feat: lazy TP weight loading via SafetensorsSlice

When tensor parallelism is active, each rank previously read the full
weight tensor from disk and then called .narrow() to extract its shard,
discarding the rest. This change introduces SafetensorsSlice, a lazy
wrapper around safetensors PySafeSlice that defers disk I/O until
materialization. The .narrow() calls from weight loaders are recorded
without touching the disk, and the actual read at materialize() time
fetches only the needed sub-region.

The __torch_function__ protocol auto-materializes when the slice is
used in any torch operation (e.g. param_data.copy_()), so no downstream
weight loader code needs changes. Unsupported operations (.reshape(),
.view(), .t(), etc.) also auto-materialize as a safety fallback.

Only affects the default "lazy" safetensors loading strategy.

Co-authored-by: Claude

https://claude.ai/code/session_01Ngt6Nm9BtPKAEied3djkuE

.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/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/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 +333,15 @@ 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/test_return_tokens_as_ids.py \
-    --ignore=entrypoints/openai/test_root_path.py \
+    --ignore=entrypoints/openai/chat_completion/test_root_path.py \
     --ignore=entrypoints/openai/test_tokenization.py \
-    --ignore=entrypoints/openai/test_prompt_validation.py "}
+    --ignore=entrypoints/openai/completion/test_prompt_validation.py "}
   fi
 
   if [[ $cmds == *" entrypoints/llm "* ]]; then
@@ -492,6 +504,8 @@ else
     -e HF_TOKEN \
     -e AWS_ACCESS_KEY_ID \
     -e AWS_SECRET_ACCESS_KEY \
+    -e BUILDKITE_PARALLEL_JOB \
+    -e BUILDKITE_PARALLEL_JOB_COUNT \
     -v "${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/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/v1/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 v1/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:
@@ -29,45 +50,48 @@ 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
+- label: Distributed Torchrun + Examples (4 GPUs)
+  timeout_in_minutes: 30
   working_dir: "/vllm-workspace/tests"
   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
+  - tests/distributed/test_torchrun_example.py
+  - tests/distributed/test_torchrun_example_moe.py
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
   - examples/offline_inference/new_weight_syncing/
   - 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
@@ -85,21 +109,6 @@ steps:
   - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
   # test with internal dp
   - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
-  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_async_llm_dp.py
-  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_eagle_dp.py
-  - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/distributed/test_external_lb_dp.py
-  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_internal_lb_dp.py
-  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py
-  - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
-  - pytest -v -s distributed/test_utils.py
-  - pytest -v -s compile/fullgraph/test_basic_correctness.py
-  - pytest -v -s 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
@@ -109,6 +118,47 @@ steps:
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 rlhf_nccl.py
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 python3 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
+  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_internal_lb_dp.py
+  - TP_SIZE=1 DP_SIZE=4 pytest -v -s v1/distributed/test_hybrid_lb_dp.py
+  - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
+  - pytest -v -s distributed/test_utils.py
+
+- 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
+
 - label: Distributed Tests (8 GPUs)(H100)
   timeout_in_minutes: 10
   device: h100

.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
@@ -62,7 +64,7 @@ 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

.buildkite/test_areas/entrypoints.yaml

@@ -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
   - pytest -v -s entrypoints/test_chat_utils.py
   mirror:
     amd:

.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/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

@@ -18,9 +18,9 @@ steps:
   - 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"
 
@@ -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:

.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/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
         ```
@@ -334,7 +334,7 @@ pull_request_rules:
     - or:
       - files~=^tests/tool_use/
       - files~=^tests/entrypoints/openai/tool_parsers/
-      - files=tests/entrypoints/openai/test_chat_with_tool_reasoning.py
+      - files=tests/entrypoints/openai/chat_completion/test_chat_with_tool_reasoning.py
       - files~=^vllm/entrypoints/openai/tool_parsers/
       - files=docs/features/tool_calling.md
       - files~=^examples/tool_chat_*
@@ -381,7 +381,7 @@ 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:

.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

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_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_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/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/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/ops.h

@@ -262,7 +262,8 @@ 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 +296,14 @@ 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/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

@@ -489,8 +489,8 @@ 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 +564,21 @@ 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/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/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/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/fused_moe_modular_kernel.md

@@ -167,9 +167,6 @@ FusedMoEExpertsModular performs the core of the FusedMoE operations. The various
 
 `FusedMoEExpertsModular::activation_formats()`: Return the supported Input and Output activation formats. i.e. Contiguous / Batched format.
 
-`FusedMoEExpertsModular::supports_chunking()`: Return True if the implementation supports chunking. Typically
-implementations that input `FusedMoEActivationFormat.Standard` support chunking and `FusedMoEActivationFormat.BatchedExperts` do not.
-
 `FusedMoEExpertsModular::supports_expert_map()`: Return True if the implementation supports expert map.
 
 `FusedMoEExpertsModular::workspace_shapes()` /
@@ -220,8 +217,8 @@ If you are adding some `FusedMoEPrepareAndFinalizeModular` / `FusedMoEExpertsMod
 
 1. Add the implementation type to `MK_ALL_PREPARE_FINALIZE_TYPES` and `MK_FUSED_EXPERT_TYPES` in [mk_objects.py](../../tests/kernels/moe/modular_kernel_tools/mk_objects.py) respectively.
 2. Update `Config::is_batched_prepare_finalize()`, `Config::is_batched_fused_experts()`, `Config::is_standard_fused_experts()`,
-`Config::is_fe_16bit_supported()`,  `Config::is_fe_fp8_supported()`, `Config::is_fe_block_fp8_supported()`,
-`Config::is_fe_supports_chunking()` methods in [/tests/kernels/moe/modular_kernel_tools/common.py](../../tests/kernels/moe/modular_kernel_tools/common.py)
+`Config::is_fe_16bit_supported()`,  `Config::is_fe_fp8_supported()`, `Config::is_fe_block_fp8_supported()`
+methods in [/tests/kernels/moe/modular_kernel_tools/common.py](../../tests/kernels/moe/modular_kernel_tools/common.py)
 
 Doing this will add the new implementation to the test suite.
 

docs/design/moe_kernel_features.md

@@ -35,7 +35,8 @@ th {
 | naive | standard | all<sup>1</sup> | G,A,T | N | <sup>6</sup> | [layer.py][vllm.model_executor.layers.fused_moe.layer.FusedMoE] |
 | deepep_high_throughput | standard | fp8 | G(128),A,T<sup>2</sup> | Y | Y | [`DeepEPHTPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ht_prepare_finalize.DeepEPHTPrepareAndFinalize] |
 | deepep_low_latency | batched | fp8 | G(128),A,T<sup>3</sup> | Y | Y | [`DeepEPLLPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.deepep_ll_prepare_finalize.DeepEPLLPrepareAndFinalize] |
-| flashinfer_all2allv | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferA2APrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_a2a_prepare_finalize.FlashInferA2APrepareAndFinalize] |
+| flashinfer_nvlink_two_sided | standard | nvfp4,fp8 | G,A,T | N | N | [`FlashInferNVLinkTwoSidedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_nvlink_two_sided_prepare_finalize.FlashInferNVLinkTwoSidedPrepareAndFinalize] |
+| flashinfer_nvlink_one_sided | standard | nvfp4 | G,A,T | N | N | [`FlashInferNVLinkOneSidedPrepareAndFinalize`][vllm.model_executor.layers.fused_moe.flashinfer_nvlink_one_sided_prepare_finalize.FlashInferNVLinkOneSidedPrepareAndFinalize] |
 
 !!! info "Table key"
     1. All types: mxfp4, nvfp4, int4, int8, fp8

docs/design/torch_compile_multimodal.md

@@ -34,9 +34,6 @@ relies on caching artifacts to reduce start time, we must properly propagate the
 with the LLM text-backbone, or other instances of the same artifact (as is the case with vision block). `is_encoder=True` is also needed for encoder
 components (see Compile Range Integration).
 
-3. `with set_forward_context` context manager should be used around the nn.Module's forward call. This will properly forward the vllm_config which is needed
-for torch.compile integration.
-
 ### CompilationConfig
 
 With the exception of `compile_mm_encoder: true`, the multimodal encoder will inherit from the same compilation config as the text LLM. We may extend

docs/features/lora.md

@@ -389,3 +389,17 @@ vllm serve model --enable-lora --max-lora-rank 64
 # Bad: unnecessarily high, wastes memory
 vllm serve model --enable-lora --max-lora-rank 256
 ```
+
+### Restricting LoRA to Specific Modules
+
+The `--lora-target-modules` parameter allows you to restrict which model modules have LoRA applied at deployment time. This is useful for performance tuning when you only need LoRA on specific layers:
+
+```bash
+# Apply LoRA only to output projection layers
+vllm serve model --enable-lora --lora-target-modules o_proj
+
+# Apply LoRA to multiple specific modules
+vllm serve model --enable-lora --lora-target-modules o_proj qkv_proj down_proj
+```
+
+When `--lora-target-modules` is not specified, LoRA will be applied to all supported modules in the model. This parameter accepts module suffixes (the last component of the module name), such as `o_proj`, `qkv_proj`, `gate_proj`, etc.

docs/features/tool_calling.md

@@ -107,6 +107,27 @@ vLLM supports the `tool_choice='none'` option in the chat completion API. When t
 !!! note
     When tools are specified in the request, vLLM includes tool definitions in the prompt by default, regardless of the `tool_choice` setting. To exclude tool definitions when `tool_choice='none'`, use the `--exclude-tools-when-tool-choice-none` option.
 
+## Constrained Decoding Behavior
+
+Whether vLLM enforces the tool parameter schema during generation depends on the `tool_choice` mode:
+
+| `tool_choice` value | Schema-constrained decoding | Behavior |
+| --- | --- | --- |
+| Named function | Yes (via structured outputs backend) | Arguments are guaranteed to be valid JSON conforming to the function's parameter schema. |
+| `"required"` | Yes (via structured outputs backend) | Same as named function. The model must produce at least one tool call. |
+| `"auto"` | No | The model generates freely. A tool-call parser extracts tool calls from the raw text. Arguments may be malformed or not match the schema. |
+| `"none"` | N/A | No tool calls are produced. |
+
+When schema conformance matters, prefer `tool_choice="required"` or named function calling over `"auto"`.
+
+### Strict Mode (`strict` parameter)
+
+The [OpenAI API](https://platform.openai.com/docs/guides/function-calling#strict-mode) supports a `strict` field on function definitions. When set to `true`, OpenAI uses constrained decoding to guarantee that tool-call arguments match the function schema, even in `tool_choice="auto"` mode.
+
+vLLM **does not implement** `strict` mode today. The `strict` field is accepted in requests (to avoid breaking clients that set it), but it has no effect on decoding behavior. In auto mode, argument validity depends entirely on the model's output quality and the parser's extraction logic.
+
+Tracking issues: [#15526](https://github.com/vllm-project/vllm/issues/15526), [#16313](https://github.com/vllm-project/vllm/issues/16313).
+
 ## Automatic Function Calling
 
 To enable this feature, you should set the following flags:
@@ -124,6 +145,9 @@ from HuggingFace; and you can find an example of this in a `tokenizer_config.jso
 
 If your favorite tool-calling model is not supported, please feel free to contribute a parser & tool use chat template!
 
+!!! note
+    With `tool_choice="auto"`, tool-call arguments are extracted from the model's raw text output by the selected parser. No schema-level constraint is applied during decoding, so arguments may occasionally be malformed or violate the function's parameter schema. See [Constrained Decoding Behavior](#constrained-decoding-behavior) for details.
+
 ### Hermes Models (`hermes`)
 
 All Nous Research Hermes-series models newer than Hermes 2 Pro should be supported.
@@ -219,7 +243,7 @@ Supported models:
 
 * `ibm-granite/granite-4.0-h-small` and other Granite 4.0 models
 
-    Recommended flags: `--tool-call-parser hermes`
+    Recommended flags: `--tool-call-parser granite4`
 
 * `ibm-granite/granite-3.0-8b-instruct`
 

docs/getting_started/installation/README.md

@@ -16,4 +16,6 @@ vLLM supports the following hardware platforms:
 
 vLLM supports third-party hardware plugins that live **outside** the main `vllm` repository. These follow the [Hardware-Pluggable RFC](../../design/plugin_system.md).
 
-A list of all supported hardware can be found on the [vllm.ai website](https://vllm.ai/#compatibility). If you want to add new hardware, please contact us on [Slack](https://slack.vllm.ai/) or [Email](mailto:collaboration@vllm.ai).
+A list of all supported hardware can be found on the vLLM website, see [Universal Compatibility - Hardware](https://vllm.ai/#compatibility).
+
+If you want to add new hardware, please contact us on [Slack](https://slack.vllm.ai/) or [Email](mailto:collaboration@vllm.ai).

docs/getting_started/installation/cpu.x86.inc.md

@@ -7,7 +7,7 @@ vLLM supports basic model inferencing and serving on x86 CPU platform, with data
 --8<-- [start:requirements]
 
 - OS: Linux
-- CPU flags: `avx512f` (Recommended), `avx512_bf16` (Optional), `avx512_vnni` (Optional)
+- CPU flags: `avx512f` (Recommended), `avx2` (Limited features)
 
 !!! tip
     Use `lscpu` to check the CPU flags.
@@ -18,7 +18,7 @@ vLLM supports basic model inferencing and serving on x86 CPU platform, with data
 --8<-- [end:set-up-using-python]
 --8<-- [start:pre-built-wheels]
 
-Pre-built vLLM wheels for x86 with AVX512 are available since version 0.13.0. To install release wheels:
+Pre-built vLLM wheels for x86 with AVX512/AVX2 are available since version 0.17.0. To install release wheels:
 
 ```bash
 export VLLM_VERSION=$(curl -s https://api.github.com/repos/vllm-project/vllm/releases/latest | jq -r .tag_name | sed 's/^v//')
@@ -108,13 +108,13 @@ VLLM_TARGET_DEVICE=cpu uv pip install . --no-build-isolation
 If you want to develop vLLM, install it in editable mode instead.
 
 ```bash
-VLLM_TARGET_DEVICE=cpu uv pip install -e . --no-build-isolation
+VLLM_TARGET_DEVICE=cpu python3 setup.py develop
 ```
 
 Optionally, build a portable wheel which you can then install elsewhere:
 
 ```bash
-VLLM_TARGET_DEVICE=cpu uv build --wheel
+VLLM_TARGET_DEVICE=cpu uv build --wheel --no-build-isolation
 ```
 
 ```bash
@@ -185,12 +185,9 @@ docker run \
     -v ~/.cache/huggingface:/root/.cache/huggingface \
     -p 8000:8000 \
     --env "HF_TOKEN=<secret>" \
-vllm/vllm-openai-cpu:latest-x86_64 <args...>
+    vllm/vllm-openai-cpu:latest-x86_64 <args...>
 ```
 
-!!! warning
-    If deploying the pre-built images on machines without `avx512f`, `avx512_bf16`, or `avx512_vnni` support, an `Illegal instruction` error may be raised. See the build-image-from-source section below for build arguments to match your target CPU capabilities.
-
 --8<-- [end:pre-built-images]
 --8<-- [start:build-image-from-source]
 
@@ -198,50 +195,11 @@ vllm/vllm-openai-cpu:latest-x86_64 <args...>
 
 ```bash
 docker build -f docker/Dockerfile.cpu \
-        --build-arg VLLM_CPU_DISABLE_AVX512=<false (default)|true> \
-        --build-arg VLLM_CPU_AVX2=<false (default)|true> \
-        --build-arg VLLM_CPU_AVX512=<false (default)|true> \
-        --build-arg VLLM_CPU_AVX512BF16=<false (default)|true> \
-        --build-arg VLLM_CPU_AVX512VNNI=<false (default)|true> \
-        --build-arg VLLM_CPU_AMXBF16=<false|true (default)> \
+        --build-arg VLLM_CPU_X86=<false (default)|true> \ # For cross-compilation
         --tag vllm-cpu-env \
         --target vllm-openai .
 ```
 
-!!! note "Auto-detection by default"
-    By default, CPU instruction sets (AVX512, AVX2, etc.) are automatically detected from the build system's CPU flags. Build arguments like `VLLM_CPU_AVX2`, `VLLM_CPU_AVX512`, `VLLM_CPU_AVX512BF16`, `VLLM_CPU_AVX512VNNI`, and `VLLM_CPU_AMXBF16` are used for cross-compilation:
-
-    - `VLLM_CPU_{ISA}=true` - Force-enable the instruction set (build with ISA regardless of build system capabilities)
-    - `VLLM_CPU_{ISA}=false` - Rely on auto-detection (default)
-
-##### Examples
-
-###### Auto-detection build (default)
-
-```bash
-docker build -f docker/Dockerfile.cpu --tag vllm-cpu-env --target vllm-openai .
-```
-
-###### Cross-compile for AVX512
-
-```bash
-docker build -f docker/Dockerfile.cpu \
-        --build-arg VLLM_CPU_AVX512=true \
-        --build-arg VLLM_CPU_AVX512BF16=true \
-        --build-arg VLLM_CPU_AVX512VNNI=true \
-        --tag vllm-cpu-avx512 \
-        --target vllm-openai .
-```
-
-###### Cross-compile for AVX2
-
-```bash
-docker build -f docker/Dockerfile.cpu \
-        --build-arg VLLM_CPU_AVX2=true \
-        --tag vllm-cpu-avx2 \
-        --target vllm-openai .
-```
-
 #### Launching the OpenAI server
 
 ```bash

docs/governance/process.md

@@ -135,6 +135,19 @@ PRs requires at least one committer review and approval. If the code is covered
 
 In case where CI didn't pass due to the failure is not related to the PR, the PR can be merged by the lead maintainers using "force merge" option that overrides the CI checks.
 
+### AI Assisted Contributions
+
+AI tools can accelerate development, but contributors remain fully responsible for all code they submit. Like the Developer Certificate of Origin, this policy centers on accountability: contributors must believe they have the right to submit their contribution under vLLM's open source license, regardless of how the code was created.
+
+All AI-assisted contributions must meet the same quality, testing, and review standards as any other code. Contributors must review and understand AI-generated code before submission—just make sure it is good code:
+
+- 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.
+- Attribution preserves legal clarity and community trust. Contributors must disclose AI assistance in pull requests and mark commits with appropriate trailers (e.g. `Co-authored-by:`).
+- Avoid one-off "busywork" PRs (single typo, isolated style cleanup, one mutable default fix, etc.). Bundle mechanical cleanups into a clear, systematic scope.
+
+!!! warning
+    These topics are outlined for agents in [AGENTS.md](../../AGENTS.md) with instructions for how to autonomously implement them.
+
 ### Slack
 
 Contributors are encouraged to join `#pr-reviews` and `#contributors` channels.

docs/models/extensions/instanttensor.md

@@ -0,0 +1,31 @@
+# Loading Model Weights with InstantTensor
+
+InstantTensor accelerates loading Safetensors weights on CUDA devices through distributed loading, pipelined prefetching, and direct I/O. InstantTensor also supports GDS (GPUDirect Storage) when available.
+For more details, see the [InstantTensor GitHub repository](https://github.com/scitix/InstantTensor).
+
+## Installation
+
+```bash
+pip install instanttensor
+```
+
+## Use InstantTensor in vLLM
+
+Add `--load-format instanttensor` as a command-line argument.
+
+For example:
+
+```bash
+vllm serve Qwen/Qwen2.5-0.5B --load-format instanttensor
+```
+
+## Benchmarks
+
+| Model | GPU | Backend | Load Time (s) | Throughput (GB/s) | Speedup |
+| --- | ---: | --- | ---: | ---: | --- |
+| Qwen3-30B-A3B | 1*H200 | Safetensors | 57.4 | 1.1 | 1x |
+| Qwen3-30B-A3B | 1*H200 | InstantTensor | 1.77 | 35 | <span style="color: green">**32.4x**</span> |
+| DeepSeek-R1 | 8*H200 | Safetensors | 160 | 4.3 | 1x |
+| DeepSeek-R1 | 8*H200 | InstantTensor | 15.3 | 45 | <span style="color: green">**10.5x**</span> |
+
+For the full benchmark results, see <https://github.com/scitix/InstantTensor/blob/main/docs/benchmark.md>.

docs/models/extensions/runai_model_streamer.md

@@ -31,6 +31,16 @@ vllm serve gs://core-llm/Llama-3-8b \
     --load-format runai_streamer
 ```
 
+To run model from Azure Blob Storage run:
+
+```bash
+AZURE_STORAGE_ACCOUNT_NAME=<account> \
+vllm serve az://<container>/<model-path> \
+    --load-format runai_streamer
+```
+
+Authentication uses `DefaultAzureCredential`, which supports `az login`, managed identity, environment variables (`AZURE_CLIENT_ID`, `AZURE_TENANT_ID`, `AZURE_CLIENT_SECRET`), and other methods.
+
 To run model from a S3 compatible object store run:
 
 ```bash

docs/models/pooling_models.md

@@ -625,6 +625,46 @@ curl -s http://localhost:8000/rerank -H "Content-Type: application/json" -d '{
 }'
 ```
 
+### ColQwen3.5 Multi-Modal Late Interaction Models
+
+ColQwen3.5 is based on [ColPali](https://arxiv.org/abs/2407.01449), extending ColBERT's late interaction approach to **multi-modal** inputs. It uses the Qwen3.5 hybrid backbone (linear + full attention) and produces per-token L2-normalized vectors for MaxSim scoring.
+
+| Architecture | Backbone | Example HF Models |
+| - | - | - |
+| `ColQwen3_5` | Qwen3.5 | `athrael-soju/colqwen3.5-4.5B` |
+
+Start the server:
+
+```shell
+vllm serve athrael-soju/colqwen3.5-4.5B --max-model-len 4096
+```
+
+Then you can use the rerank endpoint:
+
+```shell
+curl -s http://localhost:8000/rerank -H "Content-Type: application/json" -d '{
+    "model": "athrael-soju/colqwen3.5-4.5B",
+    "query": "What is machine learning?",
+    "documents": [
+        "Machine learning is a subset of artificial intelligence.",
+        "Python is a programming language.",
+        "Deep learning uses neural networks."
+    ]
+}'
+```
+
+Or the score endpoint:
+
+```shell
+curl -s http://localhost:8000/score -H "Content-Type: application/json" -d '{
+    "model": "athrael-soju/colqwen3.5-4.5B",
+    "text_1": "What is the capital of France?",
+    "text_2": ["The capital of France is Paris.", "Python is a programming language."]
+}'
+```
+
+An example can be found here: [examples/pooling/score/colqwen3_5_rerank_online.py](../../examples/pooling/score/colqwen3_5_rerank_online.py)
+
 ### BAAI/bge-m3
 
 The `BAAI/bge-m3` model comes with extra weights for sparse and colbert embeddings but unfortunately in its `config.json`

docs/models/supported_models.md

@@ -418,6 +418,7 @@ th {
 | `Grok1ForCausalLM` | Grok2 | `xai-org/grok-2` | ✅︎ | ✅︎ |
 | `HunYuanDenseV1ForCausalLM` | Hunyuan Dense | `tencent/Hunyuan-7B-Instruct` | ✅︎ | ✅︎ |
 | `HunYuanMoEV1ForCausalLM` | Hunyuan-A13B | `tencent/Hunyuan-A13B-Instruct`, `tencent/Hunyuan-A13B-Pretrain`, `tencent/Hunyuan-A13B-Instruct-FP8`, etc. | ✅︎ | ✅︎ |
+| `HyperCLOVAXForCausalLM` | HyperCLOVAX-SEED-Think-14B | `naver-hyperclovax/HyperCLOVAX-SEED-Think-14B` | ✅︎ | ✅︎ |
 | `InternLMForCausalLM` | InternLM | `internlm/internlm-7b`, `internlm/internlm-chat-7b`, etc. | ✅︎ | ✅︎ |
 | `InternLM2ForCausalLM` | InternLM2 | `internlm/internlm2-7b`, `internlm/internlm2-chat-7b`, etc. | ✅︎ | ✅︎ |
 | `InternLM3ForCausalLM` | InternLM3 | `internlm/internlm3-8b-instruct`, etc. | ✅︎ | ✅︎ |
@@ -514,6 +515,7 @@ These models primarily support the [`LLM.embed`](./pooling_models.md#llmembed) A
 | ------------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `BertModel`<sup>C</sup> | BERT-based | `BAAI/bge-base-en-v1.5`, `Snowflake/snowflake-arctic-embed-xs`, etc. | | |
 | `BertSpladeSparseEmbeddingModel` | SPLADE | `naver/splade-v3` | | |
+| `ErnieModel` | BERT-like Chinese ERNIE | `shibing624/text2vec-base-chinese-sentence` | | |
 | `Gemma2Model`<sup>C</sup> | Gemma 2-based | `BAAI/bge-multilingual-gemma2`, etc. | ✅︎ | ✅︎ |
 | `Gemma3TextModel`<sup>C</sup> | Gemma 3-based | `google/embeddinggemma-300m`, etc. | ✅︎ | ✅︎ |
 | `GritLM` | GritLM | `parasail-ai/GritLM-7B-vllm`. | ✅︎ | ✅︎ |
@@ -556,8 +558,9 @@ These models primarily support the [`LLM.classify`](./pooling_models.md#llmclass
 
 | Architecture | Models | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
 | ------------ | ------ | ----------------- | -------------------- | ------------------------- |
-| `JambaForSequenceClassification` | Jamba | `ai21labs/Jamba-tiny-reward-dev`, etc. | ✅︎ | ✅︎ |
+| `ErnieForSequenceClassification` | BERT-like Chinese ERNIE | `Forrest20231206/ernie-3.0-base-zh-cls` | | |
 | `GPT2ForSequenceClassification` | GPT2 | `nie3e/sentiment-polish-gpt2-small` | | |
+| `JambaForSequenceClassification` | Jamba | `ai21labs/Jamba-tiny-reward-dev`, etc. | ✅︎ | ✅︎ |
 | `*Model`<sup>C</sup>, `*ForCausalLM`<sup>C</sup>, etc. | Generative models | N/A | \* | \* |
 
 <sup>C</sup> Automatically converted into a classification model via `--convert classify`. ([details](./pooling_models.md#model-conversion))  
@@ -574,6 +577,7 @@ These models primarily support the [`LLM.score`](./pooling_models.md#llmscore) A
 | Architecture | Models | Example HF Models | Score template (see note) | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
 | ------------ | ------ | ----------------- | ------------------------- | --------------------------- | --------------------------------------- |
 | `BertForSequenceClassification` | BERT-based | `cross-encoder/ms-marco-MiniLM-L-6-v2`, etc. | N/A | | |
+| `ErnieForSequenceClassification` | BERT-like Chinese ERNIE | `Forrest20231206/ernie-3.0-base-zh-cls` | N/A | | |
 | `GemmaForSequenceClassification` | Gemma-based | `BAAI/bge-reranker-v2-gemma`(see note), etc. | [bge-reranker-v2-gemma.jinja](../../examples/pooling/score/template/bge-reranker-v2-gemma.jinja) | ✅︎ | ✅︎ |
 | `GteNewForSequenceClassification` | mGTE-TRM (see note) | `Alibaba-NLP/gte-multilingual-reranker-base`, etc. | N/A | | |
 | `LlamaBidirectionalForSequenceClassification`<sup>C</sup> | Llama-based with bidirectional attention | `nvidia/llama-nemotron-rerank-1b-v2`, etc. | [nemotron-rerank.jinja](../../examples/pooling/score/template/nemotron-rerank.jinja) | ✅︎ | ✅︎ |
@@ -639,6 +643,7 @@ These models primarily support the [`LLM.encode`](./pooling_models.md#llmencode)
 | Architecture | Models | Example HF Models | [LoRA](../features/lora.md) | [PP](../serving/parallelism_scaling.md) |
 | ------------ | ------ | ----------------- | --------------------------- | --------------------------------------- |
 | `BertForTokenClassification` | bert-based | `boltuix/NeuroBERT-NER` (see note), etc. | | |
+| `ErnieForTokenClassification` | BERT-like Chinese ERNIE | `gyr66/Ernie-3.0-base-chinese-finetuned-ner` | | |
 | `ModernBertForTokenClassification` | ModernBERT-based | `disham993/electrical-ner-ModernBERT-base` | | |
 
 !!! note
@@ -828,6 +833,8 @@ The following table lists those that are tested in vLLM.
 | ------------ | ------ | ------ | ----------------- | -------------------- | ------------------------- |
 | `CLIPModel` | CLIP | T / I | `openai/clip-vit-base-patch32`, `openai/clip-vit-large-patch14`, etc. | | |
 | `ColModernVBertForRetrieval` | ColModernVBERT | T / I | `ModernVBERT/colmodernvbert-merged` | | |
+| `ColPaliForRetrieval` | ColPali | T / I | `vidore/colpali-v1.3-hf` | | |
+| `ColQwen3_5` | ColQwen3.5 | T + I + V | `athrael-soju/colqwen3.5-4.5B-v3` | | |
 | `LlamaNemotronVLModel` | Llama Nemotron Embedding + SigLIP | T + I | `nvidia/llama-nemotron-embed-vl-1b-v2` | | |
 | `LlavaNextForConditionalGeneration`<sup>C</sup> | LLaVA-NeXT-based | T / I | `royokong/e5-v` | | ✅︎ |
 | `Phi3VForCausalLM`<sup>C</sup> | Phi-3-Vision-based | T + I | `TIGER-Lab/VLM2Vec-Full` | | ✅︎ |

docs/serving/expert_parallel_deployment.md

@@ -21,7 +21,8 @@ vLLM provides multiple communication backends for EP. Use `--all2all-backend` to
 | `allgather_reducescatter` | Default backend | Standard all2all using allgather/reducescatter primitives | General purpose, works with any EP+DP configuration |
 | `deepep_high_throughput` | Multi-node prefill | Grouped GEMM with continuous layout, optimized for prefill | Prefill-dominated workloads, high-throughput scenarios |
 | `deepep_low_latency` | Multi-node decode | CUDA graph support, masked layout, optimized for decode | Decode-dominated workloads, low-latency scenarios |
-| `flashinfer_all2allv` | MNNVL systems | FlashInfer alltoallv kernels for multi-node NVLink | Systems with NVLink across nodes |
+| `flashinfer_nvlink_one_sided` | MNNVL systems | FlashInfer's one-sided A2A strategy for multi-node NVLink | High-throughput workloads |
+| `flashinfer_nvlink_two_sided` | MNNVL systems | FlashInfer's two-sided A2A strategy for multi-node NVLink | Systems with NVLink across nodes |
 | `naive` | Testing/debugging | Simple broadcast-based implementation | Debugging, not recommended for production |
 
 ## Single Node Deployment

docs/serving/openai_compatible_server.md

@@ -72,6 +72,9 @@ In addition, we have the following custom APIs:
     - Only applicable to [classification models](../models/pooling_models.md).
 - [Score API](#score-api) (`/score`)
     - Applicable to [embedding models and cross-encoder models](../models/pooling_models.md).
+- [Cohere Embed API](#cohere-embed-api) (`/v2/embed`)
+    - Compatible with [Cohere's Embed API](https://docs.cohere.com/reference/embed)
+    - Works with any [embedding model](../models/pooling_models.md), including multimodal models.
 - [Re-rank API](#re-rank-api) (`/rerank`, `/v1/rerank`, `/v2/rerank`)
     - Implements [Jina AI's v1 re-rank API](https://jina.ai/reranker/)
     - Also compatible with [Cohere's v1 & v2 re-rank APIs](https://docs.cohere.com/v2/reference/rerank)
@@ -429,6 +432,137 @@ these extra parameters are supported instead:
     --8<-- "vllm/entrypoints/pooling/base/protocol.py:embed-extra-params"
     ```
 
+### Cohere Embed API
+
+Our API is also compatible with [Cohere's Embed v2 API](https://docs.cohere.com/reference/embed) which adds support for some modern embedding feature such as truncation, output dimensions, embedding types, and input types. This endpoint works with any embedding model (including multimodal models).
+
+#### Cohere Embed API request parameters
+
+| Parameter | Type | Required | Description |
+| --------- | ---- | -------- | ----------- |
+| `model` | string | Yes | Model name |
+| `input_type` | string | No | Prompt prefix key (model-dependent, see below) |
+| `texts` | list[string] | No | Text inputs (use one of `texts`, `images`, or `inputs`) |
+| `images` | list[string] | No | Base64 data URI images |
+| `inputs` | list[object] | No | Mixed text and image content objects |
+| `embedding_types` | list[string] | No | Output types (default: `["float"]`) |
+| `output_dimension` | int | No | Truncate embeddings to this dimension (Matryoshka) |
+| `truncate` | string | No | `END`, `START`, or `NONE` (default: `END`) |
+
+#### Text embedding
+
+```bash
+curl -X POST "http://localhost:8000/v2/embed" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "Snowflake/snowflake-arctic-embed-m-v1.5",
+    "input_type": "query",
+    "texts": ["Hello world", "How are you?"],
+    "embedding_types": ["float"]
+  }'
+```
+
+??? console "Response"
+
+    ```json
+    {
+      "id": "embd-...",
+      "embeddings": {
+        "float": [
+          [0.012, -0.034, ...],
+          [0.056, 0.078, ...]
+        ]
+      },
+      "texts": ["Hello world", "How are you?"],
+      "meta": {
+        "api_version": {"version": "2"},
+        "billed_units": {"input_tokens": 12}
+      }
+    }
+    ```
+
+#### Mixed text and image inputs
+
+For multimodal models, you can embed images by passing base64 data URIs. The `inputs` field accepts a list of objects with mixed text and image content:
+
+```bash
+curl -X POST "http://localhost:8000/v2/embed" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "google/siglip-so400m-patch14-384",
+    "inputs": [
+      {
+        "content": [
+          {"type": "text", "text": "A photo of a cat"},
+          {"type": "image_url", "image_url": {"url": "data:image/png;base64,iVBOR..."}}
+        ]
+      }
+    ],
+    "embedding_types": ["float"]
+  }'
+```
+
+#### Embedding types
+
+The `embedding_types` parameter controls the output format. Multiple types can be requested in a single call:
+
+| Type | Description |
+| ---- | ----------- |
+| `float` | Raw float32 embeddings (default) |
+| `binary` | Bit-packed signed binary |
+| `ubinary` | Bit-packed unsigned binary |
+| `base64` | Little-endian float32 encoded as base64 |
+
+```bash
+curl -X POST "http://localhost:8000/v2/embed" \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "Snowflake/snowflake-arctic-embed-m-v1.5",
+    "input_type": "query",
+    "texts": ["What is machine learning?"],
+    "embedding_types": ["float", "binary"]
+  }'
+```
+
+??? console "Response"
+
+    ```json
+    {
+      "id": "embd-...",
+      "embeddings": {
+        "float": [[0.012, -0.034, ...]],
+        "binary": [[42, -117, ...]]
+      },
+      "texts": ["What is machine learning?"],
+      "meta": {
+        "api_version": {"version": "2"},
+        "billed_units": {"input_tokens": 8}
+      }
+    }
+    ```
+
+#### Truncation
+
+The `truncate` parameter controls how inputs exceeding the model's maximum sequence length are handled:
+
+| Value | Behavior |
+| ----- | --------- |
+| `END` (default) | Keep the first tokens, drop the end |
+| `START` | Keep the last tokens, drop the beginning |
+| `NONE` | Return an error if the input is too long |
+
+#### Input type and prompt prefixes
+
+The `input_type` field selects a prompt prefix to prepend to each text input. The available values
+depend on the model:
+
+- **Models with `task_instructions` in `config.json`**: The keys from the `task_instructions` dict are
+  the valid `input_type` values and the corresponding value is prepended to each text.
+- **Models with `config_sentence_transformers.json` prompts**: The keys from the `prompts` dict are
+  the valid `input_type` values. For example, `Snowflake/snowflake-arctic-embed-xs` defines `"query"`,
+  so setting `input_type: "query"` prepends `"Represent this sentence for searching relevant passages: "`.
+- **Other models**: `input_type` is not accepted and will raise a validation error if passed.
+
 ### Transcriptions API
 
 Our Transcriptions API is compatible with [OpenAI's Transcriptions API](https://platform.openai.com/docs/api-reference/audio/createTranscription);

docs/usage/troubleshooting.md

@@ -91,8 +91,8 @@ If GPU/CPU communication cannot be established, you can use the following Python
     import torch
     import torch.distributed as dist
     dist.init_process_group(backend="nccl")
-    local_rank = dist.get_rank() % torch.cuda.device_count()
-    torch.cuda.set_device(local_rank)
+    local_rank = dist.get_rank() % torch.accelerator.device_count()
+    torch.accelerator.set_device_index(local_rank)
     data = torch.FloatTensor([1,] * 128).to("cuda")
     dist.all_reduce(data, op=dist.ReduceOp.SUM)
     torch.accelerator.synchronize()
@@ -337,7 +337,7 @@ import vllm
 import torch
 
 print(f"CUDA available: {torch.cuda.is_available()}")
-print(f"CUDA device count: {torch.cuda.device_count()}")
+print(f"CUDA device count: {torch.accelerator.device_count()}")
 EOF
 ```
 

examples/offline_inference/audio_language.py

@@ -70,6 +70,29 @@ def run_audioflamingo3(question: str, audio_count: int) -> ModelRequestData:
     )
 
 
+# CohereASR
+def run_cohere_asr(question: str, audio_count: int) -> ModelRequestData:
+    assert audio_count == 1, "CohereASR only support single audio input per prompt"
+    # TODO (ekagra): add HF ckpt after asr release
+    model_name = "/host/engines/vllm/audio/2b-release"
+
+    prompt = (
+        "<|startofcontext|><|startoftranscript|>"
+        "<|emo:undefined|><|en|><|en|><|pnc|><|noitn|>"
+        "<|notimestamp|><|nodiarize|>"
+    )
+    engine_args = EngineArgs(
+        model=model_name,
+        limit_mm_per_prompt={"audio": audio_count},
+        trust_remote_code=True,
+    )
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+    )
+
+
 # MusicFlamingo
 def run_musicflamingo(question: str, audio_count: int) -> ModelRequestData:
     model_name = "nvidia/music-flamingo-2601-hf"
@@ -508,14 +531,15 @@ def run_whisper(question: str, audio_count: int) -> ModelRequestData:
 
 model_example_map = {
     "audioflamingo3": run_audioflamingo3,
-    "musicflamingo": run_musicflamingo,
+    "cohere_asr": run_cohere_asr,
+    "funaudiochat": run_funaudiochat,
     "gemma3n": run_gemma3n,
     "glmasr": run_glmasr,
-    "funaudiochat": run_funaudiochat,
     "granite_speech": run_granite_speech,
     "kimi_audio": run_kimi_audio,
     "midashenglm": run_midashenglm,
     "minicpmo": run_minicpmo,
+    "musicflamingo": run_musicflamingo,
     "phi4_mm": run_phi4mm,
     "qwen2_audio": run_qwen2_audio,
     "qwen2_5_omni": run_qwen2_5_omni,

examples/online_serving/disaggregated_serving/moriio_toy_proxy_server.py

@@ -14,6 +14,10 @@ import regex as re
 import zmq
 from quart import Quart, make_response, request
 
+from vllm.distributed.kv_transfer.kv_connector.v1.moriio.moriio_common import (
+    MoRIIOConstants,
+)
+
 logger = logging.getLogger(__name__)
 logger.setLevel(logging.DEBUG)
 prefill_instances: list[dict] = []
@@ -213,6 +217,8 @@ async def handle_request():
 
         dip, dport = extract_ip_port_fast(decode_instance_endpoint["request_address"])
 
+        transfer_id = f"{MoRIIOConstants.TRANSFER_PREFIX}-{str(uuid.uuid4())}"
+
         req_data_to_prefill = copy.deepcopy(req_data)
         req_data_to_prefill["kv_transfer_params"] = {}
         req_data["kv_transfer_params"] = {}
@@ -222,6 +228,7 @@ async def handle_request():
         req_data_to_prefill["kv_transfer_params"]["remote_tp_size"] = (
             decode_instance_endpoint["tp_size"]
         )
+        req_data_to_prefill["kv_transfer_params"]["transfer_id"] = transfer_id
 
         send_prefill_task = asyncio.create_task(
             send_request_to_prefill(
@@ -267,6 +274,7 @@ async def handle_request():
 
         if selected_prefill_dp_rank is not None:
             req_data["kv_transfer_params"]["remote_dp_rank"] = selected_prefill_dp_rank
+        req_data["kv_transfer_params"]["transfer_id"] = transfer_id
 
         decode_request_task = asyncio.create_task(
             start_decode_request(

examples/online_serving/new_weight_syncing/rlhf_http_ipc.py

@@ -106,7 +106,7 @@ def main():
     # IPC requires the training model to be on the same GPU as the vLLM server
     # The server should be started on GPU 0 with reduced memory utilization
     device = "cuda:0"
-    torch.cuda.set_device(device)
+    torch.accelerator.set_device_index(device)
 
     # Load the training model on the same GPU as the server
     # Use bfloat16 to reduce memory footprint

examples/online_serving/new_weight_syncing/rlhf_http_nccl.py

@@ -131,7 +131,7 @@ def main():
     inference_world_size = get_world_size(BASE_URL)
     world_size = inference_world_size + 1  # +1 for the trainer
     device = f"cuda:{inference_world_size}"
-    torch.cuda.set_device(device)
+    torch.accelerator.set_device_index(device)
 
     # Load the training model
     print(f"Loading training model: {MODEL_NAME}")

examples/online_serving/openai_chat_completion_client_for_multimodal.py

@@ -20,9 +20,9 @@ run the script with
 python openai_chat_completion_client_for_multimodal.py --chat-type audio
 """
 
-import base64
 import os
 
+import pybase64 as base64
 import requests
 from openai import OpenAI
 from utils import get_first_model

examples/online_serving/openai_realtime_client.py

@@ -24,11 +24,11 @@ The script:
 
 import argparse
 import asyncio
-import base64
 import json
 
 import librosa
 import numpy as np
+import pybase64 as base64
 import websockets
 
 from vllm.assets.audio import AudioAsset

examples/online_serving/openai_realtime_microphone_client.py

@@ -18,13 +18,13 @@ Requirements: websockets, numpy, gradio
 
 import argparse
 import asyncio
-import base64
 import json
 import queue
 import threading
 
 import gradio as gr
 import numpy as np
+import pybase64 as base64
 import websockets
 
 SAMPLE_RATE = 16_000

examples/pooling/embed/embedding_requests_base64_online.py

@@ -7,8 +7,8 @@ NOTE:
 """
 
 import argparse
-import base64
 
+import pybase64 as base64
 import requests
 import torch
 

examples/pooling/embed/vision_embedding_online.py

@@ -7,10 +7,10 @@ Refer to each `run_*` function for the command to run the server for that model.
 """
 
 import argparse
-import base64
 import io
 from typing import Literal
 
+import pybase64 as base64
 from openai import OpenAI
 from openai._types import NOT_GIVEN, NotGiven
 from openai.types.chat import ChatCompletionMessageParam

examples/pooling/plugin/prithvi_geospatial_mae_io_processor.py

@@ -1,8 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import base64
 import os
 
+import pybase64 as base64
 import torch
 
 from vllm import LLM

examples/pooling/plugin/prithvi_geospatial_mae_online.py

@@ -1,9 +1,9 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
-import base64
 import os
 
+import pybase64 as base64
 import requests
 
 # This example shows how to perform an online inference that generates

examples/pooling/score/colqwen3_5_rerank_online.py

@@ -0,0 +1,130 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Example of using ColQwen3.5 late interaction model for reranking.
+
+ColQwen3.5 is a multi-modal ColBERT-style model based on Qwen3.5.
+It produces per-token embeddings and uses MaxSim scoring for retrieval
+and reranking. Supports both text and image inputs.
+
+Start the server with:
+    vllm serve athrael-soju/colqwen3.5-4.5B --max-model-len 4096
+
+Then run this script:
+    python colqwen3_5_rerank_online.py
+"""
+
+import requests
+
+MODEL = "athrael-soju/colqwen3.5-4.5B"
+BASE_URL = "http://127.0.0.1:8000"
+
+headers = {"accept": "application/json", "Content-Type": "application/json"}
+
+
+def rerank_text():
+    """Text-only reranking via /rerank endpoint."""
+    print("=" * 60)
+    print("1. Text reranking (/rerank)")
+    print("=" * 60)
+
+    data = {
+        "model": MODEL,
+        "query": "What is machine learning?",
+        "documents": [
+            "Machine learning is a subset of artificial intelligence.",
+            "Python is a programming language.",
+            "Deep learning uses neural networks for complex tasks.",
+            "The weather today is sunny.",
+        ],
+    }
+
+    response = requests.post(f"{BASE_URL}/rerank", headers=headers, json=data)
+
+    if response.status_code == 200:
+        result = response.json()
+        print("\n  Ranked documents (most relevant first):")
+        for item in result["results"]:
+            doc_idx = item["index"]
+            score = item["relevance_score"]
+            print(f"    [{score:.4f}] {data['documents'][doc_idx]}")
+    else:
+        print(f"  Request failed: {response.status_code}")
+        print(f"  {response.text[:300]}")
+
+
+def score_text():
+    """Text-only scoring via /score endpoint."""
+    print()
+    print("=" * 60)
+    print("2. Text scoring (/score)")
+    print("=" * 60)
+
+    query = "What is the capital of France?"
+    documents = [
+        "The capital of France is Paris.",
+        "Berlin is the capital of Germany.",
+        "Python is a programming language.",
+    ]
+
+    data = {
+        "model": MODEL,
+        "text_1": query,
+        "text_2": documents,
+    }
+
+    response = requests.post(f"{BASE_URL}/score", headers=headers, json=data)
+
+    if response.status_code == 200:
+        result = response.json()
+        print(f"\n  Query: {query}\n")
+        for item in result["data"]:
+            idx = item["index"]
+            score = item["score"]
+            print(f"    Doc {idx} (score={score:.4f}): {documents[idx]}")
+    else:
+        print(f"  Request failed: {response.status_code}")
+        print(f"  {response.text[:300]}")
+
+
+def score_text_top_n():
+    """Text reranking with top_n filtering via /rerank endpoint."""
+    print()
+    print("=" * 60)
+    print("3. Text reranking with top_n=2 (/rerank)")
+    print("=" * 60)
+
+    data = {
+        "model": MODEL,
+        "query": "What is the capital of France?",
+        "documents": [
+            "The capital of France is Paris.",
+            "Berlin is the capital of Germany.",
+            "Python is a programming language.",
+            "The Eiffel Tower is in Paris.",
+        ],
+        "top_n": 2,
+    }
+
+    response = requests.post(f"{BASE_URL}/rerank", headers=headers, json=data)
+
+    if response.status_code == 200:
+        result = response.json()
+        print(f"\n  Top {data['top_n']} results:")
+        for item in result["results"]:
+            doc_idx = item["index"]
+            score = item["relevance_score"]
+            print(f"    [{score:.4f}] {data['documents'][doc_idx]}")
+    else:
+        print(f"  Request failed: {response.status_code}")
+        print(f"  {response.text[:300]}")
+
+
+def main():
+    rerank_text()
+    score_text()
+    score_text_top_n()
+
+
+if __name__ == "__main__":
+    main()

examples/pooling/score/colqwen3_rerank_online.py

@@ -15,9 +15,9 @@ Then run this script:
     python colqwen3_rerank_online.py
 """
 
-import base64
 from io import BytesIO
 
+import pybase64 as base64
 import requests
 from PIL import Image
 

examples/pooling/token_embed/colqwen3_token_embed_online.py

@@ -21,10 +21,10 @@ Then run this script:
 """
 
 import argparse
-import base64
 from io import BytesIO
 
 import numpy as np
+import pybase64 as base64
 import requests
 from PIL import Image
 

pyproject.toml

@@ -167,6 +167,7 @@ fo = "fo"
 nd = "nd"
 eles = "eles"
 datas = "datas"
+ser = "ser"
 ure = "ure"
 
 [tool.uv]

requirements/common.txt

@@ -24,20 +24,20 @@ outlines_core == 0.2.11
 # required for outlines backend disk cache
 diskcache == 5.6.3
 lark == 1.2.2
-xgrammar == 0.1.29; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64" or platform_machine == "s390x" or platform_machine == "ppc64le"
+xgrammar >= 0.1.32, < 1.0.0; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64" or platform_machine == "s390x" or platform_machine == "ppc64le"
 typing_extensions >= 4.10
 filelock >= 3.16.1 # need to contain https://github.com/tox-dev/filelock/pull/317
 partial-json-parser # used for parsing partial JSON outputs
 pyzmq >= 25.0.0
 msgspec
 gguf >= 0.17.0
-mistral_common[image] >= 1.9.1
+mistral_common[image] >= 1.10.0
 opencv-python-headless >= 4.13.0    # required for video IO
 pyyaml
 six>=1.16.0; python_version > '3.11' # transitive dependency of pandas that needs to be the latest version for python 3.12
 setuptools>=77.0.3,<81.0.0; python_version > '3.11' # Setuptools is used by triton, we need to ensure a modern version is installed for 3.12+ so that it does not try to import distutils, which was removed in 3.12
 einops # Required for Qwen2-VL.
-compressed-tensors == 0.13.0 # required for compressed-tensors
+compressed-tensors == 0.14.0.1 # required for compressed-tensors
 depyf==0.20.0 # required for profiling and debugging with compilation config
 cloudpickle # allows pickling lambda functions in model_executor/models/registry.py
 watchfiles # required for http server to monitor the updates of TLS files

requirements/cuda.txt

@@ -9,7 +9,7 @@ torchaudio==2.10.0
 # These must be updated alongside torch
 torchvision==0.25.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version
 # FlashInfer should be updated together with the Dockerfile
-flashinfer-python==0.6.4
+flashinfer-python==0.6.6
 # Cap nvidia-cudnn-frontend (transitive dep of flashinfer) due to
 # breaking changes in 1.19.0
 nvidia-cudnn-frontend>=1.13.0,<1.19.0

requirements/lint.txt

@@ -1,2 +1,2 @@
 # formatting
-pre-commit==4.0.1
+pre-commit>=4.5.1

requirements/nightly_torch_test.txt

@@ -42,6 +42,7 @@ tritonclient>=2.51.0
 
 numba == 0.61.2 # Required for N-gram speculative decoding
 numpy
-runai-model-streamer[s3,gcs]==0.15.3
+runai-model-streamer[s3,gcs,azure]==0.15.7
 fastsafetensors>=0.2.2
+instanttensor>=0.1.5
 pydantic>=2.12 # 2.11 leads to error on python 3.13

requirements/rocm-test.txt

@@ -45,10 +45,12 @@ pystemmer==3.0.0
     # via mteb
 
 # Multi-modal processing
+av==16.1.0
+    # required for audio_in_video tests
 blobfile==3.0.0
     # Multi-Modal Models Test
 decord==0.6.0
-    # video processing, required by entrypoints/openai/test_video.py
+    # video processing, required by entrypoints/openai/chat_completion/test_video.py
 rapidfuzz==3.12.1
 
 # OpenAI compatibility and testing
@@ -95,7 +97,7 @@ transformers==4.57.5
 # Pin HF Hub version
 huggingface-hub==0.36.2
 # Pin Mistral Common
-mistral-common[image,audio]==1.9.1
+mistral-common[image,audio]==1.10.0
 # Required for Prithvi tests
 terratorch==1.2.2
 # Required for Prithvi tests

requirements/rocm.txt

@@ -15,7 +15,7 @@ tensorizer==2.10.1
 packaging>=24.2
 setuptools>=77.0.3,<80.0.0
 setuptools-scm>=8
-runai-model-streamer[s3,gcs]==0.15.3
+runai-model-streamer[s3,gcs,azure]==0.15.7
 conch-triton-kernels==1.2.1
 timm>=1.0.17
 # amd-quark: required for Quark quantization on ROCm 

requirements/test.in

@@ -10,6 +10,7 @@ pytest-cov
 
 # testing utils
 albumentations # required for Nemotron Parse in test_common.py
+av  # required for audio_in_video tests
 backoff # required for phi4mm test
 blobfile # required for kimi-vl test
 einops # required for MPT, qwen-vl
@@ -55,8 +56,9 @@ grpcio-reflection==1.78.0
 arctic-inference == 0.1.1 # Required for suffix decoding test
 numba == 0.61.2 # Required for N-gram speculative decoding
 numpy
-runai-model-streamer[s3,gcs]==0.15.3
+runai-model-streamer[s3,gcs,azure]==0.15.7
 fastsafetensors>=0.2.2 # 0.2.2 contains important fixes for multi-GPU mem usage
+instanttensor>=0.1.5
 pydantic>=2.12 # 2.11 leads to error on python 3.13
 decord==0.6.0
 terratorch >= 1.2.2 # Required for Prithvi tests

requirements/test.txt

@@ -62,6 +62,16 @@ attrs==24.2.0
     #   referencing
 audioread==3.0.1
     # via librosa
+av==16.1.0
+    # via -r requirements/test.in
+azure-core==1.38.2
+    # via
+    #   azure-identity
+    #   azure-storage-blob
+azure-identity==1.25.2
+    # via runai-model-streamer-azure
+azure-storage-blob==12.28.0
+    # via runai-model-streamer-azure
 backoff==2.2.1
     # via
     #   -r requirements/test.in
@@ -101,8 +111,10 @@ certifi==2024.8.30
     #   rasterio
     #   requests
     #   sentry-sdk
-cffi==1.17.1
-    # via soundfile
+cffi==2.0.0
+    # via
+    #   cryptography
+    #   soundfile
 chardet==5.2.0
     # via mbstrdecoder
 charset-normalizer==3.4.0
@@ -146,6 +158,12 @@ coverage==7.10.6
     # via pytest-cov
 cramjam==2.9.0
     # via fastparquet
+cryptography==46.0.5
+    # via
+    #   azure-identity
+    #   azure-storage-blob
+    #   msal
+    #   pyjwt
 cuda-bindings==12.9.4
     # via torch
 cuda-pathfinder==1.3.3
@@ -375,6 +393,10 @@ inflect==5.6.2
     # via datamodel-code-generator
 iniconfig==2.0.0
     # via pytest
+instanttensor==0.1.5
+    # via -r requirements/test.in
+isodate==0.7.2
+    # via azure-storage-blob
 isoduration==20.11.0
     # via jsonschema
 isort==5.13.2
@@ -482,12 +504,18 @@ mbstrdecoder==1.1.3
     #   typepy
 mdurl==0.1.2
     # via markdown-it-py
-mistral-common==1.9.1
+mistral-common==1.10.0
     # via -r requirements/test.in
 more-itertools==10.5.0
     # via lm-eval
 mpmath==1.3.0
     # via sympy
+msal==1.34.0
+    # via
+    #   azure-identity
+    #   msal-extensions
+msal-extensions==1.3.1
+    # via azure-identity
 msgpack==1.1.0
     # via
     #   librosa
@@ -824,6 +852,8 @@ pydantic-extra-types==2.10.5
     # via mistral-common
 pygments==2.18.0
     # via rich
+pyjwt==2.11.0
+    # via msal
 pyogrio==0.11.0
     # via geopandas
 pyparsing==3.2.0
@@ -941,6 +971,7 @@ regex==2024.9.11
     #   transformers
 requests==2.32.3
     # via
+    #   azure-core
     #   buildkite-test-collector
     #   datasets
     #   diffusers
@@ -953,6 +984,7 @@ requests==2.32.3
     #   lightly
     #   lm-eval
     #   mistral-common
+    #   msal
     #   mteb
     #   pooch
     #   ray
@@ -989,11 +1021,13 @@ rsa==4.9.1
     # via google-auth
 rtree==1.4.0
     # via torchgeo
-runai-model-streamer==0.15.3
+runai-model-streamer==0.15.7
     # via -r requirements/test.in
-runai-model-streamer-gcs==0.15.3
+runai-model-streamer-azure==0.15.7
     # via runai-model-streamer
-runai-model-streamer-s3==0.15.3
+runai-model-streamer-gcs==0.15.7
+    # via runai-model-streamer
+runai-model-streamer-s3==0.15.7
     # via runai-model-streamer
 s3transfer==0.10.3
     # via boto3
@@ -1169,6 +1203,7 @@ torch==2.10.0+cu129
     #   accelerate
     #   bitsandbytes
     #   encodec
+    #   instanttensor
     #   kornia
     #   lightly
     #   lightning
@@ -1261,6 +1296,9 @@ typing-extensions==4.15.0
     #   aiosignal
     #   albumentations
     #   alembic
+    #   azure-core
+    #   azure-identity
+    #   azure-storage-blob
     #   chz
     #   fastapi
     #   grpcio

setup.py

@@ -657,13 +657,18 @@ class precompiled_wheel_utils:
     def get_base_commit_in_main_branch() -> str:
         try:
             # Get the latest commit hash of the upstream main branch.
-            resp_json = subprocess.check_output(
-                [
-                    "curl",
-                    "-s",
-                    "https://api.github.com/repos/vllm-project/vllm/commits/main",
+            curl_cmd = [
+                "curl",
+                "-s",
+                "https://api.github.com/repos/vllm-project/vllm/commits/main",
+            ]
+            github_token = os.getenv("GH_TOKEN", os.getenv("GITHUB_TOKEN"))
+            if github_token:
+                curl_cmd += [
+                    "-H",
+                    f"Authorization: token {github_token}",
                 ]
-            ).decode("utf-8")
+            resp_json = subprocess.check_output(curl_cmd).decode("utf-8")
             upstream_main_commit = json.loads(resp_json)["sha"]
             print(f"Upstream main branch latest commit: {upstream_main_commit}")
 
@@ -920,6 +925,7 @@ if _is_cpu():
 
     if platform.machine() in ("x86_64", "AMD64"):
         ext_modules.append(CMakeExtension(name="vllm._C"))
+        ext_modules.append(CMakeExtension(name="vllm._C_AVX512"))
         ext_modules.append(CMakeExtension(name="vllm._C_AVX2"))
     else:
         ext_modules.append(CMakeExtension(name="vllm._C"))
@@ -965,10 +971,13 @@ setup(
     ext_modules=ext_modules,
     install_requires=get_requirements(),
     extras_require={
+        # AMD Zen CPU optimizations via zentorch
+        "zen": ["zentorch"],
         "bench": ["pandas", "matplotlib", "seaborn", "datasets", "scipy", "plotly"],
         "tensorizer": ["tensorizer==2.10.1"],
         "fastsafetensors": ["fastsafetensors >= 0.2.2"],
-        "runai": ["runai-model-streamer[s3,gcs] >= 0.15.3"],
+        "instanttensor": ["instanttensor >= 0.1.5"],
+        "runai": ["runai-model-streamer[s3,gcs,azure] >= 0.15.7"],
         "audio": [
             "librosa",
             "scipy",
@@ -981,9 +990,9 @@ setup(
         # Optional deps for AMD FP4 quantization support
         "petit-kernel": ["petit-kernel"],
         # Optional deps for Helion kernel development
-        "helion": ["helion"],
+        "helion": ["helion==0.3.2"],
         # Optional deps for gRPC server (vllm serve --grpc)
-        "grpc": ["smg-grpc-servicer >= 0.4.2"],
+        "grpc": ["smg-grpc-servicer[vllm] >= 0.5.0"],
         # Optional deps for OpenTelemetry tracing
         "otel": [
             "opentelemetry-sdk>=1.26.0",