[CI][ROCm] Fix test_ocp_mx_wikitext_correctness reference value (#49690)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>

.buildkite/hardware_tests/amd.yaml

@@ -18,6 +18,8 @@ steps:
       TERM: "xterm-256color"
     retry:
       automatic:
+        - exit_status: 1  # Transient Docker/BuildKit failure
+          limit: 1
         - exit_status: -1  # Agent was lost
           limit: 1
         - exit_status: -10  # Agent was lost
@@ -46,6 +48,8 @@ steps:
       VLLM_BRANCH: "$BUILDKITE_COMMIT"
     retry:
       automatic:
+        - exit_status: 1  # Transient Docker/BuildKit failure
+          limit: 1
         - exit_status: -1  # Agent was lost
           limit: 1
         - exit_status: -10  # Agent was lost
@@ -72,6 +76,8 @@ steps:
       VLLM_BRANCH: "$BUILDKITE_COMMIT"
     retry:
       automatic:
+        - exit_status: 1  # Transient Docker/BuildKit failure
+          limit: 1
         - exit_status: -1  # Agent was lost
           limit: 1
         - exit_status: -10  # Agent was lost

.buildkite/hardware_tests/cpu.yaml

@@ -18,6 +18,8 @@ steps:
   - tests/kernels/quantization/test_cpu_fp8_scaled_mm.py
   - tests/kernels/mamba/cpu/test_cpu_gdn_ops.py
   - tests/kernels/mamba/test_cpu_short_conv.py
+  - tests/kernels/mamba/test_causal_conv1d.py
+  - tests/kernels/mamba/test_mamba_ssm.py
   commands:
     - |
       bash .buildkite/scripts/hardware_ci/run-cpu-test.sh 30m "
@@ -28,7 +30,9 @@ steps:
       pytest -x -v -s tests/kernels/test_onednn.py
       pytest -x -v -s tests/kernels/test_awq_int4_to_int8.py
       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"
+      pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py
+      pytest -x -v -s tests/kernels/mamba/test_causal_conv1d.py
+      pytest -x -v -s tests/kernels/mamba/test_mamba_ssm.py"
 
 # Note: SDE can't be downloaded from CI host because of AWS WAF
 # - label: CPU-Compatibility Tests

.buildkite/hardware_tests/intel_xpu_ci/test-intel.yaml

@@ -18,7 +18,7 @@ steps:
   - label: "XPU example Test"
     depends_on:
       - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 50
     optional: true
     device: intel_gpu
     agent_tags:
@@ -39,7 +39,7 @@ steps:
   - label: "XPU V1 test"
     depends_on:
       - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 70
     optional: true
     device: intel_gpu
     agent_tags:
@@ -60,7 +60,7 @@ steps:
   - label: "XPU server test"
     depends_on:
       - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 45
     optional: true
     device: intel_gpu
     agent_tags:

.buildkite/intel_jobs/basic_correctness.yaml

@@ -3,7 +3,7 @@ depends_on:
   - image-build-xpu
 steps:
 - label: XPU Sleep Mode
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   device: intel_gpu
   agent_tags:
     label: production

.buildkite/intel_jobs/benchmarks_intel.yaml

@@ -0,0 +1,26 @@
+group: Benchmarks
+depends_on: 
+  - image-build-xpu
+steps:
+- label: Benchmarks CLI Test
+  key: benchmarks-cli-test
+  timeout_in_minutes: 40
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
+  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/benchmarks/
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s benchmarks/'

.buildkite/intel_jobs/engine_intel.yaml

@@ -2,6 +2,44 @@ group: Engine Intel
 depends_on:
   - image-build-xpu
 steps:
+- label: Engine
+  key: engine
+  timeout_in_minutes: 40
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/compilation/
+  - vllm/config/
+  - vllm/engine/
+  - vllm/entrypoints/logger.py
+  - vllm/envs.py
+  - vllm/logger.py
+  - vllm/logging_utils/
+  - vllm/platforms/
+  - vllm/sequence.py
+  - vllm/triton_utils/
+  - vllm/utils/
+  - tests/engine
+  - tests/test_sequence
+  - tests/test_config
+  - tests/test_logger
+  - tests/test_vllm_port
+  - tests/test_jit_monitor.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s engine/test_arg_utils.py test_sequence.py test_logger.py test_vllm_port.py test_jit_monitor.py'
+
 - label: Engine (1 GPU)
   timeout_in_minutes: 30
   device: intel_gpu
@@ -23,3 +61,41 @@ steps:
       bash .buildkite/scripts/hardware_ci/run-intel-test.sh
       'cd tests &&
       pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py'
+
+- label: V1 e2e (2 GPUs)
+  timeout_in_minutes: 30
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 2+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/compilation/
+    - vllm/config/
+    - vllm/distributed/
+    - vllm/engine/
+    - vllm/envs.py
+    - vllm/forward_context.py
+    - vllm/inputs/
+    - vllm/logger.py
+    - vllm/logging_utils/
+    - vllm/model_executor/
+    - vllm/multimodal/
+    - vllm/platforms/
+    - vllm/sampling_params.py
+    - vllm/transformers_utils/
+    - vllm/triton_utils/
+    - vllm/utils/
+    - vllm/v1/
+    - tests/v1/e2e/spec_decode
+  commands:
+  - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"'

.buildkite/intel_jobs/lora_intel.yaml

@@ -86,7 +86,7 @@ steps:
       pytest -v -s lora/test_punica_ops.py::test_add_lora_fused_moe_early_exit'
 
 - label: LoRA Punica FP8/XPU Ops
-  timeout_in_minutes: 45
+  timeout_in_minutes: 60
   device: intel_gpu
   agent_tags:
     label: production

.buildkite/intel_jobs/misc_intel.yaml

@@ -3,7 +3,7 @@ depends_on:
   - image-build-xpu
 steps:
 - label: V1 Core + KV + Metrics
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   device: intel_gpu
   agent_tags:
     label: production
@@ -33,7 +33,7 @@ steps:
       pytest -v -s v1/executor'
 
 - label: V1 Sample + Logits
-  timeout_in_minutes: 30
+  timeout_in_minutes: 90
   device: intel_gpu
   agent_tags:
     label: production
@@ -125,13 +125,13 @@ steps:
       pytest -v -s v1/kv_offload &&
       pytest -v -s v1/kv_connector/unit/test_offloading_connector.py'
 
-- label: NixlConnector PD accuracy (2 GPUs)
+- label: NixlConnector PD accuracy (4 GPUs)
   timeout_in_minutes: 60
-  num_devices: 2
+  num_devices: 4
   device: intel_gpu
   agent_tags:
     label: production
-    gpu: 2+
+    gpu: 4+
     mem: 16+
   no_plugin: true
   working_dir: "."
@@ -148,11 +148,14 @@ steps:
     - >-
       bash .buildkite/scripts/hardware_ci/run-intel-test.sh
       'cd tests &&
-      bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh'
+      bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
+      PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=1 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
+      PREFILLER_TP_SIZE=1 DECODER_TP_SIZE=2 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh &&
+      PREFILLER_TP_SIZE=2 DECODER_TP_SIZE=2 bash v1/kv_connector/nixl_integration/run_xpu_disagg_accuracy_test.sh'
 
 - label: Regression
   key: regression
-  timeout_in_minutes: 30
+  timeout_in_minutes: 50
   device: intel_gpu
   agent_tags:
     label: production
@@ -186,7 +189,7 @@ steps:
 
 - label: Metrics, Tracing (2 GPUs)
   key: metrics-tracing-2-gpus
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   num_devices: 2
   device: intel_gpu
   agent_tags:
@@ -222,7 +225,7 @@ steps:
 
 - label: Async Engine, Inputs, Utils, Worker
   key: async-engine-inputs-utils-worker
-  timeout_in_minutes: 30
+  timeout_in_minutes: 55
   device: intel_gpu
   agent_tags:
     label: production
@@ -259,3 +262,25 @@ steps:
     pytest -v -s detokenizer &&
     pytest -v -s -m "not cpu_test" ./multimodal &&
     pytest -v -s utils_ --ignore=utils_/test_mem_utils.py'
+
+- label: Fusion Unit Tests
+  timeout_in_minutes: 30
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/compilation/
+    - tests/compile/passes/test_qk_norm_rope_fusion.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      pytest -v -s compile/passes/test_qk_norm_rope_fusion.py'

.buildkite/intel_jobs/model_executor_intel.yaml

@@ -0,0 +1,33 @@
+group: Model Executor Intel
+depends_on:
+  - image-build-xpu
+steps:
+- label: Model Executor (Intel)
+  key: model-executor-intel
+  timeout_in_minutes: 45
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/engine/arg_utils.py
+  - vllm/config/model.py
+  - vllm/model_executor
+  - tests/model_executor
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'apt-get update && apt-get install -y curl libsodium23 &&
+      pip3 install tensorizer==2.10.1 &&
+      pip3 install runai-model-streamer[s3,gcs,azure]\>=0.15.7 &&
+      export VLLM_WORKER_MULTIPROC_METHOD=spawn &&
+      export PYTHONFAULTHANDLER=1 &&
+      cd tests &&
+      pytest -v -s model_executor -m "not slow_test" --ignore="model_executor/layers/test_rocm_unquantized_gemm.py" --deselect="tests/model_executor/model_loader/test_reload.py::test_kv_scale_reload"'

.buildkite/intel_jobs/model_runner_v2_intel.yaml

@@ -8,7 +8,7 @@ steps:
   agent_tags:
     label: production
     gpu: 2+
-    mem: 16+
+    mem: 24+
   no_plugin: true
   working_dir: "."
   env:
@@ -28,7 +28,9 @@ steps:
       'export VLLM_USE_V2_MODEL_RUNNER=1 &&
       cd tests &&
       pytest -v -s v1/engine/test_llm_engine.py -k "not test_engine_metrics" &&
+      pytest -v -s v1/e2e/general/test_context_length.py &&
       ENFORCE_EAGER=1 pytest -v -s v1/e2e/general/test_async_scheduling.py -k "not ngram" &&
+      pytest -v -s entrypoints/llm/test_struct_output_generate.py -k "xgrammar and not speculative_config6 and not speculative_config7 and not speculative_config8 and not speculative_config0" &&
       pytest -v -s v1/e2e/general/test_min_tokens.py'
 
 - label: Model Runner V2 Examples (Intel)
@@ -60,3 +62,55 @@ steps:
       python3 basic/offline_inference/generate.py --model facebook/opt-125m &&
       python3 generate/multimodal/vision_language_offline.py --seed 0 &&
       python3 features/automatic_prefix_caching/prefix_caching_offline.py'
+
+- label: Model Runner V2 Distributed (2 GPUs)
+  timeout_in_minutes: 50
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 2+
+    mem: 16+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+    - vllm/v1/worker/gpu/
+    - vllm/v1/worker/gpu_worker.py
+    - tests/basic_correctness/test_basic_correctness.py
+    - tests/v1/distributed/test_async_llm_dp.py
+    - tests/v1/distributed/test_eagle_dp.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
+      cd tests &&
+      TARGET_TEST_SUITE=L4 pytest -v -s basic_correctness/test_basic_correctness.py -m "distributed\(num_gpus=2\)" -k "not ray and not True"'
+
+- label: Model Runner V2 Spec Decode
+  timeout_in_minutes: 50
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/v1/worker/gpu/
+  - vllm/v1/worker/gpu_worker.py
+  - tests/v1/spec_decode/test_max_len.py
+  - tests/v1/spec_decode/test_rejection_sampler_utils.py
+  - tests/v1/e2e/spec_decode/test_spec_decode.py
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'export VLLM_USE_V2_MODEL_RUNNER=1 &&
+      cd tests &&
+      pytest -v -s v1/spec_decode/test_synthetic_rejection_sampler_utils.py'

.buildkite/intel_jobs/models_distributed_intel.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Distributed Model Tests (2 GPUs)
   key: distributed-model-tests-2-gpus
-  timeout_in_minutes: 50
+  timeout_in_minutes: 65
   device: intel_gpu
   agent_tags:
     label: production

.buildkite/intel_jobs/models_multimodal_intel.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: "Multi-Modal Models (Standard) 1: qwen2"
   key: multi-modal-models-standard-1-qwen2
-  timeout_in_minutes: 45
+  timeout_in_minutes: 70
   device: intel_gpu
   agent_tags:
     label: production
@@ -29,7 +29,7 @@ steps:
 
 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
   key: multi-modal-models-standard-2-qwen3-gemma
-  timeout_in_minutes: 45
+  timeout_in_minutes: 70
   device: intel_gpu
   agent_tags:
     label: production
@@ -52,7 +52,7 @@ steps:
 
 - label: "Multi-Modal Models (Standard) 3: llava + qwen2_vl"
   key: multi-modal-models-standard-3-llava-qwen2-vl
-  timeout_in_minutes: 45
+  timeout_in_minutes: 65
   device: intel_gpu
   agent_tags:
     label: production
@@ -100,7 +100,7 @@ steps:
 
 - label: Multi-Modal Processor # 44min
   key: multi-modal-processor
-  timeout_in_minutes: 45
+  timeout_in_minutes: 60
   device: intel_gpu
   agent_tags:
     label: production

.buildkite/intel_jobs/samplers_intel.yaml

@@ -0,0 +1,29 @@
+group: Samplers Intel
+depends_on:
+  - image-build-xpu
+steps:
+- label: Samplers Test (FlashInfer)
+  key: samplers-test-flashinfer-intel
+  timeout_in_minutes: 40
+  device: intel_gpu
+  agent_tags:
+    label: production
+    gpu: 1+
+    mem: 24+
+  no_plugin: true
+  working_dir: "."
+  env:
+    REGISTRY: "public.ecr.aws/q9t5s3a7"
+    REPO: "vllm-ci-test-repo"
+    VLLM_TEST_DEVICE: "xpu"
+  source_file_dependencies:
+  - vllm/model_executor/layers
+  - vllm/sampling_metadata.py
+  - tests/samplers
+  - tests/conftest.py
+  - vllm/entrypoints/generate/beam_search
+  commands:
+    - >-
+      bash .buildkite/scripts/hardware_ci/run-intel-test.sh
+      'cd tests &&
+      VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers'

.buildkite/intel_jobs/test-intel.yaml

@@ -17,7 +17,7 @@ steps:
   - label: "XPU example Test"
     depends_on:
       - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 50
     device: intel_gpu
     agent_tags:
       label: production
@@ -76,7 +76,7 @@ steps:
   - label: "XPU V1 test"
     depends_on:
       - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 70
     device: intel_gpu
     agent_tags:
       label: production
@@ -99,12 +99,13 @@ steps:
         pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py &&
         pytest -v -s v1/structured_output &&
         pytest -v -s v1/test_serial_utils.py &&
+        pytest -v -s v1/e2e/general/test_correctness_sliding_window.py --deselect="tests/v1/e2e/general/test_correctness_sliding_window.py::test_sliding_window_retrieval[True-1-5-google/gemma-3-1b-it]" &&
         pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py --ignore=v1/spec_decode/test_speculators_correctness.py &&
         pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py --ignore=v1/kv_connector/unit/test_hf3fs_client.py --ignore=v1/kv_connector/unit/test_hf3fs_connector.py --ignore=v1/kv_connector/unit/test_hf3fs_metadata_server.py --ignore=v1/kv_connector/unit/test_offloading_connector.py'
   - label: "XPU server test"
     depends_on:
       - image-build-xpu
-    timeout_in_minutes: 30
+    timeout_in_minutes: 45
     device: intel_gpu
     agent_tags:
       label: production

.buildkite/lm-eval-harness/configs/Meta-Llama-4-Maverick-17B-128E-Instruct-FP8.yaml

@@ -1,6 +1,7 @@
 # For hf script, without -t option (tensor parallel size).
 # bash .buildkite/lm-eval-harness/run-lm-eval-mmlupro-vllm-baseline.sh -m meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 -l 250 -t 8 -f 5
 model_name: "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8"
+rocm_safetensors_load_strategy: lazy
 required_gpu_arch:
   - gfx942
   - gfx950

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

@@ -72,6 +72,11 @@ def launch_lm_eval(eval_config, tp_size):
     if moe_backend is not None:
         model_args += f"moe_backend={moe_backend},"
 
+    if current_platform.is_rocm():
+        rocm_load_strategy = eval_config.get("rocm_safetensors_load_strategy")
+        if rocm_load_strategy is not None:
+            model_args += f"safetensors_load_strategy={rocm_load_strategy},"
+
     env_vars = eval_config.get("env_vars", None)
     with scoped_env_vars(env_vars):
         results = lm_eval.simple_evaluate(

.buildkite/scripts/annotate-build-artifact.sh

@@ -3,7 +3,8 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 #
 # Append a build artifact line to the Buildkite annotation.
-# Usage: annotate-build-artifact.sh <label> <value>
+# Usage: annotate-build-artifact.sh <label> <value> <context>
 set -e
 echo "- **${1}**: \`${2}\`" | \
-  buildkite-agent annotate --append --style 'info' --context 'release-artifacts'
+  buildkite-agent annotate --append --style 'info' \
+    --context "${3:?context is required}"

.buildkite/scripts/build-macos-wheel.sh

@@ -0,0 +1,35 @@
+#!/usr/bin/env bash
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+#
+# Build the macOS arm64 CPU wheel natively on a macOS agent (the `macmini`
+# queue) into artifacts/dist/ for upload-nightly-wheels.sh.
+
+set -euo pipefail
+
+# The macmini queue uses persistent checkouts, so refresh tags for setuptools-scm.
+git fetch --tags --force origin
+
+# The Rust frontend build needs protoc.
+if ! command -v protoc >/dev/null 2>&1; then
+  brew install protobuf
+fi
+
+# upload-nightly-wheels.sh expects exactly one wheel.
+rm -rf artifacts/dist
+mkdir -p artifacts/dist
+
+export VLLM_TARGET_DEVICE=cpu
+export VLLM_REQUIRE_RUST_FRONTEND=1
+export MACOSX_DEPLOYMENT_TARGET=11.0
+# uv's CPython is universal2; force an arm64-only build and tag so the wheel
+# isn't mislabelled universal2 and installed on Intel Macs where import fails.
+export ARCHFLAGS="-arch arm64"
+export _PYTHON_HOST_PLATFORM="macosx-11.0-arm64"
+export CMAKE_BUILD_PARALLEL_LEVEL="${CMAKE_BUILD_PARALLEL_LEVEL:-4}"
+
+uv venv --python 3.12
+uv pip install -r requirements/build/cpu.txt --index-strategy unsafe-best-match
+uv build --wheel --no-build-isolation -o artifacts/dist
+
+ls -l artifacts/dist/*.whl

.buildkite/scripts/ci-bake-rocm.sh

@@ -15,9 +15,9 @@ set -euo pipefail
 
 DEFAULT_REPO_SLUG="vllm-project/vllm"
 DEFAULT_CI_HCL_SOURCE="docker/ci-rocm.hcl"
-DEFAULT_CI_BASE_CONTENT_FILES="requirements/common.txt requirements/rocm.txt requirements/test/rocm.txt docker/Dockerfile.rocm_base docker/ci-rocm.hcl docker/docker-bake-rocm.hcl tools/install_torchcodec_rocm.sh tests/vllm_test_utils .buildkite/scripts/ci-bake-rocm.sh .buildkite/scripts/rocm/build-ci-base.sh"
+DEFAULT_CI_BASE_CONTENT_FILES="requirements/common.txt requirements/rocm.txt requirements/test/rocm.txt docker/Dockerfile.rocm_base docker/ci-rocm.hcl docker/docker-bake-rocm.hcl tools/install_torchcodec_rocm.sh tools/install_protoc.sh rust-toolchain.toml tests/vllm_test_utils .buildkite/scripts/ci-bake-rocm.sh .buildkite/scripts/rocm/build-ci-base.sh"
 DEFAULT_CI_BASE_DOCKERFILE="docker/Dockerfile.rocm"
-DEFAULT_CI_BASE_DOCKERFILE_STAGES="base build_rixl build_rocshmem build_deepep mori_base ci_base"
+DEFAULT_CI_BASE_DOCKERFILE_STAGES="base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust-toolchain build_nixl build_rocshmem build_deepep mori_base ci_base"
 DEFAULT_CI_BASE_METADATA_VERSION="1"
 IMAGE_EXISTED_BEFORE_BUILD=0
 
@@ -285,7 +285,7 @@ get_content_arg_names() {
     fi | awk 'NF && !seen[$0]++'
 }
 
-compute_ci_base_content_hash() {
+compute_ci_base_content_hash_once() {
     local -a content_paths=()
     local -a content_args=()
     local dockerfile="${CI_BASE_DOCKERFILE:-}"
@@ -301,7 +301,8 @@ compute_ci_base_content_hash() {
         if [[ -n "${dockerfile}" ]]; then
             printf 'dockerfile:%s\n' "${dockerfile}"
             printf 'resolved-build-args:\n'
-            hash_dockerfile_arg_values "${dockerfile}" "${content_args[@]}"
+            hash_dockerfile_arg_values "${dockerfile}" "${content_args[@]}" \
+                || return 1
             if [[ -n "${stages}" ]]; then
                 printf 'dockerfile-stages:%s\n' "${stages}"
                 if [[ -f "${dockerfile}" ]]; then
@@ -314,6 +315,53 @@ compute_ci_base_content_hash() {
     } | sha256sum | cut -d' ' -f1
 }
 
+compute_ci_base_content_hash() {
+    local attempts="${CI_BASE_HASH_ATTEMPTS:-3}"
+    local delay_secs="${CI_BASE_HASH_RETRY_DELAY:-5}"
+    local attempt=0
+    local hash=""
+    local failed=0
+    local -a hashes=()
+
+    if [[ ! "${attempts}" =~ ^[1-9][0-9]*$ ]]; then
+        echo "Invalid CI_BASE_HASH_ATTEMPTS: ${attempts}" >&2
+        return 1
+    fi
+    if [[ ! "${delay_secs}" =~ ^[0-9]+$ ]]; then
+        echo "Invalid CI_BASE_HASH_RETRY_DELAY: ${delay_secs}" >&2
+        return 1
+    fi
+
+    for ((attempt = 1; attempt <= attempts; attempt++)); do
+        if ! hash=$(compute_ci_base_content_hash_once); then
+            echo "ci_base content hash calculation ${attempt}/${attempts} failed" >&2
+            failed=1
+        else
+            hashes+=("${hash}")
+            echo "ci_base content hash calculation ${attempt}/${attempts}: ${hash}" >&2
+        fi
+
+        if ((attempt < attempts)); then
+            sleep "${delay_secs}"
+        fi
+    done
+
+    if ((failed)) || ((${#hashes[@]} != attempts)); then
+        echo "Could not calculate a reliable ci_base content hash" >&2
+        return 1
+    fi
+
+    for hash in "${hashes[@]:1}"; do
+        if [[ "${hash}" != "${hashes[0]}" ]]; then
+            echo "ci_base content hash changed between calculations" >&2
+            printf '  observed: %s\n' "${hashes[@]}" >&2
+            return 1
+        fi
+    done
+
+    printf '%s\n' "${hashes[0]}"
+}
+
 extract_dockerfile_arg_default() {
     local dockerfile="$1"
     local arg_name="$2"
@@ -366,7 +414,11 @@ hash_dockerfile_arg_values() {
         printf 'arg:%s=%s\n' "${arg_name}" "${arg_value:-<empty>}"
         if [[ "${arg_name}" == "BASE_IMAGE" && -n "${arg_value}" ]]; then
             digest=$(resolve_image_digest "${arg_value}")
-            printf 'arg:%s.digest=%s\n' "${arg_name}" "${digest:-unknown}"
+            if [[ -z "${digest}" ]]; then
+                echo "Failed to resolve digest for BASE_IMAGE=${arg_value}" >&2
+                return 1
+            fi
+            printf 'arg:%s.digest=%s\n' "${arg_name}" "${digest}"
         fi
     done
 }
@@ -764,7 +816,7 @@ configure_ci_base_image_refs() {
         fi
         set_buildkite_metadata "rocm-ci-base-image" "${CI_BASE_IMAGE_TAG}"
         set_buildkite_metadata "rocm-ci-base-image-content" "${content_tag}"
-        set_buildkite_metadata "rocm-ci-base-image-commit" "${CI_BASE_IMAGE_TAG_COMMIT:-}"
+        set_buildkite_metadata "rocm-ci-base-image-commit" "${CI_BASE_IMAGE_TAG_COMMIT_REF:-}"
         set_buildkite_metadata "rocm-ci-base-image-stable" "${CI_BASE_IMAGE_TAG_STABLE:-}"
         return 0
     fi
@@ -1107,8 +1159,8 @@ ci_base_metadata_pairs() {
     metadata_pair "vllm.rocm.nic_backend" "$(resolve_dockerfile_arg_value "${dockerfile}" "NIC_BACKEND")"
     metadata_pair "vllm.rocm.ainic_version" "$(resolve_dockerfile_arg_value "${dockerfile}" "AINIC_VERSION")"
     metadata_pair "vllm.rocm.ubuntu_codename" "$(resolve_dockerfile_arg_value "${dockerfile}" "UBUNTU_CODENAME")"
-    metadata_pair "vllm.rocm.rixl_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "RIXL_REPO")"
-    metadata_pair "vllm.rocm.rixl_commit" "${RIXL_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "RIXL_BRANCH")}"
+    metadata_pair "vllm.rocm.nixl_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "NIXL_REPO")"
+    metadata_pair "vllm.rocm.nixl_commit" "${NIXL_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "NIXL_BRANCH")}"
     metadata_pair "vllm.rocm.ucx_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "UCX_REPO")"
     metadata_pair "vllm.rocm.ucx_commit" "${UCX_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "UCX_BRANCH")}"
     metadata_pair "vllm.rocm.rocshmem_repo" "$(resolve_dockerfile_arg_value "${dockerfile}" "ROCSHMEM_REPO")"
@@ -1117,7 +1169,7 @@ ci_base_metadata_pairs() {
     metadata_pair "vllm.rocm.deepep_commit" "${DEEPEP_BRANCH:-$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_BRANCH")}"
     metadata_pair "vllm.rocm.deepep_nic" "$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_NIC")"
     metadata_pair "vllm.rocm.deepep_rocm_arch" "$(resolve_dockerfile_arg_value "${dockerfile}" "DEEPEP_ROCM_ARCH")"
-    metadata_pair "vllm.rocm.rixl_cache_key" "${RIXL_CACHE_KEY:-}"
+    metadata_pair "vllm.rocm.nixl_cache_key" "${NIXL_CACHE_KEY:-}"
     metadata_pair "vllm.rocm.rocshmem_cache_key" "${ROCSHMEM_CACHE_KEY:-}"
     metadata_pair "vllm.rocm.deepep_cache_key" "${DEEPEP_CACHE_KEY:-}"
 
@@ -1211,12 +1263,24 @@ uses_rocm_csrc_cache() {
     esac
 }
 
+uses_rocm_rust_cache() {
+    case "${TARGET}" in
+        rust-rocm-ci|test-rocm-ci|test-rocm-ci-with-wheel|test-rocm-ci-with-artifacts|export-wheel-rocm)
+            return 0
+            ;;
+        *)
+            return 1
+            ;;
+    esac
+}
+
 compute_rocm_csrc_content_hash() {
     local bake_dir=""
     local dockerfile_rocm=""
     local -a content_paths=(
         "requirements/common.txt"
         "requirements/rocm.txt"
+        "pyproject.toml"
         "setup.py"
         "CMakeLists.txt"
         "cmake"
@@ -1260,6 +1324,56 @@ compute_rocm_csrc_content_hash_if_needed() {
     echo "ROCm csrc content cache ref: ${ROCM_CSRC_CONTENT_CACHE_REF}"
 }
 
+compute_rocm_rust_content_hash() {
+    local bake_dir=""
+    local dockerfile_rocm=""
+    local -a content_paths=(
+        "requirements/build/rust.txt"
+        "rust/Cargo.lock"
+        "rust/Cargo.toml"
+        "rust/proto"
+        "rust/src"
+        "rust-toolchain.toml"
+        "tools/build_rust.py"
+        "tools/install_protoc.sh"
+        "build_rust.sh"
+    )
+    local -a content_args=()
+
+    bake_dir=$(dirname "${VLLM_BAKE_FILE}")
+    dockerfile_rocm="${bake_dir}/Dockerfile.rocm"
+    mapfile -t content_args < <(
+        get_content_arg_names "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build" "${ROCM_RUST_CONTENT_ARGS:-}"
+    )
+
+    {
+        printf 'rust-input-files-hash:%s\n' "$(compute_content_hash "${content_paths[@]}")"
+        printf 'dockerfile:%s\n' "${dockerfile_rocm}"
+        printf 'resolved-build-args:\n'
+        hash_dockerfile_arg_values "${dockerfile_rocm}" "${content_args[@]}"
+        printf 'dockerfile-stages:base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build\n'
+        if [[ -f "${dockerfile_rocm}" ]]; then
+            hash_dockerfile_stages "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build"
+        else
+            printf 'missing:%s\n' "${dockerfile_rocm}"
+        fi
+    } | sha256sum | cut -d' ' -f1
+}
+
+compute_rocm_rust_content_hash_if_needed() {
+    local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
+
+    if [[ "${ROCM_RUST_CONTENT_CACHE:-1}" == "0" ]] || ! uses_rocm_rust_cache; then
+        return 0
+    fi
+
+    ROCM_RUST_CONTENT_HASH=$(compute_rocm_rust_content_hash)
+    ROCM_RUST_CONTENT_CACHE_REF="${cache_repo}:rust-rocm-input-${ROCM_RUST_CONTENT_HASH}"
+    export ROCM_RUST_CONTENT_HASH
+    export ROCM_RUST_CONTENT_CACHE_REF
+    echo "ROCm Rust content cache ref: ${ROCM_RUST_CONTENT_CACHE_REF}"
+}
+
 write_hcl_string_list_entries() {
     local indent="$1"
     local value=""
@@ -1317,6 +1431,7 @@ write_rocm_build_arg_override() {
                 "${CI_BASE_DOCKERFILE_STAGES:-${DEFAULT_CI_BASE_DOCKERFILE_STAGES}}" \
                 "${CI_BASE_CONTENT_ARGS:-}"
             get_content_arg_names "${dockerfile_rocm}" "base csrc-build" "${ROCM_CSRC_CONTENT_ARGS:-}"
+            get_content_arg_names "${dockerfile_rocm}" "base rust_toolchain_input_0 rust_toolchain_input_1 rust-toolchain-input rust_input_0 rust_input_1 rust-input rust-toolchain rust-build" "${ROCM_RUST_CONTENT_ARGS:-}"
         } | awk 'NF && !seen[$0]++'
     )
 
@@ -1365,46 +1480,133 @@ validate_cache_export_mode() {
     esac
 }
 
+validate_content_cache_export_mode() {
+    local mode="$1"
+    local env_name="$2"
+
+    case "${mode}" in
+        missing|always|never)
+            ;;
+        *)
+            echo "Error: ${env_name} must be one of: missing, always, never"
+            exit 1
+            ;;
+    esac
+}
+
+should_export_content_cache_ref() {
+    local cache_ref="$1"
+    local cache_name="$2"
+    local mode="${ROCM_CONTENT_CACHE_EXPORT_MODE:-missing}"
+
+    case "${mode}" in
+        always)
+            echo "${cache_name} content cache export mode is always; exporting ${cache_ref}"
+            return 0
+            ;;
+        never)
+            echo "${cache_name} content cache export mode is never; not exporting ${cache_ref}"
+            return 1
+            ;;
+        missing|"")
+            if docker buildx imagetools inspect "${cache_ref}" >/dev/null 2>&1; then
+                echo "${cache_name} content cache exists; not re-exporting ${cache_ref}"
+                return 1
+            fi
+            echo "${cache_name} content cache missing; will export ${cache_ref}"
+            return 0
+            ;;
+        *)
+            echo "Error: ROCM_CONTENT_CACHE_EXPORT_MODE must be one of: missing, always, never"
+            exit 1
+            ;;
+    esac
+}
+
 write_rocm_cache_override() {
     local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
+    local content_cache_export_mode="${ROCM_CONTENT_CACHE_EXPORT_MODE:-missing}"
     local csrc_cache_to_mode="${ROCM_CSRC_CACHE_TO_MODE:-max}"
+    local rust_cache_to_mode="${ROCM_RUST_CACHE_TO_MODE:-max}"
     local rocm_cache_to_mode="${ROCM_FINAL_CACHE_TO_MODE:-min}"
-    local -a content_cache_from=()
+    local -a csrc_content_cache_from=()
+    local -a rust_content_cache_from=()
+    local -a combined_content_cache_from=()
     local -a csrc_cache_to=()
+    local -a rust_cache_to=()
     local -a rocm_cache_to=()
     local -a export_wheel_cache_to=()
+    local export_csrc_cache=1
+    local export_rust_cache=1
 
-    if ! uses_rocm_csrc_cache; then
+    if ! uses_rocm_csrc_cache && ! uses_rocm_rust_cache; then
         return 0
     fi
 
+    validate_content_cache_export_mode \
+        "${content_cache_export_mode}" \
+        "ROCM_CONTENT_CACHE_EXPORT_MODE"
     validate_cache_export_mode "${csrc_cache_to_mode}" "ROCM_CSRC_CACHE_TO_MODE"
+    validate_cache_export_mode "${rust_cache_to_mode}" "ROCM_RUST_CACHE_TO_MODE"
     validate_cache_export_mode "${rocm_cache_to_mode}" "ROCM_FINAL_CACHE_TO_MODE"
+    echo "ROCm content cache export mode: ${content_cache_export_mode}"
     echo "ROCm csrc cache export mode: ${csrc_cache_to_mode}"
+    echo "ROCm Rust cache export mode: ${rust_cache_to_mode}"
     echo "ROCm final image cache export mode: ${rocm_cache_to_mode}"
 
     if [[ -n "${ROCM_CSRC_CONTENT_CACHE_REF:-}" ]]; then
-        content_cache_from+=("type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF}")
-        csrc_cache_to+=(
-            "type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF},mode=${csrc_cache_to_mode},ignore-error=true"
-        )
+        csrc_content_cache_from+=("type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF}")
+        if should_export_content_cache_ref "${ROCM_CSRC_CONTENT_CACHE_REF}" "ROCm csrc"; then
+            csrc_cache_to+=(
+                "type=registry,ref=${ROCM_CSRC_CONTENT_CACHE_REF},mode=${csrc_cache_to_mode},ignore-error=true"
+            )
+        else
+            export_csrc_cache=0
+        fi
     fi
 
+    if [[ -n "${ROCM_RUST_CONTENT_CACHE_REF:-}" ]]; then
+        rust_content_cache_from+=("type=registry,ref=${ROCM_RUST_CONTENT_CACHE_REF}")
+        if should_export_content_cache_ref "${ROCM_RUST_CONTENT_CACHE_REF}" "ROCm Rust"; then
+            rust_cache_to+=(
+                "type=registry,ref=${ROCM_RUST_CONTENT_CACHE_REF},mode=${rust_cache_to_mode},ignore-error=true"
+            )
+        else
+            export_rust_cache=0
+        fi
+    fi
+
+    combined_content_cache_from=("${csrc_content_cache_from[@]}" "${rust_content_cache_from[@]}")
+
     # Docker Hub cache exports are best-effort. A cache-only target failure can
     # otherwise cancel the sibling image target before its manifest is pushed.
     if [[ -n "${BUILDKITE_COMMIT:-}" ]]; then
-        csrc_cache_to+=(
-            "type=registry,ref=${cache_repo}:csrc-rocm-${BUILDKITE_COMMIT},mode=${csrc_cache_to_mode},ignore-error=true"
-        )
+        if [[ ${export_csrc_cache} -eq 1 ]]; then
+            csrc_cache_to+=(
+                "type=registry,ref=${cache_repo}:csrc-rocm-${BUILDKITE_COMMIT},mode=${csrc_cache_to_mode},ignore-error=true"
+            )
+        fi
+        if [[ ${export_rust_cache} -eq 1 ]]; then
+            rust_cache_to+=(
+                "type=registry,ref=${cache_repo}:rust-rocm-${BUILDKITE_COMMIT},mode=${rust_cache_to_mode},ignore-error=true"
+            )
+        fi
         rocm_cache_to+=(
             "type=registry,ref=${cache_repo}:rocm-${BUILDKITE_COMMIT},mode=${rocm_cache_to_mode},ignore-error=true"
         )
     fi
 
     if [[ -n "${ROCM_CACHE_BRANCH_TAG:-}" ]]; then
-        csrc_cache_to+=(
-            "type=registry,ref=${cache_repo}:csrc-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${csrc_cache_to_mode},ignore-error=true"
-        )
+        if [[ ${export_csrc_cache} -eq 1 ]]; then
+            csrc_cache_to+=(
+                "type=registry,ref=${cache_repo}:csrc-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${csrc_cache_to_mode},ignore-error=true"
+            )
+        fi
+        if [[ ${export_rust_cache} -eq 1 ]]; then
+            rust_cache_to+=(
+                "type=registry,ref=${cache_repo}:rust-rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${rust_cache_to_mode},ignore-error=true"
+            )
+        fi
         rocm_cache_to+=(
             "type=registry,ref=${cache_repo}:rocm-branch-${ROCM_CACHE_BRANCH_TAG},mode=${rocm_cache_to_mode},ignore-error=true"
         )
@@ -1422,7 +1624,7 @@ target "csrc-rocm-ci" {
   cache-from = concat(
     get_cache_from_rocm_csrc(),
 EOF
-        write_hcl_string_list "    " "${content_cache_from[@]}"
+        write_hcl_string_list "    " "${csrc_content_cache_from[@]}"
         cat <<EOF
   )
 EOF
@@ -1430,11 +1632,23 @@ EOF
         cat <<EOF
 }
 
+target "rust-rocm-ci" {
+  cache-from = concat(
+    get_cache_from_rocm_rust(),
+EOF
+        write_hcl_string_list "    " "${rust_content_cache_from[@]}"
+        cat <<EOF
+  )
+EOF
+        write_hcl_string_list_attr "  " "cache-to" "${rust_cache_to[@]}"
+        cat <<EOF
+}
+
 target "test-rocm-ci" {
   cache-from = concat(
     get_cache_from_rocm(),
 EOF
-        write_hcl_string_list "    " "${content_cache_from[@]}"
+        write_hcl_string_list "    " "${combined_content_cache_from[@]}"
         cat <<EOF
   )
 EOF
@@ -1446,7 +1660,7 @@ target "export-wheel-rocm" {
   cache-from = concat(
     get_cache_from_rocm(),
 EOF
-        write_hcl_string_list "    " "${content_cache_from[@]}"
+        write_hcl_string_list "    " "${combined_content_cache_from[@]}"
         cat <<EOF
   )
 EOF
@@ -1472,7 +1686,7 @@ extract_dependency_pins() {
         return 0
     fi
 
-    for var in RIXL_BRANCH UCX_BRANCH ROCSHMEM_BRANCH DEEPEP_BRANCH; do
+    for var in NIXL_BRANCH UCX_BRANCH ROCSHMEM_BRANCH DEEPEP_BRANCH; do
         if [[ -n "${!var:-}" ]]; then
             echo "Using provided ${var}: ${!var}"
             continue
@@ -1492,30 +1706,30 @@ extract_dependency_pins() {
 compute_dependency_cache_keys() {
     local bake_dir=""
     local dockerfile_rocm=""
-    local rixl_branch=""
+    local nixl_branch=""
     local ucx_branch=""
     local rocshmem_branch=""
     local deepep_branch=""
-    local rixl_material=""
+    local nixl_material=""
     local rocshmem_material=""
     local deepep_material=""
 
     bake_dir=$(dirname "${VLLM_BAKE_FILE}")
     dockerfile_rocm="${bake_dir}/Dockerfile.rocm"
-    rixl_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "RIXL_BRANCH")
+    nixl_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "NIXL_BRANCH")
     ucx_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "UCX_BRANCH")
     rocshmem_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "ROCSHMEM_BRANCH")
     deepep_branch=$(resolve_dockerfile_arg_value "${dockerfile_rocm}" "DEEPEP_BRANCH")
 
-    if [[ -n "${rixl_branch}" && -n "${ucx_branch}" ]]; then
-        rixl_material=$(compose_stage_cache_material "${dockerfile_rocm}" "base build_rixl")
-        RIXL_CACHE_KEY=$(
+    if [[ -n "${nixl_branch}" && -n "${ucx_branch}" ]]; then
+        nixl_material=$(compose_stage_cache_material "${dockerfile_rocm}" "base build_nixl")
+        NIXL_CACHE_KEY=$(
             compose_dependency_cache_key \
-                "${rixl_branch}-ucx-${ucx_branch}" \
-                "${rixl_material}"
+                "${nixl_branch}-ucx-${ucx_branch}" \
+                "${nixl_material}"
         )
-        export RIXL_CACHE_KEY
-        echo "RIXL dependency cache key: ${RIXL_CACHE_KEY}"
+        export NIXL_CACHE_KEY
+        echo "NIXL dependency cache key: ${NIXL_CACHE_KEY}"
     fi
 
     if [[ -n "${rocshmem_branch}" ]]; then
@@ -1566,11 +1780,11 @@ dependency_cache_ref_for_target() {
     local cache_repo="${DOCKERHUB_CACHE_REPO:-rocm/vllm-ci-cache}"
 
     case "${target}" in
-        rixl-rocm-ci)
-            if [[ -n "${RIXL_CACHE_KEY:-}" ]]; then
-                printf '%s\n' "${cache_repo}:rixl-rocm-${RIXL_CACHE_KEY}"
-            elif [[ -n "${RIXL_BRANCH:-}" ]]; then
-                printf '%s\n' "${cache_repo}:rixl-rocm-${RIXL_BRANCH}-ucx-${UCX_BRANCH:-}"
+        nixl-rocm-ci)
+            if [[ -n "${NIXL_CACHE_KEY:-}" ]]; then
+                printf '%s\n' "${cache_repo}:nixl-rocm-${NIXL_CACHE_KEY}"
+            elif [[ -n "${NIXL_BRANCH:-}" ]]; then
+                printf '%s\n' "${cache_repo}:nixl-rocm-${NIXL_BRANCH}-ucx-${UCX_BRANCH:-}"
             fi
             ;;
         rocshmem-rocm-ci)
@@ -1601,7 +1815,7 @@ add_dependency_cache_target() {
 
 resolve_ci_base_dependency_targets() {
     local mode="${ROCM_DEP_CACHE_EXPORT_MODE:-missing}"
-    local rixl_ref=""
+    local nixl_ref=""
     local rocshmem_ref=""
     local deepep_ref=""
 
@@ -1610,7 +1824,7 @@ resolve_ci_base_dependency_targets() {
     case "${mode}" in
         always)
             echo "ROCM_DEP_CACHE_EXPORT_MODE=always; exporting all dependency caches serially"
-            for target in rixl-rocm-ci rocshmem-rocm-ci deepep-rocm-ci; do
+            for target in nixl-rocm-ci rocshmem-rocm-ci deepep-rocm-ci; do
                 if [[ -n "$(dependency_cache_ref_for_target "${target}")" ]]; then
                     add_dependency_cache_target "${target}"
                 fi
@@ -1630,13 +1844,13 @@ resolve_ci_base_dependency_targets() {
             ;;
     esac
 
-    if [[ "${mode}" != "always" && -n "${RIXL_CACHE_KEY:-}" ]]; then
-        rixl_ref=$(dependency_cache_ref_for_target "rixl-rocm-ci")
-        if dependency_cache_ref_exists "${rixl_ref}"; then
-            echo "RIXL dependency cache exists: ${rixl_ref}"
+    if [[ "${mode}" != "always" && -n "${NIXL_CACHE_KEY:-}" ]]; then
+        nixl_ref=$(dependency_cache_ref_for_target "nixl-rocm-ci")
+        if dependency_cache_ref_exists "${nixl_ref}"; then
+            echo "NIXL dependency cache exists: ${nixl_ref}"
         else
-            echo "RIXL dependency cache missing; will seed: ${rixl_ref}"
-            add_dependency_cache_target "rixl-rocm-ci"
+            echo "NIXL dependency cache missing; will seed: ${nixl_ref}"
+            add_dependency_cache_target "nixl-rocm-ci"
         fi
     fi
 
@@ -1736,8 +1950,8 @@ confirm_remote_image_push() {
         fi
 
         if [[ -z "${remote_revision}" \
-              && ${IMAGE_EXISTED_BEFORE_BUILD} -eq 0 \
-              && image_tag_is_commit_scoped ]]; then
+              && ${IMAGE_EXISTED_BEFORE_BUILD} -eq 0 ]] \
+            && image_tag_is_commit_scoped; then
             echo "Remote image exists under a commit-scoped tag; accepting push despite missing revision label."
             return 0
         fi
@@ -1867,36 +2081,57 @@ upload_wheel_artifacts_if_present() {
     local wheel_dir="./wheel-export"
     local artifact_dir="artifacts/vllm-rocm-install"
     local archive_name="vllm-rocm-install.tar.gz"
+    local metadata_dir="${wheel_dir}/.vllm-ci-artifact"
+    local native_base_image=""
     local whl=""
     local whl_name=""
+    local -a wheels=()
 
     if ! should_upload_wheel_artifacts; then
         return 0
     fi
 
-    if [[ ! -d "${wheel_dir}" ]] || ! ls "${wheel_dir}"/*.whl >/dev/null 2>&1; then
-        echo "No ROCm wheel artifacts found in ${wheel_dir}"
-        return 0
+    if [[ -d "${wheel_dir}" ]]; then
+        mapfile -t wheels < <(find "${wheel_dir}" -maxdepth 1 -type f -name '*.whl' -print)
+    fi
+    if [[ ${#wheels[@]} -ne 1 ]]; then
+        echo "Expected exactly one ROCm wheel in ${wheel_dir}; found ${#wheels[@]}" >&2
+        return 1
+    fi
+    whl="${wheels[0]}"
+    whl_name=$(basename "${whl}")
+    native_base_image="${CI_BASE_IMAGE_TAG_COMMIT_REF:-${CI_BASE_IMAGE:-}}"
+    if [[ -z "${native_base_image}" ]]; then
+        echo "Native ROCm artifact requires a ci_base image reference" >&2
+        return 1
     fi
 
     echo "--- :package: Uploading ROCm vLLM install artifact"
-    mkdir -p "${artifact_dir}"
+    rm -rf "${artifact_dir}" "${metadata_dir}"
+    mkdir -p "${artifact_dir}" "${metadata_dir}"
+
+    printf '%s\n' "${BUILDKITE_COMMIT:-local}" > "${metadata_dir}/commit.txt"
+    printf '%s\n' "${native_base_image}" > "${metadata_dir}/native-base-image.txt"
+    printf '%s\n' "${CI_BASE_IMAGE:-}" > "${metadata_dir}/ci-base-image.txt"
+    printf '%s\n' "${IMAGE_TAG:-}" > "${metadata_dir}/fallback-image.txt"
+    printf '%s\n' "${whl_name}" > "${metadata_dir}/wheel-filename.txt"
 
     tar -C "${wheel_dir}" -czf "${artifact_dir}/${archive_name}" .
+    (
+        cd "${artifact_dir}"
+        sha256sum "${archive_name}" > "${archive_name}.sha256"
+    )
     echo "Created ${archive_name}: $(du -sh "${artifact_dir}/${archive_name}" | cut -f1)"
-    printf '%s\n' "${CI_BASE_IMAGE:-}" > "${artifact_dir}/ci-base-image.txt"
-    printf '%s\n' "${IMAGE_TAG:-}" > "${artifact_dir}/fallback-image.txt"
-
-    for whl in "${wheel_dir}"/*.whl; do
-        [[ -f "${whl}" ]] || continue
-        whl_name=$(basename "${whl}")
-        cp "${whl}" "${artifact_dir}/${whl_name}"
-        echo "Copied ${whl_name}: $(du -sh "${artifact_dir}/${whl_name}" | cut -f1)"
-    done
+    cp "${metadata_dir}"/*.txt "${artifact_dir}/"
+    cp "${whl}" "${artifact_dir}/${whl_name}"
+    echo "Copied ${whl_name}: $(du -sh "${artifact_dir}/${whl_name}" | cut -f1)"
 
     if command -v buildkite-agent >/dev/null 2>&1; then
-        buildkite-agent artifact upload "${artifact_dir}/*"
+        buildkite-agent artifact upload "${artifact_dir}/*" || return 1
         echo "ROCm vLLM install artifacts uploaded to ${artifact_dir}/"
+    elif [[ "${BUILDKITE:-false}" == "true" ]]; then
+        echo "buildkite-agent not found; cannot upload required ROCm artifacts" >&2
+        return 1
     else
         echo "Not in Buildkite, skipping artifact upload"
     fi
@@ -1920,6 +2155,7 @@ main() {
     compute_dependency_cache_keys
     write_ci_base_label_override
     compute_rocm_csrc_content_hash_if_needed
+    compute_rocm_rust_content_hash_if_needed
     write_rocm_cache_override
     resolve_ci_base_dependency_targets
     print_bake_config
@@ -1927,6 +2163,11 @@ main() {
         echo "BAKE_PRINT_ONLY=1 set; skipping build"
         return 0
     fi
+    if should_upload_wheel_artifacts; then
+        # wheel-export is an output directory, not a BuildKit cache. Starting
+        # clean prevents a failed/retried export from packaging a stale wheel.
+        rm -rf ./wheel-export
+    fi
     seed_dependency_caches_if_needed
     run_bake
     upload_wheel_artifacts_if_present

.buildkite/scripts/generate-and-upload-nightly-index.sh

@@ -45,8 +45,10 @@ $PYTHON .buildkite/scripts/generate-nightly-index.py --version "$SUBPATH" --curr
 echo "Uploading indices to $S3_COMMIT_PREFIX"
 aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "$S3_COMMIT_PREFIX"
 
-# copy to /nightly/ only if it is on the main branch and not a PR
-if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]]; then
+# copy to /nightly/ only when enabled for a main branch build that is not a PR
+if [[ "${UPDATE_NIGHTLY_INDEX:-1}" == "1" && \
+      "$BUILDKITE_BRANCH" == "main" && \
+      "$BUILDKITE_PULL_REQUEST" == "false" ]]; then
     echo "Uploading indices to overwrite /nightly/"
     aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/nightly/"
 fi
@@ -67,7 +69,7 @@ pure_version="${version%%+*}"
 echo "Pure version (without variant): $pure_version"
 
 # re-generate and copy to /<pure_version>/ only if it does not have "dev" in the version
-if [[ "$version" != *"dev"* ]]; then
+if [[ "${UPDATE_VERSION_INDEX:-1}" == "1" && "$version" != *"dev"* ]]; then
     echo "Re-generating indices for /$pure_version/"
     rm -rf "${INDICES_OUTPUT_DIR:?}"
     mkdir -p "$INDICES_OUTPUT_DIR"

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

@@ -1,7 +1,7 @@
 #!/bin/bash
 
-# This script runs tests inside the corresponding ROCm docker container.
-# It handles both single-node and multi-node test configurations.
+# This script runs ROCm tests either directly in a native CI pod or inside the
+# corresponding Docker container. Multi-node tests continue to use Docker.
 #
 # Multi-node detection: Instead of matching on fragile group names, we detect
 # multi-node jobs structurally by looking for the bracket command syntax
@@ -34,10 +34,27 @@ set -o pipefail
 : "${CLICOLOR_FORCE:=1}"
 : "${PY_COLORS:=1}"
 : "${ROCM_DOCKER_TTY:=1}"
+: "${PYTHONFAULTHANDLER:=1}"
+: "${PYTEST_TIMEOUT:=2400}"
 if [[ " ${PYTEST_ADDOPTS:-} " != *" --color"* ]]; then
   PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--color=yes"
 fi
-export BUILDKIT_PROGRESS TERM FORCE_COLOR CLICOLOR_FORCE PY_COLORS PYTEST_ADDOPTS ROCM_DOCKER_TTY
+if [[ " ${PYTEST_ADDOPTS:-} " != *" --durations="* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--durations=25"
+fi
+if [[ " ${PYTEST_ADDOPTS:-} " != *" --durations-min="* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--durations-min=1.0"
+fi
+# Dump stacks after 25 minutes, then stop an individual test after 40 minutes.
+if [[ " ${PYTEST_ADDOPTS:-} " != *" faulthandler_timeout="* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }-o faulthandler_timeout=1500"
+fi
+if [[ " ${PYTEST_ADDOPTS:-} " != *" --timeout-method="* &&
+  " ${PYTEST_ADDOPTS:-} " != *" --timeout-method "* ]]; then
+  PYTEST_ADDOPTS="${PYTEST_ADDOPTS:+${PYTEST_ADDOPTS} }--timeout-method=thread"
+fi
+export BUILDKIT_PROGRESS TERM FORCE_COLOR CLICOLOR_FORCE PY_COLORS PYTEST_ADDOPTS PYTEST_TIMEOUT ROCM_DOCKER_TTY
+export PYTHONFAULTHANDLER
 
 # 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
@@ -53,6 +70,28 @@ report_docker_usage() {
   docker system df || true
 }
 
+clear_ci_orchestration_env() {
+  unset -v \
+    VLLM_TEST_GROUP_NAME \
+    VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE \
+    VLLM_CI_ARTIFACT_STEP \
+    VLLM_TEST_CACHE \
+    VLLM_CI_EXECUTION_MODE \
+    VLLM_CI_WORKSPACE \
+    VLLM_CI_REQUIRE_WORKSPACE_MOUNT \
+    VLLM_TEST_COMMANDS \
+    VLLM_CI_BRANCH \
+    VLLM_CI_BASE_IMAGE \
+    VLLM_CI_FALLBACK_IMAGE \
+    VLLM_CI_DOCKER_DISABLED \
+    VLLM_CI_ARTIFACT_GLOB \
+    VLLM_CI_ARTIFACT_CHECKSUM_GLOB \
+    VLLM_CI_EXPECTED_GPU_COUNT \
+    VLLM_CI_USE_ARTIFACTS \
+    VLLM_CI_RESULTS_ROOT \
+    VLLM_ALLOW_DEPRECATED_BEAM_SEARCH
+}
+
 cleanup_network() {
   local max_nodes=${NUM_NODES:-2}
   for node in $(seq 0 $((max_nodes - 1))); do
@@ -145,7 +184,11 @@ prepare_artifact_image() {
   fi
 
   cp "${wheel_dir}"/*.whl "${context_dir}/wheels/" || return 1
-  tar -C "${wheel_dir}" --exclude='*.whl' -cf - . \
+  tar -C "${wheel_dir}" \
+    --exclude='*.whl' \
+    --exclude='.vllm-ci-artifact' \
+    --exclude='./.vllm-ci-artifact' \
+    -cf - . \
     | tar -C "${workspace_dir}" -xf - || return 1
   cat > "${context_dir}/Dockerfile" <<'EOF'
 ARG BASE_IMAGE
@@ -168,6 +211,276 @@ EOF
   return 0
 }
 
+is_native_runtime() {
+  [[ "${AMD_CI_RUNTIME:-}" == "native" || "${NATIVE_CI:-}" == "true" ]]
+}
+
+validate_native_workspace() {
+  local workspace_dir="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
+  local workspace_real=""
+  local checkout_real=""
+  local workspace_mount=""
+
+  mkdir -p "${workspace_dir}" || return 1
+  workspace_real=$(readlink -m "${workspace_dir}") || return 1
+  if [[ -n "${BUILDKITE_BUILD_CHECKOUT_PATH:-}" ]]; then
+    checkout_real=$(readlink -m "${BUILDKITE_BUILD_CHECKOUT_PATH}") || return 1
+    if [[ "${checkout_real}" == "${workspace_real}" \
+      || "${checkout_real}" == "${workspace_real}/"* \
+      || "${workspace_real}" == "${checkout_real}/"* ]]; then
+      echo "Refusing to replace ${workspace_real}; it overlaps the Buildkite checkout ${checkout_real}" >&2
+      return 1
+    fi
+  fi
+  if [[ "${VLLM_CI_REQUIRE_WORKSPACE_MOUNT:-1}" == "1" ]]; then
+    if ! command -v findmnt >/dev/null 2>&1; then
+      echo "findmnt is required to verify the native workspace mount" >&2
+      return 1
+    fi
+    workspace_mount=$(findmnt -n -T "${workspace_real}" -o TARGET 2>/dev/null || true)
+    if [[ "$(readlink -m "${workspace_mount:-/}")" != "${workspace_real}" ]]; then
+      echo "Native CI requires a dedicated volume mounted at ${workspace_real}" >&2
+      return 1
+    fi
+  fi
+}
+
+prepare_native_workspace() {
+  if [[ "${VLLM_CI_USE_ARTIFACTS:-0}" != "1" ]]; then
+    echo "Native CI requires VLLM_CI_USE_ARTIFACTS=1"
+    return 1
+  fi
+  if ! command -v buildkite-agent >/dev/null 2>&1; then
+    echo "buildkite-agent not found; cannot download ROCm wheel artifact"
+    return 1
+  fi
+  validate_native_workspace || return 1
+
+  local artifact_glob="${VLLM_CI_ARTIFACT_GLOB:-artifacts/vllm-rocm-install/vllm-rocm-install.tar.gz}"
+  local artifact_checksum_glob="${VLLM_CI_ARTIFACT_CHECKSUM_GLOB:-${artifact_glob}.sha256}"
+  local artifact_step="${VLLM_CI_ARTIFACT_STEP:-image-build-amd}"
+  local archive=""
+  local checksum=""
+  local download_dir=""
+  local metadata_dir=""
+  local recorded_base=""
+  local recorded_commit=""
+  local recorded_wheel=""
+  local workspace_dir="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
+  local wheel_dir=""
+  local attempt=0
+  local attempt_dir=""
+  local -a archives=()
+  local -a checksums=()
+  local -a wheels=()
+
+  artifact_work_dir=$(mktemp -d -t vllm-rocm-artifact.XXXXXX) || return 1
+  wheel_dir="${artifact_work_dir}/wheels"
+  mkdir -p "${wheel_dir}" || return 1
+
+  echo "--- Downloading ROCm wheel artifact from ${artifact_step} (native in-pod)"
+  for attempt in 1 2 3; do
+    attempt_dir="${artifact_work_dir}/download-${attempt}"
+    rm -rf "${attempt_dir}" || return 1
+    mkdir -p "${attempt_dir}" || return 1
+    if buildkite-agent artifact download \
+      "${artifact_glob}" "${attempt_dir}" --step "${artifact_step}" \
+      && buildkite-agent artifact download \
+        "${artifact_checksum_glob}" "${attempt_dir}" --step "${artifact_step}"; then
+      download_dir="${attempt_dir}"
+      break
+    fi
+    echo "Artifact download attempt ${attempt}/3 failed"
+    if [[ "${attempt}" -lt 3 ]]; then
+      sleep $((attempt * 2))
+    fi
+  done
+  if [[ -z "${download_dir}" ]]; then
+    echo "Failed to download ${artifact_glob} and ${artifact_checksum_glob} from ${artifact_step}"
+    return 1
+  fi
+
+  mapfile -t archives < <(
+    find "${download_dir}" -name "vllm-rocm-install.tar.gz" -type f -print
+  )
+  mapfile -t checksums < <(
+    find "${download_dir}" -name "vllm-rocm-install.tar.gz.sha256" -type f -print
+  )
+  if [[ ${#archives[@]} -ne 1 || ${#checksums[@]} -ne 1 ]]; then
+    echo "Expected exactly one ROCm archive and checksum; found ${#archives[@]} archive(s) and ${#checksums[@]} checksum(s)" >&2
+    return 1
+  fi
+  archive="${archives[0]}"
+  checksum="${checksums[0]}"
+  if [[ "$(dirname "${archive}")" != "$(dirname "${checksum}")" ]]; then
+    echo "ROCm archive and checksum were downloaded to different directories" >&2
+    return 1
+  fi
+  (
+    cd "$(dirname "${archive}")"
+    sha256sum -c "$(basename "${checksum}")"
+  ) || return 1
+
+  tar --no-same-owner -xzf "${archive}" -C "${wheel_dir}" || return 1
+  mapfile -t wheels < <(
+    find "${wheel_dir}" -maxdepth 1 -type f -name '*.whl' -print
+  )
+  if [[ ${#wheels[@]} -ne 1 ]]; then
+    echo "ROCm artifact must contain exactly one top-level wheel; found ${#wheels[@]}" >&2
+    return 1
+  fi
+  metadata_dir="${wheel_dir}/.vllm-ci-artifact"
+  for metadata_file in commit.txt native-base-image.txt wheel-filename.txt; do
+    if [[ ! -s "${metadata_dir}/${metadata_file}" ]]; then
+      echo "ROCm artifact metadata is missing ${metadata_file}" >&2
+      return 1
+    fi
+  done
+  for metadata_file in ci-base-image.txt fallback-image.txt; do
+    if [[ ! -f "${metadata_dir}/${metadata_file}" ]]; then
+      echo "ROCm artifact metadata is missing ${metadata_file}" >&2
+      return 1
+    fi
+  done
+
+  recorded_commit=$(tr -d '\r\n' < "${metadata_dir}/commit.txt")
+  recorded_base=$(tr -d '\r\n' < "${metadata_dir}/native-base-image.txt")
+  recorded_wheel=$(tr -d '\r\n' < "${metadata_dir}/wheel-filename.txt")
+  if [[ -z "${BUILDKITE_COMMIT:-}" || "${recorded_commit}" != "${BUILDKITE_COMMIT}" ]]; then
+    echo "ROCm artifact commit ${recorded_commit} does not match ${BUILDKITE_COMMIT:-unset}" >&2
+    return 1
+  fi
+  if [[ -z "${VLLM_CI_BASE_IMAGE:-}" || "${recorded_base}" != "${VLLM_CI_BASE_IMAGE}" ]]; then
+    echo "ROCm artifact base ${recorded_base} does not match ${VLLM_CI_BASE_IMAGE:-unset}" >&2
+    return 1
+  fi
+  if [[ "${recorded_wheel}" != "$(basename "${wheels[0]}")" ]]; then
+    echo "ROCm artifact wheel manifest ${recorded_wheel} does not match $(basename "${wheels[0]}")" >&2
+    return 1
+  fi
+  for required_dir in tests .buildkite requirements; do
+    if [[ ! -d "${wheel_dir}/${required_dir}" ]]; then
+      echo "ROCm wheel artifact did not contain ${required_dir}/" >&2
+      return 1
+    fi
+  done
+
+  echo "--- Installing ROCm wheel into pod environment"
+  python3 -m pip install --no-deps --force-reinstall "${wheels[0]}" || return 1
+
+  echo "--- Preparing ${workspace_dir} from artifact"
+  find "${workspace_dir}" -mindepth 1 -maxdepth 1 -exec rm -rf -- {} + || return 1
+  tar -C "${wheel_dir}" \
+    --exclude='*.whl' \
+    --exclude='.vllm-ci-artifact' \
+    --exclude='./.vllm-ci-artifact' \
+    -cf - . | tar --no-same-owner -C "${workspace_dir}" -xf - || return 1
+  if [[ ! -d "${workspace_dir}/tests" ]]; then
+    echo "Failed to stage the native test workspace" >&2
+    return 1
+  fi
+
+  return 0
+}
+
+initialize_native_environment() {
+  local job_id="${BUILDKITE_JOB_ID:-${BUILDKITE_PARALLEL_JOB:-local}}"
+  local job_id_suffix=""
+  local native_root=""
+  local hf_fstype=""
+  local hf_mount=""
+
+  if [[ "$(id -u)" -ne 0 ]]; then
+    echo "Native ROCm CI currently requires the ci_base container to run as root" >&2
+    return 1
+  fi
+
+  job_id="${job_id//[^A-Za-z0-9_.-]/_}"
+  job_id_suffix="${job_id##*-}"
+  job_id_suffix="${job_id_suffix:0:12}"
+  native_root="/tmp/vllm-native-${job_id}"
+  TMPDIR="/tmp/vllm-${job_id_suffix}/tmp"
+  VLLM_RPC_BASE_PATH="/tmp"
+  TORCHINDUCTOR_CACHE_DIR="${native_root}/cache/torchinductor"
+  TRITON_CACHE_DIR="${native_root}/cache/triton"
+  VLLM_CACHE_ROOT="${native_root}/cache/vllm"
+  XDG_CACHE_HOME="${native_root}/cache/xdg"
+  : "${HF_HOME:=/home/buildkite-agent/huggingface}"
+  # datasets uses POSIX locks that are unsupported by the shared HF NFS cache.
+  # Keep processed datasets job-local while retaining the persistent Hub cache.
+  HF_DATASETS_CACHE="${native_root}/cache/huggingface/datasets"
+  : "${HF_HUB_DOWNLOAD_TIMEOUT:=300}"
+  : "${HF_HUB_ETAG_TIMEOUT:=60}"
+  export TMPDIR VLLM_RPC_BASE_PATH
+  export TORCHINDUCTOR_CACHE_DIR TRITON_CACHE_DIR VLLM_CACHE_ROOT XDG_CACHE_HOME
+  export HF_HOME HF_DATASETS_CACHE HF_HUB_DOWNLOAD_TIMEOUT HF_HUB_ETAG_TIMEOUT
+  export PYTORCH_ROCM_ARCH=""
+
+  mkdir -p "${TMPDIR}" \
+    "${TORCHINDUCTOR_CACHE_DIR}" \
+    "${TRITON_CACHE_DIR}" \
+    "${VLLM_CACHE_ROOT}" \
+    "${XDG_CACHE_HOME}" \
+    "${HF_HOME}" \
+    "${HF_DATASETS_CACHE}" || return 1
+
+  echo "Native compile caches: VLLM_CACHE_ROOT=${VLLM_CACHE_ROOT} TORCHINDUCTOR_CACHE_DIR=${TORCHINDUCTOR_CACHE_DIR}"
+
+  if [[ "${VLLM_CI_REQUIRE_PERSISTENT_HF_CACHE:-0}" == "1" ]]; then
+    if ! command -v findmnt >/dev/null 2>&1; then
+      echo "findmnt is required to verify the native Hugging Face cache mount" >&2
+      return 1
+    fi
+    hf_mount=$(findmnt -n -T "${HF_HOME}" -o TARGET 2>/dev/null || true)
+    if [[ -z "${hf_mount}" || "${hf_mount}" == "/" ]]; then
+      echo "Native CI requires a persistent volume mounted at or above ${HF_HOME}" >&2
+      return 1
+    fi
+  fi
+
+  if command -v findmnt >/dev/null 2>&1; then
+    hf_fstype=$(findmnt -n -T "${HF_HOME}" -o FSTYPE 2>/dev/null || true)
+  fi
+  if [[ "${hf_fstype}" == nfs || "${hf_fstype}" == nfs4 ]]; then
+    # Keep hf-xet state local and avoid vectored writes on shared NFS.
+    export HF_XET_CACHE="${native_root}/cache/hf-xet"
+    export HF_XET_HIGH_PERFORMANCE=0
+    export HF_XET_RECONSTRUCTION_USE_VECTORED_WRITE=0
+    mkdir -p "${HF_XET_CACHE}" || return 1
+    echo "Configured hf-xet for shared ${hf_fstype} cache at ${HF_HOME}"
+  fi
+}
+
+run_native_preflight() {
+  local expected_gpus="${VLLM_CI_EXPECTED_GPU_COUNT:-1}"
+
+  if [[ ! "${expected_gpus}" =~ ^[0-9]+$ ]]; then
+    echo "Invalid VLLM_CI_EXPECTED_GPU_COUNT=${expected_gpus}" >&2
+    return 1
+  fi
+
+  python3 -c "import encodings, importlib.metadata as im, importlib.util as iu; [im.version(d) for d in ('transformers', 'torch', 'ray', 'sympy', 'markupsafe', 'vllm')]; missing=[m for m in ('torch.utils.model_zoo', 'transformers.models.nomic_bert', 'ray.dag', 'sympy.physics', 'markupsafe._speedups') if iu.find_spec(m) is None]; assert not missing, missing" || return 1
+
+  if [[ "${expected_gpus}" == "0" ]]; then
+    echo "Native CPU-only AMD job: skipping ROCm device validation"
+    return 0
+  fi
+
+  echo "--- ROCm info"
+  rocminfo || return 1
+  VLLM_CI_EXPECTED_GPU_COUNT="${expected_gpus}" python3 - <<'PY'
+import os
+
+import torch
+
+expected = int(os.environ["VLLM_CI_EXPECTED_GPU_COUNT"])
+assert torch.version.hip, "PyTorch is not a ROCm build"
+assert torch.cuda.is_available(), "ROCm GPU is not available to PyTorch"
+actual = torch.cuda.device_count()
+assert actual == expected, f"Expected {expected} ROCm GPU(s), found {actual}"
+PY
+}
+
 is_multi_node() {
   local cmds="$1"
   # Primary signal: NUM_NODES environment variable set by the pipeline
@@ -350,7 +663,58 @@ re_quote_pytest_markers() {
 # Main
 ###############################################################################
 
-# --- GPU initialization ---
+if is_native_runtime; then
+  echo "--- Native in-pod ROCm CI (AMD_CI_RUNTIME=${AMD_CI_RUNTIME:-unset}, NATIVE_CI=${NATIVE_CI:-unset})"
+  artifact_work_dir=""
+
+  cleanup_native_workspace() {
+    if [[ -n "${artifact_work_dir}" ]]; then
+      rm -rf "${artifact_work_dir}"
+    fi
+  }
+  trap cleanup_native_workspace EXIT
+
+  if [[ -n "${VLLM_TEST_COMMANDS:-}" ]]; then
+    commands="${VLLM_TEST_COMMANDS}"
+    commands_source="env"
+  else
+    commands="$*"
+    commands_source="argv"
+    if [[ -z "$commands" ]]; then
+      echo "Error: No test commands provided for native CI." >&2
+      exit 1
+    fi
+  fi
+
+  if [[ "$commands_source" == "argv" ]]; then
+    commands=$(re_quote_pytest_markers "$commands")
+  fi
+
+  if is_multi_node "$commands"; then
+    echo "Native CI does not support multi-node jobs yet."
+    exit 1
+  fi
+
+  if ! initialize_native_environment; then
+    echo "Failed to initialize the native test environment"
+    exit 1
+  fi
+  if ! prepare_native_workspace; then
+    echo "Failed to prepare native test workspace"
+    exit 1
+  fi
+
+  export PYTHONPATH="${VLLM_CI_WORKSPACE:-/vllm-workspace}"
+
+  echo "Native test commands: $commands"
+  run_native_preflight || exit 1
+  # Keep AMD CI orchestration variables out of vLLM's runtime environment.
+  clear_ci_orchestration_env
+  /bin/bash -o pipefail -c "${commands}"
+  handle_pytest_exit "$?"
+fi
+
+# --- GPU initialization for legacy Docker execution ---
 echo "--- ROCm info"
 rocminfo
 
@@ -452,25 +816,27 @@ fi
 
 echo "Final commands: $commands"
 
-# The ROCm test image often ships /vllm-workspace without .git (artifact tarball unpack).
-# tests/standalone_tests/python_only_compile.sh uses merge-base(HEAD, origin/main) for
-# wheels.vllm.ai; compute on the agent (full git checkout) and pass into the container.
-vllm_standalone_merge_base=""
-checkout="${BUILDKITE_BUILD_CHECKOUT_PATH:-}"
-if [[ -z "${checkout}" || ! -d "${checkout}" ]]; then
-  checkout="."
+standalone_merge_base_env=()
+if [[ "$commands" == *python_only_compile.sh* ]]; then
+  # The ROCm test image often ships /vllm-workspace without .git. Resolve the
+  # wheels.vllm.ai commit from the agent checkout for this test only.
+  vllm_standalone_merge_base=""
+  checkout="${BUILDKITE_BUILD_CHECKOUT_PATH:-}"
+  if [[ -z "${checkout}" || ! -d "${checkout}" ]]; then
+    checkout="."
+  fi
+  # Pass safe.directory per-command because Buildkite uses mixed user IDs.
+  if git -c "safe.directory=${checkout}" -C "${checkout}" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
+    vllm_standalone_merge_base="$(
+      git -c "safe.directory=${checkout}" -C "${checkout}" merge-base HEAD origin/main 2>/dev/null || true
+    )"
+  fi
+  if [[ -z "${vllm_standalone_merge_base}" ]]; then
+    vllm_standalone_merge_base="${BUILDKITE_COMMIT:-}"
+  fi
+  echo "INFO: passing CI_STANDALONE_MERGE_BASE into container: ${vllm_standalone_merge_base}"
+  standalone_merge_base_env=(-e "CI_STANDALONE_MERGE_BASE=${vllm_standalone_merge_base}")
 fi
-# Pass safe.directory per-command (-c) because buildkite runs will always fail
-# the next check on git 2.35.2+ due to mixed uses of root and buildkite-agent/uids.
-if git -c "safe.directory=${checkout}" -C "${checkout}" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
-  vllm_standalone_merge_base="$(
-    git -c "safe.directory=${checkout}" -C "${checkout}" merge-base HEAD origin/main 2>/dev/null || true
-  )"
-fi
-if [[ -z "${vllm_standalone_merge_base}" ]]; then
-  vllm_standalone_merge_base="${BUILDKITE_COMMIT:-}"
-fi
-echo "INFO: passing VLLM_STANDALONE_MERGE_BASE into container: ${vllm_standalone_merge_base}"
 
 MYPYTHONPATH="/vllm-workspace"
 
@@ -501,6 +867,7 @@ else
 fi
 
 # --- Route: multi-node vs single-node ---
+clear_ci_orchestration_env
 if is_multi_node "$commands"; then
   echo "--- Multi-node job detected"
   export DCKR_VER=$(docker --version | sed 's/Docker version \(.*\), build .*/\1/')
@@ -589,7 +956,9 @@ else
     -e FORCE_COLOR \
     -e CLICOLOR_FORCE \
     -e PY_COLORS \
+    -e PYTHONFAULTHANDLER \
     -e PYTEST_ADDOPTS \
+    -e PYTEST_TIMEOUT \
     -v "${HF_CACHE}:${HF_MOUNT}" \
     -e "HF_HOME=${HF_MOUNT}" \
     -e "PYTHONPATH=${MYPYTHONPATH}" \
@@ -599,7 +968,7 @@ else
     -e "VLLM_CACHE_ROOT=${CONTAINER_CACHE_ROOT}/vllm" \
     -e "XDG_CACHE_HOME=${CONTAINER_CACHE_ROOT}/xdg" \
     -e "PYTORCH_ROCM_ARCH=" \
-    -e "VLLM_STANDALONE_MERGE_BASE=${vllm_standalone_merge_base}" \
+    "${standalone_merge_base_env[@]}" \
     --name "${container_name}" \
     "${image_name}" \
     /bin/bash -c "${CONTAINER_PREFLIGHT} && ${commands}"

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

@@ -1,10 +1,11 @@
 #!/bin/bash
 set -euox pipefail
 
-export VLLM_CPU_KVCACHE_SPACE=1 
+export VLLM_CPU_KVCACHE_SPACE=1
 export VLLM_CPU_CI_ENV=1
-# Reduce sub-processes for acceleration
-export TORCH_COMPILE_DISABLE=1 
+# Skip torch.compile via vLLM's --enforce-eager flag (passed below) instead of
+# TORCH_COMPILE_DISABLE=1, which torch 2.12 no longer treats as a silent no-op
+# when callers specify fullgraph=True.
 export VLLM_ENABLE_V1_MULTIPROCESSING=0
 
 SDE_ARCHIVE="sde-external-10.7.0-2026-02-18-lin.tar.xz"
@@ -49,15 +50,15 @@ wait_for_pid_and_check_log() {
 }
 
 # Test Sky Lake (AVX512F)
-./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_0.log 2>&1 &
+./sde/sde64 -skl -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_0.log 2>&1 &
 PID_TEST_0=$!
 
 # Test Cascade Lake (AVX512F + VNNI)
-./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_1.log 2>&1 &
+./sde/sde64 -clx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_1.log 2>&1 &
 PID_TEST_1=$!
 
 # Test Cooper Lake (AVX512F + VNNI + BF16)
-./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 > test_2.log 2>&1 &
+./sde/sde64 -cpx -- python3 examples/basic/offline_inference/generate.py --model facebook/opt-125m --dtype bfloat16 --enforce-eager > test_2.log 2>&1 &
 PID_TEST_2=$!
 
 wait_for_pid_and_check_log $PID_TEST_0 test_0.log

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

@@ -40,7 +40,9 @@ function cpu_tests() {
     pytest -x -v -s tests/kernels/moe/test_cpu_fused_moe.py
     pytest -x -v -s tests/kernels/mamba/cpu/test_cpu_gdn_ops.py
     pytest -x -v -s tests/kernels/moe/test_cpu_int4_moe.py
-    pytest -x -v -s tests/kernels/mamba/test_cpu_short_conv.py"
+    pytest -x -v -s tests/kernels/mamba/test_cpu_short_conv.py
+    pytest -x -v -s tests/kernels/mamba/test_causal_conv1d.py
+    pytest -x -v -s tests/kernels/mamba/test_mamba_ssm.py"
 
   # skip tests requiring model downloads if HF_TOKEN is not set
   # due to rate-limits
@@ -97,3 +99,4 @@ function cpu_tests() {
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
 timeout 2h bash -c cpu_tests
+

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

@@ -35,6 +35,7 @@ case "${test_suite}" in
     pytest -v -s v1/worker --ignore=v1/worker/test_gpu_model_runner.py --ignore=v1/worker/test_worker_memory_snapshot.py
     pytest -v -s v1/structured_output
     pytest -v -s v1/test_serial_utils.py
+    pytest -v -s v1/e2e/general/test_correctness_sliding_window.py --deselect="tests/v1/e2e/general/test_correctness_sliding_window.py::test_sliding_window_retrieval[True-1-5-google/gemma-3-1b-it]"
     pytest -v -s v1/spec_decode --ignore=v1/spec_decode/test_max_len.py --ignore=v1/spec_decode/test_speculators_eagle3.py --ignore=v1/spec_decode/test_acceptance_length.py --ignore=v1/spec_decode/test_speculators_correctness.py
     pytest -v -s v1/kv_connector/unit --ignore=v1/kv_connector/unit/test_multi_connector.py --ignore=v1/kv_connector/unit/test_example_connector.py --ignore=v1/kv_connector/unit/test_lmcache_integration.py --ignore=v1/kv_connector/unit/test_hf3fs_client.py --ignore=v1/kv_connector/unit/test_hf3fs_connector.py --ignore=v1/kv_connector/unit/test_hf3fs_metadata_server.py --ignore=v1/kv_connector/unit/test_offloading_connector.py
     ;;

.buildkite/scripts/rocm/build-test-image.sh

@@ -13,6 +13,18 @@ metadata_get() {
     fi
 }
 
+use_ci_base_if_present() {
+    local ci_base_image=""
+
+    ci_base_image="$(metadata_get rocm-ci-base-image)"
+    if [[ -z "${ci_base_image}" ]]; then
+        return 1
+    fi
+
+    export CI_BASE_IMAGE="${ci_base_image}"
+    echo "Using ROCm ci_base image selected by the preceding build step: ${CI_BASE_IMAGE}"
+}
+
 use_refreshed_base_if_present() {
     local base_refreshed=""
 
@@ -22,15 +34,12 @@ use_refreshed_base_if_present() {
     fi
 
     export BASE_IMAGE
-    export CI_BASE_IMAGE
     export IMAGE_TAG_LATEST
 
     BASE_IMAGE="$(metadata_get rocm-base-image)"
-    CI_BASE_IMAGE="$(metadata_get rocm-ci-base-image)"
     IMAGE_TAG_LATEST="$(metadata_get rocm-ci-image-descriptive)"
 
     echo "Using refreshed ROCm base image for test image: ${BASE_IMAGE}"
-    echo "Using refreshed ROCm ci_base image for test image: ${CI_BASE_IMAGE}"
     if [[ -n "${IMAGE_TAG_LATEST}" ]]; then
         echo "Also tagging full ROCm CI image as: ${IMAGE_TAG_LATEST}"
     fi
@@ -41,6 +50,8 @@ use_refreshed_base_if_present() {
 main() {
     local base_refreshed=0
 
+    use_ci_base_if_present || true
+
     if use_refreshed_base_if_present; then
         base_refreshed=1
     fi

.buildkite/scripts/upload-nightly-wheels.sh

@@ -6,8 +6,14 @@ set -ex
 # manylinux platform tag with auditwheel.
 # Index generation is handled separately by generate-and-upload-nightly-index.sh.
 
-# shellcheck source=lib/manylinux.sh
-source .buildkite/scripts/lib/manylinux.sh
+# auditwheel is Linux-only; macOS wheels already carry a valid tag, so skip the
+# manylinux retag for them.
+WHEEL_PLATFORM="${VLLM_WHEEL_PLATFORM:-linux}"
+
+if [[ "$WHEEL_PLATFORM" == "linux" ]]; then
+  # shellcheck source=lib/manylinux.sh
+  source .buildkite/scripts/lib/manylinux.sh
+fi
 
 BUCKET="vllm-wheels"
 SUBPATH=$BUILDKITE_COMMIT
@@ -27,8 +33,10 @@ wheel="${wheel_files[0]}"
 
 # ========= detect manylinux tag and rename ==========
 
-wheel="$(apply_manylinux_tag "$wheel")"
-echo "Renamed wheel to: $wheel"
+if [[ "$WHEEL_PLATFORM" == "linux" ]]; then
+  wheel="$(apply_manylinux_tag "$wheel")"
+  echo "Renamed wheel to: $wheel"
+fi
 
 # Extract the version from the wheel
 version=$(unzip -p "$wheel" '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)

.buildkite/scripts/upload-rocm-wheels.sh

@@ -113,8 +113,8 @@ $PYTHON .buildkite/scripts/generate-nightly-index.py \
 echo "Uploading indices to $S3_COMMIT_PREFIX"
 aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "$S3_COMMIT_PREFIX"
 
-# Update rocm/nightly/ if on main branch and not a PR
-if [[ "$BUILDKITE_BRANCH" == "main" && "$BUILDKITE_PULL_REQUEST" == "false" ]] || [[ "$NIGHTLY" == "1" ]]; then
+# Only scheduled nightly builds should update the moving nightly index.
+if [[ "${NIGHTLY:-0}" == "1" ]]; then
     echo "Updating rocm/nightly/ index..."
     aws s3 cp --recursive "$INDICES_OUTPUT_DIR/" "s3://$BUCKET/rocm/nightly/"
 fi
@@ -147,7 +147,7 @@ echo ""
 echo "Install command (by commit):"
 echo "  pip install vllm --extra-index-url https://${BUCKET}.s3.amazonaws.com/$ROCM_SUBPATH/"
 echo ""
-if [[ "$BUILDKITE_BRANCH" == "main" ]] || [[ "$NIGHTLY" == "1" ]]; then
+if [[ "${NIGHTLY:-0}" == "1" ]]; then
     echo "Install command (nightly):"
     echo "  pip install vllm --extra-index-url https://${BUCKET}.s3.amazonaws.com/rocm/nightly/"
 fi

.buildkite/test_areas/attention.yaml

@@ -16,8 +16,9 @@ steps:
   parallelism: 2
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 95
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 125
       depends_on:
       - image-build-amd
       source_file_dependencies:

.buildkite/test_areas/basic_correctness.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Basic Correctness
   key: basic-correctness
-  timeout_in_minutes: 45
+  timeout_in_minutes: 68
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -18,7 +18,8 @@ steps:
   - pytest -v -s basic_correctness/test_cpu_offload.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 70
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 60
       depends_on:
       - image-build-amd

.buildkite/test_areas/benchmarks.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: Benchmarks CLI Test
   key: benchmarks-cli-test
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -13,7 +13,9 @@ steps:
   - pytest -v -s benchmarks/
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 40
       depends_on:
       - image-build-amd
 

.buildkite/test_areas/cuda.yaml

@@ -18,6 +18,7 @@ steps:
     - pytest -v -s cuda/test_platform_no_cuda_init.py
 
 - label: Cudagraph
+  device: h200_35gb
   key: cudagraph
   timeout_in_minutes: 30
   source_file_dependencies:
@@ -25,7 +26,10 @@ steps:
   - vllm/v1/cudagraph_dispatcher.py
   - vllm/config/compilation.py
   - vllm/compilation
+  - vllm/v1/worker/encoder_cudagraph.py
+  - vllm/v1/worker/encoder_cudagraph_defs.py
   commands:
     - pytest -v -s v1/cudagraph/test_cudagraph_dispatch.py
     - pytest -v -s v1/cudagraph/test_cudagraph_mode.py
-    - pytest -v -s v1/cudagraph/test_breakable_cudagraph.py
+    - pytest -v -s v1/cudagraph/test_breakable_cudagraph.py
+    - pytest -v -s v1/cudagraph/test_encoder_cudagraph.py

.buildkite/test_areas/disaggregated.yaml

@@ -15,8 +15,9 @@ steps:
     - bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
   mirror:
     amd:
+      dind: false
       device: mi300_4
-      timeout_in_minutes: 85
+      timeout_in_minutes: 60
       depends_on:
         - image-build-amd
       source_file_dependencies:
@@ -65,8 +66,9 @@ steps:
     - DP_EP=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
   mirror:
     amd:
+      dind: false
       device: mi300_4
-      timeout_in_minutes: 60
+      timeout_in_minutes: 40
       depends_on:
         - image-build-amd
       source_file_dependencies:
@@ -90,8 +92,9 @@ steps:
     - CROSS_LAYERS_BLOCKS=True bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
   mirror:
     amd:
+      dind: false
       device: mi300_4
-      timeout_in_minutes: 85
+      timeout_in_minutes: 60
       depends_on:
         - image-build-amd
       source_file_dependencies:
@@ -115,8 +118,9 @@ steps:
     - HYBRID_SSM=1 bash v1/kv_connector/nixl_integration/config_sweep_accuracy_test.sh
   mirror:
     amd:
+      dind: false
       device: mi300_4
-      timeout_in_minutes: 80
+      timeout_in_minutes: 55
       depends_on:
         - image-build-amd
       source_file_dependencies:
@@ -171,8 +175,9 @@ steps:
     - bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
   mirror:
     amd:
+      dind: false
       device: mi300_2
-      timeout_in_minutes: 70
+      timeout_in_minutes: 45
       depends_on:
         - image-build-amd
       source_file_dependencies:
@@ -182,7 +187,7 @@ steps:
         - vllm/platforms/rocm.py
       commands:
         - uv pip install --system -r /vllm-workspace/requirements/kv_connectors_rocm.txt
-        - ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
+        - KV_CACHE_MEMORY_BYTES=8G ATTENTION_BACKEND=TRITON_ATTN bash v1/kv_connector/nixl_integration/config_sweep_spec_decode_test.sh
 
 - label: MultiConnector (Nixl+Offloading) PD edge cases (2 GPUs)
   key: multiconnector-nixl-offloading-pd-edge-cases-2-gpus
@@ -198,3 +203,25 @@ steps:
   commands:
     - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh
+
+# P TP 4 - D DPEP 4 test case for DSv4-Flash
+- label: DSv4-Flash Disaggregated DP EP
+  key: dsv4-flash-disaggregated
+  timeout_in_minutes: 60
+  device: h200
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 8
+  env:
+    ENABLE_HMA_FLAG: "1"
+    DP_EP: "1"
+    GPU_MEMORY_UTILIZATION: "0.85"
+    PREFILLER_TP_SIZE: "4"
+    DECODER_TP_SIZE: "4"
+    PREFILL_BLOCK_SIZE: "256"
+    DECODE_BLOCK_SIZE: "256"
+    MODEL_NAMES: "deepseek-ai/DeepSeek-V4-Flash"
+    VLLM_SERVE_EXTRA_ARGS: "--trust-remote-code,--kv-cache-dtype,fp8"
+  commands:
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+    - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh

.buildkite/test_areas/distributed.yaml

@@ -39,7 +39,9 @@ steps:
   - DP_SIZE=2 pytest -v -s entrypoints/openai/test_multi_api_servers.py
   mirror:
     amd:
+      dind: false
       device: mi300_2
+      timeout_in_minutes: 45
       depends_on:
       - image-build-amd
       source_file_dependencies:

.buildkite/test_areas/engine.yaml

@@ -28,8 +28,9 @@ steps:
   - 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: 50
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 40
       depends_on:
       - image-build-amd
 
@@ -44,14 +45,14 @@ steps:
     - pytest -v -s v1/engine --ignore v1/engine/test_preprocess_error_handling.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 55
+      device: mi250_1
+      timeout_in_minutes: 45
       depends_on:
       - image-build-amd
 
 - label: e2e Scheduling (1 GPU)
   key: e2e-scheduling-1-gpu
-  timeout_in_minutes: 35
+  timeout_in_minutes: 53
   device: h200_18gb
   source_file_dependencies:
     - vllm/v1/
@@ -60,8 +61,8 @@ steps:
     - pytest -v -s v1/e2e/general/test_async_scheduling.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 70
+      device: mi250_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
 
@@ -76,8 +77,8 @@ steps:
     - pytest -v -s v1/e2e/general --ignore v1/e2e/general/test_async_scheduling.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 60
+      device: mi250_1
+      timeout_in_minutes: 50
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -114,7 +115,9 @@ steps:
     - pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "tensor_parallelism"
   mirror:
     amd:
+      dind: false
       device: mi300_2
+      timeout_in_minutes: 30
       depends_on:
       - image-build-amd
 

.buildkite/test_areas/entrypoints.yaml

@@ -3,6 +3,7 @@ depends_on:
   - image-build
 steps:
 - label: Entrypoints Unit Tests
+  device: h200_35gb
   key: entrypoints-unit-tests
   timeout_in_minutes: 25
   working_dir: "/vllm-workspace/tests"
@@ -15,6 +16,7 @@ steps:
   - pytest -v -s entrypoints/weight_transfer
 
 - label: Entrypoints Integration (LLM)
+  device: h200_35gb
   key: entrypoints-integration-llm
   timeout_in_minutes: 60
   working_dir: "/vllm-workspace/tests"
@@ -28,16 +30,16 @@ steps:
   - pytest -v -s entrypoints/llm/offline_mode # Needs to avoid interference with other tests
   mirror:
     amd:
-      device: mi325_1
-      # TODO(akaratza): Test after Torch >= 2.12 bump
-      soft_fail: true
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (API Server)
   key: entrypoints-integration-api-server
   device: h200_35gb
-  timeout_in_minutes: 50
+  timeout_in_minutes: 75
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -50,13 +52,16 @@ steps:
   - pytest -v -s entrypoints/scale_out
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (API Server OpenAI - Part 1)
+  device: h200_35gb
   key: entrypoints-integration-api-server-openai-part-1
-  timeout_in_minutes: 45
+  timeout_in_minutes: 68
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -67,14 +72,16 @@ steps:
   - 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
+      dind: false
+      device: mi300_1
       timeout_in_minutes: 65
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (API Server OpenAI - Part 2)
+  device: h200_35gb
   key: entrypoints-integration-api-server-openai-part-2
-  timeout_in_minutes: 45
+  timeout_in_minutes: 83
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -86,12 +93,14 @@ steps:
   - pytest -v -s entrypoints/openai/completion --ignore=entrypoints/openai/completion/test_tensorizer_entrypoint.py
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 80
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 70
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (API Server Generate)
+  device: h200_35gb
   key: entrypoints-integration-api-server-generate
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
@@ -108,12 +117,14 @@ steps:
   - pytest -v -s entrypoints/anthropic
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
       timeout_in_minutes: 65
       depends_on:
       - image-build-amd
 
 - label: Entrypoints Integration (Responses API)
+  device: h200_35gb
   key: entrypoints-integration-responses-api
   timeout_in_minutes: 50
   working_dir: "/vllm-workspace/tests"
@@ -148,8 +159,9 @@ steps:
   - pytest -v -s entrypoints/multimodal
 
 - label: Entrypoints Integration (Pooling)
+  device: h200_35gb
   key: entrypoints-integration-pooling
-  timeout_in_minutes: 50
+  timeout_in_minutes: 75
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:
   - vllm/
@@ -169,7 +181,9 @@ steps:
   - pytest -s entrypoints/openai/correctness/
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 30
       depends_on:
       - image-build-amd
       source_file_dependencies:

.buildkite/test_areas/expert_parallelism.yaml

@@ -16,7 +16,9 @@ steps:
   - pytest -v -s distributed/test_eplb_utils.py
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 30
       depends_on:
       - image-build-amd
       source_file_dependencies:

.buildkite/test_areas/fault_tolerance.yaml

@@ -0,0 +1,26 @@
+group: Fault Tolerance
+depends_on:
+  - image-build
+steps:
+- label: Fault Tolerance E2E (2xH100)
+  key: fault-tolerance-e2e-2xh100
+  timeout_in_minutes: 35
+  device: h100
+  num_devices: 2
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/v1/fault_tolerance/
+  - vllm/v1/worker/sentinel/
+  - vllm/entrypoints/serve/fault_tolerance/
+  - vllm/distributed/elastic_ep/
+  - vllm/distributed/device_communicators/
+  - vllm/v1/engine/
+  - vllm/v1/worker/
+  - tests/v1/fault_tolerance/
+  - tests/v1/distributed/test_external_lb_dp.py
+  commands:
+  # Base image has no nixl; install it or has_nixl_ep() skips the tests.
+  - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+  # https://github.com/NVIDIA/nccl/issues/1838
+  - export NCCL_CUMEM_HOST_ENABLE=0
+  - pytest -v -s v1/fault_tolerance/test_fault_tolerance_e2e.py

.buildkite/test_areas/kernels.yaml

@@ -15,6 +15,7 @@ steps:
     - pytest -v -s tests/kernels/ir
 
 - label: Kernels Core Operation Test
+  device: h200_35gb
   key: kernels-core-operation-test
   timeout_in_minutes: 120
   source_file_dependencies:
@@ -79,7 +80,8 @@ steps:
   parallelism: 2
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
       timeout_in_minutes: 90
       depends_on:
       - image-build-amd
@@ -117,7 +119,9 @@ steps:
   parallelism: 2
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 120
       source_file_dependencies:
       - csrc/quantization/
       - vllm/model_executor/layers/quantization
@@ -147,8 +151,9 @@ steps:
   parallelism: 5
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 65
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
       source_file_dependencies:
       - csrc/quantization/cutlass_w8a8/moe/
       - csrc/moe/
@@ -163,6 +168,7 @@ steps:
       - image-build-amd
 
 - label: Kernels Mamba Test
+  device: h200_35gb
   key: kernels-mamba-test
   timeout_in_minutes: 40
   source_file_dependencies:
@@ -176,9 +182,9 @@ steps:
   timeout_in_minutes: 25
   device: h200_18gb
   source_file_dependencies:
-  - vllm/model_executor/layers/fla/ops/kda.py
-  - vllm/model_executor/layers/fla/ops/chunk_delta_h.py
-  - vllm/model_executor/layers/fla/ops/l2norm.py
+  - vllm/third_party/flash_linear_attention/ops/kda.py
+  - vllm/third_party/flash_linear_attention/ops/chunk_delta_h.py
+  - vllm/third_party/flash_linear_attention/ops/l2norm.py
   - tests/kernels/test_kda.py
   commands:
     - pytest -v -s kernels/test_kda.py
@@ -232,6 +238,15 @@ steps:
   - vllm/v1/attention/backends/mla/flashinfer_mla.py
   - vllm/v1/attention/selector.py
   - vllm/platforms/cuda.py
+  - vllm/model_executor/kernels/linear/cute_dsl/ll_bf16.py
+  - vllm/model_executor/kernels/linear/cute_dsl/_ll_bf16_dotprod.py
+  - vllm/model_executor/kernels/linear/cute_dsl/_ll_bf16_splitk.py
+  - vllm/cute_utils/
+  - vllm/model_executor/layers/mamba/ops/gdn_chunk_cutedsl/
+  - vllm/model_executor/layers/fused_moe/router/bf16x3_router_gemm_cutedsl.py
+  - tests/kernels/mamba/test_gdn_prefill_cutedsl.py
+  - tests/kernels/test_bf16x3_router_gemm_cutedsl.py
+  - tests/kernels/test_ll_bf16_gemm.py
   - tests/kernels/test_top_k_per_row.py
   commands:
     - nvidia-smi
@@ -260,6 +275,9 @@ steps:
     - pytest -v -s tests/kernels/moe/test_flashinfer_moe.py
     - pytest -v -s tests/kernels/moe/test_trtllm_nvfp4_moe.py
     - pytest -v -s tests/kernels/moe/test_cutedsl_moe.py
+    - pytest -v -s tests/kernels/mamba/test_gdn_prefill_cutedsl.py
+    - pytest -v -s tests/kernels/test_bf16x3_router_gemm_cutedsl.py
+    - pytest -v -s tests/kernels/test_ll_bf16_gemm.py
     # e2e
     - pytest -v -s tests/models/quantization/test_nvfp4.py
 

.buildkite/test_areas/lm_eval.yaml

@@ -14,8 +14,9 @@ steps:
   - 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
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 45
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -78,6 +79,28 @@ steps:
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-small-tp.txt
 
+- label: LM Eval PCP (4xB200)
+  key: lm-eval-pcp-4xb200
+  timeout_in_minutes: 360
+  device: b200-k8s
+  num_devices: 4
+  optional: true
+  source_file_dependencies:
+  - csrc/
+  - tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP2-PCP2-EP.yaml
+  - tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP1-PCP4-EP.yaml
+  - tests/evals/gsm8k/configs/models-pcp.txt
+  - vllm/model_executor/layers/quantization
+  - vllm/config/parallel.py
+  - vllm/distributed/parallel_state.py
+  - vllm/model_executor/layers/attention/mla_attention.py
+  - vllm/model_executor/layers/attention/pcp.py
+  - vllm/v1/worker/gpu/model_runner.py
+  - vllm/v1/worker/gpu/pcp_manager.py
+  autorun_on_main: true
+  commands:
+  - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-pcp.txt
+
 - label: LM Eval Large Models EP (2xB200)
   key: lm-eval-large-models-ep-2xb200
   timeout_in_minutes: 60
@@ -103,7 +126,7 @@ steps:
   - vllm/transformers_utils/configs/qwen3_5_moe.py
   - vllm/model_executor/models/qwen3_next.py
   - vllm/model_executor/models/qwen3_next_mtp.py
-  - vllm/model_executor/layers/fla/ops/
+  - vllm/third_party/flash_linear_attention/ops/
   commands:
   - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-qwen35-blackwell.txt
 
@@ -117,8 +140,9 @@ steps:
     - pytest -s -v evals/gsm8k/test_gsm8k_correctness.py --config-list-file=configs/models-h200.txt
   mirror:
     amd:
+      dind: false
       device: mi300_8
-      timeout_in_minutes: 60
+      timeout_in_minutes: 40
       depends_on:
       - image-build-amd
       commands:
@@ -313,7 +337,7 @@ steps:
 
 - label: LM Eval KV-Offload (2xH100)
   key: kv-offload-medium
-  timeout_in_minutes: 30
+  timeout_in_minutes: 45
   device: h100
   num_devices: 2
   source_file_dependencies:
@@ -323,7 +347,7 @@ steps:
   - vllm/v1/simple_kv_offload/
   - tests/evals/gsm8k/test_gsm8k_offloading.py
   commands:
-  - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "qwen3.5-35b"
+  - pytest -s -v evals/gsm8k/test_gsm8k_offloading.py -k "qwen3.5-35b or deepseek-v2-lite"
 
 - label: LM Eval KV-Offload (4xH100)
   key: kv-offload-large

.buildkite/test_areas/lora.yaml

@@ -14,9 +14,10 @@ steps:
   parallelism: 4
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
       working_dir: "/vllm-workspace/tests"
-      timeout_in_minutes: 65
+      timeout_in_minutes: 85
       source_file_dependencies:
       - vllm/lora
       - tests/lora
@@ -46,4 +47,4 @@ steps:
     - pytest -v -s -x lora/test_qwen3_with_multi_loras.py
     - pytest -v -s -x lora/test_olmoe_tp.py
     - pytest -v -s -x lora/test_gptoss_tp.py
-    - pytest -v -s -x lora/test_qwen35_densemodel_lora.py
+    - pytest -v -s -x lora/test_qwen35_densemodel_lora.py

.buildkite/test_areas/misc.yaml

@@ -23,14 +23,15 @@ steps:
     - pytest -v -s -m 'not slow_test' v1/spec_decode
   mirror:
     amd:
+      dind: false
       device: mi300_1
-      timeout_in_minutes: 75
+      timeout_in_minutes: 50
       depends_on:
       - image-build-amd
 
 - label: V1 Sample + Logits
   key: v1-sample-logits
-  timeout_in_minutes: 45
+  timeout_in_minutes: 83
   device: h200_18gb
   source_file_dependencies:
     - vllm/config/
@@ -58,13 +59,16 @@ steps:
     - pytest -v -s v1/test_outputs.py
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 70
       depends_on:
       - image-build-amd
 
 - label: V1 Core + KV + Metrics
+  device: h200_35gb
   key: v1-core-kv-metrics
-  timeout_in_minutes: 60
+  timeout_in_minutes: 80
   source_file_dependencies:
     - vllm/config/
     - vllm/distributed/
@@ -88,6 +92,7 @@ steps:
     - tests/v1/kv_offload
     - tests/v1/simple_kv_offload
     - tests/v1/worker
+    - tests/v1/streaming_input
     - tests/v1/kv_connector/unit
     - tests/v1/ec_connector/unit
     - tests/v1/metrics
@@ -101,6 +106,7 @@ steps:
     - pytest -v -s v1/kv_offload
     - pytest -v -s v1/simple_kv_offload
     - pytest -v -s v1/worker
+    - pytest -v -s v1/streaming_input
     - pytest -v -s -m 'not cpu_test' v1/kv_connector/unit
     - pytest -v -s -m 'not cpu_test' v1/ec_connector/unit
     - pytest -v -s -m 'not cpu_test' v1/metrics
@@ -109,8 +115,9 @@ steps:
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 75
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
       depends_on:
       - image-build-amd
 
@@ -141,6 +148,7 @@ steps:
     - pytest -v -s -m 'cpu_test' v1/core
     - pytest -v -s v1/structured_output
     - pytest -v -s v1/test_serial_utils.py
+    - pytest -v -s v1/cudagraph/test_cudagraph_manager.py
     - pytest -v -s -m 'cpu_test' v1/kv_connector/unit
     - pytest -v -s -m 'cpu_test' v1/metrics
 
@@ -204,7 +212,7 @@ steps:
   - vllm/multimodal
   - examples/
   commands:
-    - pip install tensorizer # for tensorizer test
+    - pip install --no-deps tensorizer # for tensorizer test
      # for basic
     - python3 basic/offline_inference/chat.py
     - python3 basic/offline_inference/generate.py --model facebook/opt-125m
@@ -228,7 +236,9 @@ steps:
     - python3 features/speculative_decoding/spec_decode_offline.py --test --method eagle3 --num_spec_tokens 3 --dataset-name hf --dataset-path philschmid/mt-bench --num-prompts 80 --temp 0 --top-p 1.0 --top-k -1 --tp 1 --enable-chunked-prefill --max-model-len 1536
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 75
       source_file_dependencies:
       - vllm/entrypoints
       - vllm/multimodal
@@ -264,10 +274,11 @@ steps:
   - pytest -v -s v1/tracing
   mirror:
     amd:
-      device: mi325_2
+      dind: false
+      device: mi300_2
+      timeout_in_minutes: 30
       depends_on:
       - image-build-amd
-      optional: true
 
 - label: Python-only Installation
   key: python-only-installation
@@ -282,8 +293,8 @@ steps:
   - bash standalone_tests/python_only_compile.sh
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 45
+      device: mi250_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
       source_file_dependencies:

.buildkite/test_areas/model_executor.yaml

@@ -3,13 +3,16 @@ depends_on:
   - image-build
 steps:
 - label: Model Executor
+  device: h200_35gb
   key: model-executor
-  timeout_in_minutes: 45
+  timeout_in_minutes: 60
   source_file_dependencies:
   - vllm/engine/arg_utils.py
   - vllm/config/model.py
   - vllm/model_executor
+  - vllm/model_executor/warmup
   - tests/model_executor
+  - tests/model_executor/test_jit_warmup.py
   - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
   commands:
     - apt-get update && apt-get install -y curl libsodium23
@@ -25,14 +28,18 @@ steps:
     - pytest -v -s entrypoints/openai/completion/test_tensorizer_entrypoint.py --timeout=900 --timeout-method=thread
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 60
       depends_on:
       - image-build-amd
       source_file_dependencies:
       - vllm/engine/arg_utils.py
       - vllm/config/model.py
       - vllm/model_executor
+      - vllm/model_executor/warmup
       - tests/model_executor
+      - tests/model_executor/test_jit_warmup.py
       - tests/entrypoints/openai/completion/test_tensorizer_entrypoint.py
       - vllm/_aiter_ops.py
       - vllm/platforms/rocm.py

.buildkite/test_areas/model_runner_v2.yaml

@@ -41,7 +41,7 @@ steps:
   commands:
     - set -x
     - export VLLM_USE_V2_MODEL_RUNNER=1
-    - pip install tensorizer # for tensorizer test
+    - pip install --no-deps tensorizer # for tensorizer test
     - python3 basic/offline_inference/chat.py # for basic
     - python3 basic/offline_inference/generate.py --model facebook/opt-125m
     #- python3 basic/offline_inference/generate.py --model meta-llama/Llama-2-13b-chat-hf --cpu-offload-gb 10  # TODO

.buildkite/test_areas/models_basic.yaml

@@ -42,10 +42,25 @@ steps:
     - pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 50
       depends_on:
       - image-build-amd
 
+- label: Inkling Unit Tests (B200)
+  key: inkling-unit-tests-b200
+  timeout_in_minutes: 40
+  device: b200-k8s
+  source_file_dependencies:
+  - vllm/models/inkling/
+  - vllm/cute_utils/
+  - cmake/external_projects/tml_fa4.cmake
+  - tests/models/inkling/
+  commands:
+    # FA4 kernel tests require SM100; the suite skips them elsewhere.
+    - pytest -v -s models/inkling
+
 - label: Basic Models Test (Other CPU) # 5min
   key: basic-models-test-other-cpu
   depends_on:

.buildkite/test_areas/models_language.yaml

@@ -15,11 +15,14 @@ steps:
     - pytest -v -s models/language -m 'core_model and (not slow_test)'
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 45
       depends_on:
       - image-build-amd
 
 - label: Language Models Tests (Extra Standard) %N
+  device: h200_35gb
   key: language-models-tests-extra-standard
   timeout_in_minutes: 40
   source_file_dependencies:
@@ -35,7 +38,9 @@ steps:
   parallelism: 2
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 40
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -49,8 +54,8 @@ steps:
       - tests/models/language/pooling/test_classification.py
       - vllm/_aiter_ops.py
       - vllm/platforms/rocm.py
-
 - label: Language Models Tests (Hybrid) %N
+  device: h200_35gb
   key: language-models-tests-hybrid
   timeout_in_minutes: 65
   source_file_dependencies:
@@ -58,16 +63,16 @@ steps:
   - tests/models/language/generation
   commands:
     # Install fast path packages for testing against transformers
-    # Note: also needed to run plamo2 model in vLLM
     - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
-    # Shard hybrid language model tests
-    - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
+    # Shard the hybrid language model tests that are numerically stable on Hopper.
+    - pytest -v -s models/language/generation -m hybrid_model -k 'not granite-4.0-tiny-preview' --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
   parallelism: 2
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 70
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 60
       depends_on:
       - image-build-amd
       commands:
@@ -75,6 +80,20 @@ steps:
       - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
       - pytest -v -s models/language/generation -m hybrid_model --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --shard-id=$$BUILDKITE_PARALLEL_JOB
 
+# Granite 4 hybrid generation is sensitive to hardware-specific Triton SSD
+# autotuning (https://github.com/vllm-project/vllm/issues/25194). Keep this one
+# correctness test on L4 until its H200 output matches the Transformers reference.
+- label: Language Models Tests (Granite L4 Compatibility)
+  key: language-models-tests-granite-l4-compatibility
+  timeout_in_minutes: 65
+  source_file_dependencies:
+  - vllm/
+  - tests/models/language/generation
+  commands:
+    - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
+    - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
+    - pytest -v -s models/language/generation -m hybrid_model -k 'granite-4.0-tiny-preview'
+
 - label: Language Models Test (Extended Generation) # 80min
   device: h200_35gb
   key: language-models-test-extended-generation
@@ -85,7 +104,6 @@ steps:
   - tests/models/language/generation
   commands:
     # Install fast path packages for testing against transformers
-    # Note: also needed to run plamo2 model in vLLM
     - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.3.0'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.6.0'
     - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
@@ -101,10 +119,10 @@ steps:
   commands:
     - pytest -v -s models/language/generation_ppl_test
 
-- label: Language Models Test (Extended Pooling)  # 36min
+- label: Language Models Test (Extended Pooling)
   device: h200_35gb
   key: language-models-test-extended-pooling
-  timeout_in_minutes: 70
+  timeout_in_minutes: 120
   optional: true
   source_file_dependencies:
   - vllm/
@@ -113,14 +131,15 @@ steps:
     - pytest -v -s models/language/pooling -m 'not core_model'
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 100
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 95
       depends_on:
       - image-build-amd
 
 - label: Language Models Test (MTEB)
   key: language-models-test-mteb
-  timeout_in_minutes: 45
+  timeout_in_minutes: 68
   device: h200_18gb
   optional: true
   source_file_dependencies:

.buildkite/test_areas/models_multimodal.yaml

@@ -4,7 +4,7 @@ depends_on:
 steps:
 - label: "Multi-Modal Models (Standard) 1: qwen2"
   key: multi-modal-models-standard-1-qwen2
-  timeout_in_minutes: 45
+  timeout_in_minutes: 68
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -14,13 +14,15 @@ steps:
     - pytest -v -s models/multimodal/generation/test_ultravox.py -m core_model
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 65
       depends_on:
       - image-build-amd
 
 - label: "Multi-Modal Models (Standard) 2: qwen3 + gemma"
   key: multi-modal-models-standard-2-qwen3-gemma
-  timeout_in_minutes: 50
+  timeout_in_minutes: 75
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -31,7 +33,9 @@ steps:
     - pytest -v -s models/multimodal/generation/test_qwen2_5_vl.py -m core_model
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
 
@@ -47,14 +51,15 @@ steps:
     - pytest -v -s models/multimodal/generation/test_qwen2_vl.py -m core_model
   mirror:
     amd:
-      device: mi325_1
+      device: mi250_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
 
 - label: "Multi-Modal Models (Standard) 4: other + whisper"
   device: h200_35gb
   key: multi-modal-models-standard-4-other-whisper
-  timeout_in_minutes: 50
+  timeout_in_minutes: 75
   source_file_dependencies:
   - vllm/
   - tests/models/multimodal
@@ -65,7 +70,9 @@ steps:
     - cd .. && VLLM_WORKER_MULTIPROC_METHOD=spawn pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 50
       depends_on:
       - image-build-amd
 
@@ -85,7 +92,7 @@ steps:
 
 - label: Multi-Modal Processor # 44min
   key: multi-modal-processor
-  timeout_in_minutes: 65
+  timeout_in_minutes: 98
   device: h200_18gb
   source_file_dependencies:
   - vllm/
@@ -107,7 +114,9 @@ steps:
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-mm-small.txt --tp-size=1
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 35
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -118,6 +127,7 @@ steps:
       - vllm/model_executor/model_loader/
 
 - label: Multi-Modal Models (Extended Generation 1)
+  device: h200_35gb
   key: multi-modal-models-extended-generation-1
   optional: true
   source_file_dependencies:
@@ -129,7 +139,9 @@ steps:
     - pytest -v -s models/multimodal/test_mapping.py
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 90
       depends_on:
       - image-build-amd
 
@@ -164,8 +176,9 @@ steps:
     - pytest -v -s models/multimodal/pooling -m 'not core_model'
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 75
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 60
       depends_on:
       - image-build-amd
       source_file_dependencies:

.buildkite/test_areas/pytorch.yaml

@@ -5,7 +5,7 @@ steps:
 - label: PyTorch Compilation Unit Tests
   device: h200_35gb
   key: pytorch-compilation-unit-tests
-  timeout_in_minutes: 90
+  timeout_in_minutes: 150
   source_file_dependencies:
     - vllm/__init__.py
     - vllm/_aiter_ops.py
@@ -107,16 +107,11 @@ steps:
     - tests/compile/passes
   commands:
   - pytest -s -v compile/passes --ignore compile/passes/distributed
-  mirror:
-    amd:
-      device: mi300_1
-      timeout_in_minutes: 65
-      depends_on:
-      - image-build-amd
 
 - label: PyTorch Fullgraph Smoke Test
+  device: h200_35gb
   key: pytorch-fullgraph-smoke-test
-  timeout_in_minutes: 60
+  timeout_in_minutes: 90
   source_file_dependencies:
   - vllm/__init__.py
   - vllm/_aiter_ops.py
@@ -148,7 +143,42 @@ steps:
   # as it is a heavy test that is covered in other steps.
   # Use `find` to launch multiple instances of pytest so that
   # they do not suffer from https://github.com/vllm-project/vllm/issues/28965
-  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_graph.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
+  - "find compile/fullgraph/ -name 'test_*.py' -not -name 'test_full_cudagraph.py' -not -name 'test_full_graph.py' -print0 | xargs -0 -n1 -I{} pytest -s -v '{}'"
+
+# Hopper-only DeepSeek-V2-Lite cases in this file require two 29.3-GiB model
+# instances and cannot fit a 35GB MIG slice. L4 retains the original coverage:
+# those SM90 cases skip while the architecture-compatible cases still run.
+- label: PyTorch Fullgraph CUDAGraph (L4 Compatibility)
+  key: pytorch-fullgraph-cudagraph-l4-compatibility
+  timeout_in_minutes: 60
+  source_file_dependencies:
+  - vllm/__init__.py
+  - vllm/_aiter_ops.py
+  - vllm/_custom_ops.py
+  - vllm/compilation/
+  - vllm/config/
+  - vllm/distributed/
+  - vllm/engine/
+  - vllm/env_override.py
+  - vllm/envs.py
+  - vllm/forward_context.py
+  - vllm/inputs/
+  - vllm/ir/
+  - vllm/kernels/
+  - vllm/logger.py
+  - vllm/model_executor/
+  - vllm/multimodal/
+  - vllm/platforms/
+  - vllm/plugins/
+  - vllm/sampling_params.py
+  - vllm/sequence.py
+  - vllm/transformers_utils/
+  - vllm/triton_utils/
+  - vllm/utils/
+  - vllm/v1/
+  - tests/compile
+  commands:
+  - pytest -s -v compile/fullgraph/test_full_cudagraph.py
 
 - label: PyTorch Fullgraph
   key: pytorch-fullgraph
@@ -197,7 +227,9 @@ steps:
   - bash standalone_tests/pytorch_nightly_dependency.sh
   mirror:
     amd:
+      dind: false
       device: mi300_1
+      timeout_in_minutes: 30
       depends_on:
       - image-build-amd
       source_file_dependencies:

.buildkite/test_areas/quantization.yaml

@@ -3,8 +3,11 @@ depends_on:
   - image-build
 steps:
 - label: Quantization
+  device: h200_35gb
   key: quantization
-  timeout_in_minutes: 60
+  timeout_in_minutes: 75
+  env:
+    VLLM_USE_V2_MODEL_RUNNER: "0"
   source_file_dependencies:
   - csrc/
   - vllm/model_executor/layers/quantization
@@ -19,9 +22,12 @@ steps:
   # TODO(jerryzh168): resolve the above comment
   - uv pip install --system torchao==0.17.0 --index-url https://download.pytorch.org/whl/cu130
   - uv pip install --system conch-triton-kernels
-  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py
+  # The SM90-only checkpoint currently contains a removed weight_chan_scale
+  # parameter. It was not exercised by the previous L4 job.
+  - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization/ --ignore quantization/test_blackwell_moe.py -k 'not test_compressed_tensors_w4a8_fp8'
 
 - label: Quantized Fusions
+  device: h200_35gb
   key: quantized-fusions
   timeout_in_minutes: 20
   source_file_dependencies:
@@ -52,8 +58,11 @@ steps:
     - pytest -s -v tests/quantization/test_blackwell_moe.py
 
 - label: Quantized Models Test
+  device: h200_35gb
   key: quantized-models-test
-  timeout_in_minutes: 50
+  timeout_in_minutes: 65
+  env:
+    VLLM_USE_V2_MODEL_RUNNER: "0"
   source_file_dependencies:
   - vllm/model_executor/layers/quantization
   - tests/models/quantization

.buildkite/test_areas/rust_frontend.yaml

@@ -81,6 +81,7 @@ steps:
   - pytest -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
 
 - label: Rust Frontend Tool Use
+  device: h200_35gb
   timeout_in_minutes: 25
   working_dir: "/vllm-workspace/tests"
   source_file_dependencies:

.buildkite/test_areas/samplers.yaml

@@ -19,8 +19,18 @@ steps:
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
   mirror:
     amd:
-      device: mi325_1
+      device: mi250_1
+      timeout_in_minutes: 40
       depends_on:
       - image-build-amd
+      source_file_dependencies:
+      - vllm/model_executor/layers
+      - vllm/sampling_metadata.py
+      - vllm/v1/sample/
+      - vllm/entrypoints/generate/beam_search/
+      - tests/samplers
+      - tests/conftest.py
+      - vllm/_aiter_ops.py
+      - vllm/platforms/rocm.py
       commands:
       - pytest -v -s samplers

.buildkite/test_areas/spec_decode.yaml

@@ -14,8 +14,9 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "eagle_correctness"
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 60
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 55
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -53,8 +54,9 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "speculators or mtp_correctness"
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 65
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 75
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -92,10 +94,9 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "ngram or suffix"
   mirror:
     amd:
-      device: mi325_1
-      timeout_in_minutes: 55
-      # TODO(akaratza): Test after Torch >= 2.12 bump
-      soft_fail: true
+      dind: false
+      device: mi300_1
+      timeout_in_minutes: 35
       depends_on:
       - image-build-amd
       source_file_dependencies:
@@ -119,7 +120,8 @@ steps:
     - pytest -v -s v1/e2e/spec_decode -k "draft_model or no_sync or batch_inference"
   mirror:
     amd:
-      device: mi325_1
+      dind: false
+      device: mi300_1
       timeout_in_minutes: 55
       depends_on:
       - image-build-amd
@@ -170,3 +172,19 @@ steps:
     - tests/v1/e2e/spec_decode/
   commands:
     - pytest -v -s v1/e2e/spec_decode -k "qwen3_5-hybrid"
+
+- label: Spec Decode DeepSeek MTP Parallel Load (B200)
+  key: spec-decode-deepseek-mtp-parallel-load-b200
+  timeout_in_minutes: 30
+  device: b200-k8s
+  optional: true
+  num_devices: 2
+  source_file_dependencies:
+    - vllm/v1/spec_decode/llm_base_proposer.py
+    - vllm/v1/spec_decode/eagle.py
+    - vllm/v1/worker/gpu/spec_decode/eagle/
+    - vllm/model_executor/models/deepseek_mtp.py
+    - vllm/model_executor/models/deepseek_v2.py
+    - tests/v1/e2e/spec_decode/test_mtp_parallel_load.py
+  commands:
+    - pytest -v -s v1/e2e/spec_decode/test_mtp_parallel_load.py

.buildkite/test_areas/weight_loading.yaml

@@ -15,7 +15,9 @@ steps:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
   mirror:
     amd:
+      dind: false
       device: mi300_2
+      timeout_in_minutes: 35
       depends_on:
       - image-build-amd
       commands:

.github/CODEOWNERS

@@ -47,6 +47,7 @@
 
 # Rust Frontend
 /rust/ @BugenZhao @njhill
+/rust/src/bench @esmeetu
 /build_rust.sh @BugenZhao @njhill
 /rust-toolchain.toml @BugenZhao @njhill
 /.buildkite/test_areas/rust* @BugenZhao @njhill
@@ -172,7 +173,7 @@ mkdocs.yaml @hmellor
 # Kernels
 /vllm/v1/attention/ops/chunked_prefill_paged_decode.py @tdoublep
 /vllm/v1/attention/ops/triton_unified_attention.py @tdoublep
-/vllm/model_executor/layers/fla @ZJY0516 @vadiklyutiy
+/vllm/third_party/flash_linear_attention @ZJY0516 @vadiklyutiy
 
 # ROCm related: specify owner with write access to notify AMD folks for careful code review
 /vllm/**/*rocm* @tjtanaa @dllehr-amd

.github/mergify.yml

@@ -19,6 +19,7 @@ pull_request_rules:
   description: Comment on PR when pre-commit check fails
   conditions:
     - check-failure=pre-commit
+    - -check-cancelled=pre-commit
     - -closed
     - -draft
     - or:
@@ -181,6 +182,18 @@ pull_request_rules:
       add:
         - performance
 
+- name: label-quantization
+  description: Automatically apply quantization label
+  conditions:
+    - label != stale
+    - or:
+      - files~=^vllm/model_executor/layers/quantization/
+      - title~=(?i)quant
+  actions:
+    label:
+      add:
+        - quantization
+
 - name: label-qwen
   description: Automatically apply qwen label
   conditions:
@@ -220,6 +233,31 @@ pull_request_rules:
       add:
         - gpt-oss
 
+- name: label-kimi
+  description: Automatically apply kimi label
+  conditions:
+    - label != stale
+    - or:
+      - files~=(?i)kimi
+      - files~=(?i)moonshot
+      - title~=(?i)(?:kimi|moonshot)
+  actions:
+    label:
+      add:
+        - kimi
+
+- name: label-k3
+  description: Automatically apply k3 label (launch triage; retire after ramp-down)
+  conditions:
+    - label != stale
+    - or:
+      - files~=(?i)kimi[-_]?k3
+      - title~=(?i)(?:kimi[-\s]?k3|\bk3\b)
+  actions:
+    label:
+      add:
+        - k3
+
 - name: label-nvidia
   description: Automatically apply nvidia label
   conditions:

.github/workflows/issue_autolabel.yml

@@ -130,6 +130,66 @@ jobs:
                   },
                 ],
               },
+              kimi: {
+                keywords: [
+                  { term: "Kimi",     searchIn: "both" },
+                  { term: "Moonshot", searchIn: "both" },
+                ],
+                substrings: [
+                  { term: "moonshotai/", searchIn: "both" },
+                  { term: "kimi",        searchIn: "title" },
+                ],
+              },
+              k3: {
+                keywords: [
+                  { term: "Kimi K3", searchIn: "both" },
+                  { term: "K3",      searchIn: "title" },
+                ],
+                substrings: [
+                  { term: "moonshotai/kimi-k3", searchIn: "both" },
+                ],
+              },
+              quantization: {
+                keywords: [
+                  {
+                    term: "quantization",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "quantized",
+                    searchIn: "both"
+                  },
+                ],
+              },
+              "intel-gpu": {
+                // Keyword search - matches whole words only (with word boundaries)
+                keywords: [
+                  {
+                    term: "B50",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "B60",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "B70",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "intel gpu",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "Arc GPU",
+                    searchIn: "both"
+                  },
+                  {
+                    term: "BMG",
+                    searchIn: "both"
+                  },
+                ],
+              },
               // Add more label configurations here as needed
               // example: {
               //   keywords: [...],
@@ -323,7 +383,7 @@ jobs:
             // {users} will be replaced with @mentions
             const ccConfig = {
               rocm: {
-                users: ['hongxiayang', 'tjtanaa', 'vllmellm'],
+                users: ['hongxiayang', 'tjtanaa', 'vllmellm', 'giuseppegrossi'],
                 message: 'CC {users} for ROCm-related issue',
               },
               mistral: {
@@ -491,4 +551,4 @@ jobs:
               issue_number: context.issue.number,
               body: message,
             });
-            core.notice(`Requested missing ROCm info from @${author}: ${missing.map(m => m.name).join(', ')}`);
+            core.notice(`Requested missing ROCm info from @${author}: ${missing.map(m => m.name).join(', ')}`);

.gitignore

@@ -18,6 +18,9 @@ vllm/third_party/deep_gemm/
 # fmha_sm100 vendored package built from source
 vllm/third_party/fmha_sm100/
 
+# tml-fa4 vendored package built from source
+vllm/third_party/tml_fa4/
+
 # triton jit
 .triton
 
@@ -170,9 +173,6 @@ venv.bak/
 
 # mkdocs documentation
 /site
-docs/argparse
-docs/examples/*
-!docs/examples/README.md
 
 # mypy
 .mypy_cache/

.markdownlint.yaml

@@ -3,6 +3,9 @@ MD007:
 MD013: false
 MD024:
   siblings_only: true
+MD025:
+  # Allow front matter title to be different from the first heading in the document.
+  front_matter_title: ""
 MD031:
   list_items: false
 MD033: false

.pre-commit-config.yaml

@@ -30,7 +30,7 @@ repos:
   - id: markdownlint-cli2
     language_version: lts
     args: [--fix]
-    exclude: ^CLAUDE\.md$
+    exclude: (^|/)CLAUDE\.md$
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:
@@ -210,7 +210,7 @@ repos:
     name: Check SPDX headers
     entry: python tools/pre_commit/check_spdx_header.py
     language: python
-    types: [python]
+    types_or: [python, rust, proto]
   - id: check-root-lazy-imports
     name: Check root lazy imports
     entry: python tools/pre_commit/check_init_lazy_imports.py
@@ -260,10 +260,6 @@ repos:
     files: ^docker/(Dockerfile|versions\.json)$
     pass_filenames: false
     additional_dependencies: [dockerfile-parse]
-  - id: attention-backend-docs
-    name: Check attention backend documentation is up to date
-    entry: python tools/pre_commit/generate_attention_backend_docs.py --check
-    language: python
   - id: check-boolean-context-manager
     name: Check for boolean ops in with-statements
     entry: python tools/pre_commit/check_boolean_context_manager.py

.yapfignore

@@ -1,2 +0,0 @@
-collect_env.py
-vllm/model_executor/layers/fla/ops/*.py

CMakeLists.txt

@@ -68,8 +68,8 @@ endif()
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.11.0")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.11.0")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.13.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.13.0")
 # TORCH_NIGHTLY=1 builds run against unpinned nightly wheels, so the supported-
 # version check would always warn. Only treat it as a nightly build when the
 # value is exactly "1" (the bootstrap exports TORCH_NIGHTLY=0 by default, which
@@ -114,6 +114,11 @@ find_package(Torch REQUIRED)
 # Supported NVIDIA architectures.
 # This check must happen after find_package(Torch) because that's when CMAKE_CUDA_COMPILER_VERSION gets defined
 if(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
+   CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.4)
+  # Rubin (10.7) can run SM100 family code, but CUDA 13.4 also supports
+  # targeting it directly.
+  set(CUDA_SUPPORTED_ARCHS "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.7;11.0;12.0")
+elseif(DEFINED CMAKE_CUDA_COMPILER_VERSION AND
    CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 13.0)
    # starting from CUDA 12.9 and Blackwell (10.0), we use family-specific targets (10.0f, 12.0f, etc)
    # to support the whole generation without specifying all sub-architectures
@@ -420,7 +425,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
 
     if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
       cuda_archs_loose_intersection(COOPERATIVE_TOPK_ARCHS
-        "9.0a;10.0f;10.1f;10.3f;11.0f;12.0f;12.1f" "${CUDA_ARCHS}")
+        "9.0a;10.0f;10.1f;10.3f;10.7f;11.0f;12.0f;12.1f" "${CUDA_ARCHS}")
     else()
       cuda_archs_loose_intersection(COOPERATIVE_TOPK_ARCHS
         "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
@@ -695,7 +700,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
 
   # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;10.7f;11.0f;12.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
   endif()
@@ -815,7 +820,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
   # require CUDA 12.8 or later
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
   endif()
@@ -899,7 +904,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   endif()
 
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
   endif()
@@ -924,7 +929,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
 
   # moe_data.cu is used by all CUTLASS MoE kernels.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;11.0f;12.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0f;10.7f;11.0f;12.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${CUDA_ARCHS}")
   endif()
@@ -981,7 +986,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   # SM10x/11x FP4 kernels. MXFP4 experts quantization is currently compiled
   # only in this block; SM12x has separate NVFP4 matmul/MoE kernels above.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(FP4_SM100_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
   endif()
@@ -1047,7 +1052,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   # Runtime dispatch is gated in
   # vllm/v1/attention/backends/mla/cutlass_mla.py.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(MLA_ARCHS "10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(MLA_ARCHS "10.0f;10.7f;11.0f" "${CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(MLA_ARCHS "10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
   endif()
@@ -1364,6 +1369,7 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
   set(VLLM_ROCM_EXT_SRC
     "csrc/rocm/torch_bindings.cpp"
     "csrc/rocm/skinny_gemms.cu"
+    "csrc/rocm/skinny_gemms_int4.cu"
     "csrc/rocm/attention.cu")
 
   set(VLLM_ROCM_HAS_GFX1100 OFF)
@@ -1407,6 +1413,7 @@ if (VLLM_GPU_LANG STREQUAL "CUDA")
     include(cmake/external_projects/fmha_sm100.cmake)
     include(cmake/external_projects/flashmla.cmake)
     include(cmake/external_projects/qutlass.cmake)
+    include(cmake/external_projects/tml_fa4.cmake)
 
     # vllm-flash-attn should be last as it overwrites some CMake functions
     include(cmake/external_projects/vllm_flash_attn.cmake)

README.md

@@ -48,7 +48,7 @@ vLLM is flexible and easy to use with:
 - Tool calling and reasoning parsers
 - OpenAI-compatible API server, plus Anthropic Messages API and gRPC support
 - Efficient multi-LoRA support for dense and MoE layers
-- Support for NVIDIA GPUs, AMD GPUs, and x86/ARM/PowerPC CPUs. Additionally, diverse hardware plugins such as Google TPUs, Intel Gaudi, IBM Spyre, Huawei Ascend, Rebellions NPU, Apple Silicon, MetaX GPU, and more.
+- Support for NVIDIA GPUs, AMD GPUs, Intel GPUs, and x86/ARM/PowerPC CPUs. Additionally, diverse hardware plugins such as Google TPUs, Intel Gaudi, IBM Spyre, Huawei Ascend, Rebellions NPU, Apple Silicon, MetaX GPU, and more.
 
 vLLM seamlessly supports 200+ model architectures on Hugging Face, including:
 

benchmarks/attention_benchmarks/benchmark.py

@@ -75,7 +75,11 @@ def run_mla_benchmark(config: BenchmarkConfig, **kwargs) -> BenchmarkResult:
     from mla_runner import run_mla_benchmark as run_mla
 
     return run_mla(
-        config.backend, config, prefill_backend=config.prefill_backend, **kwargs
+        config.backend,
+        config,
+        prefill_backend=config.prefill_backend,
+        sparse_mla_force_mqa=config.sparse_mla_force_mqa,
+        **kwargs,
     )
 
 
@@ -592,6 +596,30 @@ def main():
         default="profile",
         help="Output file name for ncu profile (default: 'profile').",
     )
+    parser.add_argument(
+        "--torch-profile",
+        action="store_true",
+        default=False,
+        help="Collect a PyTorch profiler Chrome trace for each benchmark run.",
+    )
+    parser.add_argument(
+        "--torch-profile-dir",
+        default=None,
+        help="Directory for PyTorch profiler traces.",
+    )
+    parser.add_argument(
+        "--torch-profile-iters",
+        type=int,
+        default=3,
+        help="Number of forward passes to record per PyTorch profiler trace.",
+    )
+    parser.add_argument(
+        "--sparse-mla-mha-variants",
+        nargs="+",
+        default=None,
+        choices=["dense_mha", "mqa"],
+        help="Sparse MLA variants to run in mha_vs_mqa mode. Defaults to both.",
+    )
 
     # Parameter sweep (use YAML config for advanced sweeps)
     parser.add_argument(
@@ -641,6 +669,7 @@ def main():
 
         # Prefill backends (e.g., ["fa3", "fa4"])
         args.prefill_backends = yaml_config.get("prefill_backends", None)
+        args.prefill_backend = yaml_config.get("prefill_backend", None)
 
         # FP8 output benchmark knobs; CLI wins.
         if args.fp8_output_scale is None:
@@ -683,6 +712,9 @@ def main():
             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)
+            args.max_model_len = model.get(
+                "max_model_len", getattr(args, "max_model_len", None)
+            )
             # 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)
@@ -701,6 +733,21 @@ def main():
             args.cuda_graphs = yaml_config["cuda_graphs"]
         if "ncu_profile" in yaml_config:
             args.ncu_profile = yaml_config["ncu_profile"]
+        if "torch_profile" in yaml_config:
+            args.torch_profile = yaml_config["torch_profile"]
+        if "torch_profile_dir" in yaml_config:
+            args.torch_profile_dir = yaml_config["torch_profile_dir"]
+        if "torch_profile_iters" in yaml_config:
+            args.torch_profile_iters = yaml_config["torch_profile_iters"]
+        args.sparse_mla_topk_pattern = yaml_config.get(
+            "sparse_mla_topk_pattern", "random"
+        )
+        args.sparse_mla_dense_mha_max_seq_len = yaml_config.get(
+            "sparse_mla_dense_mha_max_seq_len", None
+        )
+        args.sparse_mla_mha_variants = yaml_config.get(
+            "sparse_mla_mha_variants", args.sparse_mla_mha_variants
+        )
 
         # Parameter sweep configuration
         if "parameter_sweep" in yaml_config:
@@ -842,8 +889,6 @@ 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,
@@ -1063,6 +1108,133 @@ def main():
                     f"\n  [yellow]Prefill always faster for batch_size={bs}[/]"
                 )
 
+    # Handle MHA vs MQA comparison mode for sparse MLA
+    elif hasattr(args, "mode") and args.mode == "mha_vs_mqa":
+        console.print("[yellow]Mode: MHA vs MQA comparison for sparse MLA[/]")
+
+        sparse_mla_topk_pattern = getattr(args, "sparse_mla_topk_pattern", "random")
+        dense_mha_max_seq_len = getattr(args, "sparse_mla_dense_mha_max_seq_len", None)
+        prefill_backend = getattr(args, "prefill_backend", None)
+        if prefill_backend:
+            console.print(f"Prefill backend: {prefill_backend}")
+        available_variants = [
+            ("dense_mha", False, "dense"),
+            ("mqa", True, "auto"),
+        ]
+        requested_variants = getattr(args, "sparse_mla_mha_variants", None)
+        if requested_variants is not None:
+            valid_variants = {label for label, _, _ in available_variants}
+            invalid_variants = sorted(set(requested_variants) - valid_variants)
+            if invalid_variants:
+                raise ValueError(
+                    "Invalid sparse_mla_mha_variants entries: "
+                    f"{invalid_variants}. Valid variants are: "
+                    f"{sorted(valid_variants)}"
+                )
+            requested_variant_set = set(requested_variants)
+            variants = [
+                variant
+                for variant in available_variants
+                if variant[0] in requested_variant_set
+            ]
+        else:
+            variants = available_variants
+        formatter = ResultsFormatter(console)
+        total = 0
+        for spec in args.batch_specs:
+            q_len = max(request.q_len for request in parse_batch_spec(spec))
+            for variant_label, _, _ in variants:
+                if (
+                    variant_label == "dense_mha"
+                    and dense_mha_max_seq_len is not None
+                    and q_len > dense_mha_max_seq_len
+                ):
+                    continue
+                total += len(backends)
+
+        with tqdm(total=total, desc="Benchmarking") as pbar:
+            for spec in args.batch_specs:
+                q_len = max(request.q_len for request in parse_batch_spec(spec))
+                for backend in backends:
+                    for variant_label, force_mqa, mha_mode in variants:
+                        if (
+                            variant_label == "dense_mha"
+                            and dense_mha_max_seq_len is not None
+                            and q_len > dense_mha_max_seq_len
+                        ):
+                            continue
+                        config = BenchmarkConfig(
+                            backend=f"{backend}_{variant_label}",
+                            batch_spec=spec,
+                            num_layers=args.num_layers,
+                            head_dim=args.head_dim,
+                            num_q_heads=args.num_q_heads,
+                            num_kv_heads=args.num_kv_heads,
+                            block_size=args.block_size,
+                            device=args.device,
+                            max_model_len=getattr(args, "max_model_len", None),
+                            kv_cache_dtype=args.kv_cache_dtype,
+                            profile_memory=args.profile_memory,
+                            use_cuda_graphs=args.cuda_graphs,
+                            ncu_profile=args.ncu_profile,
+                            torch_profile=args.torch_profile,
+                            torch_profile_dir=args.torch_profile_dir,
+                            torch_profile_iters=args.torch_profile_iters,
+                            warmup_ms=args.warmup_ms,
+                            kv_lora_rank=getattr(args, "kv_lora_rank", None),
+                            qk_nope_head_dim=getattr(args, "qk_nope_head_dim", None),
+                            qk_rope_head_dim=getattr(args, "qk_rope_head_dim", None),
+                            v_head_dim=getattr(args, "v_head_dim", None),
+                            sparse_mla_force_mqa=force_mqa,
+                            sparse_mla_mha_mode=mha_mode,
+                            sparse_mla_dense_mha_max_seq_len=dense_mha_max_seq_len,
+                            sparse_mla_topk_pattern=sparse_mla_topk_pattern,
+                            prefill_backend=prefill_backend,
+                        )
+
+                        # run_mla_benchmark needs the real backend name
+                        from mla_runner import run_mla_benchmark as run_mla
+
+                        run_label = f"{backend}_{variant_label} {spec}"
+                        pbar.set_postfix_str(run_label)
+
+                        try:
+                            result = run_mla(
+                                backend,
+                                config,
+                                prefill_backend=prefill_backend,
+                                sparse_mla_force_mqa=force_mqa,
+                            )
+                        except Exception as e:
+                            result = 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),
+                            )
+
+                        all_results.append(result)
+                        if args.output_csv:
+                            formatter.save_csv(all_results, args.output_csv)
+                        if args.output_json:
+                            formatter.save_json(all_results, args.output_json)
+
+                        if not result.success:
+                            console.print(
+                                f"[red]Error {backend}_{variant_label} "
+                                f"{spec}: {result.error}[/]"
+                            )
+
+                        pbar.update(1)
+
+        # Display results with variant labels as separate "backends"
+        console.print("\n[bold green]MHA vs MQA Results:[/]")
+        variant_backends = [f"{b}_{v}" for b in backends for v, _, _ in variants]
+        formatter.print_table(all_results, variant_backends)
+
     # Handle model parameter sweep mode
     elif hasattr(args, "model_parameter_sweep") and args.model_parameter_sweep:
         # Model parameter sweep
@@ -1186,6 +1358,10 @@ def main():
                             profile_memory=args.profile_memory,
                             warmup_ms=args.warmup_ms,
                             prefill_backend=pb,
+                            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,
+                            v_head_dim=args.v_head_dim,
                         )
 
                         result = run_benchmark(config)

benchmarks/attention_benchmarks/common.py

@@ -4,8 +4,10 @@
 """Common utilities for attention benchmarking."""
 
 import csv
+import gc
 import json
 import math
+from collections.abc import Sequence
 from dataclasses import asdict, dataclass
 from pathlib import Path
 from typing import Any
@@ -44,10 +46,13 @@ def run_do_bench(
     kwargs: dict[str, Any] = {"return_mode": "all"}
     if use_cuda_graphs:
         result = triton.testing.do_bench_cudagraph(benchmark_fn, **kwargs)
+        gc.collect()
+        torch.accelerator.empty_cache()
     else:
         if warmup_ms is not None:
             kwargs["warmup"] = warmup_ms
         result = triton.testing.do_bench(benchmark_fn, **kwargs)
+    torch.accelerator.synchronize()
     return result
 
 
@@ -91,42 +96,6 @@ except ImportError:
     AttentionLayerBase = object  # Fallback
 
 
-class MockKVBProj:
-    """Mock KV projection layer for MLA prefill mode.
-
-    Mimics ColumnParallelLinear behavior for kv_b_proj in MLA backends.
-    Projects kv_c_normed to [qk_nope_head_dim + v_head_dim] per head.
-    """
-
-    def __init__(self, num_heads: int, qk_nope_head_dim: int, v_head_dim: int):
-        self.num_heads = num_heads
-        self.qk_nope_head_dim = qk_nope_head_dim
-        self.v_head_dim = v_head_dim
-        self.out_dim = qk_nope_head_dim + v_head_dim
-        self.weight = torch.empty(0, dtype=torch.bfloat16)
-
-    def __call__(self, x: torch.Tensor) -> tuple[torch.Tensor]:
-        """
-        Project kv_c_normed to output space.
-
-        Args:
-            x: Input tensor [num_tokens, kv_lora_rank]
-
-        Returns:
-            Tuple containing output tensor
-                [num_tokens, num_heads, qk_nope_head_dim + v_head_dim]
-        """
-        num_tokens = x.shape[0]
-        result = torch.randn(
-            num_tokens,
-            self.num_heads,
-            self.out_dim,
-            device=x.device,
-            dtype=x.dtype,
-        )
-        return (result,)  # Return as tuple to match ColumnParallelLinear API
-
-
 class MockIndexer:
     """Mock Indexer for sparse MLA backends.
 
@@ -158,6 +127,60 @@ class MockIndexer:
         )
         self.topk_indices_buffer[:num_tokens] = indices
 
+    def fill_indices(
+        self,
+        num_tokens: int,
+        max_kv_len: int,
+        pattern: str = "random",
+        requests: Sequence[Any] | None = None,
+    ):
+        if pattern == "random":
+            self.fill_random_indices(num_tokens, max_kv_len)
+            return
+        if pattern == "prefix":
+            indices = torch.arange(
+                self.topk_tokens,
+                dtype=torch.int32,
+                device=self.topk_indices_buffer.device,
+            )
+            indices = (indices % max_kv_len).expand(num_tokens, -1)
+            self.topk_indices_buffer[:num_tokens] = indices
+            return
+        if pattern == "sliding_window":
+            if requests is None:
+                start = max(max_kv_len - self.topk_tokens, 0)
+                indices = torch.arange(
+                    start,
+                    start + self.topk_tokens,
+                    dtype=torch.int32,
+                    device=self.topk_indices_buffer.device,
+                )
+                indices = indices.clamp(max=max_kv_len - 1).expand(num_tokens, -1)
+                self.topk_indices_buffer[:num_tokens] = indices
+                return
+
+            rows = []
+            offsets = torch.arange(
+                self.topk_tokens,
+                dtype=torch.int32,
+                device=self.topk_indices_buffer.device,
+            ) - (self.topk_tokens - 1)
+            for request in requests:
+                q_len = request.q_len
+                kv_len = request.kv_len
+                context_len = kv_len - q_len
+                positions = torch.arange(
+                    context_len,
+                    kv_len,
+                    dtype=torch.int32,
+                    device=self.topk_indices_buffer.device,
+                )
+                row_indices = positions[:, None] + offsets[None, :]
+                rows.append(row_indices.clamp(min=0, max=kv_len - 1))
+            self.topk_indices_buffer[:num_tokens] = torch.cat(rows, dim=0)
+            return
+        raise ValueError(f"Unknown sparse MLA topk pattern: {pattern}")
+
 
 class MockLayer(AttentionLayerBase):
     """Mock attention layer with scale parameters and impl.
@@ -252,10 +275,14 @@ class BenchmarkConfig:
     num_kv_heads: int
     block_size: int
     device: str
+    max_model_len: int | None = None
     dtype: torch.dtype = torch.float16
     profile_memory: bool = False
-    use_cuda_graphs: bool = False
+    use_cuda_graphs: bool = True
     ncu_profile: bool = False
+    torch_profile: bool = False
+    torch_profile_dir: str | None = None
+    torch_profile_iters: int = 3
     warmup_ms: int | None = None
 
     # "auto" or "fp8"
@@ -271,6 +298,10 @@ class BenchmarkConfig:
     # Backend-specific tuning
     num_kv_splits: int | None = None  # CUTLASS MLA
     reorder_batch_threshold: int | None = None  # FlashAttn MLA, FlashMLA
+    sparse_mla_force_mqa: bool = False  # Force MQA path for sparse MLA
+    sparse_mla_mha_mode: str = "auto"  # "auto" or "dense"
+    sparse_mla_dense_mha_max_seq_len: int | None = None
+    sparse_mla_topk_pattern: str = "random"  # "random", "prefix", "sliding_window"
     num_splits: int | None = None  # FlashAttention split-K (0=auto, 1=disabled)
 
 

benchmarks/attention_benchmarks/configs/mla_sparse_mha_vs_mqa.yaml

@@ -0,0 +1,474 @@
+# Sparse MLA benchmark: forward_mha vs forward_mqa
+#
+# Usage:
+#   python benchmark.py --config configs/mla_sparse_mha_vs_mqa.yaml
+#
+# Heatmap grid:
+#   - batch_size: 1, 2, 4, 8, 16, 32
+#   - seq_len: 32, 64, 128, 256, 512, 1024, 2048
+#   - q_len: powers of two through seq_len
+#
+# Specs with q_len < seq_len include context; the q_len == seq_len diagonal
+# covers pure prefill.
+# The model shape below is the DP case. For the TP8 run, manually change
+# model.num_q_heads from 128 to 16 before rerunning this benchmark.
+
+mode: mha_vs_mqa
+
+model:
+  name: "deepseek-v3"
+  num_layers: 60
+  num_q_heads: 128
+  num_kv_heads: 1
+  head_dim: 576
+  kv_lora_rank: 512
+  qk_nope_head_dim: 128
+  qk_rope_head_dim: 64
+  v_head_dim: 128
+  block_size: 128
+  max_model_len: 2048
+
+batch_specs:
+  # Batch size 1
+  # seq_len = 32
+  - "1q1s32"
+  - "1q2s32"
+  - "1q4s32"
+  - "1q8s32"
+  - "1q16s32"
+  - "1q32"
+  # seq_len = 64
+  - "1q1s64"
+  - "1q2s64"
+  - "1q4s64"
+  - "1q8s64"
+  - "1q16s64"
+  - "1q32s64"
+  - "1q64"
+  # seq_len = 128
+  - "1q1s128"
+  - "1q2s128"
+  - "1q4s128"
+  - "1q8s128"
+  - "1q16s128"
+  - "1q32s128"
+  - "1q64s128"
+  - "1q128"
+  # seq_len = 256
+  - "1q1s256"
+  - "1q2s256"
+  - "1q4s256"
+  - "1q8s256"
+  - "1q16s256"
+  - "1q32s256"
+  - "1q64s256"
+  - "1q128s256"
+  - "1q256"
+  # seq_len = 512
+  - "1q1s512"
+  - "1q2s512"
+  - "1q4s512"
+  - "1q8s512"
+  - "1q16s512"
+  - "1q32s512"
+  - "1q64s512"
+  - "1q128s512"
+  - "1q256s512"
+  - "1q512"
+  # seq_len = 1024
+  - "1q1s1024"
+  - "1q2s1024"
+  - "1q4s1024"
+  - "1q8s1024"
+  - "1q16s1024"
+  - "1q32s1024"
+  - "1q64s1024"
+  - "1q128s1024"
+  - "1q256s1024"
+  - "1q512s1024"
+  - "1q1024"
+  # seq_len = 2048
+  - "1q1s2048"
+  - "1q2s2048"
+  - "1q4s2048"
+  - "1q8s2048"
+  - "1q16s2048"
+  - "1q32s2048"
+  - "1q64s2048"
+  - "1q128s2048"
+  - "1q256s2048"
+  - "1q512s2048"
+  - "1q1024s2048"
+  - "1q2048"
+
+  # Batch size 2
+  # seq_len = 32
+  - "2q1s32"
+  - "2q2s32"
+  - "2q4s32"
+  - "2q8s32"
+  - "2q16s32"
+  - "2q32"
+  # seq_len = 64
+  - "2q1s64"
+  - "2q2s64"
+  - "2q4s64"
+  - "2q8s64"
+  - "2q16s64"
+  - "2q32s64"
+  - "2q64"
+  # seq_len = 128
+  - "2q1s128"
+  - "2q2s128"
+  - "2q4s128"
+  - "2q8s128"
+  - "2q16s128"
+  - "2q32s128"
+  - "2q64s128"
+  - "2q128"
+  # seq_len = 256
+  - "2q1s256"
+  - "2q2s256"
+  - "2q4s256"
+  - "2q8s256"
+  - "2q16s256"
+  - "2q32s256"
+  - "2q64s256"
+  - "2q128s256"
+  - "2q256"
+  # seq_len = 512
+  - "2q1s512"
+  - "2q2s512"
+  - "2q4s512"
+  - "2q8s512"
+  - "2q16s512"
+  - "2q32s512"
+  - "2q64s512"
+  - "2q128s512"
+  - "2q256s512"
+  - "2q512"
+  # seq_len = 1024
+  - "2q1s1024"
+  - "2q2s1024"
+  - "2q4s1024"
+  - "2q8s1024"
+  - "2q16s1024"
+  - "2q32s1024"
+  - "2q64s1024"
+  - "2q128s1024"
+  - "2q256s1024"
+  - "2q512s1024"
+  - "2q1024"
+  # seq_len = 2048
+  - "2q1s2048"
+  - "2q2s2048"
+  - "2q4s2048"
+  - "2q8s2048"
+  - "2q16s2048"
+  - "2q32s2048"
+  - "2q64s2048"
+  - "2q128s2048"
+  - "2q256s2048"
+  - "2q512s2048"
+  - "2q1024s2048"
+  - "2q2048"
+
+  # Batch size 4
+  # seq_len = 32
+  - "4q1s32"
+  - "4q2s32"
+  - "4q4s32"
+  - "4q8s32"
+  - "4q16s32"
+  - "4q32"
+  # seq_len = 64
+  - "4q1s64"
+  - "4q2s64"
+  - "4q4s64"
+  - "4q8s64"
+  - "4q16s64"
+  - "4q32s64"
+  - "4q64"
+  # seq_len = 128
+  - "4q1s128"
+  - "4q2s128"
+  - "4q4s128"
+  - "4q8s128"
+  - "4q16s128"
+  - "4q32s128"
+  - "4q64s128"
+  - "4q128"
+  # seq_len = 256
+  - "4q1s256"
+  - "4q2s256"
+  - "4q4s256"
+  - "4q8s256"
+  - "4q16s256"
+  - "4q32s256"
+  - "4q64s256"
+  - "4q128s256"
+  - "4q256"
+  # seq_len = 512
+  - "4q1s512"
+  - "4q2s512"
+  - "4q4s512"
+  - "4q8s512"
+  - "4q16s512"
+  - "4q32s512"
+  - "4q64s512"
+  - "4q128s512"
+  - "4q256s512"
+  - "4q512"
+  # seq_len = 1024
+  - "4q1s1024"
+  - "4q2s1024"
+  - "4q4s1024"
+  - "4q8s1024"
+  - "4q16s1024"
+  - "4q32s1024"
+  - "4q64s1024"
+  - "4q128s1024"
+  - "4q256s1024"
+  - "4q512s1024"
+  - "4q1024"
+  # seq_len = 2048
+  - "4q1s2048"
+  - "4q2s2048"
+  - "4q4s2048"
+  - "4q8s2048"
+  - "4q16s2048"
+  - "4q32s2048"
+  - "4q64s2048"
+  - "4q128s2048"
+  - "4q256s2048"
+  - "4q512s2048"
+  - "4q1024s2048"
+  - "4q2048"
+
+  # Batch size 8
+  # seq_len = 32
+  - "8q1s32"
+  - "8q2s32"
+  - "8q4s32"
+  - "8q8s32"
+  - "8q16s32"
+  - "8q32"
+  # seq_len = 64
+  - "8q1s64"
+  - "8q2s64"
+  - "8q4s64"
+  - "8q8s64"
+  - "8q16s64"
+  - "8q32s64"
+  - "8q64"
+  # seq_len = 128
+  - "8q1s128"
+  - "8q2s128"
+  - "8q4s128"
+  - "8q8s128"
+  - "8q16s128"
+  - "8q32s128"
+  - "8q64s128"
+  - "8q128"
+  # seq_len = 256
+  - "8q1s256"
+  - "8q2s256"
+  - "8q4s256"
+  - "8q8s256"
+  - "8q16s256"
+  - "8q32s256"
+  - "8q64s256"
+  - "8q128s256"
+  - "8q256"
+  # seq_len = 512
+  - "8q1s512"
+  - "8q2s512"
+  - "8q4s512"
+  - "8q8s512"
+  - "8q16s512"
+  - "8q32s512"
+  - "8q64s512"
+  - "8q128s512"
+  - "8q256s512"
+  - "8q512"
+  # seq_len = 1024
+  - "8q1s1024"
+  - "8q2s1024"
+  - "8q4s1024"
+  - "8q8s1024"
+  - "8q16s1024"
+  - "8q32s1024"
+  - "8q64s1024"
+  - "8q128s1024"
+  - "8q256s1024"
+  - "8q512s1024"
+  - "8q1024"
+  # seq_len = 2048
+  - "8q1s2048"
+  - "8q2s2048"
+  - "8q4s2048"
+  - "8q8s2048"
+  - "8q16s2048"
+  - "8q32s2048"
+  - "8q64s2048"
+  - "8q128s2048"
+  - "8q256s2048"
+  - "8q512s2048"
+  - "8q1024s2048"
+  - "8q2048"
+
+  # Batch size 16
+  # seq_len = 32
+  - "16q1s32"
+  - "16q2s32"
+  - "16q4s32"
+  - "16q8s32"
+  - "16q16s32"
+  - "16q32"
+  # seq_len = 64
+  - "16q1s64"
+  - "16q2s64"
+  - "16q4s64"
+  - "16q8s64"
+  - "16q16s64"
+  - "16q32s64"
+  - "16q64"
+  # seq_len = 128
+  - "16q1s128"
+  - "16q2s128"
+  - "16q4s128"
+  - "16q8s128"
+  - "16q16s128"
+  - "16q32s128"
+  - "16q64s128"
+  - "16q128"
+  # seq_len = 256
+  - "16q1s256"
+  - "16q2s256"
+  - "16q4s256"
+  - "16q8s256"
+  - "16q16s256"
+  - "16q32s256"
+  - "16q64s256"
+  - "16q128s256"
+  - "16q256"
+  # seq_len = 512
+  - "16q1s512"
+  - "16q2s512"
+  - "16q4s512"
+  - "16q8s512"
+  - "16q16s512"
+  - "16q32s512"
+  - "16q64s512"
+  - "16q128s512"
+  - "16q256s512"
+  - "16q512"
+  # seq_len = 1024
+  - "16q1s1024"
+  - "16q2s1024"
+  - "16q4s1024"
+  - "16q8s1024"
+  - "16q16s1024"
+  - "16q32s1024"
+  - "16q64s1024"
+  - "16q128s1024"
+  - "16q256s1024"
+  - "16q512s1024"
+  - "16q1024"
+  # seq_len = 2048
+  - "16q1s2048"
+  - "16q2s2048"
+  - "16q4s2048"
+  - "16q8s2048"
+  - "16q16s2048"
+  - "16q32s2048"
+  - "16q64s2048"
+  - "16q128s2048"
+  - "16q256s2048"
+  - "16q512s2048"
+  - "16q1024s2048"
+  - "16q2048"
+
+  # Batch size 32
+  # seq_len = 32
+  - "32q1s32"
+  - "32q2s32"
+  - "32q4s32"
+  - "32q8s32"
+  - "32q16s32"
+  - "32q32"
+  # seq_len = 64
+  - "32q1s64"
+  - "32q2s64"
+  - "32q4s64"
+  - "32q8s64"
+  - "32q16s64"
+  - "32q32s64"
+  - "32q64"
+  # seq_len = 128
+  - "32q1s128"
+  - "32q2s128"
+  - "32q4s128"
+  - "32q8s128"
+  - "32q16s128"
+  - "32q32s128"
+  - "32q64s128"
+  - "32q128"
+  # seq_len = 256
+  - "32q1s256"
+  - "32q2s256"
+  - "32q4s256"
+  - "32q8s256"
+  - "32q16s256"
+  - "32q32s256"
+  - "32q64s256"
+  - "32q128s256"
+  - "32q256"
+  # seq_len = 512
+  - "32q1s512"
+  - "32q2s512"
+  - "32q4s512"
+  - "32q8s512"
+  - "32q16s512"
+  - "32q32s512"
+  - "32q64s512"
+  - "32q128s512"
+  - "32q256s512"
+  - "32q512"
+  # seq_len = 1024
+  - "32q1s1024"
+  - "32q2s1024"
+  - "32q4s1024"
+  - "32q8s1024"
+  - "32q16s1024"
+  - "32q32s1024"
+  - "32q64s1024"
+  - "32q128s1024"
+  - "32q256s1024"
+  - "32q512s1024"
+  - "32q1024"
+  # seq_len = 2048
+  - "32q1s2048"
+  - "32q2s2048"
+  - "32q4s2048"
+  - "32q8s2048"
+  - "32q16s2048"
+  - "32q32s2048"
+  - "32q64s2048"
+  - "32q128s2048"
+  - "32q256s2048"
+  - "32q512s2048"
+  - "32q1024s2048"
+  - "32q2048"
+
+backends:
+  - FLASHMLA_SPARSE
+
+device: "cuda:0"
+profile_memory: false
+sparse_mla_dense_mha_max_seq_len: 2048
+sparse_mla_topk_pattern: "random"
+
+output:
+  csv: "benchmark_output/mla_sparse_mha_vs_mqa.csv"
+  json: "benchmark_output/mla_sparse_mha_vs_mqa.json"

benchmarks/attention_benchmarks/mla_runner.py

@@ -9,6 +9,8 @@ needing full VllmConfig integration.
 """
 
 import statistics
+import tempfile
+from pathlib import Path
 
 import numpy as np
 import torch
@@ -17,7 +19,6 @@ from common import (
     BenchmarkResult,
     MockHfConfig,
     MockIndexer,
-    MockKVBProj,
     MockLayer,
     run_do_bench,
     run_ncu_profile,
@@ -33,8 +34,59 @@ from vllm.config import (
     VllmConfig,
     set_current_vllm_config,
 )
+from vllm.model_executor.layers.linear import ColumnParallelLinear
 from vllm.v1.attention.backends.mla.prefill.registry import MLAPrefillBackendEnum
 
+
+def _safe_profile_name(value: str) -> str:
+    return "".join(c if c.isalnum() or c in "._-" else "_" for c in value)
+
+
+def _create_kv_b_proj(
+    mla_dims: dict,
+    device: torch.device,
+):
+    kv_b_proj = ColumnParallelLinear(
+        mla_dims["kv_lora_rank"],
+        mla_dims["num_q_heads"]
+        * (mla_dims["qk_nope_head_dim"] + mla_dims["v_head_dim"]),
+        bias=False,
+        params_dtype=torch.bfloat16,
+        quant_config=None,
+        prefix="benchmark.kv_b_proj",
+    ).to(device)
+    with torch.no_grad():
+        kv_b_proj.weight.copy_(torch.randn_like(kv_b_proj.weight))
+    return kv_b_proj
+
+
+def _ensure_single_rank_model_parallel() -> None:
+    import torch.distributed as dist
+
+    from vllm.distributed import (
+        ensure_model_parallel_initialized,
+        init_distributed_environment,
+        model_parallel_is_initialized,
+    )
+
+    if not dist.is_available():
+        return
+    if not dist.is_initialized():
+        with tempfile.NamedTemporaryFile(
+            prefix="vllm_bench_dist_", delete=False
+        ) as init_file:
+            distributed_init_method = f"file://{init_file.name}"
+        init_distributed_environment(
+            world_size=1,
+            rank=0,
+            distributed_init_method=distributed_init_method,
+            local_rank=0,
+            backend="nccl",
+        )
+    if not model_parallel_is_initialized():
+        ensure_model_parallel_initialized(1, 1)
+
+
 # ============================================================================
 # VllmConfig Creation
 # ============================================================================
@@ -66,10 +118,12 @@ def create_minimal_vllm_config(
     block_size: int = 128,
     max_num_seqs: int = 256,
     max_num_batched_tokens: int = 8192,
+    max_model_len: int = 32768,
     mla_dims: dict | None = None,
     index_topk: int | None = None,
     prefill_backend: str | None = None,
     kv_cache_dtype: str = "auto",
+    sparse_mla_force_mqa: bool = False,
 ) -> VllmConfig:
     """
     Create minimal VllmConfig for MLA benchmarks.
@@ -86,6 +140,8 @@ def create_minimal_vllm_config(
         prefill_backend: Prefill backend name (e.g., "fa3", "fa4", "flashinfer",
                         "trtllm"). Configures the attention config to force
                         the specified prefill backend.
+        sparse_mla_force_mqa: If True, forces all sparse MLA tokens through
+                    forward_mqa (even prefill tokens).
 
     Returns:
         VllmConfig for benchmarking
@@ -131,7 +187,7 @@ def create_minimal_vllm_config(
             trust_remote_code=True,
             dtype="bfloat16",
             seed=0,
-            max_model_len=32768,
+            max_model_len=max_model_len,
             quantization=None,
             enforce_eager=False,
             max_logprobs=20,
@@ -163,7 +219,7 @@ def create_minimal_vllm_config(
     scheduler_config = SchedulerConfig(
         max_num_seqs=max_num_seqs,
         max_num_batched_tokens=max(max_num_batched_tokens, max_num_seqs),
-        max_model_len=32768,
+        max_model_len=max_model_len,
         is_encoder_decoder=False,
         enable_chunked_prefill=True,
     )
@@ -192,6 +248,9 @@ def create_minimal_vllm_config(
                 "flash_attn_version"
             ]
 
+    if sparse_mla_force_mqa:
+        vllm_config.attention_config.sparse_mla_force_mqa = True
+
     return vllm_config
 
 
@@ -548,12 +607,7 @@ def _create_backend_impl(
     # Calculate scale
     scale = 1.0 / np.sqrt(mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"])
 
-    # Create mock kv_b_proj layer for prefill mode
-    mock_kv_b_proj = MockKVBProj(
-        num_heads=mla_dims["num_q_heads"],
-        qk_nope_head_dim=mla_dims["qk_nope_head_dim"],
-        v_head_dim=mla_dims["v_head_dim"],
-    )
+    kv_b_proj = _create_kv_b_proj(mla_dims, device)
 
     # Create indexer for sparse backends
     indexer = None
@@ -584,7 +638,7 @@ def _create_backend_impl(
         "qk_rope_head_dim": mla_dims["qk_rope_head_dim"],
         "qk_head_dim": mla_dims["qk_nope_head_dim"] + mla_dims["qk_rope_head_dim"],
         "v_head_dim": mla_dims["v_head_dim"],
-        "kv_b_proj": mock_kv_b_proj,
+        "kv_b_proj": kv_b_proj,
     }
 
     # Add indexer for sparse backends
@@ -785,14 +839,35 @@ def _run_single_benchmark(
     # Fill indexer with random indices for sparse backends
     is_sparse = backend_cfg.get("is_sparse", False)
     if is_sparse and indexer is not None:
-        indexer.fill_random_indices(total_q, max_kv_len)
+        indexer.fill_indices(
+            total_q,
+            max_kv_len,
+            getattr(config, "sparse_mla_topk_pattern", "random"),
+        )
 
     # Determine which forward methods to use based on metadata.
-    # Sparse MLA backends always use forward_mqa
-    has_decode = is_sparse or getattr(metadata, "decode", None) is not None
-    has_prefill = not is_sparse and getattr(metadata, "prefill", None) is not None
+    # Non-sparse backends use .decode/.prefill sub-objects.
+    # Sparse backends use num_decode_tokens/num_prefills directly.
+    #
+    # sparse_mla_force_mqa overrides: even for prefill metadata, use MQA.
+    force_mqa = getattr(config, "sparse_mla_force_mqa", False)
+    force_dense_mha = getattr(config, "sparse_mla_mha_mode", "auto") == "dense"
+    if force_mqa:
+        has_decode = True
+        has_prefill = False
+    elif is_sparse:
+        has_decode = metadata.num_decode_tokens > 0
+        has_prefill = metadata.num_prefills > 0
+    else:
+        has_decode = metadata.decode is not None
+        has_prefill = metadata.prefill is not None
     if not has_decode and not has_prefill:
         raise RuntimeError("Metadata has neither decode nor prefill metadata")
+    if is_sparse and force_dense_mha and not has_prefill:
+        raise RuntimeError(
+            "Sparse MLA dense_mha benchmark did not produce prefill metadata. "
+            "Check reorder_batch_threshold/path forcing."
+        )
 
     num_decode = (
         metadata.num_decode_tokens
@@ -871,7 +946,6 @@ def _run_single_benchmark(
                 metadata,
                 prefill_inputs["k_scale"],
                 prefill_fp8_output if fused_output else prefill_inputs["output"],
-                prefill_output_scale if fused_output else None,
             )
             if fused_output:
                 out = prefill_fp8_output
@@ -898,6 +972,48 @@ def _run_single_benchmark(
             throughput_tokens_per_sec=0.0,
         )
 
+    if config.torch_profile:
+        profile_dir = Path(
+            config.torch_profile_dir or "benchmark_outputs/torch_profiles"
+        )
+        profile_dir.mkdir(parents=True, exist_ok=True)
+        trace_name = _safe_profile_name(f"{config.backend}_{config.batch_spec}")
+        trace_path = profile_dir / f"{trace_name}.json"
+        iters = max(config.torch_profile_iters, 1)
+
+        forward_fn()
+        torch.accelerator.synchronize()
+        with torch.profiler.profile(
+            activities=[
+                torch.profiler.ProfilerActivity.CPU,
+                torch.profiler.ProfilerActivity.CUDA,
+            ],
+            record_shapes=True,
+            profile_memory=True,
+            with_stack=False,
+        ) as prof:
+            for _ in range(iters):
+                forward_fn()
+                torch.accelerator.synchronize()
+                prof.step()
+        prof.export_chrome_trace(str(trace_path))
+        print(f"Saved PyTorch profiler trace to {trace_path}")
+        print(
+            prof.key_averages().table(
+                sort_by="cuda_time_total",
+                row_limit=25,
+            )
+        )
+        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
@@ -920,6 +1036,7 @@ def _run_mla_benchmark_batched(
     configs_with_params: list[tuple],  # [(config, threshold, num_splits), ...]
     index_topk: int = 2048,
     prefill_backend: str | None = None,
+    sparse_mla_force_mqa: bool = False,
     output_scale: float | None = None,
     fuse_quant_op: bool = False,
 ) -> list[BenchmarkResult]:
@@ -940,6 +1057,8 @@ def _run_mla_benchmark_batched(
         index_topk: Topk value for sparse MLA backends (default 2048)
         prefill_backend: Prefill backend name (e.g., "fa3", "fa4").
             When set, forces the specified FlashAttention version for prefill.
+        sparse_mla_force_mqa: If True, forces all sparse MLA tokens through
+            forward_mqa (even prefill tokens).
 
     Returns:
         List of BenchmarkResult objects
@@ -980,21 +1099,41 @@ def _run_mla_benchmark_batched(
         sum(r.q_len for r in parse_batch_spec(cfg.batch_spec))
         for cfg, *_ in configs_with_params
     )
+    max_model_len = max(
+        max_total_q,
+        max(
+            getattr(cfg, "max_model_len", None) or 32768
+            for cfg, *_ in configs_with_params
+        ),
+    )
 
     # Create and set vLLM config for MLA (reused across all benchmarks)
     vllm_config = create_minimal_vllm_config(
         model_name="deepseek-v3",  # Used only for model path
         block_size=block_size,
         max_num_batched_tokens=max_total_q,
+        max_model_len=max_model_len,
         mla_dims=mla_dims,  # Use custom dims from config or default
         index_topk=index_topk if is_sparse else None,
         prefill_backend=prefill_backend,
         kv_cache_dtype=kv_cache_dtype,
+        sparse_mla_force_mqa=sparse_mla_force_mqa,
     )
 
     results = []
 
+    # Initialize workspace manager (needed by metadata builders)
+    from vllm.v1.worker.workspace import (
+        init_workspace_manager,
+        is_workspace_manager_initialized,
+    )
+
+    if not is_workspace_manager_initialized():
+        init_workspace_manager(device)
+
     with set_current_vllm_config(vllm_config):
+        _ensure_single_rank_model_parallel()
+
         # Create backend impl, layer, builder, and indexer (reused across benchmarks)
         impl, layer, builder_instance, indexer = _create_backend_impl(
             backend_cfg,
@@ -1040,9 +1179,20 @@ def _run_mla_benchmark_batched(
         for config, threshold, num_splits in configs_with_params:
             # Set threshold for this benchmark (FlashAttn/FlashMLA only)
             original_threshold = None
-            if threshold is not None and builder_instance:
+            effective_threshold = threshold
+            force_dense_mha = (
+                is_sparse
+                and getattr(config, "sparse_mla_mha_mode", "auto") == "dense"
+                and not getattr(config, "sparse_mla_force_mqa", False)
+            )
+            if force_dense_mha:
+                # Sparse MLA normally treats q_len <= 1 as decode. Use an
+                # impossible threshold so dense_mha benchmarks actually run
+                # the prefill/MHA path, including q_len=1 short extends.
+                effective_threshold = -1
+            if effective_threshold is not None and builder_instance:
                 original_threshold = builder_instance.reorder_batch_threshold
-                builder_instance.reorder_batch_threshold = threshold
+                builder_instance.reorder_batch_threshold = effective_threshold
 
             # Set num_splits for CUTLASS
             original_num_splits = None
@@ -1090,6 +1240,7 @@ def run_mla_benchmark(
     num_kv_splits: int | None = None,
     index_topk: int = 2048,
     prefill_backend: str | None = None,
+    sparse_mla_force_mqa: bool = False,
     output_scale: float | None = None,
     fuse_quant_op: bool = False,
 ) -> BenchmarkResult | list[BenchmarkResult]:
@@ -1111,6 +1262,8 @@ def run_mla_benchmark(
         index_topk: Topk value for sparse MLA backends (default 2048)
         prefill_backend: Prefill backend name (e.g., "fa3", "fa4").
             When set, forces the specified FlashAttention version for prefill.
+        sparse_mla_force_mqa: If True, forces all sparse MLA tokens through
+            forward_mqa (even prefill tokens).
         output_scale: Static per-tensor FP8 scale for prefill output (None = bf16).
         fuse_quant_op: With output_scale set, fuse the FP8 write into the prefill
             kernel vs a standalone post-quant kernel. See _run_single_benchmark.
@@ -1142,6 +1295,7 @@ def run_mla_benchmark(
         configs_with_params,
         index_topk,
         prefill_backend=prefill_backend,
+        sparse_mla_force_mqa=sparse_mla_force_mqa,
         output_scale=output_scale,
         fuse_quant_op=fuse_quant_op,
     )

benchmarks/kernels/bench_cp_gather_fp8.py

@@ -69,12 +69,11 @@ def make_inputs(total_tokens, num_reqs, block_size):
     # Output workspace
     dst = torch.zeros(total_tokens, HEAD_DIM, dtype=torch.bfloat16, device="cuda")
 
-    seq_lens_t = torch.tensor(seq_lens, dtype=torch.int32, device="cuda")
     workspace_starts_t = torch.tensor(
         workspace_starts, dtype=torch.int32, device="cuda"
     )
 
-    return cache, dst, block_table, seq_lens_t, workspace_starts_t
+    return cache, dst, block_table, workspace_starts_t
 
 
 def bench_scenario(label, num_reqs, total_tokens_list, save_path):
@@ -94,7 +93,7 @@ def bench_scenario(label, num_reqs, total_tokens_list, save_path):
         )
     )
     def bench_fn(total_tokens, provider, num_reqs):
-        cache, dst, block_table, seq_lens_t, ws_starts = make_inputs(
+        cache, dst, block_table, ws_starts = make_inputs(
             total_tokens, num_reqs, BLOCK_SIZE
         )
 
@@ -102,7 +101,7 @@ def bench_scenario(label, num_reqs, total_tokens_list, save_path):
 
         ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
             lambda: ops.cp_gather_and_upconvert_fp8_kv_cache(
-                cache, dst, block_table, seq_lens_t, ws_starts, num_reqs
+                cache, dst, block_table, ws_starts, num_reqs
             ),
             quantiles=quantiles,
             rep=500,

benchmarks/kernels/benchmark_inkling_qkvr_prep.py

@@ -0,0 +1,176 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import statistics
+
+import torch
+from tabulate import tabulate
+
+from vllm.models.inkling.nvidia.ops import qkvr_prep
+from vllm.utils.argparse_utils import FlexibleArgumentParser
+
+
+def make_inputs(tokens: int, tp_size: int, is_local: bool):
+    torch.manual_seed(0)
+    num_q_heads = 64 // tp_size
+    num_kv_heads = (16 if is_local else 8) // tp_size
+    head_dim = 128
+    d_rel = 16
+    rel_extent = 512 if is_local else 1024
+    page_size = 16
+    num_blocks = (tokens + page_size - 1) // page_size
+    q_width = num_q_heads * head_dim
+    kv_width = num_kv_heads * head_dim
+    r_width = num_q_heads * d_rel
+    device = "cuda"
+
+    qkvr = torch.randn(
+        tokens,
+        q_width + 2 * kv_width + r_width,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    k_weight = torch.randn(kv_width, 4, device=device, dtype=torch.bfloat16)
+    v_weight = torch.randn_like(k_weight)
+    q_norm_weight = torch.randn(head_dim, device=device, dtype=torch.bfloat16)
+    k_norm_weight = torch.randn_like(q_norm_weight)
+    rel_proj = torch.randn(d_rel, rel_extent, device=device, dtype=torch.bfloat16)
+    conv_cache = torch.zeros(
+        num_blocks,
+        num_kv_heads,
+        page_size,
+        2 * head_dim,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    key_cache = torch.empty(
+        num_blocks,
+        page_size,
+        num_kv_heads,
+        head_dim,
+        device=device,
+        dtype=torch.bfloat16,
+    )
+    value_cache = torch.empty_like(key_cache)
+    positions = torch.arange(tokens, device=device, dtype=torch.int64)
+    block_table = torch.arange(num_blocks, device=device, dtype=torch.int32)[None]
+    seq_idx = torch.zeros(tokens, device=device, dtype=torch.int32)
+    slots = torch.arange(tokens, device=device, dtype=torch.int64)
+    query_start = torch.zeros(tokens, device=device, dtype=torch.int32)
+    log_scaling = None
+    if not is_local:
+        effective_n = (positions + 1).to(torch.float32)
+        log_scaling = 1.0 + 0.1 * torch.log(torch.clamp(effective_n / 128000, min=1.0))
+    return (
+        qkvr,
+        k_weight,
+        v_weight,
+        q_norm_weight,
+        k_norm_weight,
+        rel_proj,
+        1e-6,
+        num_q_heads,
+        num_kv_heads,
+        head_dim,
+        d_rel,
+        conv_cache,
+        key_cache,
+        value_cache,
+        positions,
+        block_table,
+        seq_idx,
+        slots,
+        query_start,
+        slots,
+        0,
+        head_dim,
+        page_size,
+        log_scaling,
+    )
+
+
+def capture(implementation, inputs):
+    outputs = []
+
+    def run():
+        outputs[:] = implementation.fused_qkvr_prep(*inputs)
+
+    stream = torch.cuda.Stream()
+    stream.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(stream):
+        for _ in range(3):
+            run()
+    torch.cuda.current_stream().wait_stream(stream)
+    torch.accelerator.synchronize()
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        run()
+    torch.accelerator.synchronize()
+    return graph, outputs
+
+
+def time_graph(graph: torch.cuda.CUDAGraph, warmup: int, repeats: int) -> float:
+    for _ in range(warmup):
+        graph.replay()
+    torch.accelerator.synchronize()
+    start = torch.cuda.Event(enable_timing=True)
+    end = torch.cuda.Event(enable_timing=True)
+    start.record()
+    for _ in range(repeats):
+        graph.replay()
+    end.record()
+    end.synchronize()
+    return start.elapsed_time(end) * 1000 / repeats
+
+
+def benchmark(inputs, args) -> float:
+    graph, _ = capture(qkvr_prep, inputs)
+    return statistics.median(
+        time_graph(graph, args.warmup, args.repeats) for _ in range(args.trials)
+    )
+
+
+@torch.inference_mode()
+def main(args):
+    rows = []
+    for tp_size in args.tp_sizes:
+        for tokens in args.tokens:
+            for is_local in (True, False):
+                triton_us = benchmark(make_inputs(tokens, tp_size, is_local), args)
+                rows.append(
+                    [
+                        tp_size,
+                        tokens,
+                        "local" if is_local else "global",
+                        triton_us,
+                    ]
+                )
+
+    print("Inkling QKVR prep (CUDA graph, median latency)")
+    print(
+        tabulate(
+            rows,
+            headers=[
+                "TP",
+                "tokens",
+                "scope",
+                "Triton (us)",
+            ],
+            floatfmt=("d", "d", "", ".2f"),
+        )
+    )
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser()
+    parser.add_argument(
+        "--tokens",
+        type=int,
+        nargs="+",
+        default=[1 << power for power in range(15)],
+    )
+    parser.add_argument("--tp-sizes", type=int, nargs="+", default=[4, 8])
+    parser.add_argument("--warmup", type=int, default=20)
+    parser.add_argument("--repeats", type=int, default=200)
+    parser.add_argument("--trials", type=int, default=5)
+    main(parser.parse_args())

benchmarks/kernels/benchmark_rdna_hybrid_w4a16_gemm.py

@@ -0,0 +1,201 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+Benchmark the RDNAHybridW4A16LinearKernel across decode and prefill shapes.
+
+Usage:
+    python benchmark_int4_gemm.py
+    python benchmark_int4_gemm.py --models Qwen/Qwen3-4B
+    python benchmark_int4_gemm.py --group-size 128
+"""
+
+import argparse
+import copy
+import itertools
+import os
+
+import torch
+
+from vllm.triton_utils import triton
+
+# ---------------------------------------------------------------------------
+# Weight shapes: [K, N], TP_SPLIT_DIM
+# ---------------------------------------------------------------------------
+WEIGHT_SHAPES = {
+    "Qwen/Qwen3-4B": [
+        ([2560, 3840], 1),  # qkv_proj
+        ([2560, 2560], 0),  # o_proj
+        ([2560, 19456], 1),  # gate_up_proj
+        ([9728, 2560], 0),  # down_proj
+    ],
+    "Qwen/Qwen2.5-7B-Instruct": [
+        ([3584, 4608], 1),
+        ([3584, 3584], 0),
+        ([3584, 37888], 1),
+        ([18944, 3584], 0),
+    ],
+    "trymirai/SmolLM2-1.7B-Instruct-AWQ": [
+        ([2048, 6144], 1),  # qkv_proj
+        ([2048, 2048], 0),  # o_proj
+        ([2048, 16384], 1),  # gate_up_proj
+        ([8192, 2048], 0),  # down_proj
+    ],
+    "RedHatAI/Qwen3-8B-quantized.w4a16": [
+        ([4096, 6144], 1),  # qkv_proj
+        ([4096, 4096], 0),  # o_proj
+        ([4096, 24576], 1),  # gate_up_proj
+        ([12288, 4096], 0),  # down_proj
+    ],
+}
+
+
+# ---------------------------------------------------------------------------
+# Weight packing
+# ---------------------------------------------------------------------------
+def prepare_hybrid_weights(K, N, group_size, device="cuda"):
+    """Create random weights for benchmarking.
+
+    Returns (w_q_skinny, w_s_skinny, w_fp16, w_zp). The triton path derives
+    its int32 view from w_q_skinny, so no separate int32 buffer is returned.
+    """
+    num_groups = K // group_size
+
+    # Random packed weights — actual values don't matter for throughput
+    w_q_skinny_i32 = torch.randint(
+        0, 2**31, (N, K // 8), dtype=torch.int32, device=device
+    )
+    w_q_skinny = w_q_skinny_i32.view(torch.int8).contiguous()
+    w_s_skinny = torch.randn(N, num_groups, dtype=torch.float16, device=device) * 0.01
+
+    # Raw per-group zero-points for asymmetric benchmarks
+    w_zp = torch.randint(0, 16, (N, num_groups), dtype=torch.int32, device=device).to(
+        torch.float16
+    )
+
+    # FP16 baseline for F.linear
+    w_fp16 = torch.randn(N, K, dtype=torch.float16, device=device) * 0.01
+
+    return w_q_skinny, w_s_skinny, w_fp16, w_zp
+
+
+# ---------------------------------------------------------------------------
+# Benchmark
+# ---------------------------------------------------------------------------
+PROVIDERS = ["torch-fp16", "hybrid-w4a16", "hybrid-w4a16-zp"]
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size"],
+        x_vals=[1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096],
+        x_log=False,
+        line_arg="provider",
+        line_vals=PROVIDERS,
+        line_names=PROVIDERS,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="FP16 vs Hybrid W4A16",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K, group_size, weights):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.float16
+    a = torch.randn((M, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-fp16":
+        w_fp16 = weights["w_fp16"]
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, w_fp16),
+            quantiles=quantiles,
+        )
+    elif provider in ("hybrid-w4a16", "hybrid-w4a16-zp"):
+        from vllm.model_executor.kernels.linear.mixed_precision import (
+            rdna_hybrid_w4a16 as _k,
+        )
+
+        _rdna_hybrid_w4a16_apply_impl = _k._rdna_hybrid_w4a16_apply_impl
+        from vllm.utils.platform_utils import num_compute_units
+
+        w = weights
+        cu_count = num_compute_units()
+        use_zp = provider == "hybrid-w4a16-zp"
+
+        def run():
+            return _rdna_hybrid_w4a16_apply_impl(
+                a,
+                w["w_q_skinny"],
+                w["w_s_skinny"],
+                w["w_zp"] if use_zp else None,
+                None,  # bias
+                cu_count,
+                group_size,
+            )
+
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            run,
+            quantiles=quantiles,
+        )
+    else:
+        return 0.0, 0.0, 0.0
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    KN_model_names = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            KN_model_names.append(KN)
+    return KN_model_names
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Benchmark RDNAHybridW4A16LinearKernel"
+    )
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["Qwen/Qwen3-4B"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    parser.add_argument("--group-size", type=int, default=128)
+    parser.add_argument("--save-path", type=str, default=None)
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        group_size = args.group_size
+        print(f"\n{'=' * 70}")
+        print(f"{model}, N={N} K={K}, group_size={group_size}")
+        print(f"{'=' * 70}")
+
+        w_q_skinny, w_s_skinny, w_fp16, w_zp = prepare_hybrid_weights(K, N, group_size)
+
+        weights = {
+            "w_q_skinny": w_q_skinny,
+            "w_s_skinny": w_s_skinny,
+            "w_fp16": w_fp16,
+            "w_zp": w_zp,
+        }
+
+        save_path = args.save_path or f"bench_int4_res_n{N}_k{K}"
+        os.makedirs(save_path, exist_ok=True)
+        benchmark.run(
+            print_data=True,
+            show_plots=False,
+            save_path=save_path,
+            N=N,
+            K=K,
+            group_size=group_size,
+            weights=weights,
+        )
+
+    print("\nBenchmark finished!")

benchmarks/kernels/cpu/benchmark_cpu_attn.py

@@ -154,7 +154,7 @@ def main(
                 scale=scale,
                 causal=True,
                 alibi_slopes=None,
-                sliding_window=window_size,
+                sliding_window=window_size if sliding_window is not None else -1,
                 block_table=block_tables,
                 softcap=0,
                 scheduler_metadata=metadata,

benchmarks/replayssm/e2e_decode_speedup.py

@@ -0,0 +1,267 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""End-to-end autoregressive decode benchmark: ReplaySSM vs the standard SSM kernel.
+
+Loads a hybrid Mamba2 model, replicates one prompt across the batch, and times a
+long greedy decode (CUDA graphs on) once with the standard kernel and once with
+ReplaySSM, then reports the per-step / throughput speedup. The two modes run in
+separate subprocesses so each gets a clean CUDA context.
+
+The FlashInfer FP4-MoE autotuner is disabled by default (it is unstable under
+CUDA-graph capture on the pre-release Blackwell FP4 path); pass
+--no-disable-flashinfer-autotune for non-FP4 models.
+
+Examples:
+    python e2e_decode_speedup.py --model-id nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
+    python e2e_decode_speedup.py --dtype auto --buffer-len 16 \
+        --model-id nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4   # B300 NVFP4
+"""
+
+import argparse
+import json
+import os
+import subprocess
+import sys
+import time
+
+DEFAULT_PROMPT = "My cat wrote all this CUDA code for a new language model and"
+
+MODE_LABEL = {"standard": "standard", "replayssm": "ReplaySSM"}
+
+
+def parse_args():
+    p = argparse.ArgumentParser(
+        description="E2E decode speedup: ReplaySSM vs the standard SSM kernel."
+    )
+    p.add_argument("--model-id", default="nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16")
+    p.add_argument("--prompt", default=DEFAULT_PROMPT)
+    p.add_argument("--batch-size", type=int, default=256)
+    p.add_argument("--num-steps", type=int, default=1000)
+    p.add_argument("--warmup-steps", type=int, default=128)
+    p.add_argument("--repeats", type=int, default=1)
+    p.add_argument(
+        "--buffer-len", type=int, default=16, help="ReplaySSM input-buffer length."
+    )
+    p.add_argument(
+        "--dtype",
+        default="bfloat16",
+        choices=["bfloat16", "float16", "float32", "auto"],
+    )
+    p.add_argument("--gpu-memory-utilization", type=float, default=0.9)
+    p.add_argument("--max-model-len", type=int, default=None)
+    p.add_argument(
+        "--disable-flashinfer-autotune",
+        action=argparse.BooleanOptionalAction,
+        default=True,
+        help="Disable the FlashInfer FP4-MoE autotuner (default: on). "
+        "It is unstable under CUDA-graph capture on the "
+        "pre-release Blackwell FP4 path; pass "
+        "--no-disable-flashinfer-autotune for non-FP4 models.",
+    )
+    p.add_argument(
+        "--mamba-ssm-cache-dtype",
+        default="auto",
+        choices=["auto", "float32", "float16", "bfloat16"],
+        help="SSM state dtype (both modes). 'auto' = config-driven; "
+        "'float32' = fp32 state, 'bfloat16' = s16 state.",
+    )
+    p.add_argument(
+        "--baseline-ssm-config",
+        default="",
+        help="Pin the STANDARD baseline's SSM launch config as "
+        "'bsm,nw' via override_ssm_config (forces the in-process "
+        "engine so the override reaches the kernel). Empty = off.",
+    )
+    p.add_argument(
+        "--worker",
+        choices=["standard", "replayssm"],
+        default=None,
+        help=argparse.SUPPRESS,
+    )
+    return p.parse_args()
+
+
+def resolve_max_model_len(args) -> int:
+    if args.max_model_len is not None:
+        return args.max_model_len
+    return args.num_steps + 256
+
+
+def run_worker(args):
+    # override_ssm_config is a module global; it only reaches the model if the
+    # engine runs in-process (default V1 spawns a separate EngineCore). Force it.
+    if args.worker == "standard" and args.baseline_ssm_config:
+        os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"
+
+    import torch
+
+    from vllm import LLM, SamplingParams
+
+    mode = args.worker
+    max_model_len = resolve_max_model_len(args)
+
+    llm_kwargs = dict(
+        model=args.model_id,
+        tensor_parallel_size=1,
+        dtype=args.dtype,
+        max_model_len=max_model_len,
+        trust_remote_code=True,
+        enable_prefix_caching=False,
+        enable_chunked_prefill=False,
+        max_num_seqs=args.batch_size,
+        max_num_batched_tokens=max(max_model_len, args.batch_size * 64),
+        enforce_eager=False,
+        disable_log_stats=True,
+        gpu_memory_utilization=args.gpu_memory_utilization,
+        # SSM state dtype (applies to both standard and ReplaySSM).
+        mamba_ssm_cache_dtype=args.mamba_ssm_cache_dtype,
+    )
+    if args.disable_flashinfer_autotune:
+        # FP4-MoE autotuner is unstable under CUDA-graph capture on Blackwell;
+        # re-enable (--no-disable-flashinfer-autotune) only for non-FP4 models.
+        llm_kwargs["kernel_config"] = {"enable_flashinfer_autotune": False}
+    if mode == "replayssm":
+        llm_kwargs.update(use_replayssm=True, replayssm_buffer_len=args.buffer_len)
+
+    _ssm_cm = None
+    if mode == "standard" and args.baseline_ssm_config:
+        from vllm.model_executor.layers.mamba.ops.mamba_ssm import override_ssm_config
+
+        _bsm, _nw = (int(x) for x in args.baseline_ssm_config.split(","))
+        _ssm_cm = override_ssm_config((_bsm, _nw))
+        _ssm_cm.__enter__()  # active through LLM() graph capture + decode
+        print(
+            f"[{mode}] override_ssm_config -> (BLOCK_SIZE_M={_bsm}, num_warps={_nw})",
+            flush=True,
+        )
+
+    llm = LLM(**llm_kwargs)
+    prompts = [args.prompt] * args.batch_size
+
+    def timed_generate(n_tokens):
+        sp = SamplingParams(
+            n=1,
+            temperature=0.0,
+            ignore_eos=True,
+            min_tokens=n_tokens,
+            max_tokens=n_tokens,
+        )
+        if torch.accelerator.is_available():
+            torch.accelerator.synchronize()
+        t0 = time.perf_counter()
+        outs = llm.generate(prompts, sp, use_tqdm=False)
+        if torch.accelerator.is_available():
+            torch.accelerator.synchronize()
+        elapsed = time.perf_counter() - t0
+        produced = min(len(o.outputs[0].token_ids) for o in outs)
+        assert produced == n_tokens, f"expected {n_tokens} tokens, got {produced}"
+        return elapsed
+
+    timed_generate(args.warmup_steps)
+
+    best = None
+    for _ in range(args.repeats):
+        elapsed = timed_generate(args.num_steps)
+        tok_s = args.batch_size * args.num_steps / elapsed
+        per_step_ms = elapsed / args.num_steps * 1e3
+        print(
+            f"[{mode}] {elapsed:.3f}s  {tok_s:,.0f} tok/s  {per_step_ms:.3f} ms/step",
+            flush=True,
+        )
+        if best is None or elapsed < best["elapsed_s"]:
+            best = {
+                "mode": mode,
+                "elapsed_s": elapsed,
+                "tok_s": tok_s,
+                "per_step_ms": per_step_ms,
+            }
+
+    print("RESULT_JSON " + json.dumps(best), flush=True)
+    if _ssm_cm is not None:
+        _ssm_cm.__exit__(None, None, None)
+
+
+def run_one_mode(args, mode) -> dict:
+    cmd = [
+        sys.executable,
+        __file__,
+        "--worker",
+        mode,
+        "--model-id",
+        args.model_id,
+        "--prompt",
+        args.prompt,
+        "--batch-size",
+        str(args.batch_size),
+        "--num-steps",
+        str(args.num_steps),
+        "--warmup-steps",
+        str(args.warmup_steps),
+        "--repeats",
+        str(args.repeats),
+        "--buffer-len",
+        str(args.buffer_len),
+        "--dtype",
+        args.dtype,
+        "--gpu-memory-utilization",
+        str(args.gpu_memory_utilization),
+        "--mamba-ssm-cache-dtype",
+        args.mamba_ssm_cache_dtype,
+        "--baseline-ssm-config",
+        args.baseline_ssm_config,
+    ]
+    cmd.append(
+        "--disable-flashinfer-autotune"
+        if args.disable_flashinfer_autotune
+        else "--no-disable-flashinfer-autotune"
+    )
+    if args.max_model_len is not None:
+        cmd += ["--max-model-len", str(args.max_model_len)]
+
+    result = None
+    proc = subprocess.Popen(
+        cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True, bufsize=1
+    )
+    for line in proc.stdout:
+        sys.stdout.write(line)
+        sys.stdout.flush()
+        if line.startswith("RESULT_JSON "):
+            result = json.loads(line[len("RESULT_JSON ") :])
+    proc.wait()
+    if proc.returncode != 0:
+        raise RuntimeError(f"mode '{mode}' worker exited with {proc.returncode}")
+    if result is None:
+        raise RuntimeError(f"mode '{mode}' produced no RESULT_JSON line")
+    return result
+
+
+def main():
+    args = parse_args()
+    if args.worker is not None:
+        run_worker(args)
+        return
+
+    print(
+        f"model={args.model_id}  batch_size={args.batch_size}  "
+        f"steps={args.num_steps}  buffer_len={args.buffer_len}  dtype={args.dtype}"
+    )
+
+    std = run_one_mode(args, "standard")
+    fla = run_one_mode(args, "replayssm")
+    speedup = std["per_step_ms"] / fla["per_step_ms"]
+
+    print()
+    header = f"{'mode':<10}{'ms/step':>12}{'tok/s':>16}{'wall (s)':>12}"
+    print(header)
+    print("-" * len(header))
+    for r in (std, fla):
+        print(
+            f"{MODE_LABEL[r['mode']]:<10}{r['per_step_ms']:>12.3f}"
+            f"{r['tok_s']:>16,.0f}{r['elapsed_s']:>12.3f}"
+        )
+    print("-" * len(header))
+    print(f"speedup (standard / ReplaySSM, per step): {speedup:.3f}x")
+
+
+if __name__ == "__main__":
+    main()

cmake/cpu_extension.cmake

@@ -15,6 +15,7 @@ endif()
 #
 set(ENABLE_X86_ISA $ENV{VLLM_CPU_X86})
 set(ENABLE_ARM_BF16 $ENV{VLLM_CPU_ARM_BF16})
+set(ENABLE_ARM_I8MM $ENV{VLLM_CPU_ARM_I8MM})
 set(ENABLE_RVV_BF16 $ENV{VLLM_CPU_RVV_BF16})
 
 include_directories("${CMAKE_SOURCE_DIR}/csrc")
@@ -96,12 +97,14 @@ if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
     set(ENABLE_NUMA OFF)
     check_sysctl(hw.optional.neon ASIMD_FOUND)
     check_sysctl(hw.optional.arm.FEAT_BF16 ARM_BF16_FOUND)
+    check_sysctl(hw.optional.arm.FEAT_I8MM ARM_I8MM_FOUND)
 else()
     find_isa(${CPUINFO} "Power11" POWER11_FOUND)
     find_isa(${CPUINFO} "POWER10" POWER10_FOUND)
     find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
     find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
     find_isa(${CPUINFO} "bf16" ARM_BF16_FOUND) # Check for ARM BF16 support
+    find_isa(${CPUINFO} "i8mm" ARM_I8MM_FOUND) # Check for ARM I8MM support
     find_isa(${CPUINFO} "S390" S390_FOUND)
     find_isa(${CPUINFO} "zvfhmin" RVV_FP16_FOUND) # Check for RISC-V Vector FP16 support
     find_isa(${CPUINFO} "zvfbfmin" RVV_BF16_FOUND) # Check for RISC-V Vector BF16 support
@@ -111,6 +114,11 @@ else()
         set(ARM_BF16_FOUND ON)
         message(STATUS "ARM BF16 support enabled via VLLM_CPU_ARM_BF16 environment variable")
     endif()
+    if (ENABLE_ARM_I8MM)
+        set(ARM_I8MM_FOUND ON)
+        message(STATUS
+            "ARM I8MM support enabled via VLLM_CPU_ARM_I8MM environment variable")
+    endif()
     # Some kernels (e.g. Bianbu on Spacemit X100) do not report zvfbfmin
     # in /proc/cpuinfo despite hardware support. VLLM_CPU_RVV_BF16=1
     # overrides the detection result.
@@ -166,6 +174,11 @@ elseif (ASIMD_FOUND)
         message(WARNING "BF16 functionality is not available")
         set(MARCH_FLAGS "-march=armv8.2-a+dotprod+fp16")  
     endif()
+    if(ARM_I8MM_FOUND)
+        message(STATUS "I8MM extension detected")
+        string(APPEND MARCH_FLAGS "+i8mm")
+        add_compile_definitions(ARM_I8MM_SUPPORT)
+    endif()
     list(APPEND CXX_COMPILE_FLAGS ${MARCH_FLAGS})     
 elseif (S390_FOUND)
     message(STATUS "S390 detected")
@@ -430,6 +443,7 @@ set(VLLM_EXT_SRC
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
+    "csrc/cpu/mamba_cpu.cpp"
     "csrc/moe/dynamic_4bit_int_moe_cpu.cpp"
     "csrc/cpu/cpu_attn.cpp"
     "csrc/cpu/torch_bindings.cpp")
@@ -446,8 +460,13 @@ if (ASIMD_FOUND AND NOT APPLE_SILICON_FOUND)
         "csrc/cpu/shm.cpp"
         "csrc/cpu/activation_lut_bf16.cpp"
         "csrc/cpu/cpu_tanhf_neon.hpp"
-        "csrc/cpu/cpu_fused_moe.cpp"
         ${VLLM_EXT_SRC})
+    if (ARM_BF16_FOUND)
+        set(VLLM_EXT_SRC "csrc/cpu/cpu_fused_moe.cpp" ${VLLM_EXT_SRC})
+        if (ARM_I8MM_FOUND)
+            set(VLLM_EXT_SRC "csrc/cpu/cpu_fused_moe_int8.cpp" ${VLLM_EXT_SRC})
+        endif()
+    endif()
 endif()
 
 if (POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)	
@@ -489,6 +508,7 @@ if (ENABLE_X86_ISA)
         "csrc/cpu/spec_decode_utils.cpp"
         "csrc/cpu/cpu_attn.cpp"
         "csrc/cpu/dnnl_kernels.cpp"
+        "csrc/cpu/mamba_cpu.cpp"
         "csrc/cpu/torch_bindings.cpp"
         # TODO: Remove these files
         "csrc/cpu/activation.cpp"
@@ -502,6 +522,7 @@ if (ENABLE_X86_ISA)
         "csrc/cpu/utils.cpp"
         "csrc/cpu/spec_decode_utils.cpp"
         "csrc/cpu/cpu_attn.cpp"
+        "csrc/cpu/mamba_cpu.cpp"
         "csrc/cpu/dnnl_kernels.cpp"
         "csrc/cpu/torch_bindings.cpp"
         # TODO: Remove these files

cmake/external_projects/deepgemm.cmake

@@ -68,6 +68,9 @@ endif()
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.9)
     list(APPEND DEEPGEMM_SUPPORT_ARCHS "10.0f")
+    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.4)
+      list(APPEND DEEPGEMM_SUPPORT_ARCHS "10.7f")
+    endif()
   else()
     list(APPEND DEEPGEMM_SUPPORT_ARCHS "10.0a")
   endif()

cmake/external_projects/flashmla.cmake

@@ -19,7 +19,7 @@ else()
   FetchContent_Declare(
         flashmla
         GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
-        GIT_TAG b70aff3d110a2b1a037e62eac295166b5143643a
+        GIT_TAG a8f794d1251cbfd88a5011445dd5582289c727e4
         GIT_PROGRESS TRUE
         CONFIGURE_COMMAND ""
         BUILD_COMMAND ""
@@ -35,7 +35,7 @@ set(FLASHMLA_VENDOR_DIR "${CMAKE_SOURCE_DIR}/vllm/third_party/flashmla")
 file(MAKE_DIRECTORY "${FLASHMLA_VENDOR_DIR}")
 file(READ "${flashmla_SOURCE_DIR}/flash_mla/flash_mla_interface.py"
      FLASHMLA_INTERFACE_CONTENT)
-string(REPLACE "import flash_mla.cuda as flash_mla_cuda"
+string(REPLACE "flash_mla_cuda = torch.ops._flashmla_C"
                "import vllm._flashmla_C\nflash_mla_cuda = torch.ops._flashmla_C"
                FLASHMLA_INTERFACE_CONTENT
                "${FLASHMLA_INTERFACE_CONTENT}")
@@ -60,6 +60,9 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.9)
     # CUDA 12.9 has introduced "Family-Specific Architecture Features"
     # this supports all compute_10x family
     list(APPEND SUPPORT_ARCHS "10.0f")
+    if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.4)
+        list(APPEND SUPPORT_ARCHS "10.7f")
+    endif()
 elseif(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8)
     list(APPEND SUPPORT_ARCHS "10.0a")
 endif()
@@ -72,7 +75,7 @@ if(FLASH_MLA_ARCHS)
     list(APPEND VLLM_FLASHMLA_GPU_FLAGS "--expt-relaxed-constexpr" "--expt-extended-lambda" "--use_fast_math")
 
     set(FlashMLA_SOURCES
-        ${flashmla_SOURCE_DIR}/csrc/torch_api.cpp
+        ${flashmla_SOURCE_DIR}/csrc/api/api.cpp
 
         # Misc kernels for decoding
         ${flashmla_SOURCE_DIR}/csrc/smxx/decode/get_decoding_sched_meta/get_decoding_sched_meta.cu
@@ -128,6 +131,7 @@ if(FLASH_MLA_ARCHS)
 
     set(FlashMLA_Extension_INCLUDES
         ${flashmla_SOURCE_DIR}/csrc
+        ${flashmla_SOURCE_DIR}/csrc/kerutils/include
         ${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/
         ${flashmla_SOURCE_DIR}/csrc/cutlass/include
         ${flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
@@ -152,15 +156,18 @@ if(FLASH_MLA_ARCHS)
         USE_SABI 3
         WITH_SOABI)
 
-    # Keep Stable ABI for the module, but *not* for CUDA/C++ files.
-    # This prevents Py_LIMITED_API from affecting nvcc and C++ compiles.
-    # Also enable C++20 for the FlashMLA sources (required for std::span, requires, etc.)
+    # Enable C++20 for the FlashMLA sources (required for std::span, requires, etc.)
     target_compile_options(_flashmla_C PRIVATE
-        $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
-        $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>
         $<$<COMPILE_LANGUAGE:CXX>:-std=c++20>
         $<$<COMPILE_LANGUAGE:CUDA>:-std=c++20>)
 
+    # _flashmla_C is now ABI-stable torch 2.11+
+    target_compile_definitions(_flashmla_C PRIVATE
+        TORCH_TARGET_VERSION=0x020B000000000000ULL)
+    if(VLLM_GPU_LANG STREQUAL "CUDA")
+        target_compile_definitions(_flashmla_C PRIVATE USE_CUDA)
+    endif()
+
     define_extension_target(
         _flashmla_extension_C
         DESTINATION vllm
@@ -172,15 +179,15 @@ if(FLASH_MLA_ARCHS)
         USE_SABI 3
         WITH_SOABI)
 
-    # Keep Stable ABI for the module, but *not* for CUDA/C++ files.
-    # This prevents Py_LIMITED_API from affecting nvcc and C++ compiles.
-    target_compile_options(_flashmla_extension_C PRIVATE
-        $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
-        $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
+    # _flashmla_extension_C is now ABI-stable w/ torch 2.11+
+    target_compile_definitions(_flashmla_extension_C PRIVATE
+        TORCH_TARGET_VERSION=0x020B000000000000ULL)
+    if(VLLM_GPU_LANG STREQUAL "CUDA")
+        target_compile_definitions(_flashmla_extension_C PRIVATE USE_CUDA)
+    endif()
 else()
     message(STATUS "FlashMLA will not compile: unsupported CUDA architecture ${CUDA_ARCHS}")
     # Create empty targets for setup.py on unsupported systems
     add_custom_target(_flashmla_C)
     add_custom_target(_flashmla_extension_C)
 endif()
-

cmake/external_projects/fmha_sm100.cmake

@@ -17,7 +17,7 @@ else()
   FetchContent_Declare(
     fmha_sm100
     GIT_REPOSITORY https://github.com/vllm-project/MSA.git
-    GIT_TAG 2e63ec37a0fc29bc20f39cd1a52e0f5affc33a73
+    GIT_TAG 890aaa1a37a598ad17ccff0827fea21540d381fa
     GIT_PROGRESS TRUE
     CONFIGURE_COMMAND ""
     BUILD_COMMAND ""

cmake/external_projects/qutlass.cmake

@@ -22,7 +22,7 @@ if(QUTLASS_SRC_DIR)
   set(qutlass_BINARY_DIR "${CMAKE_BINARY_DIR}/qutlass-binary-dir-unused")
 else()
   set(_QUTLASS_UPSTREAM_REPO "https://github.com/IST-DASLab/qutlass.git")
-  set(_QUTLASS_UPSTREAM_TAG "830d2c4537c7396e14a02a46fbddd18b5d107c65")
+  set(_QUTLASS_UPSTREAM_TAG "e74319e3405ce6d71965732880f5dc1f52371f64")
 
   set(_qutlass_fc_root "${FETCHCONTENT_BASE_DIR}")
   if(NOT _qutlass_fc_root)
@@ -55,7 +55,11 @@ message(STATUS "[QUTLASS] QuTLASS is available at ${qutlass_SOURCE_DIR}")
 
 if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
   cuda_archs_loose_intersection(QUTLASS_SM120_ARCHS "12.0f" "${CUDA_ARCHS}")
-  cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.4)
+    cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f;10.7f" "${CUDA_ARCHS}")
+  else()
+    cuda_archs_loose_intersection(QUTLASS_SM100_ARCHS "10.0f" "${CUDA_ARCHS}")
+  endif()
 else()
   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}")
@@ -125,8 +129,6 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
     CUDA_ARCHS "${QUTLASS_ARCHS}"
   )
 
-  # QuTLASS uses legacy ATen headers and cannot be built with TORCH_TARGET_VERSION.
-  # Keep it as its own extension (registers torch.ops._qutlass_C).
   define_extension_target(
     _qutlass_C
     DESTINATION vllm
@@ -139,9 +141,11 @@ if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND QUTLASS_ARCHS)
     WITH_SOABI)
 
   target_compile_definitions(_qutlass_C PRIVATE
-    QUTLASS_DISABLE_PYBIND=1
+    QUTLASS_MINIMAL_BUILD=1
     TARGET_CUDA_ARCH=${QUTLASS_TARGET_CC}
-    CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
+    CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1
+    TORCH_TARGET_VERSION=0x020B000000000000ULL
+    USE_CUDA)
 
   set_property(SOURCE ${QUTLASS_SOURCES} APPEND PROPERTY COMPILE_OPTIONS
     $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr --use_fast_math -O3>

cmake/external_projects/tml_fa4.cmake

@@ -0,0 +1,50 @@
+include(FetchContent)
+
+if(DEFINED ENV{TML_FA4_SRC_DIR})
+  set(TML_FA4_SRC_DIR $ENV{TML_FA4_SRC_DIR})
+endif()
+
+if(TML_FA4_SRC_DIR)
+  FetchContent_Declare(
+    tml_fa4
+    SOURCE_DIR ${TML_FA4_SRC_DIR}
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND "")
+else()
+  FetchContent_Declare(
+    tml_fa4
+    GIT_REPOSITORY https://github.com/vllm-project/tml-fa4.git
+    GIT_TAG b206834606ed5b5f21f8eed6b0683f528ea9cf7d
+    GIT_PROGRESS TRUE
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND "")
+endif()
+
+FetchContent_GetProperties(tml_fa4)
+if(NOT tml_fa4_POPULATED)
+  FetchContent_Populate(tml_fa4)
+endif()
+message(STATUS "tml-fa4 is available at ${tml_fa4_SOURCE_DIR}")
+
+add_custom_target(tml_fa4)
+
+# Install into a private namespace so this implementation cannot shadow the
+# flash_attn package used by vLLM's standard attention backends.
+install(CODE "
+  file(GLOB_RECURSE TML_FA4_PY_FILES
+    \"${tml_fa4_SOURCE_DIR}/flash_attn/cute/*.py\")
+  foreach(SRC_FILE \${TML_FA4_PY_FILES})
+    file(RELATIVE_PATH REL_PATH
+      \"${tml_fa4_SOURCE_DIR}/flash_attn/cute\" \${SRC_FILE})
+    set(DST_FILE
+      \"\${CMAKE_INSTALL_PREFIX}/vllm/third_party/tml_fa4/\${REL_PATH}\")
+    get_filename_component(DST_DIR \${DST_FILE} DIRECTORY)
+    file(MAKE_DIRECTORY \${DST_DIR})
+    file(READ \${SRC_FILE} FILE_CONTENTS)
+    string(REPLACE
+      \"flash_attn.cute\"
+      \"vllm.third_party.tml_fa4\"
+      FILE_CONTENTS \"\${FILE_CONTENTS}\")
+    file(WRITE \${DST_FILE} \"\${FILE_CONTENTS}\")
+  endforeach()
+" COMPONENT tml_fa4)

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 bb9a72e7dde0dc614ffc663e052cd6a19ce73a42
+          GIT_TAG ed4b7342bc8f0489dd9b649d5288867e35fc6a32
           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

@@ -396,14 +396,24 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   # match — e.g. SRC="12.0f" matches TGT="12.1a" since SM121 is in the SM12x
   # family. The output uses TGT's value to preserve the user's compilation flags.
   set(_CUDA_ARCHS)
+  # Resolve exact base matches before family fallbacks so a generic entry such
+  # as 10.0f cannot consume a 10.7 target that has a 10.7f source entry.
+  foreach(_arch ${_SRC_CUDA_ARCHS})
+    if(_arch MATCHES "[af]$")
+      string(REGEX REPLACE "[af]$" "" _base "${_arch}")
+      if("${_base}" IN_LIST _TGT_CUDA_ARCHS)
+        list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
+        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
+        list(APPEND _CUDA_ARCHS "${_arch}")
+      endif()
+    endif()
+  endforeach()
+
   foreach(_arch ${_SRC_CUDA_ARCHS})
     if(_arch MATCHES "[af]$")
       list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
       string(REGEX REPLACE "[af]$" "" _base "${_arch}")
-      if ("${_base}" IN_LIST TGT_CUDA_ARCHS)
-        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
-        list(APPEND _CUDA_ARCHS "${_arch}")
-      elseif("${_base}a" IN_LIST _TGT_CUDA_ARCHS)
+      if("${_base}a" IN_LIST _TGT_CUDA_ARCHS)
         list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}a")
         list(APPEND _CUDA_ARCHS "${_base}a")
       elseif("${_base}f" IN_LIST _TGT_CUDA_ARCHS)
@@ -487,7 +497,7 @@ endfunction()
 
 function(cuda_archs_sm90plus OUT_CUDA_ARCHS TGT_CUDA_ARCHS)
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(_archs "9.0a;10.0f;11.0f;12.0f" "${TGT_CUDA_ARCHS}")
+    cuda_archs_loose_intersection(_archs "9.0a;10.0f;10.7f;11.0f;12.0f" "${TGT_CUDA_ARCHS}")
   else()
     cuda_archs_loose_intersection(_archs "9.0a;10.0a;10.1a;10.3a;12.0a;12.1a" "${TGT_CUDA_ARCHS}")
   endif()

csrc/cache.h

@@ -67,9 +67,8 @@ void cp_gather_and_upconvert_fp8_kv_cache(
     torch::Tensor const& src_cache,         // [NUM_BLOCKS, BLOCK_SIZE, 656]
     torch::Tensor const& dst,               // [TOT_TOKENS, 576]
     torch::Tensor const& block_table,       // [BATCH, BLOCK_INDICES]
-    torch::Tensor const& seq_lens,          // [BATCH]
     torch::Tensor const& workspace_starts,  // [BATCH]
-    int64_t batch_size);
+    int64_t batch_size, std::optional<torch::Tensor> seq_starts = std::nullopt);
 
 // Indexer K quantization and cache function
 void indexer_k_quant_and_cache(

csrc/cpu/cpu_arch_macros.h

@@ -172,4 +172,15 @@
 
 #endif  // __riscv_v
 
+// Power VSX
+#ifdef __powerpc__
+  // FP32Vec16::exp() in cpu_types_vsx.hpp delegates to FP32Vec8::exp(), which
+  // implements a vectorised 5-term minimax polynomial using VSX intrinsics.
+  #define DEFINE_FAST_EXP                                                     \
+    auto fast_exp = [&](const vec_op::FP32Vec16& vec)                         \
+                        __attribute__((always_inline)) { return vec.exp(); }; \
+    auto fast_exp_f16 = fast_exp;
+
+#endif  // __powerpc__
+
 #endif

csrc/cpu/cpu_attn_vec.hpp

@@ -102,7 +102,9 @@ class TileGemm82 {
     kv_cache_t* __restrict__ curr_b = b_tile;
 
     for (int32_t k = 0; k < dynamic_k_size; ++k) {
-      auto [fp32_b_0_reg, fp32_b_1_reg] = load_b_pair_vec(curr_b);
+      auto fp32_b_regs = load_b_pair_vec(curr_b);
+      auto fp32_b_0_reg = fp32_b_regs.first;
+      auto fp32_b_1_reg = fp32_b_regs.second;
 
       float* __restrict__ curr_m_a = curr_a;
       vec_op::unroll_loop<int32_t, M>([&](int32_t i) {

csrc/cpu/cpu_fused_moe.cpp

@@ -1,5 +1,6 @@
 #include "cpu/cpu_types.hpp"
 #include "cpu/utils.hpp"
+#include "cpu/cpu_fused_moe_activations.hpp"
 #include "cpu/micro_gemm/cpu_micro_gemm_vec.hpp"
 #include "cpu/cpu_arch_macros.h"
 
@@ -43,193 +44,9 @@
   }()
 
 namespace {
-enum class FusedMOEAct {
-  SiluAndMul,
-  SwigluOAIAndMul,
-  GeluAndMul,
-  GeluTanhAndMul,
-};
 
-FusedMOEAct get_act_type(const std::string& act) {
-  if (act == "silu") {
-    return FusedMOEAct::SiluAndMul;
-  } else if (act == "swigluoai") {
-    return FusedMOEAct::SwigluOAIAndMul;
-  } else if (act == "gelu") {
-    return FusedMOEAct::GeluAndMul;
-  } else if (act == "gelu_tanh") {
-    return FusedMOEAct::GeluTanhAndMul;
-  } else {
-    TORCH_CHECK(false, "Invalid act type: " + act);
-  }
-}
-
-template <typename scalar_t>
-void swigluoai_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                       const int32_t m_size, const int32_t n_size,
-                       const int32_t input_stride,
-                       const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-#if !defined(__aarch64__)
-  // For GPT-OSS interleaved gate-up weights
-  alignas(64) static int32_t index[16] = {0,  2,  4,  6,  8,  10, 12, 14,
-                                          16, 18, 20, 22, 24, 26, 28, 30};
-  vec_op::INT32Vec16 index_vec(index);
-#endif
-  vec_op::FP32Vec16 gate_up_max_vec(7.0);
-  vec_op::FP32Vec16 up_min_vec(-7.0);
-  vec_op::FP32Vec16 alpha_vec(1.702);
-  vec_op::FP32Vec16 one_vec(1.0);
-
-  DEFINE_FAST_EXP
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < n_size; n += 32) {
-      // Note: AdvSIMD does not support gather loads
-#if defined(__aarch64__)
-      vec_op::FP32Vec16 gate_vec(vec_op::uninit);
-      vec_op::FP32Vec16 up_vec(vec_op::uninit);
-      vec_op::FP32Vec16::load_even_odd(input + n, gate_vec, up_vec);
-#else
-      vec_op::FP32Vec16 gate_vec(input + n, index_vec);
-      vec_op::FP32Vec16 up_vec(input + n + 1, index_vec);
-#endif
-      gate_vec = gate_vec.min(gate_up_max_vec);
-      up_vec = up_vec.clamp(up_min_vec, gate_up_max_vec);
-      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec * alpha_vec));
-      auto glu = gate_vec * sigmoid_vec;
-      auto gated_output_fp32 = (one_vec + up_vec) * glu;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n / 2);
-    }
-    input += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-void silu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                  const int32_t m_size, const int32_t n_size,
-                  const int32_t input_stride, const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  const int32_t dim = n_size / 2;
-  float* __restrict__ gate = input;
-  float* __restrict__ up = input + dim;
-  vec_op::FP32Vec16 one_vec(1.0);
-
-  DEFINE_FAST_EXP
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < dim; n += 16) {
-      vec_op::FP32Vec16 gate_vec(gate + n);
-      vec_op::FP32Vec16 up_vec(up + n);
-      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec));
-      auto silu = gate_vec * sigmoid_vec;
-      auto gated_output_fp32 = up_vec * silu;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n);
-    }
-    gate += input_stride;
-    up += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-void gelu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                  const int32_t m_size, const int32_t n_size,
-                  const int32_t input_stride, const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  const int32_t dim = n_size / 2;
-  float* __restrict__ gate = input;
-  float* __restrict__ up = input + dim;
-  vec_op::FP32Vec16 one_vec(1.0);
-  vec_op::FP32Vec16 w1_vec(M_SQRT1_2);
-  vec_op::FP32Vec16 w2_vec(0.5);
-  alignas(64) float temp[16];
-
-  DEFINE_FAST_EXP
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < dim; n += 16) {
-      vec_op::FP32Vec16 gate_vec(gate + n);
-      vec_op::FP32Vec16 up_vec(up + n);
-      auto er_input_vec = gate_vec * w1_vec;
-
-      er_input_vec.save(temp);
-      for (int32_t i = 0; i < 16; ++i) {
-        temp[i] = std::erf(temp[i]);
-      }
-      vec_op::FP32Vec16 er_vec(temp);
-      auto gelu = gate_vec * w2_vec * (one_vec + er_vec);
-      auto gated_output_fp32 = up_vec * gelu;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n);
-    }
-    gate += input_stride;
-    up += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-void gelu_tanh_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
-                       const int32_t m_size, const int32_t n_size,
-                       const int32_t input_stride,
-                       const int32_t output_stride) {
-  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
-  const int32_t dim = n_size / 2;
-  float* __restrict__ gate = input;
-  float* __restrict__ up = input + dim;
-  vec_op::FP32Vec16 one_vec(1.0);
-  vec_op::FP32Vec16 w1_vec(0.7978845608028654);
-  vec_op::FP32Vec16 w2_vec(0.5);
-  vec_op::FP32Vec16 w3_vec(0.044715);
-
-  for (int32_t m = 0; m < m_size; ++m) {
-    for (int32_t n = 0; n < dim; n += 16) {
-      vec_op::FP32Vec16 gate_vec(gate + n);
-      vec_op::FP32Vec16 up_vec(up + n);
-      auto gate_pow3_vec = gate_vec * gate_vec * gate_vec;
-      auto inner_vec = w1_vec * (gate_vec + w3_vec * gate_pow3_vec);
-      // Note: can't use fast_exp form because diffusiongemma will generate
-      // wrong results
-      auto tanh_vec = inner_vec.tanh();
-      auto gelu_tanh = gate_vec * w2_vec * (one_vec + tanh_vec);
-      auto gated_output_fp32 = up_vec * gelu_tanh;
-      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
-      gated_output.save(output + n);
-    }
-    gate += input_stride;
-    up += input_stride;
-    output += output_stride;
-  }
-}
-
-template <typename scalar_t>
-FORCE_INLINE void apply_gated_act(const FusedMOEAct act,
-                                  float* __restrict__ input,
-                                  scalar_t* __restrict__ output,
-                                  const int32_t m, const int32_t n,
-                                  const int32_t input_stride,
-                                  const int32_t output_stride) {
-  switch (act) {
-    case FusedMOEAct::SwigluOAIAndMul:
-      swigluoai_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    case FusedMOEAct::SiluAndMul:
-      silu_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    case FusedMOEAct::GeluAndMul:
-      gelu_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    case FusedMOEAct::GeluTanhAndMul:
-      gelu_tanh_and_mul(input, output, m, n, input_stride, output_stride);
-      return;
-    default:
-      TORCH_CHECK(false, "Unsupported act type.");
-  }
-}
+using cpu_fused_moe_utils::apply_gated_act;
+using cpu_fused_moe_utils::FusedMOEAct;
 
 template <typename scalar_t, typename gemm_t>
 void prepack_moe_weight_impl(scalar_t* __restrict__ weight_ptr,
@@ -817,6 +634,7 @@ void fused_moe_impl(scalar_t* __restrict__ output, scalar_t* __restrict__ input,
     }
   }
 }
+
 }  // namespace
 
 void prepack_moe_weight(
@@ -864,7 +682,7 @@ void cpu_fused_moe(
   const int32_t input_size_2 = w2.size(2);
   const int32_t output_size_2 = w2.size(1);
   const int32_t topk_num = topk_id.size(1);
-  const FusedMOEAct act_type = get_act_type(act);
+  const FusedMOEAct act_type = cpu_fused_moe_utils::get_act_type(act);
   cpu_utils::ISA isa_type = cpu_utils::get_isa(isa);
   TORCH_CHECK(!skip_weighted || topk_num == 1,
               "skip_weighted is only supported for topk=1 on CPU");

csrc/cpu/cpu_fused_moe_activations.hpp

@@ -0,0 +1,204 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#ifndef CPU_FUSED_MOE_ACTIVATIONS_HPP
+#define CPU_FUSED_MOE_ACTIVATIONS_HPP
+
+#include <cmath>
+#include <cstdint>
+#include <string>
+
+#include "cpu/cpu_arch_macros.h"
+#include "cpu/utils.hpp"
+
+namespace cpu_fused_moe_utils {
+enum class FusedMOEAct {
+  SiluAndMul,
+  SwigluOAIAndMul,
+  GeluAndMul,
+  GeluTanhAndMul,
+};
+
+inline FusedMOEAct get_act_type(const std::string& act) {
+  if (act == "silu") {
+    return FusedMOEAct::SiluAndMul;
+  } else if (act == "swigluoai") {
+    return FusedMOEAct::SwigluOAIAndMul;
+  } else if (act == "gelu") {
+    return FusedMOEAct::GeluAndMul;
+  } else if (act == "gelu_tanh") {
+    return FusedMOEAct::GeluTanhAndMul;
+  } else {
+    TORCH_CHECK(false, "Invalid act type: " + act);
+  }
+}
+
+template <typename scalar_t>
+void swigluoai_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                       const int32_t m_size, const int32_t n_size,
+                       const int32_t input_stride,
+                       const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+#if !defined(__aarch64__)
+  // For GPT-OSS interleaved gate-up weights
+  alignas(64) static int32_t index[16] = {0,  2,  4,  6,  8,  10, 12, 14,
+                                          16, 18, 20, 22, 24, 26, 28, 30};
+  vec_op::INT32Vec16 index_vec(index);
+#endif
+  vec_op::FP32Vec16 gate_up_max_vec(7.0);
+  vec_op::FP32Vec16 up_min_vec(-7.0);
+  vec_op::FP32Vec16 alpha_vec(1.702);
+  vec_op::FP32Vec16 one_vec(1.0);
+
+  DEFINE_FAST_EXP
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < n_size; n += 32) {
+      // Note: AdvSIMD does not support gather loads
+#if defined(__aarch64__)
+      vec_op::FP32Vec16 gate_vec(vec_op::uninit);
+      vec_op::FP32Vec16 up_vec(vec_op::uninit);
+      vec_op::FP32Vec16::load_even_odd(input + n, gate_vec, up_vec);
+#else
+      vec_op::FP32Vec16 gate_vec(input + n, index_vec);
+      vec_op::FP32Vec16 up_vec(input + n + 1, index_vec);
+#endif
+      gate_vec = gate_vec.min(gate_up_max_vec);
+      up_vec = up_vec.clamp(up_min_vec, gate_up_max_vec);
+      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec * alpha_vec));
+      auto glu = gate_vec * sigmoid_vec;
+      auto gated_output_fp32 = (one_vec + up_vec) * glu;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n / 2);
+    }
+    input += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+void silu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                  const int32_t m_size, const int32_t n_size,
+                  const int32_t input_stride, const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  const int32_t dim = n_size / 2;
+  float* __restrict__ gate = input;
+  float* __restrict__ up = input + dim;
+  vec_op::FP32Vec16 one_vec(1.0);
+
+  DEFINE_FAST_EXP
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < dim; n += 16) {
+      vec_op::FP32Vec16 gate_vec(gate + n);
+      vec_op::FP32Vec16 up_vec(up + n);
+      auto sigmoid_vec = one_vec / (one_vec + fast_exp(-gate_vec));
+      auto silu = gate_vec * sigmoid_vec;
+      auto gated_output_fp32 = up_vec * silu;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n);
+    }
+    gate += input_stride;
+    up += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+void gelu_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                  const int32_t m_size, const int32_t n_size,
+                  const int32_t input_stride, const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  const int32_t dim = n_size / 2;
+  float* __restrict__ gate = input;
+  float* __restrict__ up = input + dim;
+  vec_op::FP32Vec16 one_vec(1.0);
+  vec_op::FP32Vec16 w1_vec(M_SQRT1_2);
+  vec_op::FP32Vec16 w2_vec(0.5);
+  alignas(64) float temp[16];
+
+  DEFINE_FAST_EXP
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < dim; n += 16) {
+      vec_op::FP32Vec16 gate_vec(gate + n);
+      vec_op::FP32Vec16 up_vec(up + n);
+      auto er_input_vec = gate_vec * w1_vec;
+
+      er_input_vec.save(temp);
+      for (int32_t i = 0; i < 16; ++i) {
+        temp[i] = std::erf(temp[i]);
+      }
+      vec_op::FP32Vec16 er_vec(temp);
+      auto gelu = gate_vec * w2_vec * (one_vec + er_vec);
+      auto gated_output_fp32 = up_vec * gelu;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n);
+    }
+    gate += input_stride;
+    up += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+void gelu_tanh_and_mul(float* __restrict__ input, scalar_t* __restrict__ output,
+                       const int32_t m_size, const int32_t n_size,
+                       const int32_t input_stride,
+                       const int32_t output_stride) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  const int32_t dim = n_size / 2;
+  float* __restrict__ gate = input;
+  float* __restrict__ up = input + dim;
+  vec_op::FP32Vec16 one_vec(1.0);
+  vec_op::FP32Vec16 w1_vec(0.7978845608028654);
+  vec_op::FP32Vec16 w2_vec(0.5);
+  vec_op::FP32Vec16 w3_vec(0.044715);
+
+  for (int32_t m = 0; m < m_size; ++m) {
+    for (int32_t n = 0; n < dim; n += 16) {
+      vec_op::FP32Vec16 gate_vec(gate + n);
+      vec_op::FP32Vec16 up_vec(up + n);
+      auto gate_pow3_vec = gate_vec * gate_vec * gate_vec;
+      auto inner_vec = w1_vec * (gate_vec + w3_vec * gate_pow3_vec);
+      // Note: can't use fast_exp form because diffusiongemma will generate
+      // wrong results
+      auto tanh_vec = inner_vec.tanh();
+      auto gelu_tanh = gate_vec * w2_vec * (one_vec + tanh_vec);
+      auto gated_output_fp32 = up_vec * gelu_tanh;
+      scalar_vec_t gated_output = scalar_vec_t(gated_output_fp32);
+      gated_output.save(output + n);
+    }
+    gate += input_stride;
+    up += input_stride;
+    output += output_stride;
+  }
+}
+
+template <typename scalar_t>
+FORCE_INLINE void apply_gated_act(const FusedMOEAct act,
+                                  float* __restrict__ input,
+                                  scalar_t* __restrict__ output,
+                                  const int32_t m, const int32_t n,
+                                  const int32_t input_stride,
+                                  const int32_t output_stride) {
+  switch (act) {
+    case FusedMOEAct::SwigluOAIAndMul:
+      swigluoai_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    case FusedMOEAct::SiluAndMul:
+      silu_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    case FusedMOEAct::GeluAndMul:
+      gelu_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    case FusedMOEAct::GeluTanhAndMul:
+      gelu_tanh_and_mul(input, output, m, n, input_stride, output_stride);
+      return;
+    default:
+      TORCH_CHECK(false, "Unsupported act type.");
+  }
+}
+}  // namespace cpu_fused_moe_utils
+
+#endif

csrc/cpu/cpu_fused_moe_int8.cpp

@@ -0,0 +1,647 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#include "cpu/cpu_arch_macros.h"
+
+#include <algorithm>
+#include <cstdint>
+#include <cstring>
+#include <optional>
+#include <string>
+
+#include "cpu/cpu_fused_moe_activations.hpp"
+#include "cpu/cpu_types.hpp"
+#include "cpu/micro_gemm/cpu_micro_gemm_impl.hpp"
+#include "cpu/utils.hpp"
+
+#if defined(ARM_I8MM_SUPPORT) && defined(ARM_BF16_SUPPORT)
+  #include "cpu/micro_gemm/cpu_micro_gemm_int8_neon.hpp"
+  #define NEON_DISPATCH(SCALAR_TYPE, ...)                                   \
+    case cpu_utils::ISA::NEON: {                                            \
+      using gemm_t =                                                        \
+          cpu_micro_gemm::MicroGemmINT8<cpu_utils::ISA::NEON, SCALAR_TYPE>; \
+      return __VA_ARGS__();                                                 \
+    }
+#else
+  #define NEON_DISPATCH(SCALAR_TYPE, ...) case cpu_utils::ISA::NEON:
+#endif
+
+#define CPU_INT8_ISA_DISPATCH_IMPL(ISA_TYPE, SCALAR_TYPE, ...) \
+  [&] {                                                        \
+    switch (ISA_TYPE) {                                        \
+      NEON_DISPATCH(SCALAR_TYPE, __VA_ARGS__)                  \
+      default: {                                               \
+        TORCH_CHECK(false, "Invalid CPU ISA type.");           \
+      }                                                        \
+    }                                                          \
+  }()
+
+namespace {
+using cpu_fused_moe_utils::apply_gated_act;
+using cpu_fused_moe_utils::FusedMOEAct;
+
+template <typename gemm_t>
+void prepack_moe_weight_int8_impl(const int8_t* __restrict__ weight_ptr,
+                                  int8_t* __restrict__ packed_weight_ptr,
+                                  const int32_t expert_num,
+                                  const int32_t output_size,
+                                  const int32_t input_size,
+                                  const int64_t expert_stride) {
+#pragma omp parallel for
+  for (int32_t e_idx = 0; e_idx < expert_num; ++e_idx) {
+    gemm_t::pack_weight(weight_ptr + expert_stride * e_idx,
+                        packed_weight_ptr + expert_stride * e_idx, output_size,
+                        input_size);
+  }
+}
+
+// INT8 MoE kernel, based on the original BF16 kernel in cpu_fused_moe.cpp
+template <typename scalar_t, typename gemm_t>
+void fused_moe_int8_impl(
+    scalar_t* __restrict__ output, const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ w13, const int8_t* __restrict__ w2,
+    const float* __restrict__ w13_scales, const float* __restrict__ w2_scales,
+    scalar_t* __restrict__ w13_bias, scalar_t* __restrict__ w2_bias,
+    const float* __restrict__ topk_weights, const int32_t* __restrict__ topk_id,
+    const FusedMOEAct act_type, const int32_t token_num,
+    const int32_t expert_num, const int32_t topk_num,
+    const int32_t input_size_13, const int32_t output_size_13,
+    const int32_t input_size_2, const int32_t output_size_2,
+    const bool skip_weighted) {
+  using scalar_vec_t = typename cpu_utils::VecTypeTrait<scalar_t>::vec_t;
+  constexpr int32_t gemm_n_tile_size = gemm_t::NSize;
+  constexpr int32_t gemm_m_tile_size = gemm_t::MaxMSize;
+  constexpr int32_t min_w13_n_tile_size = 2 * gemm_n_tile_size;
+
+  TORCH_CHECK_EQ(input_size_13 % gemm_t::K, 0);
+  TORCH_CHECK_EQ(input_size_2 % gemm_t::K, 0);
+  TORCH_CHECK_EQ(output_size_13 % min_w13_n_tile_size, 0);
+  TORCH_CHECK_EQ(output_size_2 % gemm_n_tile_size, 0);
+  TORCH_CHECK_EQ(output_size_13 / 2, input_size_2);
+
+  const int32_t thread_num = cpu_utils::get_max_threads();
+  const int32_t w13_input_buffer_size = cpu_utils::round_up<64>(
+      gemm_m_tile_size * input_size_13 * sizeof(int8_t));
+  const int32_t w2_input_buffer_size =
+      cpu_utils::round_up<64>(gemm_m_tile_size * input_size_2 * sizeof(int8_t));
+
+  const int32_t w13_n_tile_size = [&]() {
+    const int64_t cache_size = cpu_utils::get_available_l2_size();
+    const int32_t n_size_cache_limit =
+        (cache_size - w13_input_buffer_size) /
+        (gemm_m_tile_size * sizeof(float) + input_size_13 * sizeof(int8_t));
+    const int32_t n_size_thread_limit =
+        output_size_13 / std::max(1, thread_num / topk_num);
+    const int32_t n_size = cpu_utils::round_down<min_w13_n_tile_size>(
+        std::min(n_size_cache_limit, n_size_thread_limit));
+    return std::max(n_size, min_w13_n_tile_size);
+  }();
+
+  const int32_t w2_n_tile_size = [&]() {
+    const int64_t cache_size = cpu_utils::get_available_l2_size();
+    const int32_t n_size_cache_limit =
+        (cache_size - w2_input_buffer_size) / (input_size_2 * sizeof(int8_t));
+    const int32_t n_size_thread_limit =
+        output_size_2 / std::max(1, thread_num / topk_num);
+    const int32_t n_size = cpu_utils::round_down<gemm_n_tile_size>(
+        std::min(n_size_cache_limit, n_size_thread_limit));
+    return std::max(n_size, gemm_n_tile_size);
+  }();
+
+  int32_t common_buffer_offset = 0;
+  const int32_t token_num_per_group_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(expert_num * sizeof(int32_t));
+  const int32_t cu_token_num_per_group_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>((expert_num + 1) * sizeof(int32_t));
+  const int32_t expanded_token_num = token_num * topk_num;
+  const int32_t expand_token_id_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(expanded_token_num * sizeof(int32_t));
+  const int32_t expand_token_id_index_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(expanded_token_num * sizeof(int32_t));
+  const int32_t input_quant_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(token_num * input_size_13 * sizeof(int8_t));
+  const int32_t input_scale_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(token_num * sizeof(float));
+  const int32_t w13_gemm_output_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(
+      expanded_token_num * input_size_2 * sizeof(scalar_t));
+  const int32_t w13_output_scale_buffer_offset = common_buffer_offset;
+  common_buffer_offset +=
+      cpu_utils::round_up<64>(expanded_token_num * sizeof(float));
+  const int32_t w2_gemm_output_buffer_offset = common_buffer_offset;
+  common_buffer_offset += cpu_utils::round_up<64>(
+      expanded_token_num * output_size_2 * sizeof(float));
+
+  int32_t gemm_thread_buffer_offset = 0;
+  const int32_t gemm_input_buffer_offset = gemm_thread_buffer_offset;
+  gemm_thread_buffer_offset +=
+      std::max(w13_input_buffer_size, w2_input_buffer_size);
+  const int32_t gemm_output_buffer_offset = gemm_thread_buffer_offset;
+  gemm_thread_buffer_offset += cpu_utils::round_up<64>(
+      gemm_m_tile_size * std::max(w13_n_tile_size, w2_n_tile_size) *
+      sizeof(int32_t));
+
+  const int32_t ws_output_buffer_offset = 0;
+  const int32_t ws_thread_buffer_size =
+      cpu_utils::round_up<64>(output_size_2 * sizeof(float));
+  const int32_t thread_buffer_size =
+      std::max(gemm_thread_buffer_offset, ws_thread_buffer_size);
+  const int32_t buffer_size =
+      common_buffer_offset + thread_buffer_size * thread_num;
+  cpu_utils::ScratchPadManager::get_scratchpad_manager()->realloc(buffer_size);
+  uint8_t* common_buffer_start =
+      cpu_utils::ScratchPadManager::get_scratchpad_manager()
+          ->get_data<uint8_t>();
+  uint8_t* thread_buffer_start = common_buffer_start + common_buffer_offset;
+
+  int32_t* __restrict__ token_num_per_group_buffer = reinterpret_cast<int32_t*>(
+      common_buffer_start + token_num_per_group_buffer_offset);
+  int32_t* __restrict__ cu_token_num_per_group_buffer =
+      reinterpret_cast<int32_t*>(common_buffer_start +
+                                 cu_token_num_per_group_buffer_offset);
+  int32_t* __restrict__ expand_token_id_buffer = reinterpret_cast<int32_t*>(
+      common_buffer_start + expand_token_id_buffer_offset);
+  int32_t* __restrict__ expand_token_id_index_buffer =
+      reinterpret_cast<int32_t*>(common_buffer_start +
+                                 expand_token_id_index_buffer_offset);
+  int8_t* __restrict__ input_quant_buffer = reinterpret_cast<int8_t*>(
+      common_buffer_start + input_quant_buffer_offset);
+  float* __restrict__ input_scale_buffer =
+      reinterpret_cast<float*>(common_buffer_start + input_scale_buffer_offset);
+
+  std::memset(token_num_per_group_buffer, 0, expert_num * sizeof(int32_t));
+  for (int32_t i = 0; i < expanded_token_num; ++i) {
+    ++token_num_per_group_buffer[topk_id[i]];
+  }
+
+  int32_t token_num_sum = 0;
+  cu_token_num_per_group_buffer[0] = 0;
+  int32_t* token_index_buffer = cu_token_num_per_group_buffer + 1;
+  for (int32_t i = 0; i < expert_num; ++i) {
+    token_index_buffer[i] = token_num_sum;
+    token_num_sum += token_num_per_group_buffer[i];
+  }
+
+  for (int32_t i = 0; i < token_num; ++i) {
+    const int32_t* curr_topk_id = topk_id + i * topk_num;
+    int32_t* curr_index_buffer = expand_token_id_index_buffer + i * topk_num;
+    for (int32_t j = 0; j < topk_num; ++j) {
+      const int32_t curr_expert_id = curr_topk_id[j];
+      const int32_t curr_index = token_index_buffer[curr_expert_id]++;
+      expand_token_id_buffer[curr_index] = i;
+      curr_index_buffer[j] = curr_index;
+    }
+  }
+
+// quantize inputs
+#pragma omp parallel for
+  for (int32_t token_idx = 0; token_idx < token_num; ++token_idx) {
+    gemm_t::quantize_row(input + token_idx * input_size_13,
+                         input_quant_buffer + token_idx * input_size_13,
+                         input_scale_buffer[token_idx], input_size_13);
+  }
+
+  {
+    alignas(64) cpu_utils::Counter counter;
+    cpu_utils::Counter* counter_ptr = &counter;
+
+// w13 GEMM + act
+#pragma omp parallel for schedule(static, 1)
+    for (int32_t thread_id = 0; thread_id < thread_num; ++thread_id) {
+      const int32_t task_num_per_expert =
+          (output_size_13 + w13_n_tile_size - 1) / w13_n_tile_size;
+      const int32_t task_num = task_num_per_expert * expert_num;
+      uint8_t* __restrict__ thread_buffer =
+          thread_buffer_start + thread_id * thread_buffer_size;
+      int8_t* __restrict__ gemm_input_buffer =
+          reinterpret_cast<int8_t*>(thread_buffer + gemm_input_buffer_offset);
+      float* __restrict__ gemm_output_buffer =
+          reinterpret_cast<float*>(thread_buffer + gemm_output_buffer_offset);
+      auto* __restrict__ w13_gemm_output_buffer = reinterpret_cast<scalar_t*>(
+          common_buffer_start + w13_gemm_output_buffer_offset);
+      gemm_t gemm;
+
+      const int32_t w13_n_group_stride =
+          gemm_t::WeightOCGroupSize * input_size_13;
+      const int32_t w13_n_tile_stride = gemm_n_tile_size * input_size_13;
+
+      for (;;) {
+        const int32_t task_id = counter_ptr->acquire_counter();
+        if (task_id >= task_num) {
+          break;
+        }
+        const int32_t curr_expert_id = task_id / task_num_per_expert;
+        const int32_t curr_output_group_id = task_id % task_num_per_expert;
+        const int32_t curr_token_num =
+            token_num_per_group_buffer[curr_expert_id];
+        if (curr_token_num == 0) {
+          continue;
+        }
+
+        const int32_t actual_n_tile_size =
+            std::min(w13_n_tile_size,
+                     output_size_13 - curr_output_group_id * w13_n_tile_size);
+        const int32_t* __restrict__ curr_expand_token_id_buffer =
+            expand_token_id_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id];
+        scalar_t* __restrict__ curr_w13_gemm_output_buffer =
+            w13_gemm_output_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id] * input_size_2 +
+            curr_output_group_id * w13_n_tile_size / 2;
+
+        const int8_t* w13_weight_ptr_0 = nullptr;
+        const int8_t* w13_weight_ptr_1 = nullptr;
+        const float* w13_scale_ptr_0 = nullptr;
+        const float* w13_scale_ptr_1 = nullptr;
+        scalar_t* w13_bias_ptr_0 = nullptr;
+        scalar_t* w13_bias_ptr_1 = nullptr;
+        if (act_type == FusedMOEAct::SwigluOAIAndMul) {
+          const int32_t output_offset = curr_output_group_id * w13_n_tile_size;
+          w13_weight_ptr_0 = w13 +
+                             curr_expert_id * input_size_13 * output_size_13 +
+                             output_offset * input_size_13;
+          w13_weight_ptr_1 =
+              w13_weight_ptr_0 + actual_n_tile_size / 2 * input_size_13;
+          w13_scale_ptr_0 =
+              w13_scales + curr_expert_id * output_size_13 + output_offset;
+          w13_scale_ptr_1 = w13_scale_ptr_0 + actual_n_tile_size / 2;
+          if (w13_bias != nullptr) {
+            w13_bias_ptr_0 =
+                w13_bias + curr_expert_id * output_size_13 + output_offset;
+            w13_bias_ptr_1 = w13_bias_ptr_0 + actual_n_tile_size / 2;
+          }
+        } else {
+          const int32_t output_offset =
+              curr_output_group_id * (w13_n_tile_size / 2);
+          w13_weight_ptr_0 = w13 +
+                             curr_expert_id * input_size_13 * output_size_13 +
+                             output_offset * input_size_13;
+          w13_weight_ptr_1 =
+              w13_weight_ptr_0 + output_size_13 / 2 * input_size_13;
+          w13_scale_ptr_0 =
+              w13_scales + curr_expert_id * output_size_13 + output_offset;
+          w13_scale_ptr_1 = w13_scale_ptr_0 + output_size_13 / 2;
+          if (w13_bias != nullptr) {
+            w13_bias_ptr_0 =
+                w13_bias + curr_expert_id * output_size_13 + output_offset;
+            w13_bias_ptr_1 = w13_bias_ptr_0 + output_size_13 / 2;
+          }
+        }
+
+        for (int32_t token_idx = 0; token_idx < curr_token_num;
+             token_idx += gemm_m_tile_size) {
+          const int32_t actual_token_num =
+              std::min(gemm_m_tile_size, curr_token_num - token_idx);
+          const int8_t* input_rows[gemm_m_tile_size];
+          alignas(64) float input_scales[gemm_m_tile_size];
+          // gather and pack
+          for (int32_t i = 0; i < actual_token_num; ++i) {
+            const int32_t curr_token_id = curr_expand_token_id_buffer[i];
+            input_rows[i] = input_quant_buffer + curr_token_id * input_size_13;
+            input_scales[i] = input_scale_buffer[curr_token_id];
+          }
+          gemm_t::pack_input_from_rows(input_rows, gemm_input_buffer,
+                                       actual_token_num, input_size_13);
+          curr_expand_token_id_buffer += actual_token_num;
+
+          const int8_t* w13_weight_ptr_0_iter = w13_weight_ptr_0;
+          const int8_t* w13_weight_ptr_1_iter = w13_weight_ptr_1;
+          const float* w13_scale_ptr_0_iter = w13_scale_ptr_0;
+          const float* w13_scale_ptr_1_iter = w13_scale_ptr_1;
+          scalar_t* w13_bias_ptr_0_iter = w13_bias_ptr_0;
+          scalar_t* w13_bias_ptr_1_iter = w13_bias_ptr_1;
+          float* w13_output_buffer_0_iter = gemm_output_buffer;
+          float* w13_output_buffer_1_iter =
+              gemm_output_buffer + actual_n_tile_size / 2;
+
+          for (int32_t i = 0; i < actual_n_tile_size;
+               i += min_w13_n_tile_size) {
+            auto* output_0_int32 =
+                reinterpret_cast<int32_t*>(w13_output_buffer_0_iter);
+            gemm.gemm(gemm_input_buffer, w13_weight_ptr_0_iter, output_0_int32,
+                      actual_token_num, input_size_13, w13_n_group_stride,
+                      actual_n_tile_size);
+            gemm_t::dequantize_tile(output_0_int32, w13_output_buffer_0_iter,
+                                    input_scales, w13_scale_ptr_0_iter,
+                                    actual_token_num, gemm_n_tile_size,
+                                    actual_n_tile_size);
+            if (w13_bias != nullptr) {
+              cpu_micro_gemm::add_bias_epilogue<gemm_n_tile_size>(
+                  w13_output_buffer_0_iter, w13_output_buffer_0_iter,
+                  w13_bias_ptr_0_iter, actual_token_num, actual_n_tile_size,
+                  actual_n_tile_size);
+              w13_bias_ptr_0_iter += gemm_n_tile_size;
+            }
+
+            auto* output_1_int32 =
+                reinterpret_cast<int32_t*>(w13_output_buffer_1_iter);
+            gemm.gemm(gemm_input_buffer, w13_weight_ptr_1_iter, output_1_int32,
+                      actual_token_num, input_size_13, w13_n_group_stride,
+                      actual_n_tile_size);
+            gemm_t::dequantize_tile(output_1_int32, w13_output_buffer_1_iter,
+                                    input_scales, w13_scale_ptr_1_iter,
+                                    actual_token_num, gemm_n_tile_size,
+                                    actual_n_tile_size);
+            if (w13_bias != nullptr) {
+              cpu_micro_gemm::add_bias_epilogue<gemm_n_tile_size>(
+                  w13_output_buffer_1_iter, w13_output_buffer_1_iter,
+                  w13_bias_ptr_1_iter, actual_token_num, actual_n_tile_size,
+                  actual_n_tile_size);
+              w13_bias_ptr_1_iter += gemm_n_tile_size;
+            }
+
+            w13_weight_ptr_0_iter += w13_n_tile_stride;
+            w13_weight_ptr_1_iter += w13_n_tile_stride;
+            w13_scale_ptr_0_iter += gemm_n_tile_size;
+            w13_scale_ptr_1_iter += gemm_n_tile_size;
+            w13_output_buffer_0_iter += gemm_n_tile_size;
+            w13_output_buffer_1_iter += gemm_n_tile_size;
+          }
+
+          apply_gated_act(act_type, gemm_output_buffer,
+                          curr_w13_gemm_output_buffer, actual_token_num,
+                          actual_n_tile_size, actual_n_tile_size, input_size_2);
+          curr_w13_gemm_output_buffer += gemm_m_tile_size * input_size_2;
+        }
+      }
+    }
+  }
+
+  auto* __restrict__ w13_gemm_output_buffer = reinterpret_cast<scalar_t*>(
+      common_buffer_start + w13_gemm_output_buffer_offset);
+  float* __restrict__ w13_output_scale_buffer = reinterpret_cast<float*>(
+      common_buffer_start + w13_output_scale_buffer_offset);
+
+// quantize w2 inputs - in place
+#pragma omp parallel for
+  for (int32_t token_idx = 0; token_idx < expanded_token_num; ++token_idx) {
+    scalar_t* input_row = w13_gemm_output_buffer + token_idx * input_size_2;
+    int8_t* output_row = reinterpret_cast<int8_t*>(input_row);
+    gemm_t::quantize_row(input_row, output_row,
+                         w13_output_scale_buffer[token_idx], input_size_2);
+  }
+
+  {
+    alignas(64) cpu_utils::Counter counter;
+    cpu_utils::Counter* counter_ptr = &counter;
+
+// w2 gemm
+#pragma omp parallel for schedule(static, 1)
+    for (int32_t thread_id = 0; thread_id < thread_num; ++thread_id) {
+      const int32_t task_num_per_expert =
+          (output_size_2 + w2_n_tile_size - 1) / w2_n_tile_size;
+      const int32_t task_num = task_num_per_expert * expert_num;
+      uint8_t* __restrict__ thread_buffer =
+          thread_buffer_start + thread_id * thread_buffer_size;
+      int8_t* __restrict__ gemm_input_buffer =
+          reinterpret_cast<int8_t*>(thread_buffer + gemm_input_buffer_offset);
+      float* __restrict__ gemm_output_buffer =
+          reinterpret_cast<float*>(thread_buffer + gemm_output_buffer_offset);
+      float* __restrict__ w2_gemm_output_buffer = reinterpret_cast<float*>(
+          common_buffer_start + w2_gemm_output_buffer_offset);
+      gemm_t gemm;
+
+      const int32_t w2_n_group_stride =
+          gemm_t::WeightOCGroupSize * input_size_2;
+      const int32_t w2_n_tile_stride = gemm_n_tile_size * input_size_2;
+
+      for (;;) {
+        const int32_t task_id = counter_ptr->acquire_counter();
+        if (task_id >= task_num) {
+          break;
+        }
+        const int32_t curr_expert_id = task_id / task_num_per_expert;
+        const int32_t curr_output_group_id = task_id % task_num_per_expert;
+        const int32_t curr_token_num =
+            token_num_per_group_buffer[curr_expert_id];
+        if (curr_token_num == 0) {
+          continue;
+        }
+
+        const int32_t actual_n_tile_size =
+            std::min(w2_n_tile_size,
+                     output_size_2 - curr_output_group_id * w2_n_tile_size);
+        scalar_t* __restrict__ curr_w13_gemm_output_buffer =
+            w13_gemm_output_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id] * input_size_2;
+        float* __restrict__ curr_w13_output_scale_buffer =
+            w13_output_scale_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id];
+        float* __restrict__ curr_w2_gemm_output_buffer =
+            w2_gemm_output_buffer +
+            cu_token_num_per_group_buffer[curr_expert_id] * output_size_2 +
+            curr_output_group_id * w2_n_tile_size;
+        const int8_t* __restrict__ w2_weight_ptr =
+            w2 + curr_expert_id * output_size_2 * input_size_2 +
+            curr_output_group_id * w2_n_tile_size * input_size_2;
+        const float* __restrict__ w2_scale_ptr =
+            w2_scales + curr_expert_id * output_size_2 +
+            curr_output_group_id * w2_n_tile_size;
+        scalar_t* w2_bias_ptr = nullptr;
+        if (w2_bias != nullptr) {
+          w2_bias_ptr = w2_bias + curr_expert_id * output_size_2 +
+                        curr_output_group_id * w2_n_tile_size;
+        }
+
+        for (int32_t token_idx = 0; token_idx < curr_token_num;
+             token_idx += gemm_m_tile_size) {
+          const int32_t actual_token_num =
+              std::min(gemm_m_tile_size, curr_token_num - token_idx);
+          const int8_t* input_rows[gemm_m_tile_size];
+          alignas(64) float input_scales[gemm_m_tile_size];
+          for (int32_t i = 0; i < actual_token_num; ++i) {
+            input_rows[i] = reinterpret_cast<const int8_t*>(
+                curr_w13_gemm_output_buffer + i * input_size_2);
+            input_scales[i] = curr_w13_output_scale_buffer[i];
+          }
+          gemm_t::pack_input_from_rows(input_rows, gemm_input_buffer,
+                                       actual_token_num, input_size_2);
+
+          const int8_t* w2_weight_ptr_iter = w2_weight_ptr;
+          const float* w2_scale_ptr_iter = w2_scale_ptr;
+          scalar_t* w2_bias_ptr_iter = w2_bias_ptr;
+          float* curr_w2_gemm_output_buffer_iter = curr_w2_gemm_output_buffer;
+          for (int32_t i = 0; i < actual_n_tile_size; i += gemm_n_tile_size) {
+            auto* output_int32 = reinterpret_cast<int32_t*>(gemm_output_buffer);
+            gemm.gemm(gemm_input_buffer, w2_weight_ptr_iter, output_int32,
+                      actual_token_num, input_size_2, w2_n_group_stride,
+                      gemm_n_tile_size);
+            gemm_t::dequantize_tile(output_int32, gemm_output_buffer,
+                                    input_scales, w2_scale_ptr_iter,
+                                    actual_token_num, gemm_n_tile_size,
+                                    gemm_n_tile_size);
+            if (w2_bias != nullptr) {
+              cpu_micro_gemm::add_bias_epilogue<gemm_n_tile_size>(
+                  gemm_output_buffer, gemm_output_buffer, w2_bias_ptr_iter,
+                  actual_token_num, gemm_n_tile_size, gemm_n_tile_size);
+              w2_bias_ptr_iter += gemm_n_tile_size;
+            }
+            for (int32_t m_idx = 0; m_idx < actual_token_num; ++m_idx) {
+              std::memcpy(
+                  curr_w2_gemm_output_buffer_iter + m_idx * output_size_2,
+                  gemm_output_buffer + m_idx * gemm_n_tile_size,
+                  gemm_n_tile_size * sizeof(float));
+            }
+
+            w2_weight_ptr_iter += w2_n_tile_stride;
+            w2_scale_ptr_iter += gemm_n_tile_size;
+            curr_w2_gemm_output_buffer_iter += gemm_n_tile_size;
+          }
+
+          curr_w13_gemm_output_buffer += gemm_m_tile_size * input_size_2;
+          curr_w13_output_scale_buffer += gemm_m_tile_size;
+          curr_w2_gemm_output_buffer += gemm_m_tile_size * output_size_2;
+        }
+      }
+    }
+  }
+
+  {
+    alignas(64) cpu_utils::Counter counter;
+    cpu_utils::Counter* counter_ptr = &counter;
+
+#pragma omp parallel for schedule(static, 1)
+    for (int32_t thread_id = 0; thread_id < thread_num; ++thread_id) {
+      uint8_t* __restrict__ thread_buffer =
+          thread_buffer_start + thread_id * thread_buffer_size;
+      float* __restrict__ ws_output_buffer =
+          reinterpret_cast<float*>(thread_buffer + ws_output_buffer_offset);
+      float* __restrict__ w2_gemm_output_buffer = reinterpret_cast<float*>(
+          common_buffer_start + w2_gemm_output_buffer_offset);
+
+      for (;;) {
+        const int32_t token_id = counter_ptr->acquire_counter();
+        if (token_id >= token_num) {
+          break;
+        }
+        int32_t* __restrict__ curr_expand_token_id_index_buffer =
+            expand_token_id_index_buffer + token_id * topk_num;
+        const float* __restrict__ curr_weight =
+            topk_weights + token_id * topk_num;
+        const float first_weight = skip_weighted ? 1.0f : curr_weight[0];
+        scalar_t* __restrict__ curr_output_buffer =
+            output + token_id * output_size_2;
+
+        if (topk_num > 1) {
+          int32_t w2_output_idx = curr_expand_token_id_index_buffer[0];
+          float* w2_output_iter =
+              w2_gemm_output_buffer + w2_output_idx * output_size_2;
+          float* ws_output_buffer_iter = ws_output_buffer;
+          vec_op::FP32Vec16 weight_vec(first_weight);
+          for (int32_t i = 0; i < output_size_2; i += 16) {
+            vec_op::FP32Vec16 vec(w2_output_iter);
+            (vec * weight_vec).save(ws_output_buffer_iter);
+            w2_output_iter += 16;
+            ws_output_buffer_iter += 16;
+          }
+
+          for (int32_t idx = 1; idx < topk_num - 1; ++idx) {
+            w2_output_idx = curr_expand_token_id_index_buffer[idx];
+            w2_output_iter =
+                w2_gemm_output_buffer + w2_output_idx * output_size_2;
+            ws_output_buffer_iter = ws_output_buffer;
+            weight_vec = vec_op::FP32Vec16(curr_weight[idx]);
+            for (int32_t i = 0; i < output_size_2; i += 16) {
+              vec_op::FP32Vec16 vec(w2_output_iter);
+              vec_op::FP32Vec16 sum(ws_output_buffer_iter);
+              (sum + vec * weight_vec).save(ws_output_buffer_iter);
+              w2_output_iter += 16;
+              ws_output_buffer_iter += 16;
+            }
+          }
+
+          const int32_t last_idx = topk_num - 1;
+          w2_output_idx = curr_expand_token_id_index_buffer[last_idx];
+          w2_output_iter =
+              w2_gemm_output_buffer + w2_output_idx * output_size_2;
+          ws_output_buffer_iter = ws_output_buffer;
+          scalar_t* curr_output_buffer_iter = curr_output_buffer;
+          weight_vec = vec_op::FP32Vec16(curr_weight[last_idx]);
+          for (int32_t i = 0; i < output_size_2; i += 16) {
+            vec_op::FP32Vec16 vec(w2_output_iter);
+            vec_op::FP32Vec16 sum(ws_output_buffer_iter);
+            scalar_vec_t(sum + vec * weight_vec).save(curr_output_buffer_iter);
+            w2_output_iter += 16;
+            ws_output_buffer_iter += 16;
+            curr_output_buffer_iter += 16;
+          }
+        } else {
+          const int32_t w2_output_idx = curr_expand_token_id_index_buffer[0];
+          float* w2_output_iter =
+              w2_gemm_output_buffer + w2_output_idx * output_size_2;
+          scalar_t* curr_output_buffer_iter = curr_output_buffer;
+          vec_op::FP32Vec16 weight_vec(first_weight);
+          for (int32_t i = 0; i < output_size_2; i += 16) {
+            vec_op::FP32Vec16 vec(w2_output_iter);
+            scalar_vec_t(vec * weight_vec).save(curr_output_buffer_iter);
+            w2_output_iter += 16;
+            curr_output_buffer_iter += 16;
+          }
+        }
+      }
+    }
+  }
+}
+}  // namespace
+
+void prepack_moe_weight_int8(
+    const torch::Tensor& weight,  // [expert_num, output_size, input_size]
+    torch::Tensor& packed_weight, const std::string& isa) {
+  TORCH_CHECK(weight.is_contiguous());
+  const int32_t expert_num = weight.size(0);
+  const int32_t output_size = weight.size(1);
+  const int32_t input_size = weight.size(2);
+  const int64_t expert_stride = weight.stride(0);
+  const cpu_utils::ISA isa_type = cpu_utils::get_isa(isa);
+  TORCH_CHECK_EQ(output_size % 32, 0);
+
+  CPU_INT8_ISA_DISPATCH_IMPL(isa_type, c10::BFloat16, [&]() {
+    TORCH_CHECK_EQ(input_size % gemm_t::K, 0);
+    prepack_moe_weight_int8_impl<gemm_t>(
+        weight.data_ptr<int8_t>(), packed_weight.data_ptr<int8_t>(), expert_num,
+        output_size, input_size, expert_stride);
+  });
+}
+
+void cpu_fused_moe_int8(torch::Tensor& output, const torch::Tensor& input,
+                        const torch::Tensor& w13, const torch::Tensor& w2,
+                        const torch::Tensor& w13_scale,
+                        const torch::Tensor& w2_scale,
+                        const std::optional<torch::Tensor>& w13_bias,
+                        const std::optional<torch::Tensor>& w2_bias,
+                        const torch::Tensor& topk_weights,
+                        const torch::Tensor& topk_id, const bool skip_weighted,
+                        const std::string& act, const std::string& isa) {
+  const int32_t token_num = input.size(0);
+  const int32_t input_size_13 = input.size(1);
+  const int64_t input_stride = input.stride(0);
+  TORCH_CHECK_EQ(input_stride, input_size_13);
+  const int32_t expert_num = w13.size(0);
+  const int32_t output_size_13 = w13.size(1);
+  const int32_t input_size_2 = w2.size(2);
+  const int32_t output_size_2 = w2.size(1);
+  const int32_t topk_num = topk_id.size(1);
+  const FusedMOEAct act_type = cpu_fused_moe_utils::get_act_type(act);
+  const cpu_utils::ISA isa_type = cpu_utils::get_isa(isa);
+  TORCH_CHECK(!skip_weighted || topk_num == 1,
+              "skip_weighted is only supported for topk=1 on CPU");
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "cpu_fused_moe_int8", [&]() {
+        CPU_INT8_ISA_DISPATCH_IMPL(isa_type, scalar_t, [&]() {
+          fused_moe_int8_impl<scalar_t, gemm_t>(
+              output.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),
+              w13.data_ptr<int8_t>(), w2.data_ptr<int8_t>(),
+              w13_scale.data_ptr<float>(), w2_scale.data_ptr<float>(),
+              w13_bias.has_value() ? w13_bias->data_ptr<scalar_t>() : nullptr,
+              w2_bias.has_value() ? w2_bias->data_ptr<scalar_t>() : nullptr,
+              topk_weights.data_ptr<float>(), topk_id.data_ptr<int32_t>(),
+              act_type, token_num, expert_num, topk_num, input_size_13,
+              output_size_13, input_size_2, output_size_2, skip_weighted);
+        });
+      });
+}

csrc/cpu/cpu_types_vsx.hpp

@@ -287,7 +287,7 @@ struct FP32Vec4 : public Vec<FP32Vec4> {
 
   explicit FP32Vec4(__vector float data) : reg(data) {}
 
-  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {}
+  FP32Vec4(const FP32Vec4& data) : reg(data.reg) {}
 };
 
 struct FP32Vec8 : public Vec<FP32Vec8> {
@@ -316,7 +316,7 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
 
   explicit FP32Vec8(f32x4x2_t data) : reg(data) {}
 
-  explicit FP32Vec8(const FP32Vec8& data) {
+  FP32Vec8(const FP32Vec8& data) {
     reg.val[0] = data.reg.val[0];
     reg.val[1] = data.reg.val[1];
   }
@@ -336,13 +336,14 @@ struct FP32Vec8 : public Vec<FP32Vec8> {
     reg.val[1] = fp16_to_fp32_bits(raw_lo);
   }
   float reduce_sum() const {
-    AliasReg ar;
-    ar.reg = reg;
-    float result = 0;
-    unroll_loop<int, VEC_ELEM_NUM>(
-        [&result, &ar](int i) { result += ar.values[i]; });
-
-    return result;
+    // VSX horizontal reduction: 3 vector ops instead of 8 scalar adds.
+    // Step 1: pairwise sum of the two 4-wide halves
+    __vector float s = vec_add(reg.val[0], reg.val[1]);
+    // Step 2: rotate by 8 bytes (2 floats) and add
+    s = vec_add(s, vec_sld(s, s, 8));
+    // Step 3: rotate by 4 bytes (1 float) and add  => all lanes hold total
+    s = vec_add(s, vec_sld(s, s, 4));
+    return vec_extract(s, 0);
   }
   FP32Vec8 exp() const {
     f32x4x2_t out;
@@ -592,7 +593,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   explicit FP32Vec16(bool, const float* ptr) : FP32Vec16(ptr) {}
   explicit FP32Vec16(f32x4x4_t data) : reg(data) {}
 
-  explicit FP32Vec16(const FP32Vec16& data) {
+  FP32Vec16(const FP32Vec16& data) {
     reg.val[0] = data.reg.val[0];
     reg.val[1] = data.reg.val[1];
     reg.val[2] = data.reg.val[2];
@@ -746,6 +747,15 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
                                 vec_abs(reg.val[2]), vec_abs(reg.val[3])}));
   }
 
+  FP32Vec16 exp() const {
+    FP32Vec8 lo(f32x4x2_t{reg.val[0], reg.val[1]});
+    FP32Vec8 hi(f32x4x2_t{reg.val[2], reg.val[3]});
+    auto lo_e = lo.exp();
+    auto hi_e = hi.exp();
+    return FP32Vec16(f32x4x4_t{lo_e.reg.val[0], lo_e.reg.val[1],
+                               hi_e.reg.val[0], hi_e.reg.val[1]});
+  }
+
   float reduce_max() {
     __vector float max01 = vec_max(reg.val[0], reg.val[1]);
     __vector float max23 = vec_max(reg.val[2], reg.val[3]);

csrc/cpu/mamba_cpu.cpp

@@ -0,0 +1,285 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+//
+// CPU at::Tensor wrappers for Mamba decode-step kernels defined in
+// mamba_kernels.hpp.
+
+#include "cpu/mamba_kernels.hpp"
+
+#include <ATen/ATen.h>
+#include <torch/library.h>
+#include <c10/util/Optional.h>
+
+#include "cpu_types.hpp"
+
+// ---------------------------------------------------------------------------
+// causal_conv1d_update
+// ---------------------------------------------------------------------------
+at::Tensor causal_conv1d_update_cpu_impl(
+    at::Tensor& x, at::Tensor& conv_state, const at::Tensor& weight,
+    const c10::optional<at::Tensor>& bias,
+    const c10::optional<std::string>& activation,
+    const c10::optional<at::Tensor>& conv_state_indices,
+    const c10::optional<at::Tensor>& query_start_loc, int64_t pad_slot_id) {
+  bool do_silu = false;
+  if (activation.has_value()) {
+    const std::string& act = activation.value();
+    do_silu = (act == "silu" || act == "swish");
+  }
+
+  at::ScalarType dtype = x.scalar_type();
+
+  // Input x: contiguous in native dtype.
+  at::Tensor x_c = x.is_contiguous() ? x : x.contiguous();
+
+  // conv_state: NEVER copy the full paged tensor just for layout reasons.
+  // If the dtype matches we work directly on conv_state (contiguous or not)
+  // by extracting strides and passing them to the kernel.
+  // Only a dtype-conversion copy is made when types differ (rare for BF16).
+  bool state_type_ok = (conv_state.scalar_type() == dtype);
+  at::Tensor state_c = state_type_ok ? conv_state : conv_state.to(dtype);
+  // state_c and conv_state may be non-contiguous — that is intentional.
+
+  // Weight: coerce to same dtype if needed (should match in practice)
+  at::Tensor w_c =
+      (weight.scalar_type() != dtype)
+          ? weight.to(dtype).contiguous()
+          : (weight.is_contiguous() ? weight : weight.contiguous());
+
+  // Bias stays float32 (small scalar, used only for fp32 accumulation)
+  at::Tensor bias_f32;
+  if (bias.has_value() && bias.value().defined())
+    bias_f32 = bias.value().to(at::kFloat).contiguous();
+
+  int64_t batch = x_c.size(0);
+  int64_t dim = x_c.size(1);
+  int64_t seqlen = (x_c.dim() == 3) ? x_c.size(2) : 1;
+  int64_t width = w_c.size(1);
+  int64_t state_len = state_c.size(2);
+
+  // Extract strides — works for contiguous AND non-contiguous (transposed)
+  // state. stride(0): between cache slots (e.g. num_slots × dim × width-1 in
+  // contiguous) stride(1): between conv channels (dim stride) stride(2):
+  // between state elements (=1 when contiguous, =dim when transposed)
+  int64_t stride_s_slot = state_c.stride(0);
+  int64_t stride_s_dim = state_c.stride(1);
+  int64_t stride_s_state = state_c.stride(2);
+
+  at::Tensor out = x_c.clone();  // native dtype, no float32 alloc
+
+  const int32_t* cache_idx_ptr = nullptr;
+  at::Tensor cache_idx_int;
+  if (conv_state_indices.has_value()) {
+    cache_idx_int = conv_state_indices.value().to(at::kInt).contiguous();
+    cache_idx_ptr = cache_idx_int.data_ptr<int32_t>();
+  }
+
+  VLLM_DISPATCH_FLOATING_TYPES(dtype, "causal_conv1d_update", [&] {
+    mamba_cpu::causal_conv1d_update_kernel<scalar_t>(
+        x_c.data_ptr<scalar_t>(), state_c.data_ptr<scalar_t>(), stride_s_slot,
+        stride_s_dim, stride_s_state, w_c.data_ptr<scalar_t>(),
+        bias_f32.defined() ? bias_f32.data_ptr<float>() : nullptr,
+        out.data_ptr<scalar_t>(), cache_idx_ptr,
+        static_cast<int32_t>(pad_slot_id), batch, dim, seqlen, width, state_len,
+        do_silu);
+  });
+
+  // Write back only when a type-conversion copy was made.
+  // Layout-only non-contiguity is handled via strides above — no copy needed.
+  if (!state_type_ok) conv_state.copy_(state_c);
+
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// selective_state_update
+// ---------------------------------------------------------------------------
+void selective_state_update_cpu_impl(
+    at::Tensor& state,    // (nstates, nheads, dim, dstate)
+    const at::Tensor& x,  // (N, nheads, dim)
+    const at::Tensor& dt, const at::Tensor& A, const at::Tensor& B,
+    const at::Tensor& C, const c10::optional<at::Tensor>& D,
+    const c10::optional<at::Tensor>& z,
+    const c10::optional<at::Tensor>& dt_bias, bool dt_softplus,
+    const c10::optional<at::Tensor>& state_batch_indices,
+    const c10::optional<at::Tensor>& dst_state_batch_indices,
+    int64_t null_block_id, at::Tensor& out,
+    const c10::optional<at::Tensor>& num_accepted_tokens,
+    const c10::optional<at::Tensor>& cu_seqlens) {
+  at::ScalarType state_type = state.scalar_type();
+  at::ScalarType input_type = x.scalar_type();
+
+  // x, B, C must be contiguous and match input_type
+  auto ensure_input = [input_type](const at::Tensor& t) -> at::Tensor {
+    at::Tensor r = (t.scalar_type() != input_type) ? t.to(input_type) : t;
+    return r.is_contiguous() ? r : r.contiguous();
+  };
+  at::Tensor x_in = ensure_input(x);
+  at::Tensor B_in = ensure_input(B);
+  at::Tensor C_in = ensure_input(C);
+  at::Tensor z_in;
+  if (z.has_value() && z.value().defined()) z_in = ensure_input(z.value());
+
+  // A, D, dt_bias are float32 model parameters that arrive here as expanded
+  // tensors, e.g. A is (nheads, head_dim, dstate) with strides (1, 0, 0).
+  // We need just the scalar value per head as a (nheads,) 1-D array so that
+  // A_ptr[h] in the kernel correctly reads head h's value.
+  //
+  // Strategy: peel trailing expanded (stride=0) dims via .select(), which is
+  // a zero-copy view.  For A: (nheads, head_dim, dstate) strides (1,0,0)
+  //   → .select(2,0) → (nheads, head_dim) strides (1,0)
+  //   → .select(1,0) → (nheads,)          stride  (1,)   ← contiguous, free.
+  // No allocation, no type conversion (A is already float32).
+  auto to_per_head_1d_f32 = [](const at::Tensor& t) -> at::Tensor {
+    at::Tensor r = t;
+    // Peel trailing dimensions that are broadcast (stride=0 or size=1)
+    while (r.dim() > 1) r = r.select(r.dim() - 1, 0);
+    if (r.scalar_type() != at::kFloat) r = r.to(at::kFloat);
+    return r.is_contiguous() ? r : r.contiguous();
+  };
+
+  at::Tensor A_f32 = to_per_head_1d_f32(A);  // (nheads,) float32
+  at::Tensor D_f32, dt_bias_f32;
+  if (D.has_value() && D.value().defined())
+    D_f32 = to_per_head_1d_f32(D.value());
+  if (dt_bias.has_value() && dt_bias.value().defined())
+    dt_bias_f32 = to_per_head_1d_f32(dt_bias.value());
+
+  // dt: reduce (N, nheads, head_dim) expanded tensor → (N, nheads) BEFORE
+  // the type conversion so we convert head_dim x fewer elements.
+  at::Tensor dt_f32;
+  {
+    // If dt was expanded to (N, nheads, head_dim) with stride-0 in dim 2,
+    // take a zero-copy view of index 0 along that dim first.
+    at::Tensor t2 = (dt.dim() == 3) ? dt.select(2, 0) : dt;  // (N, nheads)
+    at::Tensor t3 = (t2.scalar_type() != at::kFloat) ? t2.to(at::kFloat) : t2;
+    dt_f32 = t3.is_contiguous() ? t3 : t3.contiguous();
+  }
+
+  int64_t nheads = state.size(1);
+  int64_t dim = state.size(2);
+  int64_t dstate = state.size(3);
+  int64_t N = (cu_seqlens.has_value() && cu_seqlens.value().defined())
+                  ? cu_seqlens.value().size(0) - 1
+                  : x_in.size(0);
+  int64_t ngroups = B_in.size(1);
+
+  // Strides
+  int64_t stride_state_n = state.stride(0);
+  int64_t stride_state_h = state.stride(1);
+  int64_t stride_state_d = state.stride(2);
+  int64_t stride_x_n = x_in.stride(0);
+  int64_t stride_x_h = x_in.stride(1);
+  int64_t stride_dt_n = dt_f32.stride(0);  // dt is (N, nheads)
+  int64_t stride_BC_n = B_in.stride(0);
+  int64_t stride_BC_g = B_in.stride(1);
+  int64_t stride_out_n = out.stride(0);
+  int64_t stride_out_h = out.stride(1);
+
+  // Optional index pointers
+  auto get_int32_ptr =
+      [](const c10::optional<at::Tensor>& opt) -> const int32_t* {
+    return (opt.has_value() && opt.value().defined())
+               ? opt.value().data_ptr<int32_t>()
+               : nullptr;
+  };
+  const int32_t* sbi_ptr = get_int32_ptr(state_batch_indices);
+  const int32_t* dsbi_ptr = get_int32_ptr(dst_state_batch_indices);
+  const int32_t* nat_ptr = get_int32_ptr(num_accepted_tokens);
+  const int32_t* csl_ptr = get_int32_ptr(cu_seqlens);
+
+  // Dispatch on (state_t, input_t, out_t): write directly into `out`
+  // without any intermediate float32 buffer.
+  VLLM_DISPATCH_FLOATING_TYPES(state_type, "ssu_state", [&] {
+    using state_t = scalar_t;
+    VLLM_DISPATCH_FLOATING_TYPES(input_type, "ssu_input", [&] {
+      using input_t = scalar_t;
+      VLLM_DISPATCH_FLOATING_TYPES(out.scalar_type(), "ssu_out", [&] {
+        using out_t = scalar_t;
+        mamba_cpu::selective_state_update_kernel<state_t, input_t, out_t>(
+            state.data_ptr<state_t>(), stride_state_n, stride_state_h,
+            stride_state_d, x_in.data_ptr<input_t>(), stride_x_n, stride_x_h,
+            dt_f32.data_ptr<float>(), stride_dt_n, A_f32.data_ptr<float>(),
+            B_in.data_ptr<input_t>(), C_in.data_ptr<input_t>(), stride_BC_n,
+            stride_BC_g, D_f32.defined() ? D_f32.data_ptr<float>() : nullptr,
+            z_in.defined() ? z_in.data_ptr<input_t>() : nullptr,
+            dt_bias_f32.defined() ? dt_bias_f32.data_ptr<float>() : nullptr,
+            out.data_ptr<out_t>(), stride_out_n, stride_out_h, sbi_ptr,
+            dsbi_ptr, static_cast<int32_t>(null_block_id), nat_ptr, csl_ptr, N,
+            nheads, ngroups, dim, dstate, dt_softplus);
+      });
+    });
+  });
+}
+
+// ---------------------------------------------------------------------------
+// mamba_chunk_scan_fwd_cpu
+// ---------------------------------------------------------------------------
+void mamba_chunk_scan_fwd_cpu_impl(
+    at::Tensor& out,  // [seqlen, nheads, headdim] — pre-allocated by caller
+    at::Tensor&
+        final_states,     // [batch, nheads, headdim, dstate] float32 contiguous
+    const at::Tensor& x,  // [seqlen, nheads, headdim]
+    const at::Tensor&
+        dt,  // [seqlen, nheads] float32 (preprocessed: bias+softplus+clamp)
+    const at::Tensor& A,                 // [nheads] float32
+    const at::Tensor& B,                 // [seqlen, ngroups, dstate]
+    const at::Tensor& C,                 // [seqlen, ngroups, dstate]
+    const c10::optional<at::Tensor>& D,  // [nheads] float32 (optional)
+    const c10::optional<at::Tensor>& z,  // [seqlen, nheads, headdim] (optional)
+    const at::Tensor& cu_seqlens         // [batch+1] int32
+) {
+  const at::ScalarType input_type = x.scalar_type();
+
+  auto ensure_contig = [input_type](const at::Tensor& t) -> at::Tensor {
+    at::Tensor r = (t.scalar_type() != input_type) ? t.to(input_type) : t;
+    return r.is_contiguous() ? r : r.contiguous();
+  };
+  at::Tensor x_in = ensure_contig(x);
+  at::Tensor B_in = ensure_contig(B);
+  at::Tensor C_in = ensure_contig(C);
+  at::Tensor z_in;
+  if (z.has_value() && z.value().defined()) z_in = ensure_contig(z.value());
+
+  // A and D are float32 model parameters, potentially broadcast-expanded.
+  // Strip trailing broadcast dims to get a contiguous (nheads,) array.
+  auto to_per_head_f32 = [](const at::Tensor& t) -> at::Tensor {
+    at::Tensor r = t;
+    while (r.dim() > 1) r = r.select(r.dim() - 1, 0);
+    if (r.scalar_type() != at::kFloat) r = r.to(at::kFloat);
+    return r.is_contiguous() ? r : r.contiguous();
+  };
+  at::Tensor A_f32 = to_per_head_f32(A);
+  at::Tensor D_f32;
+  if (D.has_value() && D.value().defined()) D_f32 = to_per_head_f32(D.value());
+
+  // dt: [seqlen, nheads] float32 — caller has applied bias+softplus+clamp in
+  // Python.
+  at::Tensor dt_c = dt.is_contiguous() ? dt : dt.contiguous();
+  if (dt_c.scalar_type() != at::kFloat) dt_c = dt_c.to(at::kFloat);
+
+  at::Tensor cu_int = cu_seqlens.to(at::kInt).contiguous();
+
+  const int64_t batch = final_states.size(0);
+  const int64_t nheads = final_states.size(1);
+  const int64_t headdim = final_states.size(2);
+  const int64_t dstate = final_states.size(3);
+  const int64_t ngroups = B_in.size(1);
+
+  TORCH_CHECK(final_states.is_contiguous(),
+              "mamba_chunk_scan_fwd_cpu: final_states must be contiguous");
+  TORCH_CHECK(out.is_contiguous(),
+              "mamba_chunk_scan_fwd_cpu: out must be contiguous (writes via "
+              "raw data_ptr)");
+
+  VLLM_DISPATCH_FLOATING_TYPES(input_type, "mamba_chunk_scan_fwd_cpu", [&] {
+    mamba_cpu::mamba_chunk_scan_fwd_kernel<scalar_t>(
+        final_states.data_ptr<float>(), x_in.data_ptr<scalar_t>(),
+        dt_c.data_ptr<float>(), A_f32.data_ptr<float>(),
+        B_in.data_ptr<scalar_t>(), C_in.data_ptr<scalar_t>(),
+        D_f32.defined() ? D_f32.data_ptr<float>() : nullptr,
+        z_in.defined() ? z_in.data_ptr<scalar_t>() : nullptr,
+        out.data_ptr<scalar_t>(), cu_int.data_ptr<int32_t>(), batch, nheads,
+        ngroups, headdim, dstate);
+  });
+}

csrc/cpu/mamba_kernels.hpp

@@ -0,0 +1,382 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+//
+// Fused CPU vector kernels for Mamba decode-step hotspots:
+//   - causal_conv1d_update  (depthwise 1-D conv state roll + compute)
+//   - selective_state_update (SSM recurrence, single-step)
+
+#pragma once
+
+#include "cpu_types.hpp"
+#include <cmath>
+#include <cstring>
+#include <cstdint>
+#include <algorithm>
+
+namespace mamba_cpu {
+
+// ---------------------------------------------------------------------------
+// causal_conv1d_update — templated for native BF16/FP32
+//
+// state_ptr may point to a NON-CONTIGUOUS paged KV cache tensor.
+// Explicit strides are passed so the kernel writes directly into the
+// correct memory locations without making a contiguous copy of the full
+// paged tensor (which was the source of the 34-41% direct_copy_kernel).
+//
+//   stride_s_slot  = state.stride(0)  — between cache slots
+//   stride_s_dim   = state.stride(1)  — between conv_dim channels
+//   stride_s_state = state.stride(2)  — between state elements
+//
+// When stride_s_state == 1 (contiguous), the memmove fast path is used.
+// ---------------------------------------------------------------------------
+template <typename scalar_t>
+inline void causal_conv1d_update_kernel(
+    const scalar_t* __restrict__ x_ptr, scalar_t* __restrict__ state_ptr,
+    int64_t stride_s_slot, int64_t stride_s_dim, int64_t stride_s_state,
+    const scalar_t* __restrict__ weight_ptr, const float* __restrict__ bias_ptr,
+    scalar_t* __restrict__ out_ptr, const int32_t* __restrict__ cache_idxs,
+    int32_t pad_slot_id, int64_t batch, int64_t dim, int64_t seqlen,
+    int64_t width, int64_t state_len, bool do_silu) {
+#pragma omp parallel for
+  for (int64_t b = 0; b < batch; ++b) {
+    int64_t cache_idx = (cache_idxs != nullptr) ? cache_idxs[b] : b;
+    if (cache_idx == pad_slot_id) continue;
+
+    for (int64_t t = 0; t < seqlen; ++t) {
+      const scalar_t* x_b = x_ptr + (b * dim * seqlen + t);
+      scalar_t* out_b = out_ptr + (b * dim * seqlen + t);
+      // Base of this slot in the (possibly non-contiguous) paged state
+      scalar_t* s_base = state_ptr + cache_idx * stride_s_slot;
+
+      for (int64_t d = 0; d < dim; ++d) {
+        float x_val = static_cast<float>(x_b[d * seqlen]);
+        scalar_t* sd = s_base + d * stride_s_dim;  // start of this dim's state
+        const scalar_t* w = weight_ptr + d * width;
+
+        // Accumulate in float32 for precision
+        float acc = (bias_ptr != nullptr) ? bias_ptr[d] : 0.0f;
+        for (int64_t k = 0; k < state_len; ++k) {
+          acc += static_cast<float>(w[k]) *
+                 static_cast<float>(sd[k * stride_s_state]);
+        }
+        acc += static_cast<float>(w[state_len]) * x_val;
+
+        // Shift state left and append new input.
+        // Use memmove when contiguous (stride==1); element loop otherwise.
+        if (stride_s_state == 1) {
+          if (state_len > 1)
+            std::memmove(sd, sd + 1, (state_len - 1) * sizeof(scalar_t));
+          if (state_len > 0) sd[state_len - 1] = static_cast<scalar_t>(x_val);
+        } else {
+          for (int64_t k = 0; k < state_len - 1; ++k)
+            sd[k * stride_s_state] = sd[(k + 1) * stride_s_state];
+          if (state_len > 0)
+            sd[(state_len - 1) * stride_s_state] = static_cast<scalar_t>(x_val);
+        }
+
+        if (do_silu) {
+          float sigmoid = (acc >= 0) ? 1.0f / (1.0f + std::exp(-acc))
+                                     : std::exp(acc) / (1.0f + std::exp(acc));
+          acc *= sigmoid;
+        }
+        out_b[d * seqlen] = static_cast<scalar_t>(acc);
+      }
+    }
+  }
+}
+
+// ---------------------------------------------------------------------------
+// selective_state_update
+//
+// Template parameters:
+//   state_t  - dtype of ssm_state cache (typically BFloat16)
+//   input_t  - dtype of x, B, C (typically BFloat16)
+//   out_t    - dtype of output tensor (typically BFloat16)
+//              Write directly — no float32 intermediate buffer needed.
+//
+// A, D, dt_bias are accepted as const float* (they are always float32
+// model parameters in Mamba2).  This eliminates the per-call float32→BF16
+// conversion and the .contiguous() materialisation of the broadcast-expand.
+//
+// dt is accepted as a (N, nheads) scalar-per-head tensor, not as the
+// (N, nheads, head_dim) expansion, so no .contiguous() copy is needed.
+// ---------------------------------------------------------------------------
+template <typename state_t, typename input_t, typename out_t = float>
+inline void selective_state_update_kernel(
+    state_t* __restrict__ state_ptr, int64_t stride_state_n,
+    int64_t stride_state_h, int64_t stride_state_d,
+    const input_t* __restrict__ x_ptr, int64_t stride_x_n, int64_t stride_x_h,
+    // dt: (N, nheads) — scalar per head, NOT expanded to head_dim
+    const float* __restrict__ dt_ptr, int64_t stride_dt_n,
+    // A: (nheads,) float32 — scalar per head
+    const float* __restrict__ A_ptr, const input_t* __restrict__ B_ptr,
+    const input_t* __restrict__ C_ptr, int64_t stride_BC_n, int64_t stride_BC_g,
+    // D: (nheads,) float32 — scalar per head (nullptr if not used)
+    const float* __restrict__ D_ptr,
+    // z: same shape as x (optional)
+    const input_t* __restrict__ z_ptr,
+    // dt_bias: (nheads,) float32 — scalar per head (nullptr if not used)
+    const float* __restrict__ dt_bias_ptr, out_t* __restrict__ out_ptr,
+    int64_t stride_out_n, int64_t stride_out_h,
+    const int32_t* __restrict__ state_batch_indices,
+    const int32_t* __restrict__ dst_state_batch_indices, int32_t null_block_id,
+    const int32_t* __restrict__ num_accepted_tokens,
+    const int32_t* __restrict__ cu_seqlens, int64_t N, int64_t nheads,
+    int64_t ngroups, int64_t dim, int64_t dstate, bool dt_softplus) {
+  using state_vec_t = vec_op::vec_t<state_t>;
+  using input_vec_t = vec_op::vec_t<input_t>;
+  constexpr int VEC_ELEM_NUM = 8;
+
+  int64_t nheads_per_group = nheads / ngroups;
+
+  for (int64_t seq_idx = 0; seq_idx < N; ++seq_idx) {
+    int64_t bos, seq_len;
+    if (cu_seqlens != nullptr) {
+      bos = cu_seqlens[seq_idx];
+      seq_len = cu_seqlens[seq_idx + 1] - bos;
+    } else {
+      bos = seq_idx;
+      seq_len = 1;
+    }
+
+    int64_t state_read_idx = (state_batch_indices != nullptr)
+                                 ? state_batch_indices[seq_idx]
+                                 : seq_idx;
+    if (state_read_idx == null_block_id) continue;
+
+    int64_t state_write_idx = (num_accepted_tokens == nullptr)
+                                  ? ((dst_state_batch_indices != nullptr)
+                                         ? dst_state_batch_indices[seq_idx]
+                                         : state_read_idx)
+                                  : -1;
+
+    state_t* s = state_ptr + state_read_idx * stride_state_n;
+
+    for (int64_t t = 0; t < seq_len; ++t) {
+      int64_t token_idx = bos + t;
+      const input_t* x_tok = x_ptr + token_idx * stride_x_n;
+      // dt: (N, nheads) — one float per head per token
+      const float* dt_tok = dt_ptr + token_idx * stride_dt_n;
+      const input_t* B_tok = B_ptr + token_idx * stride_BC_n;
+      const input_t* C_tok = C_ptr + token_idx * stride_BC_n;
+      out_t* out_tok = out_ptr + token_idx * stride_out_n;
+
+#pragma omp parallel for
+      for (int64_t h = 0; h < nheads; ++h) {
+        int64_t g = h / nheads_per_group;
+        const input_t* x_h = x_tok + h * stride_x_h;
+        const input_t* B_g = B_tok + g * stride_BC_g;
+        const input_t* C_g = C_tok + g * stride_BC_g;
+        out_t* out_h = out_tok + h * stride_out_h;
+        state_t* s_h = s + h * stride_state_h;
+
+        // Read scalars-per-head (A, dt, dt_bias, D) — no per-dim indexing
+        float dt_val = dt_tok[h];
+        if (dt_bias_ptr != nullptr) dt_val += dt_bias_ptr[h];
+        if (dt_softplus) {
+          dt_val = (dt_val <= 20.0f) ? std::log1p(std::exp(dt_val)) : dt_val;
+        }
+        const float A_val = A_ptr[h];  // scalar: same for all dim, dstate
+        const float D_val = (D_ptr != nullptr) ? D_ptr[h] : 0.0f;
+
+        const input_t* z_h =
+            (z_ptr != nullptr) ? z_ptr + token_idx * stride_x_n + h * stride_x_h
+                               : nullptr;
+
+        vec_op::FP32Vec8 dt_vec(dt_val);
+        // dA = exp(A * dt): A and dt are SCALARS per head, so compute once
+        // and broadcast.  This saves 7 redundant std::exp() calls that
+        // FP32Vec8::exp() would otherwise make on the broadcast vector.
+        const float dA_scalar = std::exp(A_val * dt_val);
+        vec_op::FP32Vec8 dA(dA_scalar);  // broadcast
+
+        for (int64_t d = 0; d < dim; ++d) {
+          float x_val = static_cast<float>(x_h[d]);
+
+          vec_op::FP32Vec8 out_vec(0.0f);
+          state_t* s_hd = s_h + d * stride_state_d;
+          const input_t* B_g_base = B_g;
+          const input_t* C_g_base = C_g;
+
+          vec_op::FP32Vec8 x_vec(x_val);
+          // dBx = B * x * dt — same dA for all dstate (A is scalar)
+          // s_new = s * dA + B * x * dt
+
+          int64_t n = 0;
+          for (; n <= dstate - VEC_ELEM_NUM; n += VEC_ELEM_NUM) {
+            vec_op::FP32Vec8 B_v((input_vec_t(B_g_base + n)));
+            vec_op::FP32Vec8 C_v((input_vec_t(C_g_base + n)));
+            vec_op::FP32Vec8 s_v((state_vec_t(s_hd + n)));
+
+            vec_op::FP32Vec8 dBx = B_v * x_vec * dt_vec;
+            vec_op::FP32Vec8 s_new = s_v * dA + dBx;
+
+            state_vec_t(s_new).save(s_hd + n);
+            out_vec = out_vec + s_new * C_v;
+          }
+
+          float out_val = out_vec.reduce_sum();
+          for (; n < dstate; ++n) {
+            // Reuse dA_scalar computed once per head — no exp() re-call
+            float dBx = static_cast<float>(B_g[n]) * x_val * dt_val;
+            float s_new = static_cast<float>(s_hd[n]) * dA_scalar + dBx;
+            s_hd[n] = static_cast<state_t>(s_new);
+            out_val += s_new * static_cast<float>(C_g[n]);
+          }
+
+          if (D_ptr != nullptr) out_val += x_val * D_val;
+          if (z_h != nullptr) {
+            float z_val = static_cast<float>(z_h[d]);
+            float sigmoid = (z_val >= 0)
+                                ? 1.0f / (1.0f + std::exp(-z_val))
+                                : std::exp(z_val) / (1.0f + std::exp(z_val));
+            out_val *= z_val * sigmoid;
+          }
+          out_h[d] = static_cast<out_t>(out_val);
+        }
+      }
+
+      if (num_accepted_tokens != nullptr &&
+          dst_state_batch_indices != nullptr) {
+        int64_t token_dst_idx = dst_state_batch_indices[seq_idx * seq_len + t];
+        if (token_dst_idx != null_block_id && token_dst_idx != state_read_idx) {
+          state_t* dst_s = state_ptr + token_dst_idx * stride_state_n;
+          std::memmove(dst_s, s, nheads * stride_state_h * sizeof(state_t));
+        }
+      }
+    }
+
+    if (num_accepted_tokens == nullptr && state_write_idx != null_block_id &&
+        state_write_idx != state_read_idx) {
+      state_t* dst_s = state_ptr + state_write_idx * stride_state_n;
+      std::memmove(dst_s, s, nheads * stride_state_h * sizeof(state_t));
+    }
+  }
+}
+
+// ---------------------------------------------------------------------------
+// mamba_chunk_scan_fwd
+//
+// Prefill SSM recurrence for Mamba2 / SSD models.
+//
+// Key difference from selective_state_update_kernel (decode path):
+//   - #pragma omp parallel for collapse(2) is OUTSIDE the time loop.
+//     Each thread owns a (batch, head) slice and runs the entire token
+//     sequence without any per-token OpenMP synchronisation overhead.
+//     For seqlen=256, this eliminates 256 thread-barrier launches per batch.
+//
+// `dt` arrives already processed (float32, after bias + softplus + clamp)
+// to keep this kernel simple. Preprocessing is done in the Python wrapper.
+//
+// `states_ptr` points to the [batch, nheads, headdim, dstate] float32 output
+// tensor, pre-initialised by the caller (zero or from initial_states).
+// Each (b, h) slice is private to exactly one thread via collapse(2), so
+// there are no write conflicts.
+//
+// D is treated as a scalar per head ([nheads] float32).
+// ---------------------------------------------------------------------------
+template <typename input_t>
+inline void mamba_chunk_scan_fwd_kernel(
+    float* __restrict__ states_ptr,     // [batch, nheads, headdim, dstate] f32
+    const input_t* __restrict__ x_ptr,  // [seqlen, nheads, headdim]
+    const float* __restrict__ dt_ptr,   // [seqlen, nheads] f32 (preprocessed)
+    const float* __restrict__ A_ptr,    // [nheads] f32
+    const input_t* __restrict__ B_ptr,  // [seqlen, ngroups, dstate]
+    const input_t* __restrict__ C_ptr,  // [seqlen, ngroups, dstate]
+    const float* __restrict__ D_ptr,    // [nheads] f32 (nullable)
+    const input_t* __restrict__ z_ptr,  // [seqlen, nheads, headdim] (nullable)
+    input_t* __restrict__ out_ptr,      // [seqlen, nheads, headdim]
+    const int32_t* __restrict__ cu_seqlens,  // [batch+1] int32
+    int64_t batch, int64_t nheads, int64_t ngroups, int64_t headdim,
+    int64_t dstate) {
+  using input_vec_t = vec_op::vec_t<input_t>;
+  constexpr int VEC_ELEM_NUM = 8;
+
+  const int64_t nheads_per_group = nheads / ngroups;
+  // states layout: [batch, nheads, headdim, dstate] contiguous (caller
+  // guarantee)
+  const int64_t stride_s_b = nheads * headdim * dstate;
+  const int64_t stride_s_h = headdim * dstate;
+  // stride_s_d = dstate, stride_s_n = 1
+
+#pragma omp parallel for collapse(2) schedule(static)
+  for (int64_t b = 0; b < batch; ++b) {
+    for (int64_t h = 0; h < nheads; ++h) {
+      const int64_t seq_start = cu_seqlens[b];
+      const int64_t seq_end = cu_seqlens[b + 1];
+      const int64_t g = h / nheads_per_group;
+
+      const float A_val = A_ptr[h];
+      const float D_val = (D_ptr != nullptr) ? D_ptr[h] : 0.0f;
+
+      // Working state slice: states[b, h, :, :] — float32, headdim * dstate.
+      // Fits in L1/L2 for typical dims (e.g. 64*128*4 = 32 KB).
+      float* s_bh = states_ptr + b * stride_s_b + h * stride_s_h;
+
+      for (int64_t t = seq_start; t < seq_end; ++t) {
+        const input_t* x_h = x_ptr + t * nheads * headdim + h * headdim;
+        const float* dt_h = dt_ptr + t * nheads + h;
+        const input_t* B_g = B_ptr + t * ngroups * dstate + g * dstate;
+        const input_t* C_g = C_ptr + t * ngroups * dstate + g * dstate;
+        const input_t* z_h = (z_ptr != nullptr)
+                                 ? z_ptr + t * nheads * headdim + h * headdim
+                                 : nullptr;
+        input_t* out_h = out_ptr + t * nheads * headdim + h * headdim;
+
+        const float dt_val = *dt_h;
+        const float dA_val = std::exp(A_val * dt_val);
+        const vec_op::FP32Vec8 dA_vec(dA_val);  // broadcast scalar
+        const vec_op::FP32Vec8 dt_vec(dt_val);
+
+        for (int64_t d = 0; d < headdim; ++d) {
+          const float x_val = static_cast<float>(x_h[d]);
+          float* s_bhd = s_bh + d * dstate;  // [dstate] contiguous float32
+
+          // Vectorised SSM update + readout over dstate:
+          //   s_new = s * dA + x * dt * B
+          //   y    += s_new * C
+          int64_t n = 0;
+          vec_op::FP32Vec8 y_vec(0.0f);
+          const vec_op::FP32Vec8 x_vec(x_val);
+
+          for (; n <= dstate - VEC_ELEM_NUM; n += VEC_ELEM_NUM) {
+            const vec_op::FP32Vec8 B_v((input_vec_t(B_g + n)));
+            const vec_op::FP32Vec8 C_v((input_vec_t(C_g + n)));
+            const vec_op::FP32Vec8 s_v(s_bhd + n);
+
+            const vec_op::FP32Vec8 s_new = s_v * dA_vec + x_vec * dt_vec * B_v;
+            s_new.save(s_bhd + n);
+            y_vec = y_vec + s_new * C_v;
+          }
+
+          float y_val = y_vec.reduce_sum();
+
+          // Scalar tail for remaining dstate elements
+          for (; n < dstate; ++n) {
+            const float B_n = static_cast<float>(B_g[n]);
+            const float C_n = static_cast<float>(C_g[n]);
+            const float s_new = s_bhd[n] * dA_val + x_val * dt_val * B_n;
+            s_bhd[n] = s_new;
+            y_val += s_new * C_n;
+          }
+
+          // D skip connection (scalar per head)
+          if (D_ptr != nullptr) y_val += x_val * D_val;
+
+          // z gating: out = y * z * sigmoid(z)  (SiLU)
+          if (z_h != nullptr) {
+            const float z_val = static_cast<float>(z_h[d]);
+            const float sigmoid =
+                (z_val >= 0.0f) ? 1.0f / (1.0f + std::exp(-z_val))
+                                : std::exp(z_val) / (1.0f + std::exp(z_val));
+            y_val *= z_val * sigmoid;
+          }
+
+          out_h[d] = static_cast<input_t>(y_val);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace mamba_cpu

csrc/cpu/micro_gemm/cpu_micro_gemm_impl.hpp

@@ -31,6 +31,9 @@ class MicroGemm {
   }
 };
 
+template <cpu_utils::ISA isa, typename scalar_t>
+class MicroGemmINT8;
+
 template <int32_t n_size, typename scalar_t>
 FORCE_INLINE void default_epilogue(float* __restrict__ c_ptr,
                                    scalar_t* __restrict__ d_ptr,

csrc/cpu/micro_gemm/cpu_micro_gemm_int8_neon.hpp

@@ -0,0 +1,424 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+#ifndef CPU_MICRO_GEMM_INT8_NEON_HPP
+#define CPU_MICRO_GEMM_INT8_NEON_HPP
+
+#include <algorithm>
+#include <cstdint>
+
+#include "cpu/micro_gemm/cpu_micro_gemm_impl.hpp"
+
+#include <arm_bf16.h>
+#include <arm_neon.h>
+#include <c10/util/BFloat16.h>
+#include <c10/util/Exception.h>
+#include <c10/util/Half.h>
+
+namespace cpu_micro_gemm {
+
+namespace neon_smmla {
+
+constexpr int32_t K = 8;
+constexpr int32_t Cols = 2;
+constexpr int32_t TileSize = K * Cols;
+
+FORCE_INLINE float32x4x2_t load_as_f32(const float* input) {
+  float32x4x2_t result;
+  result.val[0] = vld1q_f32(input);
+  result.val[1] = vld1q_f32(input + 4);
+  return result;
+}
+
+FORCE_INLINE float32x4x2_t load_as_f32(const c10::Half* input) {
+  const auto input_vec = vld1q_f16(reinterpret_cast<const float16_t*>(input));
+  float32x4x2_t result;
+  result.val[0] = vcvt_f32_f16(vget_low_f16(input_vec));
+  result.val[1] = vcvt_f32_f16(vget_high_f16(input_vec));
+  return result;
+}
+
+FORCE_INLINE float32x4x2_t load_as_f32(const c10::BFloat16* input) {
+  const auto input_vec = vld1q_bf16(reinterpret_cast<const bfloat16_t*>(input));
+  float32x4x2_t result;
+  result.val[0] = vcvt_f32_bf16(vget_low_bf16(input_vec));
+  result.val[1] = vcvt_f32_bf16(vget_high_bf16(input_vec));
+  return result;
+}
+
+FORCE_INLINE void store_acc_rowpair(const int32x4_t acc01,
+                                    const int32x4_t acc23,
+                                    const int32x4_t acc45,
+                                    const int32x4_t acc67,
+                                    int32_t* __restrict__ c_ptr,
+                                    const int64_t ldc, const int32_t m_rows) {
+  if (m_rows == 0) {
+    return;
+  }
+
+  vst1q_s32(c_ptr, vcombine_s32(vget_low_s32(acc01), vget_low_s32(acc23)));
+  vst1q_s32(c_ptr + 4, vcombine_s32(vget_low_s32(acc45), vget_low_s32(acc67)));
+
+  if (m_rows == 2) {
+    vst1q_s32(c_ptr + ldc,
+              vcombine_s32(vget_high_s32(acc01), vget_high_s32(acc23)));
+    vst1q_s32(c_ptr + ldc + 4,
+              vcombine_s32(vget_high_s32(acc45), vget_high_s32(acc67)));
+  }
+}
+
+FORCE_INLINE void gemm_micro_smmla_8x8_packed_a(
+    const int8_t* __restrict__ a_packed, const int8_t* __restrict__ b_packed,
+    int32_t* __restrict__ c_ptr, const int32_t m, const int32_t k_size,
+    const int64_t ldc) {
+  const int32x4_t zero = vdupq_n_s32(0);
+  int32x4_t acc0101 = zero, acc0123 = zero, acc0145 = zero, acc0167 = zero;
+  int32x4_t acc2301 = zero, acc2323 = zero, acc2345 = zero, acc2367 = zero;
+  int32x4_t acc4501 = zero, acc4523 = zero, acc4545 = zero, acc4567 = zero;
+  int32x4_t acc6701 = zero, acc6723 = zero, acc6745 = zero, acc6767 = zero;
+
+  const int8_t* __restrict__ a_tile = a_packed;
+  const int8_t* __restrict__ b_tile = b_packed;
+
+#pragma GCC unroll 8
+  for (int32_t k_idx = 0; k_idx < k_size; k_idx += K) {
+    const int8x16_t a_tile01 = vld1q_s8(a_tile);
+    const int8x16_t a_tile23 = vld1q_s8(a_tile + TileSize);
+    const int8x16_t a_tile45 = vld1q_s8(a_tile + 2 * TileSize);
+    const int8x16_t a_tile67 = vld1q_s8(a_tile + 3 * TileSize);
+
+    const int8x16_t b_tile01 = vld1q_s8(b_tile);
+    const int8x16_t b_tile23 = vld1q_s8(b_tile + TileSize);
+    const int8x16_t b_tile45 = vld1q_s8(b_tile + 2 * TileSize);
+    const int8x16_t b_tile67 = vld1q_s8(b_tile + 3 * TileSize);
+
+    acc0101 = vmmlaq_s32(acc0101, a_tile01, b_tile01);
+    acc2301 = vmmlaq_s32(acc2301, a_tile23, b_tile01);
+    acc4501 = vmmlaq_s32(acc4501, a_tile45, b_tile01);
+    acc6701 = vmmlaq_s32(acc6701, a_tile67, b_tile01);
+
+    acc0123 = vmmlaq_s32(acc0123, a_tile01, b_tile23);
+    acc2323 = vmmlaq_s32(acc2323, a_tile23, b_tile23);
+    acc4523 = vmmlaq_s32(acc4523, a_tile45, b_tile23);
+    acc6723 = vmmlaq_s32(acc6723, a_tile67, b_tile23);
+
+    acc0145 = vmmlaq_s32(acc0145, a_tile01, b_tile45);
+    acc2345 = vmmlaq_s32(acc2345, a_tile23, b_tile45);
+    acc4545 = vmmlaq_s32(acc4545, a_tile45, b_tile45);
+    acc6745 = vmmlaq_s32(acc6745, a_tile67, b_tile45);
+
+    acc0167 = vmmlaq_s32(acc0167, a_tile01, b_tile67);
+    acc2367 = vmmlaq_s32(acc2367, a_tile23, b_tile67);
+    acc4567 = vmmlaq_s32(acc4567, a_tile45, b_tile67);
+    acc6767 = vmmlaq_s32(acc6767, a_tile67, b_tile67);
+
+    a_tile += 4 * TileSize;
+    b_tile += 4 * TileSize;
+  }
+
+  store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc,
+                    std::min(2, m));
+  store_acc_rowpair(acc2301, acc2323, acc2345, acc2367, c_ptr + 2 * ldc, ldc,
+                    std::min(2, std::max(0, m - 2)));
+  store_acc_rowpair(acc4501, acc4523, acc4545, acc4567, c_ptr + 4 * ldc, ldc,
+                    std::min(2, std::max(0, m - 4)));
+  store_acc_rowpair(acc6701, acc6723, acc6745, acc6767, c_ptr + 6 * ldc, ldc,
+                    std::min(2, std::max(0, m - 6)));
+}
+
+FORCE_INLINE void gemm_micro_smmla_4x16_packed_a(
+    const int8_t* __restrict__ a_packed, const int8_t* __restrict__ b_packed,
+    int32_t* __restrict__ c_ptr, const int32_t m, const int32_t k_size,
+    const int64_t b_n_group_stride, const int64_t ldc) {
+  const int32_t m_rows_01 = std::min(2, m);
+  const int32_t m_rows_23 = std::min(2, std::max(0, m - 2));
+  const int32x4_t zero = vdupq_n_s32(0);
+
+  int32x4_t acc0101 = zero, acc0123 = zero, acc0145 = zero, acc0167 = zero;
+  int32x4_t acc2301 = zero, acc2323 = zero, acc2345 = zero, acc2367 = zero;
+  int32x4_t acc0189 = zero, acc011011 = zero, acc011213 = zero,
+            acc011415 = zero;
+  int32x4_t acc2389 = zero, acc231011 = zero, acc231213 = zero,
+            acc231415 = zero;
+
+  const int8_t* __restrict__ a_tile = a_packed;
+  // note: b packs 8 panels contiguously, so we need 2 b_tile ptrs
+  // for the 4x16 microkernel
+  const int8_t* __restrict__ b_tile0 = b_packed;
+  const int8_t* __restrict__ b_tile1 = b_packed + b_n_group_stride;
+
+#pragma GCC unroll 8
+  for (int32_t k_idx = 0; k_idx < k_size; k_idx += K) {
+    const int8x16_t a_tile01 = vld1q_s8(a_tile);
+    const int8x16_t a_tile23 = vld1q_s8(a_tile + TileSize);
+    const int8x16_t b_tile01 = vld1q_s8(b_tile0);
+    const int8x16_t b_tile23 = vld1q_s8(b_tile0 + TileSize);
+    const int8x16_t b_tile45 = vld1q_s8(b_tile0 + 2 * TileSize);
+    const int8x16_t b_tile67 = vld1q_s8(b_tile0 + 3 * TileSize);
+    const int8x16_t b_tile89 = vld1q_s8(b_tile1);
+    const int8x16_t b_tile1011 = vld1q_s8(b_tile1 + TileSize);
+    const int8x16_t b_tile1213 = vld1q_s8(b_tile1 + 2 * TileSize);
+    const int8x16_t b_tile1415 = vld1q_s8(b_tile1 + 3 * TileSize);
+
+    acc0101 = vmmlaq_s32(acc0101, a_tile01, b_tile01);
+    acc2301 = vmmlaq_s32(acc2301, a_tile23, b_tile01);
+    acc0123 = vmmlaq_s32(acc0123, a_tile01, b_tile23);
+    acc2323 = vmmlaq_s32(acc2323, a_tile23, b_tile23);
+
+    acc0145 = vmmlaq_s32(acc0145, a_tile01, b_tile45);
+    acc2345 = vmmlaq_s32(acc2345, a_tile23, b_tile45);
+    acc0167 = vmmlaq_s32(acc0167, a_tile01, b_tile67);
+    acc2367 = vmmlaq_s32(acc2367, a_tile23, b_tile67);
+
+    acc0189 = vmmlaq_s32(acc0189, a_tile01, b_tile89);
+    acc2389 = vmmlaq_s32(acc2389, a_tile23, b_tile89);
+    acc011011 = vmmlaq_s32(acc011011, a_tile01, b_tile1011);
+    acc231011 = vmmlaq_s32(acc231011, a_tile23, b_tile1011);
+
+    acc011213 = vmmlaq_s32(acc011213, a_tile01, b_tile1213);
+    acc231213 = vmmlaq_s32(acc231213, a_tile23, b_tile1213);
+    acc011415 = vmmlaq_s32(acc011415, a_tile01, b_tile1415);
+    acc231415 = vmmlaq_s32(acc231415, a_tile23, b_tile1415);
+
+    a_tile += 2 * TileSize;
+    b_tile0 += 4 * TileSize;
+    b_tile1 += 4 * TileSize;
+  }
+
+  // rows 0-1, columns 0-7
+  store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc, m_rows_01);
+  // rows 0-1, columns 8-15
+  store_acc_rowpair(acc0189, acc011011, acc011213, acc011415, c_ptr + 8, ldc,
+                    m_rows_01);
+  // rows 2-3, columns 0-7
+  store_acc_rowpair(acc2301, acc2323, acc2345, acc2367, c_ptr + 2 * ldc, ldc,
+                    m_rows_23);
+  // rows 2-3, columns 8-15
+  store_acc_rowpair(acc2389, acc231011, acc231213, acc231415,
+                    c_ptr + 2 * ldc + 8, ldc, m_rows_23);
+}
+
+}  // namespace neon_smmla
+
+template <typename scalar_t>
+class MicroGemmINT8<cpu_utils::ISA::NEON, scalar_t> {
+ public:
+  static constexpr int32_t K = neon_smmla::K;
+  static constexpr int32_t Mr = 8;
+  static constexpr int32_t Nr = 8;
+  static constexpr int32_t NrGemv = 16;
+  static constexpr int32_t MaxMSize = 8;
+  static constexpr int32_t NSize = 32;
+  static constexpr int32_t WeightOCGroupSize = Nr;
+  static_assert(MaxMSize % Mr == 0);
+
+  static FORCE_INLINE void quantize_row(const scalar_t* input, int8_t* output,
+                                        float& scale, const int32_t size) {
+    TORCH_CHECK_EQ(size % K, 0);
+    float32x4_t max_vec = vdupq_n_f32(0.0f);
+
+    for (int32_t i = 0; i < size; i += K) {
+      const float32x4x2_t input_vec = neon_smmla::load_as_f32(input + i);
+      max_vec = vmaxq_f32(max_vec, vabsq_f32(input_vec.val[0]));
+      max_vec = vmaxq_f32(max_vec, vabsq_f32(input_vec.val[1]));
+    }
+
+    const float abs_max = std::max(vmaxvq_f32(max_vec), 1.0e-7f);
+    scale = abs_max / 127.0f;
+    const float32x4_t inv_scale_vec = vdupq_n_f32(127.0f / abs_max);
+
+    for (int32_t i = 0; i < size; i += K) {
+      const float32x4x2_t input_vec = neon_smmla::load_as_f32(input + i);
+      const int32x4_t output_low =
+          vcvtnq_s32_f32(vmulq_f32(input_vec.val[0], inv_scale_vec));
+      const int32x4_t output_high =
+          vcvtnq_s32_f32(vmulq_f32(input_vec.val[1], inv_scale_vec));
+      const int16x8_t output_s16 =
+          vcombine_s16(vqmovn_s32(output_low), vqmovn_s32(output_high));
+      vst1_s8(output + i, vqmovn_s16(output_s16));
+    }
+  }
+
+  // with current code, fusing this into the gemm micro kernel didn't move the
+  // needle
+  static FORCE_INLINE void dequantize_tile(
+      int32_t* input, float* output, const float* __restrict__ input_scales,
+      const float* __restrict__ weight_scales, const int32_t m, const int32_t n,
+      const int32_t stride) {
+    TORCH_CHECK_EQ(n % 4, 0);
+    for (int32_t m_idx = 0; m_idx < m; ++m_idx) {
+      const float32x4_t input_scale_vec = vdupq_n_f32(input_scales[m_idx]);
+      for (int32_t n_idx = 0; n_idx < n; n_idx += 4) {
+        const int32x4_t input_vec = vld1q_s32(input + m_idx * stride + n_idx);
+        const float32x4_t weight_scale_vec = vld1q_f32(weight_scales + n_idx);
+        const float32x4_t output_vec =
+            vmulq_f32(vcvtq_f32_s32(input_vec),
+                      vmulq_f32(input_scale_vec, weight_scale_vec));
+        vst1q_f32(output + m_idx * stride + n_idx, output_vec);
+      }
+    }
+  }
+
+  // physical layout [
+  //  M / (8 or 4); Mr is 8 or 4
+  //  K / 8; K for smmla is 8
+  //  4,   ; 4 row-pairs for each 8 rows
+  //  2,   ; row-pair is 2 rows
+  //  4    ; 4 elements per row
+  // ]
+  static void pack_input_from_rows(const int8_t* const* __restrict__ rows,
+                                   int8_t* __restrict__ a_packed,
+                                   const int32_t m, const int32_t k) {
+    TORCH_CHECK(m > 0 && m <= MaxMSize);
+    TORCH_CHECK(k % K == 0);
+    const int8x8_t zero = vdup_n_s8(0);
+
+    for (int32_t row_base = 0; row_base < m; row_base += Mr) {
+      const int32_t panel_m = std::min(Mr, m - row_base);
+      const int8_t* const* panel_rows = rows + row_base;
+      int8_t* __restrict__ out = a_packed + row_base * k;
+
+      // fast path for full 8-row panels (fast path for 4-row panels didn't move
+      // the needle)
+      if (panel_m == Mr) {
+        const int8_t* __restrict__ row0 = panel_rows[0];
+        const int8_t* __restrict__ row1 = panel_rows[1];
+        const int8_t* __restrict__ row2 = panel_rows[2];
+        const int8_t* __restrict__ row3 = panel_rows[3];
+        const int8_t* __restrict__ row4 = panel_rows[4];
+        const int8_t* __restrict__ row5 = panel_rows[5];
+        const int8_t* __restrict__ row6 = panel_rows[6];
+        const int8_t* __restrict__ row7 = panel_rows[7];
+        int32_t k_idx = 0;
+        for (; k_idx + 2 * K <= k; k_idx += 2 * K) {
+          int8_t* __restrict__ block0 = out;
+          int8_t* __restrict__ block1 = out + 4 * neon_smmla::TileSize;
+
+          int8x16_t a0 = vld1q_s8(row0 + k_idx);
+          int8x16_t a1 = vld1q_s8(row1 + k_idx);
+          vst1q_s8(block0, vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1, vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          a0 = vld1q_s8(row2 + k_idx);
+          a1 = vld1q_s8(row3 + k_idx);
+          vst1q_s8(block0 + neon_smmla::TileSize,
+                   vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1 + neon_smmla::TileSize,
+                   vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          a0 = vld1q_s8(row4 + k_idx);
+          a1 = vld1q_s8(row5 + k_idx);
+          vst1q_s8(block0 + 2 * neon_smmla::TileSize,
+                   vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1 + 2 * neon_smmla::TileSize,
+                   vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          a0 = vld1q_s8(row6 + k_idx);
+          a1 = vld1q_s8(row7 + k_idx);
+          vst1q_s8(block0 + 3 * neon_smmla::TileSize,
+                   vcombine_s8(vget_low_s8(a0), vget_low_s8(a1)));
+          vst1q_s8(block1 + 3 * neon_smmla::TileSize,
+                   vcombine_s8(vget_high_s8(a0), vget_high_s8(a1)));
+
+          out += 8 * neon_smmla::TileSize;
+        }
+
+        for (; k_idx < k; k_idx += K) {
+          int8x8_t a0 = vld1_s8(row0 + k_idx);
+          int8x8_t a1 = vld1_s8(row1 + k_idx);
+          vst1q_s8(out, vcombine_s8(a0, a1));
+
+          a0 = vld1_s8(row2 + k_idx);
+          a1 = vld1_s8(row3 + k_idx);
+          vst1q_s8(out + neon_smmla::TileSize, vcombine_s8(a0, a1));
+
+          a0 = vld1_s8(row4 + k_idx);
+          a1 = vld1_s8(row5 + k_idx);
+          vst1q_s8(out + 2 * neon_smmla::TileSize, vcombine_s8(a0, a1));
+
+          a0 = vld1_s8(row6 + k_idx);
+          a1 = vld1_s8(row7 + k_idx);
+          vst1q_s8(out + 3 * neon_smmla::TileSize, vcombine_s8(a0, a1));
+
+          out += 4 * neon_smmla::TileSize;
+        }
+        continue;
+      }
+
+      const int32_t row_pairs = (panel_m <= 4) ? 2 : Mr / 2;
+      for (int32_t k_idx = 0; k_idx < k; k_idx += K) {
+        for (int32_t pair_idx = 0; pair_idx < row_pairs; ++pair_idx) {
+          const int32_t row_idx = pair_idx * 2;
+          const int8x8_t row0 =
+              (row_idx < panel_m) ? vld1_s8(panel_rows[row_idx] + k_idx) : zero;
+          const int8x8_t row1 = (row_idx + 1 < panel_m)
+                                    ? vld1_s8(panel_rows[row_idx + 1] + k_idx)
+                                    : zero;
+          vst1q_s8(out, vcombine_s8(row0, row1));
+          out += neon_smmla::TileSize;
+        }
+      }
+    }
+  }
+
+  // physical layout [
+  //  N / 8; Nr is 8
+  //  K / 8; K for smmla is 8
+  //  4,   ; 4 col-pairs for each 8 cols
+  //  2,   ; col-pair is 2 cols
+  //  4    ; 4 elements per col
+  // ]
+  static void pack_weight(const int8_t* __restrict__ weight,
+                          int8_t* __restrict__ packed_weight,
+                          const int32_t output_size, const int32_t input_size) {
+    TORCH_CHECK(output_size % NSize == 0);
+    TORCH_CHECK(input_size % K == 0);
+
+    for (int32_t o_idx = 0; o_idx < output_size; o_idx += Nr) {
+      int8_t* __restrict__ dst = packed_weight + o_idx * input_size;
+      for (int32_t k_idx = 0; k_idx < input_size; k_idx += K) {
+        for (int32_t pair_idx = 0; pair_idx < Nr;
+             pair_idx += neon_smmla::Cols) {
+          const int8_t* __restrict__ row0 =
+              weight + (o_idx + pair_idx) * input_size + k_idx;
+          const int8_t* __restrict__ row1 = row0 + input_size;
+          vst1q_s8(dst, vcombine_s8(vld1_s8(row0), vld1_s8(row1)));
+          dst += neon_smmla::TileSize;
+        }
+      }
+    }
+  }
+
+  void gemm(const int8_t* __restrict__ a_packed,
+            const int8_t* __restrict__ b_packed, int32_t* __restrict__ c,
+            const int32_t m, const int32_t k, const int64_t b_n_group_stride,
+            const int64_t ldc) const {
+    TORCH_CHECK(m > 0 && m <= MaxMSize);
+    TORCH_CHECK(k % K == 0);
+
+    for (int32_t n_idx = 0; n_idx < NSize; n_idx += NrGemv) {
+      const int8_t* __restrict__ b_panel = b_packed + n_idx * k;
+
+      for (int32_t row_base = 0; row_base < m; row_base += Mr) {
+        const int32_t panel_m = std::min(Mr, m - row_base);
+        const int8_t* __restrict__ a_panel = a_packed + row_base * k;
+        int32_t* __restrict__ c_panel = c + row_base * ldc + n_idx;
+
+        if (panel_m <= 4) {
+          neon_smmla::gemm_micro_smmla_4x16_packed_a(
+              a_panel, b_panel, c_panel, panel_m, k, b_n_group_stride, ldc);
+        } else {
+          neon_smmla::gemm_micro_smmla_8x8_packed_a(a_panel, b_panel, c_panel,
+                                                    panel_m, k, ldc);
+          neon_smmla::gemm_micro_smmla_8x8_packed_a(
+              a_panel, b_panel + b_n_group_stride, c_panel + Nr, panel_m, k,
+              ldc);
+        }
+      }
+    }
+  }
+};
+
+}  // namespace cpu_micro_gemm
+
+#endif

csrc/cpu/micro_gemm/cpu_micro_gemm_neon.hpp

@@ -1,3 +1,6 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 #ifndef CPU_MICRO_GEMM_NEON_HPP
 #define CPU_MICRO_GEMM_NEON_HPP
 
@@ -16,9 +19,6 @@ namespace {
 constexpr int32_t K = 4;
 constexpr int32_t Cols = 2;
 constexpr int32_t TileSize = K * Cols;
-constexpr int32_t Mr = 8;
-constexpr int32_t Nr = 8;
-constexpr int32_t Nr_gemv = 16;
 
 // a = [a0, a1, a2, a3], b = [b0, b1, b2, b3] -> [a0, a1, b0, b1]
 FORCE_INLINE float32x4_t zip1_f32x4(const float32x4_t a, const float32x4_t b) {
@@ -132,7 +132,7 @@ FORCE_INLINE void gemm_micro_bfmmla_8x8_packed_a(
     acc6767 = vbfmmlaq_f32(acc6767, a_tile67, b_tile67);
 
     a_tile += 4 * TileSize;
-    b_tile += Nr * K;
+    b_tile += 4 * TileSize;
   }
 
   store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc,
@@ -205,8 +205,8 @@ FORCE_INLINE void gemm_micro_bfmmla_4x16_packed_a(
     acc231415 = vbfmmlaq_f32(acc231415, a_tile23, b_tile1415);
 
     a_tile += 2 * TileSize;
-    b_tile0 += Nr * K;
-    b_tile1 += Nr * K;
+    b_tile0 += 4 * TileSize;
+    b_tile1 += 4 * TileSize;
   }
 
   store_acc_rowpair(acc0101, acc0123, acc0145, acc0167, c_ptr, ldc, m_rows_01);
@@ -223,6 +223,9 @@ FORCE_INLINE void gemm_micro_bfmmla_4x16_packed_a(
 template <typename scalar_t>
 class MicroGemm<cpu_utils::ISA::NEON, scalar_t> {
  public:
+  static constexpr int32_t Mr = 8;
+  static constexpr int32_t Nr = 8;
+  static constexpr int32_t NrGemv = 16;
   static constexpr int32_t MaxMSize = 8;
   static constexpr int32_t NSize = 32;
   static constexpr int32_t WeightOCGroupSize = Nr;
@@ -246,6 +249,9 @@ class MicroGemm<cpu_utils::ISA::NEON, c10::BFloat16> {
  public:
   using scalar_t = c10::BFloat16;
 
+  static constexpr int32_t Mr = 8;
+  static constexpr int32_t Nr = 8;
+  static constexpr int32_t NrGemv = 16;
   static constexpr int32_t MaxMSize = 8;
   static constexpr int32_t NSize = 32;
   static constexpr int32_t WeightOCGroupSize = Nr;
@@ -253,7 +259,7 @@ class MicroGemm<cpu_utils::ISA::NEON, c10::BFloat16> {
 
  public:
   // physical layout [
-  //  M / 8; Mr is 8
+  //  M / (8 or 4); Mr is 8 or 4
   //  K / 4; K for bfmmla is 4
   //  4,   ; 4 row-pairs for each 8 rows
   //  2,   ; row-pair is 2 rows
@@ -439,7 +445,7 @@ class MicroGemm<cpu_utils::ISA::NEON, c10::BFloat16> {
     (void)lda;  // A is packed, so lda is not needed
     TORCH_CHECK_EQ(k % K, 0);
 
-    for (int32_t n_idx = 0; n_idx < NSize; n_idx += Nr_gemv) {
+    for (int32_t n_idx = 0; n_idx < NSize; n_idx += NrGemv) {
       const bfloat16_t* __restrict__ b_panel =
           reinterpret_cast<const bfloat16_t*>(b_ptr) + n_idx * k;
 

csrc/cpu/sgl-kernels/conv.cpp

@@ -451,6 +451,90 @@ void causal_conv1d_update_kernel_impl(
   });
 }
 
+template <typename scalar_t>
+void causal_conv1d_update_multi_kernel_impl(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ input,
+    scalar_t* __restrict__ conv_states,
+    const scalar_t* __restrict__ weight,
+    const scalar_t* __restrict__ bias,
+    const int32_t* __restrict__ num_accepted_tokens,
+    const int32_t* __restrict__ conv_indices,
+    bool silu_activation,
+    int64_t batch,
+    int64_t dim,
+    int64_t seqlen,
+    int64_t width,
+    int64_t state_len,
+    int64_t conv_state_slot_stride) {
+  constexpr int64_t BLOCK_N = block_size_n() * 2;
+  const int64_t NB = div_up(dim, BLOCK_N);
+
+  AT_DISPATCH_BOOL2(bias != nullptr, has_bias, silu_activation, has_silu, [&] {
+    at::parallel_for(0, batch * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t bs{0}, nb{0};
+      data_index_init(begin, bs, batch, nb, NB);
+
+      for (int64_t i = begin; i < end; ++i) {
+        const int64_t nb_start = nb * BLOCK_N;
+        const int64_t nb_size = std::min(dim - nb_start, BLOCK_N);
+        const int32_t conv_state_index = conv_indices[bs];
+        const int32_t history_offset = num_accepted_tokens[bs] - 1;
+
+        switch (width << 4 | nb_size >> 4) {
+          case 0x42:
+            tinygemm_kernel<scalar_t, 4, 32, has_bias, has_silu>::apply(
+                input + bs * seqlen * dim + nb_start,
+                weight + nb_start * width,
+                out + bs * seqlen * dim + nb_start,
+                has_bias ? bias + nb_start : nullptr,
+                conv_states + conv_state_index * conv_state_slot_stride +
+                    history_offset * dim + nb_start,
+                true,
+                seqlen,
+                dim,
+                true);
+            break;
+          case 0x44:
+            tinygemm_kernel<scalar_t, 4, 64, has_bias, has_silu>::apply(
+                input + bs * seqlen * dim + nb_start,
+                weight + nb_start * width,
+                out + bs * seqlen * dim + nb_start,
+                has_bias ? bias + nb_start : nullptr,
+                conv_states + conv_state_index * conv_state_slot_stride +
+                    history_offset * dim + nb_start,
+                true,
+                seqlen,
+                dim,
+                true);
+            break;
+          default:
+            TORCH_CHECK(false, "Unexpected block size, ", width, " x ", nb_size);
+        }
+
+        data_index_step(bs, batch, nb, NB);
+      }
+    });
+  });
+
+  at::parallel_for(0, batch, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t bs = begin; bs < end; ++bs) {
+      const int32_t conv_state_index = conv_indices[bs];
+      const int32_t num_accepted = num_accepted_tokens[bs];
+      scalar_t* state = conv_states + conv_state_index * conv_state_slot_stride;
+
+      std::memmove(
+          state,
+          state + num_accepted * dim,
+          (state_len - seqlen) * dim * sizeof(scalar_t));
+      std::memcpy(
+          state + (state_len - seqlen) * dim,
+          input + bs * seqlen * dim,
+          seqlen * dim * sizeof(scalar_t));
+    }
+  });
+}
+
 }  // anonymous namespace
 
 // from [dim, width] or [N, K]
@@ -545,7 +629,7 @@ at::Tensor get_block_indices(const std::optional<at::Tensor>& offsets, int64_t n
 //   query_start_loc: (batch + 1) int32
 //   cache_indices: (batch)  int32
 //   has_initial_state: (batch) bool
-//   conv_states: (..., dim, width - 1) itype
+//   conv_states: (..., dim, state_len) itype, where state_len >= width - 1
 //   activation: either None or "silu" or "swish"
 //   pad_slot_id: int
 //
@@ -586,11 +670,14 @@ at::Tensor causal_conv1d_fwd_cpu(
     CHECK_EQ(conv_states_val.scalar_type(), scalar_type);
     CHECK_GE(padded_batch, batch);
     CHECK_EQ(conv_states_val.size(1), dim);
-    CHECK_EQ(conv_states_val.size(2), width - 1);
+    const int64_t state_len = conv_states_val.size(2);
+    CHECK_GE(state_len, width - 1);
 
     // adjust `conv_states` to be contiguous on `dim`
     // should happen only once
     if (conv_states_val.stride(-2) != 1) {
+      TORCH_CHECK(state_len == width - 1,
+          "causal_conv1d_fwd_cpu: wide conv_states must be contiguous on dim.");
       auto conv_states_copy = conv_states_val.clone();
       conv_states_val.as_strided_({padded_batch, dim, width - 1}, {(width - 1) * dim, 1, dim});
       conv_states_val.copy_(conv_states_copy);
@@ -651,14 +738,14 @@ at::Tensor causal_conv1d_fwd_cpu(
 
 // API aligned with GPUs
 //
-//   x: (batch, dim) or (batch, dim, seqlen)
+//   x: (batch, dim) or (batch, seqlen, dim)
 //   conv_state: (..., dim, state_len), where state_len >= width - 1
 //   weight: (dim, width)
 //   bias: (dim,)
-//   cache_seqlens: (batch,), dtype int32.
+//   num_accepted_tokens: (batch,), dtype int32.
 //   conv_state_indices: (batch,), dtype int32
 //   pad_slot_id: int
-//   out: (batch, dim) or (batch, dim, seqlen)
+//   out: (batch, dim) or (batch, seqlen, dim)
 //
 at::Tensor causal_conv1d_update_cpu(
     const at::Tensor& x,
@@ -666,7 +753,7 @@ at::Tensor causal_conv1d_update_cpu(
     const at::Tensor& weight,
     const std::optional<at::Tensor>& bias,
     bool silu_activation,
-    const std::optional<at::Tensor>& cache_seqlens,
+    const std::optional<at::Tensor>& num_accepted_tokens,
     const std::optional<at::Tensor>& conv_state_indices,
     int64_t pad_slot_id,
     bool is_vnni) {
@@ -674,13 +761,13 @@ at::Tensor causal_conv1d_update_cpu(
   CHECK_CONTIGUOUS(weight);
   auto packed_w = is_vnni ? weight : causal_conv1d_weight_pack(weight);
 
-  // TODO: add multi-token prediction support
-  TORCH_CHECK(x.dim() == 2, "causal_conv1d_update_cpu: expect x to be 2D tensor.");
-  TORCH_CHECK(!cache_seqlens.has_value(), "causal_conv1d_update_cpu: don't support cache_seqlens.");
+  TORCH_CHECK(
+      x.dim() == 2 || x.dim() == 3,
+      "causal_conv1d_update_cpu: expect x to be 2D or 3D tensor.");
 
   int64_t batch = x.size(0);
-  int64_t dim = x.size(1);
-  int64_t seqlen = 1;
+  int64_t dim = x.dim() == 2 ? x.size(1) : x.size(2);
+  int64_t seqlen = x.dim() == 2 ? 1 : x.size(1);
   int64_t width = weight.size(-1);
 
   const auto scalar_type = x.scalar_type();
@@ -690,10 +777,84 @@ at::Tensor causal_conv1d_update_cpu(
 
   CHECK_EQ(conv_states.scalar_type(), scalar_type);
   CHECK_EQ(conv_states.size(1), dim);
-  CHECK_EQ(conv_states.size(2), width - 1);
+  const int64_t state_len = conv_states.size(2);
+  CHECK_GE(state_len, width - 1);
+
+  if (x.dim() == 3) {
+    TORCH_CHECK(
+        num_accepted_tokens.has_value(),
+        "causal_conv1d_update_cpu: num_accepted_tokens is required for 3D x.");
+    TORCH_CHECK(
+        conv_state_indices.has_value(),
+        "causal_conv1d_update_cpu: conv_state_indices is required for 3D x.");
+    CHECK_OPTIONAL_SHAPE_DTYPE(num_accepted_tokens, batch, at::kInt);
+    TORCH_CHECK(
+        width == 4,
+        "causal_conv1d_update_cpu: support only width of 4 for 3D x.");
+    TORCH_CHECK(
+        seqlen > 0,
+        "causal_conv1d_update_cpu: expect non-empty sequence for 3D x.");
+    TORCH_CHECK(
+        state_len >= seqlen,
+        "causal_conv1d_update_cpu: state_len must be >= seqlen for 3D x.");
+    TORCH_CHECK(
+        conv_states.stride(-2) == 1 && conv_states.stride(-1) == dim,
+        "causal_conv1d_update_cpu: 3D x requires SD conv_states layout.");
+
+    const int32_t* accepted_counts =
+        num_accepted_tokens.value().data_ptr<int32_t>();
+    const int32_t* indices = conv_state_indices.value().data_ptr<int32_t>();
+    const int64_t num_slots = conv_states.size(0);
+    for (int64_t bs = 0; bs < batch; ++bs) {
+      const int32_t num_accepted = accepted_counts[bs];
+      const int32_t conv_state_index = indices[bs];
+      TORCH_CHECK(
+          conv_state_index != pad_slot_id,
+          "causal_conv1d_update_cpu: 3D x does not support pad slots.");
+      TORCH_CHECK(
+          conv_state_index >= 0 && conv_state_index < num_slots,
+          "causal_conv1d_update_cpu: conv_state_indices out of range.");
+      TORCH_CHECK(
+          num_accepted >= 1 && num_accepted <= seqlen,
+          "causal_conv1d_update_cpu: num_accepted_tokens must be in [1, "
+          "seqlen].");
+      TORCH_CHECK(
+          num_accepted - 1 + width - 1 <= state_len,
+          "causal_conv1d_update_cpu: history window exceeds conv_states.");
+    }
+
+    int64_t conv_state_slot_stride = conv_states.stride(0);
+    at::Tensor out = at::empty_like(x);
+    AT_DISPATCH_REDUCED_FLOATING_TYPES(
+        scalar_type, "causal_conv1d_update_multi_kernel_impl", [&] {
+          causal_conv1d_update_multi_kernel_impl<scalar_t>(
+              out.data_ptr<scalar_t>(),
+              x.data_ptr<scalar_t>(),
+              conv_states.data_ptr<scalar_t>(),
+              packed_w.data_ptr<scalar_t>(),
+              conditional_data_ptr<scalar_t>(bias),
+              accepted_counts,
+              indices,
+              silu_activation,
+              batch,
+              dim,
+              seqlen,
+              width,
+              state_len,
+              conv_state_slot_stride);
+        });
+    return out;
+  }
+
+  TORCH_CHECK(
+      !num_accepted_tokens.has_value(),
+      "causal_conv1d_update_cpu: num_accepted_tokens is only supported for 3D "
+      "x.");
 
   // adjust `conv_states` to be contiguous on `dim`
   if (conv_states.stride(-2) != 1) {
+    TORCH_CHECK(state_len == width - 1,
+        "causal_conv1d_update_cpu: wide conv_states must be contiguous on dim.");
     int64_t num_cache_lines = conv_states.size(0);
     auto conv_states_copy = conv_states.clone();
     conv_states.as_strided_({num_cache_lines, dim, width - 1}, {(width - 1) * dim, 1, dim});

csrc/cpu/sgl-kernels/fla.cpp

@@ -1116,6 +1116,164 @@ void fused_sigmoid_gating_delta_rule_update_kernel_impl(
   });
 }
 
+// Speculative-decode variant: processes a varlen batch where each sequence has
+// ``q_len`` draft tokens, runs the recurrence sequentially over those tokens
+// (inside the kernel, so one dispatch handles the whole draft block), reads the
+// initial state from cache slot ``num_accepted-1`` and stores the state *after*
+// token ``t`` into cache slot ``t`` (multi-slot rollback, matching the GPU
+// kernel). Parallelized over (sequence, v_head); the per-sequence token loop is
+// sequential as required by the recurrence.
+template <typename scalar_t, typename param_t>
+void fused_sigmoid_gating_delta_rule_update_spec_kernel_impl(
+    const scalar_t* __restrict__ q_ptr,  // [T, HK, EK]
+    const scalar_t* __restrict__ k_ptr,  // [T, HK, EK]
+    const scalar_t* __restrict__ v_ptr,  // [T, HV, EV]
+    const param_t* __restrict__ A_log_ptr,
+    const scalar_t* __restrict__ a_ptr,  // [T, HV]
+    const scalar_t* __restrict__ dt_bias_ptr,
+    const scalar_t* __restrict__ b_ptr,  // [T, HV]
+    const int32_t* __restrict__ spec_indices_ptr,  // [N, S]
+    const int32_t* __restrict__ num_accepted_ptr,  // [N]
+    const int32_t* __restrict__ cu_seqlens_ptr,     // [N + 1]
+    float* __restrict__ state_ptr,
+    scalar_t* __restrict__ o_ptr,  // [T, HV, EV]
+    float* __restrict__ qk_scale_buf,  // [2, T, HK]
+    int64_t total_tokens,
+    int64_t batch_size,
+    int64_t spec_stride,
+    int64_t num_heads,
+    int64_t head_dim,
+    int64_t v_num_heads,
+    int64_t v_head_dim,
+    int64_t q_strideT,
+    int64_t q_strideH,
+    int64_t k_strideT,
+    int64_t k_strideH,
+    int64_t v_strideT,
+    int64_t v_strideH,
+    int64_t state_slot_stride,
+    bool use_qk_l2norm_in_kernel,
+    double softplus_threshold) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int64_t VecSize = bVec::size();
+  constexpr int64_t fVecSize = fVec::size();
+  int64_t group_size = v_num_heads / num_heads;
+  double scale = 1 / std::sqrt((double)head_dim);
+  fVec scale_vec = fVec((float)scale);
+
+  if (use_qk_l2norm_in_kernel) {
+    float eps = 1e-5f;
+    at::parallel_for(0, total_tokens * num_heads, 0, [&](int64_t begin, int64_t end) {
+      for (int64_t i = begin; i < end; ++i) {
+        int64_t ti = i / num_heads;
+        int64_t ni = i % num_heads;
+        const scalar_t* qp = q_ptr + ti * q_strideT + ni * q_strideH;
+        const scalar_t* kp = k_ptr + ti * k_strideT + ni * k_strideH;
+        float sq = 0.f, sk = 0.f;
+        for (int64_t d = 0; d < head_dim; ++d) {
+          float qv = (float)qp[d];
+          sq += qv * qv;
+          float kv = (float)kp[d];
+          sk += kv * kv;
+        }
+        qk_scale_buf[ti * num_heads + ni] = 1.f / std::sqrt(sq + eps);
+        qk_scale_buf[total_tokens * num_heads + ti * num_heads + ni] = 1.f / std::sqrt(sk + eps);
+      }
+    });
+  }
+
+  at::parallel_for(0, batch_size * v_num_heads, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t idx = begin; idx < end; ++idx) {
+      int64_t bi = idx / v_num_heads;
+      int64_t ni = idx % v_num_heads;
+      int64_t kh = ni / group_size;
+      int64_t q_start = cu_seqlens_ptr[bi];
+      int64_t q_len = cu_seqlens_ptr[bi + 1] - q_start;
+      if (q_len <= 0) {
+        continue;
+      }
+      int64_t acc = (int64_t)num_accepted_ptr[bi];
+      // Clamp acc-1 to >=0: when num_accepted is 0 the unclamped index reads
+      // out of bounds and yields an arbitrary prev_slot used to index the SSM
+      // state. Mirrors the GPU guard tl.maximum(num_accepted - 1, 0).
+      int64_t prev_slot =
+          (int64_t)spec_indices_ptr[bi * spec_stride + (acc > 0 ? acc - 1 : 0)];
+      for (int64_t t = 0; t < q_len; ++t) {
+        int64_t cur_slot = (int64_t)spec_indices_ptr[bi * spec_stride + t];
+        int64_t token = q_start + t;
+        const float* src = state_ptr + prev_slot * state_slot_stride + ni * head_dim * v_head_dim;
+        float* dst = state_ptr + cur_slot * state_slot_stride + ni * head_dim * v_head_dim;
+        float g_val = -std::exp((float)A_log_ptr[ni]) *
+            softplus((float)a_ptr[token * v_num_heads + ni] + (float)dt_bias_ptr[ni], softplus_threshold);
+        float g_val_exp = std::exp(g_val);
+        fVec g_val_exp_vec = fVec(g_val_exp);
+        float beta_val = 1.f / (1.f + std::exp(-(float)b_ptr[token * v_num_heads + ni]));
+        fVec beta_vec = fVec(beta_val);
+        int64_t q_offset = token * q_strideT + kh * q_strideH;
+        int64_t k_offset = token * k_strideT + kh * k_strideH;
+        float q_scale = use_qk_l2norm_in_kernel ? qk_scale_buf[token * num_heads + kh] : 1.f;
+        float k_scale =
+            use_qk_l2norm_in_kernel ? qk_scale_buf[total_tokens * num_heads + token * num_heads + kh] : 1.f;
+        int64_t v_offset = token * v_strideT + ni * v_strideH;
+        int64_t o_offset = (token * v_num_heads + ni) * v_head_dim;
+        int64_t dvi = 0;
+        for (; dvi <= v_head_dim - VecSize; dvi += VecSize) {
+          fVec kv_mem_vec0 = fVec(0.f);
+          fVec kv_mem_vec1 = fVec(0.f);
+          for (int di = 0; di < head_dim; ++di) {
+            fVec k_val_vec = fVec((float)k_ptr[k_offset + di] * k_scale);
+            fVec sv0 = fVec::loadu(src + di * v_head_dim + dvi);
+            fVec sv1 = fVec::loadu(src + di * v_head_dim + dvi + fVecSize);
+            kv_mem_vec0 = kv_mem_vec0 + sv0 * g_val_exp_vec * k_val_vec;
+            kv_mem_vec1 = kv_mem_vec1 + sv1 * g_val_exp_vec * k_val_vec;
+          }
+          bVec v_bvec = bVec::loadu(v_ptr + v_offset + dvi);
+          fVec v_vec0, v_vec1;
+          std::tie(v_vec0, v_vec1) = at::vec::convert_to_float(v_bvec);
+          fVec dt_vec0 = (v_vec0 - kv_mem_vec0) * beta_vec;
+          fVec dt_vec1 = (v_vec1 - kv_mem_vec1) * beta_vec;
+          fVec o_vec0 = fVec(0.f);
+          fVec o_vec1 = fVec(0.f);
+          for (int di = 0; di < head_dim; ++di) {
+            fVec q_vec = fVec((float)q_ptr[q_offset + di] * q_scale);
+            fVec k_vec = fVec((float)k_ptr[k_offset + di] * k_scale);
+            fVec sv0 = fVec::loadu(src + di * v_head_dim + dvi);
+            fVec sv1 = fVec::loadu(src + di * v_head_dim + dvi + fVecSize);
+            sv0 = sv0 * g_val_exp_vec + k_vec * dt_vec0;
+            sv1 = sv1 * g_val_exp_vec + k_vec * dt_vec1;
+            o_vec0 = o_vec0 + sv0 * q_vec * scale_vec;
+            o_vec1 = o_vec1 + sv1 * q_vec * scale_vec;
+            sv0.store(dst + di * v_head_dim + dvi);
+            sv1.store(dst + di * v_head_dim + dvi + fVecSize);
+          }
+          bVec o_vec = at::vec::convert_from_float<scalar_t>(o_vec0, o_vec1);
+          o_vec.store(o_ptr + o_offset + dvi);
+        }
+        for (; dvi < v_head_dim; ++dvi) {
+          float kv_mem_val = 0.f;
+          for (int di = 0; di < head_dim; ++di) {
+            float k_val = (float)k_ptr[k_offset + di] * k_scale;
+            kv_mem_val += src[di * v_head_dim + dvi] * g_val_exp * k_val;
+          }
+          float v_val = (float)v_ptr[v_offset + dvi];
+          float dt_val = (v_val - kv_mem_val) * beta_val;
+          float o_val = 0.f;
+          for (int di = 0; di < head_dim; ++di) {
+            float q_val = (float)q_ptr[q_offset + di] * q_scale;
+            float k_val = (float)k_ptr[k_offset + di] * k_scale;
+            float ns = src[di * v_head_dim + dvi] * g_val_exp + k_val * dt_val;
+            dst[di * v_head_dim + dvi] = ns;
+            o_val += ns * q_val * scale;
+          }
+          o_ptr[o_offset + dvi] = (scalar_t)o_val;
+        }
+        prev_slot = cur_slot;
+      }
+    }
+  });
+}
+
 template <typename scalar_t>
 void fused_gdn_gating_kernel_impl(
     float* __restrict__ A_log,
@@ -1500,6 +1658,103 @@ at::Tensor fused_sigmoid_gating_delta_rule_update_cpu(
   return core_attn_out;
 }
 
+// Speculative-decode update (multi-token, multi-slot rollback).
+// q: [T, HK, EK]  k: [T, HK, EK]  v: [T, HV, EV]
+// a: [T, HV]  b: [T, HV]
+// initial_state_source: [N_slots, HV, EK, EV] FP32 (updated in place)
+// spec_state_indices: [batch, S] INT32 (S = num_spec + 1)
+// num_accepted_tokens: [batch] INT32
+// cu_seqlens: [batch + 1] INT32
+// Returns output: [T, HV, EV]
+at::Tensor fused_sigmoid_gating_delta_rule_update_spec_cpu(
+    const at::Tensor& A_log,
+    const at::Tensor& dt_bias,
+    const at::Tensor& q,
+    const at::Tensor& k,
+    const at::Tensor& v,
+    const at::Tensor& a,
+    const at::Tensor& b,
+    at::Tensor& initial_state_source,
+    const at::Tensor& spec_state_indices,
+    const at::Tensor& num_accepted_tokens,
+    const at::Tensor& cu_seqlens,
+    bool use_qk_l2norm_in_kernel,
+    double softplus_beta = 1.0,
+    double softplus_threshold = 20.0) {
+  CHECK_DIM(3, q);
+  CHECK_DIM(3, v);
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(q);
+  int64_t total_tokens = q.size(0);
+  int64_t num_heads = q.size(1);
+  int64_t head_dim = q.size(2);
+  int64_t v_num_heads = v.size(1);
+  int64_t v_head_dim = v.size(2);
+  int64_t batch_size = cu_seqlens.size(0) - 1;
+  int64_t spec_stride = spec_state_indices.stride(0);
+  CHECK_INPUT_SHAPE_DTYPE<true>(k, {total_tokens, num_heads, head_dim}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(v, {total_tokens, v_num_heads, v_head_dim}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(a, {total_tokens, v_num_heads}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(b, {total_tokens, v_num_heads}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(dt_bias, {v_num_heads}, q.scalar_type());
+  CHECK_INPUT_SHAPE_DTYPE<true>(num_accepted_tokens, {batch_size}, at::kInt);
+  CHECK_INPUT_SHAPE_DTYPE<true>(cu_seqlens, {batch_size + 1}, at::kInt);
+  CHECK_EQ(v_num_heads % num_heads, 0);
+  TORCH_CHECK(A_log.sizes() == at::IntArrayRef({v_num_heads}));
+  CHECK_INPUT_SHAPE_DTYPE<true>(
+      initial_state_source,
+      {initial_state_source.size(0), v_num_heads, head_dim, v_head_dim},
+      at::kFloat);
+  TORCH_CHECK(initial_state_source.size(0) >= batch_size,
+      "initial_state_source capacity too small: size(0)=",
+      initial_state_source.size(0), ", batch_size=", batch_size);
+
+  int64_t q_strideT = q.stride(0);
+  int64_t q_strideH = q.stride(1);
+  int64_t k_strideT = k.stride(0);
+  int64_t k_strideH = k.stride(1);
+  int64_t v_strideT = v.stride(0);
+  int64_t v_strideH = v.stride(1);
+  int64_t state_slot_stride = initial_state_source.stride(0);
+
+  at::Tensor o = at::empty({total_tokens, v_num_heads, v_head_dim}, q.options());
+  at::Tensor qk_scale_buf = at::empty({2, total_tokens, num_heads}, at::kFloat);
+
+  CPU_DISPATCH_REDUCED_FLOATING_TYPES_EXT(
+      q.scalar_type(), A_log.scalar_type(), "fused_sigmoid_gating_delta_rule_update_spec_kernel_impl", [&] {
+        fused_sigmoid_gating_delta_rule_update_spec_kernel_impl<scalar_t, param_t>(
+            q.data_ptr<scalar_t>(),
+            k.data_ptr<scalar_t>(),
+            v.data_ptr<scalar_t>(),
+            A_log.data_ptr<param_t>(),
+            a.data_ptr<scalar_t>(),
+            dt_bias.data_ptr<scalar_t>(),
+            b.data_ptr<scalar_t>(),
+            spec_state_indices.data_ptr<int32_t>(),
+            num_accepted_tokens.data_ptr<int32_t>(),
+            cu_seqlens.data_ptr<int32_t>(),
+            initial_state_source.data_ptr<float>(),
+            o.data_ptr<scalar_t>(),
+            qk_scale_buf.data_ptr<float>(),
+            total_tokens,
+            batch_size,
+            spec_stride,
+            num_heads,
+            head_dim,
+            v_num_heads,
+            v_head_dim,
+            q_strideT,
+            q_strideH,
+            k_strideT,
+            k_strideH,
+            v_strideT,
+            v_strideH,
+            state_slot_stride,
+            use_qk_l2norm_in_kernel,
+            softplus_threshold);
+      });
+  return o;
+}
+
 // A_log: [num_v_heads]
 // a: [batch, num_v_heads]
 // b: [batch, num_v_heads]

csrc/cpu/torch_bindings.cpp

@@ -120,6 +120,14 @@ at::Tensor fused_sigmoid_gating_delta_rule_update_cpu(
     bool use_qk_l2norm_in_kernel, double softplus_beta = 1.0,
     double softplus_threshold = 20.0);
 
+at::Tensor fused_sigmoid_gating_delta_rule_update_spec_cpu(
+    const at::Tensor& A_log, const at::Tensor& dt_bias, const at::Tensor& q,
+    const at::Tensor& k, const at::Tensor& v, const at::Tensor& a,
+    const at::Tensor& b, at::Tensor& initial_state_source,
+    const at::Tensor& spec_state_indices, const at::Tensor& num_accepted_tokens,
+    const at::Tensor& cu_seqlens, bool use_qk_l2norm_in_kernel,
+    double softplus_beta = 1.0, double softplus_threshold = 20.0);
+
 std::tuple<at::Tensor, at::Tensor> fused_gdn_gating_cpu(
     const at::Tensor& A_log, const at::Tensor& a, const at::Tensor& b,
     const at::Tensor& dt_bias);
@@ -139,7 +147,7 @@ at::Tensor causal_conv1d_fwd_cpu(
 at::Tensor causal_conv1d_update_cpu(
     const at::Tensor& x, const at::Tensor& conv_states,
     const at::Tensor& weight, const std::optional<at::Tensor>& bias,
-    bool silu_activation, const std::optional<at::Tensor>& cache_seqlens,
+    bool silu_activation, const std::optional<at::Tensor>& num_accepted_tokens,
     const std::optional<at::Tensor>& conv_state_indices, int64_t pad_slot_id,
     bool is_vnni);
 
@@ -199,12 +207,52 @@ void cpu_fused_moe(torch::Tensor& output, const torch::Tensor& input,
                    const torch::Tensor& topk_id, const bool skip_weighted,
                    const std::string& act, const std::string& isa);
 
+void prepack_moe_weight_int8(const torch::Tensor& weight,
+                             torch::Tensor& packed_weight,
+                             const std::string& isa);
+
+void cpu_fused_moe_int8(torch::Tensor& output, const torch::Tensor& input,
+                        const torch::Tensor& w13, const torch::Tensor& w2,
+                        const torch::Tensor& w13_scale,
+                        const torch::Tensor& w2_scale,
+                        const std::optional<torch::Tensor>& w13_bias,
+                        const std::optional<torch::Tensor>& w2_bias,
+                        const torch::Tensor& topk_weights,
+                        const torch::Tensor& topk_id, const bool skip_weighted,
+                        const std::string& act, const std::string& isa);
+
 void compute_slot_mapping_kernel_impl(const torch::Tensor query_start_loc,
                                       const torch::Tensor positions,
                                       const torch::Tensor block_table,
                                       torch::Tensor slot_mapping,
                                       const int64_t block_size);
 
+at::Tensor causal_conv1d_update_cpu_impl(
+    at::Tensor& x, at::Tensor& conv_state, const at::Tensor& weight,
+    const c10::optional<at::Tensor>& bias,
+    const c10::optional<std::string>& activation,
+    const c10::optional<at::Tensor>& conv_state_indices,
+    const c10::optional<at::Tensor>& query_start_loc, int64_t pad_slot_id);
+
+void selective_state_update_cpu_impl(
+    at::Tensor& state, const at::Tensor& x, const at::Tensor& dt,
+    const at::Tensor& A, const at::Tensor& B, const at::Tensor& C,
+    const c10::optional<at::Tensor>& D, const c10::optional<at::Tensor>& z,
+    const c10::optional<at::Tensor>& dt_bias, bool dt_softplus,
+    const c10::optional<at::Tensor>& state_batch_indices,
+    const c10::optional<at::Tensor>& dst_state_batch_indices,
+    int64_t null_block_id, at::Tensor& out,
+    const c10::optional<at::Tensor>& num_accepted_tokens,
+    const c10::optional<at::Tensor>& cu_seqlens);
+
+void mamba_chunk_scan_fwd_cpu_impl(at::Tensor& out, at::Tensor& final_states,
+                                   const at::Tensor& x, const at::Tensor& dt,
+                                   const at::Tensor& A, const at::Tensor& B,
+                                   const at::Tensor& C,
+                                   const c10::optional<at::Tensor>& D,
+                                   const c10::optional<at::Tensor>& z,
+                                   const at::Tensor& cu_seqlens);
+
 void init_cpu_memory_env(std::vector<int64_t> node_ids);
 
 namespace cpu_utils {
@@ -468,7 +516,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def(
       "causal_conv1d_update_cpu(Tensor x, Tensor(a!) conv_states, Tensor "
       "weight, Tensor? bias, bool silu_activation,"
-      "Tensor? cache_seqlens, Tensor? conv_state_indices, int pad_slot_id, "
+      "Tensor? num_accepted_tokens, Tensor? conv_state_indices, int "
+      "pad_slot_id, "
       "bool is_vnni) -> Tensor");
   ops.impl("causal_conv1d_update_cpu", torch::kCPU, &causal_conv1d_update_cpu);
 #endif
@@ -508,6 +557,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "softplus_threshold=20.0) -> Tensor");
   ops.impl("fused_sigmoid_gating_delta_rule_update_cpu", torch::kCPU,
            &fused_sigmoid_gating_delta_rule_update_cpu);
+  ops.def(
+      "fused_sigmoid_gating_delta_rule_update_spec_cpu(Tensor A_log, Tensor "
+      "dt_bias, Tensor q, Tensor k, Tensor v, Tensor a, Tensor b, "
+      "Tensor(a!) initial_state_source, Tensor spec_state_indices, "
+      "Tensor num_accepted_tokens, Tensor cu_seqlens, bool "
+      "use_qk_l2norm_in_kernel, float softplus_beta=1.0, float "
+      "softplus_threshold=20.0) -> Tensor");
+  ops.impl("fused_sigmoid_gating_delta_rule_update_spec_cpu", torch::kCPU,
+           &fused_sigmoid_gating_delta_rule_update_spec_cpu);
   ops.def(
       "fused_gdn_gating_cpu(Tensor A_log, Tensor a, Tensor b, Tensor dt_bias) "
       "-> (Tensor, Tensor)");
@@ -553,7 +611,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 #endif
 
   // fused moe
-#if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT))
+#if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT) && !defined(__APPLE__))
   ops.def(
       "prepack_moe_weight(Tensor weight, Tensor(a1!) packed_weight, str isa) "
       "-> ()");
@@ -564,7 +622,22 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "bool skip_weighted, "
       "str act, str isa) -> ()");
   ops.impl("cpu_fused_moe", torch::kCPU, &cpu_fused_moe);
-#endif  // #if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT))
+#endif  // #if defined(__AVX512F__) || (defined(ARM_BF16_SUPPORT) &&
+        // !defined(__APPLE__))
+#if defined(ARM_I8MM_SUPPORT) && defined(ARM_BF16_SUPPORT) && \
+    !defined(__APPLE__)
+  ops.def(
+      "prepack_moe_weight_int8(Tensor weight, Tensor(a1!) packed_weight, "
+      "str isa) -> ()");
+  ops.impl("prepack_moe_weight_int8", torch::kCPU, &prepack_moe_weight_int8);
+  ops.def(
+      "cpu_fused_moe_int8(Tensor(a0!) output, Tensor input, Tensor w13, "
+      "Tensor w2, Tensor w13_scale, Tensor w2_scale, Tensor? w13_bias, "
+      "Tensor? w2_bias, Tensor topk_weights, Tensor topk_id, bool "
+      "skip_weighted, str act, str isa) -> ()");
+  ops.impl("cpu_fused_moe_int8", torch::kCPU, &cpu_fused_moe_int8);
+#endif  // #if defined(ARM_I8MM_SUPPORT) && defined(ARM_BF16_SUPPORT) &&
+        // !defined(__APPLE__)
   ops.def(
       "mla_decode_kvcache("
       "   Tensor! out, Tensor query, Tensor kv_cache,"
@@ -577,6 +650,30 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "block_size) -> ()",
       &compute_slot_mapping_kernel_impl);
 
+  // Mamba CPU kernels
+  ops.def(
+      "causal_conv1d_update_cpu_vec("
+      "Tensor(a0!) x, Tensor(a1!) conv_state, Tensor weight, "
+      "Tensor? bias, str? activation, Tensor? conv_state_indices, "
+      "Tensor? query_start_loc, SymInt pad_slot_id) -> Tensor",
+      &causal_conv1d_update_cpu_impl);
+
+  ops.def(
+      "selective_state_update_cpu("
+      "Tensor(a0!) state, Tensor x, Tensor dt, Tensor A, Tensor B, Tensor C, "
+      "Tensor? D, Tensor? z, Tensor? dt_bias, bool dt_softplus, "
+      "Tensor? state_batch_indices, Tensor? dst_state_batch_indices, "
+      "SymInt null_block_id, Tensor(a13!) out, "
+      "Tensor? num_accepted_tokens, Tensor? cu_seqlens) -> ()",
+      &selective_state_update_cpu_impl);
+
+  ops.def(
+      "mamba_chunk_scan_fwd_cpu("
+      "Tensor(a0!) out, Tensor(a1!) final_states, "
+      "Tensor x, Tensor dt, Tensor A, Tensor B, Tensor C, "
+      "Tensor? D, Tensor? z, Tensor cu_seqlens) -> ()",
+      &mamba_chunk_scan_fwd_cpu_impl);
+
   ops.def("init_cpu_memory_env(SymInt[] node_ids) -> ()", &init_cpu_memory_env);
 
   // Speculative decoding kernels

csrc/libtorch_stable/cache_kernels.cu

@@ -1025,6 +1025,9 @@ __global__ void gather_and_maybe_dequant_cache(
     batch_offset += offset;
     int32_t block_table_id = batch_offset / block_size;
     int32_t slot_id = batch_offset % block_size;
+    // seq_starts may push the block index past the end of the batch's block
+    // table row.
+    if (block_table_id >= block_table_stride) continue;
     int32_t block_table_offset = batch_id * block_table_stride + block_table_id;
     int32_t block_id = block_table[block_table_offset];
     int64_t cache_offset =
@@ -1174,7 +1177,8 @@ __global__ void cp_gather_and_upconvert_fp8_kv_cache(
     const int32_t num_reqs, const int32_t block_size,
     const int32_t total_tokens, const int64_t block_table_stride,
     const int64_t cache_block_stride, const int64_t cache_entry_stride,
-    const int64_t dst_entry_stride) {
+    const int64_t dst_entry_stride,
+    const int32_t* __restrict__ seq_starts) {  // Optional source offsets
   const int flat_warp_id = (blockIdx.x * blockDim.x + threadIdx.x) >> 5;
   if (flat_warp_id >= total_tokens) return;
   const int lane_id = threadIdx.x & 31;
@@ -1192,7 +1196,8 @@ __global__ void cp_gather_and_upconvert_fp8_kv_cache(
 
   // Compute physical token address via block table
   const int out_token_id = flat_warp_id;
-  const int token_offset = out_token_id - workspace_starts[req_id];
+  int token_offset = out_token_id - workspace_starts[req_id];
+  if (seq_starts != nullptr) token_offset += seq_starts[req_id];
   const int cache_block_idx = token_offset / block_size;
   const int offset_in_block = token_offset % block_size;
   const int physical_block =
@@ -1383,9 +1388,9 @@ void cp_gather_and_upconvert_fp8_kv_cache(
     torch::stable::Tensor const& src_cache,    // [NUM_BLOCKS, BLOCK_SIZE, 656]
     torch::stable::Tensor const& dst,          // [TOT_TOKENS, 576]
     torch::stable::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
-    torch::stable::Tensor const& seq_lens,     // [BATCH]
     torch::stable::Tensor const& workspace_starts,  // [BATCH]
-    int64_t batch_size) {
+    int64_t batch_size,
+    std::optional<torch::stable::Tensor> seq_starts = std::nullopt) {
   torch::stable::accelerator::DeviceGuard device_guard(
       src_cache.get_device_index());
   const cudaStream_t stream = get_current_cuda_stream();
@@ -1396,20 +1401,25 @@ void cp_gather_and_upconvert_fp8_kv_cache(
   STD_TORCH_CHECK(
       block_table.scalar_type() == torch::headeronly::ScalarType::Int,
       "block_table must be int32");
-  STD_TORCH_CHECK(seq_lens.scalar_type() == torch::headeronly::ScalarType::Int,
-                  "seq_lens must be int32");
   STD_TORCH_CHECK(
       workspace_starts.scalar_type() == torch::headeronly::ScalarType::Int,
       "workspace_starts must be int32");
+  if (seq_starts.has_value()) {
+    STD_TORCH_CHECK(
+        seq_starts.value().scalar_type() == torch::headeronly::ScalarType::Int,
+        "seq_starts must be int32");
+  }
 
   STD_TORCH_CHECK(src_cache.device() == dst.device(),
                   "src_cache and dst must be on the same device");
   STD_TORCH_CHECK(src_cache.device() == block_table.device(),
                   "src_cache and block_table must be on the same device");
-  STD_TORCH_CHECK(src_cache.device() == seq_lens.device(),
-                  "src_cache and seq_lens must be on the same device");
   STD_TORCH_CHECK(src_cache.device() == workspace_starts.device(),
                   "src_cache and workspace_starts must be on the same device");
+  if (seq_starts.has_value()) {
+    STD_TORCH_CHECK(src_cache.device() == seq_starts.value().device(),
+                    "src_cache and seq_starts must be on the same device");
+  }
   auto dtype = src_cache.scalar_type();
   STD_TORCH_CHECK(
       dtype == torch::headeronly::ScalarType::Byte ||               // uint8
@@ -1438,6 +1448,9 @@ void cp_gather_and_upconvert_fp8_kv_cache(
   constexpr int warps_per_block = 8;
   const int grid_size = (total_tokens + warps_per_block - 1) / warps_per_block;
   const int block_size_threads = warps_per_block * 32;  // 256 threads
+  const int32_t* seq_starts_ptr =
+      seq_starts.has_value() ? seq_starts.value().const_data_ptr<int32_t>()
+                             : nullptr;
 
   vllm::cp_gather_and_upconvert_fp8_kv_cache<<<grid_size, block_size_threads, 0,
                                                stream>>>(
@@ -1446,7 +1459,7 @@ void cp_gather_and_upconvert_fp8_kv_cache(
       workspace_starts.const_data_ptr<int32_t>(),
       static_cast<int32_t>(batch_size), block_size, total_tokens,
       block_table_stride, cache_block_stride, cache_entry_stride,
-      dst_entry_stride);
+      dst_entry_stride, seq_starts_ptr);
 }
 
 // Macro to dispatch the kernel based on the data type.

csrc/libtorch_stable/dispatch_utils.h

@@ -21,6 +21,21 @@
 #define VLLM_STABLE_DISPATCH_FP8_CASE(enum_type, ...) \
   THO_PRIVATE_CASE_TYPE_USING_HINT(enum_type, fp8_t, __VA_ARGS__)
 
+// Same idea, for dispatching on an int32/int64 index tensor (e.g. topk_ids)
+// nested inside a value-type dispatch. Named 'idx_t' instead of 'scalar_t'.
+#define VLLM_STABLE_DISPATCH_IDX_CASE(enum_type, ...) \
+  THO_PRIVATE_CASE_TYPE_USING_HINT(enum_type, idx_t, __VA_ARGS__)
+
+#define VLLM_STABLE_DISPATCH_CASE_IDX_TYPES(...)                     \
+  VLLM_STABLE_DISPATCH_IDX_CASE(torch::headeronly::ScalarType::Int,  \
+                                __VA_ARGS__)                         \
+  VLLM_STABLE_DISPATCH_IDX_CASE(torch::headeronly::ScalarType::Long, \
+                                __VA_ARGS__)
+
+#define VLLM_STABLE_DISPATCH_IDX_TYPES(TYPE, NAME, ...) \
+  THO_DISPATCH_SWITCH(TYPE, NAME,                       \
+                      VLLM_STABLE_DISPATCH_CASE_IDX_TYPES(__VA_ARGS__))
+
 #define VLLM_STABLE_DISPATCH_CASE_FLOATING_TYPES(...)                  \
   THO_DISPATCH_CASE(torch::headeronly::ScalarType::Float, __VA_ARGS__) \
   THO_DISPATCH_CASE(torch::headeronly::ScalarType::Half, __VA_ARGS__)  \

csrc/libtorch_stable/moe/moe_align_sum_kernels.cu

@@ -360,12 +360,24 @@ __global__ void count_and_sort_expert_tokens_kernel(
 template <typename scalar_t>
 constexpr int MOE_SUM_VEC = 16 / sizeof(scalar_t);
 
-template <typename scalar_t, int TOPK>
+template <typename idx_t>
+__device__ __forceinline__ bool moe_sum_pad_aware_skip(
+    const idx_t* __restrict__ topk_ids, const int32_t* __restrict__ expert_map,
+    int64_t idx) {
+  int64_t expert_id = static_cast<int64_t>(topk_ids[idx]);
+  if (expert_id < 0) return true;
+  if (expert_map != nullptr && expert_map[expert_id] < 0) return true;
+  return false;
+}
+
+template <typename scalar_t, typename idx_t, int TOPK, bool PAD_AWARE>
 __global__ void moe_sum_vec_kernel(
     scalar_t* __restrict__ out,          // [num_tokens, d], contiguous
     const scalar_t* __restrict__ input,  // [num_tokens, topk, d], d contiguous
     const int64_t num_tokens, const int d, const int64_t stride_token,
-    const int64_t stride_topk) {
+    const int64_t stride_topk, const idx_t* __restrict__ topk_ids,
+    const int32_t* __restrict__ expert_map, const int64_t stride_tk_token,
+    const int64_t stride_tk_k) {
   using vec_t = vllm::vec_n_t<scalar_t, MOE_SUM_VEC<scalar_t>>;  // 16-byte pack
   constexpr int VEC = MOE_SUM_VEC<scalar_t>;
   const int64_t n_vec = d / VEC;
@@ -375,6 +387,10 @@ __global__ void moe_sum_vec_kernel(
     const int64_t token = i / n_vec;
     const int64_t v = i % n_vec;
     const scalar_t* in_tok = input + token * stride_token + v * VEC;
+    const idx_t* tk_tok = nullptr;
+    if constexpr (PAD_AWARE) {
+      tk_tok = topk_ids + token * stride_tk_token;
+    }
 
     float acc[VEC];
 #pragma unroll
@@ -382,6 +398,11 @@ __global__ void moe_sum_vec_kernel(
 
 #pragma unroll
     for (int k = 0; k < TOPK; ++k) {
+      if constexpr (PAD_AWARE) {
+        if (moe_sum_pad_aware_skip(tk_tok, expert_map, k * stride_tk_k)) {
+          continue;
+        }
+      }
       vec_t packed = *reinterpret_cast<const vec_t*>(in_tok + k * stride_topk);
 #pragma unroll
       for (int j = 0; j < VEC; ++j) acc[j] += static_cast<float>(packed.val[j]);
@@ -394,13 +415,16 @@ __global__ void moe_sum_vec_kernel(
   }
 }
 
-// Runtime-topk variant of the above.
-template <typename scalar_t>
+// Runtime-topk variant of the above, for topk values outside the templated
+// set.
+template <typename scalar_t, typename idx_t, bool PAD_AWARE>
 __global__ void moe_sum_vec_dynamic_kernel(
     scalar_t* __restrict__ out,          // [num_tokens, d], contiguous
     const scalar_t* __restrict__ input,  // [num_tokens, topk, d], d contiguous
     const int64_t num_tokens, const int d, const int topk,
-    const int64_t stride_token, const int64_t stride_topk) {
+    const int64_t stride_token, const int64_t stride_topk,
+    const idx_t* __restrict__ topk_ids, const int32_t* __restrict__ expert_map,
+    const int64_t stride_tk_token, const int64_t stride_tk_k) {
   using vec_t = vllm::vec_n_t<scalar_t, MOE_SUM_VEC<scalar_t>>;
   constexpr int VEC = MOE_SUM_VEC<scalar_t>;
   const int64_t n_vec = d / VEC;
@@ -410,12 +434,21 @@ __global__ void moe_sum_vec_dynamic_kernel(
     const int64_t token = i / n_vec;
     const int64_t v = i % n_vec;
     const scalar_t* in_tok = input + token * stride_token + v * VEC;
+    const idx_t* tk_tok = nullptr;
+    if constexpr (PAD_AWARE) {
+      tk_tok = topk_ids + token * stride_tk_token;
+    }
 
     float acc[VEC];
 #pragma unroll
     for (int j = 0; j < VEC; ++j) acc[j] = 0.f;
 
     for (int k = 0; k < topk; ++k) {
+      if constexpr (PAD_AWARE) {
+        if (moe_sum_pad_aware_skip(tk_tok, expert_map, k * stride_tk_k)) {
+          continue;
+        }
+      }
       vec_t packed = *reinterpret_cast<const vec_t*>(in_tok + k * stride_topk);
 #pragma unroll
       for (int j = 0; j < VEC; ++j) acc[j] += static_cast<float>(packed.val[j]);
@@ -430,17 +463,28 @@ __global__ void moe_sum_vec_dynamic_kernel(
 
 // Stride-aware scalar fallback: handles unaligned/non-vectorizable hidden dims
 // (including a non-contiguous hidden stride) via per-element strided reads.
-template <typename scalar_t>
+template <typename scalar_t, typename idx_t, bool PAD_AWARE>
 __global__ void moe_sum_scalar_kernel(
     scalar_t* __restrict__ out,          // [num_tokens, d], contiguous
     const scalar_t* __restrict__ input,  // [num_tokens, topk, d]
     const int d, const int topk, const int64_t stride_token,
-    const int64_t stride_topk, const int64_t stride_hidden) {
+    const int64_t stride_topk, const int64_t stride_hidden,
+    const idx_t* __restrict__ topk_ids, const int32_t* __restrict__ expert_map,
+    const int64_t stride_tk_token, const int64_t stride_tk_k) {
   const int64_t token_idx = blockIdx.x;
   const scalar_t* in_tok = input + token_idx * stride_token;
+  const idx_t* tk_tok = nullptr;
+  if constexpr (PAD_AWARE) {
+    tk_tok = topk_ids + token_idx * stride_tk_token;
+  }
   for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
     float x = 0.f;
     for (int k = 0; k < topk; ++k) {
+      if constexpr (PAD_AWARE) {
+        if (moe_sum_pad_aware_skip(tk_tok, expert_map, k * stride_tk_k)) {
+          continue;
+        }
+      }
       x += static_cast<float>(
           VLLM_LDG(&in_tok[k * stride_topk + idx * stride_hidden]));
     }
@@ -711,8 +755,9 @@ void batched_moe_align_block_size(int64_t max_tokens_per_batch,
 }
 
 void moe_sum(torch::stable::Tensor& input,   // [num_tokens, topk, hidden_size]
-             torch::stable::Tensor& output)  // [num_tokens, hidden_size]
-{
+             torch::stable::Tensor& output,  // [num_tokens, hidden_size]
+             std::optional<torch::stable::Tensor> topk_ids,
+             std::optional<torch::stable::Tensor> expert_map) {
   // Output is dense and written in place, so it must be contiguous. The input
   // is read by its strides (no copy); only the hidden dim needs to be
   // contiguous to take the vectorized path.
@@ -731,10 +776,102 @@ void moe_sum(torch::stable::Tensor& input,   // [num_tokens, topk, hidden_size]
   const cudaStream_t stream =
       get_current_cuda_stream(output.get_device_index());
 
-#define LAUNCH_MOE_SUM_VEC(TOPK)                \
-  vllm::moe::moe_sum_vec_kernel<scalar_t, TOPK> \
-      <<<grid, dim3(block), 0, stream>>>(       \
-          out_ptr, in_ptr, num_tokens, hidden_size, stride_token, stride_topk)
+  if (topk_ids.has_value()) {
+    // Pad-aware reduce path
+    const torch::stable::Tensor& tk = topk_ids.value();
+    STD_TORCH_CHECK(tk.size(0) == num_tokens && tk.size(1) == topk,
+                    "moe_sum: topk_ids must have shape [num_tokens, topk]");
+    const int64_t stride_tk_token = tk.stride(0);
+    const int64_t stride_tk_k = tk.stride(1);
+
+    const int32_t* expert_map_ptr = nullptr;
+    if (expert_map.has_value()) {
+      STD_TORCH_CHECK(
+          expert_map->scalar_type() == torch::headeronly::ScalarType::Int,
+          "moe_sum: expert_map must be int32");
+      expert_map_ptr =
+          reinterpret_cast<const int32_t*>(expert_map->const_data_ptr());
+    }
+
+#define LAUNCH_MOE_SUM_PAD_AWARE_VEC(TOPK)                                     \
+  vllm::moe::moe_sum_vec_kernel<scalar_t, idx_t, TOPK, true>                   \
+      <<<grid, dim3(block), 0, stream>>>(                                      \
+          out_ptr, in_ptr, num_tokens, hidden_size, stride_token, stride_topk, \
+          topk_ids_ptr, expert_map_ptr, stride_tk_token, stride_tk_k)
+
+    VLLM_STABLE_DISPATCH_FLOATING_TYPES(
+        input.scalar_type(), "moe_sum_pad_aware", [&] {
+          constexpr int VEC = vllm::moe::MOE_SUM_VEC<scalar_t>;
+          constexpr int WIDTH = VEC * sizeof(scalar_t);
+          auto* out_ptr =
+              reinterpret_cast<scalar_t*>(output.mutable_data_ptr());
+          auto* in_ptr =
+              reinterpret_cast<const scalar_t*>(input.const_data_ptr());
+
+          const bool can_vec =
+              (stride_hidden == 1) && (hidden_size % VEC == 0) &&
+              (stride_token % VEC == 0) && (stride_topk % VEC == 0) &&
+              (reinterpret_cast<uintptr_t>(in_ptr) % WIDTH == 0) &&
+              (reinterpret_cast<uintptr_t>(out_ptr) % WIDTH == 0);
+
+          VLLM_STABLE_DISPATCH_IDX_TYPES(
+              tk.scalar_type(), "moe_sum_pad_aware_idx", [&] {
+                auto* topk_ids_ptr =
+                    reinterpret_cast<const idx_t*>(tk.const_data_ptr());
+                if (can_vec) {
+                  const int64_t n_vec = hidden_size / VEC;
+                  const int64_t total = num_tokens * n_vec;
+                  const int block = 256;
+                  const dim3 grid(
+                      std::min<int64_t>((total + block - 1) / block, 65535));
+                  switch (topk) {
+                    case 1:
+                      LAUNCH_MOE_SUM_PAD_AWARE_VEC(1);
+                      break;
+                    case 2:
+                      LAUNCH_MOE_SUM_PAD_AWARE_VEC(2);
+                      break;
+                    case 4:
+                      LAUNCH_MOE_SUM_PAD_AWARE_VEC(4);
+                      break;
+                    case 6:
+                      LAUNCH_MOE_SUM_PAD_AWARE_VEC(6);
+                      break;
+                    case 8:
+                      LAUNCH_MOE_SUM_PAD_AWARE_VEC(8);
+                      break;
+                    case 9:
+                      LAUNCH_MOE_SUM_PAD_AWARE_VEC(9);
+                      break;
+                    default:
+                      vllm::moe::moe_sum_vec_dynamic_kernel<scalar_t, idx_t,
+                                                            true>
+                          <<<grid, dim3(block), 0, stream>>>(
+                              out_ptr, in_ptr, num_tokens, hidden_size, topk,
+                              stride_token, stride_topk, topk_ids_ptr,
+                              expert_map_ptr, stride_tk_token, stride_tk_k);
+                      break;
+                  }
+                } else {
+                  dim3 grid(num_tokens);
+                  dim3 block(std::min(hidden_size, 1024));
+                  vllm::moe::moe_sum_scalar_kernel<scalar_t, idx_t, true>
+                      <<<grid, block, 0, stream>>>(
+                          out_ptr, in_ptr, hidden_size, topk, stride_token,
+                          stride_topk, stride_hidden, topk_ids_ptr,
+                          expert_map_ptr, stride_tk_token, stride_tk_k);
+                }
+              });
+        });
+#undef LAUNCH_MOE_SUM_PAD_AWARE_VEC
+    return;
+  }
+
+#define LAUNCH_MOE_SUM_VEC(TOPK)                                      \
+  vllm::moe::moe_sum_vec_kernel<scalar_t, int32_t, TOPK, false>       \
+      <<<grid, dim3(block), 0, stream>>>(out_ptr, in_ptr, num_tokens, \
+                                         hidden_size, stride_token,   \
+                                         stride_topk, nullptr, nullptr, 0, 0)
 
   VLLM_STABLE_DISPATCH_FLOATING_TYPES(input.scalar_type(), "moe_sum", [&] {
     constexpr int VEC = vllm::moe::MOE_SUM_VEC<scalar_t>;
@@ -774,18 +911,19 @@ void moe_sum(torch::stable::Tensor& input,   // [num_tokens, topk, hidden_size]
           LAUNCH_MOE_SUM_VEC(9);
           break;
         default:
-          vllm::moe::moe_sum_vec_dynamic_kernel<scalar_t>
-              <<<grid, dim3(block), 0, stream>>>(out_ptr, in_ptr, num_tokens,
-                                                 hidden_size, topk,
-                                                 stride_token, stride_topk);
+          vllm::moe::moe_sum_vec_dynamic_kernel<scalar_t, int32_t, false>
+              <<<grid, dim3(block), 0, stream>>>(
+                  out_ptr, in_ptr, num_tokens, hidden_size, topk, stride_token,
+                  stride_topk, nullptr, nullptr, 0, 0);
           break;
       }
     } else {
       dim3 grid(num_tokens);
       dim3 block(std::min(hidden_size, 1024));
-      vllm::moe::moe_sum_scalar_kernel<scalar_t><<<grid, block, 0, stream>>>(
-          out_ptr, in_ptr, hidden_size, topk, stride_token, stride_topk,
-          stride_hidden);
+      vllm::moe::moe_sum_scalar_kernel<scalar_t, int32_t, false>
+          <<<grid, block, 0, stream>>>(out_ptr, in_ptr, hidden_size, topk,
+                                       stride_token, stride_topk, stride_hidden,
+                                       nullptr, nullptr, 0, 0);
     }
   });
 #undef LAUNCH_MOE_SUM_VEC

csrc/libtorch_stable/moe/moe_ops.h

@@ -9,14 +9,16 @@ void topk_softmax(torch::stable::Tensor& topk_weights,
                   torch::stable::Tensor& topk_indices,
                   torch::stable::Tensor& token_expert_indices,
                   torch::stable::Tensor& gating_output, bool renormalize,
-                  std::optional<torch::stable::Tensor> bias);
+                  std::optional<torch::stable::Tensor> bias,
+                  std::optional<torch::stable::Tensor> is_padding);
 
 void topk_sigmoid(torch::stable::Tensor& topk_weights,
                   torch::stable::Tensor& topk_indices,
                   torch::stable::Tensor& token_expert_indices,
                   torch::stable::Tensor& gating_output, bool renormalize,
                   std::optional<torch::stable::Tensor> bias,
-                  double routed_scaling_factor);
+                  double routed_scaling_factor,
+                  std::optional<torch::stable::Tensor> is_padding);
 
 void topk_softplus_sqrt(
     torch::stable::Tensor& topk_weights, torch::stable::Tensor& topk_indices,
@@ -25,9 +27,12 @@ void topk_softplus_sqrt(
     double routed_scaling_factor,
     const std::optional<torch::stable::Tensor>& correction_bias,
     const std::optional<torch::stable::Tensor>& input_ids,
-    const std::optional<torch::stable::Tensor>& tid2eid);
+    const std::optional<torch::stable::Tensor>& tid2eid,
+    const std::optional<torch::stable::Tensor>& is_padding);
 
-void moe_sum(torch::stable::Tensor& input, torch::stable::Tensor& output);
+void moe_sum(torch::stable::Tensor& input, torch::stable::Tensor& output,
+             std::optional<torch::stable::Tensor> topk_ids,
+             std::optional<torch::stable::Tensor> expert_map);
 
 void moe_align_block_size(
     torch::stable::Tensor topk_ids, int64_t num_experts, int64_t block_size,

csrc/libtorch_stable/moe/topk_softmax_kernels.cu

@@ -174,7 +174,8 @@ __launch_bounds__(TPB) __global__ void moeTopK(
     const int end_expert,
     const bool renormalize,
     const float* bias,
-    const double routed_scaling_factor)
+    const double routed_scaling_factor,
+    const bool* is_padding)
 {
 
     using cub_kvp = cub::KeyValuePair<int, float>;
@@ -228,12 +229,14 @@ __launch_bounds__(TPB) __global__ void moeTopK(
             const int expert = result_kvp.key;
             const bool node_uses_expert = expert >= start_expert && expert < end_expert;
             const bool should_process_row = row_is_active && node_uses_expert;
+            const bool is_pad_row = is_padding != nullptr && is_padding[block_row];
 
             const int idx = k * block_row + k_idx;
             // Return the unbiased scores for output weights
             output[idx] = inputs_after_softmax[thread_read_offset + expert];
-            indices[idx] = should_process_row ? (expert - start_expert) : num_experts;
-            assert(indices[idx] >= 0);
+            indices[idx] = is_pad_row ? static_cast<IndType>(-1)
+                                       : (should_process_row ? (expert - start_expert) : num_experts);
+            assert(is_pad_row || indices[idx] >= 0);
             source_rows[idx] = k_idx * num_rows + block_row;
             if (renormalize) {
                 selected_sum += inputs_after_softmax[thread_read_offset + expert];
@@ -277,7 +280,7 @@ template <int VPT, int NUM_EXPERTS, int WARPS_PER_CTA, int BYTES_PER_LDG, int WA
 __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
     void topkGating(const InputType* input, const bool* finished, float* output, const int num_rows, IndType* indices,
         int* source_rows, const int k, const int start_expert, const int end_expert, const bool renormalize,
-        const float* bias, const double routed_scaling_factor)
+        const float* bias, const double routed_scaling_factor, const bool* is_padding)
 {
     static_assert(std::is_same_v<InputType, float> || std::is_same_v<InputType, __nv_bfloat16> ||
                       std::is_same_v<InputType, __half>,
@@ -545,12 +548,14 @@ __launch_bounds__(WARPS_PER_CTA* WARP_SIZE_PARAM) __global__
             // Add a guard to ignore experts not included by this node
             const bool node_uses_expert = expert >= start_expert && expert < end_expert;
             const bool should_process_row = row_is_active && node_uses_expert;
+            const bool is_pad_row = is_padding != nullptr && is_padding[thread_row];
 
             // The lead thread from each sub-group will write out the final results to global memory. (This will be a
             // single) thread per row of the input/output matrices.
             const int idx = k * thread_row + k_idx;
             output[idx] = max_val;
-            indices[idx] = should_process_row ? (expert - start_expert) : NUM_EXPERTS;
+            indices[idx] = is_pad_row ? static_cast<IndType>(-1)
+                                       : (should_process_row ? (expert - start_expert) : NUM_EXPERTS);
             source_rows[idx] = k_idx * num_rows + thread_row;
             if (renormalize) {
                 selected_sum += max_val;
@@ -605,7 +610,7 @@ struct TopkConstants
 template <int EXPERTS, int WARPS_PER_TB, int WARP_SIZE_PARAM, int MAX_BYTES_PER_LDG, typename IndType, typename InputType, ScoringFunc SF>
 void topkGatingLauncherHelper(const InputType* input, const bool* finished, float* output, IndType* indices,
     int* source_row, const int num_rows, const int k, const int start_expert, const int end_expert, const bool renormalize,
-    const float* bias, const double routed_scaling_factor, cudaStream_t stream)
+    const float* bias, const double routed_scaling_factor, cudaStream_t stream, const bool* is_padding)
 {
     static constexpr int BYTES_PER_LDG = MIN(MAX_BYTES_PER_LDG, sizeof(InputType) * EXPERTS);
     using Constants = detail::TopkConstants<EXPERTS, BYTES_PER_LDG, WARP_SIZE_PARAM, InputType>;
@@ -616,7 +621,7 @@ void topkGatingLauncherHelper(const InputType* input, const bool* finished, floa
 
     dim3 block_dim(WARP_SIZE_PARAM, WARPS_PER_TB);
     topkGating<VPT, EXPERTS, WARPS_PER_TB, BYTES_PER_LDG, WARP_SIZE_PARAM, IndType, InputType, SF><<<num_blocks, block_dim, 0, stream>>>(
-        input, finished, output, num_rows, indices, source_row, k, start_expert, end_expert, renormalize, bias, routed_scaling_factor);
+        input, finished, output, num_rows, indices, source_row, k, start_expert, end_expert, renormalize, bias, routed_scaling_factor, is_padding);
 }
 
 #ifndef USE_ROCM
@@ -627,7 +632,7 @@ void topkGatingLauncherHelper(const InputType* input, const bool* finished, floa
                              IndType, InputType, SF>(                         \
         gating_output, nullptr, topk_weights, topk_indices,                   \
         token_expert_indices, num_tokens, topk, 0, num_experts, renormalize,  \
-        bias, routed_scaling_factor, stream);
+        bias, routed_scaling_factor, stream, is_padding);
 #else
   #define LAUNCH_TOPK(NUM_EXPERTS, WARPS_PER_TB, MAX_BYTES)                    \
     if (WARP_SIZE == 64) {                                                     \
@@ -635,13 +640,13 @@ void topkGatingLauncherHelper(const InputType* input, const bool* finished, floa
                                IndType, InputType, SF>(                        \
           gating_output, nullptr, topk_weights, topk_indices,                  \
           token_expert_indices, num_tokens, topk, 0, num_experts, renormalize, \
-          bias, routed_scaling_factor, stream);                                                       \
+          bias, routed_scaling_factor, stream, is_padding);                                            \
     } else if (WARP_SIZE == 32) {                                              \
       topkGatingLauncherHelper<NUM_EXPERTS, WARPS_PER_TB, 32, MAX_BYTES,       \
                                IndType, InputType, SF>(                        \
           gating_output, nullptr, topk_weights, topk_indices,                  \
           token_expert_indices, num_tokens, topk, 0, num_experts, renormalize, \
-          bias, routed_scaling_factor, stream);                                                       \
+          bias, routed_scaling_factor, stream, is_padding);                                            \
     } else {                                                                   \
       assert(false &&                                                          \
              "Unsupported warp size. Only 32 and 64 are supported for ROCm");  \
@@ -661,7 +666,8 @@ void topkGatingKernelLauncher(
     const bool renormalize,
     const float* bias,
     const double routed_scaling_factor,
-    cudaStream_t stream) {
+    cudaStream_t stream,
+    const bool* is_padding) {
     static constexpr int WARPS_PER_TB = 4;
     static constexpr int BYTES_PER_LDG_POWER_OF_2 = 16;
 #ifndef USE_ROCM
@@ -736,7 +742,7 @@ void topkGatingKernelLauncher(
             }
             moeTopK<TPB><<<num_tokens, TPB, 0, stream>>>(
                 workspace, nullptr, topk_weights, topk_indices, token_expert_indices,
-                num_experts, topk, 0, num_experts, renormalize, bias, routed_scaling_factor);
+                num_experts, topk, 0, num_experts, renormalize, bias, routed_scaling_factor, is_padding);
         }
     }
 }
@@ -755,7 +761,8 @@ void dispatch_topk_launch(
     int num_tokens, int num_experts, int topk, bool renormalize,
     std::optional<torch::stable::Tensor> bias,
     double routed_scaling_factor,
-    cudaStream_t stream)
+    cudaStream_t stream,
+    std::optional<torch::stable::Tensor> is_padding)
  {
     const float* bias_ptr = nullptr;
     if (bias.has_value()) {
@@ -769,6 +776,18 @@ void dispatch_topk_launch(
       bias_ptr = bias_tensor.const_data_ptr<float>();
     }
 
+    const bool* is_padding_ptr = nullptr;
+    if (is_padding.has_value()) {
+      const torch::stable::Tensor& is_padding_tensor = is_padding.value();
+      STD_TORCH_CHECK(is_padding_tensor.scalar_type() == torch::headeronly::ScalarType::Bool,
+                      "is_padding tensor must be bool");
+      STD_TORCH_CHECK(is_padding_tensor.dim() == 1, "is_padding tensor must be 1D");
+      STD_TORCH_CHECK(is_padding_tensor.size(0) == num_tokens,
+                      "is_padding size mismatch, expected: ", num_tokens);
+      STD_TORCH_CHECK(is_padding_tensor.is_contiguous(), "is_padding tensor must be contiguous");
+      is_padding_ptr = is_padding_tensor.const_data_ptr<bool>();
+    }
+
     if (topk_indices.scalar_type() == torch::headeronly::ScalarType::Int) {
         vllm::moe::topkGatingKernelLauncher<int, ComputeType, SF>(
             reinterpret_cast<const ComputeType*>(gating_output.const_data_ptr()),
@@ -777,7 +796,7 @@ void dispatch_topk_launch(
             token_expert_indices.mutable_data_ptr<int>(),
             softmax_workspace.mutable_data_ptr<float>(),
             num_tokens, num_experts, topk, renormalize,
-            bias_ptr, routed_scaling_factor, stream);
+            bias_ptr, routed_scaling_factor, stream, is_padding_ptr);
     } else if (topk_indices.scalar_type() == torch::headeronly::ScalarType::UInt32) {
         vllm::moe::topkGatingKernelLauncher<uint32_t, ComputeType, SF>(
             reinterpret_cast<const ComputeType*>(gating_output.const_data_ptr()),
@@ -786,7 +805,7 @@ void dispatch_topk_launch(
             token_expert_indices.mutable_data_ptr<int>(),
             softmax_workspace.mutable_data_ptr<float>(),
             num_tokens, num_experts, topk, renormalize,
-            bias_ptr, routed_scaling_factor, stream);
+            bias_ptr, routed_scaling_factor, stream, is_padding_ptr);
     } else {
         STD_TORCH_CHECK(topk_indices.scalar_type() == torch::headeronly::ScalarType::Long);
         vllm::moe::topkGatingKernelLauncher<int64_t, ComputeType, SF>(
@@ -796,7 +815,7 @@ void dispatch_topk_launch(
             token_expert_indices.mutable_data_ptr<int>(),
             softmax_workspace.mutable_data_ptr<float>(),
             num_tokens, num_experts, topk, renormalize,
-            bias_ptr, routed_scaling_factor, stream);
+            bias_ptr, routed_scaling_factor, stream, is_padding_ptr);
     }
 }
 
@@ -806,7 +825,8 @@ void topk_softmax(
     torch::stable::Tensor& token_expert_indices,        // [num_tokens, topk]
     torch::stable::Tensor& gating_output,               // [num_tokens, num_experts]
     bool renormalize,
-    std::optional<torch::stable::Tensor> bias)
+    std::optional<torch::stable::Tensor> bias,
+    std::optional<torch::stable::Tensor> is_padding)
 {
     const int num_experts = gating_output.size(-1);
     const auto num_tokens = gating_output.numel() / num_experts;
@@ -825,15 +845,15 @@ void topk_softmax(
     if (gating_output.scalar_type() == torch::headeronly::ScalarType::Float) {
         dispatch_topk_launch<float, vllm::moe::SCORING_SOFTMAX>(gating_output, topk_weights, topk_indices,
             token_expert_indices, softmax_workspace, num_tokens, num_experts, topk, renormalize,
-            bias, 1.0, stream);
+            bias, 1.0, stream, is_padding);
     } else if (gating_output.scalar_type() == torch::headeronly::ScalarType::Half) {
         dispatch_topk_launch<__half, vllm::moe::SCORING_SOFTMAX>(gating_output, topk_weights, topk_indices,
             token_expert_indices, softmax_workspace, num_tokens, num_experts, topk, renormalize,
-            bias, 1.0, stream);
+            bias, 1.0, stream, is_padding);
     } else if (gating_output.scalar_type() == torch::headeronly::ScalarType::BFloat16) {
         dispatch_topk_launch<__nv_bfloat16, vllm::moe::SCORING_SOFTMAX>(gating_output, topk_weights, topk_indices,
             token_expert_indices, softmax_workspace, num_tokens, num_experts, topk, renormalize,
-            bias, 1.0, stream);
+            bias, 1.0, stream, is_padding);
     } else {
         STD_TORCH_CHECK(false, "Unsupported gating_output data type: ", gating_output.scalar_type());
     }
@@ -846,7 +866,8 @@ void topk_sigmoid(
     torch::stable::Tensor& gating_output,               // [num_tokens, num_experts]
     bool renormalize,
     std::optional<torch::stable::Tensor> bias,
-    double routed_scaling_factor)
+    double routed_scaling_factor,
+    std::optional<torch::stable::Tensor> is_padding)
 {
     const int num_experts = gating_output.size(-1);
     const auto num_tokens = gating_output.numel() / num_experts;
@@ -865,15 +886,15 @@ void topk_sigmoid(
     if (gating_output.scalar_type() == torch::headeronly::ScalarType::Float) {
         dispatch_topk_launch<float, vllm::moe::SCORING_SIGMOID>(gating_output, topk_weights, topk_indices,
             token_expert_indices, workspace, num_tokens, num_experts, topk, renormalize,
-            bias, routed_scaling_factor, stream);
+            bias, routed_scaling_factor, stream, is_padding);
     } else if (gating_output.scalar_type() == torch::headeronly::ScalarType::Half) {
         dispatch_topk_launch<__half, vllm::moe::SCORING_SIGMOID>(gating_output, topk_weights, topk_indices,
             token_expert_indices, workspace, num_tokens, num_experts, topk, renormalize,
-            bias, routed_scaling_factor, stream);
+            bias, routed_scaling_factor, stream, is_padding);
     } else if (gating_output.scalar_type() == torch::headeronly::ScalarType::BFloat16) {
         dispatch_topk_launch<__nv_bfloat16, vllm::moe::SCORING_SIGMOID>(gating_output, topk_weights, topk_indices,
             token_expert_indices, workspace, num_tokens, num_experts, topk, renormalize,
-            bias, routed_scaling_factor, stream);
+            bias, routed_scaling_factor, stream, is_padding);
     } else {
         STD_TORCH_CHECK(false, "Unsupported gating_output data type: ", gating_output.scalar_type());
     }