Apply pre-commit formatting and use torch.accelerator API

- Apply ruff, clang-format formatting fixes
- Replace torch.cuda.set_device/device_count/synchronize with
  torch.accelerator equivalents per project convention

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signed-off-by: Alexander Matveev <alexm-redhat@dgx-b200-02.mgmt.accl-001.lab.rdu2.dc.redhat.com>

Signed-off-by:  <>

.buildkite/hardware_tests/cpu.yaml

@@ -28,18 +28,19 @@ steps:
       pytest -x -v -s tests/kernels/quantization/test_cpu_fp8_scaled_mm.py
       pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py"
 
-- label: CPU-Compatibility Tests
-  depends_on: []
-  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"
+# Note: SDE can't be downloaded from CI host because of AWS WAF
+# - label: CPU-Compatibility Tests
+#   depends_on: []
+#   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: []
@@ -90,7 +91,7 @@ steps:
   - tests/quantization/test_cpu_wna16.py
   commands:
     - |
-      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
+      bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 45m "
       pytest -x -v -s tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_logprobs
       pytest -x -v -s tests/quantization/test_cpu_wna16.py"
       

.buildkite/intel_jobs/basic_correctness.yaml

@@ -0,0 +1,22 @@
+group: Basic Correctness
+depends_on:
+  - image-build-xpu
+steps:
+- label: XPU Sleep Mode
+  timeout_in_minutes: 30
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/
+  - tests/basic_correctness/test_cumem.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      pytest -v -s basic_correctness/test_mem.py::test_end_to_end'

.buildkite/intel_jobs/expert_parallelism_intel.yaml

@@ -0,0 +1,23 @@
+group: Expert Parallelism
+depends_on: 
+  - image-build-xpu
+steps:
+- label: EPLB Algorithm
+  key: eplb-algorithm
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/distributed/eplb
+  - tests/distributed/test_eplb_algo.py
+  - tests/distributed/test_eplb_utils.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s distributed/test_eplb_algo.py'

.buildkite/intel_jobs/misc_intel.yaml

@@ -38,7 +38,17 @@ steps:
     REPO: "vllm-ci-test-repo"
     VLLM_TEST_DEVICE: "xpu"
   source_file_dependencies:
-    - vllm/
+    - vllm/config/
+    - vllm/distributed/
+    - vllm/engine/
+    - vllm/inputs/
+    - vllm/logger.py
+    - vllm/model_executor/
+    - vllm/platforms/
+    - vllm/sampling_params.py
+    - vllm/transformers_utils/
+    - vllm/utils/
+    - vllm/v1/
     - tests/v1/sample
     - tests/v1/logits_processors
     - tests/v1/test_oracle.py
@@ -52,4 +62,126 @@ steps:
       pytest -v -s v1/logits_processors --ignore=v1/logits_processors/test_custom_online.py --ignore=v1/logits_processors/test_custom_offline.py &&
       pytest -v -s v1/test_oracle.py &&
       pytest -v -s v1/test_request.py &&
-      pytest -v -s v1/test_outputs.py'
+      pytest -v -s v1/test_outputs.py &&
+      pytest -v -s v1/sample/test_topk_topp_sampler.py'
+
+- label: XPU CPU Offload
+  timeout_in_minutes: 60
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/
+    - vllm/v1/kv_offload/
+    - vllm/v1/kv_connector/
+    - tests/v1/kv_offload/
+    - tests/v1/kv_connector/unit/test_offloading_connector.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      cd tests &&
+      pytest -v -s v1/kv_offload &&
+      pytest -v -s v1/kv_connector/unit/test_offloading_connector.py'
+
+- label: Regression
+  key: regression
+  timeout_in_minutes: 30
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/config/
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/inputs/
+  - vllm/model_executor/
+  - vllm/multimodal/
+  - vllm/platforms/
+  - vllm/sampling_params.py
+  - vllm/transformers_utils/
+  - vllm/utils/
+  - vllm/v1/
+  - tests/test_regression
+  commands:
+  - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install modelscope &&
+      cd tests &&
+      pytest -v -s test_regression.py'
+
+- label: Metrics, Tracing (2 GPUs)
+  key: metrics-tracing-2-gpus
+  timeout_in_minutes: 30
+  num_devices: 2
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/config/
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/inputs/
+  - vllm/model_executor/
+  - vllm/multimodal/
+  - vllm/platforms/
+  - vllm/sampling_params.py
+  - vllm/tracing/
+  - vllm/transformers_utils/
+  - vllm/utils/
+  - vllm/v1/
+  - tests/v1/tracing
+  commands:
+  - >-
+    bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+    'pip install opentelemetry-sdk\>=1.26.0 opentelemetry-api\>=1.26.0 opentelemetry-exporter-otlp\>=1.26.0 opentelemetry-semantic-conventions-ai\>=0.4.1 &&
+    cd tests &&
+    pytest -v -s v1/tracing'
+
+- label: Async Engine, Inputs, Utils, Worker
+  key: async-engine-inputs-utils-worker
+  timeout_in_minutes: 30
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/assets/
+  - vllm/config/
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/inputs/
+  - vllm/model_executor/
+  - vllm/multimodal/
+  - vllm/platforms/
+  - vllm/sampling_params.py
+  - vllm/tokenizers/
+  - vllm/transformers_utils/
+  - vllm/utils/
+  - vllm/v1/
+  - tests/detokenizer
+  - tests/multimodal
+  - tests/utils_
+  commands:
+  - >-
+    bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+    'cd tests &&
+    pip install av &&
+    pytest -v -s detokenizer &&
+    pytest -v -s -m "not cpu_test" ./multimodal &&
+    pytest -v -s utils_ --ignore=utils_/test_mem_utils.py'

.buildkite/intel_jobs/models_multimodal_intel.yaml

@@ -0,0 +1,111 @@
+group: Models - Multimodal
+depends_on: 
+  - image-build-xpu
+steps:
+- label: "Multi-Modal Models (Standard) 1: qwen2"
+  key: multi-modal-models-standard-1-qwen2
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install av git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      cd tests &&
+      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'
+
+- label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
+  key: multi-modal-models-standard-2-qwen3-gemma
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      cd tests &&
+      pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model'
+
+- label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
+  key: multi-modal-models-standard-3-llava-qwen2-vl
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      cd tests &&
+      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'
+
+- label: "Multi-Modal Models (Standard) 4: other + whisper"
+  key: multi-modal-models-standard-4-other-whisper
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install av git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      cd tests &&
+      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/generation/test_memory_leak.py --ignore models/multimodal/processing'
+
+- label: Multi-Modal Processor # 44min
+  key: multi-modal-processor
+  timeout_in_minutes: 45
+  device: intel_gpu
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/
+  - tests/models/multimodal
+  - tests/models/registry.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'pip install av matplotlib ftfy git+https://github.com/TIGER-AI-Lab/Mantis.git &&
+      pip install open-clip-torch --no-deps &&
+      cd tests &&
+      pytest -v -s models/multimodal/processing/test_tensor_schema.py
+      --deselect "tests/models/multimodal/processing/test_tensor_schema.py::test_model_tensor_schema[mistralai/Mistral-Large-3-675B-Instruct-2512-NVFP4]"
+      --deselect "tests/models/multimodal/processing/test_tensor_schema.py::test_model_tensor_schema[Qwen/Qwen2.5-Omni-7B-AWQ]"
+      --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB'
+  parallelism: 4

.buildkite/intel_jobs/test-intel.yaml

@@ -40,7 +40,9 @@ steps:
         python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --block-size 64 --enforce-eager --max-model-len 8192 &&
         python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 &&
         python3 examples/basic/offline_inference/generate.py --model ibm-research/PowerMoE-3b --block-size 64 --enforce-eager -tp 2 --enable-expert-parallel &&
-        python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --max-model-len 8192
+        python3 examples/basic/offline_inference/generate.py --model superjob/Qwen3-4B-Instruct-2507-GPTQ-Int4 --max-model-len 8192 &&
+        VLLM_XPU_FUSED_MOE_USE_REF=1 python3 examples/basic/offline_inference/generate.py --model Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 --enforce-eager -tp 2 --max-model-len 8192 &&
+        python3 examples/basic/offline_inference/generate.py --model INCModel/Qwen3-30B-A3B-Instruct-2507-MXFP4-LLMC --enforce-eager -tp 2 --max-model-len 8192
         '
   - label: "XPU V1 test"
     depends_on:
@@ -83,5 +85,5 @@ steps:
         bash .buildkite/scripts/hardware_ci/run-intel-test.sh
         'pip install av &&
         cd tests &&
-        pytest -v -s entrypoints/openai/chat_completion/test_audio_in_video.py &&
+        pytest -v -s entrypoints/multimodal/openai/chat_completion/test_audio_in_video.py &&
         pytest -v -s benchmarks/test_serve_cli.py'

.buildkite/lm-eval-harness/configs/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8.yaml

@@ -6,9 +6,7 @@ tasks:
     value: 0.7142
   - name: "exact_match,flexible-extract"
     value: 0.4579
-env_vars:
-  VLLM_USE_FLASHINFER_MOE_FP8: "1"
-  VLLM_FLASHINFER_MOE_BACKEND: "throughput"
+moe_backend: "flashinfer_cutlass"
 limit: 1319
 num_fewshot: 5
 max_model_len: 262144

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -68,6 +68,10 @@ def launch_lm_eval(eval_config, tp_size):
     if current_platform.is_rocm() and "Nemotron-3" in eval_config["model_name"]:
         model_args += "attention_backend=TRITON_ATTN"
 
+    moe_backend = eval_config.get("moe_backend", None)
+    if moe_backend is not None:
+        model_args += f"moe_backend={moe_backend},"
+
     env_vars = eval_config.get("env_vars", None)
     with scoped_env_vars(env_vars):
         results = lm_eval.simple_evaluate(

.buildkite/release-pipeline.yaml

@@ -1,12 +1,25 @@
 # CUDA architecture lists — following PyTorch RELEASE.md
 # (https://github.com/pytorch/pytorch/blob/main/RELEASE.md)
 # SM86 included for broader Ampere coverage; SM89 for marlin fp8 support
+# These requested arches are filtered by CMake's CUDA_SUPPORTED_ARCHS before
+# per-kernel arch selection. Do not add +PTX here: top-level +PTX is stripped
+# during that filtering, so kernels that need PTX must request it locally.
 env:
-  CUDA_ARCH_X86: "7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX"
-  # aarch64 only architectures: 8.7 for Orin, 11.0 for Thor (since CUDA 13)
-  CUDA_ARCH_AARCH64: "8.0 8.7 8.9 9.0 10.0 11.0 12.0+PTX"
+  # for CUDA >=13, sm_100+ targets have family specifiers (see CMakeLists.txt)
+  # so targets like 10.3 and 12.1 are automatically supported with this list
+  CUDA_ARCH_X86: "7.5 8.0 8.6 8.9 9.0 10.0 12.0"
+  # aarch64-only targets: Orin (8.7), Thor (11.0, CUDA 13+)
+  CUDA_ARCH_AARCH64: "8.0 8.7 8.9 9.0 10.0 11.0 12.0"
+
+  # for CUDA <13, we need to specify all needed targets
+  # some targets (10.3, 12.1) are skipped to limit the wheel size (< 500MB)
+  # please use CUDA 13 wheels or compile yourself on these new devices
   CUDA_ARCH_X86_CU129: "7.5 8.0 8.6 8.9 9.0 10.0 12.0"
   CUDA_ARCH_AARCH64_CU129: "8.0 8.7 8.9 9.0 10.0 12.0"
+  
+  # pre-built mooncake wheels
+  # the manylinux_2_35 wheel has compatibility issue on Ubuntu 24.04
+  # so we use different wheels for the time being
   MOONCAKE_WHEEL_AARCH64_2_35: "https://vllm-wheels.s3.amazonaws.com/mooncake/mooncake_transfer_engine-0.3.10.post2-0da9dfea3-cp312-cp312-manylinux_2_35_aarch64.whl"
   MOONCAKE_WHEEL_AARCH64_2_39: "https://vllm-wheels.s3.amazonaws.com/mooncake/mooncake_transfer_engine-0.3.10.post2-0da9dfea3-cp312-cp312-manylinux_2_39_aarch64.whl"
   MOONCAKE_WHEEL_X86_64: "https://vllm-wheels.s3.amazonaws.com/mooncake/mooncake_transfer_engine-0.3.10.post2-0da9dfea3-cp312-cp312-manylinux_2_35_x86_64.whl"
@@ -846,7 +859,6 @@ steps:
         allow_failure: true
       - step: build-cpu-release-image-arm64
         allow_failure: true
-    if: build.env("NIGHTLY") != "1"
 
   - label: "Publish release images to DockerHub"
     depends_on:

.buildkite/scripts/ci-clean-log.sh

@@ -13,5 +13,8 @@ INPUT_FILE="$1"
 # Strip timestamps
 sed -i 's/^\[[0-9]\{4\}-[0-9]\{2\}-[0-9]\{2\}T[0-9]\{2\}:[0-9]\{2\}:[0-9]\{2\}Z\] //' "$INPUT_FILE"
 
+# Strip Buildkite inline timestamp markers (ESC _bk;t=<ms> BEL)
+sed -i 's/\x1B_bk;t=[0-9]*\x07//g' "$INPUT_FILE"
+
 # Strip colorization
 sed -i -r 's/\x1B\[[0-9;]*[mK]//g' "$INPUT_FILE"

.buildkite/scripts/ci-fetch-log.sh

@@ -1,74 +1,178 @@
 #!/bin/bash
-# Usage: ./ci-fetch-log.sh <buildkite_job_url> [output_file]
-#        ./ci-fetch-log.sh <build_number> <job_uuid> [output_file]
+# Fetch vLLM Buildkite CI logs (public; no login required).
 #
-# Downloads the raw log for a Buildkite job from the public, unauthenticated
-# /organizations/<org>/pipelines/<pipeline>/builds/<n>/jobs/<uuid>/download
-# endpoint, then strips ANSI/timestamps via ci-clean-log.sh.
+# Usage:
+#   ci-fetch-log.sh [--soft|--all] --pr [<PR>]  failed jobs in the PR's latest
+#                                               build (current branch if omitted)
+#   ci-fetch-log.sh [--soft|--all] <build_url>  failed jobs in that build
+#   ci-fetch-log.sh <job_url> [output]          one job; both #<job_uuid> and
+#                                               ?sid=<id> URL forms work
+#   ci-fetch-log.sh <build> <job_uuid> [output]
 #
-# Find <build_number> and <job_uuid> via:
-#   gh pr checks <PR> --repo vllm-project/vllm
-# Each failing row's URL is .../builds/<build_number>#<job_uuid>.
-#
-# Default output path: ci-<build>-<uuid_first_13_chars>.log (e.g.
-# ci-68478-019e6b07-daae.log). Jobs in the same build share the UUID's
-# first 8 chars, so the second segment is needed for uniqueness when
-# fetching multiple jobs in parallel. The script refuses to overwrite an
-# existing output file; pass an explicit path or set CI_FETCH_LOG_FORCE=1
-# to override.
+# --soft also fetches soft-failed jobs; --all fetches every finished job.
+# Saves each log as ci-<build>-<job-name>.log (ANSI/timestamps stripped) and
+# prints "<file>\t<job name>" per job. [output] is single-job only; "-"
+# streams to stdout. Existing files are kept; CI_FETCH_LOG_FORCE=1 refetches.
 
 set -euo pipefail
 
 ORG="vllm"
 PIPELINE="ci"
+UA="vllm-ci-fetch-log"
+UUID_RE='[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}'
 
 usage() {
-    echo "Usage: $0 <buildkite_job_url> [output_file]"
-    echo "       $0 <build_number> <job_uuid> [output_file]"
+    sed -n '2,15p' "$0" | sed 's/^# \{0,1\}//'
     exit 1
 }
 
-if [ $# -lt 1 ]; then usage; fi
+die() {
+    echo "$1" >&2
+    exit 1
+}
 
-if [[ "$1" == https://* ]]; then
+BUILD="" JOB="" SID="" OUT=""
+SCOPE="failed"
+
+while :; do
+    case "${1:-}" in
+    --soft) SCOPE="soft" ;;
+    --all) SCOPE="all" ;;
+    *) break ;;
+    esac
+    shift
+done
+
+case "${1:-}" in
+--pr)
+    PR="${2:-}"
+    # gh pr checks exits non-zero when checks are failing; that is the
+    # expected case here.
+    URL=$(gh pr checks ${PR:+"$PR"} --repo vllm-project/vllm 2>/dev/null |
+        grep -oE "https://buildkite.com/${ORG}/${PIPELINE}/builds/[0-9]+" |
+        sort -t/ -k7 -n | tail -1 || true)
+    [ -n "$URL" ] || die "No Buildkite build found via: gh pr checks ${PR:-<current branch>}"
+    BUILD="${URL##*/}"
+    ;;
+https://*)
     BUILD=$(echo "$1" | sed -nE 's#.*/builds/([0-9]+).*#\1#p')
-    JOB=$(echo "$1" | grep -oE '[0-9a-f]{8}-[0-9a-f-]+' | head -n 1)
+    JOB=$(echo "$1" | grep -oE "#${UUID_RE}" | head -n 1 | cut -c2- || true)
+    SID=$(echo "$1" | grep -oE "[?&]sid=${UUID_RE}" | head -n 1 | sed 's/.*sid=//' || true)
     OUT="${2:-}"
-else
-    if [ $# -lt 2 ]; then usage; fi
+    [ -n "$BUILD" ] || die "Could not parse build number from: $1"
+    ;;
+[0-9]*)
+    [ $# -ge 2 ] || usage
     BUILD="$1"
     JOB="$2"
     OUT="${3:-}"
-fi
-
-if [ -z "$BUILD" ] || [ -z "$JOB" ]; then
-    echo "Could not parse build number or job UUID from: $1" >&2
+    ;;
+*)
     usage
-fi
-
-# Jobs in the same build share the UUID's first segment, so include the
-# second segment (chars 9-13, e.g. "019e6b07-daae") to keep default filenames
-# unique when fetching multiple jobs from one build in parallel.
-if [ -z "$OUT" ]; then
-    OUT="ci-${BUILD}-${JOB:0:13}.log"
-fi
-
-if [ -e "$OUT" ] && [ -z "${CI_FETCH_LOG_FORCE:-}" ]; then
-    echo "Refusing to overwrite existing $OUT (set CI_FETCH_LOG_FORCE=1 or pass an explicit output path)." >&2
-    exit 1
-fi
+    ;;
+esac
 
 COOKIES=$(mktemp)
-trap 'rm -f "$COOKIES"' EXIT
+JOBS_TSV=$(mktemp)
+trap 'rm -f "$COOKIES" "$JOBS_TSV"' EXIT
 
-# Buildkite issues a session cookie on first hit; subsequent /download needs it.
-curl -fsSL -c "$COOKIES" -A "vllm-ci-fetch-log" \
+# Buildkite issues a session cookie on first hit; later requests need it.
+curl -fsSL -c "$COOKIES" -A "$UA" \
     "https://buildkite.com/${ORG}/${PIPELINE}/builds/${BUILD}" -o /dev/null
 
-curl -fsSL -b "$COOKIES" -A "vllm-ci-fetch-log" \
-    "https://buildkite.com/organizations/${ORG}/pipelines/${PIPELINE}/builds/${BUILD}/jobs/${JOB}/download" \
-    -o "$OUT"
+# The build's job list (id, step uuid, state, name) is served as JSON from
+# the user-facing /data/jobs endpoint. Flatten it to TSV for easy filtering:
+#   job_id  step_uuid  failed  soft_failed  finished  slug  name
+curl -fsSL -b "$COOKIES" -A "$UA" \
+    "https://buildkite.com/${ORG}/${PIPELINE}/builds/${BUILD}/data/jobs" |
+    python3 -c '
+import json, re, sys
 
-bash "$(dirname "$0")/ci-clean-log.sh" "$OUT"
+data = json.load(sys.stdin)
+if data.get("has_next_page"):
+    print("warning: job list is paginated; some jobs not shown", file=sys.stderr)
+for r in data["records"]:
+    if r.get("type") != "script":
+        continue
+    name = (r.get("name") or "").replace("\t", " ").replace("\n", " ")
+    slug = re.sub(r"[^a-z0-9]+", "-", name.lower()).strip("-")[:60]
+    print("\t".join([
+        r["id"],
+        r.get("step_uuid") or "",
+        str(r.get("passed") is False),
+        str(bool(r.get("soft_failed"))),
+        str(bool(r.get("finished_at"))),
+        slug,
+        name,
+    ]))
+' >"$JOBS_TSV" || die "Could not list jobs for build ${BUILD}"
 
-echo "$OUT"
+if [ -n "$SID" ] && [ -z "$JOB" ]; then
+    # The ?sid= in builds/<N>/list URLs is the *step* uuid, not the job uuid.
+    JOB=$(awk -F'\t' -v s="$SID" '$1 == s || $2 == s {print $1; exit}' "$JOBS_TSV")
+    [ -n "$JOB" ] || die "No job matching sid=${SID} in build ${BUILD}"
+fi
+
+fetch_job() { # <job_uuid> <output_file>
+    curl -fsSL -b "$COOKIES" -A "$UA" \
+        "https://buildkite.com/organizations/${ORG}/pipelines/${PIPELINE}/builds/${BUILD}/jobs/$1/download" \
+        -o "$2"
+    bash "$(dirname "$0")/ci-clean-log.sh" "$2"
+}
+
+if [ -n "$JOB" ]; then
+    # Single-job mode.
+    NAME=$(awk -F'\t' -v j="$JOB" '$1 == j {print $7; exit}' "$JOBS_TSV")
+    SLUG=$(awk -F'\t' -v j="$JOB" '$1 == j {print $6; exit}' "$JOBS_TSV")
+    [ -n "$OUT" ] || OUT="ci-${BUILD}-${SLUG:-${JOB:0:13}}.log"
+    if [ "$OUT" = "-" ]; then
+        TMP=$(mktemp)
+        fetch_job "$JOB" "$TMP"
+        cat "$TMP"
+        rm -f "$TMP"
+        exit 0
+    fi
+    if [ -e "$OUT" ] && [ -z "${CI_FETCH_LOG_FORCE:-}" ]; then
+        die "Refusing to overwrite existing ${OUT} (set CI_FETCH_LOG_FORCE=1 or pass an output path)."
+    fi
+    fetch_job "$JOB" "$OUT"
+    printf '%s\t%s\n' "$OUT" "${NAME:-$JOB}"
+    exit 0
+fi
+
+# Build-wide mode: fetch finished jobs matching $SCOPE.
+[ -z "$OUT" ] || die "[output_file] is only valid when fetching a single job."
+
+case "$SCOPE" in
+failed) FILTER='$3 == "True" && $4 == "False" && $5 == "True"' ;;
+soft) FILTER='$3 == "True" && $5 == "True"' ;;
+all) FILTER='$5 == "True"' ;;
+esac
+
+if [ "$SCOPE" = "failed" ]; then
+    SOFT=$(awk -F'\t' '$3 == "True" && $4 == "True"' "$JOBS_TSV" | wc -l)
+    [ "$SOFT" -eq 0 ] || echo "Skipping ${SOFT} soft-failed job(s); use --soft to include them." >&2
+fi
+
+FOUND=0
+EMITTED=" "
+while IFS=$'\t' read -r job_id _ _ _ _ slug name; do
+    FOUND=$((FOUND + 1))
+    out="ci-${BUILD}-${slug:-${job_id:0:13}}.log"
+    # Retries share a name with the original job; disambiguate by uuid.
+    case "$EMITTED" in
+    *" $out "*) out="ci-${BUILD}-${slug:-job}-${job_id:0:13}.log" ;;
+    esac
+    EMITTED="${EMITTED}${out} "
+    if [ -e "$out" ] && [ -z "${CI_FETCH_LOG_FORCE:-}" ]; then
+        echo "Keeping existing ${out} (set CI_FETCH_LOG_FORCE=1 to refetch)." >&2
+    elif ! fetch_job "$job_id" "$out"; then
+        echo "Failed to download log for job ${job_id} (${name})." >&2
+        continue
+    fi
+    printf '%s\t%s\n' "$out" "$name"
+done < <(awk -F'\t' "$FILTER" "$JOBS_TSV")
+
+if [ "$FOUND" -eq 0 ]; then
+    echo "No matching jobs in build ${BUILD} (scope: ${SCOPE})." >&2
+fi

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

@@ -28,8 +28,10 @@
 ###############################################################################
 set -o pipefail
 
-# Export Python path
-export PYTHONPATH=".."
+# Export Python path for commands that run directly on the host. Containerized
+# tests set this to /vllm-workspace below so spawned Python processes do not
+# depend on their current working directory.
+export PYTHONPATH="${PYTHONPATH:-..}"
 
 ###############################################################################
 # Helper Functions
@@ -377,6 +379,14 @@ HF_CACHE="$(realpath ~)/huggingface"
 mkdir -p "${HF_CACHE}"
 HF_MOUNT="/root/.cache/huggingface"
 
+# Hugging Face Hub defaults to 10s request/download timeouts, while the ROCm
+# CI image currently raises downloads to 60s. AMD model-test jobs routinely
+# start from a cold or partially-populated shared cache, and the 60s read cap
+# has still timed out before pytest reached the vLLM behavior under test.
+# Keep the CI default explicit and overridable from the Buildkite environment.
+: "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
+: "${HF_HUB_ETAG_TIMEOUT:=60}"
+
 # ---- Command source selection ----
 # Prefer VLLM_TEST_COMMANDS (preserves all inner quoting intact).
 # Fall back to $* for backward compatibility, but warn that inner
@@ -416,7 +426,14 @@ fi
 
 echo "Final commands: $commands"
 
-MYPYTHONPATH=".."
+MYPYTHONPATH="/vllm-workspace"
+
+container_job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-0}}"
+container_job_id="${container_job_id//[^A-Za-z0-9_.-]/_}"
+container_job_id_short="${container_job_id:0:8}"
+CONTAINER_TMPDIR="/tmp/vllm-${container_job_id_short}"
+CONTAINER_CACHE_ROOT="/tmp/vllm-buildkite-${container_job_id}/cache"
+CONTAINER_PREFLIGHT="mkdir -p \"\$TMPDIR\" \"\$TORCHINDUCTOR_CACHE_DIR\" \"\$TRITON_CACHE_DIR\" \"\$VLLM_CACHE_ROOT\" \"\$XDG_CACHE_HOME\" && python -c \"import encodings, importlib.metadata as im, importlib.util as iu; [im.version(d) for d in ('transformers', 'torch', 'ray', 'sympy', 'markupsafe', 'vllm')]; missing=[m for m in ('torch.utils.model_zoo', 'transformers.models.nomic_bert', 'ray.dag', 'sympy.physics', 'markupsafe._speedups') if iu.find_spec(m) is None]; assert not missing, missing\""
 
 # Verify GPU access
 render_gid=$(getent group render | cut -d: -f3)
@@ -493,6 +510,8 @@ else
     --group-add "$render_gid" \
     --rm \
     -e HF_TOKEN \
+    -e "HF_HUB_DOWNLOAD_TIMEOUT=${HF_HUB_DOWNLOAD_TIMEOUT}" \
+    -e "HF_HUB_ETAG_TIMEOUT=${HF_HUB_ETAG_TIMEOUT}" \
     -e AWS_ACCESS_KEY_ID \
     -e AWS_SECRET_ACCESS_KEY \
     -e BUILDKITE_PARALLEL_JOB \
@@ -500,10 +519,15 @@ else
     -v "${HF_CACHE}:${HF_MOUNT}" \
     -e "HF_HOME=${HF_MOUNT}" \
     -e "PYTHONPATH=${MYPYTHONPATH}" \
+    -e "TMPDIR=${CONTAINER_TMPDIR}/tmp" \
+    -e "TORCHINDUCTOR_CACHE_DIR=${CONTAINER_CACHE_ROOT}/torchinductor" \
+    -e "TRITON_CACHE_DIR=${CONTAINER_CACHE_ROOT}/triton" \
+    -e "VLLM_CACHE_ROOT=${CONTAINER_CACHE_ROOT}/vllm" \
+    -e "XDG_CACHE_HOME=${CONTAINER_CACHE_ROOT}/xdg" \
     -e "PYTORCH_ROCM_ARCH=" \
     --name "${container_name}" \
     "${image_name}" \
-    /bin/bash -c "${commands}"
+    /bin/bash -c "${CONTAINER_PREFLIGHT} && ${commands}"
 
   exit_code=$?
   handle_pytest_exit "$exit_code"

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

@@ -324,23 +324,6 @@ IMAGE="${IMAGE_TAG_XPU:-${image_name}}"
 
 echo "Using image: ${IMAGE}"
 
-if docker image inspect "${IMAGE}" >/dev/null 2>&1; then
-  echo "Image already exists locally, skipping pull"
-else
-  echo "Image not found locally, waiting for lock..."
-
-  flock /tmp/docker-pull.lock bash -c "
-    if docker image inspect '${IMAGE}' >/dev/null 2>&1; then
-      echo 'Image already pulled by another runner'
-    else
-      echo 'Pulling image...'
-      timeout 900 docker pull '${IMAGE}'
-    fi
-  "
-
-  echo "Pull step completed"
-fi
-
 remove_docker_container() {
   docker rm -f "${container_name}" || true
 }
@@ -357,9 +340,12 @@ export HF_TOKEN ZE_AFFINITY_MASK
 
 {
   flock 9
-  if ! docker image inspect "${IMAGE}" >/dev/null 2>&1; then
-    echo 'Image missing before container creation, pulling again...'
+  if docker image inspect "${IMAGE}" >/dev/null 2>&1; then
+    echo "Image already exists locally, skipping pull"
+  else
+    echo "Image not found locally, pulling image..."
     timeout 900 docker pull "${IMAGE}"
+    echo "Pull step completed"
   fi
 
   docker create \
@@ -372,6 +358,8 @@ export HF_TOKEN ZE_AFFINITY_MASK
     --entrypoint='' \
     -e HF_TOKEN \
     -e ZE_AFFINITY_MASK \
+    -e BUILDKITE_PARALLEL_JOB \
+    -e BUILDKITE_PARALLEL_JOB_COUNT \
     -e CMDS \
     --name "${container_name}" \
     "${IMAGE}" \

.buildkite/scripts/run-rust-frontend-cargo-ci.sh

@@ -110,6 +110,36 @@ install_uv() {
     | env UV_INSTALL_DIR="$CARGO_HOME/bin" sh
 }
 
+setup_pyo3_python() {
+  local python_version="${PYO3_PYTHON_VERSION:-3.12}"
+
+  log_section "Installing Python ${python_version} for PyO3 tests"
+  uv python install "$python_version"
+  PYO3_PYTHON="$(uv python find \
+    --managed-python \
+    --no-project \
+    --resolve-links \
+    "$python_version")"
+  export PYO3_PYTHON
+
+  local python_libdir
+  python_libdir="$("$PYO3_PYTHON" - <<'PY'
+import pathlib
+import sysconfig
+
+libdir = pathlib.Path(sysconfig.get_config_var("LIBDIR"))
+ldlibrary = sysconfig.get_config_var("LDLIBRARY")
+assert sysconfig.get_config_var("Py_ENABLE_SHARED") == 1
+assert ldlibrary
+assert (libdir / ldlibrary).exists(), libdir / ldlibrary
+print(libdir)
+PY
+)"
+
+  export LD_LIBRARY_PATH="${python_libdir}:${LD_LIBRARY_PATH:-}"
+  export LIBRARY_PATH="${python_libdir}:${LIBRARY_PATH:-}"
+}
+
 run_style_clippy() {
   install_cargo_sort
 
@@ -132,6 +162,7 @@ run_style_clippy() {
 
 run_tests() {
   install_uv
+  setup_pyo3_python
   install_cargo_nextest
 
   log_section "Running cargo nextest"

.buildkite/test-amd.yaml

@@ -88,16 +88,16 @@
 #     - Do NOT remove `VLLM_WORKER_MULTIPROC_METHOD=spawn` setting as ROCm requires this for certain models to function.            #
 #   * [Transformers Nightly Models]: Whisper needs `VLLM_WORKER_MULTIPROC_METHOD=spawn` to avoid deadlock.                          #
 #   * [Plugin Tests (2 GPUs)]:                                                                                                      #
-#     - {`pytest -v -s entrypoints/openai/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                                       #
+#     - {`pytest -v -s plugins_tests/test_oot_registration_online.py`}: It needs a clean process                                    #
+#     - {`pytest -v -s plugins_tests/test_oot_registration_offline.py`}: It needs a clean process                                   #
+#     - {`pytest -v -s plugins_tests/lora_resolvers`}: Unit tests for in-tree lora resolver plugins                                 #
 #   * [LoRA TP (Distributed)]:                                                                                                      #
 #     - There is some Tensor Parallelism related processing logic in LoRA that requires multi-GPU testing for validation.           #
 #     - {`pytest -v -s -x lora/test_gptoss_tp.py`}: Disabled for now because MXFP4 backend on non-cuda platform doesn't support     #
 #                                                   LoRA yet.                                                                       #
 #   * [Distributed Tests (NxGPUs)(HW-TAG)]: Don't test llama model here, it seems hf implementation is buggy. See:                  #
 #                                           https://github.com/vllm-project/vllm/pull/5689                                          #
-#   * [Distributed Tests (NxGPUs)(HW-TAG)]: Some old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293     #
+#   * [Distributed Tests (NxGPUs)(HW-TAG)]: Some old E2E tests were removed in https://github.com/vllm-project/vllm/pull/33293      #
 #                                           in favor of new tests in fusions_e2e. We avoid replicating the new jobs in              #
 #                                           this file as it's deprecated.                                                           #
 #                                                                                                                                   #
@@ -315,24 +315,6 @@ steps:
   - pytest -v -s distributed/test_pp_cudagraph.py
   - pytest -v -s distributed/test_pipeline_parallel.py
 
-#----------------------------------------------------------  mi250 · engine  -----------------------------------------------------------#
-
-- label: Engine # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx90anightly, amdmi250]
-  agent_pool: mi250_1
-  optional: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/engine
-  - tests/test_sequence
-  - tests/test_config
-  - tests/test_logger
-  - tests/test_vllm_port
-  commands:
-  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
-
 #-----------------------------------------------------------  mi250 · evals  -----------------------------------------------------------#
 
 - label: Multi-Modal Accuracy Eval (Small Models) # TBD
@@ -416,7 +398,7 @@ steps:
   - tests/kernels/helion/
   - vllm/platforms/rocm.py
   commands:
-  - pip install helion==1.0.0
+  - pip install helion==1.1.0
   - pytest -v -s kernels/helion/
 
 - label: Kernels Mamba Test # TBD
@@ -635,9 +617,9 @@ 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/chat_completion/test_oot_registration.py
-  - pytest -v -s models/test_oot_registration.py
-  - pytest -v -s plugins/lora_resolvers
+  - pytest -v -s plugins_tests/test_oot_registration_online.py # it needs a clean process
+  - pytest -v -s plugins_tests/test_oot_registration_offline.py # it needs a clean process
+  - pytest -v -s plugins_tests/lora_resolvers # unit tests for in-tree lora resolver plugins
 
 #------------------------------------------------------------  mi250 · v1  -------------------------------------------------------------#
 
@@ -822,7 +804,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - ATTENTION_BACKEND=ROCM_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
+  - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
 
 - label: V1 e2e (2 GPUs) # TBD
   timeout_in_minutes: 180
@@ -848,7 +830,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 #-------------------------------------------------------------  mi250 · misc  ------------------------------------------------------------#
 
@@ -908,10 +890,10 @@ steps:
   - vllm/
   - tests/basic_correctness/test_basic_correctness
   - tests/basic_correctness/test_cpu_offload
-  - tests/basic_correctness/test_cumem.py
+  - tests/basic_correctness/test_mem.py
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s basic_correctness/test_cumem.py
+  - pytest -v -s basic_correctness/test_mem.py
   - pytest -v -s basic_correctness/test_basic_correctness.py
   - pytest -v -s basic_correctness/test_cpu_offload.py
 
@@ -1205,7 +1187,39 @@ steps:
 
 #--------------------------------------------------------  mi300 · entrypoints  --------------------------------------------------------#
 
-- label: Entrypoints Integration (API Server 2) # TBD
+- label: Entrypoints Unit Tests # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  fast_check: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/entrypoints
+  - tests/entrypoints/unit_tests
+  - tests/entrypoints/weight_transfer
+  - vllm/platforms/rocm.py
+  commands:
+  - pytest -v -s entrypoints/unit_tests
+  - pytest -v -s entrypoints/weight_transfer
+
+- label: Entrypoints Integration (LLM) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  optional: true
+  fast_check: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/llm
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py --ignore=entrypoints/llm/offline_mode
+  - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
+  - pytest -v -s entrypoints/llm/offline_mode # Needs to avoid interference with other tests
+
+- label: Entrypoints Integration (API Server) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
@@ -1221,7 +1235,7 @@ steps:
   - pytest -v -s entrypoints/serve --ignore=entrypoints/serve/dev/rpc
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc
 
-- label: Entrypoints Integration (API Server openai - Part 1) # TBD
+- label: Entrypoints Integration (API Server OpenAI - Part 1) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
@@ -1235,90 +1249,45 @@ steps:
   - tests/entrypoints/test_chat_utils
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
+  - pytest -v -s entrypoints/openai/ --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/correctness
 
-- label: Entrypoints Integration (API Server openai - Part 2) # TBD
+- label: Entrypoints Integration (API Server OpenAI - Part 2) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
-  optional: true
   fast_check: true
   torch_nightly: true
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/openai
   - tests/entrypoints/test_chat_utils
-  - tests/entrypoints/generate
-  - tests/tool_use
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/openai/chat_completion
   - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
-  - pytest -v -s entrypoints/test_chat_utils.py
-  - pytest -v -s entrypoints/generate
+
+- label: Entrypoints Integration (API Server Generate) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  optional: true
+  fast_check: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/tool_use
+  - tests/entrypoints/tool_parsers
+  - tests/entrypoints/anthropic
+  - tests/entrypoints/generate
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s tool_use
-
-- label: Entrypoints Integration (API Server openai - Part 3) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  optional: true
-  fast_check: true
-  torch_nightly: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/test_multi_api_servers.py
-
-- label: Entrypoints Integration (Speech to Text) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  fast_check: true
-  torch_nightly: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/speech_to_text
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/speech_to_text
-
-- label: Entrypoints Integration (LLM) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  optional: true
-  fast_check: true
-  torch_nightly: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/llm
-  - tests/entrypoints/offline_mode
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
-  - pytest -v -s entrypoints/llm/test_generate.py
-  - pytest -v -s entrypoints/offline_mode
-
-- label: Entrypoints Integration (Pooling) # TBD
-  timeout_in_minutes: 180
-  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
-  agent_pool: mi300_1
-  fast_check: true
-  torch_nightly: true
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/pooling
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/pooling
+  - pytest -v -s entrypoints/tool_parsers
+  - pytest -v -s entrypoints/generate
+  - pytest -v -s entrypoints/anthropic
 
 - label: Entrypoints Integration (Responses API) # TBD
   timeout_in_minutes: 180
@@ -1334,29 +1303,57 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/openai/responses
 
-- label: Entrypoints Unit Tests # TBD
+- label: Entrypoints Integration (Speech to Text) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
   fast_check: true
+  torch_nightly: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
-  - vllm/entrypoints
-  - tests/entrypoints/
-  - vllm/platforms/rocm.py
+  - vllm/
+  - tests/entrypoints/speech_to_text
   commands:
-  - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/offline_mode --ignore=entrypoints/openai --ignore=entrypoints/serve --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling --ignore=entrypoints/speech_to_text --ignore=tests/entrypoints/generate
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/speech_to_text
+
+- label: Entrypoints Integration (Multimodal)
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  fast_check: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/multimodal
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/multimodal
+
+- label: Entrypoints Integration (Pooling) # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  fast_check: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/pooling
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/pooling
 
 - label: OpenAI API correctness # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - csrc/
   - vllm/entrypoints/openai/
-  - vllm/model_executor/models/whisper.py
   - vllm/model_executor/layers/
   - vllm/v1/attention/backends/
   - vllm/v1/attention/selector.py
@@ -1410,6 +1407,7 @@ steps:
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
+  optional: true
   working_dir: "/vllm-workspace/.buildkite/lm-eval-harness"
   source_file_dependencies:
   - csrc/
@@ -2184,10 +2182,72 @@ steps:
   commands:
   - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
 
+- label: Speculators Correctness # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/config/speculative.py
+  - vllm/engine/arg_utils.py
+  - vllm/transformers_utils/config.py
+  - vllm/transformers_utils/configs/speculators/
+  - vllm/v1/spec_decode/
+  - vllm/v1/worker/gpu/spec_decode/
+  - vllm/v1/worker/gpu_model_runner.py
+  - vllm/v1/sample/
+  - vllm/v1/attention/backends/
+  - vllm/v1/attention/selector.py
+  - vllm/model_executor/model_loader/
+  - vllm/model_executor/layers/
+  - vllm/model_executor/models/llama_eagle3.py
+  - vllm/model_executor/models/qwen3.py
+  - vllm/model_executor/models/qwen3_dflash.py
+  - vllm/model_executor/models/registry.py
+  - vllm/_aiter_ops.py
+  - tests/evals/gsm8k/
+  - tests/v1/spec_decode/test_speculators_correctness.py
+  - vllm/platforms/rocm.py
+  commands:
+  - export VLLM_ALLOW_INSECURE_SERIALIZATION=1
+  - pytest -v -s v1/spec_decode/test_speculators_correctness.py -m slow_test
+
+- label: Extract Hidden States Integration # TBD
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
+  agent_pool: mi300_1
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/config/speculative.py
+  - vllm/distributed/kv_transfer/kv_connector/
+  - vllm/model_executor/layers/attention/
+  - vllm/model_executor/layers/mamba/
+  - vllm/model_executor/model_loader/
+  - vllm/model_executor/models/extract_hidden_states.py
+  - vllm/model_executor/models/llama.py
+  - vllm/model_executor/models/qwen3_5.py
+  - vllm/model_executor/models/qwen3_next.py
+  - vllm/model_executor/models/registry.py
+  - vllm/transformers_utils/configs/extract_hidden_states.py
+  - vllm/transformers_utils/configs/qwen3_5.py
+  - vllm/v1/attention/backends/
+  - vllm/v1/attention/selector.py
+  - vllm/v1/kv_cache_interface.py
+  - vllm/v1/spec_decode/extract_hidden_states.py
+  - vllm/v1/worker/gpu_model_runner.py
+  - vllm/_aiter_ops.py
+  - tests/v1/kv_connector/extract_hidden_states_integration/
+  - vllm/platforms/rocm.py
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s v1/kv_connector/extract_hidden_states_integration
+
 - label: V1 attention (H100-MI300) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
   agent_pool: mi300_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/config/attention.py
@@ -2346,7 +2406,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - CROSS_LAYERS_BLOCKS=True ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - CROSS_LAYERS_BLOCKS=True ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: Distributed DP Tests (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -2382,7 +2442,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -2396,7 +2456,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - DP_EP=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: Hybrid SSM NixlConnector PD accuracy tests (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -2410,7 +2470,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - HYBRID_SSM=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - HYBRID_SSM=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: V1 e2e (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -2546,6 +2606,7 @@ steps:
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
   agent_pool: mi325_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -2555,7 +2616,7 @@ steps:
   - tests/test_logger
   - tests/test_vllm_port
   commands:
-  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
+  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py test_jit_monitor.py
 
 #-----------------------------------------------------------  mi325 · evals  -----------------------------------------------------------#
 
@@ -2637,6 +2698,7 @@ steps:
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
   agent_pool: mi325_1
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -2736,7 +2798,7 @@ steps:
 
 #--------------------------------------------------------  mi355 · entrypoints  --------------------------------------------------------#
 
-- label: Entrypoints Integration (API Server 2) # TBD
+- label: Entrypoints Integration (API Server) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   agent_pool: mi355_1
@@ -2752,7 +2814,7 @@ steps:
   - pytest -v -s entrypoints/serve --ignore=entrypoints/serve/dev/rpc
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/serve/dev/rpc
 
-- label: Entrypoints Integration (API Server openai - Part 1) # TBD
+- label: Entrypoints Integration (API Server OpenAI - Part 1) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   agent_pool: mi355_1
@@ -2766,9 +2828,9 @@ steps:
   - tests/entrypoints/test_chat_utils
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/correctness
 
-- label: Entrypoints Integration (API Server openai - Part 2) # TBD
+- label: Entrypoints Integration (API Server OpenAI - Part 2) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   agent_pool: mi355_1
@@ -2780,29 +2842,31 @@ steps:
   - vllm/
   - tests/entrypoints/openai
   - tests/entrypoints/test_chat_utils
-  - tests/entrypoints/generate
-  - tests/tool_use
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/openai/chat_completion
   - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
-  - pytest -v -s entrypoints/test_chat_utils.py
-  - pytest -v -s entrypoints/generate
-  - pytest -v -s tool_use
 
-- label: Entrypoints Integration (API Server openai - Part 3) # TBD
+- label: Entrypoints Integration (API Server Generate) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
   agent_pool: mi355_1
   fast_check: true
   torch_nightly: true
+  optional: true
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
+  - tests/tool_use
+  - tests/entrypoints/tool_parsers
+  - tests/entrypoints/anthropic
+  - tests/entrypoints/generate
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/test_multi_api_servers.py
+  - pytest -v -s tool_use
+  - pytest -v -s entrypoints/tool_parsers
+  - pytest -v -s entrypoints/generate
+  - pytest -v -s entrypoints/anthropic
 
 - label: Entrypoints Integration (Speech to Text) # TBD
   timeout_in_minutes: 180
@@ -2818,6 +2882,20 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/speech_to_text
 
+- label: Entrypoints Integration (Multimodal)
+  timeout_in_minutes: 180
+  mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi355]
+  agent_pool: mi355_1
+  fast_check: true
+  torch_nightly: true
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/multimodal
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/multimodal
+
 - label: Entrypoints Integration (Pooling) # TBD
   timeout_in_minutes: 180
   mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
@@ -2868,7 +2946,6 @@ steps:
   - vllm/model_executor/models/qwen3_5_mtp.py
   - vllm/transformers_utils/configs/qwen3_5.py
   - vllm/transformers_utils/configs/qwen3_5_moe.py
-  - vllm/model_executor/models/qwen.py
   - vllm/model_executor/models/qwen2.py
   - vllm/model_executor/models/qwen3.py
   - vllm/model_executor/models/qwen3_next.py
@@ -2995,7 +3072,7 @@ steps:
   - vllm/_aiter_ops.py
   commands:
   - rocm-smi
-  - python3 examples/basic/offline_inference/chat.py
+  - python3 examples/basic/offline_inference/chat.py --attention-backend TRITON_ATTN
   - pytest -v -s tests/kernels/attention/test_attention_selector.py
 
 - label: Kernels Attention Test %N # TBD
@@ -3106,7 +3183,6 @@ steps:
   - vllm/model_executor/models/qwen3_5_mtp.py
   - vllm/transformers_utils/configs/qwen3_5.py
   - vllm/transformers_utils/configs/qwen3_5_moe.py
-  - vllm/model_executor/models/qwen.py
   - vllm/model_executor/models/qwen2.py
   - vllm/model_executor/models/qwen3.py
   - vllm/model_executor/models/qwen3_next.py
@@ -3359,7 +3435,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - ATTENTION_BACKEND=ROCM_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
+  - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/spec_decode_acceptance_test.sh
 
 - label: Distributed NixlConnector PD accuracy (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -3374,7 +3450,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 - label: DP EP Distributed NixlConnector PD accuracy tests (4 GPUs) # TBD
   timeout_in_minutes: 180
@@ -3389,7 +3465,7 @@ steps:
   - vllm/platforms/rocm.py
   commands:
   - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-  - DP_EP=1 ROCM_ATTN=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
+  - DP_EP=1 ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
 #------------------------------------------------------  mi355 · weight_loading  -------------------------------------------------------#
 

.buildkite/test_areas/attention.yaml

@@ -2,8 +2,8 @@ group: Attention
 depends_on: 
   - image-build
 steps:
-- label: V1 attention (H100)
-  key: v1-attention-h100
+- label: V1 attention (H100-MI300)
+  key: v1-attention-h100-mi300
   timeout_in_minutes: 30
   device: h100
   source_file_dependencies:
@@ -13,6 +13,20 @@ steps:
     - tests/v1/attention
   commands:
     - pytest -v -s v1/attention
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 70
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/config/attention.py
+      - vllm/model_executor/layers/attention
+      - vllm/v1/attention
+      - tests/v1/attention
+      - vllm/_aiter_ops.py
+      - vllm/envs.py
+      - vllm/platforms/rocm.py
 
 - label: V1 attention (B200)
   key: v1-attention-b200

.buildkite/test_areas/basic_correctness.yaml

@@ -10,9 +10,9 @@ steps:
   - vllm/
   - tests/basic_correctness/test_basic_correctness
   - tests/basic_correctness/test_cpu_offload
-  - tests/basic_correctness/test_cumem.py
+  - tests/basic_correctness/test_mem.py
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s basic_correctness/test_cumem.py
+  - pytest -v -s basic_correctness/test_mem.py
   - pytest -v -s basic_correctness/test_basic_correctness.py
   - pytest -v -s basic_correctness/test_cpu_offload.py

.buildkite/test_areas/disaggregated.yaml

@@ -61,6 +61,20 @@ steps:
     - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
 
+- label: Hybrid SSM NixlConnector PD prefix cache test (2 GPUs)
+  key: hybrid-ssm-nixlconnector-pd-prefix-cache-2-gpus
+  timeout_in_minutes: 25
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/distributed/kv_transfer/kv_connector/v1/nixl/
+    - vllm/v1/core/sched/
+    - vllm/v1/core/kv_cache_coordinator.py
+    - tests/v1/kv_connector/nixl_integration/
+  commands:
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+    - bash v1/kv_connector/nixl_integration/run_mamba_prefix_cache_test.sh
+
 - label: MultiConnector (Nixl+Offloading) PD accuracy (2 GPUs)
   key: multiconnector-nixl-offloading-pd-accuracy-2-gpus
   timeout_in_minutes: 30

.buildkite/test_areas/engine.yaml

@@ -26,6 +26,12 @@ steps:
   - tests/test_jit_monitor.py
   commands:
   - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py test_jit_monitor.py
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 60
+      depends_on:
+      - image-build-amd
 
 - label: Engine (1 GPU)
   key: engine-1-gpu

.buildkite/test_areas/entrypoints.yaml

@@ -8,10 +8,11 @@ steps:
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/entrypoints
-  - tests/entrypoints/
+  - tests/entrypoints/unit_tests
+  - tests/entrypoints/weight_transfer
   commands:
-  - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/offline_mode --ignore=entrypoints/openai --ignore=entrypoints/serve --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling --ignore=entrypoints/speech_to_text --ignore=tests/entrypoints/generate
+  - pytest -v -s entrypoints/unit_tests
+  - pytest -v -s entrypoints/weight_transfer
 
 - label: Entrypoints Integration (LLM)
   key: entrypoints-integration-llm
@@ -20,84 +21,20 @@ steps:
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/llm
-  - tests/entrypoints/offline_mode
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py --ignore=entrypoints/llm/offline_mode
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+  - pytest -v -s entrypoints/llm/offline_mode # Needs to avoid interference with other tests
   mirror:
     amd:
       device: mi325_1
-      soft_fail: true
       depends_on:
       - image-build-amd
 
-- label: Entrypoints Integration (API Server openai - Part 1)
-  key: entrypoints-integration-api-server-openai-part-1
-  timeout_in_minutes: 50
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai/chat_completion --ignore=entrypoints/openai/chat_completion/test_oot_registration.py
-  mirror:
-    amd:
-      device: mi325_1
-      soft_fail: true
-      timeout_in_minutes: 80
-      depends_on:
-      - image-build-amd
-
-- label: Entrypoints Integration (API Server openai - Part 2)
-  key: entrypoints-integration-api-server-openai-part-2
-  timeout_in_minutes: 50
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
-  - tests/entrypoints/generate
-  - tests/tool_use
-  commands:
-  - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
-  - pytest -v -s entrypoints/test_chat_utils.py
-  - pytest -v -s entrypoints/generate
-  - pytest -v -s tool_use
-  mirror:
-    amd:
-      device: mi325_1
-      soft_fail: true
-      timeout_in_minutes: 60
-      depends_on:
-      - image-build-amd
-
-- label: Entrypoints Integration (API Server openai - Part 3)
-  key: entrypoints-integration-api-server-openai-part-3
-  timeout_in_minutes: 50
-  device: h200_18gb
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai
-  - tests/entrypoints/test_chat_utils
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/tool_parsers/ --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/test_multi_api_servers.py
-  mirror:
-    amd:
-      device: mi325_1
-      soft_fail: true
-      timeout_in_minutes: 60
-      depends_on:
-      - image-build-amd
-
-- label: Entrypoints Integration (API Server 2)
+- label: Entrypoints Integration (API Server)
+  key: entrypoints-integration-api-server
   device: h200_35gb
-  key: entrypoints-integration-api-server-2
   timeout_in_minutes: 130
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
@@ -110,10 +47,78 @@ steps:
   mirror:
     amd:
       device: mi325_1
-      soft_fail: true
       depends_on:
       - image-build-amd
 
+- label: Entrypoints Integration (API Server OpenAI - Part 1)
+  key: entrypoints-integration-api-server-openai-part-1
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/completion --ignore=entrypoints/openai/chat_completion --ignore=entrypoints/openai/responses --ignore=entrypoints/openai/correctness
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 80
+      depends_on:
+      - image-build-amd
+
+- label: Entrypoints Integration (API Server OpenAI - Part 2)
+  key: entrypoints-integration-api-server-openai-part-2
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai
+  - tests/entrypoints/test_chat_utils
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/openai/chat_completion
+  - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 80
+      depends_on:
+      - image-build-amd
+
+- label: Entrypoints Integration (API Server Generate)
+  key: entrypoints-integration-api-server-generate
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/tool_use
+  - tests/entrypoints/tool_parsers
+  - tests/entrypoints/anthropic
+  - tests/entrypoints/generate
+  commands:
+  - pytest -v -s tool_use
+  - pytest -v -s entrypoints/tool_parsers
+  - pytest -v -s entrypoints/generate
+  - pytest -v -s entrypoints/anthropic
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 60
+      depends_on:
+      - image-build-amd
+
+- label: Entrypoints Integration (Responses API)
+  key: entrypoints-integration-responses-api
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/openai/responses
+  commands:
+  - pytest -v -s entrypoints/openai/responses
+
 - label: Entrypoints Integration (Speech to Text)
   device: h200_35gb
   key: entrypoints-integration-speech_to_text
@@ -126,6 +131,18 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/speech_to_text
 
+- label: Entrypoints Integration (Multimodal)
+  device: h200_35gb
+  key: entrypoints-integration-multimodal
+  timeout_in_minutes: 50
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/
+  - tests/entrypoints/multimodal
+  commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+  - pytest -v -s entrypoints/multimodal
+
 - label: Entrypoints Integration (Pooling)
   key: entrypoints-integration-pooling
   timeout_in_minutes: 50
@@ -137,16 +154,6 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/pooling
 
-- label: Entrypoints Integration (Responses API)
-  key: entrypoints-integration-responses-api
-  timeout_in_minutes: 50
-  working_dir: "/vllm-workspace/tests"
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/openai/responses
-  commands:
-  - pytest -v -s entrypoints/openai/responses
-
 - label: OpenAI API Correctness
   key: openai-api-correctness
   timeout_in_minutes: 30
@@ -156,3 +163,20 @@ steps:
   - vllm/entrypoints/openai/
   commands: # LMEval
   - pytest -s entrypoints/openai/correctness/
+  mirror:
+    amd:
+      device: mi325_1
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - csrc/
+      - vllm/entrypoints/openai/
+      - vllm/model_executor/layers/
+      - vllm/v1/attention/backends/
+      - vllm/v1/attention/selector.py
+      - vllm/_aiter_ops.py
+      - vllm/platforms/rocm.py
+      - vllm/model_executor/model_loader/
+      commands:
+      - bash ../tools/install_torchcodec_rocm.sh || exit 1
+      - pytest -s entrypoints/openai/correctness/

.buildkite/test_areas/kernels.yaml

@@ -21,8 +21,9 @@ steps:
   - csrc/
   - tests/kernels/core
   - tests/kernels/test_concat_mla_q.py
+  - tests/kernels/test_fused_qk_norm_rope_gate.py
   commands:
-    - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py  kernels/test_concat_mla_q.py
+    - pytest -v -s kernels/core --ignore=kernels/core/test_minimax_reduce_rms.py  kernels/test_concat_mla_q.py kernels/test_fused_qk_norm_rope_gate.py
 
 - label: Kernels MiniMax Reduce RMS Test (2 GPUs)
   key: kernels-minimax-reduce-rms-test-2-gpus
@@ -74,6 +75,19 @@ steps:
     - pytest -v -s kernels/attention --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 2
 
+- label: Kernels Attention DiffKV Test (H100)
+  key: kernels-attention-diffkv-test-h100
+  timeout_in_minutes: 20
+  device: h100
+  num_devices: 1
+  source_file_dependencies:
+  - vllm/v1/attention/ops/triton_unified_attention_diffkv.py
+  - vllm/v1/attention/backends/triton_attn_diffkv.py
+  - vllm/v1/attention/backends/flash_attn_diffkv.py
+  - tests/kernels/attention/test_triton_unified_attention_diffkv.py
+  commands:
+    - pytest -v -s kernels/attention/test_triton_unified_attention_diffkv.py
+
 - label: Kernels Quantization Test %N
   key: kernels-quantization-test
   timeout_in_minutes: 90
@@ -223,7 +237,7 @@ steps:
   - vllm/utils/import_utils.py
   - tests/kernels/helion/
   commands:
-    - pip install helion==1.0.0
+    - pip install helion==1.1.0
     - pytest -v -s kernels/helion/
 
  
@@ -299,3 +313,4 @@ steps:
   - vllm/config
   commands:
     - pytest -v -s kernels/moe/test_moe_layer.py
+    - pytest -v -s kernels/moe/test_deepep_v2_moe.py

.buildkite/test_areas/lm_eval.yaml

@@ -12,6 +12,21 @@ steps:
   autorun_on_main: true
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small.txt
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 55
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - csrc/
+      - vllm/model_executor/layers/quantization
+      - vllm/model_executor/models/
+      - vllm/model_executor/model_loader/
+      - vllm/v1/attention/backends/
+      - vllm/v1/attention/selector.py
+      - vllm/_aiter_ops.py
+      - vllm/platforms/rocm.py
 
 # - label: LM Eval Large Models (4 GPUs)(A100)
 #   device: a100

.buildkite/test_areas/misc.yaml

@@ -138,11 +138,26 @@ steps:
     - vllm/v1/spec_decode/extract_hidden_states.py
     - vllm/model_executor/models/extract_hidden_states.py
     - vllm/transformers_utils/configs/extract_hidden_states.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/example_hidden_states_connector.py
     - tests/v1/kv_connector/extract_hidden_states_integration
   commands:
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
     - pytest -v -s v1/kv_connector/extract_hidden_states_integration
 
+- label: Extract Hidden States Integration (2 GPUs)
+  key: extract-hidden-states-integration-2-gpus
+  timeout_in_minutes: 20
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/v1/spec_decode/extract_hidden_states.py
+    - vllm/model_executor/models/extract_hidden_states.py
+    - vllm/transformers_utils/configs/extract_hidden_states.py
+    - vllm/distributed/kv_transfer/kv_connector/v1/example_hidden_states_connector.py
+    - tests/v1/kv_connector/extract_hidden_states_integration
+  commands:
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -v -s -m 'distributed' v1/kv_connector/extract_hidden_states_integration
+
 - label: Regression
   key: regression
   timeout_in_minutes: 20
@@ -293,6 +308,7 @@ steps:
   - vllm/transformers_utils/
   - vllm/utils/
   - vllm/v1/
+  - tests/test_envs.py
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/test_pooling_params.py
@@ -300,24 +316,25 @@ steps:
   - tests/multimodal
   - tests/renderers
   - tests/standalone_tests/lazy_imports.py
-  - tests/tokenizers_
   - tests/reasoning
   - tests/tool_parsers
+  - tests/tokenizers_
   - tests/parser
   - tests/transformers_utils
   - tests/config
   device: cpu-small
   commands:
   - python3 standalone_tests/lazy_imports.py
+  - pytest -v -s test_envs.py
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s test_pooling_params.py
   - pytest -v -s test_ray_env.py
   - pytest -v -s -m 'cpu_test' multimodal
   - pytest -v -s renderers
-  - pytest -v -s tokenizers_
   - pytest -v -s reasoning --ignore=reasoning/test_seedoss_reasoning_parser.py --ignore=reasoning/test_glm4_moe_reasoning_parser.py
   - pytest -v -s tool_parsers
+  - pytest -v -s tokenizers_
   - pytest -v -s parser
   - pytest -v -s transformers_utils
   - pytest -v -s config

.buildkite/test_areas/models_language.yaml

@@ -51,6 +51,7 @@ steps:
     torch_nightly: {}
     amd:
       device: mi325_1
+      timeout_in_minutes: 90
       depends_on:
       - image-build-amd
       commands:

.buildkite/test_areas/models_multimodal.yaml

@@ -153,3 +153,12 @@ steps:
   - tests/models/multimodal/pooling
   commands:
     - pytest -v -s models/multimodal/pooling -m 'not core_model'
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 60
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/
+      - tests/models/multimodal/pooling

.buildkite/test_areas/plugins.yaml

@@ -37,6 +37,20 @@ 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/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
+  - pytest -v -s plugins_tests/test_oot_registration_online.py # it needs a clean process
+  - pytest -v -s plugins_tests/test_oot_registration_offline.py # it needs a clean process
+  - pytest -v -s plugins_tests/lora_resolvers # unit tests for in-tree lora resolver plugins
+
+
+- label: GGUF Plugin
+  key: gguf-plugin
+  device: h200_18gb
+  timeout_in_minutes: 30
+  soft_fail: true
+  optional: true
+  source_file_dependencies:
+  - vllm/model_executor/layers/quantization
+  - tests/plugins_tests/test_gguf_plugin.py
+  commands:
+    - pip install "vllm-gguf-plugin >= 0.0.2"
+    - pytest -v -s plugins_tests/gguf

.buildkite/test_areas/quantization.yaml

@@ -21,6 +21,18 @@ steps:
   - uv pip install --system conch-triton-kernels
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
 
+- label: Quantized Fusions
+  key: quantized-fusions
+  timeout_in_minutes: 30
+  source_file_dependencies:
+  - tests/fusion
+  - vllm/model_executor/layers/fusion
+  - vllm/model_executor/kernels/linear
+  - vllm/model_executor/layers/quantization/compressed_tensors
+  - vllm/model_executor/layers/quantization/modelopt.py
+  commands:
+    - pytest -v -s fusion/
+
 - label: Quantized MoE Test (B200)
   key: quantized-moe-test-b200
   timeout_in_minutes: 60

.buildkite/test_areas/spec_decode.yaml

@@ -37,6 +37,21 @@ steps:
     - tests/v1/e2e/spec_decode/
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 65
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/spec_decode/
+      - vllm/v1/worker/gpu/spec_decode/
+      - vllm/model_executor/model_loader/
+      - vllm/v1/sample/
+      - vllm/model_executor/layers/
+      - vllm/transformers_utils/configs/speculators/
+      - tests/v1/e2e/spec_decode/
+      - vllm/platforms/rocm.py
 
 - label: Spec Decode Speculators + MTP Nightly B200
   key: spec-decode-speculators-mtp-nightly-b200
@@ -61,6 +76,20 @@ steps:
     - tests/v1/e2e/spec_decode/
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 65
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/spec_decode/
+      - vllm/v1/worker/gpu/spec_decode/
+      - vllm/model_executor/model_loader/
+      - vllm/v1/sample/
+      - vllm/model_executor/layers/
+      - tests/v1/e2e/spec_decode/
+      - vllm/platforms/rocm.py
 
 - label: Spec Decode Draft Model
   key: spec-decode-draft-model
@@ -72,6 +101,20 @@ steps:
     - tests/v1/e2e/spec_decode/
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
+  mirror:
+    amd:
+      device: mi325_1
+      timeout_in_minutes: 50
+      depends_on:
+      - image-build-amd
+      source_file_dependencies:
+      - vllm/v1/spec_decode/
+      - vllm/v1/worker/gpu/spec_decode/
+      - vllm/model_executor/model_loader/
+      - vllm/v1/sample/
+      - vllm/model_executor/layers/
+      - tests/v1/e2e/spec_decode/
+      - vllm/platforms/rocm.py
 
 - label: Spec Decode Draft Model Nightly B200
   key: spec-decode-draft-model-nightly-b200

.dockerignore

@@ -33,10 +33,3 @@ share/python-wheels/
 *.egg
 MANIFEST
 rust/target/
-# Not needed in Docker builds
-docs/
-.github/
-.pre-commit-config.yaml
-.clang-format
-.gitattributes
-format.sh

.github/CODEOWNERS

@@ -23,9 +23,14 @@
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
 # so spam a lot of people
-/vllm/config @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor @yewentao256 @ProExpertProg
+/vllm/config @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @yewentao256 @ProExpertProg
 /vllm/config/cache.py @heheda12345
 
+# Config utils
+/vllm/config/utils.py @hmellor
+/vllm/engine/arg_utils.py @hmellor
+/vllm/utils/argparse_utils.py
+
 # Entrypoints
 /vllm/entrypoints/anthropic @mgoin @DarkLight1337
 /vllm/entrypoints/cli @hmellor @mgoin @DarkLight1337 @russellb
@@ -34,10 +39,11 @@
 /vllm/entrypoints/speech_to_text/realtime @njhill
 /vllm/entrypoints/speech_to_text @NickLucche
 /vllm/entrypoints/pooling @noooop
-/vllm/entrypoints/sagemaker @DarkLight1337
+/vllm/entrypoints/serve/sagemaker @DarkLight1337
 /vllm/entrypoints/serve @njhill
 /vllm/entrypoints/*.py @njhill
 /vllm/entrypoints/chat_utils.py @DarkLight1337
+/vllm/entrypoints/offline_utils.py @DarkLight1337
 /vllm/entrypoints/llm.py @DarkLight1337
 
 # Rust Frontend
@@ -74,7 +80,7 @@
 /vllm/v1/worker/kv_connector_model_runner_mixin.py @orozery @NickLucche
 
 # Model runner V2
-/vllm/v1/worker/gpu @WoosukKwon @njhill
+/vllm/v1/worker/gpu @WoosukKwon @njhill @yewentao256
 /vllm/v1/worker/gpu/kv_connector.py @orozery
 
 # CI & building

.github/dependabot.yml

@@ -21,7 +21,6 @@ updates:
       - dependency-name: "torchvision"
       - dependency-name: "xformers"
       - dependency-name: "lm-format-enforcer"
-      - dependency-name: "gguf"
       - dependency-name: "compressed-tensors"
       - dependency-name: "ray[cgraph]" # Ray Compiled Graph
       - dependency-name: "lm-eval"

.github/mergify.yml

@@ -21,6 +21,9 @@ pull_request_rules:
     - check-failure=pre-commit
     - -closed
     - -draft
+    - or:
+      - label=ready
+      - label=verified
   actions:
     comment:
       message: |
@@ -36,18 +39,6 @@ pull_request_rules:
 
         For future commits, `pre-commit` will run automatically on changed files before each commit.
 
-        > [!TIP]
-        > <details>
-        > <summary>Is <code>mypy</code> failing?</summary>
-        > <br/>
-        > <code>mypy</code> is run differently in CI. If the failure is related to this check, please use the following command to run it locally:
-        >
-        > ```bash
-        > # For mypy (substitute "3.10" with the failing version if needed)
-        > pre-commit run --hook-stage manual mypy-3.10
-        > ```
-        > </details>
-
 - name: comment-dco-failure
   description: Comment on PR when DCO check fails
   conditions:
@@ -153,12 +144,12 @@ pull_request_rules:
     - label != stale
     - or:
       - files~=^examples/.*mistral.*\.py
-      - files~=^tests/.*mistral.*\.py
-      - files~=^vllm/model_executor/models/.*mistral.*\.py
+      - files~=^tests/.*(?:mistral|voxtral|mixtral|pixtral).*\.py
+      - files~=^vllm/model_executor/models/.*(?:mistral|voxtral|mixtral|pixtral).*\.py
       - files~=^vllm/reasoning/.*mistral.*\.py
       - files~=^vllm/tool_parsers/.*mistral.*\.py
-      - files~=^vllm/transformers_utils/.*mistral.*\.py
-      - title~=(?i)Mistral
+      - files~=^vllm/transformers_utils/.*(?:mistral|voxtral|pixtral).*\.py
+      - title~=(?i)(?:mistral|ministral|voxtral|mixtral|pixtral)
   actions:
     label:
       add:

.github/workflows/scripts/build.sh

@@ -9,7 +9,7 @@ PATH=${cuda_home}/bin:$PATH
 LD_LIBRARY_PATH=${cuda_home}/lib64:$LD_LIBRARY_PATH
 
 # Install requirements
-if [ "$(echo $2 | cut -d. -f1)" = "12" ]; then
+if [ "$(echo "$2" | cut -d. -f1)" = "12" ]; then
     sed -i 's/^nvidia-cutlass-dsl\[cu13\]>=/nvidia-cutlass-dsl>=/' requirements/cuda.txt
 fi
 $python_executable -m pip install -r requirements/build/cuda.txt -r requirements/cuda.txt
@@ -17,7 +17,10 @@ $python_executable -m pip install -r requirements/build/cuda.txt -r requirements
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1
 # Make sure release wheels are built for the following architectures
-export TORCH_CUDA_ARCH_LIST="7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX"
+# Do not add +PTX here: vLLM filters torch's top-level PTX flag when it
+# converts global gencode flags into per-kernel arch lists. If a specific
+# kernel needs PTX, add +PTX to that kernel's CMake arch list instead.
+export TORCH_CUDA_ARCH_LIST="7.5 8.0 8.6 8.9 9.0 10.0 12.0"
 
 bash tools/check_repo.sh
 

.github/workflows/stale.yml

@@ -15,7 +15,7 @@ jobs:
       actions: write
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/stale@997185467fa4f803885201cee163a9f38240193d # v10.1.1
+      - uses: actions/stale@eb5cf3af3ac0a1aa4c9c45633dd1ae542a27a899 # v10.3.0
         with:
           # Increasing this value ensures that changes to this workflow
           # propagate to all issues and PRs in days rather than months

.gitignore

@@ -15,6 +15,9 @@ vllm/third_party/flashmla/flash_mla_interface.py
 # DeepGEMM vendored package built from source
 vllm/third_party/deep_gemm/
 
+# fmha_sm100 vendored package built from source
+vllm/third_party/fmha_sm100/
+
 # triton jit
 .triton
 
@@ -233,7 +236,7 @@ actionlint
 shellcheck*/
 
 # Ignore moe/marlin_moe gen code
-csrc/moe/marlin_moe_wna16/kernel_*
+csrc/libtorch_stable/moe/marlin_moe_wna16/kernel_*
 
 # Ignore ep_kernels_workspace folder
 ep_kernels_workspace/

.pre-commit-config.yaml

@@ -21,7 +21,7 @@ repos:
   rev: v21.1.2
   hooks:
   - id: clang-format
-    exclude: 'csrc/(moe/topk_softmax_kernels.cu|libtorch_stable/quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
+    exclude: 'csrc/libtorch_stable/moe/topk_softmax_kernels.cu|vllm/third_party/.*'
     types_or: [c++, cuda]
     args: [--style=file, --verbose]
 - repo: https://github.com/DavidAnson/markdownlint-cli2
@@ -148,33 +148,27 @@ repos:
     language: python
     entry: python tools/pre_commit/generate_nightly_torch_test.py
     files: ^requirements/test/cuda\.(in|txt)$
-  - id: mypy-local
-    name: Run mypy locally for lowest supported Python version
-    entry: python tools/pre_commit/mypy.py 0 "3.10"
-    stages: [pre-commit] # Don't run in CI
+  - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
+    name: Run mypy for Python 3.10
+    entry: python tools/pre_commit/mypy.py "3.10"
     <<: &mypy_common
       language: python
       types_or: [python, pyi]
       require_serial: true
-      additional_dependencies: ["mypy[faster-cache]==1.19.1", regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
-  - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
-    name: Run mypy for Python 3.10
-    entry: python tools/pre_commit/mypy.py 1 "3.10"
-    <<: *mypy_common
-    stages: [manual] # Only run in CI
+      additional_dependencies: ["mypy==1.20.2", regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
   - id: mypy-3.11 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.11
-    entry: python tools/pre_commit/mypy.py 1 "3.11"
+    entry: python tools/pre_commit/mypy.py "3.11"
     <<: *mypy_common
     stages: [manual] # Only run in CI
   - id: mypy-3.12 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.12
-    entry: python tools/pre_commit/mypy.py 1 "3.12"
+    entry: python tools/pre_commit/mypy.py "3.12"
     <<: *mypy_common
     stages: [manual] # Only run in CI
   - id: mypy-3.13 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.13
-    entry: python tools/pre_commit/mypy.py 1 "3.13"
+    entry: python tools/pre_commit/mypy.py "3.13"
     <<: *mypy_common
     stages: [manual] # Only run in CI
   - id: shellcheck

AGENTS.md

@@ -98,11 +98,33 @@ pre-commit run --all-files
 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
+pre-commit run mypy-3.12 --all-files --hook-stage manual
 ```
 
 The line length limit for Python code is 88 characters. If you are not sure, use pre-commit to check.
 
+Use [Google-style docstrings](https://google.github.io/styleguide/pyguide.html#38-comments-and-docstrings) (`Args:`/`Returns:`/`Raises:` sections), not reStructuredText/Sphinx fields (`:param:`, `:return:`, `:rtype:`).
+
+### Coding style guidelines
+
+Follow these rules for all code changes in this repository:
+
+- Try to match existing code style.
+- Code should be self-documenting and self-explanatory.
+- Keep comments and docstrings minimal and concise.
+- Assume the reader is familiar with vLLM.
+
+### Diagnosing CI failures
+
+Buildkite logs are public; no login needed. Details: [docs/contributing/ci/failures.md](docs/contributing/ci/failures.md).
+
+```bash
+# All failed-job logs for a PR's latest build (current branch's PR if omitted):
+.buildkite/scripts/ci-fetch-log.sh --pr <PR>
+# Any Buildkite build or job URL also works:
+.buildkite/scripts/ci-fetch-log.sh "<buildkite_url>"
+```
+
 ### Commit messages
 
 Add attribution using commit trailers such as `Co-authored-by:` (other projects use `Assisted-by:` or `Generated-by:`). For example:

MANIFEST.in

@@ -4,6 +4,7 @@ include requirements/cuda.txt
 include requirements/rocm.txt
 include requirements/cpu.txt
 include CMakeLists.txt
+include tools/build_rust.py
 
 recursive-include cmake *
 recursive-include csrc *

SECURITY.md

@@ -34,6 +34,15 @@ Vulnerabilities that cause denial of service or partial disruption, but do not a
 
 Minor issues such as informational disclosures, logging errors, non-exploitable flaws, or weaknesses that require local or high-privilege access and offer negligible impact. Examples include side channel attacks or hash collisions. These issues often have CVSS scores less than 4.0
 
+## Fix disclosure policy
+
+When a security report is accepted, the fix process depends on the severity:
+
+* **CRITICAL and HIGH severity**: Fixes are developed in a private security fork and coordinated with the prenotification group before public disclosure.
+* **MODERATE and LOW severity**: Fixes are developed and submitted as public pull requests. These issues do not require embargo since they do not enable arbitrary code execution or significant data breach, and public visibility accelerates community review and adoption of the fix.
+
+The vulnerability management team reserves the right to adjust the disclosure approach on a case-by-case basis, taking into account factors such as active exploitation, unusual attack surface, or coordination requirements with downstream vendors.
+
 ## Prenotification policy
 
 For certain security issues of CRITICAL, HIGH, or MODERATE severity level, we may prenotify certain organizations or vendors that ship vLLM. The purpose of this prenotification is to allow for a coordinated release of fixes for severe issues.

benchmarks/attention_benchmarks/README.md

@@ -108,7 +108,6 @@ python benchmark.py \
     --backends flash triton flashinfer \
     --batch-specs "q2k" "8q1s1k" "2q2k_32q1s1k" \
     --num-layers 10 \
-    --repeats 5 \
     --output-csv results.csv
 ```
 
@@ -164,14 +163,17 @@ python benchmark.py \
 # Model configuration
 --num-layers N                      # Number of layers
 --head-dim N                        # Head dimension
+--v-head-dim N                      # Value head dimension (defaults to --head-dim)
 --num-q-heads N                     # Query heads
 --num-kv-heads N                    # KV heads
 --block-size N                      # Block size
+--kv-lora-rank N                    # MLA KV LoRA rank
+--qk-nope-head-dim N                # MLA non-RoPE QK head dim
+--qk-rope-head-dim N                # MLA RoPE QK head dim
 
 # Benchmark settings
 --device DEVICE                     # Device (default: cuda:0)
---repeats N                         # Repetitions
---warmup-iters N                    # Warmup iterations
+--warmup-ms N                       # Warmup window in ms for triton do_bench
 --profile-memory                    # Profile memory usage
 
 # Parameter sweeps
@@ -211,8 +213,6 @@ config = BenchmarkConfig(
     num_kv_heads=1,
     block_size=128,
     device="cuda:0",
-    repeats=5,
-    warmup_iters=3,
 )
 
 # CUTLASS MLA with specific num_kv_splits
@@ -253,14 +253,10 @@ formatter.save_json(results, "output.json")
 
 ## Tips
 
-**1. Warmup matters** - Use `--warmup-iters 10` for stable results
+**1. Save results** - Always use `--output-csv` or `--output-json`
 
-**2. Multiple repeats** - Use `--repeats 20` for low variance
+**2. Test incrementally** - Start with `--num-layers 1`
 
-**3. Save results** - Always use `--output-csv` or `--output-json`
+**3. Extended grammar** - Leverage spec decode, chunked prefill patterns
 
-**4. Test incrementally** - Start with `--num-layers 1 --repeats 1`
-
-**5. Extended grammar** - Leverage spec decode, chunked prefill patterns
-
-**6. Parameter sweeps** - Use `--sweep-param` and `--sweep-values` to find optimal values
+**4. Parameter sweeps** - Use `--sweep-param` and `--sweep-values` to find optimal values

benchmarks/attention_benchmarks/benchmark.py

@@ -26,6 +26,9 @@ Examples:
 """
 
 import argparse
+import os
+import shutil
+import subprocess
 import sys
 from dataclasses import replace
 from pathlib import Path
@@ -83,13 +86,15 @@ def run_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
         else:
             return run_standard_attention_benchmark(config)
     except Exception as e:
+        error_msg = str(e) or repr(e)
         return BenchmarkResult(
             config=config,
             mean_time=float("inf"),
+            median_time=float("inf"),
             std_time=0,
             min_time=float("inf"),
             max_time=float("inf"),
-            error=str(e),
+            error=error_msg,
         )
 
 
@@ -115,9 +120,12 @@ def run_model_parameter_sweep(
     """
     all_results = []
 
-    console.print(
-        f"[yellow]Model sweep mode: testing {sweep.param_name} = {sweep.values}[/]"
+    sweep_desc = (
+        f"{sweep.param_name} = {sweep.values}"
+        if sweep.param_name
+        else f"{len(sweep.values)} configurations"
     )
+    console.print(f"[yellow]Model sweep mode: testing {sweep_desc}[/]")
 
     total = len(backends) * len(batch_specs) * len(sweep.values)
 
@@ -125,9 +133,9 @@ def run_model_parameter_sweep(
         for backend in backends:
             for spec in batch_specs:
                 for value in sweep.values:
-                    # Create config with modified model parameter
+                    # Create config with modified model parameter(s)
                     config_args = base_config_args.copy()
-                    config_args[sweep.param_name] = value
+                    sweep.apply(config_args, value)
 
                     # Create config with original backend for running
                     clean_config = BenchmarkConfig(
@@ -144,13 +152,21 @@ def run_model_parameter_sweep(
                     all_results.append(result)
 
                     if not result.success:
+                        err_label = (
+                            f"{sweep.param_name}={value}"
+                            if sweep.param_name
+                            else f"{value}"
+                        )
                         console.print(
-                            f"[red]Error {backend} {spec} {sweep.param_name}="
-                            f"{value}: {result.error}[/]"
+                            f"[red]Error {backend} {spec} {err_label}"
+                            f": {result.error}[/]"
                         )
 
                     pbar.update(1)
 
+    if base_config_args.get("ncu_profile"):
+        return all_results
+
     # Display sweep results - create separate table for each parameter value
     console.print("\n[bold green]Model Parameter Sweep Results:[/]")
     formatter = ResultsFormatter(console)
@@ -184,7 +200,10 @@ def run_model_parameter_sweep(
     )
 
     for param_value in sorted_param_values:
-        console.print(f"\n[bold cyan]{sweep.param_name} = {param_value}[/]")
+        label = (
+            f"{sweep.param_name} = {param_value}" if sweep.param_name else param_value
+        )
+        console.print(f"\n[bold cyan]{label}[/]")
         param_results = by_param_value[param_value]
 
         # Create modified results with original backend names
@@ -200,8 +219,9 @@ def run_model_parameter_sweep(
         formatter.print_table(modified_results, backends, compare_to_fastest=True)
 
     # Show optimal backend for each (param_value, batch_spec) combination
+    sweep_name = sweep.param_name or "config"
     console.print(
-        f"\n[bold cyan]Optimal backend for each ({sweep.param_name}, batch_spec):[/]"
+        f"\n[bold cyan]Optimal backend for each ({sweep_name}, batch_spec):[/]"
     )
 
     # Group by (param_value, batch_spec)
@@ -236,7 +256,10 @@ def run_model_parameter_sweep(
     for param_value, spec in sorted_keys:
         # Print header when param value changes
         if param_value != current_param_value:
-            console.print(f"\n  [bold]{sweep.param_name}={param_value}:[/]")
+            header = (
+                f"{sweep.param_name}={param_value}" if sweep.param_name else param_value
+            )
+            console.print(f"\n  [bold]{header}:[/]")
             current_param_value = param_value
 
         results = by_param_and_spec[(param_value, spec)]
@@ -322,6 +345,9 @@ def run_parameter_sweep(
 
                     pbar.update(1)
 
+    if base_config_args.get("ncu_profile"):
+        return all_results
+
     # Display sweep results
     console.print("\n[bold green]Sweep Results:[/]")
     backend_labels = [sweep.get_label(b, v) for b in backends for v in sweep_values]
@@ -474,11 +500,35 @@ def main():
     parser.add_argument("--num-q-heads", type=int, default=32, help="Query heads")
     parser.add_argument("--num-kv-heads", type=int, default=8, help="KV heads")
     parser.add_argument("--block-size", type=int, default=16, help="Block size")
+    parser.add_argument(
+        "--v-head-dim",
+        type=int,
+        default=None,
+        help="Value head dimension (defaults to --head-dim if unset)",
+    )
+
+    # MLA-specific model dimensions
+    parser.add_argument(
+        "--kv-lora-rank", type=int, default=None, help="MLA KV LoRA rank"
+    )
+    parser.add_argument(
+        "--qk-nope-head-dim", type=int, default=None, help="MLA non-RoPE QK head dim"
+    )
+    parser.add_argument(
+        "--qk-rope-head-dim", type=int, default=None, help="MLA RoPE QK head dim"
+    )
 
     # Benchmark settings
     parser.add_argument("--device", default="cuda:0", help="Device")
-    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(
+        "--warmup-ms",
+        type=int,
+        default=None,
+        help=(
+            "Warmup window in ms for triton's do_bench (default: triton's own). "
+            "Has no effect with CUDA graphs; pass --no-cuda-graphs to use it."
+        ),
+    )
     parser.add_argument("--profile-memory", action="store_true", help="Profile memory")
     parser.add_argument(
         "--kv-cache-dtype",
@@ -491,10 +541,33 @@ def main():
         action=argparse.BooleanOptionalAction,
         default=True,
         help=(
-            "Launch kernels with CUDA graphs to eliminate CPU overhead"
-            "in measurements (default: True)"
+            "Use triton do_bench_cudagraph (True) or do_bench (False) "
+            "for timing. CUDA graphs eliminate CPU launch overhead "
+            "(default: True)"
         ),
     )
+    parser.add_argument(
+        "--num-splits",
+        type=int,
+        default=None,
+        help="FlashAttention split-K factor (0=auto heuristic, 1=disabled, >1=force N)",
+    )
+    parser.add_argument(
+        "--ncu-profile",
+        action="store_true",
+        default=False,
+        help=(
+            "Enable Nsight Compute profiling mode. Automatically wraps the "
+            "script with ncu, capturing a profile with source correlation. "
+            "Use --ncu-output to set the output file name."
+        ),
+    )
+    parser.add_argument(
+        "--ncu-output",
+        type=str,
+        default="profile",
+        help="Output file name for ncu profile (default: 'profile').",
+    )
 
     # Parameter sweep (use YAML config for advanced sweeps)
     parser.add_argument(
@@ -576,23 +649,28 @@ def main():
             model = yaml_config["model"]
             args.num_layers = model.get("num_layers", args.num_layers)
             args.head_dim = model.get("head_dim", args.head_dim)
+            args.v_head_dim = model.get("v_head_dim", args.v_head_dim)
             args.num_q_heads = model.get("num_q_heads", args.num_q_heads)
             args.num_kv_heads = model.get("num_kv_heads", args.num_kv_heads)
             args.block_size = model.get("block_size", args.block_size)
+            # MLA-specific dimensions
+            args.kv_lora_rank = model.get("kv_lora_rank", args.kv_lora_rank)
+            args.qk_nope_head_dim = model.get("qk_nope_head_dim", args.qk_nope_head_dim)
+            args.qk_rope_head_dim = model.get("qk_rope_head_dim", args.qk_rope_head_dim)
 
         # Benchmark settings (top-level keys)
         if "device" in yaml_config:
             args.device = yaml_config["device"]
-        if "repeats" in yaml_config:
-            args.repeats = yaml_config["repeats"]
-        if "warmup_iters" in yaml_config:
-            args.warmup_iters = yaml_config["warmup_iters"]
+        if "warmup_ms" in yaml_config:
+            args.warmup_ms = yaml_config["warmup_ms"]
         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"]
+        if "ncu_profile" in yaml_config:
+            args.ncu_profile = yaml_config["ncu_profile"]
 
         # Parameter sweep configuration
         if "parameter_sweep" in yaml_config:
@@ -612,7 +690,7 @@ def main():
         if "model_parameter_sweep" in yaml_config:
             sweep_config = yaml_config["model_parameter_sweep"]
             args.model_parameter_sweep = ModelParameterSweep(
-                param_name=sweep_config["param_name"],
+                param_name=sweep_config.get("param_name"),
                 values=sweep_config["values"],
                 label_format=sweep_config.get(
                     "label_format", "{backend}_{param_name}_{value}"
@@ -631,6 +709,32 @@ def main():
 
         console.print()
 
+    # Re-exec under ncu if --ncu-profile and not already inside ncu. This runs
+    # after YAML processing so ncu_profile set via config file is honored.
+    if args.ncu_profile and "_NCU_INNER" not in os.environ:
+        ncu = shutil.which("ncu")
+        if ncu is None:
+            print("Error: 'ncu' not found in PATH", file=sys.stderr)
+            sys.exit(1)
+        cmd = [
+            ncu,
+            "--profile-from-start",
+            "off",
+            "--set",
+            "full",
+            "--import-source",
+            "yes",
+            "-o",
+            args.ncu_output,
+            sys.executable,
+            *sys.argv,
+        ]
+        env = os.environ.copy()
+        env["CUTE_DSL_LINEINFO"] = "1"
+        env["_NCU_INNER"] = "1"
+        print(f"Launching: {' '.join(cmd)}")
+        sys.exit(subprocess.call(cmd, env=env))
+
     # Handle CLI-based parameter sweep (if not from YAML)
     if (
         (not hasattr(args, "parameter_sweep") or args.parameter_sweep is None)
@@ -655,6 +759,18 @@ def main():
     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}")
+    if args.warmup_ms is not None and args.cuda_graphs:
+        console.print(
+            "[yellow]Warning: --warmup-ms is ignored with CUDA graphs "
+            "(do_bench_cudagraph warms up internally). Pass --no-cuda-graphs "
+            "to use it.[/]"
+        )
+    if args.num_splits == 0 and args.cuda_graphs:
+        console.print(
+            "[yellow]Warning: --num-splits 0 (FA3 heuristic) is not CUDA-graph "
+            "compatible and may fail or fall back. Pass --no-cuda-graphs or use "
+            "--num-splits >=1.[/]"
+        )
     console.print()
 
     init_workspace_manager(args.device)
@@ -662,6 +778,15 @@ def main():
     # Run benchmarks
     all_results = []
 
+    # Under ncu profiling the kernels run only to be captured by the profiler;
+    # timings are placeholder zeros, so the result tables and saved metrics are
+    # skipped. The Nsight Compute report (--ncu-output) holds the real data.
+    if args.ncu_profile:
+        console.print(
+            "[dim]ncu profiling enabled: result tables and saved metrics are "
+            "skipped (timings are placeholder zeros).[/]"
+        )
+
     # Handle special mode: decode_vs_prefill comparison
     if hasattr(args, "mode") and args.mode == "decode_vs_prefill":
         console.print("[yellow]Mode: Decode vs Prefill pipeline comparison[/]")
@@ -708,11 +833,11 @@ def main():
                         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,
+                        ncu_profile=args.ncu_profile,
+                        warmup_ms=args.warmup_ms,
                     )
 
                     # Add decode pipeline config
@@ -749,6 +874,7 @@ def main():
                         result = BenchmarkResult(
                             config=config,
                             mean_time=timing["mean"],
+                            median_time=timing.get("median", timing["mean"]),
                             std_time=timing["std"],
                             min_time=timing["min"],
                             max_time=timing["max"],
@@ -770,6 +896,7 @@ def main():
                         result = BenchmarkResult(
                             config=config,
                             mean_time=float("inf"),
+                            median_time=float("inf"),
                             std_time=0,
                             min_time=float("inf"),
                             max_time=float("inf"),
@@ -779,6 +906,9 @@ def main():
 
                 pbar.update(1)
 
+        if args.ncu_profile:
+            return
+
         # Display decode vs prefill results
         console.print("\n[bold green]Decode vs Prefill Results:[/]")
 
@@ -858,15 +988,20 @@ def main():
         base_config_args = {
             "num_layers": args.num_layers,
             "head_dim": args.head_dim,
+            "v_head_dim": args.v_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,
+            "ncu_profile": args.ncu_profile,
+            "warmup_ms": args.warmup_ms,
+            "num_splits": args.num_splits,
+            "kv_lora_rank": args.kv_lora_rank,
+            "qk_nope_head_dim": args.qk_nope_head_dim,
+            "qk_rope_head_dim": args.qk_rope_head_dim,
         }
         all_results = run_model_parameter_sweep(
             backends,
@@ -882,15 +1017,17 @@ def main():
         base_config_args = {
             "num_layers": args.num_layers,
             "head_dim": args.head_dim,
+            "v_head_dim": args.v_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,
+            "ncu_profile": args.ncu_profile,
+            "warmup_ms": args.warmup_ms,
+            "num_splits": args.num_splits,
         }
         all_results = run_parameter_sweep(
             backends, args.batch_specs, base_config_args, args.parameter_sweep, console
@@ -914,15 +1051,17 @@ def main():
                             batch_spec=spec,
                             num_layers=args.num_layers,
                             head_dim=args.head_dim,
+                            v_head_dim=getattr(args, "v_head_dim", None),
                             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,
+                            ncu_profile=args.ncu_profile,
+                            warmup_ms=args.warmup_ms,
+                            num_splits=args.num_splits,
                         )
 
                         result = run_benchmark(config)
@@ -935,9 +1074,10 @@ def main():
 
                         pbar.update(1)
 
-            console.print("\n[bold green]Results:[/]")
-            formatter = ResultsFormatter(console)
-            formatter.print_table(decode_results, backends)
+            if not args.ncu_profile:
+                console.print("\n[bold green]Results:[/]")
+                formatter = ResultsFormatter(console)
+                formatter.print_table(decode_results, backends)
 
         # Run prefill backend comparison
         if prefill_backends:
@@ -962,9 +1102,8 @@ def main():
                             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,
+                            warmup_ms=args.warmup_ms,
                             prefill_backend=pb,
                         )
 
@@ -980,16 +1119,17 @@ def main():
 
                         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
-            )
+            if not args.ncu_profile:
+                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:
+    # Save results (skip ncu profiling runs: timings are placeholder zeros)
+    if all_results and not args.ncu_profile:
         formatter = ResultsFormatter(console)
         if args.output_csv:
             formatter.save_csv(all_results, args.output_csv)

benchmarks/attention_benchmarks/common.py

@@ -15,6 +15,8 @@ from batch_spec import get_batch_type, parse_batch_spec
 from rich.console import Console
 from rich.table import Table
 
+from vllm.triton_utils import triton
+
 
 def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:
     """
@@ -34,6 +36,30 @@ def batch_spec_sort_key(spec: str) -> tuple[int, int, int]:
         return (0, 0, 0)
 
 
+def run_do_bench(
+    benchmark_fn,
+    use_cuda_graphs: bool,
+    warmup_ms: int | None = None,
+) -> list[float]:
+    kwargs: dict[str, Any] = {"return_mode": "all"}
+    if use_cuda_graphs:
+        result = triton.testing.do_bench_cudagraph(benchmark_fn, **kwargs)
+    else:
+        if warmup_ms is not None:
+            kwargs["warmup"] = warmup_ms
+        result = triton.testing.do_bench(benchmark_fn, **kwargs)
+    return result
+
+
+def run_ncu_profile(benchmark_fn) -> None:
+    benchmark_fn()
+    torch.accelerator.synchronize()
+    torch.cuda.cudart().cudaProfilerStart()
+    benchmark_fn()
+    torch.accelerator.synchronize()
+    torch.cuda.cudart().cudaProfilerStop()
+
+
 # Mock classes for vLLM attention infrastructure
 
 
@@ -182,18 +208,37 @@ class ParameterSweep:
 
 @dataclass
 class ModelParameterSweep:
-    """Configuration for sweeping a model configuration parameter."""
+    """Configuration for sweeping model configuration parameter(s).
 
-    param_name: str  # Name of the model config parameter to sweep (e.g., "num_q_heads")
-    values: list[Any]  # List of values to test
-    label_format: str = "{backend}_{param_name}_{value}"  # Result label template
+    Supports two modes:
+    - Single param: param_name="head_dim", values=[128, 256, 512]
+    - Multi param: values=[{head_dim: 192, v_head_dim: 128}, {head_dim: 256}]
+      When values are dicts, each dict's keys are applied as config overrides.
+    """
+
+    param_name: str | None = None
+    values: list[Any] | None = None
+    label_format: str = "{backend}_{param_name}_{value}"
 
     def get_label(self, backend: str, value: Any) -> str:
         """Generate a label for a specific parameter value."""
+        if isinstance(value, dict):
+            return self.label_format.format(
+                backend=backend, param_name=self.param_name, value=value, **value
+            )
         return self.label_format.format(
             backend=backend, param_name=self.param_name, value=value
         )
 
+    def apply(self, config_args: dict, value: Any) -> None:
+        """Apply a sweep value to config args."""
+        if isinstance(value, dict):
+            config_args.update(value)
+        elif self.param_name is not None:
+            config_args[self.param_name] = value
+        else:
+            raise ValueError("param_name must be set if sweep values are not dicts")
+
 
 @dataclass
 class BenchmarkConfig:
@@ -208,10 +253,10 @@ class BenchmarkConfig:
     block_size: int
     device: str
     dtype: torch.dtype = torch.float16
-    repeats: int = 1
-    warmup_iters: int = 3
     profile_memory: bool = False
     use_cuda_graphs: bool = False
+    ncu_profile: bool = False
+    warmup_ms: int | None = None
 
     # "auto" or "fp8"
     kv_cache_dtype: str = "auto"
@@ -226,6 +271,7 @@ class BenchmarkConfig:
     # Backend-specific tuning
     num_kv_splits: int | None = None  # CUTLASS MLA
     reorder_batch_threshold: int | None = None  # FlashAttn MLA, FlashMLA
+    num_splits: int | None = None  # FlashAttention split-K (0=auto, 1=disabled)
 
 
 @dataclass
@@ -234,6 +280,7 @@ class BenchmarkResult:
 
     config: BenchmarkConfig
     mean_time: float  # seconds
+    median_time: float  # seconds
     std_time: float  # seconds
     min_time: float  # seconds
     max_time: float  # seconds
@@ -252,6 +299,7 @@ class BenchmarkResult:
         return {
             "config": asdict(self.config),
             "mean_time": self.mean_time,
+            "median_time": self.median_time,
             "std_time": self.std_time,
             "min_time": self.min_time,
             "max_time": self.max_time,

benchmarks/attention_benchmarks/configs/mla_decode.yaml

@@ -56,8 +56,6 @@ backends:
   - TOKENSPEED_MLA  # Blackwell + R1 dims + FP8 KV (use --kv-cache-dtype fp8)
 
 device: "cuda:0"
-repeats: 100
-warmup_iters: 10
 profile_memory: true
 
 # Backend-specific tuning

benchmarks/attention_benchmarks/configs/mla_mixed_batch.yaml

@@ -51,8 +51,6 @@ backends:
   - FLASHMLA         # Hopper only
 
 device: "cuda:0"
-repeats: 5
-warmup_iters: 3
 profile_memory: true
 
 # Analyze chunked prefill workspace size impact

benchmarks/attention_benchmarks/configs/mla_prefill.yaml

@@ -124,5 +124,3 @@ prefill_backends:
   - tokenspeed
 
 device: "cuda:0"
-repeats: 20
-warmup_iters: 5

benchmarks/attention_benchmarks/configs/mla_sparse_decode.yaml

@@ -53,6 +53,4 @@ backends:
   - FLASHINFER_MLA_SPARSE
 
 device: "cuda:0"
-repeats: 100
-warmup_iters: 10
 profile_memory: true

benchmarks/attention_benchmarks/configs/mla_sparse_prefill.yaml

@@ -57,6 +57,4 @@ backends:
   - FLASHINFER_MLA_SPARSE
 
 device: "cuda:0"
-repeats: 10
-warmup_iters: 3
 profile_memory: true

benchmarks/attention_benchmarks/configs/reorder_threshold.yaml

@@ -63,8 +63,6 @@ model:
 
 # Benchmark settings
 device: "cuda:0"
-repeats: 15          # More repeats for spec decode variance
-warmup_iters: 5
 profile_memory: false
 
 # Output

benchmarks/attention_benchmarks/configs/speculative_decode.yaml

@@ -49,8 +49,6 @@ backends:
 
 # Benchmark settings
 device: "cuda:0"
-repeats: 10  # More repeats for statistical significance
-warmup_iters: 5
 profile_memory: false
 
 # Test these threshold values for optimization

benchmarks/attention_benchmarks/configs/standard_attention.yaml

@@ -43,6 +43,4 @@ backends:
   - FLASHINFER
 
 device: "cuda:0"
-repeats: 5
-warmup_iters: 3
 profile_memory: false

benchmarks/attention_benchmarks/configs/standard_decode.yaml

@@ -0,0 +1,142 @@
+# Standard attention decode benchmark configuration
+# Sweeps num_q_heads and num_kv_heads to isolate effects of:
+#   1. GQA ratio (fixed num_q_heads=32, vary num_kv_heads)
+#   2. Absolute head count (fixed 4:1 ratio, vary scale)
+
+model:
+  num_layers: 32
+  num_q_heads: 32 # Base value, overridden by sweep
+  num_kv_heads: 8 # Base value, overridden by sweep
+  head_dim: 128
+  block_size: 16
+
+# Head count sweep: each entry overrides num_q_heads, num_kv_heads, and
+# head_dim where it differs from the base (128). Head counts are per-GPU
+# (i.e. after TP sharding).
+#
+# Group A — vary GQA ratio (fixed q=32, head_dim=128):
+#   32:32 (MHA), 32:8 (GQA 4:1), 32:4 (GQA 8:1), 32:1 (MQA)
+#
+# Groups B-E — real model configs at various TP degrees:
+#   Model              head_dim  Full    TP2      TP4      TP8
+#   Llama 3 8B           128    32:8   16:4      8:2      4:1
+#   Llama 3 70B          128    64:8   32:4     16:2      8:1
+#   GPT-OSS 120B          64    64:8   32:4     16:2      8:1
+#   Llama 3 405B         128   128:8   64:4     32:2     16:1
+model_parameter_sweep:
+  values:
+    # --- head_dim=128 (Llama 3 family) ---
+    - { num_q_heads: 32, num_kv_heads: 32, head_dim: 128 } # MHA 1:1
+    - { num_q_heads: 32, num_kv_heads: 1, head_dim: 128 } # MQA 32:1
+    - { num_q_heads: 4, num_kv_heads: 1, head_dim: 128 } # Llama 3 8B TP8
+    - { num_q_heads: 8, num_kv_heads: 2, head_dim: 128 } # Llama 3 8B TP4
+    - { num_q_heads: 16, num_kv_heads: 4, head_dim: 128 } # Llama 3 8B TP2
+    - { num_q_heads: 32, num_kv_heads: 8, head_dim: 128 } # Llama 3 8B TP1 / GQA 4:1
+    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 128 } # Llama 3 70B TP8
+    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 128 } # Llama 3 70B TP4
+    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 128 } # Llama 3 70B TP2 / GQA 8:1
+    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 128 } # Llama 3 70B TP1
+    - { num_q_heads: 16, num_kv_heads: 1, head_dim: 128 } # Llama 3 405B TP8
+    - { num_q_heads: 32, num_kv_heads: 2, head_dim: 128 } # Llama 3 405B TP4
+    - { num_q_heads: 64, num_kv_heads: 4, head_dim: 128 } # Llama 3 405B TP2
+    - { num_q_heads: 128, num_kv_heads: 8, head_dim: 128 } # Llama 3 405B TP1
+    # --- head_dim=64 (GPT-OSS 120B) ---
+    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 64 } # GPT-OSS 120B TP8
+    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 64 } # GPT-OSS 120B TP4
+    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 64 } # GPT-OSS 120B TP2
+    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 64 } # GPT-OSS 120B TP1
+  label_format: "{backend}_q{num_q_heads}kv{num_kv_heads}d{head_dim}"
+
+batch_specs:
+  # ---- batch_size x seq_len grid (decode: q_len=1) ----
+  # Small grid for quick iteration. Uncomment for full sweep.
+
+  # Batch size 1
+  - "q1s1k"
+  - "q1s512"
+  - "q1s2k"
+  - "q1s4k"
+  - "q1s8k"
+  - "q1s16k"
+  - "q1s32k"
+
+  # Batch size 2
+  - "2q1s512"
+  - "2q1s1k"
+  - "2q1s2k"
+  - "2q1s4k"
+  - "2q1s8k"
+  - "2q1s16k"
+  - "2q1s32k"
+
+  # Batch size 4
+  - "4q1s512"
+  - "4q1s1k"
+  - "4q1s2k"
+  - "4q1s4k"
+  - "4q1s8k"
+  - "4q1s16k"
+  - "4q1s32k"
+
+  # Batch size 8
+  - "8q1s1k"
+  - "8q1s512"
+  - "8q1s2k"
+  - "8q1s4k"
+  - "8q1s8k"
+  - "8q1s16k"
+  - "8q1s32k"
+
+  # Batch size 16
+  - "16q1s512"
+  - "16q1s1k"
+  - "16q1s2k"
+  - "16q1s4k"
+  - "16q1s8k"
+  - "16q1s16k"
+  - "16q1s32k"
+
+  # Batch size 32
+  - "32q1s512"
+  - "32q1s1k"
+  - "32q1s2k"
+  - "32q1s4k"
+  - "32q1s8k"
+  - "32q1s16k"
+  - "32q1s32k"
+
+  # Batch size 64
+  - "64q1s1k"
+  - "64q1s512"
+  - "64q1s2k"
+  - "64q1s4k"
+  - "64q1s8k"
+  - "64q1s16k"
+  - "64q1s32k"
+
+  # Batch size 128
+  - "128q1s512"
+  - "128q1s1k"
+  - "128q1s2k"
+  - "128q1s4k"
+  - "128q1s8k"
+  - "128q1s16k"
+  - "128q1s32k"
+
+  # Batch size 256
+  - "256q1s1k"
+  - "256q1s512"
+  - "256q1s2k"
+  - "256q1s4k"
+  - "256q1s8k"
+  - "256q1s16k"
+  - "256q1s32k"
+
+# Available backends: FLASH_ATTN, TRITON_ATTN, FLASHINFER
+backends:
+  - FLASH_ATTN
+  - TRITON_ATTN
+  - FLASHINFER
+
+device: "cuda:0"
+profile_memory: false

benchmarks/attention_benchmarks/configs/standard_prefill.yaml

@@ -0,0 +1,108 @@
+# Standard attention prefill benchmark configuration
+# Sweeps num_q_heads and num_kv_heads to isolate effects of:
+#   1. GQA ratio (fixed num_q_heads=32, vary num_kv_heads)
+#   2. Absolute head count (fixed 4:1 ratio, vary scale)
+
+model:
+  num_layers: 32
+  num_q_heads: 32 # Base value, overridden by sweep
+  num_kv_heads: 8 # Base value, overridden by sweep
+  head_dim: 128
+  block_size: 16
+
+# Head count sweep: each entry overrides num_q_heads, num_kv_heads, and
+# head_dim where it differs from the base (128). Head counts are per-GPU
+# (i.e. after TP sharding).
+#
+# Group A — vary GQA ratio (fixed q=32, head_dim=128):
+#   32:32 (MHA), 32:8 (GQA 4:1), 32:4 (GQA 8:1), 32:1 (MQA)
+#
+# Groups B-E — real model configs at various TP degrees:
+#   Model              head_dim  Full    TP2      TP4      TP8
+#   Llama 3 8B           128    32:8   16:4      8:2      4:1
+#   Llama 3 70B          128    64:8   32:4     16:2      8:1
+#   GPT-OSS 120B          64    64:8   32:4     16:2      8:1
+#   Llama 3 405B         128   128:8   64:4     32:2     16:1
+model_parameter_sweep:
+  values:
+    # --- head_dim=128 (Llama 3 family) ---
+    - { num_q_heads: 32, num_kv_heads: 32, head_dim: 128 } # MHA 1:1
+    - { num_q_heads: 32, num_kv_heads: 1, head_dim: 128 } # MQA 32:1
+    - { num_q_heads: 4, num_kv_heads: 1, head_dim: 128 } # Llama 3 8B TP8
+    - { num_q_heads: 8, num_kv_heads: 2, head_dim: 128 } # Llama 3 8B TP4
+    - { num_q_heads: 16, num_kv_heads: 4, head_dim: 128 } # Llama 3 8B TP2
+    - { num_q_heads: 32, num_kv_heads: 8, head_dim: 128 } # Llama 3 8B TP1 / GQA 4:1
+    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 128 } # Llama 3 70B TP8
+    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 128 } # Llama 3 70B TP4
+    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 128 } # Llama 3 70B TP2 / GQA 8:1
+    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 128 } # Llama 3 70B TP1
+    - { num_q_heads: 16, num_kv_heads: 1, head_dim: 128 } # Llama 3 405B TP8
+    - { num_q_heads: 32, num_kv_heads: 2, head_dim: 128 } # Llama 3 405B TP4
+    - { num_q_heads: 64, num_kv_heads: 4, head_dim: 128 } # Llama 3 405B TP2
+    - { num_q_heads: 128, num_kv_heads: 8, head_dim: 128 } # Llama 3 405B TP1
+    # --- head_dim=64 (GPT-OSS 120B) ---
+    - { num_q_heads: 8, num_kv_heads: 1, head_dim: 64 } # GPT-OSS 120B TP8
+    - { num_q_heads: 16, num_kv_heads: 2, head_dim: 64 } # GPT-OSS 120B TP4
+    - { num_q_heads: 32, num_kv_heads: 4, head_dim: 64 } # GPT-OSS 120B TP2
+    - { num_q_heads: 64, num_kv_heads: 8, head_dim: 64 } # GPT-OSS 120B TP1
+  label_format: "{backend}_q{num_q_heads}kv{num_kv_heads}d{head_dim}"
+
+batch_specs:
+  # ---- batch_size x prefill_len grid (prefill: q_len == seq_len) ----
+  # Total tokens = batch_size * prefill_len, and prefill compute scales with
+  # prefill_len^2, so the largest cells are expensive. Trim batch sizes or
+  # lengths for quick iteration.
+
+  # Batch size 1
+  - "q512"
+  - "q1k"
+  - "q2k"
+  - "q4k"
+  - "q8k"
+  - "q16k"
+  - "q32k"
+
+  # Batch size 2
+  - "2q512"
+  - "2q1k"
+  - "2q2k"
+  - "2q4k"
+  - "2q8k"
+  - "2q16k"
+  - "2q32k"
+
+  # Batch size 4
+  - "4q512"
+  - "4q1k"
+  - "4q2k"
+  - "4q4k"
+  - "4q8k"
+  - "4q16k"
+  - "4q32k"
+
+  # Batch size 8
+  - "8q512"
+  - "8q1k"
+  - "8q2k"
+  - "8q4k"
+  - "8q8k"
+  - "8q16k"
+  - "8q32k"
+
+  # Batch size 16
+  - "16q512"
+  - "16q1k"
+  - "16q2k"
+  - "16q4k"
+  - "16q8k"
+  - "16q16k"
+  - "16q32k"
+
+# Available backends: FLASH_ATTN, TRITON_ATTN, FLASHINFER
+backends:
+  - FLASH_ATTN
+  - TRITON_ATTN
+  - FLASHINFER
+
+device: "cuda:0"
+profile_memory: false

benchmarks/attention_benchmarks/mla_runner.py

@@ -8,6 +8,8 @@ This module provides helpers for running MLA backends without
 needing full VllmConfig integration.
 """
 
+import statistics
+
 import numpy as np
 import torch
 from batch_spec import parse_batch_spec
@@ -17,6 +19,8 @@ from common import (
     MockIndexer,
     MockKVBProj,
     MockLayer,
+    run_do_bench,
+    run_ncu_profile,
     setup_mla_dims,
 )
 
@@ -820,7 +824,7 @@ def _run_single_benchmark(
             num_prefill, mla_dims, query_fmt, device, torch.bfloat16
         )
 
-    # Build forward function
+    # Build forward function (runs a single decode/prefill pass)
     def forward_fn():
         results = []
         if has_decode:
@@ -839,44 +843,35 @@ def _run_single_benchmark(
             )
         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):
-        start = torch.cuda.Event(enable_timing=True)
-        end = torch.cuda.Event(enable_timing=True)
-
-        start.record()
+    def benchmark_fn():
         for _ in range(config.num_layers):
-            benchmark_fn()
-        end.record()
+            forward_fn()
 
-        torch.accelerator.synchronize()
-        elapsed_ms = start.elapsed_time(end)
-        times.append(elapsed_ms / 1000.0 / config.num_layers)
+    if config.ncu_profile:
+        run_ncu_profile(benchmark_fn)
+        return BenchmarkResult(
+            config=config,
+            mean_time=0.0,
+            median_time=0.0,
+            std_time=0.0,
+            min_time=0.0,
+            max_time=0.0,
+            throughput_tokens_per_sec=0.0,
+        )
+
+    all_ms = run_do_bench(benchmark_fn, config.use_cuda_graphs, config.warmup_ms)
+
+    # Convert ms to seconds per layer
+    times = [t / 1000.0 / config.num_layers for t in all_ms]
+    mean_time = statistics.mean(times)
 
-    mean_time = float(np.mean(times))
     return BenchmarkResult(
         config=config,
         mean_time=mean_time,
-        std_time=float(np.std(times)),
-        min_time=float(np.min(times)),
-        max_time=float(np.max(times)),
+        median_time=statistics.median(times),
+        std_time=statistics.stdev(times) if len(times) > 1 else 0.0,
+        min_time=min(times),
+        max_time=max(times),
         throughput_tokens_per_sec=total_q / mean_time if mean_time > 0 else 0,
     )
 

benchmarks/attention_benchmarks/runner.py

@@ -9,13 +9,20 @@ This module provides helpers for running standard attention backends
 """
 
 import logging
+import statistics
 import types
 from contextlib import contextmanager
 
-import numpy as np
 import torch
 from batch_spec import parse_batch_spec, reorder_for_flashinfer
-from common import BenchmarkConfig, BenchmarkResult, MockLayer, get_attention_scale
+from common import (
+    BenchmarkConfig,
+    BenchmarkResult,
+    MockLayer,
+    get_attention_scale,
+    run_do_bench,
+    run_ncu_profile,
+)
 
 from vllm.config import (
     CacheConfig,
@@ -208,6 +215,13 @@ def _create_backend_impl(
 
     scale = get_attention_scale(config.head_dim)
 
+    # Set v_head_dim for diff-headdim backends. Always reset (defaulting to
+    # head_dim) so a prior run's value doesn't leak into this one via the
+    # backend's class-level state.
+    if hasattr(backend_class, "set_head_size_v"):
+        v_dim = config.v_head_dim if config.v_head_dim is not None else config.head_dim
+        backend_class.set_head_size_v(v_dim)
+
     impl = backend_class.get_impl_cls()(
         num_heads=config.num_q_heads,
         head_size=config.head_dim,
@@ -300,6 +314,7 @@ def _create_input_tensors(
         from vllm.platforms import current_platform
 
         q_dtype = current_platform.fp8_dtype()
+    v_dim = config.v_head_dim if config.v_head_dim is not None else config.head_dim
     q_list = [
         torch.randn(
             total_q, config.num_q_heads, config.head_dim, device=device, dtype=dtype
@@ -313,9 +328,7 @@ def _create_input_tensors(
         for _ in range(config.num_layers)
     ]
     v_list = [
-        torch.randn(
-            total_q, config.num_kv_heads, config.head_dim, device=device, dtype=dtype
-        )
+        torch.randn(total_q, config.num_kv_heads, v_dim, device=device, dtype=dtype)
         for _ in range(config.num_layers)
     ]
     return q_list, k_list, v_list
@@ -389,14 +402,17 @@ def _run_single_benchmark(
     device: torch.device,
     dtype: torch.dtype,
 ) -> tuple:
-    """Run single benchmark iteration with warmup and timing loop."""
-    total_q = q_list[0].shape[0]
-    out = torch.empty(
-        total_q, config.num_q_heads, config.head_dim, device=device, dtype=dtype
-    )
+    """Run single benchmark using triton's do_bench_cudagraph/do_bench.
 
-    # Warmup
-    for _ in range(config.warmup_iters):
+    Returns:
+        (timing_stats, mem_stats) where timing_stats is a dict with
+        mean/std/min/max in seconds per layer.
+    """
+    total_q = q_list[0].shape[0]
+    v_dim = config.v_head_dim if config.v_head_dim is not None else config.head_dim
+    out = torch.empty(total_q, config.num_q_heads, v_dim, device=device, dtype=dtype)
+
+    def benchmark_fn():
         for i in range(config.num_layers):
             impl.forward(
                 layer,
@@ -407,52 +423,22 @@ def _run_single_benchmark(
                 attn_metadata,
                 output=out,
             )
-    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
+    if config.ncu_profile:
+        run_ncu_profile(benchmark_fn)
+        timing_stats = dict.fromkeys(("mean", "median", "std", "min", "max"), 0.0)
     else:
+        all_ms = run_do_bench(benchmark_fn, config.use_cuda_graphs, config.warmup_ms)
 
-        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):
-        start = torch.cuda.Event(enable_timing=True)
-        end = torch.cuda.Event(enable_timing=True)
-
-        start.record()
-        benchmark_fn()
-        end.record()
-
-        torch.accelerator.synchronize()
-        elapsed_ms = start.elapsed_time(end)
-        times.append(elapsed_ms / 1000.0 / config.num_layers)  # seconds per layer
+        # Convert ms to seconds per layer
+        times = [t / 1000.0 / config.num_layers for t in all_ms]
+        timing_stats = {
+            "mean": statistics.mean(times),
+            "std": statistics.stdev(times) if len(times) > 1 else 0.0,
+            "min": min(times),
+            "max": max(times),
+            "median": statistics.median(times),
+        }
 
     mem_stats = {}
     if config.profile_memory:
@@ -461,7 +447,7 @@ def _run_single_benchmark(
             "reserved_mb": torch.accelerator.memory_reserved(device) / 1024**2,
         }
 
-    return times, mem_stats
+    return timing_stats, mem_stats
 
 
 # ============================================================================
@@ -541,6 +527,12 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 common_attn_metadata=common_metadata,
             )
 
+            # Override num_splits for split-K testing (FlashAttention only)
+            if config.num_splits is not None and hasattr(
+                attn_metadata, "max_num_splits"
+            ):
+                attn_metadata.max_num_splits = config.num_splits
+
             # Only quantize queries when the impl supports it
             quantize_query = config.kv_cache_dtype.startswith("fp8") and getattr(
                 impl, "supports_quant_query_input", False
@@ -553,7 +545,7 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 config, max_num_blocks, backend_class, device, dtype
             )
 
-            times, mem_stats = _run_single_benchmark(
+            timing_stats, mem_stats = _run_single_benchmark(
                 config,
                 impl,
                 layer,
@@ -566,15 +558,16 @@ def run_attention_benchmark(config: BenchmarkConfig) -> BenchmarkResult:
                 dtype,
             )
 
-    mean_time = np.mean(times)
+    mean_time = timing_stats["mean"]
     throughput = total_q / mean_time if mean_time > 0 else 0
 
     return BenchmarkResult(
         config=config,
         mean_time=mean_time,
-        std_time=np.std(times),
-        min_time=np.min(times),
-        max_time=np.max(times),
+        median_time=timing_stats["median"],
+        std_time=timing_stats["std"],
+        min_time=timing_stats["min"],
+        max_time=timing_stats["max"],
         throughput_tokens_per_sec=throughput,
         memory_allocated_mb=mem_stats.get("allocated_mb"),
         memory_reserved_mb=mem_stats.get("reserved_mb"),

benchmarks/benchmark_hidden_state_extraction.py

@@ -92,7 +92,6 @@ def run_baseline(
     llm = LLM(
         model=model,
         enable_prefix_caching=False,
-        enable_chunked_prefill=False,
         **extra_args,
     )
     sampling_params = SamplingParams(max_tokens=1)
@@ -194,7 +193,6 @@ async def _run_extraction_async(
     engine_args = AsyncEngineArgs(
         model=model,
         enable_prefix_caching=False,
-        enable_chunked_prefill=False,
         max_num_batched_tokens=40960,
         max_model_len=40960,
         speculative_config={

benchmarks/disagg_benchmarks/disagg_overhead_benchmark.sh

@@ -1,143 +0,0 @@
-#!/bin/bash
-
-# benchmark the overhead of disaggregated prefill.
-# methodology:
-# - send all request to prefill vLLM instance. It will buffer KV cache.
-# - then send all request to decode instance.
-# - The TTFT of decode instance is the overhead.
-
-set -ex
-
-kill_gpu_processes() {
-  # kill all processes on GPU.
-  pgrep pt_main_thread | xargs -r kill -9
-  pgrep python3 | xargs -r kill -9
-  # vLLM now names the process with VLLM prefix after https://github.com/vllm-project/vllm/pull/21445
-  pgrep VLLM | xargs -r kill -9
-  sleep 10
-
-  # remove vllm config file
-  rm -rf ~/.config/vllm
-
-  # Print the GPU memory usage
-  # so that we know if all GPU processes are killed.
-  gpu_memory_usage=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i 0)
-  # The memory usage should be 0 MB.
-  echo "GPU 0 Memory Usage: $gpu_memory_usage MB"
-}
-
-wait_for_server() {
-  # wait for vllm server to start
-  # return 1 if vllm server crashes
-  local port=$1
-  timeout 1200 bash -c "
-    until curl -s localhost:${port}/v1/completions > /dev/null; do
-      sleep 1
-    done" && return 0 || return 1
-}
-
-
-benchmark() {
-
-  export VLLM_LOGGING_LEVEL=DEBUG
-  export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
-
-  # compare chunked prefill with disaggregated prefill
-
-  results_folder="./results"
-  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
-  dataset_name="sonnet"
-  dataset_path="../sonnet_4x.txt"
-  num_prompts=10
-  qps=$1
-  prefix_len=50
-  input_len=2048
-  output_len=$2
-
-
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
-    --port 8100 \
-    --max-model-len 10000 \
-    --gpu-memory-utilization 0.6 \
-    --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
-
-
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
-    --port 8200 \
-    --max-model-len 10000 \
-    --gpu-memory-utilization 0.6 \
-    --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
-
-  wait_for_server 8100
-  wait_for_server 8200
-
-  # let the prefill instance finish prefill
-  vllm bench serve \
-    --backend vllm \
-    --model $model \
-    --dataset-name $dataset_name \
-    --dataset-path $dataset_path \
-    --sonnet-input-len $input_len \
-    --sonnet-output-len "$output_len" \
-    --sonnet-prefix-len $prefix_len \
-    --num-prompts $num_prompts \
-    --port 8100 \
-    --save-result \
-    --result-dir $results_folder \
-    --result-filename disagg_prefill_tp1.json \
-    --request-rate "inf"
-
-
-  # send the request to decode.
-  # The TTFT of this command will be the overhead of disagg prefill impl.
-  vllm bench serve \
-    --backend vllm \
-    --model $model \
-    --dataset-name $dataset_name \
-    --dataset-path $dataset_path \
-    --sonnet-input-len $input_len \
-    --sonnet-output-len "$output_len" \
-    --sonnet-prefix-len $prefix_len \
-    --num-prompts $num_prompts \
-    --port 8200 \
-    --save-result \
-    --result-dir $results_folder \
-    --result-filename disagg_prefill_tp1_overhead.json \
-    --request-rate "$qps"
-  kill_gpu_processes
-
-}
-
-
-main() {
-
-  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
-  (which jq) || (apt-get -y install jq)
-  (which socat) || (apt-get -y install socat)
-
-  pip install quart httpx datasets
-
-  cd "$(dirname "$0")"
-
-  cd ..
-  # create sonnet-4x.txt
-  echo "" > sonnet_4x.txt
-  for _ in {1..4}
-  do
-    cat sonnet.txt >> sonnet_4x.txt
-  done
-  cd disagg_benchmarks
-
-  rm -rf results
-  mkdir results
-
-  default_qps=1
-  default_output_len=1
-  benchmark $default_qps $default_output_len
-
-}
-
-
-main "$@"

benchmarks/disagg_benchmarks/disagg_performance_benchmark.sh

@@ -1,157 +0,0 @@
-#!/bin/bash
-
-# Requirement: 2x GPUs.
-
-
-# Model: meta-llama/Meta-Llama-3.1-8B-Instruct
-# Query: 1024 input tokens, 6 output tokens, QPS 2/4/6/8, 100 requests
-# Resource: 2x GPU
-# Approaches:
-# 2. Chunked prefill: 2 vllm instance with tp=4, equivalent to 1 tp=4 instance with QPS 4
-# 3. Disaggregated prefill: 1 prefilling instance and 1 decoding instance
-# Prefilling instance: max_output_token=1
-# Decoding instance: force the input tokens be the same across requests to bypass prefilling
-
-set -ex
-
-kill_gpu_processes() {
-  # kill all processes on GPU.
-  pgrep pt_main_thread | xargs -r kill -9
-  pgrep python3 | xargs -r kill -9
-  # vLLM now names the process with VLLM prefix after https://github.com/vllm-project/vllm/pull/21445
-  pgrep VLLM | xargs -r kill -9
-  for port in 8000 8100 8200; do lsof -t -i:$port | xargs -r kill -9; done
-  sleep 1
-}
-
-wait_for_server() {
-  # wait for vllm server to start
-  # return 1 if vllm server crashes
-  local port=$1
-  timeout 1200 bash -c "
-    until curl -s localhost:${port}/v1/completions > /dev/null; do
-      sleep 1
-    done" && return 0 || return 1
-}
-
-
-launch_chunked_prefill() {
-  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
-  # disagg prefill
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
-    --port 8100 \
-    --max-model-len 10000 \
-    --enable-chunked-prefill \
-    --gpu-memory-utilization 0.6 &
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
-    --port 8200 \
-    --max-model-len 10000 \
-    --enable-chunked-prefill \
-    --gpu-memory-utilization 0.6 &
-  wait_for_server 8100
-  wait_for_server 8200
-  python3 round_robin_proxy.py &
-  sleep 1
-}
-
-
-launch_disagg_prefill() {
-  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
-  # disagg prefill
-  CUDA_VISIBLE_DEVICES=0 vllm serve $model \
-    --port 8100 \
-    --max-model-len 10000 \
-    --gpu-memory-utilization 0.6 \
-    --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_producer","kv_rank":0,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
-
-  CUDA_VISIBLE_DEVICES=1 vllm serve $model \
-    --port 8200 \
-    --max-model-len 10000 \
-    --gpu-memory-utilization 0.6 \
-    --kv-transfer-config \
-    '{"kv_connector":"P2pNcclConnector","kv_role":"kv_consumer","kv_rank":1,"kv_parallel_size":2,"kv_buffer_size":5e9}' &
-
-  wait_for_server 8100
-  wait_for_server 8200
-  python3 disagg_prefill_proxy_server.py &
-  sleep 1
-}
-
-
-benchmark() {
-  results_folder="./results"
-  model="meta-llama/Meta-Llama-3.1-8B-Instruct"
-  dataset_name="sonnet"
-  dataset_path="../sonnet_4x.txt"
-  num_prompts=100
-  qps=$1
-  prefix_len=50
-  input_len=1024
-  output_len=$2
-  tag=$3
-
-  vllm bench serve \
-    --backend vllm \
-    --model $model \
-    --dataset-name $dataset_name \
-    --dataset-path $dataset_path \
-    --sonnet-input-len $input_len \
-    --sonnet-output-len "$output_len" \
-    --sonnet-prefix-len $prefix_len \
-    --num-prompts $num_prompts \
-    --port 8000 \
-    --save-result \
-    --result-dir $results_folder \
-    --result-filename "$tag"-qps-"$qps".json \
-    --request-rate "$qps"
-
-  sleep 2
-}
-
-
-main() {
-
-  (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
-  (which jq) || (apt-get -y install jq)
-  (which socat) || (apt-get -y install socat)
-  (which lsof) || (apt-get -y install lsof)
-
-  pip install quart httpx matplotlib aiohttp datasets
-
-  cd "$(dirname "$0")"
-
-  cd ..
-  # create sonnet-4x.txt so that we can sample 2048 tokens for input
-  echo "" > sonnet_4x.txt
-  for _ in {1..4}
-  do
-    cat sonnet.txt >> sonnet_4x.txt
-  done
-  cd disagg_benchmarks
-
-  rm -rf results
-  mkdir results
-
-  default_output_len=6
-
-  export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
-
-  launch_chunked_prefill
-  for qps in 2 4 6 8; do
-  benchmark $qps $default_output_len chunked_prefill
-  done
-  kill_gpu_processes
-
-  launch_disagg_prefill
-  for qps in 2 4 6 8; do
-  benchmark $qps $default_output_len disagg_prefill
-  done
-  kill_gpu_processes
-
-  python3 visualize_benchmark_results.py
-
-}
-
-
-main "$@"

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

@@ -1,260 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import argparse
-import asyncio
-import logging
-import os
-import time
-import uuid
-from urllib.parse import urlparse
-
-import aiohttp
-from quart import Quart, Response, make_response, request
-
-# Configure logging
-logging.basicConfig(level=logging.INFO)
-logger = logging.getLogger(__name__)
-
-
-def parse_args():
-    """parse command line arguments"""
-    parser = argparse.ArgumentParser(description="vLLM P/D disaggregation proxy server")
-
-    # Add args
-    parser.add_argument(
-        "--timeout",
-        type=float,
-        default=6 * 60 * 60,
-        help="Timeout for backend service requests in seconds (default: 21600)",
-    )
-    parser.add_argument(
-        "--port",
-        type=int,
-        default=8000,
-        help="Port to run the server on (default: 8000)",
-    )
-    parser.add_argument(
-        "--prefill-url",
-        type=str,
-        default="http://localhost:8100",
-        help="Prefill service base URL (protocol + host[:port])",
-    )
-    parser.add_argument(
-        "--decode-url",
-        type=str,
-        default="http://localhost:8200",
-        help="Decode service base URL (protocol + host[:port])",
-    )
-    parser.add_argument(
-        "--kv-host",
-        type=str,
-        default="localhost",
-        help="Hostname or IP used by KV transfer (default: localhost)",
-    )
-    parser.add_argument(
-        "--prefill-kv-port",
-        type=int,
-        default=14579,
-        help="Prefill KV port (default: 14579)",
-    )
-    parser.add_argument(
-        "--decode-kv-port",
-        type=int,
-        default=14580,
-        help="Decode KV port (default: 14580)",
-    )
-
-    return parser.parse_args()
-
-
-def main():
-    """parse command line arguments"""
-    args = parse_args()
-
-    # Initialize configuration using command line parameters
-    AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=args.timeout)
-    PREFILL_SERVICE_URL = args.prefill_url
-    DECODE_SERVICE_URL = args.decode_url
-    PORT = args.port
-
-    PREFILL_KV_ADDR = f"{args.kv_host}:{args.prefill_kv_port}"
-    DECODE_KV_ADDR = f"{args.kv_host}:{args.decode_kv_port}"
-
-    logger.info(
-        "Proxy resolved KV addresses -> prefill: %s, decode: %s",
-        PREFILL_KV_ADDR,
-        DECODE_KV_ADDR,
-    )
-
-    app = Quart(__name__)
-
-    # Attach the configuration object to the application instance so helper
-    # coroutines can read the resolved backend URLs and timeouts without using
-    # globals.
-    app.config.update(
-        {
-            "AIOHTTP_TIMEOUT": AIOHTTP_TIMEOUT,
-            "PREFILL_SERVICE_URL": PREFILL_SERVICE_URL,
-            "DECODE_SERVICE_URL": DECODE_SERVICE_URL,
-            "PREFILL_KV_ADDR": PREFILL_KV_ADDR,
-            "DECODE_KV_ADDR": DECODE_KV_ADDR,
-        }
-    )
-
-    def _normalize_base_url(url: str) -> str:
-        """Remove any trailing slash so path joins behave predictably."""
-        return url.rstrip("/")
-
-    def _get_host_port(url: str) -> str:
-        """Return the hostname:port portion for logging and KV headers."""
-        parsed = urlparse(url)
-        host = parsed.hostname or "localhost"
-        port = parsed.port
-        if port is None:
-            port = 80 if parsed.scheme == "http" else 443
-        return f"{host}:{port}"
-
-    PREFILL_BASE = _normalize_base_url(PREFILL_SERVICE_URL)
-    DECODE_BASE = _normalize_base_url(DECODE_SERVICE_URL)
-    KV_TARGET = _get_host_port(DECODE_SERVICE_URL)
-
-    def _build_headers(request_id: str) -> dict[str, str]:
-        """Construct the headers expected by vLLM's P2P disagg connector."""
-        headers: dict[str, str] = {"X-Request-Id": request_id, "X-KV-Target": KV_TARGET}
-        api_key = os.environ.get("OPENAI_API_KEY")
-        if api_key:
-            headers["Authorization"] = f"Bearer {api_key}"
-        return headers
-
-    async def _run_prefill(
-        request_path: str,
-        payload: dict,
-        headers: dict[str, str],
-        request_id: str,
-    ):
-        url = f"{PREFILL_BASE}{request_path}"
-        start_ts = time.perf_counter()
-        logger.info("[prefill] start request_id=%s url=%s", request_id, url)
-        try:
-            async with (
-                aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session,
-                session.post(url=url, json=payload, headers=headers) as resp,
-            ):
-                if resp.status != 200:
-                    error_text = await resp.text()
-                    raise RuntimeError(
-                        f"Prefill backend error {resp.status}: {error_text}"
-                    )
-                await resp.read()
-                logger.info(
-                    "[prefill] done request_id=%s status=%s elapsed=%.2fs",
-                    request_id,
-                    resp.status,
-                    time.perf_counter() - start_ts,
-                )
-        except asyncio.TimeoutError as exc:
-            raise RuntimeError(f"Prefill service timeout at {url}") from exc
-        except aiohttp.ClientError as exc:
-            raise RuntimeError(f"Prefill service unavailable at {url}") from exc
-
-    async def _stream_decode(
-        request_path: str,
-        payload: dict,
-        headers: dict[str, str],
-        request_id: str,
-    ):
-        url = f"{DECODE_BASE}{request_path}"
-        # Stream tokens from the decode service once the prefill stage has
-        # materialized KV caches on the target workers.
-        logger.info("[decode] start request_id=%s url=%s", request_id, url)
-        try:
-            async with (
-                aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session,
-                session.post(url=url, json=payload, headers=headers) as resp,
-            ):
-                if resp.status != 200:
-                    error_text = await resp.text()
-                    logger.error(
-                        "Decode backend error %s - %s", resp.status, error_text
-                    )
-                    err_msg = (
-                        '{"error": "Decode backend error ' + str(resp.status) + '"}'
-                    )
-                    yield err_msg.encode()
-                    return
-                logger.info(
-                    "[decode] streaming response request_id=%s status=%s",
-                    request_id,
-                    resp.status,
-                )
-                async for chunk_bytes in resp.content.iter_chunked(1024):
-                    yield chunk_bytes
-                logger.info("[decode] finished streaming request_id=%s", request_id)
-        except asyncio.TimeoutError:
-            logger.error("Decode service timeout at %s", url)
-            yield b'{"error": "Decode service timeout"}'
-        except aiohttp.ClientError as exc:
-            logger.error("Decode service error at %s: %s", url, exc)
-            yield b'{"error": "Decode service unavailable"}'
-
-    async def process_request():
-        """Process a single request through prefill and decode stages"""
-        try:
-            original_request_data = await request.get_json()
-
-            # Create prefill request (max_tokens=1)
-            prefill_request = original_request_data.copy()
-            prefill_request["max_tokens"] = 1
-            if "max_completion_tokens" in prefill_request:
-                prefill_request["max_completion_tokens"] = 1
-
-            # Execute prefill stage
-            # The request id encodes both KV socket addresses so the backend can
-            # shuttle tensors directly via NCCL once the prefill response
-            # completes.
-            request_id = (
-                f"___prefill_addr_{PREFILL_KV_ADDR}___decode_addr_"
-                f"{DECODE_KV_ADDR}_{uuid.uuid4().hex}"
-            )
-
-            headers = _build_headers(request_id)
-            await _run_prefill(request.path, prefill_request, headers, request_id)
-
-            # Execute decode stage and stream response
-            # Pass the unmodified user request so the decode phase can continue
-            # sampling with the already-populated KV cache.
-            generator = _stream_decode(
-                request.path, original_request_data, headers, request_id
-            )
-            response = await make_response(generator)
-            response.timeout = None  # Disable timeout for streaming response
-            return response
-
-        except Exception:
-            logger.exception("Error processing request")
-            return Response(
-                response=b'{"error": "Internal server error"}',
-                status=500,
-                content_type="application/json",
-            )
-
-    @app.route("/v1/completions", methods=["POST"])
-    async def handle_request():
-        """Handle incoming API requests with concurrency and rate limiting"""
-        try:
-            return await process_request()
-        except asyncio.CancelledError:
-            logger.warning("Request cancelled")
-            return Response(
-                response=b'{"error": "Request cancelled"}',
-                status=503,
-                content_type="application/json",
-            )
-
-    # Start the Quart server with host can be set to 0.0.0.0
-    app.run(port=PORT)
-
-
-if __name__ == "__main__":
-    main()

benchmarks/disagg_benchmarks/round_robin_proxy.py

@@ -1,63 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import asyncio
-import itertools
-
-import aiohttp
-from aiohttp import web
-
-
-class RoundRobinProxy:
-    def __init__(self, target_ports):
-        self.target_ports = target_ports
-        self.port_cycle = itertools.cycle(self.target_ports)
-
-    async def handle_request(self, request):
-        target_port = next(self.port_cycle)
-        target_url = f"http://localhost:{target_port}{request.path_qs}"
-
-        async with aiohttp.ClientSession() as session:
-            try:
-                # Forward the request
-                async with session.request(
-                    method=request.method,
-                    url=target_url,
-                    headers=request.headers,
-                    data=request.content,
-                ) as response:
-                    # Start sending the response
-                    resp = web.StreamResponse(
-                        status=response.status, headers=response.headers
-                    )
-                    await resp.prepare(request)
-
-                    # Stream the response content
-                    async for chunk in response.content.iter_any():
-                        await resp.write(chunk)
-
-                    await resp.write_eof()
-                    return resp
-
-            except Exception as e:
-                return web.Response(text=f"Error: {str(e)}", status=500)
-
-
-async def main():
-    proxy = RoundRobinProxy([8100, 8200])
-    app = web.Application()
-    app.router.add_route("*", "/{path:.*}", proxy.handle_request)
-
-    runner = web.AppRunner(app)
-    await runner.setup()
-    site = web.TCPSite(runner, "localhost", 8000)
-    await site.start()
-
-    print("Proxy server started on http://localhost:8000")
-
-    # Keep the server running
-    await asyncio.Event().wait()
-
-
-if __name__ == "__main__":
-    asyncio.run(main())

benchmarks/disagg_benchmarks/visualize_benchmark_results.py

@@ -1,47 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import json
-
-import matplotlib.pyplot as plt
-import pandas as pd
-
-if __name__ == "__main__":
-    data = []
-    for name in ["disagg_prefill", "chunked_prefill"]:
-        for qps in [2, 4, 6, 8]:
-            with open(f"results/{name}-qps-{qps}.json") as f:
-                x = json.load(f)
-                x["name"] = name
-                x["qps"] = qps
-                data.append(x)
-
-    df = pd.DataFrame.from_dict(data)
-    dis_df = df[df["name"] == "disagg_prefill"]
-    chu_df = df[df["name"] == "chunked_prefill"]
-
-    plt.style.use("bmh")
-    plt.rcParams["font.size"] = 20
-
-    for key in [
-        "mean_ttft_ms",
-        "median_ttft_ms",
-        "p99_ttft_ms",
-        "mean_itl_ms",
-        "median_itl_ms",
-        "p99_itl_ms",
-    ]:
-        fig, ax = plt.subplots(figsize=(11, 7))
-        plt.plot(
-            dis_df["qps"], dis_df[key], label="disagg_prefill", marker="o", linewidth=4
-        )
-        plt.plot(
-            chu_df["qps"], chu_df[key], label="chunked_prefill", marker="o", linewidth=4
-        )
-        ax.legend()
-
-        ax.set_xlabel("QPS")
-        ax.set_ylabel(key)
-        ax.set_ylim(bottom=0)
-        fig.savefig(f"results/{key}.png")
-        plt.close(fig)

benchmarks/kernels/benchmark_device_communicators.py

@@ -33,6 +33,7 @@ from vllm.distributed.device_communicators.custom_all_reduce import CustomAllred
 from vllm.distributed.device_communicators.flashinfer_all_reduce import (
     FlashInferAllReduce,
 )
+from vllm.distributed.device_communicators.push_all_reduce import PushAllReduce
 from vllm.distributed.device_communicators.pynccl import (
     PyNcclCommunicator,
     register_nccl_symmetric_ops,
@@ -80,6 +81,7 @@ class CommunicatorBenchmark:
 
         # Initialize communicators
         self.custom_allreduce = None
+        self.push_ar_comm = None
         self.pynccl_comm = None
         self.symm_mem_comm = None
         self.symm_mem_comm_multimem = None
@@ -106,6 +108,23 @@ class CommunicatorBenchmark:
             )
             self.custom_allreduce = None
 
+        try:
+            self.push_ar_comm = PushAllReduce(
+                group=self.cpu_group,
+                device=self.device,
+                max_size=self.max_size_override,
+            )
+            if not self.push_ar_comm.disabled:
+                logger.info("Rank %s: PushAllReduce initialized", self.rank)
+            else:
+                logger.info("Rank %s: PushAllReduce disabled", self.rank)
+                self.push_ar_comm = None
+        except Exception as e:
+            logger.warning(
+                "Rank %s: Failed to initialize PushAllReduce: %s", self.rank, e
+            )
+            self.push_ar_comm = None
+
         try:
             self.pynccl_comm = PyNcclCommunicator(
                 group=self.cpu_group, device=self.device
@@ -216,6 +235,19 @@ class CommunicatorBenchmark:
                 )
             )
 
+        if self.push_ar_comm is not None:
+            comm = self.push_ar_comm
+            communicators.append(
+                (
+                    "push_ar",
+                    lambda t, c=comm: c.all_reduce(t),
+                    lambda t, c=comm: c.should_use(t),
+                    comm.capture(),
+                    {},
+                    None,
+                )
+            )
+
         if self.pynccl_comm is not None:
             comm = self.pynccl_comm
             communicators.append(

benchmarks/kernels/benchmark_flydsl_moe_w4a16.py

@@ -0,0 +1,277 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# Copyright (c) 2025 FlyDSL Project Contributors
+
+import json
+import os
+
+import torch
+from aiter.test_common import run_perftest
+
+from vllm.model_executor.layers.fused_moe import fused_experts
+from vllm.model_executor.layers.fused_moe.activation import MoEActivation
+from vllm.model_executor.layers.fused_moe.config import (
+    int4_w4a16_moe_quant_config,
+)
+from vllm.model_executor.layers.fused_moe.fused_flydsl_moe import fused_flydsl_moe
+from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe import (  # noqa: E501
+    compressed_tensors_moe_w4a16_flydsl,
+)
+from vllm.platforms import current_platform
+
+RoutingBuffers = tuple[
+    torch.Tensor,  # sorted_token_ids
+    torch.Tensor,  # sorted_weights
+    torch.Tensor,  # sorted_expert_ids
+    torch.Tensor,  # num_valid_ids (shape [1], i32)
+    int,  # sorted_size
+    int,  # blocks
+]
+
+MODEL_PARAMS_TO_TUNE = [
+    # (num_experts, inter_dim, hidden_size, topk)
+    (384, 256, 7168, 8),  # Kimi K2.5 TP=8
+    (384, 512, 7168, 8),  # Kimi K2.5 TP=4
+]
+
+NUM_TOKENS_TO_TUNE = [
+    1,
+    2,
+    4,
+    8,
+    16,
+    24,
+    32,
+    48,
+    64,
+    128,
+    256,
+    512,
+    1024,
+    2048,
+    4096,
+    8192,
+]
+
+TILE_M_SEARCH_SPACE = [16, 32, 64, 128, 256]
+TILE_N_SEARCH_SPACE = [16, 32, 64, 128, 256]
+TILE_K_SEARCH_SPACE = [16, 32, 64, 128, 256, 512]
+TILE_N2_SEARCH_SPACE = [16, 32, 64, 128, 256]
+TILE_K2_SEARCH_SPACE = [16, 32, 64, 128, 256, 512]
+
+TILE_CONFIGS = []
+for tile_m in TILE_M_SEARCH_SPACE:
+    for tile_n in TILE_N_SEARCH_SPACE:
+        for tile_k in TILE_K_SEARCH_SPACE:
+            for tile_n2 in TILE_N2_SEARCH_SPACE:
+                for tile_k2 in TILE_K2_SEARCH_SPACE:
+                    TILE_CONFIGS.append(
+                        {
+                            "tile_m": tile_m,
+                            "tile_n": tile_n,
+                            "tile_k": tile_k,
+                            "tile_n2": tile_n2,
+                            "tile_k2": tile_k2,
+                        }
+                    )
+
+
+def tune_flydsl_moe_w4a16(
+    device: str = "cuda", num_iters: int = 100, num_warmup: int = 10
+):
+    packed_factor = 8
+    w13_num_shards = 2
+    params_dtype = torch.bfloat16
+    group_size = 32
+    scale_factor = 0.01
+
+    for model_params in MODEL_PARAMS_TO_TUNE:
+        num_experts = model_params[0]
+        inter_dim = model_params[1]
+        hidden_size = model_params[2]
+        topk = model_params[3]
+        print(
+            f"\nTuning: num_experts={num_experts}, inter_dim={inter_dim}, "
+            f"hidden_size={hidden_size}, topk={topk}...\n"
+        )
+
+        w2_scales_size = inter_dim
+        num_groups_w2 = w2_scales_size // group_size
+        num_groups_w13 = hidden_size // group_size
+
+        w13_weight = torch.randint(
+            0,
+            255,
+            (num_experts, hidden_size // packed_factor, w13_num_shards * inter_dim),
+            dtype=torch.int32,
+            device=device,
+        )
+
+        w2_weight = torch.randint(
+            0,
+            255,
+            (num_experts, inter_dim // packed_factor, hidden_size),
+            dtype=torch.int32,
+            device=device,
+        )
+        w13_scale = scale_factor * torch.randn(
+            num_experts,
+            num_groups_w13,
+            w13_num_shards * inter_dim,
+            dtype=params_dtype,
+            device=device,
+        )
+        w2_scale = scale_factor * torch.randn(
+            num_experts, num_groups_w2, hidden_size, dtype=params_dtype, device=device
+        )
+
+        w13 = w13_weight
+        w13 = compressed_tensors_moe_w4a16_flydsl._gptq_int32_to_flydsl_packed(w13)
+        w13 = w13.view(-1).contiguous()
+
+        w2 = w2_weight
+        w2 = compressed_tensors_moe_w4a16_flydsl._gptq_int32_to_flydsl_packed(w2)
+        w2 = w2.view(-1).contiguous()
+
+        w13_scale_flydsl = w13_scale
+        w2_scale_flydsl = w2_scale
+
+        if group_size > 0 and w13_scale.dim() == 3 and w13_scale.shape[1] > 1:
+            E, G, N = w13_scale.shape
+            w13_scale_flydsl = (
+                w13_scale_flydsl.view(E, G // 2, 2, N)
+                .permute(0, 1, 3, 2)
+                .contiguous()
+                .view(-1)
+                .contiguous()
+            )
+        elif w13_scale.dim() == 3 and w13_scale.shape[1] == 1:
+            w13_scale_flydsl = w13_scale_flydsl.squeeze(1)
+
+        if group_size > 0 and w2_scale.dim() == 3 and w2_scale.shape[1] > 1:
+            E, G, N = w2_scale.shape
+            w2_scale_flydsl = (
+                w2_scale_flydsl.view(E, G // 2, 2, N)
+                .permute(0, 1, 3, 2)
+                .contiguous()
+                .view(-1)
+                .contiguous()
+            )
+        elif w2_scale.dim() == 3 and w2_scale.shape[1] == 1:
+            w2_scale_flydsl = w2_scale_flydsl.squeeze(1)
+
+        w13_scale_flydsl = w13_scale_flydsl.contiguous()
+        w2_scale_flydsl = w2_scale_flydsl.contiguous()
+
+        w13.is_shuffled = True
+        w2.is_shuffled = True
+
+        w13_weight_scale = w13_scale.transpose(1, 2).contiguous()
+        w2_weight_scale = w2_scale.transpose(1, 2).contiguous()
+        w13_weight_packed = w13_weight.transpose(1, 2).contiguous().view(torch.uint8)
+        w2_weight_packed = w2_weight.transpose(1, 2).contiguous().view(torch.uint8)
+
+        moe_quant_config = int4_w4a16_moe_quant_config(
+            w1_scale=w13_weight_scale,
+            w2_scale=w2_weight_scale,
+            w1_zp=None,
+            w2_zp=None,
+            block_shape=[0, group_size],
+        )
+
+        tuned_config = {}
+
+        for num_tokens in NUM_TOKENS_TO_TUNE:
+            score = torch.rand(
+                (num_tokens, num_experts), device=device, dtype=torch.float32
+            )
+            topk_vals, topk_ids = torch.topk(score, k=topk, dim=1)
+            topk_weights = torch.softmax(topk_vals, dim=1).to(torch.float32)
+            x = torch.randn(
+                (num_tokens, hidden_size), dtype=torch.bfloat16, device=device
+            )
+            us_best = float("inf")
+            for tile_config in TILE_CONFIGS:
+                try:
+                    tile_m = tile_config["tile_m"]
+                    tile_n = tile_config["tile_n"]
+                    tile_k = tile_config["tile_k"]
+                    tile_n2 = tile_config["tile_n2"]
+                    tile_k2 = tile_config["tile_k2"]
+
+                    model_dim = x.shape[1]
+                    assert model_dim % 64 == 0
+                    assert model_dim % tile_k == 0
+                    assert inter_dim % tile_n == 0
+                    assert model_dim % tile_n2 == 0
+                    assert inter_dim % tile_k2 == 0
+                    assert ((tile_m * tile_k2) % 256) == 0
+                    bytes_per_thread_x = (tile_m * tile_k2) // 256
+                    assert (bytes_per_thread_x % 4) == 0
+
+                    out, _us = run_perftest(
+                        fused_flydsl_moe,
+                        x,
+                        w13,
+                        w2,
+                        num_experts,
+                        inter_dim,
+                        topk_weights,
+                        topk_ids,
+                        num_iters=num_iters,
+                        num_warmup=num_warmup,
+                        w1_scale=w13_scale_flydsl,
+                        w2_scale=w2_scale_flydsl,
+                        topk=topk_weights.shape[-1],
+                        group_size=group_size,
+                        doweight_stage1=False,
+                        scale_is_bf16=True,
+                        config=tile_config,
+                    )
+                    torch.accelerator.synchronize()
+                except Exception:
+                    torch.accelerator.synchronize()
+                    continue
+                else:
+                    us = _us.item()
+                    if us < us_best:
+                        out_ref = fused_experts(
+                            x,
+                            w13_weight_packed,
+                            w2_weight_packed,
+                            topk_weights=topk_weights,
+                            topk_ids=topk_ids,
+                            activation=MoEActivation.SILU,
+                            apply_router_weight_on_input=False,
+                            global_num_experts=num_experts,
+                            expert_map=None,
+                            quant_config=moe_quant_config,
+                        )
+                        try:
+                            assert torch.allclose(out, out_ref, atol=0.5, rtol=0.1)
+                        except Exception:
+                            continue
+                        else:
+                            print(
+                                f"For [num_tokens={num_tokens}, num_experts={num_experts}, "  # noqa: E501
+                                f"inter_dim={inter_dim}] found new best "  # noqa: E501
+                                f"config={tile_config}, us={us:0.3f}"
+                            )
+                            us_best = us
+                            tuned_config[str(num_tokens)] = tile_config
+            device_name = current_platform.get_device_name().replace(" ", "_")
+            tuned_config_file_name = (
+                f"E={num_experts},N={inter_dim},device_name={device_name},"
+                f"dtype=int4_w4a16,backend=flydsl.json"
+            )
+            tuner_dir_path = os.path.dirname(os.path.realpath(__file__))
+            store_path = os.path.join(tuner_dir_path, tuned_config_file_name)
+            with open(store_path, "w") as f:
+                json.dump(tuned_config, f, indent=4)
+            print(
+                f"\nTuned config for num_tokens={num_tokens} was stored at {store_path}\n"  # noqa: E501
+            )
+
+
+if __name__ == "__main__":
+    tune_flydsl_moe_w4a16(device="cuda")

benchmarks/kernels/benchmark_moe.py

@@ -250,7 +250,7 @@ def benchmark_config(
                     num_experts=num_experts,
                     experts_per_token=topk,
                     hidden_dim=hidden_size,
-                    intermediate_size_per_partition=shard_intermediate_size,
+                    intermediate_size=shard_intermediate_size,
                     num_local_experts=num_experts,
                     num_logical_experts=num_experts,
                     activation=MoEActivation.SILU,
@@ -271,7 +271,6 @@ def benchmark_config(
                     moe_config=moe_config,
                     quant_config=quant_config,
                 ),
-                inplace=not disable_inplace(),
             )
 
         with override_config(config):
@@ -279,7 +278,6 @@ def benchmark_config(
                 x, input_gating, topk, renormalize=not use_deep_gemm
             )
 
-            inplace = not disable_inplace()
             if use_deep_gemm:
                 return deep_gemm_experts.apply(
                     x,
@@ -298,7 +296,6 @@ def benchmark_config(
                 w2,
                 topk_weights,
                 topk_ids,
-                inplace=inplace,
                 quant_config=quant_config,
             )
 
@@ -795,6 +792,12 @@ def get_model_params(config):
         topk = text_config.num_experts_per_tok
         intermediate_size = text_config.moe_intermediate_size
         hidden_size = text_config.hidden_size
+    elif architecture == "DiffusionGemmaForBlockDiffusion":
+        text_config = config.get_text_config()
+        E = text_config.num_experts
+        topk = text_config.top_k_experts
+        intermediate_size = text_config.moe_intermediate_size
+        hidden_size = text_config.hidden_size
     elif architecture == "HunYuanMoEV1ForCausalLM":
         E = config.num_experts
         topk = config.moe_topk[0]

benchmarks/kv_cache_watermark.sh

@@ -0,0 +1,248 @@
+#!/bin/bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Reproducible demonstration of the KV cache watermark (`--watermark`) for
+# reducing preemption thrashing.
+#
+# The watermark is the fraction of total KV cache blocks the scheduler keeps
+# free when admitting a waiting/preempted request into the running queue.
+#
+# Why this workload triggers thrashing:
+#   Requests are admitted based on the KV cache they need *at admission time*.
+#   With `--scheduler-reserve-full-isl` (default) the input length is reserved up
+#   front, but the *output* length is unknown and unreserved. A decode-heavy
+#   workload (output >> input) at high concurrency therefore over-admits while
+#   requests are short, then runs out of KV cache as they all grow during decode
+#   -> the scheduler preempts (recompute) recently-admitted requests, re-prefills
+#   them later, and repeats. The watermark keeps a block of KV cache free so
+#   running requests can grow into it instead of triggering this churn.
+#
+# This script launches `vllm serve` under a deliberately KV-constrained config
+# and a decode-heavy workload, sweeping the watermark across several values, and
+# reports the preemption count (scraped from /metrics), throughput, and latency
+# percentiles for each. It then plots the results.
+#
+# Default workload: concurrency 200, input ~300 tokens, output ~4000 tokens
+# (+/- 20% variance), sized to run each config for ~5 minutes.
+#
+# Usage:
+#   benchmarks/kv_cache_watermark.sh
+#   MODEL=Qwen/Qwen2.5-14B-Instruct TP=2 benchmarks/kv_cache_watermark.sh
+#
+# Run inside the vLLM virtualenv (so `vllm` and `python` resolve to it).
+set -euo pipefail
+
+# ---- Config (override via environment) -------------------------------------
+MODEL=${MODEL:-Qwen/Qwen2.5-7B-Instruct}
+TP=${TP:-1}
+PORT=${PORT:-8000}
+URL="http://127.0.0.1:${PORT}"
+# Constrain the KV cache to a *near-critical* size: large enough that the engine
+# can run stably, but small enough that greedy over-admission tips it into
+# preemption thrashing. (Independent of GPU size, so the demo is reproducible.)
+# At the default workload this fits ~1.5x the mean concurrent KV demand.
+KV_CACHE_MEMORY_GB=${KV_CACHE_MEMORY_GB:-16}
+MAX_MODEL_LEN=${MAX_MODEL_LEN:-8192}
+MAX_NUM_SEQS=${MAX_NUM_SEQS:-256}
+# Optional weight loader (e.g. fastsafetensors on the GCP cluster).
+LOAD_FORMAT=${LOAD_FORMAT:-auto}
+# Decode-heavy workload: moderate input, long output, with length variance. The
+# long output means preempted requests have generated a lot before eviction, so
+# resuming them re-prefills a long sequence (high recomputation cost).
+INPUT_LEN=${INPUT_LEN:-1000}
+OUTPUT_LEN=${OUTPUT_LEN:-5000}
+RANGE_RATIO=${RANGE_RATIO:-0.2}
+CONCURRENCY=${CONCURRENCY:-128}
+# Enough prompts to keep each config saturated for ~5+ minutes.
+NUM_PROMPTS=${NUM_PROMPTS:-450}
+OUTDIR=${OUTDIR:-./watermark_bench_results}
+# Watermark fractions compared. "label value" per line; value=0 disables it.
+CONFIGS=${CONFIGS:-"off    0
+w0.02  0.02
+w0.05  0.05
+w0.10  0.10
+w0.15  0.15"}
+
+KV_CACHE_MEMORY_BYTES=$((KV_CACHE_MEMORY_GB * 1024 * 1024 * 1024))
+mkdir -p "$OUTDIR"
+
+SERVER_PID=""
+cleanup() { [[ -n "$SERVER_PID" ]] && kill "$SERVER_PID" 2>/dev/null || true; }
+trap cleanup EXIT
+
+scrape_preemptions() {
+  # Sum the vllm:num_preemptions_total counter across engines.
+  python - "${URL}/metrics" <<'PY'
+import sys, urllib.request
+total = 0.0
+try:
+    body = urllib.request.urlopen(sys.argv[1], timeout=10).read().decode("utf-8", "replace")
+    for line in body.splitlines():
+        if line.startswith("vllm:num_preemptions_total"):
+            total += float(line.rsplit(" ", 1)[-1])
+except Exception as e:  # noqa: BLE001
+    print(f"scrape error: {e}", file=sys.stderr)
+print(int(total))
+PY
+}
+
+wait_for_server() {
+  for _ in $(seq 1 300); do
+    if curl -s "${URL}/health" >/dev/null 2>&1; then return 0; fi
+    if ! kill -0 "$SERVER_PID" 2>/dev/null; then
+      echo "ERROR: server process exited during startup" >&2; return 1
+    fi
+    sleep 5
+  done
+  echo "ERROR: server did not become ready" >&2; return 1
+}
+
+run_one() {
+  local label=$1 watermark=$2
+  echo
+  echo "==================== watermark: ${label} (${watermark}) ===================="
+  vllm serve "$MODEL" \
+    --tensor-parallel-size "$TP" \
+    --load-format "$LOAD_FORMAT" \
+    --kv-cache-memory-bytes "$KV_CACHE_MEMORY_BYTES" \
+    --max-model-len "$MAX_MODEL_LEN" \
+    --max-num-seqs "$MAX_NUM_SEQS" \
+    --no-enable-prefix-caching \
+    --watermark "$watermark" \
+    --port "$PORT" >"${OUTDIR}/serve_${label}.log" 2>&1 &
+  SERVER_PID=$!
+  wait_for_server
+  sleep 5
+
+  local pre post
+  pre=$(scrape_preemptions)
+  vllm bench serve \
+    --backend vllm \
+    --base-url "$URL" \
+    --model "$MODEL" \
+    --dataset-name random \
+    --random-input-len "$INPUT_LEN" \
+    --random-output-len "$OUTPUT_LEN" \
+    --random-range-ratio "$RANGE_RATIO" \
+    --ignore-eos \
+    --num-prompts "$NUM_PROMPTS" \
+    --max-concurrency "$CONCURRENCY" \
+    --percentile-metrics "ttft,tpot,itl,e2el" \
+    --metric-percentiles "50,90,99" \
+    --save-result \
+    --result-dir "$OUTDIR" \
+    --result-filename "bench_${label}.json"
+  post=$(scrape_preemptions)
+  echo "${label} ${watermark} $((post - pre))" >>"${OUTDIR}/preemptions.txt"
+
+  kill "$SERVER_PID" 2>/dev/null || true
+  for _ in $(seq 1 60); do curl -s "${URL}/health" >/dev/null 2>&1 || break; sleep 2; done
+  SERVER_PID=""
+  sleep 10
+}
+
+: >"${OUTDIR}/preemptions.txt"
+while read -r label watermark; do
+  [[ -z "${label:-}" ]] && continue
+  run_one "$label" "$watermark"
+done <<<"$CONFIGS"
+
+echo
+echo "==================== summary ===================="
+python - "$OUTDIR" <<'PY'
+import json, os, sys
+outdir = sys.argv[1]
+pre = {}
+order = []
+for line in open(os.path.join(outdir, "preemptions.txt")):
+    label, watermark, n = line.split()
+    pre[label] = (float(watermark), int(n))
+    order.append(label)
+
+def g(d, *names):
+    for n in names:
+        if d.get(n) is not None:
+            return d[n]
+    return float("nan")
+
+cols = ["watermark", "frac", "preempt", "out_tok/s", "req/s",
+        "TTFT_p50", "TTFT_p99", "ITL_p99", "E2EL_p50"]
+print("  ".join(f"{c:>10}" for c in cols))
+rows = []
+for label in order:
+    watermark, n = pre[label]
+    d = json.load(open(os.path.join(outdir, f"bench_{label}.json")))
+    rows.append(dict(
+        label=label, watermark=watermark, preempt=n,
+        out_tok_s=g(d, "output_throughput"),
+        req_s=g(d, "request_throughput"),
+        ttft_p50=g(d, "p50_ttft_ms", "median_ttft_ms"),
+        ttft_p99=g(d, "p99_ttft_ms"),
+        itl_p99=g(d, "p99_itl_ms"),
+        e2el_p50=g(d, "p50_e2el_ms", "median_e2el_ms"),
+    ))
+    print("  ".join(f"{str(v):>10}" for v in [
+        label, watermark, n,
+        f"{rows[-1]['out_tok_s']:.0f}",
+        f"{rows[-1]['req_s']:.3f}",
+        f"{rows[-1]['ttft_p50']/1000:.2f}",
+        f"{rows[-1]['ttft_p99']/1000:.2f}",
+        f"{rows[-1]['itl_p99']:.2f}",
+        f"{rows[-1]['e2el_p50']/1000:.1f}",
+    ]))
+print("\n(TTFT/E2EL in seconds; ITL in ms. Lower preempt is better.)")
+
+# ---- Plot -------------------------------------------------------------------
+try:
+    import matplotlib
+    matplotlib.use("Agg")
+    import matplotlib.pyplot as plt
+except Exception as e:  # noqa: BLE001
+    print(f"\n(skip plot: matplotlib unavailable: {e})")
+    sys.exit(0)
+
+x = [r["watermark"] for r in rows]
+xt = [f"{r['watermark']:g}\n({r['label']})" for r in rows]
+idx = list(range(len(rows)))
+
+fig, axes = plt.subplots(2, 2, figsize=(12, 8))
+fig.suptitle(
+    f"KV cache watermark sweep — {os.path.basename(os.path.abspath(outdir))}",
+    fontsize=12,
+)
+
+ax = axes[0][0]
+ax.bar(idx, [r["preempt"] for r in rows], color="tab:red")
+ax.set_title("Preemptions (lower is better)")
+ax.set_ylabel("preemptions")
+ax.set_xticks(idx); ax.set_xticklabels(xt)
+
+ax = axes[0][1]
+ax.plot(idx, [r["out_tok_s"] for r in rows], "o-", color="tab:green")
+ax.set_title("Output throughput (higher is better)")
+ax.set_ylabel("tokens/s")
+ax.set_xticks(idx); ax.set_xticklabels(xt)
+
+ax = axes[1][0]
+ax.plot(idx, [r["itl_p99"] for r in rows], "o-", color="tab:blue")
+ax.set_title("Inter-token latency p99 (lower is better)")
+ax.set_ylabel("ITL p99 (ms)")
+ax.set_xlabel("watermark fraction")
+ax.set_xticks(idx); ax.set_xticklabels(xt)
+
+ax = axes[1][1]
+ax.plot(idx, [r["ttft_p50"] / 1000 for r in rows], "o-", label="TTFT p50")
+ax.plot(idx, [r["ttft_p99"] / 1000 for r in rows], "o-", label="TTFT p99")
+ax.plot(idx, [r["e2el_p50"] / 1000 for r in rows], "o-", label="E2EL p50")
+ax.set_title("Latency (lower is better)")
+ax.set_ylabel("seconds")
+ax.set_xlabel("watermark fraction")
+ax.set_xticks(idx); ax.set_xticklabels(xt)
+ax.legend()
+
+fig.tight_layout(rect=(0, 0, 1, 0.95))
+out_png = os.path.join(outdir, "watermark_results.png")
+fig.savefig(out_png, dpi=120)
+print(f"\nWrote plot: {out_png}")
+PY

benchmarks/multi_turn/benchmark_serving_multi_turn.py

@@ -65,6 +65,32 @@ class RequestArgs(NamedTuple):
     limit_min_tokens: int  # Use negative value for no limit
     limit_max_tokens: int  # Use negative value for no limit
     timeout_sec: int
+    send_conversation_id: bool
+    headers: dict[str, str]
+
+
+def parse_custom_header(header: str) -> tuple[str, str]:
+    separators = (":", "=")
+    for separator in separators:
+        if separator in header:
+            key, value = header.split(separator, 1)
+            key = key.strip()
+            value = value.strip()
+            if key:
+                return key, value
+            break
+    raise argparse.ArgumentTypeError(
+        "Headers must be provided as 'Header-Name: value' or 'Header-Name=value'"
+    )
+
+
+def build_request_headers(
+    api_key: str | None, custom_headers: list[tuple[str, str]] | None
+) -> dict[str, str]:
+    headers = dict(custom_headers or [])
+    if api_key:
+        headers["Authorization"] = f"Bearer {api_key}"
+    return headers
 
 
 class BenchmarkArgs(NamedTuple):
@@ -218,12 +244,11 @@ async def send_request(
     max_tokens: int | None = None,
     timeout_sec: int = 120,
     conversation_id: str | None = None,
+    headers: dict[str, str] | None = None,
 ) -> ServerResponse:
     payload = {
         "model": model,
         "messages": messages,
-        "seed": 0,
-        "temperature": 0.0,
     }
 
     if conversation_id is not None:
@@ -233,13 +258,17 @@ async def send_request(
         payload["stream"] = True
         payload["stream_options"] = {"include_usage": False}
 
-    if min_tokens is not None:
-        payload["min_tokens"] = min_tokens
+    # if min_tokens is not None:
+    #     payload["min_tokens"] = min_tokens
 
     if max_tokens is not None:
         payload["max_tokens"] = max_tokens
 
-    headers = {"Content-Type": "application/json"}
+    request_headers = {"Content-Type": "application/json"}
+    if conversation_id is not None:
+        request_headers["X-Session-ID"] = str(conversation_id)
+    if headers is not None:
+        request_headers.update(headers)
 
     # Calculate the timeout for the request
     if max_tokens is not None:
@@ -265,7 +294,7 @@ async def send_request(
     most_recent_timestamp: int = start_time
 
     async with session.post(
-        url=chat_url, json=payload, headers=headers, timeout=timeout
+        url=chat_url, json=payload, headers=request_headers, timeout=timeout
     ) as response:
         http_status = HTTPStatus(response.status)
         if http_status == HTTPStatus.OK:
@@ -317,6 +346,8 @@ async def send_request(
             latency = time.perf_counter_ns() - start_time
 
     if ttft is None:
+        if stream:
+            valid_response = False
         # The response was a single chunk
         ttft = latency
 
@@ -423,7 +454,8 @@ async def send_turn(
         min_tokens,
         max_tokens,
         req_args.timeout_sec,
-        conversation_id=conv_id,
+        conversation_id=conv_id if req_args.send_conversation_id else None,
+        headers=req_args.headers,
     )
 
     if response.valid is False:
@@ -872,6 +904,7 @@ def get_client_config(
     # Arguments for API requests
     chat_url = f"{args.url}/v1/chat/completions"
     model_name = args.served_model_name if args.served_model_name else args.model
+    headers = build_request_headers(args.api_key, args.header)
 
     req_args = RequestArgs(
         chat_url=chat_url,
@@ -880,6 +913,8 @@ def get_client_config(
         limit_min_tokens=args.limit_min_tokens,
         limit_max_tokens=args.limit_max_tokens,
         timeout_sec=args.request_timeout_sec,
+        send_conversation_id=args.send_conversation_id,
+        headers=headers,
     )
 
     return client_args, req_args
@@ -1245,19 +1280,19 @@ def process_statistics(
         )
 
 
-async def get_server_info(url: str) -> None:
+async def get_server_info(url: str, headers: dict[str, str] | None = None) -> None:
     logger.info(f"{Color.BLUE}Collecting information from server: {url}{Color.RESET}")
     async with aiohttp.ClientSession() as session:
         # Get server version (not mandatory, "version" endpoint may not exist)
         url_version = f"{url}/version"
-        async with session.get(url_version) as response:
+        async with session.get(url_version, headers=headers) as response:
             if HTTPStatus(response.status) == HTTPStatus.OK:
                 text = await response.text()
                 logger.info(f"{Color.BLUE}Server version: {text}{Color.RESET}")
 
         # Get available models
         url_models = f"{url}/v1/models"
-        async with session.get(url_models) as response:
+        async with session.get(url_models, headers=headers) as response:
             if HTTPStatus(response.status) == HTTPStatus.OK:
                 text = await response.text()
                 logger.info(f"{Color.BLUE}Models:{Color.RESET}")
@@ -1323,6 +1358,22 @@ async def main() -> None:
         help="Base URL for the LLM API server",
     )
 
+    parser.add_argument(
+        "--api-key",
+        type=str,
+        default=None,
+        help="API key to send as an Authorization bearer token",
+    )
+    parser.add_argument(
+        "--header",
+        action="append",
+        type=parse_custom_header,
+        default=None,
+        metavar="KEY=VALUE",
+        help="Custom request header. Can be specified multiple times. "
+        "Accepts 'Header-Name: value' or 'Header-Name=value'.",
+    )
+
     parser.add_argument(
         "-p",
         "--num-clients",
@@ -1437,6 +1488,22 @@ async def main() -> None:
         help="Disable stream/streaming mode (set 'stream' to False in the API request)",
     )
 
+    parser.add_argument(
+        "--send-conversation-id",
+        default=False,
+        action="store_true",
+        help=(
+            "Inject a `conversation_id` field into each Chat Completions "
+            "payload. This is a non-standard OpenAI extension consumed by "
+            "vLLM's disaggregated multi-turn proxy "
+            "(examples/disaggregated/disaggregated_serving/"
+            "disagg_proxy_multiturn.py) to key cross-turn KV cache reuse. "
+            "Leave disabled (default) when targeting strict "
+            "OpenAI-compatible endpoints; enable when benchmarking the "
+            "disaggregated proxy."
+        ),
+    )
+
     parser.add_argument(
         "-e",
         "--excel-output",
@@ -1525,7 +1592,8 @@ async def main() -> None:
         args.model, trust_remote_code=args.trust_remote_code
     )
 
-    await get_server_info(args.url)
+    headers = build_request_headers(args.api_key, args.header)
+    await get_server_info(args.url, headers=headers)
 
     # Load the input file (either conversations of configuration file)
     logger.info(f"Reading input file: {args.input_file}")

build_rust.sh

@@ -1,5 +1,5 @@
 #!/bin/bash
-# Build the vllm-rs Rust frontend binary and install it into the vllm package.
+# Build vLLM Rust artifacts and install them into the vllm package.
 # Usage: ./build_rust.sh [--debug]
 #
 # By default builds in release mode. Pass --debug for faster compile times
@@ -8,8 +8,6 @@
 set -euo pipefail
 
 REPO_ROOT="$(cd "$(dirname "$0")" && pwd)"
-RUST_DIR="$REPO_ROOT/rust"
-TARGET_PATH="${VLLM_RS_TARGET_PATH:-$REPO_ROOT/vllm/vllm-rs}"
 
 # Read the required toolchain from rust-toolchain.toml.
 TOOLCHAIN=$(grep '^channel' "$REPO_ROOT/rust-toolchain.toml" | sed 's/.*= *"\(.*\)"/\1/')
@@ -27,18 +25,9 @@ if ! rustup run "$TOOLCHAIN" rustc --version &>/dev/null; then
 fi
 
 if [[ "${1:-}" == "--debug" ]]; then
-    PROFILE_ARGS=()
-    PROFILE_DIR="debug"
+    PROFILE_ARG="--debug"
 else
-    PROFILE_ARGS=(--release)
-    PROFILE_DIR="release"
+    PROFILE_ARG="--release"
 fi
 
-cargo +"$TOOLCHAIN" build "${PROFILE_ARGS[@]}" \
-    --manifest-path "$RUST_DIR/Cargo.toml" \
-    --bin vllm-rs \
-    --features native-tls-vendored
-
-mkdir -p "$(dirname "$TARGET_PATH")"
-cp "$RUST_DIR/target/$PROFILE_DIR/vllm-rs" "$TARGET_PATH"
-echo "Installed vllm-rs to $TARGET_PATH"
+python3 "$REPO_ROOT/tools/build_rust.py" "$PROFILE_ARG"

cmake/cpu_extension.cmake

@@ -166,6 +166,10 @@ elseif (S390_FOUND)
         "-mtune=native")
 elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
     message(STATUS "RISC-V detected")
+    if(DEFINED VLLM_RVV_VLEN AND NOT VLLM_RVV_VLEN GREATER 0)
+        message(FATAL_ERROR
+            "VLLM_RVV_VLEN must be a positive integer; got '${VLLM_RVV_VLEN}'")
+    endif()
     # VLLM_RVV_VLEN selects the target VLEN. Auto-detected from /proc/cpuinfo
     # by default; override with -DVLLM_RVV_VLEN=128 or -DVLLM_RVV_VLEN=256.
     if(NOT DEFINED VLLM_RVV_VLEN)
@@ -189,8 +193,7 @@ elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
                 "RISC-V RVV is available but VLEN could not be auto-detected. "
                 "Please specify VLEN explicitly:\n"
                 "  -DVLLM_RVV_VLEN=128   (for VLEN=128 hardware)\n"
-                "  -DVLLM_RVV_VLEN=256   (for VLEN=256 hardware, e.g. Spacemit X100)\n"
-                "  -DVLLM_RVV_VLEN=0     (force scalar, no RVV)")
+                "  -DVLLM_RVV_VLEN=256   (for VLEN=256 hardware, e.g. Spacemit X100)")
         endif()
     endif()
     if(VLLM_RVV_VLEN AND VLLM_RVV_VLEN GREATER 0)
@@ -219,7 +222,7 @@ endif()
 
 
 # Build oneDNN for GEMM kernels
-if (ENABLE_X86_ISA OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
+if (ENABLE_X86_ISA OR (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND) OR POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND OR RVV_FP16_FOUND OR RVV_BF16_FOUND)
     # Fetch and build Arm Compute Library (ACL) as oneDNN's backend for AArch64
     # TODO [fadara01]: remove this once ACL can be fetched and built automatically as a dependency of oneDNN
     set(ONEDNN_AARCH64_USE_ACL OFF CACHE BOOL "")
@@ -435,6 +438,12 @@ if(USE_ONEDNN)
         ${VLLM_EXT_SRC})
 endif()
 
+if (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
+    set(VLLM_EXT_SRC
+        "csrc/cpu/sgl-kernels/gemm_int4.cpp"
+        ${VLLM_EXT_SRC})
+endif()
+
 if (ENABLE_X86_ISA)
     set(VLLM_EXT_SRC_SGL
         "csrc/cpu/sgl-kernels/conv.cpp"

cmake/external_projects/fmha_sm100.cmake

@@ -0,0 +1,50 @@
+include(FetchContent)
+
+# If FMHA_SM100_SRC_DIR is set, fmha_sm100 is installed from that directory
+# instead of downloading. This is useful for local MSA development.
+if(DEFINED ENV{FMHA_SM100_SRC_DIR})
+  set(FMHA_SM100_SRC_DIR $ENV{FMHA_SM100_SRC_DIR})
+endif()
+
+if(FMHA_SM100_SRC_DIR)
+  FetchContent_Declare(
+    fmha_sm100
+    SOURCE_DIR ${FMHA_SM100_SRC_DIR}
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND ""
+  )
+else()
+  FetchContent_Declare(
+    fmha_sm100
+    GIT_REPOSITORY https://github.com/vllm-project/MSA.git
+    GIT_TAG 544eee5e09ae2dfa774d5b06739013f9b7402c57
+    GIT_PROGRESS TRUE
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND ""
+  )
+endif()
+
+FetchContent_GetProperties(fmha_sm100)
+if(NOT fmha_sm100_POPULATED)
+  FetchContent_Populate(fmha_sm100)
+endif()
+message(STATUS "fmha_sm100 is available at ${fmha_sm100_SOURCE_DIR}")
+
+add_custom_target(fmha_sm100)
+
+install(FILES
+  "${fmha_sm100_SOURCE_DIR}/python/fmha_sm100/__init__.py"
+  "${fmha_sm100_SOURCE_DIR}/python/fmha_sm100/sparse.py"
+  DESTINATION vllm/third_party/fmha_sm100
+  COMPONENT fmha_sm100)
+
+install(DIRECTORY "${fmha_sm100_SOURCE_DIR}/python/fmha_sm100/cute/"
+  DESTINATION vllm/third_party/fmha_sm100/cute
+  COMPONENT fmha_sm100
+  FILES_MATCHING
+    REGEX "/__pycache__(/.*)?$" EXCLUDE
+    REGEX ".*\\.pyc$" EXCLUDE
+    PATTERN "example.py" EXCLUDE
+    PATTERN "test_*.py" EXCLUDE
+    PATTERN "*.py"
+    PATTERN "build_k2q_csr.cu")

cmake/external_projects/qutlass.cmake

@@ -32,21 +32,33 @@ endif()
 message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}")
 
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-  cuda_archs_loose_intersection(QUTLASS_ARCHS "10.0f;12.0f" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0f" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f" "${CUDA_ARCHS}")
 else()
-  cuda_archs_loose_intersection(QUTLASS_ARCHS "12.0a;12.1a;10.0a;10.3a" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0a;12.1a" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0a;10.3a" "${CUDA_ARCHS}")
+endif()
+
+# QUTLASS uses TARGET_CUDA_ARCH as a single preprocessor selector for all its
+# sources. Do not compile a mixed SM100/SM120 arch list with one selector; prefer
+# SM100 when both families are requested because that is the primary deployed
+# target for this extension today.
+if(QUTLASS_SM100_ARCHS)
+  set(QUTLASS_ARCHS "${QUTLASS_SM100_ARCHS}")
+  set(QUTLASS_TARGET_CC 100)
+  if(QUTLASS_SM120_ARCHS)
+    message(WARNING
+      "[QUTLASS] Both SM100 and SM120 archs were requested; selecting SM100 "
+      "because TARGET_CUDA_ARCH is a single compile-time selector.")
+  endif()
+elseif(QUTLASS_SM120_ARCHS)
+  set(QUTLASS_ARCHS "${QUTLASS_SM120_ARCHS}")
+  set(QUTLASS_TARGET_CC 120)
+else()
+  set(QUTLASS_ARCHS)
 endif()
 
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
-
-  if(QUTLASS_ARCHS MATCHES "10\\.(0a|3a|0f)")
-    set(QUTLASS_TARGET_CC 100)
-  elseif(QUTLASS_ARCHS MATCHES "12\\.[01][af]?")
-    set(QUTLASS_TARGET_CC 120)
-  else()
-    message(FATAL_ERROR "[QUTLASS] internal error parsing CUDA_ARCHS='${QUTLASS_ARCHS}'.")
-  endif()
-
   set(QUTLASS_SOURCES
     ${qutlass_SOURCE_DIR}/qutlass/csrc/bindings.cpp
     ${qutlass_SOURCE_DIR}/qutlass/csrc/gemm.cu

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 dd62dac706b1cf7895bd99b18c6cb7e7e117ee25
+          GIT_TAG 803020a8fa15407871341d41eba4919ade2ee1ee
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn

cmake/utils.cmake

@@ -487,9 +487,9 @@ endfunction()
 
 function(cuda_archs_sm90plus OUT_CUDA_ARCHS TGT_CUDA_ARCHS)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(_archs "9.0a;10.0f;11.0f" "${TGT_CUDA_ARCHS}")
+    cuda_archs_loose_intersection(_archs "9.0a;10.0f;11.0f;12.0f" "${TGT_CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(_archs "9.0a;10.0a;10.1a;10.3a" "${TGT_CUDA_ARCHS}")
+    cuda_archs_loose_intersection(_archs "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${TGT_CUDA_ARCHS}")
   endif()
   set(${OUT_CUDA_ARCHS} ${_archs} PARENT_SCOPE)
 endfunction()

csrc/cpu/cpu_attn_impl.hpp

@@ -822,8 +822,8 @@ struct AttentionInput {
       logits_buffer_t *__restrict__ logits_buffer,                          \
       float *__restrict__ partial_q_buffer, float *__restrict__ max_buffer, \
       float *__restrict__ sum_buffer, int32_t *__restrict__ block_table,    \
-      const int32_t kv_tile_start_pos, const int32_t kv_tile_end_pos,       \
-      const int32_t kv_tile_token_num,                                      \
+      const int32_t kv_end_pos, const int32_t kv_tile_start_pos,            \
+      const int32_t kv_tile_end_pos, const int32_t kv_tile_token_num,       \
       const int64_t kv_cache_num_blocks_stride, const int32_t q_head_num,   \
       const int32_t q_token_num, const int32_t q_tile_start_pos,            \
       const int32_t q_heads_per_kv, const int32_t block_size,               \
@@ -834,7 +834,7 @@ struct AttentionInput {
 
 #define CPU_ATTENTION_PARAMS                                                  \
   q_heads_buffer, k_head_cache_ptr, v_head_cache_ptr, logits_buffer,          \
-      partial_q_buffer, max_buffer, sum_buffer, block_table,                  \
+      partial_q_buffer, max_buffer, sum_buffer, block_table, kv_end_pos,      \
       kv_tile_start_pos, kv_tile_end_pos, kv_tile_token_num,                  \
       kv_cache_num_blocks_stride, q_head_num, q_token_num, q_tile_start_pos,  \
       q_heads_per_kv, block_size, left_window_size, right_window_size, scale, \
@@ -917,6 +917,7 @@ class AttentionMainLoop {
     //  - max_buffer: [MaxQHeadNumPerIteration, 1], store max logits
     //  - sum_buffer: [MaxQHeadNumPerIteration, 1], store sum of exp
     //  - block_table
+    //  - kv_end_pos: un-aligned end position of KV cache
     //  - kv_tile_start_pos: start position of KV cache, aligned to
     //  BlockSizeAlignment
     //  - kv_tile_end_pos: end position of KV cache, aligned to
@@ -1043,7 +1044,7 @@ class AttentionMainLoop {
         }
 
         apply_mask(logits_buffer, kv_tile_token_num, q_tile_start_pos,
-                   kv_tile_start_pos, kv_tile_end_pos, q_token_num,
+                   kv_end_pos, kv_tile_start_pos, kv_tile_end_pos, q_token_num,
                    q_heads_per_kv, left_window_size, right_window_size);
 
         // if (debug_info){
@@ -1126,7 +1127,7 @@ class AttentionMainLoop {
 
     void apply_mask(logits_buffer_t* __restrict__ logits_buffer,
                     const int64_t logits_buffer_stride,
-                    const int32_t q_tile_start_pos,
+                    const int32_t q_tile_start_pos, const int32_t kv_end_pos,
                     const int32_t kv_tile_start_pos,
                     const int32_t kv_tile_end_pos, const int32_t q_token_num,
                     const int32_t q_heads_per_kv,
@@ -1154,7 +1155,7 @@ class AttentionMainLoop {
                            std::max(kv_tile_start_pos,
                                     curr_token_pos + sliding_window_right + 1));
           }
-          return pos;
+          return std::min(pos, kv_end_pos);
         }();
 
         int32_t left_invalid_token_num = left_kv_pos - kv_tile_start_pos;
@@ -1789,7 +1790,7 @@ class AttentionMainLoop {
                   attn_impl.template execute_attention<Attention>(
                       curr_q_heads_buffer, curr_k_cache, curr_v_cache,
                       logits_buffer, curr_partial_q_buffer, curr_max_buffer,
-                      curr_sum_buffer, curr_block_table,
+                      curr_sum_buffer, curr_block_table, kv_end_pos,
                       aligned_actual_kv_tile_pos_left,
                       aligned_actual_kv_tile_pos_right, actual_kv_token_num,
                       kv_cache_block_num_stride, q_tile_head_num,

csrc/cpu/cpu_types_riscv_defs.hpp

@@ -57,6 +57,10 @@ typedef RVVTYPE(vfloat32, LMUL_512, _t) fixed_fp32x16_t
 typedef RVVTYPE(vfloat32, LMUL_1024, _t) fixed_fp32x32_t
     __attribute__((riscv_rvv_vector_bits(1024)));
 
+// int8
+typedef RVVTYPE(vint8, LMUL_128, _t) fixed_i8x16_t
+    __attribute__((riscv_rvv_vector_bits(128)));
+
 // int32
 typedef RVVTYPE(vint32, LMUL_256, _t) fixed_i32x8_t
     __attribute__((riscv_rvv_vector_bits(256)));

csrc/cpu/cpu_types_riscv_impl.hpp

@@ -9,10 +9,14 @@
 
 #include <algorithm>
 #include <cmath>
+#include <cstdint>
 #include <cstring>
 #include <iostream>
 #include <limits>
 #include <torch/all.h>
+
+#include "float_convert.hpp"
+
 namespace vec_op {
 
 // FP8 KV cache is not supported on RISC-V. These tag types and the
@@ -245,8 +249,7 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
     float tmp[8];
     for (int i = 0; i < 8; ++i) {
-      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
-      std::memcpy(&tmp[i], &v, 4);
+      tmp[i] = bf16_to_float(u16[i]);
     }
     reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_256)(tmp, 8);
   }
@@ -256,9 +259,7 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     RVVI(__riscv_vse32_v_f32, LMUL_256)(tmp, reg_fp32, 8);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < 8; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
   void save(void* ptr, int elem_num) const {
@@ -266,9 +267,7 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     RVVI(__riscv_vse32_v_f32, LMUL_256)(tmp, reg_fp32, 8);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < elem_num; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
   void save_strided(void* ptr, ptrdiff_t stride) const {
@@ -277,10 +276,8 @@ struct BF16Vec8 : public Vec<BF16Vec8> {
     uint8_t* u8 = static_cast<uint8_t*>(ptr);
     ptrdiff_t byte_stride = stride * sizeof(uint16_t);
     for (int i = 0; i < 8; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      uint16_t val = static_cast<uint16_t>(v >> 16);
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) =
+          float_to_bf16(tmp[i]);
     }
   }
 };
@@ -292,8 +289,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
     float tmp[16];
     for (int i = 0; i < 16; ++i) {
-      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
-      std::memcpy(&tmp[i], &v, 4);
+      tmp[i] = bf16_to_float(u16[i]);
     }
     reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_512)(tmp, 16);
   }
@@ -306,9 +302,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     RVVI(__riscv_vse32_v_f32, LMUL_512)(tmp, reg_fp32, 16);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < 16; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
   void save(void* ptr, int elem_num) const {
@@ -316,9 +310,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     RVVI(__riscv_vse32_v_f32, LMUL_512)(tmp, reg_fp32, 16);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < elem_num; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
   void save_strided(void* ptr, ptrdiff_t stride) const {
@@ -327,10 +319,8 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
     uint8_t* u8 = static_cast<uint8_t*>(ptr);
     ptrdiff_t byte_stride = stride * sizeof(uint16_t);
     for (int i = 0; i < 16; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      uint16_t val = static_cast<uint16_t>(v >> 16);
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) =
+          float_to_bf16(tmp[i]);
     }
   }
 };
@@ -343,8 +333,7 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     const uint16_t* u16 = static_cast<const uint16_t*>(ptr);
     float tmp[32];
     for (int i = 0; i < 32; ++i) {
-      uint32_t v = static_cast<uint32_t>(u16[i]) << 16;
-      std::memcpy(&tmp[i], &v, 4);
+      tmp[i] = bf16_to_float(u16[i]);
     }
     reg_fp32 = RVVI(__riscv_vle32_v_f32, LMUL_1024)(tmp, 32);
   }
@@ -371,9 +360,7 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     RVVI(__riscv_vse32_v_f32, LMUL_1024)(tmp, reg_fp32, 32);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < 32; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
 
@@ -382,9 +369,7 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     RVVI(__riscv_vse32_v_f32, LMUL_1024)(tmp, reg_fp32, 32);
     uint16_t* u16 = static_cast<uint16_t*>(ptr);
     for (int i = 0; i < elem_num; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      u16[i] = static_cast<uint16_t>(v >> 16);
+      u16[i] = float_to_bf16(tmp[i]);
     }
   }
 
@@ -394,10 +379,8 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
     uint8_t* u8 = static_cast<uint8_t*>(ptr);
     ptrdiff_t byte_stride = stride * sizeof(uint16_t);
     for (int i = 0; i < 32; ++i) {
-      uint32_t v;
-      std::memcpy(&v, &tmp[i], 4);
-      uint16_t val = static_cast<uint16_t>(v >> 16);
-      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) = val;
+      *reinterpret_cast<uint16_t*>(u8 + i * byte_stride) =
+          float_to_bf16(tmp[i]);
     }
   }
 };
@@ -734,10 +717,18 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
     return FP32Vec16(
         RVVI(__riscv_vfmax_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
   }
+  FP32Vec16 max(const FP32Vec16& b, const int elem_num) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfmax_vv_f32, LMUL_512)(reg, b.reg, elem_num));
+  }
   FP32Vec16 min(const FP32Vec16& b) const {
     return FP32Vec16(
         RVVI(__riscv_vfmin_vv_f32, LMUL_512)(reg, b.reg, VEC_ELEM_NUM));
   }
+  FP32Vec16 min(const FP32Vec16& b, const int elem_num) const {
+    return FP32Vec16(
+        RVVI(__riscv_vfmin_vv_f32, LMUL_512)(reg, b.reg, elem_num));
+  }
   FP32Vec16 abs() const {
     return FP32Vec16(RVVI(__riscv_vfabs_v_f32, LMUL_512)(reg, VEC_ELEM_NUM));
   }
@@ -867,6 +858,27 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   }
 };
 
+struct INT8Vec16 : public Vec<INT8Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  fixed_i8x16_t reg;
+
+  explicit INT8Vec16(const FP32Vec16& vec) {
+    auto i32_vec =
+        RVVI(__riscv_vfcvt_x_f_v_i32, LMUL_512)(vec.reg, VEC_ELEM_NUM);
+    auto i16_vec = RVVI(__riscv_vnclip_wx_i16, LMUL_256)(
+        i32_vec, 0, __RISCV_VXRM_RNU, VEC_ELEM_NUM);
+    reg = RVVI(__riscv_vnclip_wx_i8, LMUL_128)(i16_vec, 0, __RISCV_VXRM_RNU,
+                                               VEC_ELEM_NUM);
+  }
+
+  void save(int8_t* ptr) const {
+    RVVI(__riscv_vse8_v_i8, LMUL_128)(ptr, reg, VEC_ELEM_NUM);
+  }
+  void save(int8_t* ptr, int elem_num) const {
+    RVVI(__riscv_vse8_v_i8, LMUL_128)(ptr, reg, elem_num);
+  }
+};
+
 // ============================================================================
 // Type Traits & Global Helpers
 // ============================================================================
@@ -956,9 +968,7 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16& v)
 #else
 template <>
 inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
-  uint32_t val;
-  std::memcpy(&val, &v, 4);
-  *reinterpret_cast<uint16_t*>(ptr) = static_cast<uint16_t>(val >> 16);
+  *reinterpret_cast<uint16_t*>(ptr) = float_to_bf16(v);
 }
 inline BF16Vec8::BF16Vec8(const FP32Vec8& v) : reg_fp32(v.reg) {}
 inline BF16Vec16::BF16Vec16(const FP32Vec16& v) : reg_fp32(v.reg) {}

csrc/cpu/cpu_types_vxe.hpp

@@ -3,7 +3,9 @@
 #define CPU_TYPES_VXE_HPP
 
 #include <vecintrin.h>
+#include <bit>
 #include <cmath>
+#include <cstdint>
 #include <limits>
 #include <torch/all.h>
 namespace vec_op {
@@ -817,8 +819,7 @@ inline void storeFP32<::c10::Half>(float v, ::c10::Half* ptr) {
   // intrinsics for FP32 to FP16 conversion does not use IEEE rounding and can
   // produce incorrect results for some inputs. Process each of the 4 vectors
   // separately.
-  uint32_t in;
-  std::memcpy(&in, &v, sizeof(in));
+  uint32_t in = std::bit_cast<uint32_t>(v);
 
   uint32_t s = (in & 0x80000000) >> 16;  // Sign
   uint32_t e = (in & 0x7F800000) >> 23;  // Exponent

csrc/cpu/float_convert.hpp

@@ -1,14 +1,15 @@
+#pragma once
 
-static float bf16_to_float(uint16_t bf16) {
+#include <bit>
+#include <cstdint>
+
+inline float bf16_to_float(uint16_t bf16) {
   uint32_t bits = static_cast<uint32_t>(bf16) << 16;
-  float fp32;
-  std::memcpy(&fp32, &bits, sizeof(fp32));
-  return fp32;
+  return std::bit_cast<float>(bits);
 }
 
-static uint16_t float_to_bf16(float fp32) {
-  uint32_t bits;
-  std::memcpy(&bits, &fp32, sizeof(fp32));
+inline uint16_t float_to_bf16(float fp32) {
+  uint32_t bits = std::bit_cast<uint32_t>(fp32);
   return static_cast<uint16_t>(bits >> 16);
 }
 
@@ -18,14 +19,13 @@ static uint16_t float_to_bf16(float fp32) {
  * Codes below copied from
  * https://github.com/PrincetonVision/marvin/tree/master/tools/tensorIO_matlab
  *************************************************/
-static uint16_t float_to_fp16(float fp32) {
+inline uint16_t float_to_fp16(float fp32) {
   uint16_t fp16;
 
-  unsigned x;
   unsigned u, remainder, shift, lsb, lsb_s1, lsb_m1;
   unsigned sign, exponent, mantissa;
 
-  std::memcpy(&x, &fp32, sizeof(fp32));
+  uint32_t x = std::bit_cast<uint32_t>(fp32);
   u = (x & 0x7fffffff);
 
   // Get rid of +NaN/-NaN case first.
@@ -77,12 +77,11 @@ static uint16_t float_to_fp16(float fp32) {
   return fp16;
 }
 
-static float fp16_to_float(uint16_t fp16) {
+inline float fp16_to_float(uint16_t fp16) {
   unsigned sign = ((fp16 >> 15) & 1);
   unsigned exponent = ((fp16 >> 10) & 0x1f);
   unsigned mantissa = ((fp16 & 0x3ff) << 13);
-  int temp;
-  float fp32;
+  uint32_t temp;
   if (exponent == 0x1f) { /* NaN or Inf */
     mantissa = (mantissa ? (sign = 0, 0x7fffff) : 0);
     exponent = 0xff;
@@ -101,6 +100,5 @@ static float fp16_to_float(uint16_t fp16) {
     exponent += 0x70;
   }
   temp = ((sign << 31) | (exponent << 23) | mantissa);
-  std::memcpy(&fp32, &temp, sizeof(temp));
-  return fp32;
+  return std::bit_cast<float>(temp);
 }

csrc/cpu/generate_cpu_attn_dispatch.py

@@ -11,7 +11,7 @@ import os
 HEAD_DIMS_32 = [32, 64, 96, 128, 160, 192, 224, 256, 512]
 
 # Head dimensions divisible by 16 but not 32 (VEC16 only)
-HEAD_DIMS_16 = [80, 112]
+HEAD_DIMS_16 = [48, 80, 112]
 
 # ISA types
 ISA_TYPES = {

csrc/cpu/sgl-kernels/gemm_int4.cpp

@@ -268,6 +268,23 @@ void _dequant_gemm_accum_small_M(
   _dequant_gemm_accum_small_M<M, N, ldb, sym_quant_act>(C, A, scales_a, qzeros_a, B, scales_b, qzeros_b, K, lda, ldc);
 #endif
 
+template <int64_t N, int64_t ldb>
+inline int32_t load_uint4_vnni(const uint8_t* __restrict__ B, int64_t k, int64_t n) {
+  // B is packed as [_block_k / 4, N / 2, 4] for VNNI4. Each byte stores two
+  // columns from adjacent 8-column groups for one K lane.
+  constexpr int64_t n_group_size = 8;
+  constexpr int64_t vnni_size = 4;
+  static_assert(N % (2 * n_group_size) == 0);
+
+  int64_t n_group = n / n_group_size;
+  int64_t ni = n % n_group_size;
+  int64_t ki = k % vnni_size;
+  int64_t k_base = k - ki;
+  int64_t packed_n = (n_group / 2) * n_group_size + ni;
+  uint8_t packed = B[k_base * ldb + packed_n * vnni_size + ki];
+  return (n_group % 2 == 0) ? (packed & 0x0f) : ((packed >> 4) & 0x0f);
+}
+
 template <int64_t N, int64_t ldb, bool sym_quant_act>
 void _dequant_gemm_accum(
     float* C,
@@ -321,7 +338,24 @@ void _dequant_gemm_accum(
   } else
 #endif
   {
-    TORCH_CHECK(false, "tinygemm_kernel: scalar path not implemented!");
+    for (int64_t m = 0; m < M; ++m) {
+      for (int64_t n = 0; n < N; ++n) {
+        int32_t acc = 0;
+        for (int64_t k = 0; k < K; ++k) {
+          int32_t b = load_uint4_vnni<N, ldb>(B, k, n) - qzeros_b[n];
+          if constexpr (sym_quant_act) {
+            const int8_t* A_s8 = reinterpret_cast<const int8_t*>(A);
+            acc += static_cast<int32_t>(A_s8[m * lda + k]) * b;
+          } else {
+            acc += static_cast<int32_t>(A[m * lda + k]) * b;
+          }
+        }
+        if constexpr (!sym_quant_act) {
+          acc -= qzeros_a[m] * compensation[n];
+        }
+        C[m * ldc + n] += static_cast<float>(acc) * scales_a[m] * scales_b[n];
+      }
+    }
   }
 }
 
@@ -496,9 +530,11 @@ void _da8w4_linear_impl(
         store_out<out_dtype, BLOCK_N>(C_tmp, output + mci * block_m * N + nc * BLOCK_N, m_size, N /*lda*/);
       }
     }
+#if defined(CPU_CAPABILITY_AVX512)
     if (use_brgemm) {
       at::native::cpublas::brgemm_release();
     }
+#endif
   });
 }
 

csrc/cpu/sgl-kernels/vec.h

@@ -245,7 +245,7 @@ quantize_row_int8(uint8_t* __restrict__ Aq, float& As, const scalar_t* __restric
 
   for (int64_t k = 0; k < K; ++k) {
     const float val = static_cast<float>(A[k]) * inv_scale;
-    Aq[k] = (uint8_t)(std::round(val)) + 128;
+    Aq[k] = static_cast<uint8_t>(static_cast<int32_t>(std::round(val)) + 128);
   }
   As = scale;
 }

csrc/cpu/torch_bindings.cpp

@@ -329,8 +329,9 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("rotary_embedding", torch::kCPU, &rotary_embedding);
 
   // Quantization
-#if defined(__AVX512F__) || defined(__AVX2__) || \
-    (defined(__aarch64__) && !defined(__APPLE__)) || defined(__powerpc64__)
+#if defined(__AVX512F__) || defined(__AVX2__) ||                               \
+    (defined(__aarch64__) && !defined(__APPLE__)) || defined(__powerpc64__) || \
+    defined(__riscv_v)
   // Helper function to release oneDNN handlers
   ops.def("release_dnnl_matmul_handler(int handler) -> ()",
           &release_dnnl_matmul_handler);
@@ -428,19 +429,6 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("int8_scaled_mm_with_quant", torch::kCPU,
            &int8_scaled_mm_with_quant);
 
-  // Adapted from sglang: INT4 W4A8 kernels
-  ops.def(
-      "convert_weight_packed_scale_zp(Tensor weight, Tensor qzeros, Tensor "
-      "scales, int quant_method_4bit) -> (Tensor, "
-      "Tensor, Tensor)");
-  ops.impl("convert_weight_packed_scale_zp", torch::kCPU,
-           &convert_weight_packed_scale_zp);
-
-  ops.def(
-      "int4_scaled_mm_cpu(Tensor(a0!) x, Tensor(a1!) w, Tensor(a2!) w_zeros, "
-      "Tensor(a3!) w_scales, Tensor? bias) -> Tensor");
-  ops.impl("int4_scaled_mm_cpu", torch::kCPU, &int4_scaled_mm_cpu);
-
   // Adapted from sglang: FP8 W8A16 kernel
   ops.def(
       "fp8_scaled_mm_cpu(Tensor(a0!) mat1, Tensor(a1!) mat2, Tensor(a2!) "
@@ -467,6 +455,23 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.impl("causal_conv1d_update_cpu", torch::kCPU, &causal_conv1d_update_cpu);
 #endif
 
+#if (defined(__AVX512BF16__) && defined(__AVX512F__) && \
+     defined(__AVX512VNNI__)) ||                        \
+    defined(__riscv)
+  // Adapted from sglang: INT4 W4A8 kernels
+  ops.def(
+      "convert_weight_packed_scale_zp(Tensor weight, Tensor qzeros, Tensor "
+      "scales, int quant_method_4bit) -> (Tensor, "
+      "Tensor, Tensor)");
+  ops.impl("convert_weight_packed_scale_zp", torch::kCPU,
+           &convert_weight_packed_scale_zp);
+
+  ops.def(
+      "int4_scaled_mm_cpu(Tensor(a0!) x, Tensor(a1!) w, Tensor(a2!) w_zeros, "
+      "Tensor(a3!) w_scales, Tensor? bias) -> Tensor");
+  ops.impl("int4_scaled_mm_cpu", torch::kCPU, &int4_scaled_mm_cpu);
+#endif
+
   // Adapted from sglang: GDN kernels
   ops.def(
       "chunk_gated_delta_rule_cpu(Tensor query, Tensor key, Tensor value, "

csrc/cuda_view.cu

@@ -1,3 +1,4 @@
+// TODO: Remove this once ROCm upgrade to torch 2.11.
 #include <torch/all.h>
 #include <torch/cuda.h>
 #include <cuda_runtime.h>

csrc/cumem_allocator.cpp

@@ -9,6 +9,7 @@
 static const char* PYARGS_PARSE = "KKKK";
 #else
   #include <cstdlib>
+  #include <cstdint>
   #include <cerrno>
   #include <climits>
 
@@ -46,6 +47,29 @@ static inline unsigned long long my_min(unsigned long long a,
   return a < b ? a : b;
 }
 
+static CUresult reserve_rocm_address(CUdeviceptr* d_mem, size_t size,
+                                     size_t alignment) {
+  CUresult status = cuMemAddressReserve(d_mem, size, alignment, 0, 0);
+  if (status == CUresult(0) || alignment == 0) {
+    return status;
+  }
+
+  // Some ROCm stacks can report OOM while reserving VA with an explicit
+  // alignment even when physical VRAM is free. Let HIP choose the default
+  // alignment, then verify that the returned address still satisfies the
+  // requested alignment before accepting it.
+  status = cuMemAddressReserve(d_mem, size, 0, 0, 0);
+  if (status != CUresult(0)) {
+    return status;
+  }
+  if (((std::uintptr_t)(*d_mem) % alignment) == 0) {
+    return status;
+  }
+
+  (void)cuMemAddressFree(*d_mem, size);
+  return hipErrorNotSupported;
+}
+
 static const char* PYARGS_PARSE = "KKKO";
 #endif
 
@@ -325,7 +349,7 @@ void* my_malloc(ssize_t size, int device, CUstream stream) {
     return nullptr;
   }
 #else
-  CUDA_CHECK(cuMemAddressReserve(&d_mem, alignedSize, granularity, 0, 0));
+  CUDA_CHECK(reserve_rocm_address(&d_mem, alignedSize, granularity));
   if (error_code != 0) {
     return nullptr;
   }

csrc/cutlass_extensions/vllm_cutlass_library_extension.py

@@ -57,13 +57,13 @@ VLLMDataTypeVLLMScalarTypeTag: dict[VLLMDataType | DataType, str] = {
 }
 
 VLLMDataTypeTorchDataTypeTag: dict[VLLMDataType | DataType, str] = {
-    DataType.u8: "at::ScalarType::Byte",
-    DataType.s8: "at::ScalarType::Char",
-    DataType.e4m3: "at::ScalarType::Float8_e4m3fn",
-    DataType.s32: "at::ScalarType::Int",
-    DataType.f16: "at::ScalarType::Half",
-    DataType.bf16: "at::ScalarType::BFloat16",
-    DataType.f32: "at::ScalarType::Float",
+    DataType.u8: "torch::headeronly::ScalarType::Byte",
+    DataType.s8: "torch::headeronly::ScalarType::Char",
+    DataType.e4m3: "torch::headeronly::ScalarType::Float8_e4m3fn",
+    DataType.s32: "torch::headeronly::ScalarType::Int",
+    DataType.f16: "torch::headeronly::ScalarType::Half",
+    DataType.bf16: "torch::headeronly::ScalarType::BFloat16",
+    DataType.f32: "torch::headeronly::ScalarType::Float",
 }
 
 VLLMKernelScheduleTag: dict[MixedInputKernelScheduleType | KernelScheduleType, str] = {

csrc/libtorch_stable/activation_kernels.cu

@@ -10,11 +10,20 @@
 
 namespace vllm {
 
-template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&),
+// `alpha` and `beta` are applied to opposite operands:
+//   - alpha lives INSIDE the activation (the activated half): the gated
+//     activation computes act_half * sigmoid(alpha * act_half).
+//   - beta is added to the OTHER (non-activated) half before the multiply.
+// So the result is always ACT(act_half, alpha) * (other_half + beta).
+// Which half is which depends on `act_first` (see below). Defaults
+// alpha=1.0, beta=0.0 reproduce the plain SwiGLU/GeGLU behavior.
+template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&, const float),
           bool act_first, bool HAS_CLAMP>
 __device__ __forceinline__ scalar_t compute(const scalar_t& x,
                                             const scalar_t& y,
-                                            const float limit) {
+                                            const float limit,
+                                            const float alpha,
+                                            const float beta) {
   if constexpr (act_first) {
     scalar_t gate = x;
     scalar_t up = y;
@@ -22,7 +31,9 @@ __device__ __forceinline__ scalar_t compute(const scalar_t& x,
       gate = (scalar_t)fminf((float)gate, limit);
       up = (scalar_t)fmaxf(fminf((float)up, limit), -limit);
     }
-    return ACT_FN(gate) * up;
+    // act_first: gate is the activated half -> alpha applies to gate;
+    // beta is added to up (the non-activated half).
+    return (scalar_t)(ACT_FN(gate, alpha) * ((float)up + beta));
   } else {
     scalar_t gate = x;
     scalar_t up = y;
@@ -30,55 +41,68 @@ __device__ __forceinline__ scalar_t compute(const scalar_t& x,
       gate = (scalar_t)fmaxf(fminf((float)gate, limit), -limit);
       up = (scalar_t)fminf((float)up, limit);
     }
-    return gate * ACT_FN(up);
+    // !act_first: up is the activated half -> alpha applies to up;
+    // beta is added to gate (the non-activated half).
+    return (scalar_t)(((float)gate + beta) * ACT_FN(up, alpha));
   }
 }
 
-template <typename packed_t, packed_t (*PACKED_ACT_FN)(const packed_t&),
+template <typename packed_t,
+          packed_t (*PACKED_ACT_FN)(const packed_t&, const float),
           bool act_first, bool HAS_CLAMP>
 __device__ __forceinline__ packed_t packed_compute(const packed_t& x,
                                                    const packed_t& y,
-                                                   const float limit) {
+                                                   const float limit,
+                                                   const float alpha,
+                                                   const float beta) {
   if constexpr (act_first) {
     packed_t gate = x;
     packed_t up = y;
+    float2 u = cast_to_float2(up);
     if constexpr (HAS_CLAMP) {
       float2 g = cast_to_float2(gate);
-      float2 u = cast_to_float2(up);
       g.x = fminf(g.x, limit);
       g.y = fminf(g.y, limit);
       u.x = fmaxf(fminf(u.x, limit), -limit);
       u.y = fmaxf(fminf(u.y, limit), -limit);
       gate = cast_to_packed<packed_t>(g);
-      up = cast_to_packed<packed_t>(u);
     }
-    return packed_mul(PACKED_ACT_FN(gate), up);
+    // act_first: gate is the activated half -> alpha applies to gate;
+    // beta is added to up (the non-activated half).
+    float2 activated = cast_to_float2(PACKED_ACT_FN(gate, alpha));
+    activated.x *= u.x + beta;
+    activated.y *= u.y + beta;
+    return cast_to_packed<packed_t>(activated);
   } else {
     packed_t gate = x;
     packed_t up = y;
+    float2 g = cast_to_float2(gate);
     if constexpr (HAS_CLAMP) {
-      float2 g = cast_to_float2(gate);
       float2 u = cast_to_float2(up);
       g.x = fmaxf(fminf(g.x, limit), -limit);
       g.y = fmaxf(fminf(g.y, limit), -limit);
       u.x = fminf(u.x, limit);
       u.y = fminf(u.y, limit);
-      gate = cast_to_packed<packed_t>(g);
       up = cast_to_packed<packed_t>(u);
     }
-    return packed_mul(gate, PACKED_ACT_FN(up));
+    // !act_first: up is the activated half -> alpha applies to up;
+    // beta is added to gate (the non-activated half).
+    float2 activated = cast_to_float2(PACKED_ACT_FN(up, alpha));
+    activated.x *= g.x + beta;
+    activated.y *= g.y + beta;
+    return cast_to_packed<packed_t>(activated);
   }
 }
 
 // Activation and gating kernel template.
 template <typename scalar_t, typename packed_t,
-          scalar_t (*ACT_FN)(const scalar_t&),
-          packed_t (*PACKED_ACT_FN)(const packed_t&), bool act_first,
-          bool use_vec, bool HAS_CLAMP, bool use_256b = false>
+          scalar_t (*ACT_FN)(const scalar_t&, const float),
+          packed_t (*PACKED_ACT_FN)(const packed_t&, const float),
+          bool act_first, bool use_vec, bool HAS_CLAMP, bool use_256b = false>
 __global__ void act_and_mul_kernel(
     scalar_t* __restrict__ out,          // [..., d]
     const scalar_t* __restrict__ input,  // [..., 2, d]
-    const int d, const float limit) {
+    const int d, const float limit, const float alpha, const float beta) {
   const scalar_t* x_ptr = input + blockIdx.x * 2 * d;
   const scalar_t* y_ptr = x_ptr + d;
   scalar_t* out_ptr = out + blockIdx.x * d;
@@ -105,7 +129,7 @@ __global__ void act_and_mul_kernel(
       for (int j = 0; j < pvec_t::NUM_ELTS; j++) {
         x.elts[j] =
             packed_compute<packed_t, PACKED_ACT_FN, act_first, HAS_CLAMP>(
-                x.elts[j], y.elts[j], limit);
+                x.elts[j], y.elts[j], limit, alpha, beta);
       }
       if constexpr (use_256b) {
         st256(x, &out_vec[i]);
@@ -118,29 +142,34 @@ __global__ void act_and_mul_kernel(
     for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
       const scalar_t x = VLLM_LDG(&x_ptr[idx]);
       const scalar_t y = VLLM_LDG(&y_ptr[idx]);
-      out_ptr[idx] =
-          compute<scalar_t, ACT_FN, act_first, HAS_CLAMP>(x, y, limit);
+      out_ptr[idx] = compute<scalar_t, ACT_FN, act_first, HAS_CLAMP>(
+          x, y, limit, alpha, beta);
     }
   }
 }
 
+// Gated activations take an `alpha` argument that scales the sigmoid input
+// (`x * sigmoid(alpha * x)`). alpha defaults to 1.0 at all call sites, which
+// is exactly SiLU; only the clamp path (silu_and_mul_with_clamp) passes a
+// non-default alpha. Activations that do not use alpha simply ignore it.
 template <typename T>
-__device__ __forceinline__ T silu_kernel(const T& x) {
-  // x * sigmoid(x)
-  return (T)(((float)x) / (1.0f + expf((float)-x)));
+__device__ __forceinline__ T silu_kernel(const T& x, const float alpha) {
+  // x * sigmoid(alpha * x)
+  return (T)(((float)x) / (1.0f + expf((float)-x * alpha)));
 }
 
 template <typename packed_t>
-__device__ __forceinline__ packed_t packed_silu_kernel(const packed_t& val) {
-  // x * sigmoid(x)
+__device__ __forceinline__ packed_t packed_silu_kernel(const packed_t& val,
+                                                       const float alpha) {
+  // x * sigmoid(alpha * x)
   float2 fval = cast_to_float2(val);
-  fval.x = fval.x / (1.0f + expf(-fval.x));
-  fval.y = fval.y / (1.0f + expf(-fval.y));
+  fval.x = fval.x / (1.0f + expf(-fval.x * alpha));
+  fval.y = fval.y / (1.0f + expf(-fval.y * alpha));
   return cast_to_packed<packed_t>(fval);
 }
 
 template <typename T>
-__device__ __forceinline__ T gelu_kernel(const T& x) {
+__device__ __forceinline__ T gelu_kernel(const T& x, const float /*alpha*/) {
   // Equivalent to PyTorch GELU with 'none' approximation.
   // Refer to:
   // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L36-L38
@@ -150,7 +179,8 @@ __device__ __forceinline__ T gelu_kernel(const T& x) {
 }
 
 template <typename packed_t>
-__device__ __forceinline__ packed_t packed_gelu_kernel(const packed_t& val) {
+__device__ __forceinline__ packed_t packed_gelu_kernel(const packed_t& val,
+                                                       const float /*alpha*/) {
   // Equivalent to PyTorch GELU with 'none' approximation.
   // Refer to:
   // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L36-L38
@@ -162,7 +192,8 @@ __device__ __forceinline__ packed_t packed_gelu_kernel(const packed_t& val) {
 }
 
 template <typename T>
-__device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
+__device__ __forceinline__ T gelu_tanh_kernel(const T& x,
+                                              const float /*alpha*/) {
   // Equivalent to PyTorch GELU with 'tanh' approximation.
   // Refer to:
   // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L25-L30
@@ -176,7 +207,7 @@ __device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
 
 template <typename packed_t>
 __device__ __forceinline__ packed_t
-packed_gelu_tanh_kernel(const packed_t& val) {
+packed_gelu_tanh_kernel(const packed_t& val, const float /*alpha*/) {
   // Equivalent to PyTorch GELU with 'tanh' approximation.
   // Refer to:
   // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L25-L30
@@ -202,7 +233,7 @@ packed_gelu_tanh_kernel(const packed_t& val) {
 // clamped (max only) and up input is clamped (both sides) before the
 // activation function is applied.
 #define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL, PACKED_KERNEL, ACT_FIRST,        \
-                                      HAS_CLAMP, LIMIT)                        \
+                                      HAS_CLAMP, LIMIT, ALPHA, BETA)           \
   auto dtype = input.scalar_type();                                            \
   int d = input.size(-1) / 2;                                                  \
   int64_t num_tokens = input.numel() / input.size(-1);                         \
@@ -230,7 +261,7 @@ packed_gelu_tanh_kernel(const packed_t& val) {
             PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
             ACT_FIRST, true, HAS_CLAMP, true><<<grid, block, 0, stream>>>(     \
             out.mutable_data_ptr<scalar_t>(),                                  \
-            input.const_data_ptr<scalar_t>(), d, LIMIT);                       \
+            input.const_data_ptr<scalar_t>(), d, LIMIT, ALPHA, BETA);          \
       });                                                                      \
     } else {                                                                   \
       VLLM_STABLE_DISPATCH_FLOATING_TYPES(dtype, "act_and_mul_kernel", [&] {   \
@@ -240,7 +271,7 @@ packed_gelu_tanh_kernel(const packed_t& val) {
             PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>, \
             ACT_FIRST, true, HAS_CLAMP, false><<<grid, block, 0, stream>>>(    \
             out.mutable_data_ptr<scalar_t>(),                                  \
-            input.const_data_ptr<scalar_t>(), d, LIMIT);                       \
+            input.const_data_ptr<scalar_t>(), d, LIMIT, ALPHA, BETA);          \
       });                                                                      \
     }                                                                          \
   } else {                                                                     \
@@ -252,7 +283,7 @@ packed_gelu_tanh_kernel(const packed_t& val) {
           PACKED_KERNEL<typename vllm::PackedTypeConverter<scalar_t>::Type>,   \
           ACT_FIRST, false, HAS_CLAMP><<<grid, block, 0, stream>>>(            \
           out.mutable_data_ptr<scalar_t>(), input.const_data_ptr<scalar_t>(),  \
-          d, LIMIT);                                                           \
+          d, LIMIT, ALPHA, BETA);                                              \
     });                                                                        \
   }
 
@@ -260,14 +291,18 @@ void silu_and_mul(torch::stable::Tensor& out,    // [..., d]
                   torch::stable::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
-                                true, false, 0.0f);
+                                true, false, 0.0f, 1.0f, 0.0f);
 }
 
 void silu_and_mul_clamp(torch::stable::Tensor& out,    // [..., d]
                         torch::stable::Tensor& input,  // [..., 2 * d]
-                        double limit) {
+                        double limit, double alpha, double beta) {
+  // out = (gate.clamp(max=limit) * sigmoid(alpha * gate.clamp(max=limit)))
+  //       * (up.clamp(+-limit) + beta)
+  // alpha=1.0, beta=0.0 reduce this to silu(gate) * up.
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
-                                true, true, (float)limit);
+                                true, true, (float)limit, (float)alpha,
+                                (float)beta);
 }
 
 void mul_and_silu(torch::stable::Tensor& out,    // [..., d]
@@ -276,21 +311,22 @@ void mul_and_silu(torch::stable::Tensor& out,    // [..., d]
   // The difference between mul_and_silu and silu_and_mul is that mul_and_silu
   // applies the silu to the latter half of the input.
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel, vllm::packed_silu_kernel,
-                                false, false, 0.0f);
+                                false, false, 0.0f, 1.0f, 0.0f);
 }
 
 void gelu_and_mul(torch::stable::Tensor& out,    // [..., d]
                   torch::stable::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel, vllm::packed_gelu_kernel,
-                                true, false, 0.0f);
+                                true, false, 0.0f, 1.0f, 0.0f);
 }
 
 void gelu_tanh_and_mul(torch::stable::Tensor& out,    // [..., d]
                        torch::stable::Tensor& input)  // [..., 2 * d]
 {
-  LAUNCH_ACTIVATION_GATE_KERNEL(
-      vllm::gelu_tanh_kernel, vllm::packed_gelu_tanh_kernel, true, false, 0.0f);
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel,
+                                vllm::packed_gelu_tanh_kernel, true, false,
+                                0.0f, 1.0f, 0.0f);
 }
 
 namespace vllm {

csrc/libtorch_stable/concat_mla_q.cuh

@@ -1,9 +1,6 @@
 #ifndef CONCAT_MLA_Q_CUH_
 #define CONCAT_MLA_Q_CUH_
 
-#include <cuda_bf16.h>
-#include <cuda_fp16.h>
-
 #include "cuda_vec_utils.cuh"
 
 namespace vllm {

csrc/libtorch_stable/cuda_vec_utils.cuh

@@ -21,7 +21,7 @@
 // together enable 256-bit (v8.u32) PTX load/store instructions.
 // Use for PTX instruction selection with architecture fallback paths.
 #if !defined(USE_ROCM) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 1000 && \
-    defined(CUDA_VERSION) && CUDA_VERSION >= 12090
+    defined(CUDART_VERSION) && CUDART_VERSION >= 12090
   #define VLLM_256B_PTX_ENABLED 1
 #else
   #define VLLM_256B_PTX_ENABLED 0

csrc/libtorch_stable/cuda_view.cu

@@ -0,0 +1,76 @@
+#include <torch/csrc/stable/tensor.h>
+#include <torch/csrc/stable/ops.h>
+#include <torch/csrc/stable/accelerator.h>
+#include <torch/headeronly/core/ScalarType.h>
+#include <torch/csrc/stable/device.h>
+#include <torch/csrc/stable/c/shim.h>
+#include <torch/headeronly/version.h>
+#include <cuda_runtime.h>
+
+#include <array>
+#include <optional>
+
+// This function assumes that `cpu_tensor` is a CPU tensor,
+// and that UVA (Unified Virtual Addressing) is enabled.
+torch::stable::Tensor get_cuda_view_from_cpu_tensor(
+    torch::stable::Tensor& cpu_tensor) {
+  STD_TORCH_CHECK(cpu_tensor.device().is_cpu(), "Input tensor must be on CPU");
+
+  const auto dtype = cpu_tensor.scalar_type();
+  const auto layout = cpu_tensor.layout();
+  const torch::stable::Device cuda_dev(torch::headeronly::DeviceType::CUDA);
+
+  // handle empty tensor
+  if (cpu_tensor.numel() == 0) {
+    return torch::stable::empty(cpu_tensor.sizes(), dtype, layout, cuda_dev);
+  }
+
+  std::array<StableIValue, 2> is_pinned_stack{
+      torch::stable::detail::from(cpu_tensor),
+      torch::stable::detail::from(std::nullopt)};
+  TORCH_ERROR_CODE_CHECK(torch_call_dispatcher(
+      "aten::is_pinned", "", is_pinned_stack.data(), TORCH_ABI_VERSION));
+  if (torch::stable::detail::to<bool>(is_pinned_stack[0])) {
+    // If CPU tensor is pinned, directly get the device pointer.
+    void* host_ptr = const_cast<void*>(cpu_tensor.mutable_data_ptr());
+    void* device_ptr = nullptr;
+    cudaError_t err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
+    STD_TORCH_CHECK(err == cudaSuccess, "cudaHostGetDevicePointer failed: ",
+                    cudaGetErrorString(err));
+
+    return torch::stable::from_blob(
+        device_ptr, cpu_tensor.sizes(), cpu_tensor.strides(), cuda_dev, dtype,
+        [base = cpu_tensor](void*) {});  // keep cpu tensor alive
+  }
+
+  // If CPU tensor is not pinned, allocate a new pinned memory buffer.
+  torch::stable::Tensor contiguous_cpu = torch::stable::contiguous(cpu_tensor);
+  size_t nbytes = contiguous_cpu.numel() * contiguous_cpu.element_size();
+
+  void* host_ptr = nullptr;
+  cudaError_t err = cudaHostAlloc(&host_ptr, nbytes, cudaHostAllocMapped);
+  if (err != cudaSuccess) {
+    STD_TORCH_CHECK(false, "cudaHostAlloc failed: ", cudaGetErrorString(err));
+  }
+
+  err = cudaMemcpy(host_ptr, contiguous_cpu.const_data_ptr(), nbytes,
+                   cudaMemcpyDefault);
+  if (err != cudaSuccess) {
+    cudaFreeHost(host_ptr);
+    STD_TORCH_CHECK(false, "cudaMemcpy failed: ", cudaGetErrorString(err));
+  }
+
+  void* device_ptr = nullptr;
+  err = cudaHostGetDevicePointer(&device_ptr, host_ptr, 0);
+  if (err != cudaSuccess) {
+    cudaFreeHost(host_ptr);
+    STD_TORCH_CHECK(
+        false, "cudaHostGetDevicePointer failed: ", cudaGetErrorString(err));
+  }
+
+  auto deleter = [host_ptr](void*) { cudaFreeHost(host_ptr); };
+
+  return torch::stable::from_blob(device_ptr, contiguous_cpu.sizes(),
+                                  contiguous_cpu.strides(), cuda_dev,
+                                  contiguous_cpu.scalar_type(), deleter);
+}

csrc/libtorch_stable/cutlass_extensions/common.cpp

@@ -1,4 +1,4 @@
-#include "cutlass_extensions/common.hpp"
+#include "common.hpp"
 
 int32_t get_sm_version_num() {
   int32_t major_capability, minor_capability;

csrc/libtorch_stable/dispatch_utils.h

@@ -30,6 +30,28 @@
   THO_DISPATCH_SWITCH(TYPE, NAME,                            \
                       VLLM_STABLE_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
 
+#define VLLM_STABLE_DISPATCH_CASE_INTEGRAL_TYPES(...)                  \
+  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Byte, __VA_ARGS__)  \
+  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Char, __VA_ARGS__)  \
+  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Short, __VA_ARGS__) \
+  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Int, __VA_ARGS__)   \
+  THO_DISPATCH_CASE(torch::headeronly::ScalarType::Long, __VA_ARGS__)
+
+#define VLLM_STABLE_DISPATCH_CASE_INTEGRAL_AND_UNSIGNED_TYPES(...)      \
+  VLLM_STABLE_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__)                 \
+  THO_DISPATCH_CASE(torch::headeronly::ScalarType::UInt16, __VA_ARGS__) \
+  THO_DISPATCH_CASE(torch::headeronly::ScalarType::UInt32, __VA_ARGS__) \
+  THO_DISPATCH_CASE(torch::headeronly::ScalarType::UInt64, __VA_ARGS__)
+
+#define VLLM_STABLE_DISPATCH_INTEGRAL_TYPES(TYPE, NAME, ...) \
+  THO_DISPATCH_SWITCH(TYPE, NAME,                            \
+                      VLLM_STABLE_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__))
+
+#define VLLM_STABLE_DISPATCH_INTEGRAL_AND_UNSIGNED_TYPES(TYPE, NAME, ...) \
+  THO_DISPATCH_SWITCH(                                                    \
+      TYPE, NAME,                                                         \
+      VLLM_STABLE_DISPATCH_CASE_INTEGRAL_AND_UNSIGNED_TYPES(__VA_ARGS__))
+
 // FP8 type dispatch - ROCm uses FNUZ format, CUDA uses OCP format
 #ifdef USE_ROCM
   #define VLLM_STABLE_DISPATCH_CASE_FP8_TYPES(...)                 \

csrc/libtorch_stable/fp32_router_gemm.cu

@@ -175,49 +175,52 @@ void invokeFp32RouterGemm(float* output, InputT const* mat_a,
 }
 
 // ---------------------------------------------------------------------------
-// Explicit instantiations: M=1..32, E=256, H=3072, for both input types
+// Explicit instantiations: M=1..32, for both input types, for the supported
+// (E, H) pairs:  (256, 3072) [MiniMax-M2/M2.5]  and  (128, 6144) [MiniMax-M3].
 // ---------------------------------------------------------------------------
 
-#define INSTANTIATE(T, M)                              \
-  template void invokeFp32RouterGemm<T, M, 256, 3072>( \
-      float*, T const*, float const*, cudaStream_t);
+#define INSTANTIATE(T, M, E, H)                                    \
+  template void invokeFp32RouterGemm<T, M, E, H>(float*, T const*, \
+                                                 float const*, cudaStream_t);
 
-#define INSTANTIATE_ALL(T) \
-  INSTANTIATE(T, 1)        \
-  INSTANTIATE(T, 2)        \
-  INSTANTIATE(T, 3)        \
-  INSTANTIATE(T, 4)        \
-  INSTANTIATE(T, 5)        \
-  INSTANTIATE(T, 6)        \
-  INSTANTIATE(T, 7)        \
-  INSTANTIATE(T, 8)        \
-  INSTANTIATE(T, 9)        \
-  INSTANTIATE(T, 10)       \
-  INSTANTIATE(T, 11)       \
-  INSTANTIATE(T, 12)       \
-  INSTANTIATE(T, 13)       \
-  INSTANTIATE(T, 14)       \
-  INSTANTIATE(T, 15)       \
-  INSTANTIATE(T, 16)       \
-  INSTANTIATE(T, 17)       \
-  INSTANTIATE(T, 18)       \
-  INSTANTIATE(T, 19)       \
-  INSTANTIATE(T, 20)       \
-  INSTANTIATE(T, 21)       \
-  INSTANTIATE(T, 22)       \
-  INSTANTIATE(T, 23)       \
-  INSTANTIATE(T, 24)       \
-  INSTANTIATE(T, 25)       \
-  INSTANTIATE(T, 26)       \
-  INSTANTIATE(T, 27)       \
-  INSTANTIATE(T, 28)       \
-  INSTANTIATE(T, 29)       \
-  INSTANTIATE(T, 30)       \
-  INSTANTIATE(T, 31)       \
-  INSTANTIATE(T, 32)
+#define INSTANTIATE_ALL(T, E, H) \
+  INSTANTIATE(T, 1, E, H)        \
+  INSTANTIATE(T, 2, E, H)        \
+  INSTANTIATE(T, 3, E, H)        \
+  INSTANTIATE(T, 4, E, H)        \
+  INSTANTIATE(T, 5, E, H)        \
+  INSTANTIATE(T, 6, E, H)        \
+  INSTANTIATE(T, 7, E, H)        \
+  INSTANTIATE(T, 8, E, H)        \
+  INSTANTIATE(T, 9, E, H)        \
+  INSTANTIATE(T, 10, E, H)       \
+  INSTANTIATE(T, 11, E, H)       \
+  INSTANTIATE(T, 12, E, H)       \
+  INSTANTIATE(T, 13, E, H)       \
+  INSTANTIATE(T, 14, E, H)       \
+  INSTANTIATE(T, 15, E, H)       \
+  INSTANTIATE(T, 16, E, H)       \
+  INSTANTIATE(T, 17, E, H)       \
+  INSTANTIATE(T, 18, E, H)       \
+  INSTANTIATE(T, 19, E, H)       \
+  INSTANTIATE(T, 20, E, H)       \
+  INSTANTIATE(T, 21, E, H)       \
+  INSTANTIATE(T, 22, E, H)       \
+  INSTANTIATE(T, 23, E, H)       \
+  INSTANTIATE(T, 24, E, H)       \
+  INSTANTIATE(T, 25, E, H)       \
+  INSTANTIATE(T, 26, E, H)       \
+  INSTANTIATE(T, 27, E, H)       \
+  INSTANTIATE(T, 28, E, H)       \
+  INSTANTIATE(T, 29, E, H)       \
+  INSTANTIATE(T, 30, E, H)       \
+  INSTANTIATE(T, 31, E, H)       \
+  INSTANTIATE(T, 32, E, H)
 
-INSTANTIATE_ALL(float)
-INSTANTIATE_ALL(__nv_bfloat16)
+INSTANTIATE_ALL(float, 256, 3072)
+INSTANTIATE_ALL(__nv_bfloat16, 256, 3072)
+INSTANTIATE_ALL(float, 128, 6144)
+INSTANTIATE_ALL(__nv_bfloat16, 128, 6144)
 
 #undef INSTANTIATE_ALL
 #undef INSTANTIATE

csrc/libtorch_stable/fp32_router_gemm_entry.cu

@@ -22,36 +22,42 @@ inline int getSMVersion() {
 
 }  // namespace
 
-static constexpr int FP32_NUM_EXPERTS = 256;
-static constexpr int FP32_HIDDEN_DIM = 3072;
 static constexpr int FP32_MAX_TOKENS = 32;
 
+// Supported (hidden_dim, num_experts) pairs (must match the instantiations in
+// fp32_router_gemm.cu): (3072, 256) for MiniMax-M2/M2.5, (6144, 128) for M3.
+static inline bool fp32_router_gemm_supported(int hidden_dim, int num_experts) {
+  return (hidden_dim == 3072 && num_experts == 256) ||
+         (hidden_dim == 6144 && num_experts == 128);
+}
+
 // Forward declarations — 4 template params must match fp32_router_gemm.cu
 template <typename InputT, int kNumTokens, int kNumExperts, int kHiddenDim>
 void invokeFp32RouterGemm(float* output, InputT const* mat_a,
                           float const* mat_b, cudaStream_t stream);
 
-// LoopUnroller templated on InputT
-template <typename InputT, int kBegin, int kEnd>
+// LoopUnroller templated on InputT, kNumExperts and kHiddenDim
+template <typename InputT, int kNumExperts, int kHiddenDim, int kBegin,
+          int kEnd>
 struct Fp32LoopUnroller {
   static void unroll(int num_tokens, float* output, InputT const* mat_a,
                      float const* mat_b, cudaStream_t stream) {
     if (num_tokens == kBegin) {
-      invokeFp32RouterGemm<InputT, kBegin, FP32_NUM_EXPERTS, FP32_HIDDEN_DIM>(
+      invokeFp32RouterGemm<InputT, kBegin, kNumExperts, kHiddenDim>(
           output, mat_a, mat_b, stream);
     } else {
-      Fp32LoopUnroller<InputT, kBegin + 1, kEnd>::unroll(num_tokens, output,
-                                                         mat_a, mat_b, stream);
+      Fp32LoopUnroller<InputT, kNumExperts, kHiddenDim, kBegin + 1,
+                       kEnd>::unroll(num_tokens, output, mat_a, mat_b, stream);
     }
   }
 };
 
-template <typename InputT, int kEnd>
-struct Fp32LoopUnroller<InputT, kEnd, kEnd> {
+template <typename InputT, int kNumExperts, int kHiddenDim, int kEnd>
+struct Fp32LoopUnroller<InputT, kNumExperts, kHiddenDim, kEnd, kEnd> {
   static void unroll(int num_tokens, float* output, InputT const* mat_a,
                      float const* mat_b, cudaStream_t stream) {
     if (num_tokens == kEnd) {
-      invokeFp32RouterGemm<InputT, kEnd, FP32_NUM_EXPERTS, FP32_HIDDEN_DIM>(
+      invokeFp32RouterGemm<InputT, kEnd, kNumExperts, kHiddenDim>(
           output, mat_a, mat_b, stream);
     } else {
       throw std::invalid_argument(
@@ -60,6 +66,23 @@ struct Fp32LoopUnroller<InputT, kEnd, kEnd> {
   }
 };
 
+// Dispatch over the supported (num_experts, hidden_dim) pairs.
+template <typename InputT>
+void dispatchFp32RouterGemm(int num_experts, int hidden_dim, int num_tokens,
+                            float* output, InputT const* mat_a,
+                            float const* mat_b, cudaStream_t stream) {
+  if (num_experts == 256 && hidden_dim == 3072) {
+    Fp32LoopUnroller<InputT, 256, 3072, 1, FP32_MAX_TOKENS>::unroll(
+        num_tokens, output, mat_a, mat_b, stream);
+  } else if (num_experts == 128 && hidden_dim == 6144) {
+    Fp32LoopUnroller<InputT, 128, 6144, 1, FP32_MAX_TOKENS>::unroll(
+        num_tokens, output, mat_a, mat_b, stream);
+  } else {
+    throw std::invalid_argument(
+        "fp32_router_gemm: unsupported (hidden_dim, num_experts) pair");
+  }
+}
+
 void fp32_router_gemm(
     torch::stable::Tensor& output,       // [num_tokens, num_experts]
     torch::stable::Tensor const& mat_a,  // [num_tokens, hidden_dim]
@@ -85,10 +108,10 @@ void fp32_router_gemm(
   STD_TORCH_CHECK(
       mat_a.size(1) == mat_b.size(1),
       "fp32_router_gemm: mat_a and mat_b must have the same hidden_dim");
-  STD_TORCH_CHECK(hidden_dim == FP32_HIDDEN_DIM,
-                  "fp32_router_gemm: expected hidden_dim=3072");
-  STD_TORCH_CHECK(num_experts == FP32_NUM_EXPERTS,
-                  "fp32_router_gemm: expected num_experts=256");
+  STD_TORCH_CHECK(
+      fp32_router_gemm_supported(hidden_dim, num_experts),
+      "fp32_router_gemm: supported (hidden_dim, num_experts) pairs are "
+      "(3072, 256) and (6144, 128)");
   STD_TORCH_CHECK(num_tokens <= FP32_MAX_TOKENS,
                   "fp32_router_gemm: num_tokens must be in [0, 32]");
   STD_TORCH_CHECK(
@@ -113,12 +136,13 @@ void fp32_router_gemm(
   if (mat_a.scalar_type() == torch::headeronly::ScalarType::BFloat16) {
     auto const* mat_a_ptr =
         reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr());
-    Fp32LoopUnroller<__nv_bfloat16, 1, FP32_MAX_TOKENS>::unroll(
-        num_tokens, out_ptr, mat_a_ptr, mat_b_ptr, stream);
+    dispatchFp32RouterGemm<__nv_bfloat16>(num_experts, hidden_dim, num_tokens,
+                                          out_ptr, mat_a_ptr, mat_b_ptr,
+                                          stream);
   } else {
     auto const* mat_a_ptr = reinterpret_cast<float const*>(mat_a.data_ptr());
-    Fp32LoopUnroller<float, 1, FP32_MAX_TOKENS>::unroll(
-        num_tokens, out_ptr, mat_a_ptr, mat_b_ptr, stream);
+    dispatchFp32RouterGemm<float>(num_experts, hidden_dim, num_tokens, out_ptr,
+                                  mat_a_ptr, mat_b_ptr, stream);
   }
 }
 

csrc/libtorch_stable/fused_deepseek_v4_qnorm_rope_kv_insert_kernel.cu

@@ -102,6 +102,35 @@ constexpr float NUM_TOKEN_CUTOFF = 1024;
 constexpr int kNumLanes = 32;
 constexpr int kElemsPerLane = kHeadDim / kNumLanes;  // 16
 
+// Pack this lane's 16 fp32 elements into per-tensor E4M3 FP8 (one uint4 = 16
+// B), scaling by `scale` (a reciprocal scale) and saturating to ±448.  Used by
+// the FlashInfer full-cache path for both the Q and KV stores.
+__device__ __forceinline__ uint4 packFp8E4M3x16(float const* values,
+                                                float const scale) {
+#ifndef USE_ROCM
+  uint4 out;
+  auto* out2 = reinterpret_cast<__nv_fp8x2_storage_t*>(&out);
+  #pragma unroll
+  for (int i = 0; i < kElemsPerLane / 2; i++) {
+    float2 scaled =
+        make_float2(values[2 * i] * scale, values[2 * i + 1] * scale);
+    scaled.x = fminf(fmaxf(scaled.x, -kFp8Max), kFp8Max);
+    scaled.y = fminf(fmaxf(scaled.y, -kFp8Max), kFp8Max);
+    out2[i] = __nv_cvt_float2_to_fp8x2(scaled, __NV_SATFINITE, __NV_E4M3);
+  }
+  return out;
+#else
+  uint8_t out_bytes[kElemsPerLane];
+  #pragma unroll
+  for (int i = 0; i < kElemsPerLane; i++) {
+    float scaled = values[i] * scale;
+    scaled = fminf(fmaxf(scaled, -kFp8Max), kFp8Max);
+    out_bytes[i] = rocm_cvt_float_to_fp8_e4m3(scaled);
+  }
+  return *reinterpret_cast<uint4 const*>(out_bytes);
+#endif
+}
+
 // ────────────────────────────────────────────────────────────────────────────
 // Small inline helpers
 // ────────────────────────────────────────────────────────────────────────────
@@ -649,6 +678,257 @@ void launchFusedDeepseekV4QNormRopeKVRopeQuantInsert(
 #undef DISPATCH
 }
 
+// ────────────────────────────────────────────────────────────────────────────
+// FlashInfer full-cache kernel
+// ────────────────────────────────────────────────────────────────────────────
+//
+// Sibling to the FlashMLA kernel above, used by the FlashInfer V4 sparse-MLA
+// backend.  Differences from the legacy path:
+//   * No Q head padding — output Q layout matches the input num_heads_q.
+//   * KV is written as a *contiguous* 512-wide row per token (token-strided),
+//     not the legacy UE8M0 paged layout with a separate scale tail.
+//   * Q/KV are stored either as bf16 or as per-tensor E4M3 FP8 (one global
+//     scale), selected by the STORE_Q_FP8 / STORE_KV_FP8 template flags.
+//
+// Grid: 1D, gridDim.x = ceil(num_tokens_full * (num_heads_q + 1) / warps).
+// Each warp handles one (token, slot): slot < num_heads_q → Q, slot ==
+// num_heads_q → KV.
+template <typename scalar_t_in, bool STORE_Q_FP8, bool STORE_KV_FP8>
+__global__ void fusedDeepseekV4FullCacheKernel(
+    scalar_t_in* __restrict__ q_inout,          // [N, H, 512], in place (bf16)
+    uint8_t* __restrict__ q_fp8_out,            // [N, H, 512] fp8, optional
+    int64_t const q_fp8_stride0,                // elements (fp8 == bytes)
+    int64_t const q_fp8_stride1,                // elements (fp8 == bytes)
+    scalar_t_in const* __restrict__ kv_in,      // [N, 512] bf16
+    uint8_t* __restrict__ k_cache,              // contiguous bf16 or fp8 cache
+    int64_t const* __restrict__ slot_mapping,   // [num_tokens_insert] i64
+    int64_t const* __restrict__ position_ids,   // [N] i64
+    float const* __restrict__ cos_sin_cache,    // [max_pos, 64] fp32
+    float const* __restrict__ fp8_scale_ptr,    // scalar, KV fp8 only
+    float const* __restrict__ q_fp8_scale_inv,  // scalar, Q fp8 only
+    float const eps,
+    int const num_tokens_full,      // = q.size(0) = kv.size(0)
+    int const num_tokens_insert,    // = slot_mapping.size(0)
+    int const num_heads_q,          // H (no padding)
+    int const cache_block_size,     // tokens per cache block
+    int64_t const kv_block_stride,  // bytes per cache block
+    int64_t const kv_token_stride) {  // bytes per cache token
+#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
+  if constexpr (std::is_same_v<scalar_t_in, c10::BFloat16>) {
+    return;
+  } else {
+#endif
+    using Converter = vllm::_typeConvert<scalar_t_in>;
+    int const warpsPerBlock = blockDim.x / 32;
+    int const warpId = threadIdx.x / 32;
+    int const laneId = threadIdx.x % 32;
+    int const globalWarpIdx = blockIdx.x * warpsPerBlock + warpId;
+
+    int const slotsPerToken = num_heads_q + 1;
+    int const tokenIdx = globalWarpIdx / slotsPerToken;
+    int const slotIdx = globalWarpIdx % slotsPerToken;
+    if (tokenIdx >= num_tokens_full) return;
+    bool const isKV = (slotIdx == num_heads_q);
+    // KV branch: skip DP-padded tokens (no slot reserved for them).
+    if (isKV && tokenIdx >= num_tokens_insert) return;
+
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
+    cudaGridDependencySynchronize();
+#endif
+
+    int const dim_base = laneId * kElemsPerLane;  // in [0, 512) step 16
+    scalar_t_in const* src_ptr;
+    if (isKV) {
+      src_ptr = kv_in + static_cast<int64_t>(tokenIdx) * kHeadDim + dim_base;
+    } else {
+      src_ptr = q_inout +
+                (static_cast<int64_t>(tokenIdx) * num_heads_q + slotIdx) *
+                    kHeadDim +
+                dim_base;
+    }
+    uint4 const v0 = *reinterpret_cast<uint4 const*>(src_ptr);
+    uint4 const v1 = *reinterpret_cast<uint4 const*>(src_ptr + 8);
+
+    // ── Decode bf16 → 16 fp32 registers ───────────────────────────────────
+    float elements[kElemsPerLane];
+    {
+      auto const* p0 =
+          reinterpret_cast<typename Converter::packed_hip_type const*>(&v0);
+      auto const* p1 =
+          reinterpret_cast<typename Converter::packed_hip_type const*>(&v1);
+#pragma unroll
+      for (int i = 0; i < 4; i++) {
+        float2 f2 = Converter::convert(p0[i]);
+        elements[2 * i] = f2.x;
+        elements[2 * i + 1] = f2.y;
+      }
+#pragma unroll
+      for (int i = 0; i < 4; i++) {
+        float2 f2 = Converter::convert(p1[i]);
+        elements[8 + 2 * i] = f2.x;
+        elements[8 + 2 * i + 1] = f2.y;
+      }
+    }
+
+    // ── Q branch: RMSNorm (no weight) ─────────────────────────────────────
+    if (!isKV) {
+      float sumOfSquares = 0.0f;
+#pragma unroll
+      for (int i = 0; i < kElemsPerLane; i++) {
+        sumOfSquares += elements[i] * elements[i];
+      }
+      sumOfSquares = warpSum<float>(sumOfSquares);
+      float const rms_rcp =
+          rsqrtf(sumOfSquares / static_cast<float>(kHeadDim) + eps);
+#pragma unroll
+      for (int i = 0; i < kElemsPerLane; i++) {
+        elements[i] = elements[i] * rms_rcp;
+      }
+    }
+
+    // ── GPT-J RoPE on dims [NOPE_DIM, HEAD_DIM) ───────────────────────────
+    bool const is_rope_lane = dim_base >= kNopeDim;
+    if (is_rope_lane) {
+      int64_t const pos = position_ids[tokenIdx];
+      constexpr int kHalfRope = kRopeDim / 2;
+      float const* cos_ptr = cos_sin_cache + pos * kRopeDim;
+      float const* sin_ptr = cos_ptr + kHalfRope;
+      int const rope_local_base = dim_base - kNopeDim;
+      int const half_base = rope_local_base >> 1;
+      float4 const c0 = *reinterpret_cast<float4 const*>(cos_ptr + half_base);
+      float4 const c1 = *reinterpret_cast<float4 const*>(cos_ptr + half_base + 4);
+      float4 const s0 = *reinterpret_cast<float4 const*>(sin_ptr + half_base);
+      float4 const s1 = *reinterpret_cast<float4 const*>(sin_ptr + half_base + 4);
+      float const cos_arr[8] = {c0.x, c0.y, c0.z, c0.w, c1.x, c1.y, c1.z, c1.w};
+      float const sin_arr[8] = {s0.x, s0.y, s0.z, s0.w, s1.x, s1.y, s1.z, s1.w};
+#pragma unroll
+      for (int p = 0; p < kElemsPerLane / 2; p++) {
+        float const x_even = elements[2 * p];
+        float const x_odd = elements[2 * p + 1];
+        elements[2 * p] = x_even * cos_arr[p] - x_odd * sin_arr[p];
+        elements[2 * p + 1] = x_even * sin_arr[p] + x_odd * cos_arr[p];
+      }
+    }
+
+    // ── Store ─────────────────────────────────────────────────────────────
+    if (!isKV) {
+      if constexpr (STORE_Q_FP8) {
+        float const scale_inv = VLLM_LDG(q_fp8_scale_inv);
+        uint4 const out = packFp8E4M3x16(elements, scale_inv);
+        uint8_t* dst = q_fp8_out +
+                       static_cast<int64_t>(tokenIdx) * q_fp8_stride0 +
+                       static_cast<int64_t>(slotIdx) * q_fp8_stride1 + dim_base;
+        *reinterpret_cast<uint4*>(dst) = out;
+      } else {
+        uint4 out0, out1;
+        auto* po0 = reinterpret_cast<typename Converter::packed_hip_type*>(&out0);
+        auto* po1 = reinterpret_cast<typename Converter::packed_hip_type*>(&out1);
+#pragma unroll
+        for (int i = 0; i < 4; i++) {
+          po0[i] = Converter::convert(
+              make_float2(elements[2 * i], elements[2 * i + 1]));
+        }
+#pragma unroll
+        for (int i = 0; i < 4; i++) {
+          po1[i] = Converter::convert(
+              make_float2(elements[8 + 2 * i], elements[8 + 2 * i + 1]));
+        }
+        scalar_t_in* dst =
+            q_inout +
+            (static_cast<int64_t>(tokenIdx) * num_heads_q + slotIdx) * kHeadDim +
+            dim_base;
+        *reinterpret_cast<uint4*>(dst) = out0;
+        *reinterpret_cast<uint4*>(dst + 8) = out1;
+      }
+    } else {
+      int64_t const slot_id = slot_mapping[tokenIdx];
+      if (slot_id >= 0) {
+        int64_t const block_idx = slot_id / cache_block_size;
+        int64_t const pos_in_block = slot_id % cache_block_size;
+        uint8_t* cache_row =
+            k_cache + block_idx * kv_block_stride + pos_in_block * kv_token_stride;
+        if constexpr (STORE_KV_FP8) {
+          float const inv_scale = 1.0f / VLLM_LDG(fp8_scale_ptr);
+          uint4 const out = packFp8E4M3x16(elements, inv_scale);
+          *reinterpret_cast<uint4*>(cache_row + dim_base) = out;
+        } else {
+          uint4 out0, out1;
+          auto* po0 =
+              reinterpret_cast<typename Converter::packed_hip_type*>(&out0);
+          auto* po1 =
+              reinterpret_cast<typename Converter::packed_hip_type*>(&out1);
+#pragma unroll
+          for (int i = 0; i < 4; i++) {
+            po0[i] = Converter::convert(
+                make_float2(elements[2 * i], elements[2 * i + 1]));
+          }
+#pragma unroll
+          for (int i = 0; i < 4; i++) {
+            po1[i] = Converter::convert(
+                make_float2(elements[8 + 2 * i], elements[8 + 2 * i + 1]));
+          }
+          scalar_t_in* dst = reinterpret_cast<scalar_t_in*>(cache_row) + dim_base;
+          *reinterpret_cast<uint4*>(dst) = out0;
+          *reinterpret_cast<uint4*>(dst + 8) = out1;
+        }
+      }
+    }
+
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
+    cudaTriggerProgrammaticLaunchCompletion();
+#endif
+#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
+  }
+#endif
+}
+
+// Configure + launch helper shared by the bf16 and fp8 full-cache launchers.
+template <typename scalar_t_in, bool STORE_Q_FP8, bool STORE_KV_FP8>
+static void launchFullCacheKernel(
+    scalar_t_in* q_inout, uint8_t* q_fp8_out, int64_t q_fp8_stride0,
+    int64_t q_fp8_stride1, scalar_t_in const* kv_in, uint8_t* k_cache,
+    int64_t const* slot_mapping, int64_t const* position_ids,
+    float const* cos_sin_cache, float const* fp8_scale,
+    float const* q_fp8_scale_inv, float const eps, int const num_tokens_full,
+    int const num_tokens_insert, int const num_heads_q,
+    int const cache_block_size, int64_t const kv_block_stride,
+    int64_t const kv_token_stride, char const* op_name, cudaStream_t stream) {
+  constexpr int kBlockSize = 256;
+  constexpr int kWarpsPerBlock = kBlockSize / 32;
+  int64_t const total_warps =
+      static_cast<int64_t>(num_tokens_full) * (num_heads_q + 1);
+  int const grid =
+      static_cast<int>((total_warps + kWarpsPerBlock - 1) / kWarpsPerBlock);
+  auto* kernel =
+      fusedDeepseekV4FullCacheKernel<scalar_t_in, STORE_Q_FP8, STORE_KV_FP8>;
+#ifndef USE_ROCM
+  static int const sm_version = getSMVersion();
+  STD_TORCH_CHECK(sm_version >= 80, op_name,
+                  " requires sm_80+ (Ampere or newer); got sm_", sm_version);
+  cudaLaunchConfig_t config;
+  config.gridDim = dim3(grid);
+  config.blockDim = dim3(kBlockSize);
+  config.dynamicSmemBytes = 0;
+  config.stream = stream;
+  cudaLaunchAttribute attrs[1];
+  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+  attrs[0].val.programmaticStreamSerializationAllowed = 1;
+  config.attrs = attrs;
+  config.numAttrs = (sm_version >= 90) ? 1 : 0;
+  cudaLaunchKernelEx(&config, kernel, q_inout, q_fp8_out, q_fp8_stride0,
+                     q_fp8_stride1, kv_in, k_cache, slot_mapping, position_ids,
+                     cos_sin_cache, fp8_scale, q_fp8_scale_inv, eps,
+                     num_tokens_full, num_tokens_insert, num_heads_q,
+                     cache_block_size, kv_block_stride, kv_token_stride);
+#else
+  kernel<<<grid, kBlockSize, 0, stream>>>(
+      q_inout, q_fp8_out, q_fp8_stride0, q_fp8_stride1, kv_in, k_cache,
+      slot_mapping, position_ids, cos_sin_cache, fp8_scale, q_fp8_scale_inv,
+      eps, num_tokens_full, num_tokens_insert, num_heads_q, cache_block_size,
+      kv_block_stride, kv_token_stride);
+#endif
+}
+
 }  // namespace deepseek_v4_fused_ops
 }  // namespace vllm
 
@@ -735,3 +1015,167 @@ torch::stable::Tensor fused_deepseek_v4_qnorm_rope_kv_rope_quant_insert(
       });
   return q_out;
 }
+
+// ────────────────────────────────────────────────────────────────────────────
+// FlashInfer full-cache torch ops
+// ────────────────────────────────────────────────────────────────────────────
+void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert(
+    torch::stable::Tensor& q,                    // [N, H, 512] bf16, in place
+    torch::stable::Tensor const& kv,             // [N, 512] bf16, read-only
+    torch::stable::Tensor& k_cache,              // [num_blocks, bs, 512] bf16
+    torch::stable::Tensor const& slot_mapping,   // [num_tokens_insert] int64
+    torch::stable::Tensor const& position_ids,   // [N] int64
+    torch::stable::Tensor const& cos_sin_cache,  // [max_pos, 64] float32
+    double eps, int64_t cache_block_size) {
+  using torch::headeronly::ScalarType;
+  STD_TORCH_CHECK(q.device().is_cuda() && q.is_contiguous(),
+                  "q must be contiguous CUDA");
+  STD_TORCH_CHECK(kv.device().is_cuda() && kv.is_contiguous(),
+                  "kv must be contiguous CUDA");
+  STD_TORCH_CHECK(k_cache.device().is_cuda(), "k_cache must be CUDA");
+  STD_TORCH_CHECK(slot_mapping.device().is_cuda() &&
+                      slot_mapping.scalar_type() == ScalarType::Long,
+                  "slot_mapping must be int64 CUDA");
+  STD_TORCH_CHECK(position_ids.device().is_cuda() &&
+                      position_ids.scalar_type() == ScalarType::Long,
+                  "position_ids must be int64 CUDA");
+  STD_TORCH_CHECK(cos_sin_cache.device().is_cuda() &&
+                      cos_sin_cache.scalar_type() == ScalarType::Float &&
+                      cos_sin_cache.dim() == 2 && cos_sin_cache.size(1) == 64,
+                  "cos_sin_cache shape [max_pos, 64] float32");
+  STD_TORCH_CHECK(q.dim() == 3 && q.size(2) == 512, "q shape [N, H, 512]");
+  STD_TORCH_CHECK(kv.dim() == 2 && kv.size(1) == 512, "kv shape [N, 512]");
+  STD_TORCH_CHECK(q.scalar_type() == ScalarType::BFloat16 &&
+                      kv.scalar_type() == ScalarType::BFloat16,
+                  "q and kv must be bfloat16");
+  STD_TORCH_CHECK(k_cache.dim() == 3 && k_cache.size(1) == cache_block_size &&
+                      k_cache.size(2) == 512 && k_cache.stride(2) == 1,
+                  "k_cache shape [num_blocks, cache_block_size, 512] contiguous");
+  STD_TORCH_CHECK(k_cache.scalar_type() == ScalarType::BFloat16,
+                  "k_cache must be bfloat16");
+
+  int const num_tokens_full = static_cast<int>(q.size(0));
+  int const num_tokens_insert = static_cast<int>(slot_mapping.size(0));
+  STD_TORCH_CHECK(static_cast<int>(kv.size(0)) == num_tokens_full &&
+                      static_cast<int>(position_ids.size(0)) == num_tokens_full,
+                  "q/kv/position_ids row counts must match");
+  STD_TORCH_CHECK(num_tokens_insert <= num_tokens_full,
+                  "slot_mapping must not exceed q row count");
+  int const num_heads_q = static_cast<int>(q.size(1));
+
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      q.get_device_index());
+  const cudaStream_t stream = get_current_cuda_stream(q.get_device_index());
+
+  // bf16 cache: 2 bytes/element -> byte strides for the uint8-addressed kernel.
+  int64_t const kv_block_stride = k_cache.stride(0) * 2;
+  int64_t const kv_token_stride = k_cache.stride(1) * 2;
+
+  VLLM_STABLE_DISPATCH_HALF_TYPES(
+      q.scalar_type(),
+      "fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert", [&] {
+        vllm::deepseek_v4_fused_ops::launchFullCacheKernel<scalar_t, false,
+                                                           false>(
+            reinterpret_cast<scalar_t*>(q.mutable_data_ptr()), nullptr, 0, 0,
+            reinterpret_cast<scalar_t const*>(kv.const_data_ptr()),
+            reinterpret_cast<uint8_t*>(k_cache.mutable_data_ptr()),
+            slot_mapping.const_data_ptr<int64_t>(),
+            position_ids.const_data_ptr<int64_t>(),
+            cos_sin_cache.const_data_ptr<float>(), nullptr, nullptr,
+            static_cast<float>(eps), num_tokens_full, num_tokens_insert,
+            num_heads_q, static_cast<int>(cache_block_size), kv_block_stride,
+            kv_token_stride,
+            "fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_bf16_insert",
+            stream);
+      });
+}
+
+void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert(
+    torch::stable::Tensor const& q,                // [N, H, 512] bf16, read-only
+    torch::stable::Tensor const& kv,               // [N, 512] bf16, read-only
+    torch::stable::Tensor& q_fp8,                  // [N, H, 512] fp8 e4m3
+    torch::stable::Tensor& k_cache,                // [num_blocks, bs, 512] fp8
+    torch::stable::Tensor const& slot_mapping,     // [num_tokens_insert] int64
+    torch::stable::Tensor const& position_ids,     // [N] int64
+    torch::stable::Tensor const& cos_sin_cache,    // [max_pos, 64] float32
+    torch::stable::Tensor const& fp8_scale,        // scalar float32 (KV scale)
+    torch::stable::Tensor const& q_fp8_scale_inv,  // scalar float32 (1 / Q scale)
+    double eps, int64_t cache_block_size) {
+  using torch::headeronly::ScalarType;
+  STD_TORCH_CHECK(q.device().is_cuda() && q.is_contiguous(),
+                  "q must be contiguous CUDA");
+  STD_TORCH_CHECK(kv.device().is_cuda() && kv.is_contiguous(),
+                  "kv must be contiguous CUDA");
+  STD_TORCH_CHECK(q_fp8.device().is_cuda() && q_fp8.is_contiguous() &&
+                      q_fp8.scalar_type() == ScalarType::Float8_e4m3fn &&
+                      q_fp8.dim() == 3 && q_fp8.size(0) == q.size(0) &&
+                      q_fp8.size(1) == q.size(1) && q_fp8.size(2) == q.size(2),
+                  "q_fp8 must be a contiguous float8_e4m3fn tensor matching q");
+  STD_TORCH_CHECK(k_cache.device().is_cuda(), "k_cache must be CUDA");
+  STD_TORCH_CHECK(slot_mapping.device().is_cuda() &&
+                      slot_mapping.scalar_type() == ScalarType::Long,
+                  "slot_mapping must be int64 CUDA");
+  STD_TORCH_CHECK(position_ids.device().is_cuda() &&
+                      position_ids.scalar_type() == ScalarType::Long,
+                  "position_ids must be int64 CUDA");
+  STD_TORCH_CHECK(cos_sin_cache.device().is_cuda() &&
+                      cos_sin_cache.scalar_type() == ScalarType::Float &&
+                      cos_sin_cache.dim() == 2 && cos_sin_cache.size(1) == 64,
+                  "cos_sin_cache shape [max_pos, 64] float32");
+  STD_TORCH_CHECK(fp8_scale.device().is_cuda() &&
+                      fp8_scale.scalar_type() == ScalarType::Float &&
+                      fp8_scale.size(0) == 1,
+                  "fp8_scale must be a scalar float32 CUDA tensor");
+  STD_TORCH_CHECK(q_fp8_scale_inv.device().is_cuda() &&
+                      q_fp8_scale_inv.scalar_type() == ScalarType::Float &&
+                      q_fp8_scale_inv.size(0) == 1,
+                  "q_fp8_scale_inv must be a scalar float32 CUDA tensor");
+  STD_TORCH_CHECK(q.dim() == 3 && q.size(2) == 512, "q shape [N, H, 512]");
+  STD_TORCH_CHECK(kv.dim() == 2 && kv.size(1) == 512, "kv shape [N, 512]");
+  STD_TORCH_CHECK(q.scalar_type() == kv.scalar_type(),
+                  "q and kv dtype must match");
+  STD_TORCH_CHECK(k_cache.dim() == 3 && k_cache.size(1) == cache_block_size &&
+                      k_cache.size(2) == 512 && k_cache.stride(2) == 1,
+                  "k_cache shape [num_blocks, cache_block_size, 512] contiguous");
+  STD_TORCH_CHECK(k_cache.scalar_type() == ScalarType::Float8_e4m3fn,
+                  "k_cache must be float8_e4m3fn");
+
+  int const num_tokens_full = static_cast<int>(q.size(0));
+  int const num_tokens_insert = static_cast<int>(slot_mapping.size(0));
+  STD_TORCH_CHECK(static_cast<int>(kv.size(0)) == num_tokens_full &&
+                      static_cast<int>(position_ids.size(0)) == num_tokens_full,
+                  "q/kv/position_ids row counts must match");
+  STD_TORCH_CHECK(num_tokens_insert <= num_tokens_full,
+                  "slot_mapping must not exceed q row count");
+  int const num_heads_q = static_cast<int>(q.size(1));
+
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      q.get_device_index());
+  const cudaStream_t stream = get_current_cuda_stream(q.get_device_index());
+
+  VLLM_STABLE_DISPATCH_HALF_TYPES(
+      q.scalar_type(),
+      "fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert", [&] {
+        vllm::deepseek_v4_fused_ops::launchFullCacheKernel<scalar_t, true,
+                                                           true>(
+            // q is read-only in the fp8 path (the kernel writes q_fp8); the
+            // launcher signature is non-const, so cast away const on the ptr.
+            reinterpret_cast<scalar_t*>(
+                const_cast<void*>(q.const_data_ptr())),
+            reinterpret_cast<uint8_t*>(q_fp8.mutable_data_ptr()),
+            q_fp8.stride(0), q_fp8.stride(1),
+            reinterpret_cast<scalar_t const*>(kv.const_data_ptr()),
+            reinterpret_cast<uint8_t*>(k_cache.mutable_data_ptr()),
+            slot_mapping.const_data_ptr<int64_t>(),
+            position_ids.const_data_ptr<int64_t>(),
+            cos_sin_cache.const_data_ptr<float>(),
+            fp8_scale.const_data_ptr<float>(),
+            q_fp8_scale_inv.const_data_ptr<float>(), static_cast<float>(eps),
+            num_tokens_full, num_tokens_insert, num_heads_q,
+            static_cast<int>(cache_block_size),
+            // fp8 cache: 1 byte/element -> stride already in bytes.
+            k_cache.stride(0), k_cache.stride(1),
+            "fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert",
+            stream);
+      });
+}

csrc/libtorch_stable/fused_minimax_m3_qknorm_rope_kv_insert_kernel.cu

@@ -0,0 +1,635 @@
+/*
+ * SPDX-License-Identifier: Apache-2.0
+ * SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+ *
+ * Horizontally-fused MiniMax-M3 attention pre-processing kernel.
+ *
+ * Replaces the per-token Python sequence in
+ * ``MiniMaxM3SparseAttention.forward`` / ``MiniMaxM3Attention.forward``:
+ *
+ *     q  = q_norm(q);  k = k_norm(k);  q, k = rotary_emb(pos, q, k)
+ *     index_q = index_q_norm(index_q);  index_k = index_k_norm(index_k)
+ *     index_q, index_k = rotary_emb(pos, index_q, index_k)
+ *     _insert_kv(k, v, index_k)
+ *
+ * All branches share head_dim=128 and the *same* partial-NeoX RoPE table
+ * (``rotary_dim`` rotated, the trailing dims pass through).  The four norms
+ * are Gemma-style RMSNorm (``x * rsqrt(mean(x^2)+eps) * (1 + weight)``) with
+ * independent weights.
+ *
+ * Everything lives in a single fused ``qkv`` tensor.  The sparse layer's
+ * fused projection (MinimaxM3QKVParallelLinearWithIndexer) emits, per token::
+ *
+ *     [ q | k | v | index_q | index_k ]   (the "5 results")
+ *
+ * while the dense layer emits just ``[ q | k | v ]``.  The kernel reads the
+ * index branch straight out of that packed row -- no separate index tensors.
+ *
+ * One kernel, one grid; each warp owns one (token, head-slot) pair.  Slot
+ * enumeration per token:
+ *     [0, nq)                         Q  heads   -> norm(q_w)  + RoPE, write
+ * qkv [nq, nq+nkv)                    K  heads   -> norm(k_w)  + RoPE, write
+ * qkv
+ *                                                  (+ insert into key cache)
+ *     [nq+nkv, nq+2*nkv)              V  heads   -> insert into value cache
+ *     IQ heads (niq)                             -> norm(iq_w) + RoPE, write iq
+ *     IK       (1)                               -> norm(ik_w) + RoPE
+ *                                                  (+ insert into index cache)
+ *
+ * The IQ/IK warps address the index_q/index_k sub-blocks *inside* qkv at the
+ * fixed physical offsets (nq+2*nkv)*128 and (nq+2*nkv+niq)*128.
+ *
+ * Dense vs sparse is a compile-time choice via the ``kIsSparse``/``kInsertKV``
+ * template bools (3 instantiations: dense <false,false>, sparse-profiling
+ * <true,false>, sparse-serving <true,true>), so the index slots, the V slots
+ * and the cache inserts fold away entirely on paths that don't use them. The
+ * dense layer passes no caches/index: norm+RoPE happens in place and the
+ * generic ``Attention`` layer owns the cache write.
+ *
+ * Q/K and (sparse) index_q/index_k are all rewritten in place inside the fused
+ * ``qkv`` tensor.  Caches (bf16) are scatter-written by slot.
+ */
+
+#include <cmath>
+#include <cuda_runtime.h>
+#include <type_traits>
+
+#include "torch_utils.h"
+
+#include "../cuda_compat.h"
+#include "../type_convert.cuh"
+#include "dispatch_utils.h"
+
+#ifndef FINAL_MASK
+  #ifdef USE_ROCM
+    #define FINAL_MASK 0xffffffffffffffffULL
+  #else
+    #define FINAL_MASK 0xffffffffu
+  #endif
+#endif
+
+namespace vllm {
+namespace minimax_m3_fused_ops {
+
+namespace {
+inline int getSMVersion() {
+  auto* props = get_device_prop();
+  return props->major * 10 + props->minor;
+}
+}  // namespace
+
+// ────────────────────────────────────────────────────────────────────────────
+// Constants (hard-coded for MiniMax-M3-preview).
+// ────────────────────────────────────────────────────────────────────────────
+constexpr int kHeadDim = 128;
+constexpr int kNumLanes = 32;
+constexpr int kElemsPerLane = kHeadDim / kNumLanes;  // 4
+
+// ────────────────────────────────────────────────────────────────────────────
+// Helpers
+// ────────────────────────────────────────────────────────────────────────────
+__device__ __forceinline__ float warpReduceSum(float val) {
+#pragma unroll
+  for (int mask = 16; mask > 0; mask >>= 1) {
+    val += __shfl_xor_sync(FINAL_MASK, val, mask, 32);
+  }
+  return val;
+}
+
+// Gemma RMSNorm over the full head (no-op when ``weight == nullptr``), rounded
+// back to scalar_t like the materialized unfused norm output, followed by
+// partial NeoX RoPE on the leading ``rotary_dim`` dims. Each lane owns
+// ``kElemsPerLane`` contiguous dims [laneId*4, laneId*4+4).
+template <typename scalar_t>
+__device__ __forceinline__ void normAndRope(
+    float (&elems)[kElemsPerLane], int const laneId, float const eps,
+    scalar_t const* __restrict__ weight,  // [kHeadDim] or nullptr (no norm)
+    bool const do_rope, int const rotary_dim,
+    scalar_t const* __restrict__ cos_ptr,  // cos_sin_cache + pos*rotary_dim
+    bool const apply_norm) {
+  // ── Gemma RMSNorm: x * rsqrt(mean(x^2)+eps) * (1 + w) ──────────────────
+  if (apply_norm) {
+    float sumsq = 0.0f;
+#pragma unroll
+    for (int i = 0; i < kElemsPerLane; i++) sumsq += elems[i] * elems[i];
+    sumsq = warpReduceSum(sumsq);
+    float const rms_rcp = rsqrtf(sumsq / static_cast<float>(kHeadDim) + eps);
+#pragma unroll
+    for (int i = 0; i < kElemsPerLane; i++) {
+      int const dim = laneId * kElemsPerLane + i;
+      float const w = 1.0f + static_cast<float>(weight[dim]);
+      elems[i] = elems[i] * rms_rcp * w;
+    }
+  }
+
+  // ── Partial NeoX RoPE on dims [0, rotary_dim) ──────────────────────────
+  // half = rotary_dim/2.  Pair (i, i+half) for i in [0, half).  Lane L owns
+  // dims [4L, 4L+4); since half is a multiple of 4, a lane lies wholly in the
+  // first half (own=x[i]) or second half (own=x[i+half]); its partner lives
+  // ``half/4`` lanes away (XOR with that distance).
+  if (do_rope) {
+    int const half = rotary_dim / 2;
+    int const dim0 = laneId * kElemsPerLane;
+    bool const in_rope = dim0 < rotary_dim;
+    int const lane_xor = half / kElemsPerLane;  // partner-lane distance
+
+    float partner[kElemsPerLane];
+#pragma unroll
+    for (int i = 0; i < kElemsPerLane; i++) {
+      partner[i] = __shfl_xor_sync(FINAL_MASK, elems[i], lane_xor, 32);
+    }
+    if (in_rope) {
+      bool const first_half = dim0 < half;
+      int const i_base = first_half ? dim0 : (dim0 - half);  // cos/sin index
+      scalar_t const* sin_ptr = cos_ptr + half;
+#pragma unroll
+      for (int i = 0; i < kElemsPerLane; i++) {
+        float const c = static_cast<float>(cos_ptr[i_base + i]);
+        float const s = static_cast<float>(sin_ptr[i_base + i]);
+        if (first_half) {
+          elems[i] = elems[i] * c - partner[i] * s;
+        } else {
+          elems[i] = elems[i] * c + partner[i] * s;
+        }
+      }
+    }
+  }
+}
+
+// Load 4 contiguous bf16 -> 4 fp32 registers.
+template <typename scalar_t>
+__device__ __forceinline__ void loadElems(scalar_t const* __restrict__ src,
+                                          float (&elems)[kElemsPerLane]) {
+  using Converter = vllm::_typeConvert<scalar_t>;
+  uint2 v = *reinterpret_cast<uint2 const*>(src);
+  auto const* p =
+      reinterpret_cast<typename Converter::packed_hip_type const*>(&v);
+#pragma unroll
+  for (int i = 0; i < kElemsPerLane / 2; i++) {
+    float2 f2 = Converter::convert(p[i]);
+    elems[2 * i] = f2.x;
+    elems[2 * i + 1] = f2.y;
+  }
+}
+
+// Store 4 fp32 registers -> 4 contiguous bf16.
+template <typename scalar_t>
+__device__ __forceinline__ void storeElems(
+    scalar_t* __restrict__ dst, float const (&elems)[kElemsPerLane]) {
+  using Converter = vllm::_typeConvert<scalar_t>;
+  uint2 v;
+  auto* p = reinterpret_cast<typename Converter::packed_hip_type*>(&v);
+#pragma unroll
+  for (int i = 0; i < kElemsPerLane / 2; i++) {
+    p[i] = Converter::convert(make_float2(elems[2 * i], elems[2 * i + 1]));
+  }
+  *reinterpret_cast<uint2*>(dst) = v;
+}
+
+// ────────────────────────────────────────────────────────────────────────────
+// Kernel
+// ────────────────────────────────────────────────────────────────────────────
+// Grid: 1D, ceil(num_tokens * slots_per_token / warps_per_block).
+// Each warp = one (token, slot).
+//
+// `kIsSparse` and `kInsertKV` are compile-time template bools, so all the
+// branch decisions that distinguish the dense layer from the sparse layer
+// (index slots, KV/index inserts, V slots) fold away per instantiation.
+// Three instantiations are built: dense <false,false>, sparse-profiling
+// <true,false> and sparse-serving <true,true>.  Slots per token:
+//     Q : nq                            (always — norm+RoPE)
+//     K : nkv                           (always — norm+RoPE; +K-cache insert)
+//     V : nkv  only if kInsertKV        (V-cache insert; no warps in dense)
+//     IQ: niq  only if kIsSparse        (norm+RoPE)
+//     IK: 1    only if kIsSparse        (norm+RoPE; +index-cache insert)
+template <typename scalar_t, bool kIsSparse, bool kInsertKV>
+__global__ void fusedMiniMaxM3QNormRopeKVInsertKernel(
+    scalar_t* __restrict__ qkv,  // [N, qkv_row] in/out (packs index if sparse)
+    scalar_t* __restrict__ q_out,        // [N, nq*128] contiguous, or nullptr
+    scalar_t* __restrict__ index_q_out,  // [N, niq*128] contiguous, or nullptr
+    scalar_t const* __restrict__ q_norm_w,
+    scalar_t const* __restrict__ k_norm_w,
+    scalar_t const* __restrict__ iq_norm_w,
+    scalar_t const* __restrict__ ik_norm_w,
+    scalar_t const* __restrict__ cos_sin_cache,  // [max_pos, rotary_dim]
+    int64_t const* __restrict__ positions,       // [N] i64
+    int64_t const* __restrict__ slot_mapping,    // main K/V slots or nullptr
+    int64_t const* __restrict__ index_slot_mapping,  // index K slots/nullptr
+    scalar_t* __restrict__ kv_cache,     // [nb,2,bs,nkv,128] or nullptr
+    scalar_t* __restrict__ index_cache,  // [nb*bs, 128] or nullptr
+    float const eps, int const rotary_dim, int const num_tokens, int const nq,
+    int const nkv, int const niq, int const block_size,
+    // kv_cache strides (in elements) for logical shape [nb, 2, bs, nkv, 128].
+    // The head_dim (last) dim is always innermost-contiguous (stride 1), so the
+    // NHD/HND layout choice is fully captured by these four strides: NHD keeps
+    // s_token < s_head, HND swaps them. dim_base addresses head_dim directly.
+    int64_t const kv_s_block, int64_t const kv_s_kv, int64_t const kv_s_token,
+    int64_t const kv_s_head) {
+#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
+  // _typeConvert<BFloat16> is unavailable on pre-Ampere; the M3 kernel only
+  // runs with bf16/fp16 inputs in practice.  Discard the bf16 body there.
+  if constexpr (std::is_same_v<scalar_t, c10::BFloat16>) {
+    return;
+  } else {
+#endif
+    int const warpsPerBlock = blockDim.x / 32;
+    int const laneId = threadIdx.x % 32;
+    int const globalWarpIdx = blockIdx.x * warpsPerBlock + (threadIdx.x / 32);
+
+    // Slot layout (compile-time gated: dense has neither V nor index slots).
+    int const v_slots = kInsertKV ? nkv : 0;
+    int const idx_slots = kIsSparse ? niq + 1 : 0;
+    int const slots_per_token = nq + nkv + v_slots + idx_slots;
+
+    int const tokenIdx = globalWarpIdx / slots_per_token;
+    int const slot = globalWarpIdx % slots_per_token;
+    if (tokenIdx >= num_tokens) return;
+
+    // Slot boundaries.
+    int const k_begin = nq;
+    int const v_begin = nq + nkv;             // valid only when kInsertKV
+    int const iq_begin = nq + nkv + v_slots;  // index block start
+    int const ik_slot = iq_begin + niq;       // valid only when kIsSparse
+
+    bool const isQ = slot < k_begin;
+    bool const isK = slot >= k_begin && slot < v_begin;
+    bool isV = false;
+    if constexpr (kInsertKV) isV = slot >= v_begin && slot < v_begin + nkv;
+    bool isIQ = false, isIK = false;
+    if constexpr (kIsSparse) {
+      isIQ = slot >= iq_begin && slot < ik_slot;
+      isIK = slot == ik_slot;
+    }
+
+    int const dim_base = laneId * kElemsPerLane;
+    // Physical row width of qkv: the dense layer packs [q|k|v]; the sparse
+    // layer additionally packs [index_q (niq heads) | index_k (1 head)].
+    int const qkv_row = (nq + 2 * nkv + (kIsSparse ? (niq + 1) : 0)) * kHeadDim;
+
+    // ── Resolve source pointer + per-branch parameters. ────────────────────
+    scalar_t* row_ptr = nullptr;       // in-place output location
+    scalar_t const* norm_w = nullptr;  // nullptr -> skip norm (V)
+    bool do_rope = true;
+    int head = 0;  // kv head index for inserts
+
+    if (isQ) {
+      row_ptr =
+          qkv + static_cast<int64_t>(tokenIdx) * qkv_row + slot * kHeadDim;
+      norm_w = q_norm_w;
+    } else if (isK) {
+      head = slot - k_begin;
+      row_ptr =
+          qkv + static_cast<int64_t>(tokenIdx) * qkv_row + slot * kHeadDim;
+      norm_w = k_norm_w;
+    } else if (isV) {
+      // qkv V section starts at slot index (nq + nkv): slot * kHeadDim is the
+      // correct in-tensor offset.
+      head = slot - v_begin;
+      row_ptr =
+          qkv + static_cast<int64_t>(tokenIdx) * qkv_row + slot * kHeadDim;
+      norm_w = nullptr;  // V: no norm, no rope
+      do_rope = false;
+    } else if (isIQ) {
+      // index_q sub-block lives at physical offset (nq+2*nkv)*128 in qkv.
+      int const ih = slot - iq_begin;
+      row_ptr = qkv + static_cast<int64_t>(tokenIdx) * qkv_row +
+                (nq + 2 * nkv + ih) * kHeadDim;
+      norm_w = iq_norm_w;
+    } else {  // isIK -- single shared index key at (nq+2*nkv+niq)*128.
+      row_ptr = qkv + static_cast<int64_t>(tokenIdx) * qkv_row +
+                (nq + 2 * nkv + niq) * kHeadDim;
+      norm_w = ik_norm_w;
+    }
+
+    // Store destination.  Q and index_q are gathered into dedicated contiguous
+    // output buffers (when provided) so the downstream SM100 sparse kernel's
+    // flat TMA descriptor can address them as [tokens*heads, head_dim]; this
+    // folds the de-interleaving into the store the kernel already does, instead
+    // of a separate q.contiguous() copy.  Everything else stays in place.
+    scalar_t* store_ptr = row_ptr;
+    if (isQ && q_out != nullptr) {
+      store_ptr = q_out + static_cast<int64_t>(tokenIdx) * nq * kHeadDim +
+                  slot * kHeadDim;
+    } else if (isIQ && index_q_out != nullptr) {
+      store_ptr = index_q_out +
+                  static_cast<int64_t>(tokenIdx) * niq * kHeadDim +
+                  (slot - iq_begin) * kHeadDim;
+    }
+
+    // PDL: wait for the predecessor kernel (the qkv-projection GEMM that
+    // produces ``qkv``) to finish before touching any global memory.  No-op
+    // when PDL is not enabled on the launch.  The CUDA runtime wrapper emits
+    // the griddepcontrol.wait PTX with the required memory clobber internally.
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
+    cudaGridDependencySynchronize();
+#endif
+
+    // ── Load -> norm+rope (fp32) -> store back in place. ───────────────────
+    float elems[kElemsPerLane];
+    loadElems<scalar_t>(row_ptr + dim_base, elems);
+
+    if (!isV) {
+      int64_t const pos = positions[tokenIdx];
+      scalar_t const* cos_ptr = cos_sin_cache + pos * rotary_dim;
+      normAndRope<scalar_t>(elems, laneId, eps, norm_w, do_rope, rotary_dim,
+                            cos_ptr, /*apply_norm=*/norm_w != nullptr);
+      storeElems<scalar_t>(store_ptr + dim_base, elems);
+    }
+
+    // ── Cache inserts (sparse serving only). ───────────────────────────────
+    if constexpr (kInsertKV) {
+      // Guard (not early-return) so every thread reaches the PDL trigger below.
+      int64_t const sm = (isK || isV)
+                             ? slot_mapping[tokenIdx]
+                             : (isIK ? index_slot_mapping[tokenIdx] : -1);
+      if (sm >= 0) {  // skip padded / unscheduled tokens
+        if (isIK) {
+          scalar_t* dst = index_cache + sm * kHeadDim + dim_base;
+          storeElems<scalar_t>(dst, elems);
+        } else if (isK || isV) {
+          // kv_cache logical shape [num_blocks, 2, block_size, nkv, head_dim].
+          // Paging is logical (block = sm/block_size, token = sm%block_size);
+          // the physical NHD/HND layout is honoured via the passed strides.
+          int64_t const b = sm / block_size;
+          int64_t const t = sm % block_size;
+          int const kv = isK ? 0 : 1;
+          int64_t const off =
+              b * kv_s_block + kv * kv_s_kv + t * kv_s_token + head * kv_s_head;
+          storeElems<scalar_t>(kv_cache + off + dim_base, elems);
+        }
+      }
+    }
+
+    // PDL: signal that this kernel is done so a dependent successor may launch
+    // early.  No-op when PDL is not enabled on the launch.
+#if defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)
+    cudaTriggerProgrammaticLaunchCompletion();
+#endif
+#if (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ < 800) && !defined(USE_ROCM)
+  }
+#endif
+}
+
+// ────────────────────────────────────────────────────────────────────────────
+// Launch wrapper
+// ────────────────────────────────────────────────────────────────────────────
+template <typename scalar_t>
+void launchFusedMiniMaxM3(scalar_t* qkv, scalar_t* q_out, scalar_t* index_q_out,
+                          scalar_t const* q_norm_w, scalar_t const* k_norm_w,
+                          scalar_t const* iq_norm_w, scalar_t const* ik_norm_w,
+                          scalar_t const* cos_sin_cache,
+                          int64_t const* positions, int64_t const* slot_mapping,
+                          int64_t const* index_slot_mapping, scalar_t* kv_cache,
+                          scalar_t* index_cache, float const eps,
+                          int const rotary_dim, int const num_tokens,
+                          int const nq, int const nkv, int const niq,
+                          int const block_size, int64_t const kv_s_block,
+                          int64_t const kv_s_kv, int64_t const kv_s_token,
+                          int64_t const kv_s_head, bool const has_index,
+                          bool const insert_kv, cudaStream_t stream) {
+  // Slot count must match the kernel's compile-time gating.
+  int const v_slots = insert_kv ? nkv : 0;
+  int const idx_slots = has_index ? niq + 1 : 0;
+  int const slots_per_token = nq + nkv + v_slots + idx_slots;
+
+  constexpr int kBlockSize = 256;
+  constexpr int kWarpsPerBlock = kBlockSize / 32;
+  int64_t const total_warps =
+      static_cast<int64_t>(num_tokens) * slots_per_token;
+  int const grid =
+      static_cast<int>((total_warps + kWarpsPerBlock - 1) / kWarpsPerBlock);
+  if (grid == 0) return;
+
+#ifndef USE_ROCM
+  // PDL: enable programmatic stream serialization whenever the hardware
+  // supports it (SM90+).  On pre-Hopper GPUs the attribute is unavailable, so
+  // leave numAttrs = 0 and launch as a regular kernel via cudaLaunchKernelEx.
+  static int const sm_version = getSMVersion();
+  cudaLaunchConfig_t config;
+  config.gridDim = dim3(grid);
+  config.blockDim = dim3(kBlockSize);
+  config.dynamicSmemBytes = 0;
+  config.stream = stream;
+  cudaLaunchAttribute attrs[1];
+  attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
+  attrs[0].val.programmaticStreamSerializationAllowed = 1;
+  config.attrs = attrs;
+  config.numAttrs = (sm_version >= 90) ? 1 : 0;
+
+  #define LAUNCH(IS_SPARSE, INSERT)                                           \
+    cudaLaunchKernelEx(                                                       \
+        &config,                                                              \
+        fusedMiniMaxM3QNormRopeKVInsertKernel<scalar_t, IS_SPARSE, INSERT>,   \
+        qkv, q_out, index_q_out, q_norm_w, k_norm_w, iq_norm_w, ik_norm_w,    \
+        cos_sin_cache, positions, slot_mapping, index_slot_mapping, kv_cache, \
+        index_cache, eps, rotary_dim, num_tokens, nq, nkv, niq, block_size,   \
+        kv_s_block, kv_s_kv, kv_s_token, kv_s_head)
+#else
+  // ROCm: standard kernel launch syntax (no PDL/stream serialization).
+  // clang-format off
+  #define LAUNCH(IS_SPARSE, INSERT)                                          \
+    fusedMiniMaxM3QNormRopeKVInsertKernel<scalar_t, IS_SPARSE, INSERT>       \
+        <<<grid, kBlockSize, 0, stream>>>(                                   \
+            qkv, q_out, index_q_out, q_norm_w, k_norm_w, iq_norm_w,          \
+            ik_norm_w, cos_sin_cache, positions, slot_mapping,               \
+            index_slot_mapping, kv_cache, index_cache, eps, rotary_dim,      \
+            num_tokens, nq, nkv, niq, block_size, kv_s_block, kv_s_kv,       \
+            kv_s_token, kv_s_head)
+  // clang-format on
+#endif
+
+  if (has_index) {
+    if (insert_kv) {
+      LAUNCH(true, true);  // sparse serving
+    } else {
+      LAUNCH(true, false);  // sparse profiling
+    }
+  } else {
+    // Dense layer: never has an index branch and never inserts here (the
+    // generic Attention layer owns the KV insert).
+    LAUNCH(false, false);
+  }
+#undef LAUNCH
+}
+
+}  // namespace minimax_m3_fused_ops
+}  // namespace vllm
+
+// ────────────────────────────────────────────────────────────────────────────
+// Torch op wrapper
+// ────────────────────────────────────────────────────────────────────────────
+void fused_minimax_m3_qknorm_rope_kv_insert(
+    torch::stable::Tensor& qkv,  // [N, qkv_row] (packs index if sparse)
+    torch::stable::Tensor const& q_norm_weight,  // [128]
+    torch::stable::Tensor const& k_norm_weight,  // [128]
+    torch::stable::Tensor const& cos_sin_cache,  // [max_pos, rotary_dim]
+    torch::stable::Tensor const& positions,      // [N] i64
+    int64_t num_heads, int64_t num_kv_heads, int64_t rotary_dim, double eps,
+    std::optional<torch::stable::Tensor> index_q_norm_weight,  // [128]
+    std::optional<torch::stable::Tensor> index_k_norm_weight,  // [128]
+    int64_t num_index_heads,                                  // niq; 0 => dense
+    std::optional<torch::stable::Tensor> slot_mapping,        // [N] i64
+    std::optional<torch::stable::Tensor> index_slot_mapping,  // [N] i64
+    std::optional<torch::stable::Tensor> kv_cache,     // [nb,2,bs,nkv,128]
+    std::optional<torch::stable::Tensor> index_cache,  // [nb,bs,128]
+    int64_t block_size,
+    std::optional<torch::stable::Tensor> q_out,  // [N, nq*128] contiguous
+    std::optional<torch::stable::Tensor>
+        index_q_out) {  // [N, niq*128] contiguous
+  STD_TORCH_CHECK(qkv.is_cuda() && qkv.is_contiguous(),
+                  "qkv must be contiguous CUDA");
+  STD_TORCH_CHECK(
+      positions.is_cuda() &&
+          positions.scalar_type() == torch::headeronly::ScalarType::Long,
+      "positions must be int64 CUDA");
+  STD_TORCH_CHECK(cos_sin_cache.is_cuda() && cos_sin_cache.is_contiguous(),
+                  "cos_sin_cache must be contiguous CUDA");
+  STD_TORCH_CHECK(cos_sin_cache.scalar_type() == qkv.scalar_type(),
+                  "cos_sin_cache dtype must match qkv");
+  STD_TORCH_CHECK(
+      cos_sin_cache.dim() == 2 && cos_sin_cache.size(1) == rotary_dim,
+      "cos_sin_cache shape [max_pos, rotary_dim]");
+
+  STD_TORCH_CHECK(q_norm_weight.scalar_type() == qkv.scalar_type() &&
+                      k_norm_weight.scalar_type() == qkv.scalar_type(),
+                  "q/k norm weight dtype must match qkv");
+  STD_TORCH_CHECK(
+      q_norm_weight.numel() == vllm::minimax_m3_fused_ops::kHeadDim &&
+          k_norm_weight.numel() == vllm::minimax_m3_fused_ops::kHeadDim,
+      "q/k norm weight must have 128 elements");
+  STD_TORCH_CHECK(rotary_dim > 0 && rotary_dim % 8 == 0 &&
+                      rotary_dim <= vllm::minimax_m3_fused_ops::kHeadDim,
+                  "rotary_dim must be a positive multiple of 8 and <= 128");
+
+  int const num_tokens = static_cast<int>(qkv.size(0));
+  int const nq = static_cast<int>(num_heads);
+  int const nkv = static_cast<int>(num_kv_heads);
+  int const niq = static_cast<int>(num_index_heads);
+
+  // The sparse layer packs the index branch ([index_q (niq heads) | index_k
+  // (1 head)]) right after [q|k|v] in the same row; the dense layer does not.
+  bool const has_index = niq > 0;
+  bool const insert_kv = kv_cache.has_value();
+  int const kHeadDim = vllm::minimax_m3_fused_ops::kHeadDim;
+  int const expected_row =
+      (nq + 2 * nkv + (has_index ? niq + 1 : 0)) * kHeadDim;
+  STD_TORCH_CHECK(qkv.size(1) == expected_row,
+                  "qkv last dim must be (num_heads + 2*num_kv_heads"
+                  " + num_index_heads + 1) * 128 for sparse, "
+                  "(num_heads + 2*num_kv_heads) * 128 for dense");
+
+  // Only the sparse layer inserts here (dense lets the generic Attention layer
+  // own the KV write); there is no dense+insert kernel instantiation.
+  STD_TORCH_CHECK(
+      !insert_kv || has_index,
+      "insert mode (kv_cache) requires the index branch (sparse layer)");
+  if (has_index) {
+    STD_TORCH_CHECK(
+        index_q_norm_weight.has_value() && index_k_norm_weight.has_value(),
+        "index branch requires both index norm weights");
+    STD_TORCH_CHECK(index_q_norm_weight->scalar_type() == qkv.scalar_type() &&
+                        index_k_norm_weight->scalar_type() == qkv.scalar_type(),
+                    "index norm weights dtype must match qkv");
+    STD_TORCH_CHECK(index_q_norm_weight->numel() == kHeadDim &&
+                        index_k_norm_weight->numel() == kHeadDim,
+                    "index norm weights must have 128 elements");
+  }
+  // kv_cache strides (logical shape [nb, 2, bs, nkv, head_dim]). Read straight
+  // off the tensor so the kernel honours whatever physical layout the attention
+  // backend allocated (NHD: stride order (0,1,2,3,4); HND: (0,1,3,2,4)). No new
+  // op argument is needed -- the strides ride along with the tensor itself.
+  int64_t kv_s_block = 0, kv_s_kv = 0, kv_s_token = 0, kv_s_head = 0;
+  torch::stable::Tensor const* effective_index_slot_mapping = nullptr;
+  if (insert_kv) {
+    STD_TORCH_CHECK(
+        slot_mapping.has_value() && slot_mapping->is_cuda() &&
+            slot_mapping->scalar_type() == torch::headeronly::ScalarType::Long,
+        "insert mode requires int64 CUDA slot_mapping");
+    STD_TORCH_CHECK(
+        !index_slot_mapping.has_value() ||
+            (index_slot_mapping->is_cuda() &&
+             index_slot_mapping->scalar_type() ==
+                 torch::headeronly::ScalarType::Long &&
+             index_slot_mapping->numel() == slot_mapping->numel()),
+        "index_slot_mapping must be int64 CUDA with slot_mapping length");
+    STD_TORCH_CHECK(kv_cache->scalar_type() == qkv.scalar_type(),
+                    "kv_cache dtype must match qkv (bf16 cache only)");
+    STD_TORCH_CHECK(index_cache.has_value() &&
+                        index_cache->scalar_type() == qkv.scalar_type(),
+                    "insert mode requires matching index_cache");
+    STD_TORCH_CHECK(kv_cache->dim() == 5 && kv_cache->stride(4) == 1,
+                    "kv_cache must be [nb,2,bs,nkv,head_dim] with contiguous "
+                    "head_dim (stride(4)==1)");
+    kv_s_block = kv_cache->stride(0);
+    kv_s_kv = kv_cache->stride(1);
+    kv_s_token = kv_cache->stride(2);
+    kv_s_head = kv_cache->stride(3);
+    effective_index_slot_mapping = index_slot_mapping.has_value()
+                                       ? &index_slot_mapping.value()
+                                       : &slot_mapping.value();
+  }
+  // Optional contiguous gather targets: when given, the normed/roped q (and
+  // index_q) are written here instead of in place, so callers avoid a separate
+  // .contiguous() copy.  index_q_out only makes sense on the sparse path.
+  if (q_out.has_value()) {
+    STD_TORCH_CHECK(
+        q_out->is_cuda() && q_out->is_contiguous() &&
+            q_out->scalar_type() == qkv.scalar_type(),
+        "q_out must be a contiguous CUDA tensor matching qkv dtype");
+    STD_TORCH_CHECK(
+        q_out->numel() == static_cast<int64_t>(num_tokens) * nq * kHeadDim,
+        "q_out must have num_tokens * num_heads * 128 elements");
+  }
+  if (index_q_out.has_value()) {
+    STD_TORCH_CHECK(
+        has_index,
+        "index_q_out requires the index branch (num_index_heads > 0)");
+    STD_TORCH_CHECK(
+        index_q_out->is_cuda() && index_q_out->is_contiguous() &&
+            index_q_out->scalar_type() == qkv.scalar_type(),
+        "index_q_out must be a contiguous CUDA tensor matching qkv dtype");
+    STD_TORCH_CHECK(index_q_out->numel() ==
+                        static_cast<int64_t>(num_tokens) * niq * kHeadDim,
+                    "index_q_out must have num_tokens * num_index_heads * 128 "
+                    "elements");
+  }
+
+  const torch::stable::accelerator::DeviceGuard device_guard(
+      qkv.get_device_index());
+  auto stream = get_current_cuda_stream(qkv.get_device_index());
+
+  VLLM_STABLE_DISPATCH_HALF_TYPES(
+      qkv.scalar_type(), "fused_minimax_m3_qknorm_rope_kv_insert", [&] {
+        using st = scalar_t;
+        vllm::minimax_m3_fused_ops::launchFusedMiniMaxM3<st>(
+            reinterpret_cast<st*>(qkv.data_ptr()),
+            q_out.has_value() ? reinterpret_cast<st*>(q_out->data_ptr())
+                              : nullptr,
+            index_q_out.has_value()
+                ? reinterpret_cast<st*>(index_q_out->data_ptr())
+                : nullptr,
+            reinterpret_cast<st const*>(q_norm_weight.data_ptr()),
+            reinterpret_cast<st const*>(k_norm_weight.data_ptr()),
+            has_index
+                ? reinterpret_cast<st const*>(index_q_norm_weight->data_ptr())
+                : nullptr,
+            has_index
+                ? reinterpret_cast<st const*>(index_k_norm_weight->data_ptr())
+                : nullptr,
+            reinterpret_cast<st const*>(cos_sin_cache.data_ptr()),
+            reinterpret_cast<int64_t const*>(positions.data_ptr()),
+            insert_kv
+                ? reinterpret_cast<int64_t const*>(slot_mapping->data_ptr())
+                : nullptr,
+            insert_kv ? reinterpret_cast<int64_t const*>(
+                            effective_index_slot_mapping->data_ptr())
+                      : nullptr,
+            insert_kv ? reinterpret_cast<st*>(kv_cache->data_ptr()) : nullptr,
+            (insert_kv && has_index)
+                ? reinterpret_cast<st*>(index_cache->data_ptr())
+                : nullptr,
+            static_cast<float>(eps), static_cast<int>(rotary_dim), num_tokens,
+            nq, nkv, niq, static_cast<int>(block_size), kv_s_block, kv_s_kv,
+            kv_s_token, kv_s_head, has_index, insert_kv, stream);
+      });
+}

csrc/libtorch_stable/moe/dsv3_router_gemm_bf16_out.cu

@@ -18,14 +18,11 @@
  * limitations under the License.
  */
 
-#include <ATen/ATen.h>
-#include <ATen/cuda/CUDAContext.h>
+#include <cstdint>
 
 #include <cuda_bf16.h>
 #include <cuda_runtime.h>
 
-#include "dsv3_router_gemm_utils.h"
-
 // Custom FMA implementation using PTX assembly instructions
 __device__ __forceinline__ void fma(float2& d, float2 const& a, float2 const& b,
                                     float2 const& c) {

csrc/libtorch_stable/moe/dsv3_router_gemm_entry.cu

@@ -18,15 +18,25 @@
  * limitations under the License.
  */
 
-#include <ATen/ATen.h>
-#include <ATen/cuda/CUDAContext.h>
-#include <torch/all.h>
+#include <torch/csrc/stable/library.h>
+#include <torch/csrc/stable/tensor.h>
+#include <torch/headeronly/core/ScalarType.h>
+
+#include "libtorch_stable/torch_utils.h"
 
 #include <cuda_bf16.h>
 #include <cuda_runtime.h>
 
-#include "core/registration.h"
-#include "dsv3_router_gemm_utils.h"
+#include <stdexcept>
+
+namespace {
+
+inline int getSMVersion() {
+  auto* props = get_device_prop();
+  return props->major * 10 + props->minor;
+}
+
+}  // namespace
 
 static constexpr int DEFAULT_NUM_EXPERTS = 256;
 static constexpr int KIMI_K2_NUM_EXPERTS = 384;
@@ -98,40 +108,47 @@ struct LoopUnroller<kEnd, kEnd, kNumExperts, kHiddenDim> {
   }
 };
 
-void dsv3_router_gemm(at::Tensor& output,       // [num_tokens, num_experts]
-                      const at::Tensor& mat_a,  // [num_tokens, hidden_dim]
-                      const at::Tensor& mat_b   // [num_experts, hidden_dim]
+void dsv3_router_gemm(
+    torch::stable::Tensor& output,       // [num_tokens, num_experts]
+    torch::stable::Tensor const& mat_a,  // [num_tokens, hidden_dim]
+    torch::stable::Tensor const& mat_b   // [num_experts, hidden_dim]
 ) {
-  TORCH_CHECK(output.dim() == 2 && mat_a.dim() == 2 && mat_b.dim() == 2);
+  STD_TORCH_CHECK(output.dim() == 2 && mat_a.dim() == 2 && mat_b.dim() == 2);
 
   const int num_tokens = mat_a.size(0);
   const int num_experts = mat_b.size(0);
   const int hidden_dim = mat_a.size(1);
 
-  TORCH_CHECK(mat_a.size(1) == mat_b.size(1),
-              "mat_a and mat_b must have the same hidden_dim");
-  TORCH_CHECK(hidden_dim == DEFAULT_HIDDEN_DIM,
-              "Expected hidden_dim=", DEFAULT_HIDDEN_DIM,
-              ", but got hidden_dim=", hidden_dim);
-  TORCH_CHECK(
+  STD_TORCH_CHECK(mat_a.size(1) == mat_b.size(1),
+                  "mat_a and mat_b must have the same hidden_dim");
+  STD_TORCH_CHECK(hidden_dim == DEFAULT_HIDDEN_DIM,
+                  "Expected hidden_dim=", DEFAULT_HIDDEN_DIM,
+                  ", but got hidden_dim=", hidden_dim);
+  STD_TORCH_CHECK(
       num_experts == DEFAULT_NUM_EXPERTS || num_experts == KIMI_K2_NUM_EXPERTS,
       "Expected num_experts=", DEFAULT_NUM_EXPERTS,
       " or num_experts=", KIMI_K2_NUM_EXPERTS,
       ", but got num_experts=", num_experts);
-  TORCH_CHECK(num_tokens >= 1 && num_tokens <= 16,
-              "currently num_tokens must be less than or equal to 16 for "
-              "router_gemm");
-  TORCH_CHECK(mat_a.dtype() == at::kBFloat16, "mat_a must be bf16");
-  TORCH_CHECK(mat_b.dtype() == at::kBFloat16, "mat_b must be bf16");
-  TORCH_CHECK(output.dtype() == at::kFloat || output.dtype() == at::kBFloat16,
-              "output must be float32 or bf16");
+  STD_TORCH_CHECK(num_tokens >= 1 && num_tokens <= 16,
+                  "currently num_tokens must be less than or equal to 16 for "
+                  "router_gemm");
+  STD_TORCH_CHECK(
+      mat_a.scalar_type() == torch::headeronly::ScalarType::BFloat16,
+      "mat_a must be bf16");
+  STD_TORCH_CHECK(
+      mat_b.scalar_type() == torch::headeronly::ScalarType::BFloat16,
+      "mat_b must be bf16");
+  STD_TORCH_CHECK(
+      output.scalar_type() == torch::headeronly::ScalarType::Float ||
+          output.scalar_type() == torch::headeronly::ScalarType::BFloat16,
+      "output must be float32 or bf16");
 
-  auto const sm = getSMVersion();
-  TORCH_CHECK(sm >= 90 && sm <= 103, "required SM_103 >= CUDA ARCH >= SM_90");
+  const int sm = getSMVersion();
+  STD_TORCH_CHECK(sm >= 90, "required CUDA ARCH >= SM_90");
 
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  const cudaStream_t stream = get_current_cuda_stream(mat_a.get_device_index());
 
-  if (output.dtype() == at::kFloat) {
+  if (output.scalar_type() == torch::headeronly::ScalarType::Float) {
     if (num_experts == DEFAULT_NUM_EXPERTS) {
       LoopUnroller<1, 16, DEFAULT_NUM_EXPERTS, DEFAULT_HIDDEN_DIM>::
           unroll_float_output(
@@ -145,7 +162,7 @@ void dsv3_router_gemm(at::Tensor& output,       // [num_tokens, num_experts]
               reinterpret_cast<__nv_bfloat16 const*>(mat_a.data_ptr()),
               reinterpret_cast<__nv_bfloat16 const*>(mat_b.data_ptr()), stream);
     }
-  } else if (output.dtype() == at::kBFloat16) {
+  } else if (output.scalar_type() == torch::headeronly::ScalarType::BFloat16) {
     if (num_experts == DEFAULT_NUM_EXPERTS) {
       LoopUnroller<1, 16, DEFAULT_NUM_EXPERTS, DEFAULT_HIDDEN_DIM>::
           unroll_bf16_output(
@@ -164,6 +181,6 @@ void dsv3_router_gemm(at::Tensor& output,       // [num_tokens, num_experts]
   }
 }
 
-TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
-  m.impl("dsv3_router_gemm", &dsv3_router_gemm);
+STABLE_TORCH_LIBRARY_IMPL(_moe_C, CUDA, m) {
+  m.impl("dsv3_router_gemm", TORCH_BOX(&dsv3_router_gemm));
 }

csrc/libtorch_stable/moe/dsv3_router_gemm_float_out.cu

@@ -18,14 +18,11 @@
  * limitations under the License.
  */
 
-#include <ATen/ATen.h>
-#include <ATen/cuda/CUDAContext.h>
+#include <cstdint>
 
 #include <cuda_bf16.h>
 #include <cuda_runtime.h>
 
-#include "dsv3_router_gemm_utils.h"
-
 // Custom FMA implementation using PTX assembly instructions
 __device__ __forceinline__ void fma(float2& d, float2 const& a, float2 const& b,
                                     float2 const& c) {

csrc/libtorch_stable/moe/grouped_topk_kernels.cu

@@ -18,9 +18,14 @@
  * limitations under the License.
  */
 #include "moeTopKFuncs.cuh"
-#include <c10/cuda/CUDAStream.h>
-#include <torch/all.h>
+
+#include <torch/csrc/stable/tensor.h>
+#include <torch/headeronly/core/ScalarType.h>
+
+#include "libtorch_stable/torch_utils.h"
+
 #include <cmath>
+#include <tuple>
 #include <cuda_fp16.h>
 #include <cuda_bf16.h>
 #include <cuda/std/limits>
@@ -1001,38 +1006,40 @@ INSTANTIATE_NOAUX_TC(__nv_bfloat16, __nv_bfloat16, int32_t, SCORING_NONE);
 }  // end namespace moe
 }  // namespace vllm
 
-std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
-    torch::Tensor const& scores, int64_t n_group, int64_t topk_group,
+std::tuple<torch::stable::Tensor, torch::stable::Tensor> grouped_topk(
+    torch::stable::Tensor const& scores, int64_t n_group, int64_t topk_group,
     int64_t topk, bool renormalize, double routed_scaling_factor,
-    torch::Tensor const& bias, int64_t scoring_func = 0) {
-  auto data_type = scores.scalar_type();
-  auto bias_type = bias.scalar_type();
-  auto input_size = scores.sizes();
-  int64_t num_tokens = input_size[0];
-  int64_t num_experts = input_size[1];
-  TORCH_CHECK(input_size.size() == 2, "scores must be a 2D Tensor");
-  TORCH_CHECK(n_group > 0, "n_group must be positive");
-  TORCH_CHECK(topk > 0, "topk must be positive");
-  TORCH_CHECK(topk_group > 0, "topk_group must be positive");
-  TORCH_CHECK(topk_group <= n_group, "topk_group must be <= n_group");
-  TORCH_CHECK(num_experts % n_group == 0,
-              "num_experts should be divisible by n_group");
-  TORCH_CHECK(n_group <= 32,
-              "n_group should be smaller than or equal to 32 for now");
-  TORCH_CHECK(topk <= 32, "topk should be smaller than or equal to 32 for now");
-  TORCH_CHECK(topk <= topk_group * (num_experts / n_group),
-              "topk must be <= topk_group * (num_experts / n_group)");
-  TORCH_CHECK(scoring_func == vllm::moe::SCORING_NONE ||
-                  scoring_func == vllm::moe::SCORING_SIGMOID,
-              "scoring_func must be SCORING_NONE (0) or SCORING_SIGMOID (1)");
+    torch::stable::Tensor const& bias, int64_t scoring_func = 0) {
+  const auto data_type = scores.scalar_type();
+  const auto bias_type = bias.scalar_type();
+  STD_TORCH_CHECK(scores.dim() == 2, "scores must be a 2D Tensor");
+  const int64_t num_tokens = scores.size(0);
+  const int64_t num_experts = scores.size(1);
+  STD_TORCH_CHECK(n_group > 0, "n_group must be positive");
+  STD_TORCH_CHECK(topk > 0, "topk must be positive");
+  STD_TORCH_CHECK(topk_group > 0, "topk_group must be positive");
+  STD_TORCH_CHECK(topk_group <= n_group, "topk_group must be <= n_group");
+  STD_TORCH_CHECK(num_experts % n_group == 0,
+                  "num_experts should be divisible by n_group");
+  STD_TORCH_CHECK(n_group <= 32,
+                  "n_group should be smaller than or equal to 32 for now");
+  STD_TORCH_CHECK(topk <= 32,
+                  "topk should be smaller than or equal to 32 for now");
+  STD_TORCH_CHECK(topk <= topk_group * (num_experts / n_group),
+                  "topk must be <= topk_group * (num_experts / n_group)");
+  STD_TORCH_CHECK(
+      scoring_func == vllm::moe::SCORING_NONE ||
+          scoring_func == vllm::moe::SCORING_SIGMOID,
+      "scoring_func must be SCORING_NONE (0) or SCORING_SIGMOID (1)");
 
   // Always output float32 for topk_values (eliminates Python-side conversion)
-  torch::Tensor topk_values = torch::empty(
-      {num_tokens, topk}, torch::dtype(torch::kFloat32).device(torch::kCUDA));
-  torch::Tensor topk_indices = torch::empty(
-      {num_tokens, topk}, torch::dtype(torch::kInt32).device(torch::kCUDA));
+  auto topk_values = torch::stable::new_empty(
+      scores, {num_tokens, topk}, torch::headeronly::ScalarType::Float);
+  auto topk_indices = torch::stable::new_empty(
+      scores, {num_tokens, topk}, torch::headeronly::ScalarType::Int);
 
-  auto stream = c10::cuda::getCurrentCUDAStream(scores.get_device());
+  const cudaStream_t stream =
+      get_current_cuda_stream(scores.get_device_index());
   auto const sf = static_cast<vllm::moe::ScoringFunc>(scoring_func);
 
 #define LAUNCH_KERNEL_SF(T, BiasT, IdxT)                                      \
@@ -1057,7 +1064,7 @@ std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
             routed_scaling_factor, false, stream);                            \
         break;                                                                \
       default:                                                                \
-        throw std::invalid_argument("Unsupported scoring_func");              \
+        STD_TORCH_CHECK(false, "Unsupported scoring_func");                   \
         break;                                                                \
     }                                                                         \
   } while (0)
@@ -1065,17 +1072,18 @@ std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
 #define LAUNCH_KERNEL(T, IdxT)                                         \
   do {                                                                 \
     switch (bias_type) {                                               \
-      case torch::kFloat16:                                            \
+      case torch::headeronly::ScalarType::Half:                        \
         LAUNCH_KERNEL_SF(T, half, IdxT);                               \
         break;                                                         \
-      case torch::kFloat32:                                            \
+      case torch::headeronly::ScalarType::Float:                       \
         LAUNCH_KERNEL_SF(T, float, IdxT);                              \
         break;                                                         \
-      case torch::kBFloat16:                                           \
+      case torch::headeronly::ScalarType::BFloat16:                    \
         LAUNCH_KERNEL_SF(T, __nv_bfloat16, IdxT);                      \
         break;                                                         \
       default:                                                         \
-        throw std::invalid_argument(                                   \
+        STD_TORCH_CHECK(                                               \
+            false,                                                     \
             "Invalid bias dtype, only supports float16, float32, and " \
             "bfloat16");                                               \
         break;                                                         \
@@ -1083,22 +1091,22 @@ std::tuple<torch::Tensor, torch::Tensor> grouped_topk(
   } while (0)
 
   switch (data_type) {
-    case torch::kFloat16:
+    case torch::headeronly::ScalarType::Half:
       // Handle Float16
       LAUNCH_KERNEL(half, int32_t);
       break;
-    case torch::kFloat32:
+    case torch::headeronly::ScalarType::Float:
       // Handle Float32
       LAUNCH_KERNEL(float, int32_t);
       break;
-    case torch::kBFloat16:
+    case torch::headeronly::ScalarType::BFloat16:
       // Handle BFloat16
       LAUNCH_KERNEL(__nv_bfloat16, int32_t);
       break;
     default:
       // Handle other data types
-      throw std::invalid_argument(
-          "Invalid dtype, only supports float16, float32, and bfloat16");
+      STD_TORCH_CHECK(
+          false, "Invalid dtype, only supports float16, float32, and bfloat16");
       break;
   }
 #undef LAUNCH_KERNEL