"using {decode_backend} for decode impl[/]"
+            )
+
+            with tqdm(total=total, desc="Prefill benchmarking") as pbar:
+                for spec in args.batch_specs:
+                    for pb in prefill_backends:
+                        config = BenchmarkConfig(
+                            backend=decode_backend,
+                            batch_spec=spec,
+                            num_layers=args.num_layers,
+                            head_dim=args.head_dim,
+                            num_q_heads=args.num_q_heads,
+                            num_kv_heads=args.num_kv_heads,
+                            block_size=args.block_size,
+                            device=args.device,
+                            repeats=args.repeats,
+                            warmup_iters=args.warmup_iters,
+                            profile_memory=args.profile_memory,
+                            prefill_backend=pb,
+                        )
+
+                        result = run_benchmark(config)
+
+                        # Label result with prefill backend name for display
+                        labeled_config = replace(result.config, backend=pb)
+                        result = replace(result, config=labeled_config)
+                        prefill_results.append(result)
+
+                        if not result.success:
+                            console.print(f"[red]Error {pb} {spec}: {result.error}[/]")
+
+                        pbar.update(1)
+
+            console.print("\n[bold green]Prefill Backend Results:[/]")
+            formatter = ResultsFormatter(console)
+            formatter.print_table(
+                prefill_results, prefill_backends, compare_to_fastest=True
+            )
+
+        all_results = decode_results + prefill_results
 
     # Save results
     if all_results:

benchmarks/attention_benchmarks/common.py

@@ -77,6 +77,7 @@ class MockKVBProj:
         self.qk_nope_head_dim = qk_nope_head_dim
         self.v_head_dim = v_head_dim
         self.out_dim = qk_nope_head_dim + v_head_dim
+        self.weight = torch.empty(0, dtype=torch.bfloat16)
 
     def __call__(self, x: torch.Tensor) -> tuple[torch.Tensor]:
         """
@@ -212,7 +213,11 @@ class BenchmarkConfig:
     profile_memory: bool = False
     use_cuda_graphs: bool = False
 
+    # "auto" or "fp8"
+    kv_cache_dtype: str = "auto"
+
     # MLA-specific
+    prefill_backend: str | None = None
     kv_lora_rank: int | None = None
     qk_nope_head_dim: int | None = None
     qk_rope_head_dim: int | None = None
@@ -367,6 +372,7 @@ class ResultsFormatter:
                     "backend",
                     "batch_spec",
                     "num_layers",
+                    "kv_cache_dtype",
                     "mean_time",
                     "std_time",
                     "throughput",
@@ -380,6 +386,7 @@ class ResultsFormatter:
                         "backend": r.config.backend,
                         "batch_spec": r.config.batch_spec,
                         "num_layers": r.config.num_layers,
+                        "kv_cache_dtype": r.config.kv_cache_dtype,
                         "mean_time": r.mean_time,
                         "std_time": r.std_time,
                         "throughput": r.throughput_tokens_per_sec or 0,

benchmarks/attention_benchmarks/configs/mla_mixed_batch.yaml

@@ -30,9 +30,9 @@ batch_specs:
   - "2q16k_32q1s4k"         # 2 very large prefill + 32 decode
 
   # Context extension + decode
-  - "2q1kkv2k_16q1s1k"       # 2 extend + 16 decode
-  - "4q2kkv4k_32q1s2k"       # 4 extend + 32 decode
-  - "2q1kkv8k_32q1s2k"       # 2 large extend + 32 decode
+  - "2q1ks2k_16q1s1k"       # 2 extend + 16 decode
+  - "4q2ks4k_32q1s2k"       # 4 extend + 32 decode
+  - "2q1ks8k_32q1s2k"       # 2 large extend + 32 decode
 
   # Explicitly chunked prefill
   - "q8k"           # 8k prefill with chunking hint

benchmarks/attention_benchmarks/configs/mla_prefill.yaml

@@ -1,4 +1,19 @@
-# MLA prefill-only benchmark configuration for sparse backends
+# MLA prefill backend comparison
+#
+# Compares all available MLA prefill backends:
+#   FA backends:  fa2, fa3, fa4 (FlashAttention versions)
+#   Non-FA:       flashinfer, cudnn, trtllm (Blackwell-only, require flashinfer)
+#
+# Uses cutlass_mla as the decode backend for impl construction
+# (only the prefill path is exercised).
+#
+# Backends that aren't available on the current platform will report errors
+# in the results table (e.g., fa3 on Blackwell, cudnn without artifactory).
+#
+# Usage:
+#   python benchmark.py --config configs/mla_prefill.yaml
+
+description: "MLA prefill backend comparison"
 
 model:
   name: "deepseek-v3"
@@ -12,20 +27,25 @@ model:
   v_head_dim: 128
   block_size: 128
 
-# Model parameter sweep: simulate tensor parallelism by varying num_q_heads
-# TP=1: 128 heads, TP=2: 64 heads, TP=4: 32 heads, TP=8: 16 heads
-model_parameter_sweep:
-  param_name: "num_q_heads"
-  values: [128, 64, 32, 16]
-  label_format: "{backend}_{value}h"
+# model:
+#   name: "deepseek-v2-lite"
+#   num_layers: 27
+#   num_q_heads: 16
+#   num_kv_heads: 1
+#   head_dim: 576
+#   kv_lora_rank: 512
+#   qk_nope_head_dim: 128
+#   qk_rope_head_dim: 64
+#   v_head_dim: 128
+#   block_size: 128
 
 batch_specs:
   # Pure prefill
-  - "1q512"
-  - "1q1k"
-  - "1q2k"
-  - "1q4k"
-  - "1q8k"
+  - "q512"
+  - "q1k"
+  - "q2k"
+  - "q4k"
+  - "q8k"
 
   # Batched pure prefill
   - "2q512"
@@ -44,19 +64,63 @@ batch_specs:
   - "8q4k"
   - "8q8k"
 
-  # Extend
-  - "1q512s4k"
-  - "1q512s8k"
-  - "1q1ks8k"
-  - "1q2ks8k"
-  - "1q2ks16k"
-  - "1q4ks16k"
+  # Chunked prefill / extend
+  # Short context
+  - "q128s1k"
+  - "q256s2k"
+  - "q512s4k"
+  - "q1ks4k"
+  - "q2ks8k"
+  - "2q128s1k"
+  - "2q256s2k"
+  - "2q512s4k"
+  - "2q1ks4k"
+  - "2q2ks8k"
+  - "4q128s1k"
+  - "4q256s2k"
+  - "4q512s4k"
+  - "4q1ks4k"
+  - "4q2ks8k"
+  - "8q128s1k"
+  - "8q256s2k"
+  - "8q512s4k"
+  - "8q1ks4k"
 
-backends:
-  - FLASHMLA_SPARSE
-  - FLASHINFER_MLA_SPARSE
+  # Medium context
+  - "q128s16k"
+  - "q512s16k"
+  - "q1ks16k"
+  - "q2ks16k"
+  - "2q128s16k"
+  - "2q512s16k"
+  - "2q1ks16k"
+  - "2q2ks16k"
+  - "4q128s16k"
+  - "4q512s16k"
+  - "4q1ks16k"
+  - "4q2ks16k"
+
+  # Long context
+  - "q128s64k"
+  - "q512s64k"
+  - "q1ks64k"
+  - "q2ks64k"
+  - "2q128s64k"
+  - "2q512s64k"
+  - "2q1ks64k"
+  - "2q2ks64k"
+
+decode_backends:
+  - CUTLASS_MLA
+
+prefill_backends:
+  - fa2
+  - fa3
+  - fa4
+  - flashinfer
+  - cudnn
+  - trtllm
 
 device: "cuda:0"
-repeats: 10
-warmup_iters: 3
-profile_memory: true
+repeats: 20
+warmup_iters: 5

benchmarks/attention_benchmarks/configs/mla_sparse_decode.yaml

@@ -0,0 +1,58 @@
+# MLA decode-only benchmark configuration
+
+model:
+  name: "deepseek-v3"
+  num_layers: 60
+  num_q_heads: 128  # Base value, can be swept for TP simulation
+  num_kv_heads: 1  # MLA uses single latent KV
+  head_dim: 576
+  kv_lora_rank: 512
+  qk_nope_head_dim: 128
+  qk_rope_head_dim: 64
+  v_head_dim: 128
+  block_size: 128  # CUTLASS MLA and FlashAttn MLA use 128
+
+# Model parameter sweep: simulate tensor parallelism by varying num_q_heads
+# TP=1: 128 heads, TP=2: 64 heads, TP=4: 32 heads, TP=8: 16 heads
+model_parameter_sweep:
+  param_name: "num_q_heads"
+  values: [128, 64, 32, 16]
+  label_format: "{backend}_{value}h"
+
+batch_specs:
+  # Small batches, varying sequence lengths
+  - "16q1s512"     # 16 requests, 512 KV cache
+  - "16q1s1k"      # 16 requests, 1k KV cache
+  - "16q1s2k"      # 16 requests, 2k KV cache
+  - "16q1s4k"      # 16 requests, 4k KV cache
+
+  # Medium batches
+  - "32q1s1k"      # 32 requests, 1k KV cache
+  - "32q1s2k"      # 32 requests, 2k KV cache
+  - "32q1s4k"      # 32 requests, 4k KV cache
+  - "32q1s8k"      # 32 requests, 8k KV cache
+
+  # Large batches
+  - "64q1s1k"      # 64 requests, 1k KV cache
+  - "64q1s2k"      # 64 requests, 2k KV cache
+  - "64q1s4k"      # 64 requests, 4k KV cache
+  - "64q1s8k"      # 64 requests, 8k KV cache
+
+  # Very large batches
+  - "128q1s1k"     # 128 requests, 1k KV cache
+  - "128q1s2k"     # 128 requests, 2k KV cache
+  - "128q1s4k"     # 128 requests, 4k KV cache
+  - "128q1s8k"     # 128 requests, 8k KV cache
+
+  # Long context
+  - "32q1s16k"     # 32 requests, 16k KV cache
+  - "32q1s32k"     # 32 requests, 32k KV cache
+
+backends:
+  - FLASHMLA_SPARSE
+  - FLASHINFER_MLA_SPARSE
+
+device: "cuda:0"
+repeats: 100
+warmup_iters: 10
+profile_memory: true

benchmarks/attention_benchmarks/configs/mla_sparse_prefill.yaml

@@ -0,0 +1,62 @@
+# MLA prefill-only benchmark configuration for sparse backends
+
+model:
+  name: "deepseek-v3"
+  num_layers: 60
+  num_q_heads: 128
+  num_kv_heads: 1
+  head_dim: 576
+  kv_lora_rank: 512
+  qk_nope_head_dim: 128
+  qk_rope_head_dim: 64
+  v_head_dim: 128
+  block_size: 128
+
+# Model parameter sweep: simulate tensor parallelism by varying num_q_heads
+# TP=1: 128 heads, TP=2: 64 heads, TP=4: 32 heads, TP=8: 16 heads
+model_parameter_sweep:
+  param_name: "num_q_heads"
+  values: [128, 64, 32, 16]
+  label_format: "{backend}_{value}h"
+
+batch_specs:
+  # Pure prefill
+  - "1q512"
+  - "1q1k"
+  - "1q2k"
+  - "1q4k"
+  - "1q8k"
+
+  # Batched pure prefill
+  - "2q512"
+  - "2q1k"
+  - "2q2k"
+  - "2q4k"
+  - "2q8k"
+  - "4q512"
+  - "4q1k"
+  - "4q2k"
+  - "4q4k"
+  - "4q8k"
+  - "8q512"
+  - "8q1k"
+  - "8q2k"
+  - "8q4k"
+  - "8q8k"
+
+  # Extend
+  - "1q512s4k"
+  - "1q512s8k"
+  - "1q1ks8k"
+  - "1q2ks8k"
+  - "1q2ks16k"
+  - "1q4ks16k"
+
+backends:
+  - FLASHMLA_SPARSE
+  - FLASHINFER_MLA_SPARSE
+
+device: "cuda:0"
+repeats: 10
+warmup_iters: 3
+profile_memory: true

benchmarks/attention_benchmarks/mla_runner.py

@@ -60,8 +60,11 @@ def create_minimal_vllm_config(
     model_name: str = "deepseek-v3",
     block_size: int = 128,
     max_num_seqs: int = 256,
+    max_num_batched_tokens: int = 8192,
     mla_dims: dict | None = None,
     index_topk: int | None = None,
+    prefill_backend: str | None = None,
+    kv_cache_dtype: str = "auto",
 ) -> VllmConfig:
     """
     Create minimal VllmConfig for MLA benchmarks.
@@ -75,6 +78,9 @@ def create_minimal_vllm_config(
                   setup_mla_dims(model_name)
         index_topk: Optional topk value for sparse MLA backends. If provided,
                     the config will include index_topk for sparse attention.
+        prefill_backend: Prefill backend name (e.g., "fa3", "fa4", "flashinfer",
+                        "cudnn", "trtllm"). Configures the attention config to
+                        force the specified prefill backend.
 
     Returns:
         VllmConfig for benchmarking
@@ -145,13 +151,13 @@ def create_minimal_vllm_config(
     cache_config = CacheConfig(
         block_size=block_size,
         gpu_memory_utilization=0.9,
-        cache_dtype="auto",
+        cache_dtype=kv_cache_dtype,
         enable_prefix_caching=False,
     )
 
     scheduler_config = SchedulerConfig(
         max_num_seqs=max_num_seqs,
-        max_num_batched_tokens=8192,
+        max_num_batched_tokens=max(max_num_batched_tokens, max_num_seqs),
         max_model_len=32768,
         is_encoder_decoder=False,
         enable_chunked_prefill=True,
@@ -163,7 +169,7 @@ def create_minimal_vllm_config(
 
     compilation_config = CompilationConfig()
 
-    return VllmConfig(
+    vllm_config = VllmConfig(
         model_config=model_config,
         cache_config=cache_config,
         parallel_config=parallel_config,
@@ -171,9 +177,84 @@ def create_minimal_vllm_config(
         compilation_config=compilation_config,
     )
 
+    if prefill_backend is not None:
+        prefill_cfg = get_prefill_backend_config(prefill_backend)
+        if prefill_cfg["flash_attn_version"] is not None:
+            vllm_config.attention_config.flash_attn_version = prefill_cfg[
+                "flash_attn_version"
+            ]
+        vllm_config.attention_config.disable_flashinfer_prefill = prefill_cfg[
+            "disable_flashinfer_prefill"
+        ]
+        vllm_config.attention_config.use_cudnn_prefill = prefill_cfg[
+            "use_cudnn_prefill"
+        ]
+        vllm_config.attention_config.use_trtllm_ragged_deepseek_prefill = prefill_cfg[
+            "use_trtllm_ragged_deepseek_prefill"
+        ]
+
+    return vllm_config
+
 
 # ============================================================================
-# Backend Configuration
+# Prefill Backend Configuration
+# ============================================================================
+
+# Maps prefill backend names to attention config overrides.
+# FA backends set flash_attn_version and disable non-FA paths.
+# Non-FA backends enable their specific path and disable others.
+_PREFILL_BACKEND_CONFIG: dict[str, dict] = {
+    "fa2": {
+        "flash_attn_version": 2,
+        "disable_flashinfer_prefill": True,
+        "use_cudnn_prefill": False,
+        "use_trtllm_ragged_deepseek_prefill": False,
+    },
+    "fa3": {
+        "flash_attn_version": 3,
+        "disable_flashinfer_prefill": True,
+        "use_cudnn_prefill": False,
+        "use_trtllm_ragged_deepseek_prefill": False,
+    },
+    "fa4": {
+        "flash_attn_version": 4,
+        "disable_flashinfer_prefill": True,
+        "use_cudnn_prefill": False,
+        "use_trtllm_ragged_deepseek_prefill": False,
+    },
+    "flashinfer": {
+        "flash_attn_version": None,
+        "disable_flashinfer_prefill": False,
+        "use_cudnn_prefill": False,
+        "use_trtllm_ragged_deepseek_prefill": False,
+    },
+    "cudnn": {
+        "flash_attn_version": None,
+        "disable_flashinfer_prefill": True,
+        "use_cudnn_prefill": True,
+        "use_trtllm_ragged_deepseek_prefill": False,
+    },
+    "trtllm": {
+        "flash_attn_version": None,
+        "disable_flashinfer_prefill": True,
+        "use_cudnn_prefill": False,
+        "use_trtllm_ragged_deepseek_prefill": True,
+    },
+}
+
+
+def get_prefill_backend_config(prefill_backend: str) -> dict:
+    """Get attention config overrides for a prefill backend."""
+    if prefill_backend not in _PREFILL_BACKEND_CONFIG:
+        raise ValueError(
+            f"Unknown prefill backend: {prefill_backend!r}. "
+            f"Available: {list(_PREFILL_BACKEND_CONFIG.keys())}"
+        )
+    return _PREFILL_BACKEND_CONFIG[prefill_backend]
+
+
+# ============================================================================
+# Decode Backend Configuration
 # ============================================================================
 
 
@@ -203,6 +284,7 @@ def _get_backend_config(backend: str) -> dict:
     Returns:
         Dict with backend configuration
     """
+    from vllm.v1.attention.backend import MultipleOf
     from vllm.v1.attention.backends.registry import AttentionBackendEnum
 
     try:
@@ -219,8 +301,8 @@ def _get_backend_config(backend: str) -> dict:
     block_sizes = backend_class.get_supported_kernel_block_sizes()
     # Use first supported block size (backends typically support one for MLA)
     block_size = block_sizes[0] if block_sizes else None
-    if hasattr(block_size, "value"):
-        # Handle MultipleOf enum
+    if isinstance(block_size, MultipleOf):
+        # No fixed block size; fall back to config value
         block_size = None
 
     # Check if sparse via class method if available
@@ -455,6 +537,7 @@ def _create_backend_impl(
     device: torch.device,
     max_num_tokens: int = 8192,
     index_topk: int | None = None,
+    kv_cache_dtype: str = "auto",
 ):
     """
     Create backend implementation instance.
@@ -503,7 +586,7 @@ def _create_backend_impl(
         "num_kv_heads": mla_dims["num_kv_heads"],
         "alibi_slopes": None,
         "sliding_window": None,
-        "kv_cache_dtype": "auto",
+        "kv_cache_dtype": kv_cache_dtype,
         "logits_soft_cap": None,
         "attn_type": "decoder",
         "kv_sharing_target_layer_name": None,
@@ -621,6 +704,7 @@ def _run_single_benchmark(
     mla_dims: dict,
     device: torch.device,
     indexer=None,
+    kv_cache_dtype: str | None = None,
 ) -> BenchmarkResult:
     """
     Run a single benchmark iteration.
@@ -654,54 +738,124 @@ def _run_single_benchmark(
     )
 
     # Create KV cache
-    kv_cache = torch.zeros(
-        num_blocks,
-        block_size,
-        mla_dims["kv_lora_rank"] + mla_dims["qk_rope_head_dim"],
-        device=device,
-        dtype=torch.bfloat16,
-    )
+    if kv_cache_dtype is None:
+        kv_cache_dtype = getattr(config, "kv_cache_dtype", "auto")
+    head_size = mla_dims["kv_lora_rank"] + mla_dims["qk_rope_head_dim"]
+    if kv_cache_dtype == "fp8_ds_mla":
+        # FlashMLA sparse custom format: 656 bytes per token, stored as uint8.
+        # Layout: kv_lora_rank fp8 bytes + 4 float32 tile scales
+        #         + 2*rope_dim bf16 bytes
+        # = 512 + 16 + 128 = 656 bytes for DeepSeek dims.
+        kv_cache = torch.zeros(
+            num_blocks,
+            block_size,
+            656,
+            device=device,
+            dtype=torch.uint8,
+        )
+    elif kv_cache_dtype == "fp8":
+        from vllm.platforms import current_platform
 
-    # Create input tensors for both decode and prefill modes
-    decode_inputs, prefill_inputs = _create_input_tensors(
-        total_q,
-        mla_dims,
-        backend_cfg["query_format"],
-        device,
-        torch.bfloat16,
-    )
+        kv_cache = torch.zeros(
+            num_blocks,
+            block_size,
+            head_size,
+            device=device,
+            dtype=torch.uint8,
+        ).view(current_platform.fp8_dtype())
+    else:
+        kv_cache = torch.zeros(
+            num_blocks,
+            block_size,
+            head_size,
+            device=device,
+            dtype=torch.bfloat16,
+        )
 
     # Fill indexer with random indices for sparse backends
     is_sparse = backend_cfg.get("is_sparse", False)
     if is_sparse and indexer is not None:
         indexer.fill_random_indices(total_q, max_kv_len)
 
-    # Determine which forward method to use
-    if is_sparse:
-        # Sparse backends use forward_mqa
-        forward_fn = lambda: impl.forward_mqa(decode_inputs, kv_cache, metadata, layer)
-    elif metadata.decode is not None:
-        forward_fn = lambda: impl._forward_decode(
-            decode_inputs, kv_cache, metadata, layer
-        )
-    elif metadata.prefill is not None:
-        forward_fn = lambda: impl._forward_prefill(
-            prefill_inputs["q"],
-            prefill_inputs["k_c_normed"],
-            prefill_inputs["k_pe"],
-            kv_cache,
-            metadata,
-            prefill_inputs["k_scale"],
-            prefill_inputs["output"],
-        )
-    else:
+    # Determine which forward methods to use based on metadata.
+    # Sparse MLA backends always use forward_mqa
+    has_decode = is_sparse or getattr(metadata, "decode", None) is not None
+    has_prefill = not is_sparse and getattr(metadata, "prefill", None) is not None
+    if not has_decode and not has_prefill:
         raise RuntimeError("Metadata has neither decode nor prefill metadata")
 
+    num_decode = (
+        metadata.num_decode_tokens
+        if (has_decode and has_prefill)
+        else total_q
+        if has_decode
+        else 0
+    )
+    num_prefill = total_q - num_decode
+
+    # Some backends requires fp8 queries when using fp8 KV cache.
+    is_fp8_kvcache = kv_cache_dtype.startswith("fp8")
+    quantize_query = is_fp8_kvcache and getattr(
+        impl, "supports_quant_query_input", False
+    )
+
+    # quantize_query forces concat format
+    query_fmt = "concat" if quantize_query else backend_cfg["query_format"]
+
+    # Create decode query tensors
+    if has_decode:
+        decode_inputs, _ = _create_input_tensors(
+            num_decode, mla_dims, query_fmt, device, torch.bfloat16
+        )
+        # Cast decode query to fp8 if the backend supports it
+        if quantize_query:
+            from vllm.platforms import current_platform
+
+            if isinstance(decode_inputs, tuple):
+                decode_inputs = torch.cat(list(decode_inputs), dim=-1)
+            decode_inputs = decode_inputs.to(current_platform.fp8_dtype())
+
+    # Create prefill input tensors
+    if has_prefill:
+        _, prefill_inputs = _create_input_tensors(
+            num_prefill, mla_dims, query_fmt, device, torch.bfloat16
+        )
+
+    # Build forward function
+    def forward_fn():
+        results = []
+        if has_decode:
+            results.append(impl.forward_mqa(decode_inputs, kv_cache, metadata, layer))
+        if has_prefill:
+            results.append(
+                impl.forward_mha(
+                    prefill_inputs["q"],
+                    prefill_inputs["k_c_normed"],
+                    prefill_inputs["k_pe"],
+                    kv_cache,
+                    metadata,
+                    prefill_inputs["k_scale"],
+                    prefill_inputs["output"],
+                )
+            )
+        return results[0] if len(results) == 1 else tuple(results)
+
     # Warmup
     for _ in range(config.warmup_iters):
         forward_fn()
     torch.accelerator.synchronize()
 
+    # Optionally capture a CUDA graph after warmup.
+    # Graph replay eliminates CPU launch overhead so timings reflect pure
+    # kernel time.
+    if config.use_cuda_graphs:
+        graph = torch.cuda.CUDAGraph()
+        with torch.cuda.graph(graph):
+            forward_fn()
+        benchmark_fn = graph.replay
+    else:
+        benchmark_fn = forward_fn
+
     # Benchmark
     times = []
     for _ in range(config.repeats):
@@ -710,7 +864,7 @@ def _run_single_benchmark(
 
         start.record()
         for _ in range(config.num_layers):
-            forward_fn()
+            benchmark_fn()
         end.record()
 
         torch.accelerator.synchronize()
@@ -732,6 +886,7 @@ def _run_mla_benchmark_batched(
     backend: str,
     configs_with_params: list[tuple],  # [(config, threshold, num_splits), ...]
     index_topk: int = 2048,
+    prefill_backend: str | None = None,
 ) -> list[BenchmarkResult]:
     """
     Unified batched MLA benchmark runner for all backends.
@@ -743,11 +898,13 @@ def _run_mla_benchmark_batched(
     to avoid setup/teardown overhead.
 
     Args:
-        backend: Backend name
+        backend: Backend name (decode backend used for impl construction)
         configs_with_params: List of (config, threshold, num_splits) tuples
             - threshold: reorder_batch_threshold (FlashAttn/FlashMLA only)
             - num_splits: num_kv_splits (CUTLASS only)
         index_topk: Topk value for sparse MLA backends (default 2048)
+        prefill_backend: Prefill backend name (e.g., "fa3", "fa4").
+            When set, forces the specified FlashAttention version for prefill.
 
     Returns:
         List of BenchmarkResult objects
@@ -757,7 +914,7 @@ def _run_mla_benchmark_batched(
 
     backend_cfg = _get_backend_config(backend)
     device = torch.device(configs_with_params[0][0].device)
-    torch.cuda.set_device(device)
+    torch.accelerator.set_device_index(device)
 
     # Determine block size
     config_block_size = configs_with_params[0][0].block_size
@@ -774,26 +931,91 @@ def _run_mla_benchmark_batched(
     # Determine if this is a sparse backend
     is_sparse = backend_cfg.get("is_sparse", False)
 
+    # Extract kv_cache_dtype from the first config
+    kv_cache_dtype = getattr(first_config, "kv_cache_dtype", "auto")
+
+    # FlashMLA sparse only supports "fp8_ds_mla" internally (not generic "fp8").
+    # Remap here so the user can pass --kv-cache-dtype fp8 regardless of backend.
+    if backend.upper() == "FLASHMLA_SPARSE" and kv_cache_dtype == "fp8":
+        kv_cache_dtype = "fp8_ds_mla"
+
+    # Compute max total_q across all configs so the metadata builder buffer
+    # and scheduler config are large enough for all batch specs.
+    max_total_q = max(
+        sum(r.q_len for r in parse_batch_spec(cfg.batch_spec))
+        for cfg, *_ in configs_with_params
+    )
+
     # Create and set vLLM config for MLA (reused across all benchmarks)
     vllm_config = create_minimal_vllm_config(
         model_name="deepseek-v3",  # Used only for model path
         block_size=block_size,
+        max_num_batched_tokens=max_total_q,
         mla_dims=mla_dims,  # Use custom dims from config or default
         index_topk=index_topk if is_sparse else None,
+        prefill_backend=prefill_backend,
+        kv_cache_dtype=kv_cache_dtype,
     )
 
     results = []
 
     with set_current_vllm_config(vllm_config):
+        # Clear cached prefill backend detection functions so they re-evaluate
+        # with the current VllmConfig. These are @functools.cache decorated and
+        # would otherwise return stale results from a previous backend's config.
+        from vllm.model_executor.layers.attention.mla_attention import (
+            use_cudnn_prefill,
+            use_flashinfer_prefill,
+            use_trtllm_ragged_deepseek_prefill,
+        )
+
+        use_flashinfer_prefill.cache_clear()
+        use_cudnn_prefill.cache_clear()
+        use_trtllm_ragged_deepseek_prefill.cache_clear()
+
         # Create backend impl, layer, builder, and indexer (reused across benchmarks)
         impl, layer, builder_instance, indexer = _create_backend_impl(
             backend_cfg,
             mla_dims,
             vllm_config,
             device,
+            max_num_tokens=max_total_q,
             index_topk=index_topk if is_sparse else None,
+            kv_cache_dtype=kv_cache_dtype,
         )
 
+        # Verify the actual prefill backend matches what was requested
+        if prefill_backend is not None:
+            prefill_cfg = get_prefill_backend_config(prefill_backend)
+            fa_version = prefill_cfg["flash_attn_version"]
+
+            if fa_version is not None:
+                # FA backend: verify the impl's FA version
+                actual_fa_version = getattr(impl, "vllm_flash_attn_version", None)
+                if actual_fa_version != fa_version:
+                    raise RuntimeError(
+                        f"Prefill backend '{prefill_backend}' requested FA "
+                        f"version {fa_version}, but the impl is using FA "
+                        f"version {actual_fa_version}. Check "
+                        f"vllm/v1/attention/backends/fa_utils.py."
+                    )
+            else:
+                # Non-FA backend: verify the builder picked the right path
+                expected_flags = {
+                    "flashinfer": "_use_fi_prefill",
+                    "cudnn": "_use_cudnn_prefill",
+                    "trtllm": "_use_trtllm_ragged_prefill",
+                }
+                flag_name = expected_flags.get(prefill_backend)
+                if flag_name and not getattr(builder_instance, flag_name, False):
+                    raise RuntimeError(
+                        f"Prefill backend '{prefill_backend}' was requested "
+                        f"but the metadata builder did not enable it. This "
+                        f"usually means a dependency is missing (e.g., "
+                        f"flashinfer not installed) or the platform doesn't "
+                        f"support it."
+                    )
+
         # Run each benchmark with the shared impl
         for config, threshold, num_splits in configs_with_params:
             # Set threshold for this benchmark (FlashAttn/FlashMLA only)
@@ -818,6 +1040,7 @@ def _run_mla_benchmark_batched(
                     mla_dims,
                     device,
                     indexer=indexer,
+                    kv_cache_dtype=kv_cache_dtype,
                 )
                 results.append(result)
 
@@ -844,6 +1067,7 @@ def run_mla_benchmark(
     reorder_batch_threshold: int | None = None,
     num_kv_splits: int | None = None,
     index_topk: int = 2048,
+    prefill_backend: str | None = None,
 ) -> BenchmarkResult | list[BenchmarkResult]:
     """
     Unified MLA benchmark runner for all backends.
@@ -861,6 +1085,8 @@ def run_mla_benchmark(
                                  (single config mode only)
         num_kv_splits: Number of KV splits for CUTLASS (single config mode only)
         index_topk: Topk value for sparse MLA backends (default 2048)
+        prefill_backend: Prefill backend name (e.g., "fa3", "fa4").
+            When set, forces the specified FlashAttention version for prefill.
 
     Returns:
         BenchmarkResult (single mode) or list of BenchmarkResult (batched mode)
@@ -884,7 +1110,9 @@ def run_mla_benchmark(
         return_single = True
 
     # Use unified batched execution
-    results = _run_mla_benchmark_batched(backend, configs_with_params, index_topk)
+    results = _run_mla_benchmark_batched(
+        backend, configs_with_params, index_topk, prefill_backend=prefill_backend
+    )
 
     # Return single result or list based on input
     return results[0] if return_single else results

benchmarks/attention_benchmarks/runner.py

@@ -140,7 +140,7 @@ def _create_vllm_config(
 
     cache_config = CacheConfig(
         block_size=config.block_size,
-        cache_dtype="auto",
+        cache_dtype=config.kv_cache_dtype,
     )
     cache_config.num_gpu_blocks = max_num_blocks
     cache_config.num_cpu_blocks = 0
@@ -215,7 +215,7 @@ def _create_backend_impl(
         num_kv_heads=config.num_kv_heads,
         alibi_slopes=None,
         sliding_window=None,
-        kv_cache_dtype="auto",
+        kv_cache_dtype=config.kv_cache_dtype,
     )
 
     kv_cache_spec = FullAttentionSpec(
@@ -288,12 +288,22 @@ def _create_input_tensors(
     total_q: int,
     device: torch.device,
     dtype: torch.dtype,
+    quantize_query: bool = False,
 ) -> tuple:
-    """Create Q, K, V input tensors for all layers."""
+    """Create Q, K, V input tensors for all layers.
+
+    When quantize_query is True, queries are cast to fp8 to match backends
+    that require query/key/value dtype consistency.
+    """
+    q_dtype = dtype
+    if quantize_query:
+        from vllm.platforms import current_platform
+
+        q_dtype = current_platform.fp8_dtype()
     q_list = [
         torch.randn(
             total_q, config.num_q_heads, config.head_dim, device=device, dtype=dtype
-        )
+        ).to(q_dtype)
         for _ in range(config.num_layers)
     ]
     k_list = [
@@ -344,10 +354,17 @@ def _create_kv_cache(
     # Compute inverse permutation to get back to logical view
     inv_order = [stride_order.index(i) for i in range(len(stride_order))]
 
+    # Use fp8 dtype for cache when requested.
+    cache_dtype = dtype
+    if config.kv_cache_dtype == "fp8":
+        from vllm.platforms import current_platform
+
+        cache_dtype = current_platform.fp8_dtype()
+
     cache_list = []
     for _ in range(config.num_layers):
         # Allocate in physical layout order (contiguous in memory)
-        cache = torch.zeros(*physical_shape, device=device, dtype=dtype)
+        cache = torch.zeros(*physical_shape, device=device, dtype=cache_dtype)
         # Permute to logical view
         cache = cache.permute(*inv_order)
         cache_list.append(cache)
@@ -392,6 +409,37 @@ def _run_single_benchmark(
             )
     torch.accelerator.synchronize()
 
+    # Optionally capture a CUDA graph after warmup.
+    # Graph replay eliminates CPU launch overhead so timings reflect pure
+    # kernel time.
+    if config.use_cuda_graphs:
+        graph = torch.cuda.CUDAGraph()
+        with torch.cuda.graph(graph):
+            for i in range(config.num_layers):
+                impl.forward(
+                    layer,
+                    q_list[i],
+                    k_list[i],
+                    v_list[i],
+                    cache_list[i],
+                    attn_metadata,
+                    output=out,
+                )
+        benchmark_fn = graph.replay
+    else:
+
+        def benchmark_fn():
+            for i in range(config.num_layers):
+                impl.forward(
+                    layer,
+                    q_list[i],
+                    k_list[i],
+                    v_list[i],
+                    cache_list[i],
+                    attn_metadata,
+                    output=out,
+                )
+
     # Benchmark
     times = []
     for _ in range(config.repeats):
@@ -399,16 +447,7 @@ def _run_single_benchmark(
         end = torch.cuda.Event(enable_timing=True)
 
         start.record()
-        for i in range(config.num_layers):
-            impl.forward(
-                layer,
-                q_list[i],
-                k_list[i],
-                v_list[i],
-                cache_list[i],
-                attn_metadata,
-                output=out,
-            )
+        benchmark_fn()
         end.record()
 
         torch.accelerator.synchronize()
@@ -418,8 +457,8 @@ def _run_single_benchmark(
     mem_stats = {}
     if config.profile_memory:
         mem_stats = {
-            "allocated_mb": torch.cuda.memory_allocated(device) / 1024**2,
-            "reserved_mb": torch.cuda.memory_reserved(device) / 1024**2,
+            "allocated_mb": torch.accelerator.memory_allocated(device) / 1024**2,
+            "reserved_mb": torch.accelerator.memory_reserved(device) / 1024**2,
         }
 
     return times, mem_stats
@@ -443,7 +482,7 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
         BenchmarkResult with timing and memory statistics
     """
     device = torch.device(config.device)
-    torch.cuda.set_device(device)
+    torch.accelerator.set_device_index(device)
 
     backend_cfg = _get_backend_config(config.backend)
 
@@ -502,8 +541,12 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 common_attn_metadata=common_metadata,
             )
 
+            # Only quantize queries when the impl supports it
+            quantize_query = config.kv_cache_dtype.startswith("fp8") and getattr(
+                impl, "supports_quant_query_input", False
+            )
             q_list, k_list, v_list = _create_input_tensors(
-                config, total_q, device, dtype
+                config, total_q, device, dtype, quantize_query=quantize_query
             )
 
             cache_list = _create_kv_cache(

benchmarks/benchmark_long_document_qa_throughput.py

@@ -40,9 +40,9 @@ LLM engine. You can refer to the `vllm.engine.arg_utils.EngineArgs` for more
 details.
 """
 
-import dataclasses
 import random
 import time
+from dataclasses import fields
 
 from vllm import LLM, SamplingParams
 from vllm.engine.arg_utils import EngineArgs
@@ -124,7 +124,7 @@ def main(args):
 
     # Create the LLM engine
     engine_args = EngineArgs.from_cli_args(args)
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
     sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
 
     print("------warm up------")

benchmarks/benchmark_prefix_caching.py

@@ -32,6 +32,7 @@ import dataclasses
 import json
 import random
 import time
+from dataclasses import fields
 
 from transformers import PreTrainedTokenizerBase
 
@@ -196,7 +197,7 @@ def main(args):
 
     engine_args = EngineArgs.from_cli_args(args)
 
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
 
     sampling_params = SamplingParams(
         temperature=0,

benchmarks/benchmark_prioritization.py

@@ -3,10 +3,10 @@
 """Benchmark offline prioritization."""
 
 import argparse
-import dataclasses
 import json
 import random
 import time
+from dataclasses import fields
 
 from transformers import AutoTokenizer, PreTrainedTokenizerBase
 
@@ -79,7 +79,7 @@ def run_vllm(
 ) -> float:
     from vllm import LLM, SamplingParams
 
-    llm = LLM(**dataclasses.asdict(engine_args))
+    llm = LLM(**{f.name: getattr(engine_args, f.name) for f in fields(engine_args)})
 
     assert all(
         llm.llm_engine.model_config.max_model_len >= (request[1] + request[2])

benchmarks/benchmark_topk_topp.py

@@ -95,13 +95,16 @@ def create_logits(
 def measure_memory() -> tuple[int, int]:
     """Return (allocated, reserved) memory in bytes."""
     torch.accelerator.synchronize()
-    return torch.cuda.memory_allocated(), torch.cuda.max_memory_allocated()
+    return (
+        torch.accelerator.memory_allocated(),
+        torch.accelerator.max_memory_allocated(),
+    )
 
 
 def reset_memory_stats():
     """Reset peak memory statistics."""
     reset_buffer_cache()
-    torch.cuda.reset_peak_memory_stats()
+    torch.accelerator.reset_peak_memory_stats()
     torch.accelerator.empty_cache()
     gc.collect()
 

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -64,7 +64,7 @@ def bench_run(
     per_out_ch: bool,
     mkn: tuple[int, int, int],
 ):
-    init_workspace_manager(torch.cuda.current_device())
+    init_workspace_manager(torch.accelerator.current_device_index())
     (m, k, n) = mkn
 
     dtype = torch.half

benchmarks/kernels/benchmark_device_communicators.py

@@ -495,7 +495,7 @@ def main():
 
     # Set device
     device = torch.device(f"cuda:{rank}")
-    torch.cuda.set_device(device)
+    torch.accelerator.set_device_index(device)
 
     # Get CPU process group
     cpu_group = dist.new_group(backend="gloo")

benchmarks/kernels/benchmark_fused_collective.py

@@ -392,7 +392,7 @@ def benchmark_operation(
     num_op_per_cudagraph = 10
 
     # Use vLLM's graph_capture to make tensor_model_parallel_all_reduce graph-safe
-    device = torch.device(f"cuda:{torch.cuda.current_device()}")
+    device = torch.device(f"cuda:{torch.accelerator.current_device_index()}")
     with graph_capture(device=device), torch.cuda.graph(graph):
         for _ in range(num_op_per_cudagraph):
             operation_func(*args, **kwargs)
@@ -984,7 +984,7 @@ def main():
     world_size = int(os.environ["WORLD_SIZE"])
 
     device = torch.device(f"cuda:{rank}")
-    torch.cuda.set_device(device)
+    torch.accelerator.set_device_index(device)
     torch.set_default_device(device)
 
     init_distributed_environment()

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -50,7 +50,7 @@ def bench_run(
     per_out_ch: bool,
     mkn: tuple[int, int, int],
 ):
-    init_workspace_manager(torch.cuda.current_device())
+    init_workspace_manager(torch.accelerator.current_device_index())
     label = "Quant Matmul"
 
     sub_label = (

benchmarks/kernels/benchmark_moe.py

@@ -750,17 +750,20 @@ def get_weight_block_size_safety(config, default_value=None):
 
 
 def get_model_params(config):
-    if config.architectures[0] == "DbrxForCausalLM":
+    architectures = getattr(config, "architectures", None) or [type(config).__name__]
+    architecture = architectures[0]
+
+    if architecture == "DbrxForCausalLM":
         E = config.ffn_config.moe_num_experts
         topk = config.ffn_config.moe_top_k
         intermediate_size = config.ffn_config.ffn_hidden_size
         hidden_size = config.hidden_size
-    elif config.architectures[0] == "JambaForCausalLM":
+    elif architecture == "JambaForCausalLM":
         E = config.num_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         hidden_size = config.hidden_size
-    elif config.architectures[0] in (
+    elif architecture in (
         "DeepseekV2ForCausalLM",
         "DeepseekV3ForCausalLM",
         "DeepseekV32ForCausalLM",
@@ -774,7 +777,7 @@ def get_model_params(config):
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
         hidden_size = config.hidden_size
-    elif config.architectures[0] in (
+    elif architecture in (
         "Qwen2MoeForCausalLM",
         "Qwen3MoeForCausalLM",
         "Qwen3NextForCausalLM",
@@ -783,23 +786,27 @@ def get_model_params(config):
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
         hidden_size = config.hidden_size
-    elif config.architectures[0] == "Qwen3VLMoeForConditionalGeneration":
+    elif architecture in (
+        "Qwen3VLMoeForConditionalGeneration",
+        "Qwen3_5MoeForConditionalGeneration",
+        "Qwen3_5MoeTextConfig",
+    ):
         text_config = config.get_text_config()
         E = text_config.num_experts
         topk = text_config.num_experts_per_tok
         intermediate_size = text_config.moe_intermediate_size
         hidden_size = text_config.hidden_size
-    elif config.architectures[0] == "HunYuanMoEV1ForCausalLM":
+    elif architecture == "HunYuanMoEV1ForCausalLM":
         E = config.num_experts
         topk = config.moe_topk[0]
         intermediate_size = config.moe_intermediate_size[0]
         hidden_size = config.hidden_size
-    elif config.architectures[0] == "Qwen3OmniMoeForConditionalGeneration":
+    elif architecture == "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
-    elif config.architectures[0] == "PixtralForConditionalGeneration":
+    elif architecture == "PixtralForConditionalGeneration":
         # Pixtral can contain different LLM architectures,
         # recurse to get their parameters
         return get_model_params(config.get_text_config())
@@ -814,6 +821,23 @@ def get_model_params(config):
     return E, topk, intermediate_size, hidden_size
 
 
+def resolve_dtype(config) -> torch.dtype:
+    if current_platform.is_rocm():
+        return torch.float16
+
+    dtype = getattr(config, "dtype", None)
+    if dtype is not None:
+        return dtype
+
+    if hasattr(config, "get_text_config"):
+        text_config = config.get_text_config()
+        dtype = getattr(text_config, "dtype", None)
+        if dtype is not None:
+            return dtype
+
+    return torch.bfloat16
+
+
 def get_quantization_group_size(config) -> int | None:
     """Extract the quantization group size from the HF model config.
 
@@ -861,7 +885,7 @@ def main(args: argparse.Namespace):
     else:
         ensure_divisibility(intermediate_size, args.tp_size, "intermediate_size")
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    dtype = torch.float16 if current_platform.is_rocm() else config.dtype
+    dtype = resolve_dtype(config)
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
     use_int8_w8a16 = args.dtype == "int8_w8a16"
     use_int4_w4a16 = args.dtype == "int4_w4a16"

benchmarks/kernels/benchmark_router_gemm.py

@@ -0,0 +1,134 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import torch
+import torch.nn.functional as F
+
+from vllm import _custom_ops as ops
+from vllm.platforms import current_platform
+from vllm.transformers_utils.config import get_config
+from vllm.triton_utils import triton
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+# Dimensions supported by the DSV3 specialized kernel
+DSV3_SUPPORTED_NUM_EXPERTS = [256, 384]
+DSV3_SUPPORTED_HIDDEN_SIZES = [7168]
+
+# Dimensions supported by the gpt-oss specialized kernel
+GPT_OSS_SUPPORTED_NUM_EXPERTS = [32, 128]
+GPT_OSS_SUPPORTED_HIDDEN_SIZES = [2880]
+
+
+def get_batch_size_range(max_batch_size):
+    return [2**x for x in range(14) if 2**x <= max_batch_size]
+
+
+def get_model_params(config):
+    if config.architectures[0] in (
+        "DeepseekV2ForCausalLM",
+        "DeepseekV3ForCausalLM",
+        "DeepseekV32ForCausalLM",
+    ):
+        num_experts = config.n_routed_experts
+        hidden_size = config.hidden_size
+    elif config.architectures[0] in ("GptOssForCausalLM",):
+        num_experts = config.num_local_experts
+        hidden_size = config.hidden_size
+    else:
+        raise ValueError(f"Unsupported architecture: {config.architectures}")
+    return num_experts, hidden_size
+
+
+def get_benchmark(model, max_batch_size, trust_remote_code):
+    @triton.testing.perf_report(
+        triton.testing.Benchmark(
+            x_names=["batch_size"],
+            x_vals=get_batch_size_range(max_batch_size),
+            x_log=False,
+            line_arg="provider",
+            line_vals=[
+                "torch",
+                "vllm",
+            ],
+            line_names=["PyTorch", "vLLM"],
+            styles=([("blue", "-"), ("red", "-")]),
+            ylabel="TFLOPs",
+            plot_name=f"{model} router gemm throughput",
+            args={},
+        )
+    )
+    def benchmark(batch_size, provider):
+        config = get_config(model=model, trust_remote_code=trust_remote_code)
+        num_experts, hidden_size = get_model_params(config)
+
+        mat_a = torch.randn(
+            (batch_size, hidden_size), dtype=torch.bfloat16, device="cuda"
+        ).contiguous()
+        mat_b = torch.randn(
+            (num_experts, hidden_size), dtype=torch.bfloat16, device="cuda"
+        ).contiguous()
+        bias = torch.randn(
+            num_experts, dtype=torch.bfloat16, device="cuda"
+        ).contiguous()
+
+        is_hopper_or_blackwell = current_platform.is_device_capability(
+            90
+        ) or current_platform.is_device_capability_family(100)
+        allow_dsv3_router_gemm = (
+            is_hopper_or_blackwell
+            and num_experts in DSV3_SUPPORTED_NUM_EXPERTS
+            and hidden_size in DSV3_SUPPORTED_HIDDEN_SIZES
+        )
+        allow_gpt_oss_router_gemm = (
+            is_hopper_or_blackwell
+            and num_experts in GPT_OSS_SUPPORTED_NUM_EXPERTS
+            and hidden_size in GPT_OSS_SUPPORTED_HIDDEN_SIZES
+        )
+
+        has_bias = False
+        if allow_gpt_oss_router_gemm:
+            has_bias = True
+
+        quantiles = [0.5, 0.2, 0.8]
+
+        if provider == "torch":
+
+            def runner():
+                if has_bias:
+                    F.linear(mat_a, mat_b, bias)
+                else:
+                    F.linear(mat_a, mat_b)
+        elif provider == "vllm":
+
+            def runner():
+                if allow_dsv3_router_gemm:
+                    ops.dsv3_router_gemm(mat_a, mat_b, torch.bfloat16)
+                elif allow_gpt_oss_router_gemm:
+                    ops.gpt_oss_router_gemm(mat_a, mat_b, bias)
+                else:
+                    raise ValueError("Unsupported router gemm")
+
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            runner, quantiles=quantiles
+        )
+
+        def tflops(t_ms):
+            flops = 2 * batch_size * hidden_size * num_experts
+            return flops / (t_ms * 1e-3) / 1e12
+
+        return tflops(ms), tflops(max_ms), tflops(min_ms)
+
+    return benchmark
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+    parser.add_argument("--model", type=str, default="openai/gpt-oss-20b")
+    parser.add_argument("--max-batch-size", default=16, type=int)
+    parser.add_argument("--trust-remote-code", action="store_true")
+    args = parser.parse_args()
+
+    # Get the benchmark function
+    benchmark = get_benchmark(args.model, args.max_batch_size, args.trust_remote_code)
+    # Run performance benchmark
+    benchmark.run(print_data=True)

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -285,7 +285,7 @@ def tune_on_gpu(args_dict):
     weight_shapes = args_dict["weight_shapes"]
     args = args_dict["args"]
 
-    torch.cuda.set_device(gpu_id)
+    torch.accelerator.set_device_index(gpu_id)
     print(f"Starting tuning on GPU {gpu_id} with batch sizes {batch_sizes}")
 
     block_n = args.block_n
@@ -334,7 +334,7 @@ def distribute_batch_sizes(batch_sizes, num_gpus):
 
 def main(args):
     print(args)
-    num_gpus = torch.cuda.device_count()
+    num_gpus = torch.accelerator.device_count()
     if num_gpus == 0:
         raise RuntimeError("No GPU available for tuning")
     print(f"Found {num_gpus} GPUs for parallel tuning")

benchmarks/kernels/cpu/benchmark_cpu_attn.py

@@ -27,7 +27,7 @@ def get_attn_isa(
     else:
         if current_platform.get_cpu_architecture() == CpuArchEnum.ARM:
             return "neon"
-        elif torch._C._cpu._is_amx_tile_supported():
+        elif torch.cpu._is_amx_tile_supported():
             return "amx"
         else:
             return "vec"

benchmarks/kernels/cpu/benchmark_cpu_fused_moe.py

@@ -24,7 +24,7 @@ except (ImportError, AttributeError) as e:
     sys.exit(1)
 
 # ISA selection following test_cpu_fused_moe.py pattern
-ISA_CHOICES = ["amx", "vec"] if torch._C._cpu._is_amx_tile_supported() else ["vec"]
+ISA_CHOICES = ["amx", "vec"] if torch.cpu._is_amx_tile_supported() else ["vec"]
 
 
 @torch.inference_mode()

cmake/cpu_extension.cmake

@@ -102,11 +102,13 @@ if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64" OR ENABLE_X86_ISA)
         "-mavx512f"
         "-mavx512vl"
         "-mavx512bw"
-        "-mavx512dq"
-        "-mavx512bf16"
-        "-mavx512vnni"
+        "-mavx512dq")
+    list(APPEND CXX_COMPILE_FLAGS_AVX512_AMX 
+        ${CXX_COMPILE_FLAGS_AVX512}
         "-mamx-bf16"
-        "-mamx-tile")
+        "-mamx-tile"
+        "-mavx512bf16"
+        "-mavx512vnni")
     list(APPEND CXX_COMPILE_FLAGS_AVX2
         "-mavx2")
 elseif (POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
@@ -314,7 +316,8 @@ endif()
 
 # TODO: Refactor this
 if (ENABLE_X86_ISA)
-    message(STATUS "CPU extension (AVX512) compile flags: ${CXX_COMPILE_FLAGS_AVX512}")
+    message(STATUS "CPU extension (AVX512F + BF16 + VNNI + AMX) compile flags: ${CXX_COMPILE_FLAGS_AVX512_AMX}")
+    message(STATUS "CPU extension (AVX512F) compile flags: ${CXX_COMPILE_FLAGS_AVX512}")
     message(STATUS "CPU extension (AVX2) compile flags: ${CXX_COMPILE_FLAGS_AVX2}")
 else()
     message(STATUS "CPU extension compile flags: ${CXX_COMPILE_FLAGS}")
@@ -366,13 +369,15 @@ if(USE_ONEDNN)
 endif()
 
 if (ENABLE_X86_ISA)
-    set(VLLM_EXT_SRC_AVX512
+    set(VLLM_EXT_SRC_SGL
         "csrc/cpu/sgl-kernels/gemm.cpp"
         "csrc/cpu/sgl-kernels/gemm_int8.cpp"
         "csrc/cpu/sgl-kernels/gemm_fp8.cpp"
         "csrc/cpu/sgl-kernels/moe.cpp"
         "csrc/cpu/sgl-kernels/moe_int8.cpp"
-        "csrc/cpu/sgl-kernels/moe_fp8.cpp"
+        "csrc/cpu/sgl-kernels/moe_fp8.cpp")
+
+    set(VLLM_EXT_SRC_AVX512
         "csrc/cpu/shm.cpp"
         "csrc/cpu/cpu_wna16.cpp"
         "csrc/cpu/cpu_fused_moe.cpp"
@@ -398,31 +403,48 @@ if (ENABLE_X86_ISA)
         "csrc/cpu/pos_encoding.cpp"
         "csrc/moe/dynamic_4bit_int_moe_cpu.cpp") 
 
-    message(STATUS "CPU extension (AVX512) source files: ${VLLM_EXT_SRC_AVX512}")
+    message(STATUS "CPU extension (AVX512F + BF16 + VNNI + AMX) source files: ${VLLM_EXT_SRC_AVX512} ${VLLM_EXT_SRC_SGL}")
+    message(STATUS "CPU extension (AVX512F) source files: ${VLLM_EXT_SRC_AVX512}")
     message(STATUS "CPU extension (AVX2) source files: ${VLLM_EXT_SRC_AVX2}")
 
+    set(_C_LIBS numa dnnl_ext)
+    set(_C_AVX512_LIBS numa dnnl_ext)
+    set(_C_AVX2_LIBS numa)
+
+    # AMX + AVX512F + AVX512BF16 + AVX512VNNI
     define_extension_target(
         _C
         DESTINATION vllm
         LANGUAGE CXX
+        SOURCES ${VLLM_EXT_SRC_AVX512} ${VLLM_EXT_SRC_SGL}
+        LIBRARIES ${_C_LIBS}
+        COMPILE_FLAGS ${CXX_COMPILE_FLAGS_AVX512_AMX}
+        USE_SABI 3
+        WITH_SOABI
+    )
+
+    # For AMX kernels
+    target_compile_definitions(_C PRIVATE "-DCPU_CAPABILITY_AMXBF16")
+
+    # AVX512F 
+    define_extension_target(
+        _C_AVX512
+        DESTINATION vllm
+        LANGUAGE CXX
         SOURCES ${VLLM_EXT_SRC_AVX512}
-        LIBRARIES ${LIBS}
+        LIBRARIES ${_C_AVX512_LIBS}
         COMPILE_FLAGS ${CXX_COMPILE_FLAGS_AVX512}
         USE_SABI 3
         WITH_SOABI
     )
 
-    # For SGL kernels
-    target_compile_definitions(_C PRIVATE "-DCPU_CAPABILITY_AVX512")
-    # For AMX kernels
-    target_compile_definitions(_C PRIVATE "-DCPU_CAPABILITY_AMXBF16")
-
+    # AVX2 
     define_extension_target(
         _C_AVX2
         DESTINATION vllm
         LANGUAGE CXX
         SOURCES ${VLLM_EXT_SRC_AVX2}
-        LIBRARIES ${LIBS}
+        LIBRARIES ${_C_AVX2_LIBS}
         COMPILE_FLAGS ${CXX_COMPILE_FLAGS_AVX2}
         USE_SABI 3
         WITH_SOABI

cmake/external_projects/vllm_flash_attn.cmake

@@ -39,7 +39,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 140c00c0241bb60cc6e44e7c1be9998d4b20d8d2
+          GIT_TAG 29210221863736a08f71a866459e368ad1ac4a95
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

csrc/cpu/utils.cpp

@@ -173,10 +173,13 @@ ScratchPadManager::ScratchPadManager() : size_(0), ptr_(nullptr) {
 void ScratchPadManager::realloc(size_t new_size) {
   new_size = round(new_size);
   if (new_size > size_) {
+    void* new_ptr = std::aligned_alloc(64, new_size);
+    TORCH_CHECK(new_ptr != nullptr,
+                "ScratchPadManager: aligned_alloc failed for size ", new_size);
     if (ptr_ != nullptr) {
       std::free(ptr_);
     }
-    ptr_ = std::aligned_alloc(64, new_size);
+    ptr_ = new_ptr;
     size_ = new_size;
   }
 }

csrc/cuda_vec_utils.cuh

@@ -196,6 +196,7 @@ __forceinline__ __device__ u32x8_t ld256_cs(const u32x8_t* addr) {
   return val;
 #else
   assert(false && "ld256_cs requires SM100+ with CUDA 12.9+");
+  return u32x8_t{};
 #endif
 }
 

csrc/cumem_allocator.cpp

@@ -109,16 +109,18 @@ void create_and_map(unsigned long long device, ssize_t size, CUdeviceptr d_mem,
 
 #ifndef USE_ROCM
   int flag = 0;
-  CUDA_CHECK(cuDeviceGetAttribute(
+  CUresult rdma_result = cuDeviceGetAttribute(
       &flag, CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WITH_CUDA_VMM_SUPPORTED,
-      device));
-  if (flag) {  // support GPUDirect RDMA if possible
+      device);
+  if (rdma_result == CUDA_SUCCESS &&
+      flag) {  // support GPUDirect RDMA if possible
     prop.allocFlags.gpuDirectRDMACapable = 1;
   }
   int fab_flag = 0;
-  CUDA_CHECK(cuDeviceGetAttribute(
-      &fab_flag, CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_FABRIC_SUPPORTED, device));
-  if (fab_flag) {  // support fabric handle if possible
+  CUresult fab_result = cuDeviceGetAttribute(
+      &fab_flag, CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_FABRIC_SUPPORTED, device);
+  if (fab_result == CUDA_SUCCESS &&
+      fab_flag) {  // support fabric handle if possible
     prop.requestedHandleTypes = CU_MEM_HANDLE_TYPE_FABRIC;
   }
 #endif

csrc/libtorch_stable/ops.h

@@ -0,0 +1,9 @@
+#pragma once
+
+#include <torch/csrc/stable/library.h>
+#include <torch/csrc/stable/tensor.h>
+
+#ifndef USE_ROCM
+torch::stable::Tensor permute_cols(torch::stable::Tensor const& A,
+                                   torch::stable::Tensor const& perm);
+#endif

csrc/libtorch_stable/permute_cols.cu

@@ -1,10 +1,13 @@
-#include <torch/all.h>
-
-#include <ATen/cuda/CUDAContext.h>
-#include <c10/cuda/CUDAGuard.h>
+#include <torch/csrc/stable/library.h>
+#include <torch/csrc/stable/tensor.h>
+#include <torch/csrc/stable/accelerator.h>
+#include <torch/csrc/stable/ops.h>
+#include <torch/headeronly/core/ScalarType.h>
 
 #include <cuda_fp16.h>
 
+#include "torch_utils.h"
+
 static constexpr int default_threads = 256;
 static constexpr int div_ceil(int a, int b) { return (a + b - 1) / b; }
 
@@ -64,19 +67,22 @@ __global__ void permute_cols_kernel(int4 const* __restrict__ a_int4_ptr,
 
 // More efficient version of A[..., perm]
 //  taken from gptq_marlin.cu
-torch::Tensor permute_cols(torch::Tensor const& A, torch::Tensor const& perm) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(A));
-  auto dev = A.get_device();
-  auto stream = at::cuda::getCurrentCUDAStream(dev);
+torch::stable::Tensor permute_cols(torch::stable::Tensor const& A,
+                                   torch::stable::Tensor const& perm) {
+  const int32_t dev = A.get_device_index();
+  const torch::stable::accelerator::DeviceGuard device_guard(dev);
+  const auto stream = get_current_cuda_stream(dev);
 
-  TORCH_CHECK(A.scalar_type() == at::kHalf || A.scalar_type() == at::kBFloat16,
-              "Currently only 16bit types are supported");
-  TORCH_CHECK(A.is_contiguous(), "A must be contiguous");
-  TORCH_CHECK(A.size(-1) % 8 == 0,
-              "A columns must be a multiple of 8 (128bits)");
-  auto A_2d = A.view({-1, A.size(-1)});
+  STD_TORCH_CHECK(
+      A.scalar_type() == torch::headeronly::ScalarType::Half ||
+          A.scalar_type() == torch::headeronly::ScalarType::BFloat16,
+      "Currently only 16bit types are supported");
+  STD_TORCH_CHECK(A.is_contiguous(), "A must be contiguous");
+  STD_TORCH_CHECK(A.size(-1) % 8 == 0,
+                  "A columns must be a multiple of 8 (128bits)");
+  auto A_2d = torch::stable::view(A, {-1, A.size(-1)});
 
-  torch::Tensor D = torch::empty_like(A);
+  torch::stable::Tensor D = torch::stable::empty_like(A);
   int sms;
   cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, dev);
   int block_rows = div_ceil(A_2d.size(0), sms);

csrc/libtorch_stable/torch_bindings.cpp

@@ -0,0 +1,21 @@
+#include "ops.h"
+#include "core/registration.h"
+
+#include <torch/csrc/stable/library.h>
+
+// Register ops with STABLE_TORCH_LIBRARY for libtorch stable ABI compatibility.
+// Note: We register under namespace "_C" so ops are accessible as
+// torch.ops._C.<op_name> for compatibility with existing code.
+STABLE_TORCH_LIBRARY_FRAGMENT(_C, m) {
+#ifndef USE_ROCM
+  m.def("permute_cols(Tensor A, Tensor perm) -> Tensor");
+#endif
+}
+
+STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, m) {
+#ifndef USE_ROCM
+  m.impl("permute_cols", TORCH_BOX(&permute_cols));
+#endif
+}
+
+REGISTER_EXTENSION(_C_stable_libtorch)

csrc/libtorch_stable/torch_utils.h

@@ -0,0 +1,13 @@
+#pragma once
+
+#include <torch/csrc/inductor/aoti_torch/c/shim.h>
+#include <cuda_runtime.h>
+
+// Utility to get the current CUDA stream for a given device using stable APIs.
+// Returns a cudaStream_t for use in kernel launches.
+inline cudaStream_t get_current_cuda_stream(int32_t device_index) {
+  void* stream_ptr = nullptr;
+  TORCH_ERROR_CODE_CHECK(
+      aoti_torch_get_current_cuda_stream(device_index, &stream_ptr));
+  return reinterpret_cast<cudaStream_t>(stream_ptr);
+}

csrc/moe/gpt_oss_router_gemm.cu

@@ -0,0 +1,144 @@
+/*
+ * Adapted from
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/v1.3.0rc7/cpp/tensorrt_llm/kernels/tinygemm2/tinygemm2_cuda.cu
+ * Copyright (c) 2025, The vLLM team.
+ * SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION.
+ * All rights reserved. SPDX-License-Identifier: Apache-2.0
+ *
+ * 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 <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAStream.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <torch/all.h>
+#include "gpt_oss_router_gemm.cuh"
+
+void launch_gpt_oss_router_gemm(__nv_bfloat16* gA, __nv_bfloat16* gB,
+                                __nv_bfloat16* gC, __nv_bfloat16* bias,
+                                int batch_size, int output_features,
+                                int input_features, cudaStream_t stream) {
+  static int const WARP_TILE_M = 16;
+  static int const TILE_M = WARP_TILE_M;
+  static int const TILE_N = 8;
+  static int const TILE_K = 64;
+  static int const STAGES = 16;
+  static int const STAGE_UNROLL = 4;
+  static bool const PROFILE = false;
+
+  CUtensorMap weight_map{};
+  CUtensorMap activation_map{};
+
+  constexpr uint32_t rank = 2;
+  uint64_t size[rank] = {(uint64_t)input_features, (uint64_t)output_features};
+  uint64_t stride[rank - 1] = {input_features * sizeof(__nv_bfloat16)};
+  uint32_t box_size[rank] = {TILE_K, TILE_M};
+  uint32_t elem_stride[rank] = {1, 1};
+
+  CUresult res = cuTensorMapEncodeTiled(
+      &weight_map, CUtensorMapDataType::CU_TENSOR_MAP_DATA_TYPE_BFLOAT16, rank,
+      gB, size, stride, box_size, elem_stride,
+      CUtensorMapInterleave::CU_TENSOR_MAP_INTERLEAVE_NONE,
+      CUtensorMapSwizzle::CU_TENSOR_MAP_SWIZZLE_128B,
+      CUtensorMapL2promotion::CU_TENSOR_MAP_L2_PROMOTION_NONE,
+      CUtensorMapFloatOOBfill::CU_TENSOR_MAP_FLOAT_OOB_FILL_NONE);
+  TORCH_CHECK(res == CUDA_SUCCESS,
+              "cuTensorMapEncodeTiled failed for weight_map, error code=",
+              static_cast<int>(res));
+
+  size[1] = batch_size;
+  box_size[1] = TILE_N;
+
+  res = cuTensorMapEncodeTiled(
+      &activation_map, CUtensorMapDataType::CU_TENSOR_MAP_DATA_TYPE_BFLOAT16,
+      rank, gA, size, stride, box_size, elem_stride,
+      CUtensorMapInterleave::CU_TENSOR_MAP_INTERLEAVE_NONE,
+      CUtensorMapSwizzle::CU_TENSOR_MAP_SWIZZLE_128B,
+      CUtensorMapL2promotion::CU_TENSOR_MAP_L2_PROMOTION_NONE,
+      CUtensorMapFloatOOBfill::CU_TENSOR_MAP_FLOAT_OOB_FILL_NONE);
+  TORCH_CHECK(res == CUDA_SUCCESS,
+              "cuTensorMapEncodeTiled failed for activation_map, error code=",
+              static_cast<int>(res));
+
+  int smem_size = STAGES * STAGE_UNROLL *
+                  (TILE_M * TILE_K * sizeof(__nv_bfloat16) +
+                   TILE_N * TILE_K * sizeof(__nv_bfloat16));
+
+  gpuErrChk(cudaFuncSetAttribute(
+      gpt_oss_router_gemm_kernel<WARP_TILE_M, TILE_M, TILE_N, TILE_K, STAGES,
+                                 STAGE_UNROLL, PROFILE>,
+      cudaFuncAttributeMaxDynamicSharedMemorySize, smem_size));
+
+  int tiles_m = (output_features + TILE_M - 1) / TILE_M;
+  int tiles_n = (batch_size + TILE_N - 1) / TILE_N;
+
+  dim3 grid(tiles_m, tiles_n);
+  dim3 block(384);
+
+  cudaLaunchConfig_t config;
+  cudaLaunchAttribute attrs[1];
+  config.gridDim = grid;
+  config.blockDim = block;
+  config.dynamicSmemBytes = smem_size;
+  config.stream = stream;
+  config.attrs = attrs;
+  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+  attrs[0].val.programmaticStreamSerializationAllowed = 1;
+  config.numAttrs = 1;
+
+  cudaLaunchKernelEx(
+      &config,
+      &gpt_oss_router_gemm_kernel<WARP_TILE_M, TILE_M, TILE_N, TILE_K, STAGES,
+                                  STAGE_UNROLL, PROFILE>,
+      gC, gA, gB, bias, output_features, batch_size, input_features, weight_map,
+      activation_map, nullptr);
+}
+
+void gpt_oss_router_gemm_cuda_forward(torch::Tensor& output,
+                                      torch::Tensor input, torch::Tensor weight,
+                                      torch::Tensor bias) {
+  auto const batch_size = input.size(0);
+  auto const input_dim = input.size(1);
+  auto const output_dim = weight.size(0);
+
+  auto stream = at::cuda::getCurrentCUDAStream();
+
+  if (input.scalar_type() == at::ScalarType::BFloat16) {
+    launch_gpt_oss_router_gemm((__nv_bfloat16*)input.data_ptr(),
+                               (__nv_bfloat16*)weight.data_ptr(),
+                               (__nv_bfloat16*)output.mutable_data_ptr(),
+                               (__nv_bfloat16*)bias.data_ptr(), batch_size,
+                               output_dim, input_dim, stream);
+  } else {
+    throw std::invalid_argument("Unsupported dtype, only supports bfloat16");
+  }
+}
+
+void gpt_oss_router_gemm(torch::Tensor& output, torch::Tensor input,
+                         torch::Tensor weight, torch::Tensor bias) {
+  TORCH_CHECK(input.dim() == 2, "input must be 2D");
+  TORCH_CHECK(weight.dim() == 2, "weight must be 2D");
+  TORCH_CHECK(bias.dim() == 1, "bias must be 1D");
+  TORCH_CHECK(input.sizes()[1] == weight.sizes()[1],
+              "input.size(1) must match weight.size(1)");
+  TORCH_CHECK(weight.sizes()[0] == bias.sizes()[0],
+              "weight.size(0) must match bias.size(0)");
+  TORCH_CHECK(input.scalar_type() == at::ScalarType::BFloat16,
+              "input tensor must be bfloat16");
+  TORCH_CHECK(weight.scalar_type() == at::ScalarType::BFloat16,
+              "weight tensor must be bfloat16");
+  TORCH_CHECK(bias.scalar_type() == at::ScalarType::BFloat16,
+              "bias tensor must be bfloat16");
+  gpt_oss_router_gemm_cuda_forward(output, input, weight, bias);
+}

csrc/moe/gpt_oss_router_gemm.cuh

@@ -0,0 +1,447 @@
+/*
+ * Adapted from
+ * https://github.com/NVIDIA/TensorRT-LLM/blob/v1.3.0rc7/cpp/tensorrt_llm/kernels/tinygemm2/tinygemm2_kernel.cuh
+ * Copyright (c) 2025, The vLLM team.
+ * SPDX-FileCopyrightText: Copyright (c) 2025, NVIDIA CORPORATION.
+ * All rights reserved. SPDX-License-Identifier: Apache-2.0
+ *
+ * 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 "cuda_bf16.h"
+#include <stdint.h>
+#include <stdio.h>
+#include <vector>
+
+#include "cuda_pipeline.h"
+#include <cuda.h>
+#include <cuda/barrier>
+#include <cuda/std/utility>
+#include <cuda_runtime.h>
+
+using barrier = cuda::barrier<cuda::thread_scope_block>;
+namespace cde = cuda::device::experimental;
+namespace ptx = cuda::ptx;
+
+#define gpuErrChk(ans)                    \
+  {                                       \
+    gpuAssert((ans), __FILE__, __LINE__); \
+  }
+
+inline void gpuAssert(cudaError_t code, char const* file, int line,
+                      bool abort = true) {
+  if (code != cudaSuccess) {
+    fprintf(stderr, "GPUassert: %s %s %d\n", cudaGetErrorString(code), file,
+            line);
+    if (abort) {
+      throw std::runtime_error(cudaGetErrorString(code));
+    }
+  }
+}
+
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
+__device__ uint64_t gclock64() {
+  unsigned long long int rv;
+  asm volatile("mov.u64 %0, %%globaltimer;" : "=l"(rv));
+  return rv;
+}
+
+__device__ void ldmatrix(__nv_bfloat16 rv[2], uint32_t smem_ptr) {
+  int dst;
+  asm volatile("ldmatrix.sync.aligned.x1.m8n8.shared.b16 {%0}, [%1];\n"
+               : "=r"(dst)
+               : "r"(smem_ptr));
+  int* rvi = reinterpret_cast<int*>(&rv[0]);
+  rvi[0] = dst;
+}
+
+__device__ void ldmatrix2(__nv_bfloat16 rv[4], uint32_t smem_ptr) {
+  int x, y;
+  asm volatile("ldmatrix.sync.aligned.x2.m8n8.shared.b16 {%0, %1}, [%2];\n"
+               : "=r"(x), "=r"(y)
+               : "r"(smem_ptr));
+
+  int* rvi = reinterpret_cast<int*>(&rv[0]);
+  rvi[0] = x;
+  rvi[1] = y;
+}
+
+__device__ void ldmatrix4(__nv_bfloat16 rv[8], uint32_t smem_ptr) {
+  int x, y, z, w;
+  asm volatile(
+      "ldmatrix.sync.aligned.x4.m8n8.shared.b16 {%0, %1, %2, %3}, [%4];"
+      : "=r"(x), "=r"(y), "=r"(z), "=r"(w)
+      : "r"(smem_ptr));
+  int* rvi = reinterpret_cast<int*>(&rv[0]);
+  rvi[0] = x;
+  rvi[1] = y;
+  rvi[2] = z;
+  rvi[3] = w;
+}
+
+__device__ void HMMA_1688(float d[4], __nv_bfloat16 a[4], __nv_bfloat16 b[2],
+                          float c[4]) {
+  uint32_t const* A = reinterpret_cast<uint32_t const*>(&a[0]);
+  uint32_t const* B = reinterpret_cast<uint32_t const*>(&b[0]);
+  float const* C = reinterpret_cast<float const*>(&c[0]);
+  float* D = reinterpret_cast<float*>(&d[0]);
+
+  asm volatile(
+      "mma.sync.aligned.m16n8k8.row.col.f32.bf16.bf16.f32 "
+      "{%0,%1,%2,%3}, {%4,%5}, {%6}, {%7,%8,%9,%10};\n"
+      : "=f"(D[0]), "=f"(D[1]), "=f"(D[2]), "=f"(D[3])
+      : "r"(A[0]), "r"(A[1]), "r"(B[0]), "f"(C[0]), "f"(C[1]), "f"(C[2]),
+        "f"(C[3]));
+}
+
+__device__ void HMMA_16816(float d[4], __nv_bfloat16 a[8], __nv_bfloat16 b[4],
+                           float c[4]) {
+  uint32_t const* A = reinterpret_cast<uint32_t const*>(&a[0]);
+  uint32_t const* B = reinterpret_cast<uint32_t const*>(&b[0]);
+  float const* C = reinterpret_cast<float const*>(&c[0]);
+  float* D = reinterpret_cast<float*>(&d[0]);
+
+  asm volatile(
+      "mma.sync.aligned.m16n8k16.row.col.f32.bf16.bf16.f32 "
+      "{%0,%1,%2,%3}, {%4,%5,%6,%7}, {%8,%9}, {%10,%11,%12,%13};\n"
+      : "=f"(D[0]), "=f"(D[1]), "=f"(D[2]), "=f"(D[3])
+      : "r"(A[0]), "r"(A[1]), "r"(A[2]), "r"(A[3]), "r"(B[0]), "r"(B[1]),
+        "f"(C[0]), "f"(C[1]), "f"(C[2]), "f"(C[3]));
+}
+
+__device__ void bar_wait(uint32_t bar_ptr, int phase) {
+  asm volatile(
+      "{\n"
+      ".reg .pred                P1;\n"
+      "LAB_WAIT:\n"
+      "mbarrier.try_wait.parity.shared::cta.b64 P1, [%0], %1;\n"
+      "@P1                       bra.uni DONE;\n"
+      "bra.uni                   LAB_WAIT;\n"
+      "DONE:\n"
+      "}\n" ::"r"(bar_ptr),
+      "r"(phase));
+}
+
+__device__ bool bar_try_wait(uint32_t bar_ptr, int phase) {
+  uint32_t success;
+  #ifdef INTERNAL
+  asm volatile(".pragma \"set knob DontInsertYield\";\n" : : : "memory");
+  #endif
+  asm volatile(
+      "{\n\t"
+      ".reg .pred P1; \n\t"
+      "mbarrier.try_wait.parity.shared::cta.b64 P1, [%1], %2; \n\t"
+      "selp.b32 %0, 1, 0, P1; \n\t"
+      "}"
+      : "=r"(success)
+      : "r"(bar_ptr), "r"(phase));
+  return success;
+}
+
+__device__ uint32_t elect_one_sync() {
+  uint32_t pred = 0;
+  uint32_t laneid = 0;
+  asm volatile(
+      "{\n"
+      ".reg .b32 %%rx;\n"
+      ".reg .pred %%px;\n"
+      "     elect.sync %%rx|%%px, %2;\n"
+      "@%%px mov.s32 %1, 1;\n"
+      "     mov.s32 %0, %%rx;\n"
+      "}\n"
+      : "+r"(laneid), "+r"(pred)
+      : "r"(0xFFFFFFFF));
+  return pred;
+}
+#endif
+
+struct Profile {
+  uint64_t start;
+  uint64_t weight_load_start;
+  uint64_t act_load_start;
+  uint64_t compute_start;
+  uint64_t complete;
+};
+
+template <int WARP_TILE_M, int TILE_M, int TILE_N, int TILE_K, int STAGES,
+          int STAGE_UNROLL, bool PROFILE>
+__global__ __launch_bounds__(384, 1) void gpt_oss_router_gemm_kernel(
+    __nv_bfloat16* output, __nv_bfloat16* weights, __nv_bfloat16* activations,
+    __nv_bfloat16* bias, int M, int N, int K,
+    const __grid_constant__ CUtensorMap weight_map,
+    const __grid_constant__ CUtensorMap activation_map,
+    Profile* profile = nullptr) {
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
+
+  if (PROFILE && threadIdx.x == 0 && blockIdx.y == 0)
+    profile[blockIdx.x].start = gclock64();
+
+  extern __shared__ __align__(128) char smem[];
+
+  __nv_bfloat16* sh_weights = (__nv_bfloat16*)&smem[0];
+  __nv_bfloat16* sh_activations =
+      (__nv_bfloat16*)&smem[STAGES * STAGE_UNROLL * TILE_M * TILE_K *
+                            sizeof(__nv_bfloat16)];
+
+  #pragma nv_diag_suppress static_var_with_dynamic_init
+  __shared__ barrier bar_wt_ready[STAGES];
+  __shared__ barrier bar_act_ready[STAGES];
+  __shared__ barrier bar_data_consumed[STAGES];
+
+  __shared__ float4 reduction_buffer[128];
+
+  __shared__ nv_bfloat16 sh_bias[TILE_M];
+
+  if (threadIdx.x == 0) {
+    for (int i = 0; i < STAGES; i++) {
+      init(&bar_wt_ready[i], 1);
+      init(&bar_act_ready[i], 1);
+      init(&bar_data_consumed[i], 32);
+    }
+    ptx::fence_proxy_async(ptx::space_shared);
+    asm volatile("prefetch.tensormap [%0];"
+                 :
+                 : "l"(reinterpret_cast<uint64_t>(&weight_map))
+                 : "memory");
+    asm volatile("prefetch.tensormap [%0];"
+                 :
+                 : "l"(reinterpret_cast<uint64_t>(&activation_map))
+                 : "memory");
+  }
+  __syncthreads();
+
+  int warp_id = threadIdx.x / 32;
+  int lane_id = threadIdx.x % 32;
+
+  int phase = 0;
+
+  int mib = blockIdx.x * TILE_M;
+  int ni = blockIdx.y * TILE_N;
+
+  float accum[4];
+  for (int i = 0; i < 4; i++) accum[i] = 0.f;
+
+  int const K_LOOPS_DMA =
+      (K + 4 * TILE_K * STAGE_UNROLL - 1) / (4 * (TILE_K * STAGE_UNROLL));
+  int const K_LOOPS_COMPUTE = K_LOOPS_DMA;
+
+  // Data loading thread
+  if (warp_id >= 4 && elect_one_sync()) {
+    int stage = warp_id % 4;
+
+    bool weight_warp = warp_id < 8;
+    if (!weight_warp) {
+      cudaGridDependencySynchronize();
+      cudaTriggerProgrammaticLaunchCompletion();
+    }
+
+    for (int ki = 0; ki < K_LOOPS_DMA; ki++) {
+      int k = (ki * 4 + (warp_id % 4)) * TILE_K * STAGE_UNROLL;
+
+      uint64_t desc_ptr_wt = reinterpret_cast<uint64_t>(&weight_map);
+      uint64_t desc_ptr_act = reinterpret_cast<uint64_t>(&activation_map);
+
+      uint32_t bar_ptr_wt = __cvta_generic_to_shared(&bar_wt_ready[stage]);
+      uint32_t bar_ptr_act = __cvta_generic_to_shared(&bar_act_ready[stage]);
+      int bytes_wt = TILE_M * TILE_K * sizeof(__nv_bfloat16);
+      int bytes_act = TILE_N * TILE_K * sizeof(__nv_bfloat16);
+
+      bar_wait(__cvta_generic_to_shared(&bar_data_consumed[stage]), phase ^ 1);
+
+      if (weight_warp)
+        asm volatile("mbarrier.arrive.expect_tx.shared.b64 _, [%0], %1;"
+                     :
+                     : "r"(bar_ptr_wt), "r"(STAGE_UNROLL * bytes_wt));
+      if (!weight_warp)
+        asm volatile("mbarrier.arrive.expect_tx.shared.b64 _, [%0], %1;"
+                     :
+                     : "r"(bar_ptr_act), "r"(STAGE_UNROLL * bytes_act));
+
+      if (PROFILE && blockIdx.y == 0 && ki == 0 && weight_warp)
+        profile[blockIdx.x].weight_load_start = gclock64();
+      if (PROFILE && blockIdx.y == 0 && ki == 0 && !weight_warp)
+        profile[blockIdx.x].act_load_start = gclock64();
+
+      for (int i = 0; i < STAGE_UNROLL; i++) {
+        uint32_t smem_ptr_wt = __cvta_generic_to_shared(
+            &sh_weights[(stage * STAGE_UNROLL + i) * TILE_M * TILE_K]);
+        uint32_t crd0 = k + i * TILE_K;
+        uint32_t crd1 = mib;
+        if (weight_warp)
+          asm volatile(
+              "cp.async.bulk.tensor.2d.shared::cta.global.mbarrier::complete_"
+              "tx::bytes [%0], [%1, {%3,%4}], "
+              "[%2];"
+              :
+              : "r"(smem_ptr_wt), "l"(desc_ptr_wt), "r"(bar_ptr_wt), "r"(crd0),
+                "r"(crd1)
+              : "memory");
+
+        uint32_t smem_ptr_act = __cvta_generic_to_shared(
+            &sh_activations[(stage * STAGE_UNROLL + i) * TILE_N * TILE_K]);
+        crd0 = k + i * TILE_K;
+        crd1 = ni;
+        if (!weight_warp)
+          asm volatile(
+              "cp.async.bulk.tensor.2d.shared::cta.global.mbarrier::complete_"
+              "tx::bytes [%0], [%1, {%3,%4}], "
+              "[%2];"
+              :
+              : "r"(smem_ptr_act), "l"(desc_ptr_act), "r"(bar_ptr_act),
+                "r"(crd0), "r"(crd1)
+              : "memory");
+      }
+
+      stage += 4;
+      if (stage >= STAGES) {
+        stage = warp_id % 4;
+        phase ^= 1;
+      }
+    }
+    // Wait for pending loads to be consumed before exiting, to avoid race
+    for (int i = 0; i < (STAGES / 4) - 1; i++) {
+      bar_wait(__cvta_generic_to_shared(&bar_data_consumed[stage]), phase ^ 1);
+      stage += 4;
+      if (stage >= STAGES) {
+        stage = warp_id % 4;
+        phase ^= 1;
+      }
+    }
+  }
+  // Compute threads
+  else if (warp_id < 4) {
+    // Sneak the bias load into the compute warps since they're just waiting for
+    // stuff anyway
+    if (threadIdx.x < TILE_M) sh_bias[threadIdx.x] = bias[mib + threadIdx.x];
+
+    int stage = warp_id;
+
+    int phase = 0;
+    int lane_id_div8 = lane_id / 8;
+    int lane_id_mod8 = lane_id % 8;
+
+    int lane_row_offset_wt = (lane_id_div8 % 2) ? 8 : 0;
+    int lane_col_offset_wt = (lane_id_div8 / 2) ? 1 : 0;
+
+    int row_wt = lane_id_mod8 + lane_row_offset_wt;
+    int row_act = lane_id_mod8;
+
+    int row_offset_wt = (reinterpret_cast<uintptr_t>(sh_weights) / 128) % 8;
+    int row_offset_act = row_offset_wt;
+
+    uint32_t bar_ptr_wt = __cvta_generic_to_shared(&bar_wt_ready[stage]);
+    uint32_t bar_ptr_act = __cvta_generic_to_shared(&bar_act_ready[stage]);
+
+    bool weight_ready = bar_try_wait(bar_ptr_wt, phase);
+    bool act_ready = bar_try_wait(bar_ptr_act, phase);
+
+  #pragma unroll 2
+    for (int ki = 0; ki < K_LOOPS_COMPUTE; ki++) {
+      int next_stage = stage + 4;
+      int next_phase = phase;
+      if (next_stage >= STAGES) {
+        next_stage = warp_id;
+        next_phase ^= 1;
+      }
+
+      while (!weight_ready || !act_ready) {
+        weight_ready = bar_try_wait(bar_ptr_wt, phase);
+        act_ready = bar_try_wait(bar_ptr_act, phase);
+      }
+
+      if (PROFILE && blockIdx.y == 0 && threadIdx.x == 0 && ki == 0)
+        profile[blockIdx.x].compute_start = gclock64();
+
+      if (ki + 1 < K_LOOPS_COMPUTE) {
+        weight_ready = bar_try_wait(
+            __cvta_generic_to_shared(&bar_wt_ready[next_stage]), next_phase);
+        act_ready = bar_try_wait(
+            __cvta_generic_to_shared(&bar_act_ready[next_stage]), next_phase);
+      }
+
+  #pragma unroll
+      for (int su = 0; su < STAGE_UNROLL; su++) {
+        __nv_bfloat16* ptr_weights =
+            &sh_weights[(stage * STAGE_UNROLL + su) * TILE_M * TILE_K];
+        __nv_bfloat16* ptr_act =
+            &sh_activations[(stage * STAGE_UNROLL + su) * TILE_N * TILE_K];
+
+  #pragma unroll
+        for (int kii = 0; kii < TILE_K / 16; kii++) {
+          __nv_bfloat16 a[8];
+          __nv_bfloat16 b[4];
+
+          int col = 2 * kii + lane_col_offset_wt;
+          int col_sw = ((row_wt + row_offset_wt) % 8) ^ col;
+
+          ldmatrix4(a, __cvta_generic_to_shared(
+                           &ptr_weights[row_wt * TILE_K + col_sw * 8]));
+
+          col = 2 * kii + lane_id_div8;
+          col_sw = ((row_act + row_offset_act) % 8) ^ col;
+
+          ldmatrix2(b, __cvta_generic_to_shared(
+                           &ptr_act[row_act * TILE_K + 8 * col_sw]));
+
+          HMMA_16816(accum, a, b, accum);
+        }
+      }
+
+      uint32_t bar_c = __cvta_generic_to_shared(&bar_data_consumed[stage]);
+      asm volatile("mbarrier.arrive.shared::cta.b64 _, [%0];" : : "r"(bar_c));
+
+      stage = next_stage;
+      phase = next_phase;
+    }
+
+    float4 accum4;
+    accum4.x = accum[0];
+    accum4.y = accum[1];
+    accum4.z = accum[2];
+    accum4.w = accum[3];
+    reduction_buffer[threadIdx.x] = accum4;
+
+    __syncthreads();
+
+    if (warp_id == 0) {
+      int mi = mib + warp_id * WARP_TILE_M;
+      int tm = mi + lane_id / 4;
+      int tn = ni + 2 * (lane_id % 4);
+
+      float4 accum1 = reduction_buffer[32 + threadIdx.x];
+      float4 accum2 = reduction_buffer[64 + threadIdx.x];
+      float4 accum3 = reduction_buffer[96 + threadIdx.x];
+
+      accum[0] = accum[0] + accum1.x + accum2.x + accum3.x;
+      accum[1] = accum[1] + accum1.y + accum2.y + accum3.y;
+      accum[2] = accum[2] + accum1.z + accum2.z + accum3.z;
+      accum[3] = accum[3] + accum1.w + accum2.w + accum3.w;
+
+      float bias_lo = __bfloat162float(sh_bias[tm - mib]);
+      float bias_hi = __bfloat162float(sh_bias[tm + 8 - mib]);
+
+      if (tn < N && tm < M)
+        output[tn * M + tm] = __float2bfloat16(accum[0] + bias_lo);
+      if (tn + 1 < N && tm < M)
+        output[(tn + 1) * M + tm] = __float2bfloat16(accum[1] + bias_lo);
+      if (tn < N && tm + 8 < M)
+        output[tn * M + tm + 8] = __float2bfloat16(accum[2] + bias_hi);
+      if (tn + 1 < N && tm + 8 < M)
+        output[(tn + 1) * M + tm + 8] = __float2bfloat16(accum[3] + bias_hi);
+
+      if (PROFILE && blockIdx.y == 0 && threadIdx.x == 0)
+        profile[blockIdx.x].complete = gclock64();
+    }
+  }
+#endif  // end if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
+}

csrc/moe/moe_ops.h

@@ -70,4 +70,8 @@ torch::Tensor router_gemm_bf16_fp32(torch::Tensor const& input,
 // Supports num_tokens in [1, 16], num_experts in {256, 384}, hidden_dim = 7168
 void dsv3_router_gemm(torch::Tensor& output, const torch::Tensor& mat_a,
                       const torch::Tensor& mat_b);
+
+// gpt-oss optimized router GEMM kernel for SM90+
+void gpt_oss_router_gemm(torch::Tensor& output, torch::Tensor input,
+                         torch::Tensor weight, torch::Tensor bias);
 #endif

csrc/moe/moe_permute_unpermute_op.cu

@@ -73,10 +73,9 @@ void moe_permute(
   MOE_DISPATCH(input.scalar_type(), [&] {
     expandInputRowsKernelLauncher<scalar_t>(
         get_ptr<scalar_t>(input), get_ptr<scalar_t>(permuted_input),
-        get_ptr<int>(permuted_experts_id), get_ptr<int>(sorted_row_idx),
-        get_ptr<int>(inv_permuted_idx), get_ptr<int>(permuted_idx),
-        get_ptr<int64_t>(expert_first_token_offset), n_token, valid_num_ptr,
-        n_hidden, topk, n_local_expert, stream);
+        get_ptr<int>(sorted_row_idx), get_ptr<int>(inv_permuted_idx),
+        get_ptr<int>(permuted_idx), get_ptr<int64_t>(expert_first_token_offset),
+        n_token, valid_num_ptr, n_hidden, topk, n_local_expert, stream);
   });
 }
 

csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.h

@@ -57,7 +57,7 @@ void sortAndScanExpert(const int* expert_for_source_row, const int* source_rows,
 
 template <typename T>
 void expandInputRowsKernelLauncher(
-    T const* unpermuted_input, T* permuted_output, int* sorted_experts,
+    T const* unpermuted_input, T* permuted_output,
     int const* expanded_dest_row_to_expanded_source_row,
     int* expanded_source_row_to_expanded_dest_row, int* permuted_idx,
     int64_t const* expert_first_token_offset, int64_t const num_rows,

csrc/moe/permute_unpermute_kernels/moe_permute_unpermute_kernel.inl

@@ -2,7 +2,7 @@
 
 template <typename T, bool CHECK_SKIPPED>
 __global__ void expandInputRowsKernel(
-    T const* unpermuted_input, T* permuted_output, int* sorted_experts,
+    T const* unpermuted_input, T* permuted_output,
     int const* expanded_dest_row_to_expanded_source_row,
     int* expanded_source_row_to_expanded_dest_row, int* permuted_idx,
     int64_t const* expert_first_token_offset, int64_t const num_rows,
@@ -16,7 +16,6 @@ __global__ void expandInputRowsKernel(
   int64_t expanded_dest_row = blockIdx.x;
   int64_t const expanded_source_row =
       expanded_dest_row_to_expanded_source_row[expanded_dest_row];
-  int expert_id = sorted_experts[expanded_dest_row];
 
   if (threadIdx.x == 0) {
     assert(expanded_dest_row <= INT32_MAX);
@@ -54,7 +53,7 @@ __global__ void expandInputRowsKernel(
 
 template <typename T>
 void expandInputRowsKernelLauncher(
-    T const* unpermuted_input, T* permuted_output, int* sorted_experts,
+    T const* unpermuted_input, T* permuted_output,
     int const* expanded_dest_row_to_expanded_source_row,
     int* expanded_source_row_to_expanded_dest_row, int* permuted_idx,
     int64_t const* expert_first_token_offset, int64_t const num_rows,
@@ -70,12 +69,12 @@ void expandInputRowsKernelLauncher(
   bool is_check_skip = num_valid_tokens_ptr != nullptr;
   auto func = func_map[is_check_skip];
 
-  func<<<blocks, threads, 0, stream>>>(
-      unpermuted_input, permuted_output, sorted_experts,
-      expanded_dest_row_to_expanded_source_row,
-      expanded_source_row_to_expanded_dest_row, permuted_idx,
-      expert_first_token_offset, num_rows, num_valid_tokens_ptr, cols, k,
-      num_local_experts);
+  func<<<blocks, threads, 0, stream>>>(unpermuted_input, permuted_output,
+                                       expanded_dest_row_to_expanded_source_row,
+                                       expanded_source_row_to_expanded_dest_row,
+                                       permuted_idx, expert_first_token_offset,
+                                       num_rows, num_valid_tokens_ptr, cols, k,
+                                       num_local_experts);
 }
 
 template <class T, class U>

csrc/moe/torch_bindings.cpp

@@ -132,6 +132,12 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
   // DeepSeek V3 optimized router GEMM for SM90+
   m.def("dsv3_router_gemm(Tensor! output, Tensor mat_a, Tensor mat_b) -> ()");
   // conditionally compiled so impl registration is in source file
+
+  // gpt-oss optimized router GEMM kernel for SM90+
+  m.def(
+      "gpt_oss_router_gemm(Tensor! output, Tensor input, Tensor weights, "
+      "Tensor bias) -> ()");
+  m.impl("gpt_oss_router_gemm", torch::kCUDA, &gpt_oss_router_gemm);
 #endif
 }
 

csrc/ops.h

@@ -201,7 +201,6 @@ torch::Tensor awq_dequantize(torch::Tensor _kernel,
                              torch::Tensor _zeros, int64_t split_k_iters,
                              int64_t thx, int64_t thy);
 
-torch::Tensor permute_cols(torch::Tensor const& A, torch::Tensor const& perm);
 #endif
 
 torch::Tensor ggml_dequantize(torch::Tensor W, int64_t type, int64_t m,
@@ -262,7 +261,8 @@ void get_cutlass_moe_mm_data(
     torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
     torch::Tensor& input_permutation, torch::Tensor& output_permutation,
     const int64_t num_experts, const int64_t n, const int64_t k,
-    const std::optional<torch::Tensor>& blockscale_offsets);
+    const std::optional<torch::Tensor>& blockscale_offsets,
+    const bool is_gated);
 
 void get_cutlass_moe_mm_problem_sizes_from_expert_offsets(
     const torch::Tensor& expert_first_token_offset,
@@ -295,10 +295,14 @@ void cutlass_scaled_sparse_mm(torch::Tensor& out, torch::Tensor const& a,
 
 std::vector<torch::Tensor> cutlass_sparse_compress(torch::Tensor const& a);
 
-void scaled_fp4_quant(torch::Tensor& output, torch::Tensor const& input,
-                      torch::Tensor& output_scale,
-                      torch::Tensor const& input_scale,
-                      bool is_sf_swizzled_layout);
+std::tuple<torch::Tensor, torch::Tensor> scaled_fp4_quant_func(
+    torch::Tensor const& input, torch::Tensor const& input_scale,
+    bool is_sf_swizzled_layout);
+
+void scaled_fp4_quant_out(torch::Tensor const& input,
+                          torch::Tensor const& input_scale,
+                          bool is_sf_swizzled_layout, torch::Tensor& output,
+                          torch::Tensor& output_scale);
 
 void scaled_fp4_experts_quant(
     torch::Tensor& output, torch::Tensor& output_scale,

csrc/quantization/fp4/nvfp4_quant_entry.cu

@@ -16,6 +16,8 @@
 
 #include <torch/all.h>
 
+#include "nvfp4_utils.cuh"
+
 #if (defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100) || \
     (defined(ENABLE_NVFP4_SM120) && ENABLE_NVFP4_SM120)
 void scaled_fp4_quant_sm1xxa(torch::Tensor const& output,
@@ -51,9 +53,10 @@ void silu_and_mul_scaled_fp4_experts_quant_sm1xxa(
     torch::Tensor const& output_scale_offset_by_experts);
 #endif
 
-void scaled_fp4_quant(torch::Tensor& output, torch::Tensor const& input,
-                      torch::Tensor& output_sf, torch::Tensor const& input_sf,
-                      bool is_sf_swizzled_layout) {
+void scaled_fp4_quant_out(torch::Tensor const& input,
+                          torch::Tensor const& input_sf,
+                          bool is_sf_swizzled_layout, torch::Tensor& output,
+                          torch::Tensor& output_sf) {
 #if (defined(ENABLE_NVFP4_SM100) && ENABLE_NVFP4_SM100) || \
     (defined(ENABLE_NVFP4_SM120) && ENABLE_NVFP4_SM120)
   return scaled_fp4_quant_sm1xxa(output, input, output_sf, input_sf,
@@ -62,6 +65,34 @@ void scaled_fp4_quant(torch::Tensor& output, torch::Tensor const& input,
   TORCH_CHECK_NOT_IMPLEMENTED(false, "No compiled nvfp4 quantization kernel");
 }
 
+std::tuple<torch::Tensor, torch::Tensor> scaled_fp4_quant_func(
+    torch::Tensor const& input, torch::Tensor const& input_sf,
+    bool is_sf_swizzled_layout) {
+  int64_t n = input.size(-1);
+  int64_t m = input.numel() / n;
+  auto device = input.device();
+
+  // Two fp4 values packed into a uint8
+  auto output = torch::empty(
+      {m, n / 2}, torch::TensorOptions().device(device).dtype(torch::kUInt8));
+
+  torch::Tensor output_sf;
+  if (is_sf_swizzled_layout) {
+    auto [sf_m, sf_n] = vllm::computeSwizzledSFShape(m, n);
+    output_sf = torch::empty(
+        {sf_m, sf_n},
+        torch::TensorOptions().device(device).dtype(torch::kInt32));
+  } else {
+    output_sf = torch::empty(
+        {m, n / CVT_FP4_SF_VEC_SIZE},
+        torch::TensorOptions().device(device).dtype(torch::kUInt8));
+  }
+
+  scaled_fp4_quant_out(input, input_sf, is_sf_swizzled_layout, output,
+                       output_sf);
+  return {output, output_sf};
+}
+
 void scaled_fp4_experts_quant(
     torch::Tensor& output, torch::Tensor& output_scale,
     torch::Tensor const& input, torch::Tensor const& input_global_scale,

csrc/quantization/fp4/nvfp4_utils.cuh

@@ -18,6 +18,7 @@
 
 #include <cuda_runtime.h>
 #include <cuda_fp8.h>
+#include <utility>
 
 #include "../../cuda_vec_utils.cuh"
 
@@ -54,6 +55,18 @@ inline int computeEffectiveRows(int m) {
   return round_up(m, ROW_TILE);
 }
 
+// Compute the shape of the swizzled SF output tensor.
+// Returns (rounded_m, rounded_n / 4) where:
+//   rounded_m = round_up(m, 128)
+//   rounded_n = round_up(n / CVT_FP4_SF_VEC_SIZE, 4)
+inline std::pair<int64_t, int64_t> computeSwizzledSFShape(int64_t m,
+                                                          int64_t n) {
+  int64_t rounded_m = round_up(m, static_cast<int64_t>(128));
+  int64_t scale_n = n / CVT_FP4_SF_VEC_SIZE;
+  int64_t rounded_n = round_up(scale_n, static_cast<int64_t>(4));
+  return {rounded_m, rounded_n / 4};
+}
+
 // Convert 8 float32 values into 8 e2m1 values (represented as one uint32_t).
 inline __device__ uint32_t fp32_vec8_to_e2m1(float (&array)[8]) {
   uint32_t val;

csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu

@@ -286,6 +286,15 @@ void rms_norm_per_block_quant(torch::Tensor& out, torch::Tensor const& input,
                 "Outer scale stride must be 1 when scales are not transposed");
   }
 
+  int64_t hidden_size = input.size(-1);
+  TORCH_CHECK(hidden_size > 0 && hidden_size % group_size == 0,
+              "hidden_size must be a positive multiple of group_size");
+  int64_t num_tokens = input.numel() / hidden_size;
+  int64_t num_groups = hidden_size / group_size;
+  TORCH_CHECK(scales.numel() >= num_tokens * num_groups,
+              "scales buffer too small: need ", num_tokens * num_groups,
+              " elements, got ", scales.numel());
+
   rms_norm_per_block_quant_dispatch(out, input, weight, scales, group_size,
                                     var_epsilon, scale_ub, residual,
                                     is_scale_transposed);

csrc/quantization/w8a8/cutlass/moe/moe_data.cu

@@ -17,8 +17,11 @@ __global__ void compute_problem_sizes(const int32_t* __restrict__ topk_ids,
                                       int32_t* problem_sizes2,
                                       int32_t* atomic_buffer,
                                       const int topk_length, const int n,
-                                      const int k) {
+                                      const int k, const bool is_gated) {
   int expert_id = blockIdx.x;
+  // For gated activations (gate + up), first GEMM output is 2*n.
+  // For non-gated activations (up only), first GEMM output is n.
+  int const n1 = is_gated ? 2 * n : n;
 
   int occurrences = 0;
   for (int i = threadIdx.x; i < topk_length; i += THREADS_PER_EXPERT) {
@@ -31,13 +34,13 @@ __global__ void compute_problem_sizes(const int32_t* __restrict__ topk_ids,
     int final_occurrences = atomic_buffer[expert_id];
     if constexpr (!SWAP_AB) {
       problem_sizes1[expert_id * 3] = final_occurrences;
-      problem_sizes1[expert_id * 3 + 1] = 2 * n;
+      problem_sizes1[expert_id * 3 + 1] = n1;
       problem_sizes1[expert_id * 3 + 2] = k;
       problem_sizes2[expert_id * 3] = final_occurrences;
       problem_sizes2[expert_id * 3 + 1] = k;
       problem_sizes2[expert_id * 3 + 2] = n;
     } else {
-      problem_sizes1[expert_id * 3] = 2 * n;
+      problem_sizes1[expert_id * 3] = n1;
       problem_sizes1[expert_id * 3 + 1] = final_occurrences;
       problem_sizes1[expert_id * 3 + 2] = k;
       problem_sizes2[expert_id * 3] = k;
@@ -107,13 +110,11 @@ __global__ void compute_arg_sorts(const int32_t* __restrict__ topk_ids,
 }
 
 namespace {
-inline void launch_compute_problem_sizes(const torch::Tensor& topk_ids,
-                                         torch::Tensor& problem_sizes1,
-                                         torch::Tensor& problem_sizes2,
-                                         torch::Tensor& atomic_buffer,
-                                         int64_t num_experts, int64_t n,
-                                         int64_t k, cudaStream_t stream,
-                                         const bool swap_ab) {
+inline void launch_compute_problem_sizes(
+    const torch::Tensor& topk_ids, torch::Tensor& problem_sizes1,
+    torch::Tensor& problem_sizes2, torch::Tensor& atomic_buffer,
+    int64_t num_experts, int64_t n, int64_t k, cudaStream_t stream,
+    const bool swap_ab, const bool is_gated) {
   int num_threads = min(THREADS_PER_EXPERT, topk_ids.numel());
 
   auto const* topk_ptr = topk_ids.data_ptr<int32_t>();
@@ -125,7 +126,7 @@ inline void launch_compute_problem_sizes(const torch::Tensor& topk_ids,
     compute_problem_sizes<SwapAB><<<num_experts, num_threads, 0, stream>>>(
         topk_ptr, ps1_ptr, ps2_ptr, atomic_ptr,
         static_cast<int>(topk_ids.numel()), static_cast<int>(n),
-        static_cast<int>(k));
+        static_cast<int>(k), is_gated);
   });
 }
 }  // namespace
@@ -222,7 +223,8 @@ void get_cutlass_moe_mm_data_caller(
     torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
     torch::Tensor& input_permutation, torch::Tensor& output_permutation,
     const int64_t num_experts, const int64_t n, const int64_t k,
-    const std::optional<torch::Tensor>& blockscale_offsets) {
+    const std::optional<torch::Tensor>& blockscale_offsets,
+    const bool is_gated) {
   auto stream = at::cuda::getCurrentCUDAStream(topk_ids.device().index());
   auto options_int32 =
       torch::TensorOptions().dtype(torch::kInt32).device(topk_ids.device());
@@ -236,7 +238,7 @@ void get_cutlass_moe_mm_data_caller(
 
   launch_compute_problem_sizes(topk_ids, problem_sizes1, problem_sizes2,
                                atomic_buffer, num_experts, n, k, stream,
-                               may_swap_ab);
+                               may_swap_ab, is_gated);
 
   if (blockscale_offsets.has_value()) {
     // fp4 path

csrc/quantization/w8a8/cutlass/scaled_mm_entry.cu

@@ -75,7 +75,8 @@ void get_cutlass_moe_mm_data_caller(
     torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
     torch::Tensor& input_permutation, torch::Tensor& output_permutation,
     const int64_t num_experts, const int64_t n, const int64_t k,
-    const std::optional<torch::Tensor>& blockscale_offsets);
+    const std::optional<torch::Tensor>& blockscale_offsets,
+    const bool is_gated);
 
 void get_cutlass_moe_mm_problem_sizes_from_expert_offsets_caller(
     const torch::Tensor& expert_first_token_offset,
@@ -278,7 +279,8 @@ void get_cutlass_moe_mm_data(
     torch::Tensor& problem_sizes1, torch::Tensor& problem_sizes2,
     torch::Tensor& input_permutation, torch::Tensor& output_permutation,
     const int64_t num_experts, const int64_t n, const int64_t k,
-    const std::optional<torch::Tensor>& blockscale_offsets) {
+    const std::optional<torch::Tensor>& blockscale_offsets,
+    const bool is_gated) {
   // This function currently gets compiled only if we have a valid cutlass moe
   // mm to run it for.
   int32_t version_num = get_sm_version_num();
@@ -288,7 +290,7 @@ void get_cutlass_moe_mm_data(
   get_cutlass_moe_mm_data_caller(topk_ids, expert_offsets, problem_sizes1,
                                  problem_sizes2, input_permutation,
                                  output_permutation, num_experts, n, k,
-                                 blockscale_offsets);
+                                 blockscale_offsets, is_gated);
   return;
 #endif
   TORCH_CHECK_NOT_IMPLEMENTED(

csrc/rocm/skinny_gemms.cu

@@ -26,6 +26,16 @@
   #define __HIP__GFX9__
 #endif
 
+#if defined(__HIPCC__) &&                                                    \
+    (defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1150__) || \
+     defined(__gfx1151__) || defined(__gfx1200__) || defined(__gfx1201__))
+  #define __HIP__GFX1X__
+#endif
+
+#if defined(__HIPCC__) && (defined(__gfx1200__) || defined(__gfx1201__))
+  #define __HIP__GFX12__
+#endif
+
 #if defined(__HIPCC__) && (defined(__gfx942__) || defined(__gfx950__))
   #define __HIP__MI3XX__
 #endif
@@ -37,15 +47,31 @@
 #endif
 
 int get_lds_size() {
-  static bool is_cached = false;
-  static int result;
-  if (is_cached == false) {
-    auto dprops = at::cuda::getCurrentDeviceProperties();
-    std::string device_arch = dprops->gcnArchName;
-    size_t substring = device_arch.find("gfx95");
-    result = (substring == std::string::npos ? 64 * 1024 : 160 * 1024);
-    is_cached = true;
-  }
+  static const int result = [] {
+    const auto* dprops = at::cuda::getCurrentDeviceProperties();
+    const std::string device_arch = dprops->gcnArchName;
+    return device_arch.find("gfx95") == std::string::npos ? 64 * 1024
+                                                          : 160 * 1024;
+  }();
+  return result;
+}
+
+bool on_gfx1x() {
+  static const bool result = [] {
+    const auto* dprops = at::cuda::getCurrentDeviceProperties();
+    const std::string device_arch = dprops->gcnArchName;
+    return device_arch.find("gfx11") != std::string::npos ||
+           device_arch.find("gfx12") != std::string::npos;
+  }();
+  return result;
+}
+
+bool on_gfx12() {
+  static const bool result = [] {
+    const auto* dprops = at::cuda::getCurrentDeviceProperties();
+    const std::string device_arch = dprops->gcnArchName;
+    return device_arch.find("gfx12") != std::string::npos;
+  }();
   return result;
 }
 
@@ -286,21 +312,35 @@ torch::Tensor LLMM1(at::Tensor& in_a, at::Tensor& in_b,
   return out_c;
 }
 
-#define DOT2C(V0, V2, V3)                                                     \
-  if constexpr (std::is_same_v<scalar_t, half>) {                             \
-    asm("v_dot2c_f32_f16 %0, %2, %3" : "=v"(V0) : "0"(V0), "v"(V2), "v"(V3)); \
-  } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {            \
-    float2 s = __bfloat1622float2(*((__hip_bfloat162*)(&(V2)))) *             \
-               __bfloat1622float2(*((__hip_bfloat162*)(&(V3))));              \
-    V0 += (s.x + s.y);                                                        \
-  }
+#if defined(__HIP__GFX9__) && !defined(__HIP__GFX1X__)
+  #define DOT2C(V0, V2, V3)                                          \
+    if constexpr (std::is_same_v<scalar_t, half>) {                  \
+      asm("v_dot2c_f32_f16 %0, %2, %3"                               \
+          : "=v"(V0)                                                 \
+          : "0"(V0), "v"(V2), "v"(V3));                              \
+    } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) { \
+      float2 s = __bfloat1622float2(*((__hip_bfloat162*)(&(V2)))) *  \
+                 __bfloat1622float2(*((__hip_bfloat162*)(&(V3))));   \
+      V0 += (s.x + s.y);                                             \
+    }
+#elif defined(__HIP__GFX1X__)
+  // gfx1x: v_dot2_f32_f16 (VOP3-P, dot10-insts, available on gfx11+gfx12)
+  #define DOT2C(V0, V2, V3)                                               \
+    if constexpr (std::is_same_v<scalar_t, half>) {                       \
+      asm("v_dot2_f32_f16 %0, %1, %2, %0" : "+v"(V0) : "v"(V2), "v"(V3)); \
+    } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {      \
+      float2 s = __bfloat1622float2(*((__hip_bfloat162*)(&(V2)))) *       \
+                 __bfloat1622float2(*((__hip_bfloat162*)(&(V3))));        \
+      V0 += (s.x + s.y);                                                  \
+    }
+#endif
 
 // To avoid LLVM silently upcasting to double
 __device__ inline unsigned int min__(uint32_t a, uint32_t b) {
   return min(a, b);
 }
 
-#if defined(__HIP__GFX9__)  // TODO: Add NAVI support
+#if defined(__HIP__GFX9__) || defined(__HIP__GFX1X__)
 // This version targets cases where A[] fits LDS capacity
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
@@ -442,14 +482,18 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
                                                 1);  // row_shr2
           sum[n][y] += __builtin_amdgcn_mov_dpp(sum[n][y], 0x111, 0xf, 0xf,
                                                 1);  // row_shr1
+  #if defined(__HIP__GFX9__)
           sum[n][y] += __builtin_amdgcn_mov_dpp(sum[n][y], 0x142, 0xf, 0xf,
                                                 1);  // ROW_BCAST15
           sum[n][y] += __builtin_amdgcn_mov_dpp(sum[n][y], 0x143, 0xf, 0xf,
                                                 1);  // ROW_BCAST31
+  #else
+          sum[n][y] += __shfl_xor(sum[n][y], 16);
+  #endif
         }
       }
 
-      if (threadIdx.x == 63) {
+      if (threadIdx.x == (THRDS - 1)) {
         scalar_t biases[N][YTILE] = {};
         if (BIAS)
           for (int n = 0; n < N; n++) {
@@ -469,9 +513,10 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         }
       }
     } else {
-  #pragma unroll
+  #ifdef __HIP__GFX9__
+    #pragma unroll
       for (int n = 0; n < N; n++) {
-  #pragma unroll
+    #pragma unroll
         for (int y = 0; y < YTILE; y++) {
           /*float accm1 = 0;
            for (int i=0; i<64; i++)
@@ -498,7 +543,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
           sum4[n][y][0] = accm;
         }
       }
-      if (threadIdx.x == 63) {
+      if (threadIdx.x == (THRDS - 1)) {
         scalar_t biases[N][YTILE] = {};
         if (BIAS)
           for (int n = 0; n < N; n++) {
@@ -513,11 +558,12 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
           }
         }
       }
+  #endif  // __HIP__GFX9__ (MFMA path)
     }
     m += CuCount * _WvPrGrp * YTILE;
   }
 }
-#else   // !defined(__HIP__GFX9__) TODO: Add NAVI support
+#else
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
 __global__ void wvSplitK_hf_sml_(const int K, const int Kbp, const int Kap,
@@ -528,9 +574,9 @@ __global__ void wvSplitK_hf_sml_(const int K, const int Kbp, const int Kap,
                                  const int _WvPrGrp, const int CuCount) {
   UNREACHABLE_CODE
 }
-#endif  // defined(__HIP__GFX9__) TODO: Add NAVI support
+#endif
 
-#if defined(__HIP__GFX9__)  // TODO: Add NAVI support
+#if defined(__HIP__GFX9__) || defined(__HIP__GFX1X__)
 // This version targets cases where A[] marginally exceeds LDS capacity
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
@@ -657,14 +703,18 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
                                                 1);  // row_shr2
           sum[n][y] += __builtin_amdgcn_mov_dpp(sum[n][y], 0x111, 0xf, 0xf,
                                                 1);  // row_shr1
+  #if defined(__HIP__GFX9__)
           sum[n][y] += __builtin_amdgcn_mov_dpp(sum[n][y], 0x142, 0xf, 0xf,
                                                 1);  // ROW_BCAST15
           sum[n][y] += __builtin_amdgcn_mov_dpp(sum[n][y], 0x143, 0xf, 0xf,
                                                 1);  // ROW_BCAST31
+  #else
+          sum[n][y] += __shfl_xor(sum[n][y], 16);
+  #endif
         }
       }
 
-      if (threadIdx.x == 63) {
+      if (threadIdx.x == (THRDS - 1)) {
         scalar_t biases[N][YTILE] = {};
         if (BIAS)
           for (int n = 0; n < N; n++) {
@@ -686,9 +736,10 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         }
       }
     } else {
-  #pragma unroll
+  #ifdef __HIP__GFX9__
+    #pragma unroll
       for (int n = 0; n < N; n++) {
-  #pragma unroll
+    #pragma unroll
         for (int y = 0; y < YTILE; y++) {
           // float accm1 = 0;
           // for (int i=0; i<64; i++)
@@ -713,7 +764,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
           sum4[n][y][0] = accm;
         }
       }
-      if (threadIdx.x == 63) {
+      if (threadIdx.x == (THRDS - 1)) {
         scalar_t biases[N][YTILE] = {};
         if (BIAS)
           for (int n = 0; n < N; n++) {
@@ -730,6 +781,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
           }
         }
       }
+  #endif  // __HIP__GFX9__ (MFMA path)
     }
 
     m += CuCount * _WvPrGrp * YTILE;
@@ -746,7 +798,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   }
 }
 
-#else   // !defined(__HIP__GFX9__) TODO: Add NAVI support
+#else
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
 __global__ void wvSplitK_hf_(const int K, const int Kbp, const int Kap,
@@ -756,9 +808,9 @@ __global__ void wvSplitK_hf_(const int K, const int Kbp, const int Kap,
                              const int _WvPrGrp, const int CuCount) {
   UNREACHABLE_CODE
 }
-#endif  // defined(__HIP__GFX9__) TODO: Add NAVI support
+#endif
 
-#if defined(__HIP__GFX9__)  // TODO: Add NAVI support
+#if defined(__HIP__GFX9__) || defined(__HIP__GFX1X__)
 // This version targets big A[] cases, where it is much larger than LDS capacity
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
@@ -1004,14 +1056,18 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
                                                 1);  // row_shr2
           sum[n][y] += __builtin_amdgcn_mov_dpp(sum[n][y], 0x111, 0xf, 0xf,
                                                 1);  // row_shr1
+  #if defined(__HIP__GFX9__)
           sum[n][y] += __builtin_amdgcn_mov_dpp(sum[n][y], 0x142, 0xf, 0xf,
                                                 1);  // ROW_BCAST15
           sum[n][y] += __builtin_amdgcn_mov_dpp(sum[n][y], 0x143, 0xf, 0xf,
                                                 1);  // ROW_BCAST31
+  #else
+          sum[n][y] += __shfl_xor(sum[n][y], 16);
+  #endif
         }
       }
 
-      if (threadIdx.x == 63) {
+      if (threadIdx.x == (THRDS - 1)) {
         scalar_t biases[N][YTILE] = {};
         if (BIAS)
           for (int n = 0; n < N; n++) {
@@ -1033,9 +1089,10 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         }
       }
     } else {
-  #pragma unroll
+  #ifdef __HIP__GFX9__
+    #pragma unroll
       for (int n = 0; n < N; n++) {
-  #pragma unroll
+    #pragma unroll
         for (int y = 0; y < YTILE; y++) {
           float accm = sum4[n][y][0];
           accm += __builtin_amdgcn_mov_dpp(sum4[n][y][1], 0x101, 0xf, 0xf,
@@ -1057,7 +1114,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
           sum4[n][y][0] = accm;
         }
       }
-      if (threadIdx.x == 63) {
+      if (threadIdx.x == (THRDS - 1)) {
         scalar_t biases[N][YTILE] = {};
         if (BIAS)
           for (int n = 0; n < N; n++) {
@@ -1074,6 +1131,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
           }
         }
       }
+  #endif  // __HIP__GFX9__ (MFMA path)
     }
 
     m += CuCount * _WvPrGrp * YTILE;
@@ -1090,7 +1148,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     }
   }
 }
-#else   // !defined(__HIP__GFX9__) TODO: Add NAVI support
+#else
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N>
 __global__ void wvSplitK_hf_big_(const int K, const int Kbp, const int Kap,
@@ -1101,7 +1159,7 @@ __global__ void wvSplitK_hf_big_(const int K, const int Kbp, const int Kap,
                                  const int _WvPrGrp, const int CuCount) {
   UNREACHABLE_CODE
 }
-#endif  // defined(__HIP__GFX9__) TODO: Add NAVI support
+#endif
 
 // Find the min val of div2 that doesn't increase N/(div1*div2)
 int mindiv(int N, int div1, int div2) {
@@ -1148,40 +1206,40 @@ torch::Tensor wvSplitK(const at::Tensor& in_a, const at::Tensor& in_b,
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   const int max_lds_len = get_lds_size() / 2;
 
-#define WVSPLITK(_YTILE, _UNRL, _N)                                           \
+#define WVSPLITK_CFG(_THRDS, _WVPRGRP, _YTILE, _UNRL, _N)                     \
   {                                                                           \
-    dim3 block(64, 16);                                                       \
-    int __wvPrGrp = mindiv(M_in, CuCount * _YTILE, 16);                       \
+    dim3 block(_THRDS, _WVPRGRP);                                             \
+    int __wvPrGrp = mindiv(M_in, CuCount * _YTILE, _WVPRGRP);                 \
     if ((Kbp_in * N_in <= max_lds_len) && (M_in % _YTILE == 0))               \
-      wvSplitK_hf_sml_<fptype, 64, _YTILE, 16, 8, _UNRL, _N>                  \
+      wvSplitK_hf_sml_<fptype, _THRDS, _YTILE, _WVPRGRP, 8, _UNRL, _N>        \
           <<<grid, block, 0, stream>>>(K_in, Kap_in, Kbp_in, M_in, Bx_in,     \
                                        By_in, af4, bf4, biasf4, c, __wvPrGrp, \
                                        CuCount);                              \
     else if (Kbp_in * N_in <= max_lds_len * 1.2)                              \
-      wvSplitK_hf_<fptype, 64, _YTILE, 16, 8, _UNRL, _N>                      \
+      wvSplitK_hf_<fptype, _THRDS, _YTILE, _WVPRGRP, 8, _UNRL, _N>            \
           <<<grid, block, 0, stream>>>(K_in, Kap_in, Kbp_in, M_in, Bx_in,     \
                                        By_in, af4, bf4, biasf4, c, __wvPrGrp, \
                                        CuCount);                              \
     else                                                                      \
-      wvSplitK_hf_big_<fptype, 64, _YTILE, 16, 8, _UNRL, _N>                  \
+      wvSplitK_hf_big_<fptype, _THRDS, _YTILE, _WVPRGRP, 8, _UNRL, _N>        \
           <<<grid, block, 0, stream>>>(K_in, Kap_in, Kbp_in, M_in, Bx_in,     \
                                        By_in, af4, bf4, biasf4, c, __wvPrGrp, \
                                        CuCount);                              \
   }
 
-#define WVSPLIT_TILE(_sYT, __N)                           \
+#define WVSPLIT_TILE_CFG(_THRDS, _WVPRGRP, _sYT, __N)     \
   {                                                       \
     bool fit_lds = (Kbp_in * N_in <= max_lds_len);        \
     if (_sYT <= 1)                                        \
-      WVSPLITK(1, 4, __N)                                 \
+      WVSPLITK_CFG(_THRDS, _WVPRGRP, 1, 4, __N)           \
     else if ((__N == 1) || (!fit_lds) || (_sYT <= 4 * 2)) \
-      WVSPLITK(2, 2, __N)                                 \
+      WVSPLITK_CFG(_THRDS, _WVPRGRP, 2, 2, __N)           \
     else if (_sYT <= 4 * 3)                               \
-      WVSPLITK(3, 2, __N)                                 \
+      WVSPLITK_CFG(_THRDS, _WVPRGRP, 3, 2, __N)           \
     else if (__N == 4)                                    \
-      WVSPLITK(4, 1, __N)                                 \
+      WVSPLITK_CFG(_THRDS, _WVPRGRP, 4, 1, __N)           \
     else                                                  \
-      WVSPLITK(4, 2, __N)                                 \
+      WVSPLITK_CFG(_THRDS, _WVPRGRP, 4, 2, __N)           \
   }
 
   AT_DISPATCH_REDUCED_FLOATING_TYPES(in_b.scalar_type(), "wvSplitK", [&] {
@@ -1198,18 +1256,31 @@ torch::Tensor wvSplitK(const at::Tensor& in_a, const at::Tensor& in_b,
     // then cut the active waves to balance their distribution...
     int sYT = (M_in + CuCount * 4 - 1) / (CuCount * 4);
 
+    const bool use_wave32 = on_gfx1x();
     switch (N_in) {
       case 1:
-        WVSPLIT_TILE(sYT, 1)
+        if (use_wave32)
+          WVSPLIT_TILE_CFG(32, 16, sYT, 1)
+        else
+          WVSPLIT_TILE_CFG(64, 16, sYT, 1)
         break;
       case 2:
-        WVSPLIT_TILE(sYT, 2)
+        if (use_wave32)
+          WVSPLIT_TILE_CFG(32, 16, sYT, 2)
+        else
+          WVSPLIT_TILE_CFG(64, 16, sYT, 2)
         break;
       case 3:
-        WVSPLIT_TILE(sYT, 3)
+        if (use_wave32)
+          WVSPLIT_TILE_CFG(32, 16, sYT, 3)
+        else
+          WVSPLIT_TILE_CFG(64, 16, sYT, 3)
         break;
       case 4:
-        WVSPLIT_TILE(sYT, 4)
+        if (use_wave32)
+          WVSPLIT_TILE_CFG(32, 16, sYT, 4)
+        else
+          WVSPLIT_TILE_CFG(64, 16, sYT, 4)
         break;
       default:
         throw std::runtime_error(
@@ -1653,7 +1724,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   #endif
   }
 }
-#else   // !defined(__HIP__GFX9__) TODO: Add NAVI support
+#else
 template <typename scalar_t, int THRDS, int YTILE, int WvPrGrp, int A_CHUNK,
           int UNRL, int N, int GrpsShrB, int CHUNKK, int DTRMNSTC>
 __global__ void wvSplitKrc_(const int actlN, const int K, const int Kap,
@@ -1688,6 +1759,8 @@ torch::Tensor wvSplitKrc(const at::Tensor& in_a, const at::Tensor& in_b,
   TORCH_CHECK(in_a.dtype() == torch::kFloat16 ||
               in_a.dtype() == torch::kBFloat16);
 
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(in_a));
+
   auto out_c = torch::empty(
       {N_in, M_in},
       torch::TensorOptions().dtype(in_a.dtype()).device(in_a.device()));
@@ -1696,7 +1769,6 @@ torch::Tensor wvSplitKrc(const at::Tensor& in_a, const at::Tensor& in_b,
 
   dim3 grid(CuCount);
 
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(in_a));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   // const int max_lds_len = get_lds_size() / 2;
 
@@ -1773,7 +1845,7 @@ torch::Tensor wvSplitKrc(const at::Tensor& in_a, const at::Tensor& in_b,
   return out_c;
 }
 
-#if defined(__HIP__MI3XX__)  // TODO: Add NAVI support
+#if defined(__HIP__MI3XX__) || defined(__HIP__GFX12__)
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
@@ -1817,12 +1889,17 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
 
   uint32_t m = (blockIdx.x * _WvPrGrp + (threadIdx.y % _WvPrGrp)) * YTILE;
 
-  using floatx16 = __attribute__((__vector_size__(16 * sizeof(float)))) float;
   float sA = *s_A;
   float sB = *s_B;
 
   while (m < M) {
+  #ifdef __HIP__GFX12__
+    // gfx12: per-lane scalar accumulation via v_dot4_f32_fp8_fp8
+    float sum[N][YTILE] = {};
+  #else
+    // gfx9: MFMA accumulation
     scalar8 sum[N][YTILE] = {};
+  #endif
     for (uint32_t k1 = 0; k1 < K; k1 += THRDS * A_CHUNK * UNRL) {
       bigType bigA[N][UNRL] = {};
       bigType bigB[YTILE][UNRL];
@@ -1854,6 +1931,17 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   #pragma unroll
       for (uint32_t k2 = 0; k2 < UNRL; k2++) {
         for (uint32_t n = 0; n < N; n++) {
+  #ifdef __HIP__GFX12__
+          // gfx12: 4 x dot4 per A_CHUNK=16 bytes (4 FP8 per dot4)
+          for (int y = 0; y < YTILE; ++y) {
+    #pragma unroll
+            for (int i = 0; i < A_CHUNK / 4; i++) {
+              sum[n][y] = __builtin_amdgcn_dot4_f32_fp8_fp8(
+                  bigA[n][k2].i[i], bigB[y][k2].i[i], sum[n][y]);
+            }
+          }
+  #else
+          // gfx9: MFMA path
           for (int i = 0; i < A_CHUNK; i += 8) {
             for (int y = 0; y < YTILE; ++y) {
               sum[n][y] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(
@@ -1861,11 +1949,33 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
                   0);
             }
           }
+  #endif
         }
       }
     }
 
     // Final reduction
+  #ifdef __HIP__GFX12__
+    // gfx12 wave32: DPP row_shr within 16-lane rows + cross-row shuffle
+    for (int n = 0; n < N; n++) {
+      for (int y = 0; y < YTILE; y++) {
+        asm("s_nop 0\n\tv_add_f32 %0, %2, %3 row_shr:8 bound_ctrl:0 "
+            : "=v"(sum[n][y])
+            : "0"(sum[n][y]), "v"(sum[n][y]), "v"(sum[n][y]));
+        asm("s_nop 0\n\tv_add_f32 %0, %2, %3 row_shr:4 bound_ctrl:0 "
+            : "=v"(sum[n][y])
+            : "0"(sum[n][y]), "v"(sum[n][y]), "v"(sum[n][y]));
+        asm("s_nop 0\n\tv_add_f32 %0, %2, %3 row_shr:2 bound_ctrl:0 "
+            : "=v"(sum[n][y])
+            : "0"(sum[n][y]), "v"(sum[n][y]), "v"(sum[n][y]));
+        asm("s_nop 0\n\tv_add_f32 %0, %2, %3 row_shr:1 bound_ctrl:0 "
+            : "=v"(sum[n][y])
+            : "0"(sum[n][y]), "v"(sum[n][y]), "v"(sum[n][y]));
+        sum[n][y] += __shfl_xor(sum[n][y], 16);
+      }
+    }
+  #else
+    // gfx9 MFMA reduction
     for (int n = 0; n < N; n++) {
       for (int y = 0; y < YTILE; y++) {
         float accm0 = sum[n][y][0];
@@ -1880,8 +1990,15 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         sum[n][y][0] = accm0;
       }
     }
+  #endif
 
-    if (threadIdx.x == 0) {
+    const bool writeback_lane =
+  #ifdef __HIP__GFX12__
+        threadIdx.x == (THRDS - 1);
+  #else
+        threadIdx.x == 0;
+  #endif
+    if (writeback_lane) {
       scalar_t biases[N][YTILE] = {};
       if (BIAS)
         for (int n = 0; n < N; n++) {
@@ -1892,13 +2009,17 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       for (int n = 0; n < N; n++) {
         for (int y = 0; y < YTILE; y++) {
           if (y + m >= M) break;  // To avoid mem access fault.
-          sum[n][y][0] *= sA * sB;
+  #ifdef __HIP__GFX12__
+          float result = sum[n][y] * sA * sB;
+  #else
+          float result = sum[n][y][0] * sA * sB;
+  #endif
           if constexpr (std::is_same_v<scalar_t, half>) {
-            sum[n][y][0] += __half2float(biases[n][y]);
+            result += __half2float(biases[n][y]);
           } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
-            sum[n][y][0] += __bfloat162float(biases[n][y]);
+            result += __bfloat162float(biases[n][y]);
           }
-          C[m + y + n * M] = __float2s<scalar_t>(sum[n][y][0]);
+          C[m + y + n * M] = __float2s<scalar_t>(result);
         }
       }
     }
@@ -1906,7 +2027,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     m += CuCount * _WvPrGrp * YTILE;
   }
 }
-#else   // !defined(__HIP__MI3XX__) TODO: Add NAVI support
+#else   // !defined(__HIP__MI3XX__) && !defined(__HIP__GFX12__)
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
 __global__ void wvSplitKQ_hf_sml_(const int K, const int Kap, const int Kbp,
@@ -1918,9 +2039,9 @@ __global__ void wvSplitKQ_hf_sml_(const int K, const int Kap, const int Kbp,
                                   const int _WvPrGrp, const int CuCount) {
   UNREACHABLE_CODE
 }
-#endif  // defined(__HIP__MI3XX__) TODO: Add NAVI support
+#endif  // defined(__HIP__MI3XX__) || defined(__HIP__GFX12__)
 
-#if defined(__HIP__MI3XX__)  // TODO: Add NAVI support
+#if defined(__HIP__MI3XX__) || defined(__HIP__GFX12__)
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
 __global__ void __launch_bounds__(WvPrGrp* THRDS)
@@ -1963,12 +2084,17 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
 
   uint32_t m = (blockIdx.x * _WvPrGrp + (threadIdx.y % _WvPrGrp)) * YTILE;
 
-  using floatx16 = __attribute__((__vector_size__(16 * sizeof(float)))) float;
   float sA = *s_A;
   float sB = *s_B;
 
   while (m < M) {
+  #ifdef __HIP__GFX12__
+    // gfx12: per-lane scalar accumulation via v_dot4_f32_fp8_fp8
+    float sum[N][YTILE] = {};
+  #else
+    // gfx9: MFMA accumulation
     scalar8 sum[N][YTILE] = {};
+  #endif
     for (uint32_t k1 = 0; k1 < K; k1 += THRDS * A_CHUNK * UNRL) {
       bigType bigA[N][UNRL] = {};
       bigType bigB[YTILE][UNRL];
@@ -2002,6 +2128,17 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
   #pragma unroll
       for (uint32_t k2 = 0; k2 < UNRL; k2++) {
         for (uint32_t n = 0; n < N; n++) {
+  #ifdef __HIP__GFX12__
+          // gfx12: 4 x dot4 per A_CHUNK=16 bytes (4 FP8 per dot4)
+          for (int y = 0; y < YTILE; ++y) {
+    #pragma unroll
+            for (int i = 0; i < A_CHUNK / 4; i++) {
+              sum[n][y] = __builtin_amdgcn_dot4_f32_fp8_fp8(
+                  bigA[n][k2].i[i], bigB[y][k2].i[i], sum[n][y]);
+            }
+          }
+  #else
+          // gfx9: MFMA path
           for (int i = 0; i < A_CHUNK; i += 8) {
             for (int y = 0; y < YTILE; ++y) {
               sum[n][y] = __builtin_amdgcn_mfma_f32_16x16x32_fp8_fp8(
@@ -2009,11 +2146,33 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
                   0);
             }
           }
+  #endif
         }
       }
     }
 
     // Final reduction
+  #ifdef __HIP__GFX12__
+    // gfx12 wave32: DPP row_shr within 16-lane rows + cross-row shuffle
+    for (int n = 0; n < N; n++) {
+      for (int y = 0; y < YTILE; y++) {
+        asm("s_nop 0\n\tv_add_f32 %0, %2, %3 row_shr:8 bound_ctrl:0 "
+            : "=v"(sum[n][y])
+            : "0"(sum[n][y]), "v"(sum[n][y]), "v"(sum[n][y]));
+        asm("s_nop 0\n\tv_add_f32 %0, %2, %3 row_shr:4 bound_ctrl:0 "
+            : "=v"(sum[n][y])
+            : "0"(sum[n][y]), "v"(sum[n][y]), "v"(sum[n][y]));
+        asm("s_nop 0\n\tv_add_f32 %0, %2, %3 row_shr:2 bound_ctrl:0 "
+            : "=v"(sum[n][y])
+            : "0"(sum[n][y]), "v"(sum[n][y]), "v"(sum[n][y]));
+        asm("s_nop 0\n\tv_add_f32 %0, %2, %3 row_shr:1 bound_ctrl:0 "
+            : "=v"(sum[n][y])
+            : "0"(sum[n][y]), "v"(sum[n][y]), "v"(sum[n][y]));
+        sum[n][y] += __shfl_xor(sum[n][y], 16);
+      }
+    }
+  #else
+    // gfx9 MFMA reduction
     for (int n = 0; n < N; n++) {
       for (int y = 0; y < YTILE; y++) {
         float accm0 = sum[n][y][0];
@@ -2028,8 +2187,15 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
         sum[n][y][0] = accm0;
       }
     }
+  #endif
 
-    if (threadIdx.x == 0) {
+    const bool writeback_lane =
+  #ifdef __HIP__GFX12__
+        threadIdx.x == (THRDS - 1);
+  #else
+        threadIdx.x == 0;
+  #endif
+    if (writeback_lane) {
       scalar_t biases[N][YTILE] = {};
       if (BIAS)
         for (int n = 0; n < N; n++) {
@@ -2040,13 +2206,17 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
       for (int n = 0; n < N; n++) {
         for (int y = 0; y < YTILE; y++) {
           if (y + m >= M) break;  // To avoid mem access fault.
-          sum[n][y][0] *= sA * sB;
+  #ifdef __HIP__GFX12__
+          float result = sum[n][y] * sA * sB;
+  #else
+          float result = sum[n][y][0] * sA * sB;
+  #endif
           if constexpr (std::is_same_v<scalar_t, half>) {
-            sum[n][y][0] += __half2float(biases[n][y]);
+            result += __half2float(biases[n][y]);
           } else if constexpr (std::is_same_v<scalar_t, __hip_bfloat16>) {
-            sum[n][y][0] += __bfloat162float(biases[n][y]);
+            result += __bfloat162float(biases[n][y]);
           }
-          C[m + y + n * M] = __float2s<scalar_t>(sum[n][y][0]);
+          C[m + y + n * M] = __float2s<scalar_t>(result);
         }
       }
     }
@@ -2054,7 +2224,7 @@ __global__ void __launch_bounds__(WvPrGrp* THRDS)
     m += CuCount * _WvPrGrp * YTILE;
   }
 }
-#else   // !defined(__HIP__MI3XX__) TODO: Add NAVI support
+#else   // !defined(__HIP__MI3XX__) && !defined(__HIP__GFX12__)
 template <typename scalar_t, typename fp8_t, int THRDS, int YTILE, int WvPrGrp,
           int A_CHUNK, int UNRL, int N>
 __global__ void wvSplitKQ_hf_(const int K, const int Kap, const int Kbp,
@@ -2066,7 +2236,7 @@ __global__ void wvSplitKQ_hf_(const int K, const int Kap, const int Kbp,
                               const int CuCount) {
   UNREACHABLE_CODE
 }
-#endif  // defined(__HIP__MI3XX__) TODO: Add NAVI support
+#endif  // defined(__HIP__MI3XX__) || defined(__HIP__GFX12__)
 
 void wvSplitKQ(const at::Tensor& in_b, const at::Tensor& in_a,
                const std::optional<at::Tensor>& in_bias, at::Tensor& out_c,
@@ -2099,24 +2269,30 @@ void wvSplitKQ(const at::Tensor& in_b, const at::Tensor& in_a,
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   const int max_lds_len = get_lds_size();
 
-#define WVSPLITKQ(_WvPrGrp, _YTILEs, _YTILEm, _UNRLs, _UNRLm, _N)             \
-  {                                                                           \
-    dim3 block(64, _WvPrGrp);                                                 \
-    if ((Kap_in * N_in <= max_lds_len) && (M_in % _YTILEs == 0)) {            \
-      int __wvPrGrp = min(_WvPrGrp, mindiv(M_in, CuCount * _YTILEs, 16));     \
-      wvSplitKQ_hf_sml_<fptype, fp8_t, 64, _YTILEs, _WvPrGrp, 16, _UNRLs, _N> \
-          <<<grid, block, 0, stream>>>(K_in, Kap_in, Kbp_in, M_in, Bx_in,     \
-                                       By_in, b_ptr, a_ptr, bias_ptr, c_ptr,  \
-                                       s_a, s_b, __wvPrGrp, CuCount);         \
-    } else {                                                                  \
-      int __wvPrGrp = min(_WvPrGrp, mindiv(M_in, CuCount * _YTILEm, 16));     \
-      wvSplitKQ_hf_<fptype, fp8_t, 64, _YTILEm, _WvPrGrp, 16, _UNRLm, _N>     \
-          <<<grid, block, 0, stream>>>(K_in, Kap_in, Kbp_in, M_in, Bx_in,     \
-                                       By_in, b_ptr, a_ptr, bias_ptr, c_ptr,  \
-                                       s_a, s_b, __wvPrGrp, CuCount);         \
-    }                                                                         \
+#define WVSPLITKQ_IMPL(_THRDS, _WvPrGrp, _YTILEs, _YTILEm, _UNRLs, _UNRLm, _N) \
+  {                                                                            \
+    dim3 block(_THRDS, _WvPrGrp);                                              \
+    if ((Kap_in * N_in <= max_lds_len) && (M_in % _YTILEs == 0)) {             \
+      int __wvPrGrp = min(_WvPrGrp, mindiv(M_in, CuCount * _YTILEs, 16));      \
+      wvSplitKQ_hf_sml_<fptype, fp8_t, _THRDS, _YTILEs, _WvPrGrp, 16, _UNRLs,  \
+                        _N><<<grid, block, 0, stream>>>(                       \
+          K_in, Kap_in, Kbp_in, M_in, Bx_in, By_in, b_ptr, a_ptr, bias_ptr,    \
+          c_ptr, s_a, s_b, __wvPrGrp, CuCount);                                \
+    } else {                                                                   \
+      int __wvPrGrp = min(_WvPrGrp, mindiv(M_in, CuCount * _YTILEm, 16));      \
+      wvSplitKQ_hf_<fptype, fp8_t, _THRDS, _YTILEm, _WvPrGrp, 16, _UNRLm, _N>  \
+          <<<grid, block, 0, stream>>>(K_in, Kap_in, Kbp_in, M_in, Bx_in,      \
+                                       By_in, b_ptr, a_ptr, bias_ptr, c_ptr,   \
+                                       s_a, s_b, __wvPrGrp, CuCount);          \
+    }                                                                          \
   }
 
+#define WVSPLITKQ(_WvPrGrp, _YTILEs, _YTILEm, _UNRLs, _UNRLm, _N)      \
+  if (on_gfx12())                                                      \
+    WVSPLITKQ_IMPL(32, _WvPrGrp, _YTILEs, _YTILEm, _UNRLs, _UNRLm, _N) \
+  else                                                                 \
+    WVSPLITKQ_IMPL(64, _WvPrGrp, _YTILEs, _YTILEm, _UNRLs, _UNRLm, _N)
+
   AT_DISPATCH_REDUCED_FLOATING_TYPES(out_c.scalar_type(), "wvSplitKQ", [&] {
     using fptype = typename scalar<scalar_t>::type;
     auto c_ptr = reinterpret_cast<fptype*>(out_c.data_ptr());
@@ -2136,10 +2312,10 @@ void wvSplitKQ(const at::Tensor& in_b, const at::Tensor& in_a,
           WVSPLITKQ(16, 2, 2, 2, 2, 2)
           break;
         case 3:
-          WVSPLITKQ(16, 2, 2, 2, 2, 3)
+          WVSPLITKQ(16, 2, 2, 1, 1, 3)
           break;
         case 4:
-          WVSPLITKQ(16, 2, 2, 2, 2, 4)
+          WVSPLITKQ(16, 2, 2, 1, 1, 4)
           break;
         default:
           throw std::runtime_error(

csrc/sampler.cu

@@ -575,7 +575,7 @@ static __global__ __launch_bounds__(kNumThreadsPerBlock) void topKPerRowDecode(
   // The range of logits within the row.
   int rowStart = 0;
   int seq_len = seqLens[rowIdx / next_n];
-  int rowEnd = seq_len - next_n + (rowIdx % next_n) + 1;
+  int rowEnd = max(0, seq_len - next_n + (rowIdx % next_n) + 1);
 
   // Local pointers to this block
   if constexpr (!multipleBlocksPerRow && !mergeBlocks) {

csrc/torch_bindings.cpp

@@ -303,9 +303,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       ") -> Tensor");
   // conditionally compiled so impl registration is in source file
 
-  ops.def("permute_cols(Tensor A, Tensor perm) -> Tensor");
-  ops.impl("permute_cols", torch::kCUDA, &permute_cols);
-
   // Marlin Optimized Quantized GEMM (supports GPTQ, AWQ, FP8, NVFP4, MXFP4).
   ops.def(
       "marlin_gemm(Tensor a, Tensor? c_or_none, Tensor b_q_weight, "
@@ -489,8 +486,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "                        Tensor! problem_sizes1, Tensor! problem_sizes2, "
       "                        Tensor! input_permutation, "
       "                        Tensor! output_permutation, int num_experts, "
-      "                        int n, int k, Tensor? blockscale_offsets) -> "
-      "()");
+      "                        int n, int k, Tensor? blockscale_offsets, "
+      "                        bool is_gated) -> ()");
   ops.impl("get_cutlass_moe_mm_data", torch::kCUDA, &get_cutlass_moe_mm_data);
 
   // compute per-expert problem sizes from expert_first_token_offset
@@ -564,10 +561,21 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 
   // Compute NVFP4 block quantized tensor.
   ops.def(
-      "scaled_fp4_quant(Tensor! output, Tensor input,"
-      "                 Tensor! output_scale, Tensor input_scale, bool "
-      "is_sf_swizzled_layout) -> ()");
-  ops.impl("scaled_fp4_quant", torch::kCUDA, &scaled_fp4_quant);
+      "scaled_fp4_quant(Tensor input,"
+      "                 Tensor input_scale, bool "
+      "is_sf_swizzled_layout) -> (Tensor, Tensor)");
+  ops.impl("scaled_fp4_quant", torch::kCUDA, &scaled_fp4_quant_func);
+
+  // Out variant
+  // TODO: Add {at::Tag::out_variant} tag and update all call sites
+  // to use the functional variant once vLLM upgrades PyTorch.
+  // See pytorch/pytorch#176117.
+  ops.def(
+      "scaled_fp4_quant.out(Tensor input,"
+      "                     Tensor input_scale, bool "
+      "is_sf_swizzled_layout, *, Tensor(a!) output, Tensor(b!) output_scale) "
+      "-> ()");
+  ops.impl("scaled_fp4_quant.out", torch::kCUDA, &scaled_fp4_quant_out);
 
   // Compute NVFP4 experts quantization.
   ops.def(

docker/Dockerfile

@@ -586,7 +586,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # This is ~1.1GB and only changes when FlashInfer version bumps
 # https://docs.flashinfer.ai/installation.html
 # From versions.json: .flashinfer.version
-ARG FLASHINFER_VERSION=0.6.4
+ARG FLASHINFER_VERSION=0.6.6
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system flashinfer-cubin==${FLASHINFER_VERSION} \
     && uv pip install --system flashinfer-jit-cache==${FLASHINFER_VERSION} \
@@ -620,7 +620,7 @@ RUN set -eux; \
 ARG BITSANDBYTES_VERSION_X86=0.46.1
 ARG BITSANDBYTES_VERSION_ARM64=0.42.0
 ARG TIMM_VERSION=">=1.0.17"
-ARG RUNAI_MODEL_STREAMER_VERSION=">=0.15.3"
+ARG RUNAI_MODEL_STREAMER_VERSION=">=0.15.7"
 RUN --mount=type=cache,target=/root/.cache/uv \
     if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
         BITSANDBYTES_VERSION="${BITSANDBYTES_VERSION_ARM64}"; \
@@ -628,7 +628,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
         BITSANDBYTES_VERSION="${BITSANDBYTES_VERSION_X86}"; \
     fi; \
     uv pip install --system accelerate hf_transfer modelscope \
-        "bitsandbytes>=${BITSANDBYTES_VERSION}" "timm${TIMM_VERSION}" "runai-model-streamer[s3,gcs]${RUNAI_MODEL_STREAMER_VERSION}"
+        "bitsandbytes>=${BITSANDBYTES_VERSION}" "timm${TIMM_VERSION}" "runai-model-streamer[s3,gcs,azure]${RUNAI_MODEL_STREAMER_VERSION}"
 
 # ============================================================
 # VLLM INSTALLATION (depends on build stage)

docker/Dockerfile.cpu

@@ -9,17 +9,13 @@
 #
 # Build targets:
 #   vllm-openai (default): used for serving deployment
+#   vllm-openai-zen: vLLM from source + zentorch from PyPI via vllm[zen]
 #   vllm-test: used for CI tests
 #   vllm-dev: used for development
 #
 # Build arguments:
 #   PYTHON_VERSION=3.13|3.12 (default)|3.11|3.10
-#   VLLM_CPU_DISABLE_AVX512=false (default)|true
-#   VLLM_CPU_AVX2=false (default)|true (for cross-compilation)
-#   VLLM_CPU_AVX512=false (default)|true (for cross-compilation)
-#   VLLM_CPU_AVX512BF16=false (default)|true (for cross-compilation)
-#   VLLM_CPU_AVX512VNNI=false (default)|true (for cross-compilation)
-#   VLLM_CPU_AMXBF16=false (default)|true (for cross-compilation)
+#   VLLM_CPU_X86=false (default)|true (for cross-compilation)
 #   VLLM_CPU_ARM_BF16=false (default)|true (for cross-compilation)
 #
 
@@ -36,7 +32,7 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt,sharing=locked \
     apt-get update -y \
     && apt-get install -y --no-install-recommends sudo ccache git curl wget ca-certificates \
-    gcc-12 g++-12 libtcmalloc-minimal4 libnuma-dev ffmpeg libsm6 libxext6 libgl1 jq lsof \
+    gcc-12 g++-12 libtcmalloc-minimal4 libnuma-dev ffmpeg libsm6 libxext6 libgl1 jq lsof make xz-utils \
     && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12 \
     && curl -LsSf https://astral.sh/uv/install.sh | sh
 
@@ -91,24 +87,9 @@ ARG max_jobs=32
 ENV MAX_JOBS=${max_jobs}
 
 ARG GIT_REPO_CHECK=0
-# Support for building with non-AVX512 vLLM: docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" ...
-ARG VLLM_CPU_DISABLE_AVX512=0
-ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
-# Support for cross-compilation with AVX2 ISA: docker build --build-arg VLLM_CPU_AVX2="1" ...
-ARG VLLM_CPU_AVX2=0
-ENV VLLM_CPU_AVX2=${VLLM_CPU_AVX2}
-# Support for cross-compilation with AVX512 ISA: docker build --build-arg VLLM_CPU_AVX512="1" ...
-ARG VLLM_CPU_AVX512=0
-ENV VLLM_CPU_AVX512=${VLLM_CPU_AVX512}
-# Support for building with AVX512BF16 ISA: docker build --build-arg VLLM_CPU_AVX512BF16="true" ...
-ARG VLLM_CPU_AVX512BF16=0
-ENV VLLM_CPU_AVX512BF16=${VLLM_CPU_AVX512BF16}
-# Support for building with AVX512VNNI ISA: docker build --build-arg VLLM_CPU_AVX512VNNI="true" ...
-ARG VLLM_CPU_AVX512VNNI=0
-ENV VLLM_CPU_AVX512VNNI=${VLLM_CPU_AVX512VNNI}
-# Support for building with AMXBF16 ISA: docker build --build-arg VLLM_CPU_AMXBF16="true" ...
-ARG VLLM_CPU_AMXBF16=1
-ENV VLLM_CPU_AMXBF16=${VLLM_CPU_AMXBF16}
+# Support for cross-compilation with x86 ISA including AVX2 and AVX512: docker build --build-arg VLLM_CPU_X86="true" ...
+ARG VLLM_CPU_X86=0
+ENV VLLM_CPU_X86=${VLLM_CPU_X86}
 # Support for cross-compilation with ARM BF16 ISA: docker build --build-arg VLLM_CPU_ARM_BF16="true" ...
 ARG VLLM_CPU_ARM_BF16=0
 ENV VLLM_CPU_ARM_BF16=${VLLM_CPU_ARM_BF16}
@@ -116,7 +97,7 @@ ENV VLLM_CPU_ARM_BF16=${VLLM_CPU_ARM_BF16}
 WORKDIR /vllm-workspace
 
 # Validate build arguments - prevent mixing incompatible ISA flags
-RUN if [ "$TARGETARCH" = "arm64" ] && { [ "$VLLM_CPU_AVX2" != "0" ] || [ "$VLLM_CPU_AVX512" != "0" ] || [ "$VLLM_CPU_AVX512BF16" != "0" ] || [ "$VLLM_CPU_AVX512VNNI" != "0" ]; }; then \
+RUN if [ "$TARGETARCH" = "arm64" ] && [ "$VLLM_CPU_X86" != "0" ]; then \
         echo "ERROR: Cannot use x86-specific ISA flags (AVX2, AVX512, etc.) when building for ARM64 (--platform=linux/arm64)"; \
         exit 1; \
     fi && \
@@ -174,7 +155,7 @@ WORKDIR /vllm-workspace
 
 RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     --mount=type=cache,target=/var/lib/apt,sharing=locked \
-    apt-get install -y --no-install-recommends vim numactl xz-utils make clangd-14
+    apt-get install -y --no-install-recommends vim numactl clangd-14
 
 RUN ln -s /usr/bin/clangd-14 /usr/bin/clangd
 
@@ -232,23 +213,29 @@ LABEL org.opencontainers.image.source="https://github.com/vllm-project/vllm"
 
 # Build configuration labels
 ARG TARGETARCH
-ARG VLLM_CPU_DISABLE_AVX512
-ARG VLLM_CPU_AVX2
-ARG VLLM_CPU_AVX512
-ARG VLLM_CPU_AVX512BF16
-ARG VLLM_CPU_AVX512VNNI
-ARG VLLM_CPU_AMXBF16
+ARG VLLM_CPU_X86
 ARG VLLM_CPU_ARM_BF16
 ARG PYTHON_VERSION
 
 LABEL ai.vllm.build.target-arch="${TARGETARCH}"
-LABEL ai.vllm.build.cpu-disable-avx512="${VLLM_CPU_DISABLE_AVX512:-false}"
-LABEL ai.vllm.build.cpu-avx2="${VLLM_CPU_AVX2:-false}"
-LABEL ai.vllm.build.cpu-avx512="${VLLM_CPU_AVX512:-false}"
-LABEL ai.vllm.build.cpu-avx512bf16="${VLLM_CPU_AVX512BF16:-false}"
-LABEL ai.vllm.build.cpu-avx512vnni="${VLLM_CPU_AVX512VNNI:-false}"
-LABEL ai.vllm.build.cpu-amxbf16="${VLLM_CPU_AMXBF16:-false}"
+LABEL ai.vllm.build.cpu-x86="${VLLM_CPU_X86:-false}"
 LABEL ai.vllm.build.cpu-arm-bf16="${VLLM_CPU_ARM_BF16:-false}"
 LABEL ai.vllm.build.python-version="${PYTHON_VERSION:-3.12}"
 
 ENTRYPOINT ["vllm", "serve"]
+
+
+######################### ZEN CPU PYPI IMAGE #########################
+FROM vllm-openai AS vllm-openai-zen
+
+ARG TARGETARCH
+
+RUN if [ "$TARGETARCH" != "amd64" ]; then \
+        echo "ERROR: vllm-openai-amd only supports --platform=linux/amd64"; \
+        exit 1; \
+    fi
+
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install "vllm[zen]"
+
+ENTRYPOINT ["vllm", "serve"]

docker/Dockerfile.nightly_torch

@@ -217,13 +217,13 @@ RUN pip install setuptools==75.6.0 packaging==23.2 ninja==1.11.1.3 build==1.2.2.
 
 
 # build flashinfer for torch nightly from source around 10 mins
-# release version: v0.6.4
+# release version: v0.6.6
 # todo(elainewy): cache flashinfer build result for faster build
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
     --mount=type=cache,target=/root/.cache/uv \
     echo "git clone flashinfer..." \
-    && git clone --depth 1 --branch v0.6.4 --recursive https://github.com/flashinfer-ai/flashinfer.git \
+    && git clone --depth 1 --branch v0.6.6 --recursive https://github.com/flashinfer-ai/flashinfer.git \
     && cd flashinfer \
     && git submodule update --init --recursive \
     && echo "finish git clone flashinfer..." \

docker/Dockerfile.rocm

@@ -184,6 +184,34 @@ RUN cd /opt/rixl && mkdir -p /app/install && \
         --ucx-plugins-dir ${UCX_HOME}/lib/ucx \
         --nixl-plugins-dir ${RIXL_HOME}/lib/x86_64-linux-gnu/plugins
 
+# DeepEP build stage
+FROM base AS build_deep
+ARG ROCSHMEM_BRANCH="ba0bf0f3"
+ARG ROCSHMEM_REPO="https://github.com/ROCm/rocm-systems.git"
+ARG DEEPEP_BRANCH="e84464ec"
+ARG DEEPEP_REPO="https://github.com/ROCm/DeepEP.git"
+ARG DEEPEP_NIC="cx7"
+ENV ROCSHMEM_DIR=/opt/rocshmem
+
+RUN git clone ${ROCSHMEM_REPO} \
+ && cd rocm-systems \
+ && git checkout ${ROCSHMEM_BRANCH} \
+ && mkdir -p projects/rocshmem/build \
+ && cd projects/rocshmem/build \
+ && cmake .. \
+    -DCMAKE_INSTALL_PREFIX="${ROCSHMEM_DIR}" \
+    -DROCM_PATH=/opt/rocm \
+    -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+    -DUSE_EXTERNAL_MPI=OFF \
+ && make -j \
+ && make install
+
+# Build DeepEP wheel.
+# DeepEP looks for rocshmem at ROCSHMEM_DIR.
+RUN git clone ${DEEPEP_REPO} \
+ && cd DeepEP \
+ && git checkout ${DEEPEP_BRANCH} \
+ && python3 setup.py --variant rocm --nic ${DEEPEP_NIC} bdist_wheel --dist-dir=/app/deep_install
 
 # -----------------------
 # vLLM wheel release build stage (for building distributable wheels)
@@ -305,6 +333,11 @@ RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
 RUN --mount=type=bind,from=build_rixl,src=/app/install,target=/rixl_install \
     uv pip install --system /rixl_install/*.whl
 
+# Install DeepEP wheel
+RUN --mount=type=bind,from=build_deep,src=/app/deep_install,target=/deep_install \
+    uv pip install --system /deep_install/*.whl
+COPY --from=build_deep /opt/rocshmem /opt/rocshmem
+
 # RIXL/MoRIIO runtime dependencies (RDMA userspace libraries)
 RUN apt-get update -q -y && apt-get install -q -y \
     librdmacm1 \

docker/Dockerfile.rocm_base

@@ -44,7 +44,7 @@ ENV DEBIAN_FRONTEND=noninteractive
 
 # Install Python and other dependencies
 RUN apt-get update -y \
-    && apt-get install -y software-properties-common git curl sudo vim less libgfortran5 libopenmpi-dev libpci-dev \
+    && apt-get install -y software-properties-common git curl sudo vim less libgfortran5 libopenmpi-dev libpci-dev liblzma-dev pkg-config \
     && for i in 1 2 3; do \
         add-apt-repository -y ppa:deadsnakes/ppa && break || \
         { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \

docker/Dockerfile.xpu

@@ -76,19 +76,22 @@ ENV UV_LINK_MODE="copy"
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,src=requirements/common.txt,target=/workspace/vllm/requirements/common.txt \
     --mount=type=bind,src=requirements/xpu.txt,target=/workspace/vllm/requirements/xpu.txt \
+    --mount=type=bind,src=requirements/xpu-test.in,target=/workspace/vllm/requirements/xpu-test.in \
     uv pip install --upgrade pip && \
-    uv pip install -r requirements/xpu.txt
-
- # used for suffix method speculative decoding
- # build deps for proto + nanobind-based extensions to set up the build environment
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install grpcio-tools protobuf nanobind
- # arctic-inference is built from source which needs torch-xpu properly installed first
-RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install -r requirements/xpu.txt && \
+    uv pip compile /workspace/vllm/requirements/xpu-test.in \
+        -o /workspace/vllm/requirements/xpu-test.txt \
+        -c /workspace/vllm/requirements/xpu.txt \
+        --index-strategy unsafe-best-match \
+        --extra-index-url ${PIP_EXTRA_INDEX_URL} \
+        --python-version ${PYTHON_VERSION} && \
+    uv pip install grpcio-tools protobuf nanobind && \
     source /opt/intel/oneapi/setvars.sh --force && \
     source /opt/intel/oneapi/ccl/2021.15/env/vars.sh --force && \
-    export CMAKE_PREFIX_PATH="$(python -c 'import site; print(site.getsitepackages()[0])'):${CMAKE_PREFIX_PATH}" && \
-    uv pip install --no-build-isolation arctic-inference==0.1.1
+    export CMAKE_PREFIX_PATH="$(python3 -c 'import site; print(site.getsitepackages()[0])'):${CMAKE_PREFIX_PATH}" && \
+    uv pip install --no-build-isolation -r /workspace/vllm/requirements/xpu-test.txt
+
+
 
 ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/lib/"
 

docker/versions.json

@@ -65,7 +65,7 @@
       "default": "true"
     },
     "FLASHINFER_VERSION": {
-      "default": "0.6.4"
+      "default": "0.6.6"
     },
     "GDRCOPY_CUDA_VERSION": {
       "default": "12.8"
@@ -83,7 +83,7 @@
       "default": ">=1.0.17"
     },
     "RUNAI_MODEL_STREAMER_VERSION": {
-      "default": ">=0.15.3"
+      "default": ">=0.15.7"
     }
   }
 }

docs/.nav.yml

@@ -25,7 +25,7 @@ nav:
     - Models:
       - models/supported_models.md
       - models/generative_models.md
-      - models/pooling_models.md
+      - Pooling Models: models/pooling_models
       - models/extensions
       - Hardware Supported Models:
         - models/hardware_supported_models/*

docs/configuration/conserving_memory.md

@@ -15,7 +15,7 @@ llm = LLM(model="ibm-granite/granite-3.1-8b-instruct", tensor_parallel_size=2)
 ```
 
 !!! warning
-    To ensure that vLLM initializes CUDA correctly, you should avoid calling related functions (e.g. [torch.cuda.set_device][])
+    To ensure that vLLM initializes CUDA correctly, you should avoid calling related functions (e.g. [torch.accelerator.set_device_index][])
     before initializing vLLM. Otherwise, you may run into an error like `RuntimeError: Cannot re-initialize CUDA in forked subprocess`.
 
     To control which devices are used, please instead set the `CUDA_VISIBLE_DEVICES` environment variable.

docs/contributing/README.md

@@ -75,7 +75,7 @@ For an optimized workflow when iterating on C++/CUDA kernels, see the [Increment
 vLLM uses `pre-commit` to lint and format the codebase. See <https://pre-commit.com/#usage> if `pre-commit` is new to you. Setting up `pre-commit` is as easy as:
 
 ```bash
-uv pip install pre-commit
+uv pip install pre-commit>=4.5.1
 pre-commit install
 ```
 
@@ -187,6 +187,30 @@ Using `-s` with `git commit` will automatically add this header.
     - **VSCode**: Open the [Settings editor](https://code.visualstudio.com/docs/configure/settings)
       and enable the `Git: Always Sign Off` (`git.alwaysSignOff`) field.
 
+### AI Assisted Contributions
+
+Before making an AI assisted contribution, you must:
+
+1. **Be involved**: Do not submit "pure agent" PRs. The human submitter is responsible for reviewing all changed lines, validating behavior end-to-end, and running relevant tests.
+2. **Ensure significance**: Avoid one-off "busywork" PRs (single typo, isolated style cleanup, one mutable default fix, etc.). Bundle mechanical cleanups into a clear, systematic scope.
+
+When AI tools provide non-trivial assistance in generating or modifying code, you must:
+
+1. **Review thoroughly**: You remain responsible for all code you submit. Review and understand AI-generated code with the same care as code you write manually.
+2. **Disclose in PR**: Always mention when a pull request includes AI-generated code. Add a note in the PR description.
+3. **Mark commits**: Add attribution using commit trailers such as `Co-authored-by:` (other projects use `Assisted-by:` or `Generated-by:`). For example:
+
+   ```text
+   Your commit message here
+
+   Co-authored-by: GitHub Copilot
+   Co-authored-by: Claude
+   Co-authored-by: gemini-code-assist
+   Signed-off-by: Your Name <your.email@example.com>
+   ```
+
+AI-assisted code must meet all quality standards: proper testing, documentation, adherence to style guides, and thorough review. Attribution helps reviewers evaluate contributions in context and maintains legal clarity for the project.
+
 ### PR Title and Classification
 
 Only specific types of PRs will be reviewed. The PR title is prefixed

docs/contributing/model/tests.md

@@ -37,7 +37,7 @@ For [generative models](../../models/generative_models.md), there are two levels
 
 #### Pooling models
 
-For [pooling models](../../models/pooling_models.md), we simply check the cosine similarity, as defined in [tests/models/utils.py](../../../tests/models/utils.py).
+For [pooling models](../../models/pooling_models/README.md), we simply check the cosine similarity, as defined in [tests/models/utils.py](../../../tests/models/utils.py).
 
 ### Multi-modal processing
 

docs/contributing/profiling.md

@@ -3,6 +3,10 @@
 !!! warning
     Profiling is only intended for vLLM developers and maintainers to understand the proportion of time spent in different parts of the codebase. **vLLM end-users should never turn on profiling** as it will significantly slow down the inference.
 
+!!! tip "Choosing a profiler"
+    - Use **Nsight Systems** for low-overhead, performance-critical profiling.
+    - Use **PyTorch Profiler** for medium-overhead profiling with richer debugging information (e.g., stack traces, memory, shapes). Note that enabling these features adds overhead and is not recommended for benchmarking.
+
 ## Profile with PyTorch Profiler
 
 We support tracing vLLM workers using different profilers. You can enable profiling by setting the `--profiler-config` flag when launching the server.

docs/design/attention_backends.md

@@ -127,8 +127,8 @@ Priority is **1 = highest** (tried first).
 | 3 | `FLASH_ATTN_MLA` |
 | 4 | `FLASHMLA` |
 | 5 | `TRITON_MLA` |
-| 6 | `FLASHMLA_SPARSE` |
-| 7 | `FLASHINFER_MLA_SPARSE` |
+| 6 | `FLASHINFER_MLA_SPARSE`**\*** |
+| 7 | `FLASHMLA_SPARSE` |
 
 **Ampere/Hopper (SM 8.x-9.x):**
 
@@ -140,6 +140,8 @@ Priority is **1 = highest** (tried first).
 | 4 | `TRITON_MLA` |
 | 5 | `FLASHMLA_SPARSE` |
 
+> **\*** For sparse MLA, FP8 KV cache always prefers `FLASHINFER_MLA_SPARSE`. With BF16 KV cache, `FLASHINFER_MLA_SPARSE` is preferred for low query-head counts (<= 16), while `FLASHMLA_SPARSE` is preferred otherwise.
+>
 > **Note:** ROCm and CPU platforms have their own selection logic. See the platform-specific documentation for details.
 
 ## Legend
@@ -164,18 +166,18 @@ Priority is **1 = highest** (tried first).
 | Backend | Version | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | MM Prefix | DCP | Attention Types | Compute Cap. |
 | ------- | ------- | ------ | --------- | ----------- | ---------- | ---- | --------- | --- | --------------- | ------------ |
 | `CPU_ATTN` | | fp16, bf16, fp32 | `auto` | Any | 32, 64, 80, 96, 112, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | All | N/A |
-| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ❌ | ❌ | ✅ | Decoder | 7.x-9.x |
-| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ✅ | ❌ | ✅ | Decoder | 10.x |
-| `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥8.0 |
-| `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ❌ | ✅ | All | 9.x |
-| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥10.0 |
+| `FLASHINFER` | Native† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ❌ | ❌ | ✅ | Decoder | 7.x-9.x |
+| `FLASHINFER` | TRTLLM† | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32, 64 | 64, 128, 256 | ✅ | ❌ | ✅ | Decoder | 10.x |
+| `FLASH_ATTN` | FA2* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥8.0 |
+| `FLASH_ATTN` | FA3* | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ❌ | ✅ | All | 9.x |
+| `FLASH_ATTN` | FA4* | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ✅ | All | ≥10.0 |
 | `FLASH_ATTN_DIFFKV` | | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ✅ | Decoder | Any |
-| `FLEX_ATTENTION` | | fp16, bf16, fp32 | `auto`, `bfloat16` | Any | Any | ❌ | ✅ | ❌ | Decoder, Encoder Only | Any |
-| `ROCM_AITER_FA` | | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder, Enc-Dec | N/A |
+| `FLEX_ATTENTION` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16` | Any | Any | ❌ | ✅ | ❌ | Decoder, Encoder Only | Any |
+| `ROCM_AITER_FA` | | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 16, 32 | 64, 128, 256 | ❌ | ❌ | ❌ | Decoder, Enc-Dec | N/A |
 | `ROCM_AITER_UNIFIED_ATTN` | | fp16, bf16 | `auto` | %16 | Any | ✅ | ✅ | ❌ | All | N/A |
-| `ROCM_ATTN` | | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ✅ | ✅ | ❌ | All | N/A |
-| `TREE_ATTN` | | fp16, bf16 | `auto` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | Any |
-| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
+| `ROCM_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | 32, 64, 80, 96, 128, 160, 192, 224, 256 | ✅ | ✅ | ❌ | All | N/A |
+| `TREE_ATTN` | | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | 32, 64, 96, 128, 160, 192, 224, 256 | ❌ | ❌ | ❌ | Decoder | Any |
+| `TRITON_ATTN` | | fp16, bf16, fp32 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | %16 | Any | ✅ | ✅ | ❌ | All | Any |
 
 > **†** FlashInfer uses TRTLLM attention on Blackwell (SM100), which supports sinks. Disable via `--attention-config.use_trtllm_attention=0`.
 >
@@ -204,14 +206,14 @@ configuration.
 
 | Backend | Dtypes | KV Dtypes | Block Sizes | Head Sizes | Sink | Sparse | MM Prefix | DCP | Attention Types | Compute Cap. |
 | ------- | ------ | --------- | ----------- | ---------- | ---- | ------ | --------- | --- | --------------- | ------------ |
-| `CUTLASS_MLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 128 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
-| `FLASHINFER_MLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | 10.x |
-| `FLASHINFER_MLA_SPARSE` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 10.x |
-| `FLASHMLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3` | 64 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x-10.x |
+| `CUTLASS_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 128 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 10.x |
+| `FLASHINFER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | 10.x |
+| `FLASHINFER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 32, 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 10.x |
+| `FLASHMLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | 64 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x-10.x |
 | `FLASHMLA_SPARSE` | bf16 | `auto`, `bfloat16`, `fp8_ds_mla` | 64 | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | 9.x-10.x |
-| `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
-| `ROCM_AITER_MLA` | fp16, bf16 | `auto`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 1 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto`, `bfloat16` | 1 | Any | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
+| `FLASH_ATTN_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | 9.x |
+| `ROCM_AITER_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3`, `fp8_e5m2` | 1 | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
+| `ROCM_AITER_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | 1 | Any | ❌ | ✅ | ❌ | ❌ | Decoder | N/A |
 | `ROCM_AITER_TRITON_MLA` | fp16, bf16 | `auto` | Any | Any | ❌ | ❌ | ❌ | ❌ | Decoder | N/A |
-| `TRITON_MLA` | fp16, bf16 | `auto`, `bfloat16` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
-| `XPU_MLA_SPARSE` | fp16, bf16 | `auto`, `bfloat16` | Any | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | Any |
+| `TRITON_MLA` | fp16, bf16 | `auto`, `float16`, `bfloat16`, `fp8`, `fp8_e4m3` | %16 | Any | ❌ | ❌ | ❌ | ✅ | Decoder | Any |
+| `XPU_MLA_SPARSE` | fp16, bf16 | `auto`, `float16`, `bfloat16` | Any | 576 | ❌ | ✅ | ❌ | ❌ | Decoder | Any |

docs/design/custom_op.md

@@ -51,11 +51,8 @@ For example:
 **1. Attention:**
 
 ```python
---8<-- "vllm/model_executor/layers/attention/mm_encoder_attention.py:mm_encoder_attn"
-
 --8<-- "vllm/model_executor/layers/mla.py:multi_head_latent_attention"
 
---8<-- "vllm/model_executor/models/deepencoder.py:rel_pos_attention"
 ```
 
 **2. Activation:**
@@ -170,6 +167,16 @@ For example:
 --8<-- "vllm/model_executor/layers/rotary_embedding/common.py:apply_rotary_emb"
 ```
 
+**12. Encoder:**
+
+```python
+--8<-- "vllm/model_executor/models/deepencoder2.py:qwen2_decoder"
+
+--8<-- "vllm/model_executor/layers/attention/mm_encoder_attention.py:mm_encoder_attn"
+
+--8<-- "vllm/model_executor/models/deepencoder.py:rel_pos_attention"
+```
+
 ## Guidelines for Implementing a New CustomOp
 
 ### Implement a New CustomOp in vLLM