updated

Signed-off-by: Robert Shaw <robshaw@redhat.com>

.buildkite/nightly-benchmarks/nightly-descriptions.md

@@ -8,7 +8,7 @@ This benchmark aims to:
 
 Latest results: [results link](https://blog.vllm.ai/2024/09/05/perf-update.html), scroll to the end.
 
-Latest reproduction guide: [github issue link](https://github.com/vllm-project/vllm/issues/8176)
+Latest reproduction guilde: [github issue link](https://github.com/vllm-project/vllm/issues/8176)
 
 ## Setup
 

.buildkite/release-pipeline.yaml

@@ -1,22 +1,24 @@
 steps:
   # aarch64 + CUDA builds. PyTorch 2.8 aarch64 + CUDA wheel is only available on CUDA 12.9
   - label: "Build arm64 wheel - CUDA 12.9"
-    depends_on: ~
     id: build-wheel-arm64-cuda-12-9
     agents:
       queue: arm64_cpu_queue_postmerge
     commands:
       # #NOTE: torch_cuda_arch_list is derived from upstream PyTorch build files here:
       # https://github.com/pytorch/pytorch/blob/main/.ci/aarch64_linux/aarch64_ci_build.sh#L7
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg VLLM_MAIN_CUDA_VERSION=12.9 --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
       - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
+  - block: "Build CUDA 12.8 wheel"
+    key: block-build-cu128-wheel
+
   - label: "Build wheel - CUDA 12.8"
-    depends_on: ~
+    depends_on: block-build-cu128-wheel
     id: build-wheel-cuda-12-8
     agents:
       queue: cpu_queue_postmerge
@@ -28,8 +30,12 @@ steps:
     env:
       DOCKER_BUILDKIT: "1"
 
-  - label: "Build wheel - CUDA 12.6"
+  - block: "Build CUDA 12.6 wheel"
+    key: block-build-cu126-wheel
     depends_on: ~
+
+  - label: "Build wheel - CUDA 12.6"
+    depends_on: block-build-cu126-wheel
     id: build-wheel-cuda-12-6
     agents:
       queue: cpu_queue_postmerge
@@ -76,7 +82,7 @@ steps:
       queue: arm64_cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg FLASHINFER_AOT_COMPILE=true --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.9.1 --build-arg torch_cuda_arch_list='8.7 9.0 10.0+PTX 12.0' --build-arg INSTALL_KV_CONNECTORS=true --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m) --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT-$(uname -m)"
 
   # Add job to create multi-arch manifest
@@ -96,6 +102,8 @@ steps:
     depends_on:
       - create-multi-arch-manifest
       - build-wheel-cuda-12-8
+      - build-wheel-cuda-12-6
+      - build-wheel-cuda-12-9
     id: annotate-release-workflow
     agents:
       queue: cpu_queue_postmerge

.buildkite/scripts/annotate-release.sh

@@ -14,33 +14,18 @@ buildkite-agent annotate --style 'info' --context 'release-workflow' << EOF
 To download the wheel:
 \`\`\`
 aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}/vllm-${RELEASE_VERSION}-cp38-abi3-manylinux2014_aarch64.whl .
-
 aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}+cu126/vllm-${RELEASE_VERSION}+cu126-cp38-abi3-manylinux1_x86_64.whl .
-aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}+cu129/vllm-${RELEASE_VERSION}+cu129-cp38-abi3-manylinux1_x86_64.whl .
+aws s3 cp s3://vllm-wheels/${RELEASE_VERSION}+cu118/vllm-${RELEASE_VERSION}+cu118-cp38-abi3-manylinux1_x86_64.whl . 
 \`\`\`
 
 To download and upload the image:
 
 \`\`\`
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64
-docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-x86_64 vllm/vllm-openai:x86_64
-docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:latest-x86_64
-docker tag vllm/vllm-openai:x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
-docker push vllm/vllm-openai:latest-x86_64
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-x86_64
-
-docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}-aarch64 vllm/vllm-openai:aarch64
-docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:latest-aarch64
-docker tag vllm/vllm-openai:aarch64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
-docker push vllm/vllm-openai:latest-aarch64
-docker push vllm/vllm-openai:v${RELEASE_VERSION}-aarch64
-
-docker manifest create vllm/vllm-openai:latest vllm/vllm-openai:latest-x86_64 vllm/vllm-openai:latest-aarch64 --amend
-docker manifest create vllm/vllm-openai:v${RELEASE_VERSION} vllm/vllm-openai:v${RELEASE_VERSION}-x86_64 vllm/vllm-openai:v${RELEASE_VERSION}-aarch64 --amend
-docker manifest push vllm/vllm-openai:latest
-docker manifest push vllm/vllm-openai:v${RELEASE_VERSION}
+docker pull public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT}
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${BUILDKITE_COMMIT} vllm/vllm-openai
+docker tag vllm/vllm-openai vllm/vllm-openai:latest
+docker tag vllm/vllm-openai vllm/vllm-openai:v${RELEASE_VERSION}
+docker push vllm/vllm-openai:latest
+docker push vllm/vllm-openai:v${RELEASE_VERSION}
 \`\`\`
 EOF 

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

@@ -86,6 +86,10 @@ if [[ $commands == *"pytest -v -s models/test_registry.py"* ]]; then
   commands=${commands//"pytest -v -s models/test_registry.py"/"pytest -v -s models/test_registry.py -k 'not BambaForCausalLM and not GritLM and not Mamba2ForCausalLM and not Zamba2ForCausalLM'"}
 fi
 
+if [[ $commands == *"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'"* ]]; then
+  commands=${commands//"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'"/"VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2 and not BambaForCausalLM and not Gemma2ForCausalLM and not Grok1ModelForCausalLM and not Zamba2ForCausalLM and not Gemma2Model and not GritLM'"}
+fi
+
 if [[ $commands == *"pytest -v -s compile/test_basic_correctness.py"* ]]; then
   commands=${commands//"pytest -v -s compile/test_basic_correctness.py"/"VLLM_USE_TRITON_FLASH_ATTN=0 pytest -v -s compile/test_basic_correctness.py"}
 fi
@@ -163,6 +167,12 @@ if [[ $commands == *" entrypoints/llm "* ]]; then
   --ignore=entrypoints/llm/test_prompt_validation.py "}
 fi
 
+#Obsolete currently
+##ignore certain Entrypoints/llm tests
+#if [[ $commands == *" && pytest -v -s entrypoints/llm/test_guided_generate.py"* ]]; then
+#  commands=${commands//" && pytest -v -s entrypoints/llm/test_guided_generate.py"/" "}
+#fi
+
 # --ignore=entrypoints/openai/test_encoder_decoder.py \
 # --ignore=entrypoints/openai/test_embedding.py \
 # --ignore=entrypoints/openai/test_oot_registration.py

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

@@ -58,11 +58,15 @@ function cpu_tests() {
     # pytest -x -v -s tests/kernels/attention/test_cache.py -m cpu_model
     # pytest -x -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
 
-    pytest -x -v -s tests/models/language/generation -m cpu_model
-    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model
+    # Note: disable Bart until supports V1
+    pytest -x -v -s tests/models/language/generation -m cpu_model \
+                --ignore=tests/models/language/generation/test_bart.py
+    VLLM_CPU_SGL_KERNEL=1 pytest -x -v -s tests/models/language/generation -m cpu_model \
+                --ignore=tests/models/language/generation/test_bart.py
 
     pytest -x -v -s tests/models/language/pooling -m cpu_model
     pytest -x -v -s tests/models/multimodal/generation \
+                --ignore=tests/models/multimodal/generation/test_mllama.py \
                 --ignore=tests/models/multimodal/generation/test_pixtral.py \
                 -m cpu_model"
 

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

@@ -62,7 +62,7 @@ echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
     && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
-    && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
+    && python3 -m pip install --progress-bar off hf-transfer
 echo "--- Python dependencies installed ---"
 export VLLM_USE_V1=1
 export VLLM_XLA_CHECK_RECOMPILATION=1

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

@@ -62,7 +62,7 @@ echo "--- Installing Python dependencies ---"
 python3 -m pip install --progress-bar off git+https://github.com/thuml/depyf.git \
     && python3 -m pip install --progress-bar off pytest pytest-asyncio tpu-info \
     && python3 -m pip install --progress-bar off "lm-eval @ git+https://github.com/EleutherAI/lm-evaluation-harness.git@206b7722158f58c35b7ffcd53b035fdbdda5126d" \
-    && python3 -m pip install --progress-bar off hf-transfer tblib==3.1.0
+    && python3 -m pip install --progress-bar off hf-transfer
 echo "--- Python dependencies installed ---"
 export VLLM_USE_V1=1
 export VLLM_XLA_CHECK_RECOMPILATION=1

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

@@ -30,12 +30,11 @@ docker run \
     bash -c '
     set -e
     echo $ZE_AFFINITY_MASK
-    pip install tblib==3.1.0
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 -O3 -O.cudagraph_mode=NONE
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend ray
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager -tp 2 --distributed-executor-backend mp
-    VLLM_ATTENTION_BACKEND=TRITON_ATTN python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
+    VLLM_ATTENTION_BACKEND=TRITON_ATTN_VLLM_V1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
     cd tests
     pytest -v -s v1/core
     pytest -v -s v1/engine

.buildkite/scripts/run-prime-rl-test.sh

@@ -1,59 +0,0 @@
-#!/bin/bash
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-# Setup script for Prime-RL integration tests
-# This script prepares the environment for running Prime-RL tests with nightly vLLM
-
-set -euo pipefail
-
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd "${SCRIPT_DIR}/../.." && pwd)"
-PRIME_RL_REPO="https://github.com/PrimeIntellect-ai/prime-rl.git"
-PRIME_RL_DIR="${REPO_ROOT}/prime-rl"
-
-echo "Setting up Prime-RL integration test environment..."
-
-# Clean up any existing Prime-RL directory
-if [ -d "${PRIME_RL_DIR}" ]; then
-    echo "Removing existing Prime-RL directory..."
-    rm -rf "${PRIME_RL_DIR}"
-fi
-
-# Install UV if not available
-if ! command -v uv &> /dev/null; then
-    echo "Installing UV package manager..."
-    curl -LsSf https://astral.sh/uv/install.sh | sh
-    source $HOME/.local/bin/env
-fi
-
-# Clone Prime-RL repository at specific branch for reproducible tests
-PRIME_RL_BRANCH="integ-vllm-main"
-echo "Cloning Prime-RL repository at branch: ${PRIME_RL_BRANCH}..."
-git clone --branch "${PRIME_RL_BRANCH}" --single-branch "${PRIME_RL_REPO}" "${PRIME_RL_DIR}"
-cd "${PRIME_RL_DIR}"
-
-echo "Setting up UV project environment..."
-export UV_PROJECT_ENVIRONMENT=/usr/local
-ln -s /usr/bin/python3 /usr/local/bin/python
-
-# Remove vllm pin from pyproject.toml
-echo "Removing vllm pin from pyproject.toml..."
-sed -i '/vllm==/d' pyproject.toml
-
-# Sync Prime-RL dependencies
-echo "Installing Prime-RL dependencies..."
-uv sync --inexact && uv sync --inexact --all-extras
-
-# Verify installation
-echo "Verifying installations..."
-uv run python -c "import vllm; print(f'vLLM version: {vllm.__version__}')"
-uv run python -c "import prime_rl; print('Prime-RL imported successfully')"
-
-echo "Prime-RL integration test environment setup complete!"
-
-echo "Running Prime-RL integration tests..."
-export WANDB_MODE=offline # this makes this test not require a WANDB_API_KEY
-uv run pytest -vs tests/integration/test_rl.py -m gpu
-
-echo "Prime-RL integration tests completed!"

.buildkite/test-pipeline.yaml

@@ -6,28 +6,24 @@
 # to generate the final pipeline yaml file.
 
 # Documentation
-# label(str): the name of the test. emojis allowed.
-# fast_check(bool): whether to run this on each commit on the fastcheck pipeline.
-# torch_nightly(bool): whether to run this on vllm against the torch nightly pipeline.
-# fast_check_only(bool): run this test on the fastcheck pipeline only
-# optional(bool): never run this test by default (i.e. need to unblock manually) unless it's a scheduled nightly run.
-# soft_fail(bool): allow this step to fail without failing the entire pipeline (useful for flaky or experimental tests).
+# label(str): the name of the test. emoji allowed.
+# fast_check(bool): whether to run this on each commit on fastcheck pipeline.
+# torch_nightly(bool): whether to run this on vllm against torch nightly pipeline.
+# fast_check_only(bool): run this test on fastcheck pipeline only
+# optional(bool): never run this test by default (i.e. need to unblock manually) unless it's scheduled nightly run.
 # command(str): the single command to run for tests. incompatible with commands.
-# commands(list): the list of commands to run for the test. incompatible with command.
-# mirror_hardwares(list): the list of hardware to run the test on as well. currently only supports [amdexperimental]
-# gpu(str): override the GPU selection for the test. default is L4 GPUs. supports a100, b200, h200
-# num_gpus(int): override the number of GPUs for the test. defaults to 1 GPU. currently supports 2,4.
-# num_nodes(int): whether to simulate multi-node setup by launching multiple containers on one host,
-#     in this case, commands must be specified. the first command runs on the first host, the second
+# commands(list): the list of commands to run for test. incompatbile with command.
+# mirror_hardwares(list): the list of hardwares to run the test on as well. currently only supports [amd]
+# gpu(str): override the GPU selection for the test. default is on L4 GPUs. currently only supports a100
+# num_gpus(int): override the number of GPUs for the test. default to 1 GPU. currently support 2,4.
+# num_nodes(int): whether to simulate multi-node setup by launch multiple containers on one host,
+#     in this case, commands must be specified. the first command runs on first host, the second
 #     command runs on the second host.
-# timeout_in_minutes(int): sets a timeout for the step in minutes. if not specified, uses the default timeout.
-# parallelism(int): number of parallel jobs to run for this step. enables test sharding using $$BUILDKITE_PARALLEL_JOB
-#     and $$BUILDKITE_PARALLEL_JOB_COUNT environment variables.
-# working_dir(str): specify the place where the command should execute, default to /vllm-workspace/tests
-# source_file_dependencies(list): the list of prefixes to opt-in the test for, if empty, the test will always run.
+# working_dir(str): specify the place where command should execute, default to /vllm-workspace/tests
+# source_file_dependencies(list): the list of prefix to opt-in the test for, if empty, the test will always run.
 
 # When adding a test
-# - If the test belongs to an existing group, add it there
+# - If the test belong to an existing group, add it there
 # - If the test is short, add to any existing step
 # - If the test takes more than 10min, then it is okay to create a new step.
 #   Note that all steps execute in parallel.
@@ -50,19 +46,23 @@ steps:
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
   - vllm/
+  - tests/mq_llm_engine
+  - tests/async_engine
   - tests/test_inputs.py
   - tests/test_outputs.py
   - tests/multimodal
   - tests/utils_
+  - tests/worker
   - tests/standalone_tests/lazy_imports.py
-  - tests/transformers_utils
   commands:
   - python3 standalone_tests/lazy_imports.py
+  - pytest -v -s mq_llm_engine # MQLLMEngine
+  - pytest -v -s async_engine # AsyncLLMEngine
   - pytest -v -s test_inputs.py
   - pytest -v -s test_outputs.py
   - pytest -v -s multimodal
   - pytest -v -s utils_ # Utils
-  - pytest -v -s transformers_utils # transformers_utils
+  - pytest -v -s worker # Worker
 
 - label: Python-only Installation Test # 10min
   timeout_in_minutes: 20
@@ -82,25 +82,27 @@ steps:
   - vllm/
   - tests/basic_correctness/test_basic_correctness
   - tests/basic_correctness/test_cpu_offload
+  - tests/basic_correctness/test_preemption
   - tests/basic_correctness/test_cumem.py
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s basic_correctness/test_cumem.py
   - pytest -v -s basic_correctness/test_basic_correctness.py
   - pytest -v -s basic_correctness/test_cpu_offload.py
+  - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
 
-- label: Entrypoints Unit Tests # 5min
-  timeout_in_minutes: 10
-  working_dir: "/vllm-workspace/tests"
+- label: Core Test # 22min
+  timeout_in_minutes: 35
+  mirror_hardwares: [amdexperimental]
   fast_check: true
   source_file_dependencies:
-  - vllm/entrypoints
-  - tests/entrypoints/
+  - vllm/core
+  - vllm/distributed
+  - tests/core
   commands:
-  - pytest -v -s entrypoints/openai/tool_parsers
-  - pytest -v -s entrypoints/ --ignore=entrypoints/llm --ignore=entrypoints/openai --ignore=entrypoints/offline_mode --ignore=entrypoints/test_chat_utils.py  --ignore=entrypoints/pooling
+  - pytest -v -s core
 
-- label: Entrypoints Integration Test (LLM) # 30min
+- label: Entrypoints Test (LLM) # 30min
   timeout_in_minutes: 40
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
@@ -112,11 +114,12 @@ steps:
   - tests/entrypoints/offline_mode
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
+  - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
+  - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
-- label: Entrypoints Integration Test (API Server) # 100min
+- label: Entrypoints Test (API Server) # 100min
   timeout_in_minutes: 130
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
@@ -129,22 +132,9 @@ steps:
   commands:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - PYTHONPATH=/vllm-workspace pytest -v -s entrypoints/openai/test_collective_rpc.py # PYTHONPATH is needed to import custom Worker extension
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/test_collective_rpc.py --ignore=entrypoints/openai/tool_parsers/
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/test_tensorizer_entrypoint.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/test_collective_rpc.py
   - pytest -v -s entrypoints/test_chat_utils.py
 
-- label: Entrypoints Integration Test (Pooling)
-  timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
-  working_dir: "/vllm-workspace/tests"
-  fast_check: true
-  torch_nightly: true
-  source_file_dependencies:
-  - vllm/
-  - tests/entrypoints/pooling
-  commands:
-  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  - pytest -v -s entrypoints/pooling
-
 - label: Distributed Tests (4 GPUs) # 35min
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
@@ -152,6 +142,7 @@ steps:
   num_gpus: 4
   source_file_dependencies:
   - vllm/distributed/
+  - vllm/core/
   - tests/distributed/test_utils
   - tests/distributed/test_pynccl
   - tests/distributed/test_events
@@ -164,20 +155,12 @@ steps:
   - tests/v1/test_internal_lb_dp.py
   - tests/v1/test_hybrid_lb_dp.py
   - tests/v1/engine/test_engine_core_client.py
-  - tests/distributed/test_symm_mem_allreduce.py
   commands:
-  # test with torchrun tp=2 and external_dp=2
+  # test with tp=2 and external_dp=2
+  - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
   - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
-  # test with torchrun tp=2 and pp=2
+  # test with tp=2 and pp=2
   - PP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
-  # test with torchrun tp=4 and dp=1
-  - TP_SIZE=4 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
-  # test with torchrun tp=2, pp=2 and dp=1
-  - PP_SIZE=2 TP_SIZE=2 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
-  # test with torchrun tp=1 and dp=4 with ep
-  - DP_SIZE=4 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
-  # test with torchrun tp=2 and dp=2 with ep
-  - TP_SIZE=2 DP_SIZE=2 ENABLE_EP=1 torchrun --nproc-per-node=4 distributed/test_torchrun_example_moe.py
   # test with internal dp
   - python3 ../examples/offline_inference/data_parallel.py --enforce-eager
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
@@ -189,7 +172,6 @@ steps:
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s distributed/test_events.py
-  - pytest -v -s distributed/test_symm_mem_allreduce.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
   - pushd ../examples/offline_inference
@@ -222,14 +204,16 @@ steps:
   num_gpus: 2
   source_file_dependencies:
   - vllm/
-  - tests/v1/tracing
+  - tests/metrics
+  - tests/tracing
   commands:
+  - pytest -v -s metrics
   - "pip install \
       'opentelemetry-sdk>=1.26.0' \
       'opentelemetry-api>=1.26.0' \
       'opentelemetry-exporter-otlp>=1.26.0' \
       'opentelemetry-semantic-conventions-ai>=0.4.1'"
-  - pytest -v -s v1/tracing
+  - pytest -v -s tracing
 
 ##### fast check tests  #####
 #####  1 GPU test  #####
@@ -292,7 +276,6 @@ steps:
     # split the test to avoid interference
     - pytest -v -s v1/core
     - pytest -v -s v1/executor
-    - pytest -v -s v1/kv_offload
     - pytest -v -s v1/sample
     - pytest -v -s v1/logits_processors
     - pytest -v -s v1/worker
@@ -300,12 +283,10 @@ steps:
     - pytest -v -s v1/spec_decode
     - pytest -v -s v1/kv_connector/unit
     - pytest -v -s v1/metrics
-    - pytest -v -s v1/test_kv_sharing.py
-    - pytest -v -s v1/test_metrics_reader.py
-    - pytest -v -s v1/test_oracle.py
-    - pytest -v -s v1/test_request.py
     - pytest -v -s v1/test_serial_utils.py
     - pytest -v -s v1/test_utils.py
+    - pytest -v -s v1/test_oracle.py
+    - pytest -v -s v1/test_metrics_reader.py
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-redhat/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
@@ -328,13 +309,13 @@ steps:
     - python3 offline_inference/vision_language.py --seed 0
     - python3 offline_inference/vision_language_pooling.py --seed 0
     - python3 offline_inference/vision_language_multi_image.py --seed 0
-    - python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
+    - VLLM_USE_V1=0 python3 others/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 others/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
+    - python3 offline_inference/encoder_decoder.py
     - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
     - python3 offline_inference/basic/classify.py
     - python3 offline_inference/basic/embed.py
     - python3 offline_inference/basic/score.py
-    - python3 offline_inference/spec_decode.py --test --method eagle --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 2048
-    - python3 offline_inference/spec_decode.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 2048
+    - VLLM_USE_V1=0 python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Platform Tests (CUDA) # 4min
   timeout_in_minutes: 15
@@ -388,7 +369,6 @@ steps:
     - pytest -v -s compile/test_async_tp.py
     - pytest -v -s compile/test_fusion_all_reduce.py
     - pytest -v -s compile/test_decorator.py
-    - pytest -v -s compile/test_noop_elimination.py
 
 - label: PyTorch Fullgraph Smoke Test # 15min
   timeout_in_minutes: 30
@@ -399,7 +379,11 @@ steps:
   - tests/compile
   commands:
   - pytest -v -s compile/test_basic_correctness.py
-  - pytest -v -s compile/piecewise/
+  # these tests need to be separated, cannot combine
+  - pytest -v -s compile/piecewise/test_simple.py
+  - pytest -v -s compile/piecewise/test_toy_llama.py
+  - pytest -v -s compile/piecewise/test_full_cudagraph.py
+  - pytest -v -s compile/piecewise/test_multiple_graphs.py
 
 - label: PyTorch Fullgraph Test # 20min
   timeout_in_minutes: 30
@@ -517,10 +501,6 @@ steps:
   commands:
   # temporary install here since we need nightly, will move to requirements/test.in
   # after torchao 0.12 release, and pin a working version of torchao nightly here
-
-  # since torchao nightly is only compatible with torch nightly currently
-  # https://github.com/pytorch/ao/issues/2919, we'll have to skip new torchao tests for now
-  # we can only upgrade after this is resolved
   - pip install --pre torchao==0.13.0.dev20250814 --index-url https://download.pytorch.org/whl/nightly/cu128
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization
 
@@ -543,6 +523,15 @@ steps:
   commands: # LMEval+Transcription WER check
   - pytest -s entrypoints/openai/correctness/
 
+- label: Encoder Decoder tests # 12min
+  timeout_in_minutes: 20
+  mirror_hardwares: [amdexperimental]
+  source_file_dependencies:
+  - vllm/
+  - tests/encoder_decoder
+  commands:
+    - pytest -v -s encoder_decoder
+
 - label: OpenAI-Compatible Tool Use # 23 min
   timeout_in_minutes: 35
   mirror_hardwares: [amdexperimental]
@@ -557,85 +546,36 @@ steps:
 
 #####  models test  #####
 
-- label: Basic Models Tests (Initialization)
-  timeout_in_minutes: 45
+- label: Basic Models Test # 57min
+  timeout_in_minutes: 75
   mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
   - vllm/
-  - tests/models/test_initialization.py
+  - tests/models
   commands:
-    # Run a subset of model initialization tests
-    - pytest -v -s models/test_initialization.py::test_can_initialize_small_subset
+    - pytest -v -s models/test_transformers.py
+    - pytest -v -s models/test_registry.py
+    - pytest -v -s models/test_utils.py
+    - pytest -v -s models/test_vision.py
+    - pytest -v -s models/test_initialization.py
 
-- label: Basic Models Tests (Extra Initialization) %N
+- label: Language Models Test (Standard) # 35min
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
-  - vllm/model_executor/models/
-  - tests/models/test_initialization.py
-  commands:
-    # Only when vLLM model source is modified - test initialization of a large
-    # subset of supported models (the complement of the small subset in the above
-    # test.) Also run if model initialization test file is modified
-    - pytest -v -s models/test_initialization.py \
-             -k 'not test_can_initialize_small_subset' \
-             --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT \
-             --shard-id=$$BUILDKITE_PARALLEL_JOB
-  parallelism: 2
-
-- label: Basic Models Tests (Other)
-  timeout_in_minutes: 45
-  mirror_hardwares: [amdexperimental]
-  torch_nightly: true
-  source_file_dependencies:
-  - vllm/
-  - tests/models/test_transformers.py
-  - tests/models/test_registry.py
-  - tests/models/test_utils.py
-  - tests/models/test_vision.py
-  commands:
-    - pytest -v -s models/test_transformers.py \
-                   models/test_registry.py \
-                   models/test_utils.py \
-                   models/test_vision.py
-
-- label: Language Models Tests (Standard)
-  timeout_in_minutes: 25
-  mirror_hardwares: [amdexperimental]
-  torch_nightly: true
-  source_file_dependencies:
   - vllm/
   - tests/models/language
   commands:
-    # Test standard language models, excluding a subset of slow tests
     - pip freeze | grep -E 'torch'
-    - pytest -v -s models/language -m 'core_model and (not slow_test)'
+    - pytest -v -s models/language -m core_model
 
-- label: Language Models Tests (Extra Standard) %N
+- label: Language Models Test (Hybrid) # 35 min
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental]
   torch_nightly: true
   source_file_dependencies:
-  - vllm/model_executor/models/
-  - tests/models/language/pooling/test_embedding.py
-  - tests/models/language/generation/test_common.py
-  - tests/models/language/pooling/test_classification.py
-  commands:
-    # Shard slow subset of standard language models tests. Only run when model
-    # source is modified, or when specified test files are modified
-    - pip freeze | grep -E 'torch'
-    - pytest -v -s models/language -m 'core_model and slow_test' \
-             --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT \
-             --shard-id=$$BUILDKITE_PARALLEL_JOB
-  parallelism: 2
-
-- label: Language Models Tests (Hybrid) %N
-  timeout_in_minutes: 75
-  mirror_hardwares: [amdexperimental]
-  torch_nightly: true
-  source_file_dependencies:
   - vllm/
   - tests/models/language/generation
   commands:
@@ -643,12 +583,7 @@ steps:
     # Note: also needed to run plamo2 model in vLLM
     - uv pip install --system --no-build-isolation 'git+https://github.com/state-spaces/mamba@v2.2.5'
     - uv pip install --system --no-build-isolation 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.2'
-    # 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
-  parallelism: 2
+    - pytest -v -s models/language/generation -m hybrid_model
 
 - label: Language Models Test (Extended Generation) # 80min
   timeout_in_minutes: 110
@@ -662,16 +597,6 @@ steps:
     - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
     - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
-- label: Language Models Test (PPL)
-  timeout_in_minutes: 110
-  mirror_hardwares: [amdexperimental]
-  optional: true
-  source_file_dependencies:
-  - vllm/
-  - tests/models/language/generation_ppl_test
-  commands:
-    - pytest -v -s models/language/generation_ppl_test
-
 - label: Language Models Test (Extended Pooling)  # 36min
   timeout_in_minutes: 50
   mirror_hardwares: [amdexperimental]
@@ -682,16 +607,6 @@ steps:
   commands:
     - pytest -v -s models/language/pooling -m 'not core_model'
 
-- label: Language Models Test (MTEB)
-  timeout_in_minutes: 110
-  mirror_hardwares: [amdexperimental]
-  optional: true
-  source_file_dependencies:
-  - vllm/
-  - tests/models/language/pooling_mteb_test
-  commands:
-    - pytest -v -s models/language/pooling_mteb_test
-
 - label: Multi-Modal Processor Test # 44min
   timeout_in_minutes: 60
   source_file_dependencies:
@@ -712,7 +627,7 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pip freeze | grep -E 'torch'
     - pytest -v -s models/multimodal -m core_model --ignore models/multimodal/generation/test_whisper.py --ignore models/multimodal/processing
-    - 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
+    - cd .. && pytest -v -s tests/models/multimodal/generation/test_whisper.py -m core_model  # Otherwise, mp_method="spawn" doesn't work
 
 - label: Multi-Modal Models Test (Extended) 1
   mirror_hardwares: [amdexperimental]
@@ -772,9 +687,8 @@ steps:
     - pytest -v -s tests/models/multimodal/processing/
     - pytest -v -s tests/models/multimodal/test_mapping.py
     - python3 examples/offline_inference/basic/chat.py
+    - python3 examples/offline_inference/audio_language.py --model-type whisper
     - python3 examples/offline_inference/vision_language.py --model-type qwen2_5_vl
-    # Whisper needs spawn method to avoid deadlock
-    - VLLM_WORKER_MULTIPROC_METHOD=spawn python3 examples/offline_inference/audio_language.py --model-type whisper
 
 - label: Blackwell Test # 38 min
   timeout_in_minutes: 60
@@ -799,12 +713,11 @@ steps:
     # num_heads2 broken by https://github.com/flashinfer-ai/flashinfer/issues/1353
     - pytest -v -s tests/kernels/attention/test_flashinfer.py -k 'not num_heads2'
     - pytest -v -s tests/kernels/attention/test_flashinfer_trtllm_attention.py
-    - pytest -v -s tests/kernels/attention/test_cutlass_mla_decode.py
-    - pytest -v -s tests/kernels/attention/test_flashinfer_mla_decode.py
+    - pytest -v -s tests/kernels/test_cutlass_mla_decode.py
     # Quantization
     - pytest -v -s tests/kernels/quantization/test_cutlass_scaled_mm.py -k 'fp8'
     - pytest -v -s tests/kernels/quantization/test_nvfp4_quant.py
-    - pytest -v -s tests/kernels/quantization/test_silu_mul_nvfp4_quant.py
+    - pytest -v -s tests/kernels/quantization/test_silu_nvfp4_quant_fusion.py
     - pytest -v -s tests/kernels/quantization/test_nvfp4_scaled_mm.py
     - pytest -v -s tests/kernels/quantization/test_flashinfer_scaled_mm.py
     - pytest -v -s tests/kernels/quantization/test_flashinfer_nvfp4_scaled_mm.py
@@ -816,20 +729,6 @@ steps:
     - pytest -v -s tests/kernels/moe/test_flashinfer.py
     - pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
 
-- label: GPT-OSS Eval (Blackwell)
-  timeout_in_minutes: 60
-  working_dir: "/vllm-workspace/"
-  gpu: b200
-  optional: true # disable while debugging
-  source_file_dependencies:
-  - tests/evals/gpt_oss
-  - vllm/model_executor/models/gpt_oss.py
-  - vllm/model_executor/layers/quantization/mxfp4.py
-  - vllm/v1/attention/backends/flashinfer.py
-  commands:
-    - uv pip install --system 'gpt-oss[eval]==0.0.5'
-    - pytest -s -v tests/evals/gpt_oss/test_gpqa_correctness.py --model openai/gpt-oss-20b --metric 0.58 --server-args '--tensor-parallel-size 2'
-
 #####  1 GPU test  #####
 #####  multi gpus test  #####
 
@@ -844,8 +743,6 @@ steps:
   commands:
   - pytest -v -s distributed/test_comm_ops.py
   - pytest -v -s distributed/test_shm_broadcast.py
-  - pytest -v -s distributed/test_shm_buffer.py
-  - pytest -v -s distributed/test_shm_storage.py
 
 - label: 2 Node Tests (4 GPUs in total) # 16min
   timeout_in_minutes: 30
@@ -872,28 +769,26 @@ steps:
     - NUM_NODES=2 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_node_count.py | grep 'Node count test passed'
     - python3 ../examples/offline_inference/data_parallel.py --dp-size=2 --tp-size=1 --node-size=2 --node-rank=1 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
 
-- label: Distributed Tests (2 GPUs) # 68min
-  timeout_in_minutes: 90
+- label: Distributed Tests (2 GPUs) # 110min
+  timeout_in_minutes: 150
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
-  - vllm/compilation/
   - vllm/distributed/
   - vllm/engine/
   - vllm/executor/
-  - vllm/worker/worker_base.py
-  - vllm/v1/engine/
-  - vllm/v1/worker/
-  - tests/compile/test_basic_correctness.py
-  - tests/compile/test_wrapper.py
+  - vllm/model_executor/models/
   - tests/distributed/
-  - tests/entrypoints/llm/test_collective_rpc.py
+  - vllm/compilation
+  - vllm/worker/worker_base.py
+  - vllm/worker/worker.py
+  - vllm/worker/model_runner.py
+  - entrypoints/llm/test_collective_rpc.py
   - tests/v1/test_async_llm_dp.py
   - tests/v1/test_external_lb_dp.py
   - tests/v1/entrypoints/openai/test_multi_api_servers.py
-  - tests/v1/shutdown
-  - tests/v1/worker/test_worker_memory_snapshot.py
+  - vllm/v1/engine/
   commands:
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_external_lb_dp.py
@@ -902,29 +797,18 @@ steps:
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
-  - pytest -v -s distributed/test_sequence_parallel.py
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
-  - pytest -v -s v1/worker/test_worker_memory_snapshot.py
-
-- label: Distributed Model Tests (2 GPUs) # 37min
-  timeout_in_minutes: 50
-  mirror_hardwares: [amdexperimental]
-  working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
-  source_file_dependencies:
-  - vllm/model_executor/model_loader/sharded_state_loader.py
-  - vllm/model_executor/models/
-  - tests/basic_correctness/
-  - tests/model_executor/model_loader/test_sharded_state_loader.py
-  - tests/models/
-  commands:
   - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s model_executor/model_loader/test_sharded_state_loader.py
   # Avoid importing model tests that cause CUDA reinitialization error
   - pytest models/test_transformers.py -v -s -m 'distributed(num_gpus=2)'
   - pytest models/language -v -s -m 'distributed(num_gpus=2)'
-  - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)' --ignore models/multimodal/generation/test_whisper.py
-  - VLLM_WORKER_MULTIPROC_METHOD=spawn pytest models/multimodal/generation/test_whisper.py -v -s -m 'distributed(num_gpus=2)'
+  - pytest models/multimodal -v -s -m 'distributed(num_gpus=2)'
+  # test sequence parallel
+  - pytest -v -s distributed/test_sequence_parallel.py
+  # this test fails consistently.
+  # TODO: investigate and fix
+  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
+  - pytest -v -s models/multimodal/generation/test_maverick.py
 
 - label: Plugin Tests (2 GPUs) # 40min
   timeout_in_minutes: 60
@@ -943,7 +827,7 @@ steps:
   # begin io_processor plugins test, all the code in between uses the prithvi_io_processor plugin
   - pip install -e ./plugins/prithvi_io_processor_plugin
   - pytest -v -s plugins_tests/test_io_processor_plugins.py
-  - pip uninstall prithvi_io_processor_plugin -y
+  - pip uninstall prithvi_io_processor_plugin -y 
   # end io_processor plugins test
   # other tests continue here:
   - pytest -v -s plugins_tests/test_scheduler_plugins.py
@@ -967,6 +851,7 @@ steps:
   commands:
   - pytest -v -s distributed/test_pp_cudagraph.py
   - pytest -v -s distributed/test_pipeline_parallel.py
+  # - pytest -v -s distributed/test_context_parallel.py # TODO: enable it on Hopper runners or add triton MLA support
 
 - label: LoRA TP Test (Distributed) # 17 min
   timeout_in_minutes: 30
@@ -990,7 +875,7 @@ steps:
   timeout_in_minutes: 45
   mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
-  num_gpus: 2
+  num_gpus: 2 
   optional: true
   source_file_dependencies:
   - vllm/
@@ -1040,34 +925,9 @@ steps:
   - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
 
-##### H200 test #####
-- label: Distrubted Tests (H200) # optional
+- label: Qwen MoE EP Test # optional
   gpu: h200
   optional: true
-  working_dir: "/vllm-workspace/"
   num_gpus: 2
   commands:
-    - pytest -v -s tests/distributed/test_context_parallel.py
-    - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048
-
-##### B200 test #####
-- label: Distributed Tests (B200) # optional
-  gpu: b200
-  optional: true
-  working_dir: "/vllm-workspace/"
-  num_gpus: 2
-  commands:
-    - pytest -v -s tests/distributed/test_context_parallel.py
-    - pytest -v -s tests/distributed/test_nccl_symm_mem_allreduce.py
-
-##### RL Integration Tests #####
-- label: Prime-RL Integration Test # 15min
-  timeout_in_minutes: 30
-  optional: true
-  num_gpus: 2
-  working_dir: "/vllm-workspace"
-  source_file_dependencies:
-  - vllm/
-  - .buildkite/scripts/run-prime-rl-test.sh
-  commands:
-    - bash .buildkite/scripts/run-prime-rl-test.sh
+    - CUDA_VISIBLE_DEVICES=1,2 VLLM_ALL2ALL_BACKEND=deepep_high_throughput VLLM_USE_DEEP_GEMM=1 VLLM_LOGGING_LEVEL=DEBUG python3 /vllm-workspace/examples/offline_inference/data_parallel.py --model Qwen/Qwen1.5-MoE-A2.7B --tp-size=1  --dp-size=2 --max-model-len 2048

.coveragerc

@@ -1,32 +0,0 @@
-[run]
-source = vllm
-omit =
-    */tests/*
-    */test_*
-    */__pycache__/*
-    */build/*
-    */dist/*
-    */vllm.egg-info/*
-    */third_party/*
-    */examples/*
-    */benchmarks/*
-    */docs/*
-
-[report]
-exclude_lines =
-    pragma: no cover
-    def __repr__
-    if self.debug:
-    if settings.DEBUG
-    raise AssertionError
-    raise NotImplementedError
-    if 0:
-    if __name__ == .__main__.:
-    class .*\bProtocol\):
-    @(abc\.)?abstractmethod
-
-[html]
-directory = htmlcov
-
-[xml]
-output = coverage.xml

.github/.bc-linter.yml

@@ -1,24 +0,0 @@
-# doc: https://github.com/pytorch/test-infra/blob/main/tools/stronghold/docs/bc_linter_config.md
-version: 1
-paths:
-# We temporarily disable globally, and will only enable with `annotations.include`
-# include:
-#   - "vllm/v1/attetion/*.py"
-#   - "vllm/v1/core/*.py"
-exclude:
-  - "**/*.py"
-
-scan:
-  functions: true        # check free functions and methods
-  classes: true          # check classes/dataclasses
-  public_only: true      # ignore names starting with "_" at any level
-
-annotations:
-  include:               # decorators that force‑include a symbol
-    - name: "bc_linter_include"  # matched by simple name or dotted suffix
-      propagate_to_members: false # for classes, include methods/inner classes
-  exclude:               # decorators that force‑exclude a symbol
-    - name: "bc_linter_skip"     # matched by simple name or dotted suffix
-      propagate_to_members: true  # for classes, exclude methods/inner classes
-
-excluded_violations: []  # e.g. ["ParameterRenamed", "FieldTypeChanged"]

.github/CODEOWNERS

@@ -2,24 +2,23 @@
 # for more info about CODEOWNERS file
 
 # This lists cover the "core" components of vLLM that require careful review
-/vllm/attention @LucasWilkinson
 /vllm/attention/backends/abstract.py @WoosukKwon @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/core @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/engine/llm_engine.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/executor/executor_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @22quinn
 /vllm/worker/worker_base.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @22quinn
-/vllm/model_executor/layers/fused_moe @mgoin
-/vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill @NickLucche
+/vllm/worker/worker.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
+/vllm/model_executor/layers/sampler.py @zhuohan123 @youkaichao @alexm-redhat @comaniac @njhill
 /vllm/model_executor/layers/quantization @mgoin @robertgshaw2-redhat @tlrmchlsmth @yewentao256
 /vllm/model_executor/layers/mamba @tdoublep
 /vllm/model_executor/model_loader @22quinn
-/vllm/multimodal @DarkLight1337 @ywang96 @NickLucche
-/vllm/v1/attention @LucasWilkinson
+/vllm/multimodal @DarkLight1337 @ywang96
 /vllm/v1/sample @22quinn @houseroad
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
-/vllm/reasoning @aarnphm @chaunceyjiang
-/vllm/entrypoints @aarnphm @chaunceyjiang
+/vllm/reasoning @aarnphm
+/vllm/entrypoints @aarnphm
 /vllm/compilation @zou3519 @youkaichao @ProExpertProg
-/vllm/distributed/kv_transfer @NickLucche @ApostaC
 CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 
 # Any change to the VllmConfig changes can have a large user-facing impact,
@@ -30,59 +29,40 @@ CMakeLists.txt @tlrmchlsmth @LucasWilkinson
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
 /vllm/v1/structured_output @mgoin @russellb @aarnphm @benchislett
 /vllm/v1/spec_decode @benchislett @luccafong
-/vllm/v1/attention/backends/flashinfer.py @mgoin
 /vllm/v1/attention/backends/triton_attn.py @tdoublep
-/vllm/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat @heheda12345 @ApostaC
-/vllm/v1/kv_cache_interface.py @heheda12345
-/vllm/v1/offloading @ApostaC
 
 # Test ownership
 /.buildkite/lm-eval-harness @mgoin @simon-mo
+/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
 /tests/distributed/test_multi_node_assignment.py @youkaichao
 /tests/distributed/test_pipeline_parallel.py @youkaichao
 /tests/distributed/test_same_node.py @youkaichao
-/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm @NickLucche
-/tests/evals @mgoin
-/tests/kernels @mgoin @tlrmchlsmth @WoosukKwon @yewentao256
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm
+/tests/kernels @tlrmchlsmth @WoosukKwon @yewentao256
 /tests/models @DarkLight1337 @ywang96
-/tests/multimodal @DarkLight1337 @ywang96 @NickLucche
+/tests/multimodal @DarkLight1337 @ywang96
+/tests/prefix_caching @comaniac @KuntaiDu
 /tests/quantization @mgoin @robertgshaw2-redhat @yewentao256
 /tests/test_inputs.py @DarkLight1337 @ywang96
 /tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
 /tests/v1/structured_output @mgoin @russellb @aarnphm
-/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat @heheda12345 @ApostaC
 /tests/weight_loading @mgoin @youkaichao @yewentao256
 /tests/lora @jeejeelee
 /tests/models/language/generation/test_hybrid.py @tdoublep
-/tests/v1/kv_connector/nixl_integration @NickLucche 
-/tests/v1/kv_connector @ApostaC
-/tests/v1/offloading @ApostaC
-
-# Transformers backend
-/vllm/model_executor/models/transformers.py @hmellor
-/tests/models/test_transformers.py @hmellor
 
 # Docs
-/docs/mkdocs @hmellor
-/docs/**/*.yml @hmellor
-/requirements/docs.txt @hmellor
-.readthedocs.yaml @hmellor
+/docs @hmellor
 mkdocs.yaml @hmellor
 
-# Linting
-.markdownlint.yaml @hmellor
-.pre-commit-config.yaml @hmellor
-/tools/pre_commit @hmellor
-
 # CPU
-/vllm/v1/worker/cpu* @bigPYJ1151
+/vllm/v1/worker/^cpu @bigPYJ1151
 /csrc/cpu @bigPYJ1151
 /vllm/platforms/cpu.py @bigPYJ1151
 /cmake/cpu_extension.cmake @bigPYJ1151
 /docker/Dockerfile.cpu @bigPYJ1151
 
 # Intel GPU
-/vllm/v1/worker/xpu* @jikunshang
+/vllm/v1/worker/^xpu @jikunshang
 /vllm/platforms/xpu.py @jikunshang
 /docker/Dockerfile.xpu @jikunshang
 
@@ -111,12 +91,3 @@ mkdocs.yaml @hmellor
 /vllm/v1/attention/backends/mla/rocm*.py @gshtras
 /vllm/attention/ops/rocm*.py @gshtras
 /vllm/model_executor/layers/fused_moe/rocm*.py @gshtras
-
-# TPU
-/vllm/v1/worker/tpu* @NickLucche
-/vllm/platforms/tpu.py @NickLucche
-/vllm/v1/sample/tpu @NickLucche
-/vllm/tests/v1/tpu @NickLucche
-
-# KVConnector installation files
-/requirements/kv_connectors.txt @NickLucche

.github/ISSUE_TEMPLATE/750-RFC.yml

@@ -43,6 +43,10 @@ body:
       Any other things you would like to mention.
   validations:
     required: false
+- type: markdown
+  attributes:
+    value: >
+      Thanks for contributing 🎉! The vLLM core team hosts a biweekly RFC review session at 9:30AM Pacific Time, while most RFCs can be discussed online, you can optionally sign up for a slot to discuss your RFC online [here](https://docs.google.com/document/d/1CiLVBZeIVfR7_PNAKVSusxpceywkoOOB78qoWqHvSZc/edit).
 - type: checkboxes
   id: askllm
   attributes:

.github/mergify.yml

@@ -124,16 +124,9 @@ pull_request_rules:
     - or:
       - files~=^examples/.*gpt[-_]?oss.*\.py
       - files~=^tests/.*gpt[-_]?oss.*\.py
-      - files~=^tests/entrypoints/openai/test_response_api_with_harmony.py
-      - files~=^tests/entrypoints/test_context.py
       - files~=^vllm/model_executor/models/.*gpt[-_]?oss.*\.py
       - files~=^vllm/model_executor/layers/.*gpt[-_]?oss.*\.py
-      - files~=^vllm/entrypoints/harmony_utils.py
-      - files~=^vllm/entrypoints/tool_server.py
-      - files~=^vllm/entrypoints/tool.py
-      - files~=^vllm/entrypoints/context.py
       - title~=(?i)gpt[-_]?oss
-      - title~=(?i)harmony
   actions:
     label:
       add:
@@ -171,7 +164,7 @@ pull_request_rules:
       - files=examples/online_serving/openai_chat_completion_structured_outputs.py
       - files=examples/online_serving/openai_chat_completion_structured_outputs_with_reasoning.py
       - files~=^tests/v1/structured_output/
-      - files=tests/v1/entrypoints/llm/test_struct_output_generate.py
+      - files=tests/v1/entrypoints/llm/test_guided_generate.py
       - files~=^vllm/v1/structured_output/
   actions:
     label:
@@ -302,20 +295,3 @@ pull_request_rules:
     label:
       remove:
         - needs-rebase
-
-- name: label-kv-connector
-  description: Automatically apply kv-connector label
-  conditions:
-    - or:
-      - files~=^examples/online_serving/disaggregated[^/]*/.*
-      - files~=^examples/offline_inference/disaggregated[^/]*/.*
-      - files~=^examples/others/lmcache/
-      - files~=^tests/v1/kv_connector/
-      - files~=^vllm/distributed/kv_transfer/
-      - title~=(?i)\bP/?D\b
-      - title~=(?i)NIXL
-      - title~=(?i)LMCache
-  actions:
-    label:
-      add:
-        - kv-connector

.github/workflows/add_label_automerge.yml

@@ -10,7 +10,7 @@ jobs:
         runs-on: ubuntu-latest
         steps:
             -   name: Add label
-                uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+                uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
                 with:
                     script: |
                         github.rest.issues.addLabels({

.github/workflows/bc-lint.yml

@@ -1,29 +0,0 @@
-name: BC Lint
-
-on:
-  pull_request:
-    types:
-      - opened
-      - synchronize
-      - reopened
-      - labeled
-      - unlabeled
-
-jobs:
-  bc_lint:
-    if: github.repository_owner == 'vllm-project'
-    runs-on: ubuntu-latest
-    steps:
-      - name: Run BC Lint Action
-        uses: pytorch/test-infra/.github/actions/bc-lint@main
-        with:
-          repo: ${{ github.event.pull_request.head.repo.full_name }}
-          base_sha: ${{ github.event.pull_request.base.sha }}
-          head_sha: ${{ github.event.pull_request.head.sha }}
-          suppression: ${{ contains(github.event.pull_request.labels.*.name, 'suppress-bc-linter') }}
-          docs_link: 'https://github.com/pytorch/test-infra/wiki/BC-Linter'
-          config_dir: .github
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}
-  cancel-in-progress: true

.github/workflows/issue_autolabel.yml

@@ -13,7 +13,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Label issues based on keywords
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd  # v8.0.0
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea  # v7.0.1
         with:
           script: |
             // Configuration: Add new labels and keywords here

.github/workflows/reminder_comment.yml

@@ -9,7 +9,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Remind to run full CI on PR
-        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8.0.0
+        uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
         with:
           script: |
             try {

.github/workflows/stale.yml

@@ -13,7 +13,7 @@ jobs:
       actions: write
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/stale@3a9db7e6a41a89f618792c92c0e97cc736e1b13f # v10.0.0
+      - uses: actions/stale@5bef64f19d7facfb25b37b414482c7164d639639 # v9.1.0
         with:
           # Increasing this value ensures that changes to this workflow
           # propagate to all issues and PRs in days rather than months

.gitignore

@@ -4,7 +4,7 @@
 # vllm-flash-attn built from source
 vllm/vllm_flash_attn/*
 
-# triton jit
+# triton jit 
 .triton
 
 # Byte-compiled / optimized / DLL files
@@ -177,14 +177,6 @@ cython_debug/
 # VSCode
 .vscode/
 
-# Claude
-CLAUDE.md
-.claude/
-
-# Codex
-AGENTS.md
-.codex/
-
 # DS Store
 .DS_Store
 
@@ -217,4 +209,4 @@ shellcheck*/
 csrc/moe/marlin_moe_wna16/kernel_*
 
 # Ignore ep_kernels_workspace folder
-ep_kernels_workspace/
+ep_kernels_workspace/

.pre-commit-config.yaml

@@ -49,7 +49,7 @@ repos:
   rev: 0.6.17
   hooks:
     - id: pip-compile
-      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128, --python-platform, x86_64-manylinux_2_28]
+      args: [requirements/test.in, -o, requirements/test.txt, --index-strategy, unsafe-best-match, --torch-backend, cu128]
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
@@ -60,32 +60,38 @@ repos:
     files: ^requirements/test\.(in|txt)$
   - id: mypy-local
     name: Run mypy for local Python installation
-    entry: python tools/pre_commit/mypy.py 0 "local"
+    entry: tools/mypy.sh 0 "local"
+    language: python
+    types: [python]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, types-setuptools, types-PyYAML, types-requests, pydantic]
     stages: [pre-commit] # Don't run in CI
-    <<: &mypy_common
-      language: python
-      types_or: [python, pyi]
-      require_serial: true
-      additional_dependencies: [mypy==1.11.1, regex, types-cachetools, types-setuptools, types-PyYAML, types-requests, types-torch, pydantic]
   - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.9
-    entry: python tools/pre_commit/mypy.py 1 "3.9"
-    <<: *mypy_common
+    entry: tools/mypy.sh 1 "3.9"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
   - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.10
-    entry: python tools/pre_commit/mypy.py 1 "3.10"
-    <<: *mypy_common
+    entry: tools/mypy.sh 1 "3.10"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
   - id: mypy-3.11 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.11
-    entry: python tools/pre_commit/mypy.py 1 "3.11"
-    <<: *mypy_common
+    entry: tools/mypy.sh 1 "3.11"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
   - id: mypy-3.12 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.12
-    entry: python tools/pre_commit/mypy.py 1 "3.12"
-    <<: *mypy_common
+    entry: tools/mypy.sh 1 "3.12"
+    language: python
+    types: [python]
+    additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
   - id: shellcheck
     name: Lint shell scripts
@@ -149,15 +155,18 @@ repos:
     additional_dependencies: [regex]
   - id: check-pickle-imports
     name: Prevent new pickle/cloudpickle imports
-    entry: python tools/pre_commit/check_pickle_imports.py
+    entry: python tools/check_pickle_imports.py
     language: python
     types: [python]
-    additional_dependencies: [regex]
+    pass_filenames: false
+    additional_dependencies: [pathspec, regex]
   - id: validate-config
     name: Validate configuration has default values and that each field has a docstring
     entry: python tools/validate_config.py
     language: python
-    additional_dependencies: [regex]
+    types: [python]
+    pass_filenames: true
+    files: vllm/config.py|tests/test_config.py|vllm/entrypoints/openai/cli_args.py
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

.readthedocs.yaml

@@ -13,7 +13,6 @@ build:
 
 mkdocs:
   configuration: mkdocs.yaml
-  fail_on_warning: true
 
 # Optionally declare the Python requirements required to build your docs
 python:

.yapfignore

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

CMakeLists.txt

@@ -13,10 +13,6 @@ cmake_minimum_required(VERSION 3.26)
 # cmake --install . --component _C
 project(vllm_extensions LANGUAGES CXX)
 
-set(CMAKE_CXX_STANDARD 17)
-set(CMAKE_CXX_STANDARD_REQUIRED ON)
-
-
 # CUDA by default, can be overridden by using -DVLLM_TARGET_DEVICE=... (used by setup.py)
 set(VLLM_TARGET_DEVICE "cuda" CACHE STRING "Target device backend for vLLM")
 message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
@@ -175,16 +171,6 @@ if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
   list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
 endif()
 
-#
-# Set CUDA include flags for CXX compiler.
-#
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -I${CUDA_TOOLKIT_ROOT_DIR}/include")
-  if(CUDA_VERSION VERSION_GREATER_EQUAL 13.0)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -I${CUDA_TOOLKIT_ROOT_DIR}/include/cccl")
-  endif()
-endif()
-
 #
 # Use FetchContent for C++ dependencies that are compiled as part of vLLM's build process.
 # setup.py will override FETCHCONTENT_BASE_DIR to play nicely with sccache.
@@ -308,6 +294,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu"
     "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
     "csrc/cutlass_extensions/common.cpp"
+    "csrc/attention/mla/cutlass_mla_entry.cu"
     "csrc/quantization/fp8/per_token_group_quant.cu")
 
   set_gencode_flags_for_srcs(
@@ -594,6 +581,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   cuda_archs_loose_intersection(MLA_ARCHS "10.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.8 AND MLA_ARCHS)
     set(SRCS
+      "csrc/attention/mla/cutlass_mla_kernels.cu"
       "csrc/attention/mla/sm100_cutlass_mla_kernel.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
@@ -791,17 +779,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  # Hadacore kernels
-  cuda_archs_loose_intersection(HADACORE_ARCHS "8.0;8.9;9.0" "${CUDA_ARCHS}")
-  if(HADACORE_ARCHS)
-    set(SRCS "csrc/quantization/hadamard/hadacore/hadamard_transform_cuda.cu")
-    set_gencode_flags_for_srcs(
-      SRCS "${SRCS}"
-      CUDA_ARCHS "${HADACORE_ARCHS}")
-    list(APPEND VLLM_EXT_SRC "${SRCS}")
-    message(STATUS "Building hadacore")
-  endif()
-
 # if CUDA endif
 endif()
 

README.md

@@ -14,9 +14,6 @@ Easy, fast, and cheap LLM serving for everyone
 | <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://blog.vllm.ai/"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
----
-Join us at the [PyTorch Conference, October 22-23](https://events.linuxfoundation.org/pytorch-conference/) and [Ray Summit, November 3-5](https://www.anyscale.com/ray-summit/2025) in San Francisco for our latest updates on vLLM and to meet the vLLM team! Register now for the largest vLLM community events of the year!
-
 ---
 
 *Latest News* 🔥
@@ -81,7 +78,7 @@ vLLM is flexible and easy to use with:
 - Tensor, pipeline, data and expert parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support for NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, and TPU. Additionally, support for diverse hardware plugins such as Intel Gaudi, IBM Spyre and Huawei Ascend.
+- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron
 - Prefix caching support
 - Multi-LoRA support
 

benchmarks/README.md

@@ -1,20 +1,807 @@
-# Benchmarks
+# Benchmarking vLLM
 
-This directory used to contain vLLM's benchmark scripts and utilities for performance testing and evaluation.
+This README guides you through running benchmark tests with the extensive
+datasets supported on vLLM. It’s a living document, updated as new features and datasets
+become available.
 
-## Contents
+## Dataset Overview
 
-- **Serving benchmarks**: Scripts for testing online inference performance (latency, throughput)
-- **Throughput benchmarks**: Scripts for testing offline batch inference performance
-- **Specialized benchmarks**: Tools for testing specific features like structured output, prefix caching, long document QA, request prioritization, and multi-modal inference
-- **Dataset utilities**: Framework for loading and sampling from various benchmark datasets (ShareGPT, HuggingFace datasets, synthetic data, etc.)
+<table style="width:100%; border-collapse: collapse;">
+  <thead>
+    <tr>
+      <th style="width:15%; text-align: left;">Dataset</th>
+      <th style="width:10%; text-align: center;">Online</th>
+      <th style="width:10%; text-align: center;">Offline</th>
+      <th style="width:65%; text-align: left;">Data Path</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td><strong>ShareGPT</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json</code></td>
+    </tr>
+    <tr>
+      <td><strong>ShareGPT4V (Image)</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>
+        <code>wget https://huggingface.co/datasets/Lin-Chen/ShareGPT4V/blob/main/sharegpt4v_instruct_gpt4-vision_cap100k.json</code>
+        <br>
+        <div>Note that the images need to be downloaded separately. For example, to download COCO's 2017 Train images:</div>
+        <code>wget http://images.cocodataset.org/zips/train2017.zip</code>
+      </td>
+    </tr>
+        <tr>
+      <td><strong>ShareGPT4Video (Video)</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>
+        <code>git clone https://huggingface.co/datasets/ShareGPT4Video/ShareGPT4Video</code>
+      </td>
+    </tr>
+    <tr>
+      <td><strong>BurstGPT</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv</code></td>
+    </tr>
+    <tr>
+      <td><strong>Sonnet (deprecated)</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>Local file: <code>benchmarks/sonnet.txt</code></td>
+    </tr>
+    <tr>
+      <td><strong>Random</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>synthetic</code></td>
+    </tr>
+    <tr>
+      <td><strong>RandomMultiModal (Image/Video)</strong></td>
+      <td style="text-align: center;">🟡</td>
+      <td style="text-align: center;">🚧</td>
+      <td><code>synthetic</code> </td>
+    </tr>
+    <tr>
+      <td><strong>Prefix Repetition</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>synthetic</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-VisionArena</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmarena-ai/VisionArena-Chat</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-InstructCoder</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>likaixin/InstructCoder</code></td>
+    </tr>
+      <tr>
+      <td><strong>HuggingFace-AIMO</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>AI-MO/aimo-validation-aime</code> , <code>AI-MO/NuminaMath-1.5</code>, <code>AI-MO/NuminaMath-CoT</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-Other</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmms-lab/LLaVA-OneVision-Data</code>, <code>Aeala/ShareGPT_Vicuna_unfiltered</code></td>
+    </tr>
+    <tr>
+      <td><strong>Custom</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>Local file: <code>data.jsonl</code></td>
+    </tr>
+  </tbody>
+</table>
 
-## Usage
+✅: supported
 
-For detailed usage instructions, examples, and dataset information, see the [Benchmark CLI documentation](https://docs.vllm.ai/en/latest/contributing/benchmarks.html#benchmark-cli).
+🟡: Partial support
 
-For full CLI reference see:
+🚧: to be supported
 
-- <https://docs.vllm.ai/en/latest/cli/bench/latency.html>
-- <https://docs.vllm.ai/en/latest/cli/bench/serve.html>
-- <https://docs.vllm.ai/en/latest/cli/bench/throughput.html>
+**Note**: HuggingFace dataset's `dataset-name` should be set to `hf`.
+For local `dataset-path`, please set `hf-name` to its Hugging Face ID like
+
+```bash
+--dataset-path /datasets/VisionArena-Chat/ --hf-name lmarena-ai/VisionArena-Chat
+```
+
+## 🚀 Example - Online Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+First start serving your model
+
+```bash
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B
+```
+
+Then run the benchmarking script
+
+```bash
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+vllm bench serve \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --num-prompts 10
+```
+
+If successful, you will see the following output
+
+```text
+============ Serving Benchmark Result ============
+Successful requests:                     10
+Benchmark duration (s):                  5.78
+Total input tokens:                      1369
+Total generated tokens:                  2212
+Request throughput (req/s):              1.73
+Output token throughput (tok/s):         382.89
+Total Token throughput (tok/s):          619.85
+---------------Time to First Token----------------
+Mean TTFT (ms):                          71.54
+Median TTFT (ms):                        73.88
+P99 TTFT (ms):                           79.49
+-----Time per Output Token (excl. 1st token)------
+Mean TPOT (ms):                          7.91
+Median TPOT (ms):                        7.96
+P99 TPOT (ms):                           8.03
+---------------Inter-token Latency----------------
+Mean ITL (ms):                           7.74
+Median ITL (ms):                         7.70
+P99 ITL (ms):                            8.39
+==================================================
+```
+
+### Custom Dataset
+
+If the dataset you want to benchmark is not supported yet in vLLM, even then you can benchmark on it using `CustomDataset`. Your data needs to be in `.jsonl` format and needs to have "prompt" field per entry, e.g., data.jsonl
+
+```json
+{"prompt": "What is the capital of India?"}
+{"prompt": "What is the capital of Iran?"}
+{"prompt": "What is the capital of China?"}
+```
+
+```bash
+# start server
+VLLM_USE_V1=1 vllm serve meta-llama/Llama-3.1-8B-Instruct
+```
+
+```bash
+# run benchmarking script
+vllm bench serve --port 9001 --save-result --save-detailed \
+  --backend vllm \
+  --model meta-llama/Llama-3.1-8B-Instruct \
+  --endpoint /v1/completions \
+  --dataset-name custom \
+  --dataset-path <path-to-your-data-jsonl> \
+  --custom-skip-chat-template \
+  --num-prompts 80 \
+  --max-concurrency 1 \
+  --temperature=0.3 \
+  --top-p=0.75 \
+  --result-dir "./log/"
+```
+
+You can skip applying chat template if your data already has it by using `--custom-skip-chat-template`.
+
+### VisionArena Benchmark for Vision Language Models
+
+```bash
+# need a model with vision capability here
+vllm serve Qwen/Qwen2-VL-7B-Instruct
+```
+
+```bash
+vllm bench serve \
+  --backend openai-chat \
+  --endpoint-type openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --hf-split train \
+  --num-prompts 1000
+```
+
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
+```
+
+``` bash
+vllm bench serve \
+    --model meta-llama/Meta-Llama-3-8B-Instruct \
+    --dataset-name hf \
+    --dataset-path likaixin/InstructCoder \
+    --num-prompts 2048
+```
+
+### Other HuggingFaceDataset Examples
+
+```bash
+vllm serve Qwen/Qwen2-VL-7B-Instruct
+```
+
+`lmms-lab/LLaVA-OneVision-Data`:
+
+```bash
+vllm bench serve \
+  --backend openai-chat \
+  --endpoint-type openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+`Aeala/ShareGPT_Vicuna_unfiltered`:
+
+```bash
+vllm bench serve \
+  --backend openai-chat \
+  --endpoint-type openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+`AI-MO/aimo-validation-aime`:
+
+``` bash
+vllm bench serve \
+    --model Qwen/QwQ-32B \
+    --dataset-name hf \
+    --dataset-path AI-MO/aimo-validation-aime \
+    --num-prompts 10 \
+    --seed 42
+```
+
+`philschmid/mt-bench`:
+
+``` bash
+vllm bench serve \
+    --model Qwen/QwQ-32B \
+    --dataset-name hf \
+    --dataset-path philschmid/mt-bench \
+    --num-prompts 80
+```
+
+### Running With Sampling Parameters
+
+When using OpenAI-compatible backends such as `vllm`, optional sampling
+parameters can be specified. Example client command:
+
+```bash
+vllm bench serve \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --top-k 10 \
+  --top-p 0.9 \
+  --temperature 0.5 \
+  --num-prompts 10
+```
+
+### Running With Ramp-Up Request Rate
+
+The benchmark tool also supports ramping up the request rate over the
+duration of the benchmark run. This can be useful for stress testing the
+server or finding the maximum throughput that it can handle, given some latency budget.
+
+Two ramp-up strategies are supported:
+
+- `linear`: Increases the request rate linearly from a start value to an end value.
+- `exponential`: Increases the request rate exponentially.
+
+The following arguments can be used to control the ramp-up:
+
+- `--ramp-up-strategy`: The ramp-up strategy to use (`linear` or `exponential`).
+- `--ramp-up-start-rps`: The request rate at the beginning of the benchmark.
+- `--ramp-up-end-rps`: The request rate at the end of the benchmark.
+
+</details>
+
+## 📈 Example - Offline Throughput Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+```bash
+vllm bench throughput \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset-name sonnet \
+  --dataset-path vllm/benchmarks/sonnet.txt \
+  --num-prompts 10
+```
+
+If successful, you will see the following output
+
+```text
+Throughput: 7.15 requests/s, 4656.00 total tokens/s, 1072.15 output tokens/s
+Total num prompt tokens:  5014
+Total num output tokens:  1500
+```
+
+### VisionArena Benchmark for Vision Language Models
+
+```bash
+vllm bench throughput \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --num-prompts 1000 \
+  --hf-split train
+```
+
+The `num prompt tokens` now includes image token counts
+
+```text
+Throughput: 2.55 requests/s, 4036.92 total tokens/s, 326.90 output tokens/s
+Total num prompt tokens:  14527
+Total num output tokens:  1280
+```
+
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+VLLM_USE_V1=1 \
+vllm bench throughput \
+    --dataset-name=hf \
+    --dataset-path=likaixin/InstructCoder \
+    --model=meta-llama/Meta-Llama-3-8B-Instruct \
+    --input-len=1000 \
+    --output-len=100 \
+    --num-prompts=2048 \
+    --async-engine \
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
+```
+
+```text
+Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s
+Total num prompt tokens:  261136
+Total num output tokens:  204800
+```
+
+### Other HuggingFaceDataset Examples
+
+`lmms-lab/LLaVA-OneVision-Data`:
+
+```bash
+vllm bench throughput \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+`Aeala/ShareGPT_Vicuna_unfiltered`:
+
+```bash
+vllm bench throughput \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+`AI-MO/aimo-validation-aime`:
+
+```bash
+vllm bench throughput \
+  --model Qwen/QwQ-32B \
+  --backend vllm \
+  --dataset-name hf \
+  --dataset-path AI-MO/aimo-validation-aime \
+  --hf-split train \
+  --num-prompts 10
+```
+
+Benchmark with LoRA adapters:
+
+``` bash
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+vllm bench throughput \
+  --model meta-llama/Llama-2-7b-hf \
+  --backend vllm \
+  --dataset_path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --dataset_name sharegpt \
+  --num-prompts 10 \
+  --max-loras 2 \
+  --max-lora-rank 8 \
+  --enable-lora \
+  --lora-path yard1/llama-2-7b-sql-lora-test
+  ```
+
+</details>
+
+## 🛠️ Example - Structured Output Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+Benchmark the performance of structured output generation (JSON, grammar, regex).
+
+### Server Setup
+
+```bash
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B
+```
+
+### JSON Schema Benchmark
+
+```bash
+python3 benchmarks/benchmark_serving_structured_output.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset json \
+  --structured-output-ratio 1.0 \
+  --request-rate 10 \
+  --num-prompts 1000
+```
+
+### Grammar-based Generation Benchmark
+
+```bash
+python3 benchmarks/benchmark_serving_structured_output.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset grammar \
+  --structure-type grammar \
+  --request-rate 10 \
+  --num-prompts 1000
+```
+
+### Regex-based Generation Benchmark
+
+```bash
+python3 benchmarks/benchmark_serving_structured_output.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset regex \
+  --request-rate 10 \
+  --num-prompts 1000
+```
+
+### Choice-based Generation Benchmark
+
+```bash
+python3 benchmarks/benchmark_serving_structured_output.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset choice \
+  --request-rate 10 \
+  --num-prompts 1000
+```
+
+### XGrammar Benchmark Dataset
+
+```bash
+python3 benchmarks/benchmark_serving_structured_output.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset xgrammar_bench \
+  --request-rate 10 \
+  --num-prompts 1000
+```
+
+</details>
+
+## 📚 Example - Long Document QA Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+Benchmark the performance of long document question-answering with prefix caching.
+
+### Basic Long Document QA Test
+
+```bash
+python3 benchmarks/benchmark_long_document_qa_throughput.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --enable-prefix-caching \
+  --num-documents 16 \
+  --document-length 2000 \
+  --output-len 50 \
+  --repeat-count 5
+```
+
+### Different Repeat Modes
+
+```bash
+# Random mode (default) - shuffle prompts randomly
+python3 benchmarks/benchmark_long_document_qa_throughput.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --enable-prefix-caching \
+  --num-documents 8 \
+  --document-length 3000 \
+  --repeat-count 3 \
+  --repeat-mode random
+
+# Tile mode - repeat entire prompt list in sequence
+python3 benchmarks/benchmark_long_document_qa_throughput.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --enable-prefix-caching \
+  --num-documents 8 \
+  --document-length 3000 \
+  --repeat-count 3 \
+  --repeat-mode tile
+
+# Interleave mode - repeat each prompt consecutively
+python3 benchmarks/benchmark_long_document_qa_throughput.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --enable-prefix-caching \
+  --num-documents 8 \
+  --document-length 3000 \
+  --repeat-count 3 \
+  --repeat-mode interleave
+```
+
+</details>
+
+## 🗂️ Example - Prefix Caching Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+Benchmark the efficiency of automatic prefix caching.
+
+### Fixed Prompt with Prefix Caching
+
+```bash
+python3 benchmarks/benchmark_prefix_caching.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --enable-prefix-caching \
+  --num-prompts 1 \
+  --repeat-count 100 \
+  --input-length-range 128:256
+```
+
+### ShareGPT Dataset with Prefix Caching
+
+```bash
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+
+python3 benchmarks/benchmark_prefix_caching.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --dataset-path /path/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --enable-prefix-caching \
+  --num-prompts 20 \
+  --repeat-count 5 \
+  --input-length-range 128:256
+```
+
+### Prefix Repetition Dataset
+
+```bash
+vllm bench serve \
+  --backend openai \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --dataset-name prefix_repetition \
+  --num-prompts 100 \
+  --prefix-repetition-prefix-len 512 \
+  --prefix-repetition-suffix-len 128 \
+  --prefix-repetition-num-prefixes 5 \
+  --prefix-repetition-output-len 128
+```
+
+</details>
+
+## ⚡ Example - Request Prioritization Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+Benchmark the performance of request prioritization in vLLM.
+
+### Basic Prioritization Test
+
+```bash
+python3 benchmarks/benchmark_prioritization.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --input-len 128 \
+  --output-len 64 \
+  --num-prompts 100 \
+  --scheduling-policy priority
+```
+
+### Multiple Sequences per Prompt
+
+```bash
+python3 benchmarks/benchmark_prioritization.py \
+  --model meta-llama/Llama-2-7b-chat-hf \
+  --input-len 128 \
+  --output-len 64 \
+  --num-prompts 100 \
+  --scheduling-policy priority \
+  --n 2
+```
+
+</details>
+
+## 👁️ Example - Multi-Modal Benchmark
+
+<details>
+<summary>Show more</summary>
+
+<br/>
+
+Benchmark the performance of multi-modal requests in vLLM.
+
+### Images (ShareGPT4V)
+
+Start vLLM:
+
+```bash
+python -m vllm.entrypoints.openai.api_server \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dtype bfloat16 \
+  --limit-mm-per-prompt '{"image": 1}' \
+  --allowed-local-media-path /path/to/sharegpt4v/images
+```
+
+Send requests with images:
+
+```bash
+python benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dataset-name sharegpt \
+  --dataset-path /path/to/ShareGPT4V/sharegpt4v_instruct_gpt4-vision_cap100k.json \
+  --num-prompts 100 \
+  --save-result \
+  --result-dir ~/vllm_benchmark_results \
+  --save-detailed \
+  --endpoint /v1/chat/completion
+```
+
+### Videos (ShareGPT4Video)
+
+Start vLLM:
+
+```bash
+python -m vllm.entrypoints.openai.api_server \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dtype bfloat16 \
+  --limit-mm-per-prompt '{"video": 1}' \
+  --allowed-local-media-path /path/to/sharegpt4video/videos
+```
+
+Send requests with videos:
+
+```bash
+python benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2.5-VL-7B-Instruct \
+  --dataset-name sharegpt \
+  --dataset-path /path/to/ShareGPT4Video/llava_v1_5_mix665k_with_video_chatgpt72k_share4video28k.json \
+  --num-prompts 100 \
+  --save-result \
+  --result-dir ~/vllm_benchmark_results \
+  --save-detailed \
+  --endpoint /v1/chat/completion
+```
+
+### Synthetic Random Images (random-mm)
+
+Generate synthetic image inputs alongside random text prompts to stress-test vision models without external datasets.
+
+Notes:
+
+- Works only with online benchmark via the OpenAI  backend (`--backend openai-chat`) and endpoint `/v1/chat/completions`.
+- Video sampling is not yet implemented.
+
+Start the server (example):
+
+```bash
+vllm serve Qwen/Qwen2.5-VL-3B-Instruct \
+  --dtype bfloat16 \
+  --max-model-len 16384 \
+  --limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --mm-processor-kwargs max_pixels=1003520
+```
+
+Benchmark. It is recommended to use the flag `--ignore-eos` to simulate real responses. You can set the size of the output via the arg `random-output-len`.
+
+Ex.1: Fixed number of items and a single image resolution, enforcing generation of approx 40 tokens:
+
+```bash
+vllm bench serve \
+  --backend openai-chat \
+  --model Qwen/Qwen2.5-VL-3B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name random-mm \
+  --num-prompts 100 \
+  --max-concurrency 10 \
+  --random-prefix-len 25 \
+  --random-input-len 300 \
+  --random-output-len 40 \
+  --random-range-ratio 0.2 \
+  --random-mm-base-items-per-request 2 \
+  --random-mm-limit-mm-per-prompt '{"image": 3, "video": 0}' \
+  --random-mm-bucket-config '{(224, 224, 1): 1.0}' \
+  --request-rate inf \
+  --ignore-eos \
+  --seed 42
+```
+
+The number of items per request can be controlled by passing multiple image buckets:
+
+```bash
+  --random-mm-base-items-per-request 2 \
+  --random-mm-num-mm-items-range-ratio 0.5 \
+  --random-mm-limit-mm-per-prompt '{"image": 4, "video": 0}' \
+  --random-mm-bucket-config '{(256, 256, 1): 0.7, (720, 1280, 1): 0.3}' \
+```
+
+Flags specific to `random-mm`:
+
+- `--random-mm-base-items-per-request`: base number of multimodal items per request.
+- `--random-mm-num-mm-items-range-ratio`: vary item count uniformly in the closed integer range [floor(n·(1−r)), ceil(n·(1+r))]. Set r=0 to keep it fixed; r=1 allows 0 items.
+- `--random-mm-limit-mm-per-prompt`: per-modality hard caps, e.g. '{"image": 3, "video": 0}'.
+- `--random-mm-bucket-config`: dict mapping (H, W, T) → probability. Entries with probability 0 are removed; remaining probabilities are renormalized to sum to 1. Use T=1 for images. Set any T>1 for videos (video sampling not yet supported).
+
+Behavioral notes:
+
+- If the requested base item count cannot be satisfied under the provided per-prompt limits, the tool raises an error rather than silently clamping.
+
+How sampling works:
+
+- Determine per-request item count k by sampling uniformly from the integer range defined by `--random-mm-base-items-per-request` and `--random-mm-num-mm-items-range-ratio`, then clamp k to at most the sum of per-modality limits.
+- For each of the k items, sample a bucket (H, W, T) according to the normalized probabilities in `--random-mm-bucket-config`, while tracking how many items of each modality have been added.
+- If a modality (e.g., image) reaches its limit from `--random-mm-limit-mm-per-prompt`, all buckets of that modality are excluded and the remaining bucket probabilities are renormalized before continuing.
+This should be seen as an edge case, and if this behavior can be avoided by setting `--random-mm-limit-mm-per-prompt` to a large number. Note that this might result in errors due to engine config `--limit-mm-per-prompt`.
+- The resulting request contains synthetic image data in `multi_modal_data` (OpenAI Chat format). When `random-mm` is used with the OpenAI Chat backend, prompts remain text and MM content is attached via `multi_modal_data`.
+
+</details>

benchmarks/auto_tune/README.md

@@ -149,70 +149,3 @@ The script follows a systematic process to find the optimal parameters:
 4. **Track Best Result**: Throughout the process, the script tracks the parameter combination that has yielded the highest valid throughput so far.
 
 5. **Profile Collection**: For the best-performing run, the script saves the vLLM profiler output, which can be used for deep-dive performance analysis with tools like TensorBoard.
-
-## Batched `auto_tune`
-
-The `batch_auto_tune.sh` script allows you to run multiple `auto_tune.sh` experiments sequentially from a single configuration file. It iterates through a list of parameter sets, executes `auto_tune.sh` for each, and records the results back into the input file.
-
-### Prerequisites
-
-- **jq**: This script requires `jq` to parse the JSON configuration file.
-- **gcloud**: If you plan to upload results to Google Cloud Storage, the `gcloud` CLI must be installed and authenticated.
-
-### How to Run
-
-1. **Create a JSON configuration file**: Create a file (e.g., `runs_config.json`) containing an array of JSON objects. Each object defines the parameters for a single `auto_tune.sh` run.
-
-2. **Execute the script**:
-
-    ```bash
-    bash batch_auto_tune.sh <path_to_json_file> [gcs_upload_path]
-    ```
-
-    - `<path_to_json_file>`: **Required.** Path to your JSON configuration file.
-    - `[gcs_upload_path]`: **Optional.** A GCS path (e.g., `gs://my-bucket/benchmark-results`) where the detailed results and profiles for each run will be uploaded. If this is empty, the results will be available on the local filesystem (see the log for `RESULT_FILE=/path/to/results/file.txt`).
-
-### Configuration File
-
-The JSON configuration file should contain an array of objects. Each object's keys correspond to the configuration variables for `auto_tune.sh` (see the [Configuration table above](#configuration)). These keys will be converted to uppercase environment variables for each run.
-
-Here is an example `runs_config.json` with two benchmark configurations:
-
-```json
-[
-  {
-    "base": "/home/user",
-    "model": "meta-llama/Llama-3.1-8B-Instruct",
-    "system": "TPU", # OR GPU
-    "tp": 8,
-    "input_len": 128,
-    "output_len": 2048,
-    "max_model_len": 2300,
-    "num_seqs_list": "128 256",
-    "num_batched_tokens_list": "8192 16384"
-  },
-  {
-    "base": "/home/user",
-    "model": "meta-llama/Llama-3.1-70B-Instruct",
-    "system": "TPU", # OR GPU
-    "tp": 8,
-    "input_len": 4000,
-    "output_len": 16,
-    "max_model_len": 4096,
-    "num_seqs_list": "64 128",
-    "num_batched_tokens_list": "4096 8192",
-    "max_latency_allowed_ms": 500
-  }
-]
-```
-
-### Output
-
-The script modifies the input JSON file in place, adding the results of each run to the corresponding object. The following fields are added:
-
-- `run_id`: A unique identifier for the run, derived from the timestamp.
-- `status`: The outcome of the run (`SUCCESS`, `FAILURE`, or `WARNING_NO_RESULT_FILE`).
-- `results`: The content of the `result.txt` file from the `auto_tune.sh` run.
-- `gcs_results`: The GCS URL where the run's artifacts are stored (if a GCS path was provided).
-
-A summary of successful and failed runs is also printed to the console upon completion.

benchmarks/auto_tune/auto_tune.sh

@@ -103,15 +103,10 @@ start_server() {
         VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 \
             vllm serve "${common_args_array[@]}" > "$vllm_log" 2>&1 &
     fi
-    local server_pid=$!
 
     # wait for 10 minutes...
     server_started=0
     for i in {1..60}; do
-        # This line checks whether the server is still alive or not,
-        # since that we should always have permission to send signal to the server process.
-        kill -0 $server_pid 2> /dev/null || break
-
         RESPONSE=$(curl -s -X GET "http://0.0.0.0:8004/health" -w "%{http_code}" -o /dev/stdout)
         STATUS_CODE=$(echo "$RESPONSE" | tail -n 1)
         if [[ "$STATUS_CODE" -eq 200 ]]; then
@@ -123,7 +118,7 @@ start_server() {
     done
 
     if (( ! server_started )); then
-        echo "server did not start within 10 minutes or crashed. Please check server log at $vllm_log".
+        echo "server did not start within 10 minutes. Please check server log at $vllm_log".
         return 1
     else
         return 0

benchmarks/auto_tune/batch_auto_tune.sh

@@ -1,128 +0,0 @@
-#!/bin/bash
-
-INPUT_JSON="$1"
-GCS_PATH="$2" # Optional GCS path for uploading results for each run
-
-SCRIPT_DIR=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)
-AUTOTUNE_SCRIPT="$SCRIPT_DIR/auto_tune.sh"
-
-if [[ -z "$INPUT_JSON" ]]; then
-  echo "Error: Input JSON file not provided."
-  echo "Usage: $0 <path_to_json_file> [gcs_upload_path]"
-  exit 1
-fi
-
-if [[ ! -f "$INPUT_JSON" ]]; then
-  echo "Error: File not found at '$INPUT_JSON'"
-  exit 1
-fi
-
-if ! command -v jq &> /dev/null; then
-    echo "Error: 'jq' command not found. Please install jq to process the JSON input."
-    exit 1
-fi
-
-if [[ -n "$GCS_PATH" ]] && ! command -v gcloud &> /dev/null; then
-    echo "Error: 'gcloud' command not found, but a GCS_PATH was provided."
-    exit 1
-fi
-
-SUCCESS_COUNT=0
-FAILURE_COUNT=0
-FAILED_RUNS=()
-SCRIPT_START_TIME=$(date +%s)
-
-json_content=$(cat "$INPUT_JSON")
-if ! num_runs=$(echo "$json_content" | jq 'length'); then
-  echo "Error: Invalid JSON in $INPUT_JSON. 'jq' failed to get array length." >&2
-  exit 1
-fi
-
-echo "Found $num_runs benchmark configurations in $INPUT_JSON."
-echo "Starting benchmark runs..."
-echo "--------------------------------------------------"
-
-for i in $(seq 0 $(($num_runs - 1))); do
-  run_object=$(echo "$json_content" | jq ".[$i]")
-
-  RUN_START_TIME=$(date +%s)
-  ENV_VARS_ARRAY=()
-  # Dynamically create env vars from the JSON object's keys
-  for key in $(echo "$run_object" | jq -r 'keys_unsorted[]'); do
-    value=$(echo "$run_object" | jq -r ".$key")
-    var_name=$(echo "$key" | tr '[:lower:]' '[:upper:]' | tr -cd 'A-Z0-9_')
-    ENV_VARS_ARRAY+=("${var_name}=${value}")
-  done
-
-  echo "Executing run #$((i+1))/$num_runs with parameters: ${ENV_VARS_ARRAY[*]}"
-
-  # Execute auto_tune.sh and capture output
-  RUN_OUTPUT_FILE=$(mktemp)
-  if env "${ENV_VARS_ARRAY[@]}" bash "$AUTOTUNE_SCRIPT" > >(tee -a "$RUN_OUTPUT_FILE") 2>&1; then
-    STATUS="SUCCESS"
-    ((SUCCESS_COUNT++))
-  else
-    STATUS="FAILURE"
-    ((FAILURE_COUNT++))
-    FAILED_RUNS+=("Run #$((i+1)): $(echo $run_object | jq -c .)")
-  fi
-
-  RUN_OUTPUT=$(<"$RUN_OUTPUT_FILE")
-  rm "$RUN_OUTPUT_FILE"
-
-  # Parse results and optionally upload them to GCS
-  RUN_ID=""
-  RESULTS=""
-  GCS_RESULTS_URL=""
-  if [[ "$STATUS" == "SUCCESS" ]]; then
-    RESULT_FILE_PATH=$(echo "$RUN_OUTPUT" | grep 'RESULT_FILE=' | tail -n 1 | cut -d'=' -f2 | tr -s '/' || true)
-
-    if [[ -n "$RESULT_FILE_PATH" && -f "$RESULT_FILE_PATH" ]]; then
-      RUN_ID=$(basename "$(dirname "$RESULT_FILE_PATH")")
-      RESULT_DIR=$(dirname "$RESULT_FILE_PATH")
-      RESULTS=$(cat "$RESULT_FILE_PATH")
-
-      if [[ -n "$GCS_PATH" ]]; then
-        GCS_RESULTS_URL="${GCS_PATH}/${RUN_ID}"
-        echo "Uploading results to GCS..."
-        if gcloud storage rsync --recursive "$RESULT_DIR/" "$GCS_RESULTS_URL"; then
-          echo "GCS upload successful."
-        else
-          echo "Warning: GCS upload failed for RUN_ID $RUN_ID."
-        fi
-      fi
-    else
-      echo "Warning: Could not find result file for a successful run."
-      STATUS="WARNING_NO_RESULT_FILE"
-    fi
-  fi
-
-  # Add the results back into the JSON object for this run
-  json_content=$(echo "$json_content" | jq --argjson i "$i" --arg run_id "$RUN_ID" --arg status "$STATUS" --arg results "$RESULTS" --arg gcs_results "$GCS_RESULTS_URL" \
-    '.[$i] += {run_id: $run_id, status: $status, results: $results, gcs_results: $gcs_results}')
-
-  RUN_END_TIME=$(date +%s)
-  echo "Run finished in $((RUN_END_TIME - RUN_START_TIME)) seconds. Status: $STATUS"
-  echo "--------------------------------------------------"
-
-  # Save intermediate progress back to the file
-  echo "$json_content" > "$INPUT_JSON.tmp" && mv "$INPUT_JSON.tmp" "$INPUT_JSON"
-
-done
-
-SCRIPT_END_TIME=$(date +%s)
-echo "All benchmark runs completed in $((SCRIPT_END_TIME - SCRIPT_START_TIME)) seconds."
-echo
-echo "====================== SUMMARY ======================"
-echo "Successful runs: $SUCCESS_COUNT"
-echo "Failed runs:     $FAILURE_COUNT"
-echo "==================================================="
-
-if [[ $FAILURE_COUNT -gt 0 ]]; then
-  echo "Details of failed runs (see JSON file for full parameters):"
-  for failed in "${FAILED_RUNS[@]}"; do
-    echo "  - $failed"
-  done
-fi
-
-echo "Updated results have been saved to '$INPUT_JSON'."

benchmarks/benchmark_latency.py

@@ -1,17 +1,191 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import sys
+"""Benchmark the latency of processing a single batch of requests."""
+
+import argparse
+import dataclasses
+import json
+import os
+import time
+from typing import Any, Optional
+
+import numpy as np
+from tqdm import tqdm
+from typing_extensions import deprecated
+
+import vllm.envs as envs
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
+from vllm import LLM, SamplingParams
+from vllm.engine.arg_utils import EngineArgs
+from vllm.inputs import PromptType
+from vllm.sampling_params import BeamSearchParams
+from vllm.utils import FlexibleArgumentParser
+
+
+def save_to_pytorch_benchmark_format(
+    args: argparse.Namespace, results: dict[str, Any]
+) -> None:
+    pt_records = convert_to_pytorch_benchmark_format(
+        args=args,
+        metrics={"latency": results["latencies"]},
+        extra_info={k: results[k] for k in ["avg_latency", "percentiles"]},
+    )
+    if pt_records:
+        pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
+        write_to_json(pt_file, pt_records)
+
+
+@deprecated(
+    "benchmark_latency.py is deprecated and will be removed in a "
+    "future version. Please use 'vllm bench latency' instead.",
+)
+def main(args: argparse.Namespace):
+    print(args)
+
+    engine_args = EngineArgs.from_cli_args(args)
+
+    # NOTE(woosuk): If the request cannot be processed in a single batch,
+    # the engine will automatically process the request in multiple batches.
+    llm = LLM(**dataclasses.asdict(engine_args))
+    assert llm.llm_engine.model_config.max_model_len >= (
+        args.input_len + args.output_len
+    ), (
+        "Please ensure that max_model_len is greater than"
+        " the sum of input_len and output_len."
+    )
+
+    sampling_params = SamplingParams(
+        n=args.n,
+        temperature=1.0,
+        top_p=1.0,
+        ignore_eos=True,
+        max_tokens=args.output_len,
+        detokenize=not args.disable_detokenize,
+    )
+    print(sampling_params)
+    dummy_prompt_token_ids = np.random.randint(
+        10000, size=(args.batch_size, args.input_len)
+    )
+    dummy_prompts: list[PromptType] = [
+        {"prompt_token_ids": batch} for batch in dummy_prompt_token_ids.tolist()
+    ]
+
+    def llm_generate():
+        if not args.use_beam_search:
+            llm.generate(dummy_prompts, sampling_params=sampling_params, use_tqdm=False)
+        else:
+            llm.beam_search(
+                dummy_prompts,
+                BeamSearchParams(
+                    beam_width=args.n,
+                    max_tokens=args.output_len,
+                    ignore_eos=True,
+                ),
+            )
+
+    def run_to_completion(profile_dir: Optional[str] = None):
+        if profile_dir:
+            llm.start_profile()
+            llm_generate()
+            llm.stop_profile()
+        else:
+            start_time = time.perf_counter()
+            llm_generate()
+            end_time = time.perf_counter()
+            latency = end_time - start_time
+            return latency
+
+    print("Warming up...")
+    for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
+        run_to_completion(profile_dir=None)
+
+    if args.profile:
+        profile_dir = envs.VLLM_TORCH_PROFILER_DIR
+        print(f"Profiling (results will be saved to '{profile_dir}')...")
+        run_to_completion(profile_dir=profile_dir)
+        return
+
+    # Benchmark.
+    latencies = []
+    for _ in tqdm(range(args.num_iters), desc="Profiling iterations"):
+        latencies.append(run_to_completion(profile_dir=None))
+    latencies = np.array(latencies)
+    percentages = [10, 25, 50, 75, 90, 99]
+    percentiles = np.percentile(latencies, percentages)
+    print(f"Avg latency: {np.mean(latencies)} seconds")
+    for percentage, percentile in zip(percentages, percentiles):
+        print(f"{percentage}% percentile latency: {percentile} seconds")
+
+    # Output JSON results if specified
+    if args.output_json:
+        results = {
+            "avg_latency": np.mean(latencies),
+            "latencies": latencies.tolist(),
+            "percentiles": dict(zip(percentages, percentiles.tolist())),
+        }
+        with open(args.output_json, "w") as f:
+            json.dump(results, f, indent=4)
+        save_to_pytorch_benchmark_format(args, results)
+
+
+def create_argument_parser():
+    parser = FlexibleArgumentParser(
+        description="Benchmark the latency of processing a single batch of "
+        "requests till completion."
+    )
+    parser.add_argument("--input-len", type=int, default=32)
+    parser.add_argument("--output-len", type=int, default=128)
+    parser.add_argument("--batch-size", type=int, default=8)
+    parser.add_argument(
+        "--n",
+        type=int,
+        default=1,
+        help="Number of generated sequences per prompt.",
+    )
+    parser.add_argument("--use-beam-search", action="store_true")
+    parser.add_argument(
+        "--num-iters-warmup",
+        type=int,
+        default=10,
+        help="Number of iterations to run for warmup.",
+    )
+    parser.add_argument(
+        "--num-iters", type=int, default=30, help="Number of iterations to run."
+    )
+    parser.add_argument(
+        "--profile",
+        action="store_true",
+        help="profile the generation process of a single batch",
+    )
+    parser.add_argument(
+        "--output-json",
+        type=str,
+        default=None,
+        help="Path to save the latency results in JSON format.",
+    )
+    parser.add_argument(
+        "--disable-detokenize",
+        action="store_true",
+        help=(
+            "Do not detokenize responses (i.e. do not include "
+            "detokenization time in the latency measurement)"
+        ),
+    )
+
+    parser = EngineArgs.add_cli_args(parser)
+    # V1 enables prefix caching by default which skews the latency
+    # numbers. We need to disable prefix caching by default.
+    parser.set_defaults(enable_prefix_caching=False)
+
+    return parser
+
 
 if __name__ == "__main__":
-    print("""DEPRECATED: This script has been moved to the vLLM CLI.
-
-Please use the following command instead:
-    vllm bench latency
-
-For help with the new command, run:
-    vllm bench latency --help
-
-Alternatively, you can run the new command directly with:
-    python -m vllm.entrypoints.cli.main bench latency --help
-""")
-    sys.exit(1)
+    parser = create_argument_parser()
+    args = parser.parse_args()
+    if args.profile and not envs.VLLM_TORCH_PROFILER_DIR:
+        raise OSError(
+            "The environment variable 'VLLM_TORCH_PROFILER_DIR' is not set. "
+            "Please set it to a valid path to use torch profiler."
+        )
+    main(args)

benchmarks/benchmark_ngram_proposer.py

@@ -1,31 +1,17 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import gc
-import time
-from unittest import mock
 
 import numpy as np
 from tabulate import tabulate
 
 from benchmark_utils import TimeCollector
-from vllm.config import (
-    CacheConfig,
-    DeviceConfig,
-    LoadConfig,
-    ModelConfig,
-    ParallelConfig,
-    SchedulerConfig,
-    SpeculativeConfig,
-    VllmConfig,
-)
-from vllm.platforms import current_platform
+from vllm.config import ModelConfig, SpeculativeConfig, VllmConfig
 from vllm.utils import FlexibleArgumentParser
 from vllm.v1.spec_decode.ngram_proposer import NgramProposer
-from vllm.v1.worker.gpu_input_batch import InputBatch
-from vllm.v1.worker.gpu_model_runner import GPUModelRunner
 
 
-def benchmark_propose(args):
+def main(args):
     rows = []
     for max_ngram in args.max_ngram:
         collector = TimeCollector(TimeCollector.US)
@@ -83,88 +69,10 @@ def benchmark_propose(args):
     )
 
 
-def benchmark_batched_propose(args):
-    NUM_SPECULATIVE_TOKENS_NGRAM = 10
-    PROMPT_LOOKUP_MIN = 5
-    PROMPT_LOOKUP_MAX = 15
-    MAX_MODEL_LEN = int(1e7)
-    DEVICE = current_platform.device_type
-
-    model_config = ModelConfig(model="facebook/opt-125m", runner="generate")
-
-    speculative_config = SpeculativeConfig(
-        target_model_config=model_config,
-        target_parallel_config=ParallelConfig(),
-        method="ngram",
-        num_speculative_tokens=NUM_SPECULATIVE_TOKENS_NGRAM,
-        prompt_lookup_max=PROMPT_LOOKUP_MAX,
-        prompt_lookup_min=PROMPT_LOOKUP_MIN,
-    )
-
-    vllm_config = VllmConfig(
-        model_config=model_config,
-        cache_config=CacheConfig(),
-        speculative_config=speculative_config,
-        device_config=DeviceConfig(device=current_platform.device_type),
-        parallel_config=ParallelConfig(),
-        load_config=LoadConfig(),
-        scheduler_config=SchedulerConfig(),
-    )
-
-    # monkey patch vllm.v1.worker.gpu_model_runner.get_pp_group
-    mock_pp_group = mock.MagicMock()
-    mock_pp_group.world_size = 1
-    with mock.patch(
-        "vllm.v1.worker.gpu_model_runner.get_pp_group", return_value=mock_pp_group
-    ):
-        runner = GPUModelRunner(vllm_config, DEVICE)
-
-        # hack max model len
-        runner.max_model_len = MAX_MODEL_LEN
-        runner.drafter.max_model_len = MAX_MODEL_LEN
-
-        dummy_input_batch = InputBatch(
-            max_num_reqs=args.num_req,
-            max_model_len=MAX_MODEL_LEN,
-            max_num_batched_tokens=args.num_req * args.num_token,
-            device=DEVICE,
-            pin_memory=False,
-            vocab_size=256000,
-            block_sizes=[16],
-        )
-        dummy_input_batch._req_ids = list(str(id) for id in range(args.num_req))
-        dummy_input_batch.spec_decode_unsupported_reqs = ()
-        dummy_input_batch.num_tokens_no_spec = [args.num_token] * args.num_req
-        dummy_input_batch.token_ids_cpu = np.random.randint(
-            0, 20, (args.num_req, args.num_token)
-        )
-
-        runner.input_batch = dummy_input_batch
-
-        sampled_token_ids = [[0]] * args.num_req
-
-        print("Starting benchmark")
-        # first run is warmup so ignore it
-        for _ in range(args.num_iteration):
-            start = time.time()
-            runner.drafter.propose(
-                sampled_token_ids,
-                dummy_input_batch.req_ids,
-                dummy_input_batch.num_tokens_no_spec,
-                dummy_input_batch.token_ids_cpu,
-                dummy_input_batch.spec_decode_unsupported_reqs,
-            )
-            end = time.time()
-            print(f"Iteration time (s): {end - start}")
-
-
 def invoke_main() -> None:
     parser = FlexibleArgumentParser(
         description="Benchmark the performance of N-gram speculative decode drafting"
     )
-    parser.add_argument(
-        "--batched", action="store_true", help="consider time to prepare batch"
-    )  # noqa: E501
     parser.add_argument(
         "--num-iteration",
         type=int,
@@ -197,17 +105,8 @@ def invoke_main() -> None:
         help="Number of speculative tokens to generate",
     )
     args = parser.parse_args()
-
-    if not args.batched:
-        benchmark_propose(args)
-    else:
-        benchmark_batched_propose(args)
+    main(args)
 
 
-"""
-# Example command lines:
-# time python3 benchmarks/benchmark_ngram_proposer.py
-# time python3 benchmarks/benchmark_ngram_proposer.py --batched --num-iteration 4 --num-token 1000000 --num-req 128
-"""  # noqa: E501
 if __name__ == "__main__":
     invoke_main()  # pragma: no cover

benchmarks/benchmark_serving_structured_output.py

@@ -449,8 +449,7 @@ async def benchmark(
     def prepare_extra_body(request) -> dict:
         extra_body = {}
         # Add the schema to the extra_body
-        extra_body["structured_outputs"] = {}
-        extra_body["structured_outputs"][request.structure_type] = request.schema
+        extra_body[request.structure_type] = request.schema
         return extra_body
 
     print("Starting initial single prompt test run...")
@@ -697,11 +696,11 @@ def evaluate(ret, args):
         return re.match(args.regex, actual) is not None
 
     def _eval_correctness(expected, actual):
-        if args.structure_type == "json":
+        if args.structure_type == "guided_json":
             return _eval_correctness_json(expected, actual)
-        elif args.structure_type == "regex":
+        elif args.structure_type == "guided_regex":
             return _eval_correctness_regex(expected, actual)
-        elif args.structure_type == "choice":
+        elif args.structure_type == "guided_choice":
             return _eval_correctness_choice(expected, actual)
         else:
             return None
@@ -781,18 +780,18 @@ def main(args: argparse.Namespace):
     )
 
     if args.dataset == "grammar":
-        args.structure_type = "grammar"
+        args.structure_type = "guided_grammar"
     elif args.dataset == "regex":
-        args.structure_type = "regex"
+        args.structure_type = "guided_regex"
     elif args.dataset == "choice":
-        args.structure_type = "choice"
+        args.structure_type = "guided_choice"
     else:
-        args.structure_type = "json"
+        args.structure_type = "guided_json"
 
     if args.no_structured_output:
         args.structured_output_ratio = 0
     if args.save_results:
-        result_file_name = f"{args.structured_output_ratio}so"
+        result_file_name = f"{args.structured_output_ratio}guided"
         result_file_name += f"_{backend}"
         result_file_name += f"_{args.request_rate}qps"
         result_file_name += f"_{args.model.split('/')[-1]}"

benchmarks/benchmark_throughput.py

@@ -1,17 +1,741 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-import sys
+"""Benchmark offline inference throughput."""
+
+import argparse
+import dataclasses
+import json
+import os
+import random
+import time
+import warnings
+from typing import Any, Optional, Union
+
+import torch
+import uvloop
+from tqdm import tqdm
+from transformers import AutoModelForCausalLM, AutoTokenizer, PreTrainedTokenizerBase
+from typing_extensions import deprecated
+
+from benchmark_dataset import (
+    AIMODataset,
+    BurstGPTDataset,
+    ConversationDataset,
+    InstructCoderDataset,
+    RandomDataset,
+    SampleRequest,
+    ShareGPTDataset,
+    SonnetDataset,
+    VisionArenaDataset,
+)
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
+from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
+from vllm.entrypoints.openai.api_server import (
+    build_async_engine_client_from_engine_args,
+)
+from vllm.inputs import TextPrompt, TokensPrompt
+from vllm.lora.request import LoRARequest
+from vllm.outputs import RequestOutput
+from vllm.sampling_params import BeamSearchParams
+from vllm.utils import FlexibleArgumentParser, merge_async_iterators
+
+
+def run_vllm(
+    requests: list[SampleRequest],
+    n: int,
+    engine_args: EngineArgs,
+    disable_detokenize: bool = False,
+) -> tuple[float, Optional[list[RequestOutput]]]:
+    from vllm import LLM, SamplingParams
+
+    llm = LLM(**dataclasses.asdict(engine_args))
+    assert all(
+        llm.llm_engine.model_config.max_model_len
+        >= (request.prompt_len + request.expected_output_len)
+        for request in requests
+    ), (
+        "Please ensure that max_model_len is greater than the sum of"
+        " prompt_len and expected_output_len for all requests."
+    )
+    # Add the requests to the engine.
+    prompts: list[Union[TextPrompt, TokensPrompt]] = []
+    sampling_params: list[SamplingParams] = []
+    for request in requests:
+        prompts.append(
+            TokensPrompt(
+                prompt_token_ids=request.prompt["prompt_token_ids"],
+                multi_modal_data=request.multi_modal_data,
+            )
+            if "prompt_token_ids" in request.prompt
+            else TextPrompt(
+                prompt=request.prompt, multi_modal_data=request.multi_modal_data
+            )
+        )
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=1.0,
+                top_p=1.0,
+                ignore_eos=True,
+                max_tokens=request.expected_output_len,
+                detokenize=not disable_detokenize,
+            )
+        )
+    lora_requests: Optional[list[LoRARequest]] = None
+    if engine_args.enable_lora:
+        lora_requests = [request.lora_request for request in requests]
+
+    use_beam_search = False
+
+    outputs = None
+    if not use_beam_search:
+        start = time.perf_counter()
+        outputs = llm.generate(
+            prompts, sampling_params, lora_request=lora_requests, use_tqdm=True
+        )
+        end = time.perf_counter()
+    else:
+        assert lora_requests is None, "BeamSearch API does not support LoRA"
+        # output_len should be the same for all requests.
+        output_len = requests[0].expected_output_len
+        for request in requests:
+            assert request.expected_output_len == output_len
+        start = time.perf_counter()
+        llm.beam_search(
+            prompts,
+            BeamSearchParams(
+                beam_width=n,
+                max_tokens=output_len,
+                ignore_eos=True,
+            ),
+        )
+        end = time.perf_counter()
+    return end - start, outputs
+
+
+def run_vllm_chat(
+    requests: list[SampleRequest],
+    n: int,
+    engine_args: EngineArgs,
+    disable_detokenize: bool = False,
+) -> tuple[float, list[RequestOutput]]:
+    """
+    Run vLLM chat benchmark. This function is recommended ONLY for benchmarking
+    multimodal models as it properly handles multimodal inputs and chat
+    formatting. For non-multimodal models, use run_vllm() instead.
+    """
+    from vllm import LLM, SamplingParams
+
+    llm = LLM(**dataclasses.asdict(engine_args))
+
+    assert all(
+        llm.llm_engine.model_config.max_model_len
+        >= (request.prompt_len + request.expected_output_len)
+        for request in requests
+    ), (
+        "Please ensure that max_model_len is greater than the sum of "
+        "prompt_len and expected_output_len for all requests."
+    )
+
+    prompts = []
+    sampling_params: list[SamplingParams] = []
+    for request in requests:
+        prompts.append(request.prompt)
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=1.0,
+                top_p=1.0,
+                ignore_eos=True,
+                max_tokens=request.expected_output_len,
+                detokenize=not disable_detokenize,
+            )
+        )
+    start = time.perf_counter()
+    outputs = llm.chat(prompts, sampling_params, use_tqdm=True)
+    end = time.perf_counter()
+    return end - start, outputs
+
+
+async def run_vllm_async(
+    requests: list[SampleRequest],
+    n: int,
+    engine_args: AsyncEngineArgs,
+    disable_frontend_multiprocessing: bool = False,
+    disable_detokenize: bool = False,
+) -> float:
+    from vllm import SamplingParams
+
+    async with build_async_engine_client_from_engine_args(
+        engine_args,
+        disable_frontend_multiprocessing=disable_frontend_multiprocessing,
+    ) as llm:
+        model_config = await llm.get_model_config()
+        assert all(
+            model_config.max_model_len
+            >= (request.prompt_len + request.expected_output_len)
+            for request in requests
+        ), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " prompt_len and expected_output_len for all requests."
+        )
+
+        # Add the requests to the engine.
+        prompts: list[Union[TextPrompt, TokensPrompt]] = []
+        sampling_params: list[SamplingParams] = []
+        lora_requests: list[Optional[LoRARequest]] = []
+        for request in requests:
+            prompts.append(
+                TokensPrompt(
+                    prompt_token_ids=request.prompt["prompt_token_ids"],
+                    multi_modal_data=request.multi_modal_data,
+                )
+                if "prompt_token_ids" in request.prompt
+                else TextPrompt(
+                    prompt=request.prompt, multi_modal_data=request.multi_modal_data
+                )
+            )
+            sampling_params.append(
+                SamplingParams(
+                    n=n,
+                    temperature=1.0,
+                    top_p=1.0,
+                    ignore_eos=True,
+                    max_tokens=request.expected_output_len,
+                    detokenize=not disable_detokenize,
+                )
+            )
+            lora_requests.append(request.lora_request)
+
+        generators = []
+        start = time.perf_counter()
+        for i, (prompt, sp, lr) in enumerate(
+            zip(prompts, sampling_params, lora_requests)
+        ):
+            generator = llm.generate(prompt, sp, lora_request=lr, request_id=f"test{i}")
+            generators.append(generator)
+        all_gens = merge_async_iterators(*generators)
+        async for i, res in all_gens:
+            pass
+        end = time.perf_counter()
+        return end - start
+
+
+def run_hf(
+    requests: list[SampleRequest],
+    model: str,
+    tokenizer: PreTrainedTokenizerBase,
+    n: int,
+    max_batch_size: int,
+    trust_remote_code: bool,
+    disable_detokenize: bool = False,
+) -> float:
+    llm = AutoModelForCausalLM.from_pretrained(
+        model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code
+    )
+    if llm.config.model_type == "llama":
+        # To enable padding in the HF backend.
+        tokenizer.pad_token = tokenizer.eos_token
+    llm = llm.cuda()
+
+    pbar = tqdm(total=len(requests))
+    start = time.perf_counter()
+    batch: list[str] = []
+    max_prompt_len = 0
+    max_output_len = 0
+    for i in range(len(requests)):
+        prompt = requests[i].prompt
+        prompt_len = requests[i].prompt_len
+        output_len = requests[i].expected_output_len
+        # Add the prompt to the batch.
+        batch.append(prompt)
+        max_prompt_len = max(max_prompt_len, prompt_len)
+        max_output_len = max(max_output_len, output_len)
+        if len(batch) < max_batch_size and i != len(requests) - 1:
+            # Check if we can add more requests to the batch.
+            next_prompt_len = requests[i + 1].prompt_len
+            next_output_len = requests[i + 1].expected_output_len
+            if (
+                max(max_prompt_len, next_prompt_len)
+                + max(max_output_len, next_output_len)
+            ) <= 2048:
+                # We can add more requests to the batch.
+                continue
+
+        # Generate the sequences.
+        input_ids = tokenizer(batch, return_tensors="pt", padding=True).input_ids
+        llm_outputs = llm.generate(
+            input_ids=input_ids.cuda(),
+            do_sample=True,
+            num_return_sequences=n,
+            temperature=1.0,
+            top_p=1.0,
+            use_cache=True,
+            max_new_tokens=max_output_len,
+        )
+        if not disable_detokenize:
+            # Include the decoding time.
+            tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
+        pbar.update(len(batch))
+
+        # Clear the batch.
+        batch = []
+        max_prompt_len = 0
+        max_output_len = 0
+    end = time.perf_counter()
+    return end - start
+
+
+def run_mii(
+    requests: list[SampleRequest],
+    model: str,
+    tensor_parallel_size: int,
+    output_len: int,
+) -> float:
+    from mii import client, serve
+
+    llm = serve(model, tensor_parallel=tensor_parallel_size)
+    prompts = [request.prompt for request in requests]
+
+    start = time.perf_counter()
+    llm.generate(prompts, max_new_tokens=output_len)
+    end = time.perf_counter()
+    client = client(model)
+    client.terminate_server()
+    return end - start
+
+
+def save_to_pytorch_benchmark_format(
+    args: argparse.Namespace, results: dict[str, Any]
+) -> None:
+    pt_records = convert_to_pytorch_benchmark_format(
+        args=args,
+        metrics={
+            "requests_per_second": [results["requests_per_second"]],
+            "tokens_per_second": [results["tokens_per_second"]],
+        },
+        extra_info={
+            k: results[k] for k in ["elapsed_time", "num_requests", "total_num_tokens"]
+        },
+    )
+    if pt_records:
+        # Don't use json suffix here as we don't want CI to pick it up
+        pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
+        write_to_json(pt_file, pt_records)
+
+
+def get_requests(args, tokenizer):
+    # Common parameters for all dataset types.
+    common_kwargs = {
+        "dataset_path": args.dataset_path,
+        "random_seed": args.seed,
+    }
+    sample_kwargs = {
+        "tokenizer": tokenizer,
+        "lora_path": args.lora_path,
+        "max_loras": args.max_loras,
+        "num_requests": args.num_prompts,
+        "input_len": args.input_len,
+        "output_len": args.output_len,
+    }
+
+    if args.dataset_path is None or args.dataset_name == "random":
+        sample_kwargs["range_ratio"] = args.random_range_ratio
+        sample_kwargs["prefix_len"] = args.prefix_len
+        dataset_cls = RandomDataset
+    elif args.dataset_name == "sharegpt":
+        dataset_cls = ShareGPTDataset
+        if args.backend == "vllm-chat":
+            sample_kwargs["enable_multimodal_chat"] = True
+    elif args.dataset_name == "sonnet":
+        assert tokenizer.chat_template or tokenizer.default_chat_template, (
+            "Tokenizer/model must have chat template for sonnet dataset."
+        )
+        dataset_cls = SonnetDataset
+        sample_kwargs["prefix_len"] = args.prefix_len
+        sample_kwargs["return_prompt_formatted"] = True
+    elif args.dataset_name == "burstgpt":
+        dataset_cls = BurstGPTDataset
+    elif args.dataset_name == "hf":
+        common_kwargs["no_stream"] = args.no_stream
+        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = VisionArenaDataset
+            common_kwargs["dataset_subset"] = None
+            common_kwargs["dataset_split"] = "train"
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = InstructCoderDataset
+            common_kwargs["dataset_split"] = "train"
+        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = ConversationDataset
+            common_kwargs["dataset_subset"] = args.hf_subset
+            common_kwargs["dataset_split"] = args.hf_split
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = AIMODataset
+            common_kwargs["dataset_subset"] = None
+            common_kwargs["dataset_split"] = "train"
+    else:
+        raise ValueError(f"Unknown dataset name: {args.dataset_name}")
+    # Remove None values
+    sample_kwargs = {k: v for k, v in sample_kwargs.items() if v is not None}
+    return dataset_cls(**common_kwargs).sample(**sample_kwargs)
+
+
+@deprecated(
+    "benchmark_throughput.py is deprecated and will be removed in a "
+    "future version. Please use 'vllm bench throughput' instead.",
+)
+def main(args: argparse.Namespace):
+    if args.seed is None:
+        args.seed = 0
+    print(args)
+    random.seed(args.seed)
+    # Sample the requests.
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.tokenizer, trust_remote_code=args.trust_remote_code
+    )
+    requests = get_requests(args, tokenizer)
+    is_multi_modal = any(request.multi_modal_data is not None for request in requests)
+    request_outputs: Optional[list[RequestOutput]] = None
+    if args.backend == "vllm":
+        if args.async_engine:
+            elapsed_time = uvloop.run(
+                run_vllm_async(
+                    requests,
+                    args.n,
+                    AsyncEngineArgs.from_cli_args(args),
+                    args.disable_frontend_multiprocessing,
+                    args.disable_detokenize,
+                )
+            )
+        else:
+            elapsed_time, request_outputs = run_vllm(
+                requests,
+                args.n,
+                EngineArgs.from_cli_args(args),
+                args.disable_detokenize,
+            )
+    elif args.backend == "hf":
+        assert args.tensor_parallel_size == 1
+        elapsed_time = run_hf(
+            requests,
+            args.model,
+            tokenizer,
+            args.n,
+            args.hf_max_batch_size,
+            args.trust_remote_code,
+            args.disable_detokenize,
+        )
+    elif args.backend == "mii":
+        elapsed_time = run_mii(
+            requests, args.model, args.tensor_parallel_size, args.output_len
+        )
+    elif args.backend == "vllm-chat":
+        elapsed_time, request_outputs = run_vllm_chat(
+            requests, args.n, EngineArgs.from_cli_args(args), args.disable_detokenize
+        )
+    else:
+        raise ValueError(f"Unknown backend: {args.backend}")
+
+    if request_outputs:
+        # Note: with the vllm and vllm-chat backends,
+        # we have request_outputs, which we use to count tokens.
+        total_prompt_tokens = 0
+        total_output_tokens = 0
+        for ro in request_outputs:
+            if not isinstance(ro, RequestOutput):
+                continue
+            total_prompt_tokens += (
+                len(ro.prompt_token_ids) if ro.prompt_token_ids else 0
+            )
+            total_output_tokens += sum(len(o.token_ids) for o in ro.outputs if o)
+        total_num_tokens = total_prompt_tokens + total_output_tokens
+    else:
+        total_num_tokens = sum(r.prompt_len + r.expected_output_len for r in requests)
+        total_output_tokens = sum(r.expected_output_len for r in requests)
+        total_prompt_tokens = total_num_tokens - total_output_tokens
+
+    if is_multi_modal and args.backend != "vllm-chat":
+        print(
+            "\033[91mWARNING\033[0m: Multi-modal request with "
+            f"{args.backend} backend detected. The "
+            "following metrics are not accurate because image tokens are not"
+            " counted. See vllm-project/vllm/issues/9778 for details."
+        )
+        # TODO(vllm-project/vllm/issues/9778): Count multi-modal token length.
+        # vllm-chat backend counts the image tokens now
+
+    print(
+        f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
+        f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
+        f"{total_output_tokens / elapsed_time:.2f} output tokens/s"
+    )
+    print(f"Total num prompt tokens:  {total_prompt_tokens}")
+    print(f"Total num output tokens:  {total_output_tokens}")
+
+    # Output JSON results if specified
+    if args.output_json:
+        results = {
+            "elapsed_time": elapsed_time,
+            "num_requests": len(requests),
+            "total_num_tokens": total_num_tokens,
+            "requests_per_second": len(requests) / elapsed_time,
+            "tokens_per_second": total_num_tokens / elapsed_time,
+        }
+        with open(args.output_json, "w") as f:
+            json.dump(results, f, indent=4)
+        save_to_pytorch_benchmark_format(args, results)
+
+
+def validate_args(args):
+    """
+    Validate command-line arguments.
+    """
+
+    # === Deprecation and Defaulting ===
+    if args.dataset is not None:
+        warnings.warn(
+            "The '--dataset' argument will be deprecated in the next release. "
+            "Please use '--dataset-name' and '--dataset-path' instead.",
+            stacklevel=2,
+        )
+        args.dataset_path = args.dataset
+
+    if not getattr(args, "tokenizer", None):
+        args.tokenizer = args.model
+
+    # === Backend Validation ===
+    valid_backends = {"vllm", "hf", "mii", "vllm-chat"}
+    if args.backend not in valid_backends:
+        raise ValueError(f"Unsupported backend: {args.backend}")
+
+    # === Dataset Configuration ===
+    if not args.dataset and not args.dataset_path:
+        print("When dataset path is not set, it will default to random dataset")
+        args.dataset_name = "random"
+        if args.input_len is None:
+            raise ValueError("input_len must be provided for a random dataset")
+
+    # === Dataset Name Specific Checks ===
+    # --hf-subset and --hf-split: only used
+    # when dataset_name is 'hf'
+    if args.dataset_name != "hf" and (
+        getattr(args, "hf_subset", None) is not None
+        or getattr(args, "hf_split", None) is not None
+    ):
+        warnings.warn(
+            "--hf-subset and --hf-split will be ignored \
+                since --dataset-name is not 'hf'.",
+            stacklevel=2,
+        )
+    elif args.dataset_name == "hf":
+        if args.dataset_path in (
+            VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys()
+            | ConversationDataset.SUPPORTED_DATASET_PATHS
+        ):
+            assert args.backend == "vllm-chat", (
+                f"{args.dataset_path} needs to use vllm-chat as the backend."
+            )  # noqa: E501
+        elif args.dataset_path in (
+            InstructCoderDataset.SUPPORTED_DATASET_PATHS
+            | AIMODataset.SUPPORTED_DATASET_PATHS
+        ):
+            assert args.backend == "vllm", (
+                f"{args.dataset_path} needs to use vllm as the backend."
+            )  # noqa: E501
+        else:
+            raise ValueError(f"{args.dataset_path} is not supported by hf dataset.")
+
+    # --random-range-ratio: only used when dataset_name is 'random'
+    if args.dataset_name != "random" and args.random_range_ratio is not None:
+        warnings.warn(
+            "--random-range-ratio will be ignored since \
+                --dataset-name is not 'random'.",
+            stacklevel=2,
+        )
+
+    # --prefix-len: only used when dataset_name is 'random', 'sonnet', or not
+    # set.
+    if (
+        args.dataset_name not in {"random", "sonnet", None}
+        and args.prefix_len is not None
+    ):
+        warnings.warn(
+            "--prefix-len will be ignored since --dataset-name\
+                 is not 'random', 'sonnet', or not set.",
+            stacklevel=2,
+        )
+
+    # === LoRA Settings ===
+    if getattr(args, "enable_lora", False) and args.backend != "vllm":
+        raise ValueError("LoRA benchmarking is only supported for vLLM backend")
+    if getattr(args, "enable_lora", False) and args.lora_path is None:
+        raise ValueError("LoRA path must be provided when enable_lora is True")
+
+    # === Backend-specific Validations ===
+    if args.backend == "hf" and args.hf_max_batch_size is None:
+        raise ValueError("HF max batch size is required for HF backend")
+    if args.backend != "hf" and args.hf_max_batch_size is not None:
+        raise ValueError("HF max batch size is only for HF backend.")
+
+    if (
+        args.backend in {"hf", "mii"}
+        and getattr(args, "quantization", None) is not None
+    ):
+        raise ValueError("Quantization is only for vLLM backend.")
+
+    if args.backend == "mii" and args.dtype != "auto":
+        raise ValueError("dtype must be auto for MII backend.")
+    if args.backend == "mii" and args.n != 1:
+        raise ValueError("n must be 1 for MII backend.")
+    if args.backend == "mii" and args.tokenizer != args.model:
+        raise ValueError("Tokenizer must be the same as the model for MII backend.")
+
+    # --data-parallel is not supported currently.
+    # https://github.com/vllm-project/vllm/issues/16222
+    if args.data_parallel_size > 1:
+        raise ValueError(
+            "Data parallel is not supported in offline benchmark, "
+            "please use benchmark serving instead"
+        )
+
+
+def create_argument_parser():
+    parser = FlexibleArgumentParser(description="Benchmark the throughput.")
+    parser.add_argument(
+        "--backend",
+        type=str,
+        choices=["vllm", "hf", "mii", "vllm-chat"],
+        default="vllm",
+    )
+    parser.add_argument(
+        "--dataset-name",
+        type=str,
+        choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"],
+        help="Name of the dataset to benchmark on.",
+        default="sharegpt",
+    )
+    parser.add_argument(
+        "--no-stream",
+        action="store_true",
+        help="Do not load the dataset in streaming mode.",
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default=None,
+        help="Path to the ShareGPT dataset, will be deprecated in\
+            the next release. The dataset is expected to "
+        "be a json in form of list[dict[..., conversations: "
+        "list[dict[..., value: <prompt_or_response>]]]]",
+    )
+    parser.add_argument(
+        "--dataset-path", type=str, default=None, help="Path to the dataset"
+    )
+    parser.add_argument(
+        "--input-len",
+        type=int,
+        default=None,
+        help="Input prompt length for each request",
+    )
+    parser.add_argument(
+        "--output-len",
+        type=int,
+        default=None,
+        help="Output length for each request. Overrides the "
+        "output length from the dataset.",
+    )
+    parser.add_argument(
+        "--n", type=int, default=1, help="Number of generated sequences per prompt."
+    )
+    parser.add_argument(
+        "--num-prompts", type=int, default=1000, help="Number of prompts to process."
+    )
+    parser.add_argument(
+        "--hf-max-batch-size",
+        type=int,
+        default=None,
+        help="Maximum batch size for HF backend.",
+    )
+    parser.add_argument(
+        "--output-json",
+        type=str,
+        default=None,
+        help="Path to save the throughput results in JSON format.",
+    )
+    parser.add_argument(
+        "--async-engine",
+        action="store_true",
+        default=False,
+        help="Use vLLM async engine rather than LLM class.",
+    )
+    parser.add_argument(
+        "--disable-frontend-multiprocessing",
+        action="store_true",
+        default=False,
+        help="Disable decoupled async engine frontend.",
+    )
+    parser.add_argument(
+        "--disable-detokenize",
+        action="store_true",
+        help=(
+            "Do not detokenize the response (i.e. do not include "
+            "detokenization time in the measurement)"
+        ),
+    )
+    # LoRA
+    parser.add_argument(
+        "--lora-path",
+        type=str,
+        default=None,
+        help="Path to the LoRA adapters to use. This can be an absolute path, "
+        "a relative path, or a Hugging Face model identifier.",
+    )
+    parser.add_argument(
+        "--prefix-len",
+        type=int,
+        default=None,
+        help=f"Number of prefix tokens to be used in RandomDataset "
+        "and SonnetDataset. For RandomDataset, the total input "
+        "length is the sum of prefix-len (default: "
+        f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length "
+        "sampled from [input_len * (1 - range_ratio), "
+        "input_len * (1 + range_ratio)]. For SonnetDataset, "
+        f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) "
+        "controls how much of the input is fixed lines versus "
+        "random lines, but the total input length remains approximately "
+        "input_len tokens.",
+    )
+    # random dataset
+    parser.add_argument(
+        "--random-range-ratio",
+        type=float,
+        default=None,
+        help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) "
+        "for sampling input/output length, "
+        "used only for RandomDataset. Must be in the range [0, 1) to "
+        "define a symmetric sampling range "
+        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
+    )
+
+    # hf dataset
+    parser.add_argument(
+        "--hf-subset", type=str, default=None, help="Subset of the HF dataset."
+    )
+    parser.add_argument(
+        "--hf-split", type=str, default=None, help="Split of the HF dataset."
+    )
+
+    parser = AsyncEngineArgs.add_cli_args(parser)
+
+    return parser
+
 
 if __name__ == "__main__":
-    print("""DEPRECATED: This script has been moved to the vLLM CLI.
-
-Please use the following command instead:
-    vllm bench throughput
-
-For help with the new command, run:
-    vllm bench throughput --help
-
-Alternatively, you can run the new command directly with:
-    python -m vllm.entrypoints.cli.main bench throughput --help
-""")
-    sys.exit(1)
+    parser = create_argument_parser()
+    args = parser.parse_args()
+    if args.tokenizer is None:
+        args.tokenizer = args.model
+    validate_args(args)
+    main(args)

benchmarks/kernels/bench_block_fp8_gemm.py

@@ -4,10 +4,7 @@
 import torch
 
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
-    apply_w8a8_block_fp8_linear,
-)
-from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
-    CUTLASS_BLOCK_FP8_SUPPORTED,
+    w8a8_block_fp8_matmul,
 )
 from vllm.platforms import current_platform
 from vllm.triton_utils import triton as vllm_triton
@@ -32,7 +29,7 @@ DEEPSEEK_V3_SHAPES = [
 ]
 
 
-def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
+def build_w8a8_block_fp8_runner(M, N, K, block_size, device):
     """Build runner function for w8a8 block fp8 matmul."""
     factor_for_scale = 1e-2
 
@@ -40,54 +37,37 @@ def build_w8a8_block_fp8_runner(M, N, K, block_size, device, use_cutlass):
     fp8_max, fp8_min = fp8_info.max, fp8_info.min
 
     # Create random FP8 tensors
-    A_ref = (torch.rand(M, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
+    A_fp32 = (torch.rand(M, K, dtype=torch.float32, device=device) - 0.5) * 2 * fp8_max
+    A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
 
-    B_ref = (torch.rand(N, K, dtype=torch.bfloat16, device=device) - 0.5) * 2 * fp8_max
-    B = B_ref.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+    B_fp32 = (torch.rand(N, K, dtype=torch.float32, device=device) - 0.5) * 2 * fp8_max
+    B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
 
     # Create scales
     block_n, block_k = block_size[0], block_size[1]
     n_tiles = (N + block_n - 1) // block_n
     k_tiles = (K + block_k - 1) // block_k
 
+    As = torch.rand(M, k_tiles, dtype=torch.float32, device=device) * factor_for_scale
     Bs = (
         torch.rand(n_tiles, k_tiles, dtype=torch.float32, device=device)
         * factor_for_scale
     )
 
-    # SM90 CUTLASS requires row-major format for scales
-    if use_cutlass and current_platform.is_device_capability(90):
-        Bs = Bs.T.contiguous()
-
     def run():
-        if use_cutlass:
-            return apply_w8a8_block_fp8_linear(
-                A_ref, B, block_size, Bs, cutlass_block_fp8_supported=True
-            )
-        else:
-            return apply_w8a8_block_fp8_linear(
-                A_ref, B, block_size, Bs, cutlass_block_fp8_supported=False
-            )
+        return w8a8_block_fp8_matmul(A, B, As, Bs, block_size, torch.bfloat16)
 
     return run
 
 
-# Determine available providers
-available_providers = ["torch-bf16", "w8a8-block-fp8-triton"]
-plot_title = "BF16 vs W8A8 Block FP8 GEMMs"
-
-if CUTLASS_BLOCK_FP8_SUPPORTED:
-    available_providers.append("w8a8-block-fp8-cutlass")
-
-
 @vllm_triton.testing.perf_report(
     vllm_triton.testing.Benchmark(
         x_names=["batch_size"],
         x_vals=[1, 16, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384],
         x_log=False,
         line_arg="provider",
-        line_vals=available_providers,
-        line_names=available_providers,
+        line_vals=["torch-bf16", "w8a8-block-fp8"],
+        line_names=["torch-bf16", "w8a8-block-fp8"],
         ylabel="TFLOP/s (larger is better)",
         plot_name="BF16 vs W8A8 Block FP8 GEMMs",
         args={},
@@ -105,22 +85,11 @@ def benchmark_tflops(batch_size, provider, N, K, block_size=(128, 128)):
         ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
             lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
         )
-    elif provider == "w8a8-block-fp8-triton":
-        run_w8a8_triton = build_w8a8_block_fp8_runner(
-            M, N, K, block_size, device, use_cutlass=False
-        )
+    else:  # w8a8-block-fp8
+        run_w8a8 = build_w8a8_block_fp8_runner(M, N, K, block_size, device)
         ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
-            lambda: run_w8a8_triton(), quantiles=quantiles
+            lambda: run_w8a8(), quantiles=quantiles
         )
-    elif provider == "w8a8-block-fp8-cutlass":
-        run_w8a8_cutlass = build_w8a8_block_fp8_runner(
-            M, N, K, block_size, device, use_cutlass=True
-        )
-        ms, min_ms, max_ms = vllm_triton.testing.do_bench_cudagraph(
-            lambda: run_w8a8_cutlass(), quantiles=quantiles
-        )
-    else:
-        raise ValueError(f"Unknown provider: {provider}")
 
     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)

benchmarks/kernels/bench_nvfp4_gemm.py

@@ -3,7 +3,6 @@
 import argparse
 import copy
 import itertools
-import os
 
 import torch
 from weight_shapes import WEIGHT_SHAPES
@@ -24,45 +23,21 @@ PROVIDER_CFGS = {
     "torch-bf16": dict(enabled=True),
     "nvfp4": dict(no_a_quant=False, enabled=True),
     "nvfp4-noquant": dict(no_a_quant=True, enabled=True),
-    "fbgemm-nvfp4": dict(fbgemm=True, no_a_quant=False, enabled=True),
-    "fbgemm-nvfp4-noquant": dict(fbgemm=True, no_a_quant=True, enabled=True),
 }
 
-_needs_fbgemm = any(
-    v.get("fbgemm", False) for v in PROVIDER_CFGS.values() if v.get("enabled", False)
-)
-if _needs_fbgemm:
-    try:
-        from fbgemm_gpu.experimental.gemm.triton_gemm.fp4_quantize import (
-            triton_scale_nvfp4_quant,
-        )
-    except ImportError:
-        print(
-            "WARNING: FBGEMM providers are enabled but fbgemm_gpu is not installed. "
-            "These providers will be skipped. Please install fbgemm_gpu with: "
-            "'pip install fbgemm-gpu-genai' to run them."
-        )
-        # Disable FBGEMM providers so the benchmark can run.
-        for cfg in PROVIDER_CFGS.values():
-            if cfg.get("fbgemm"):
-                cfg["enabled"] = False
-
 _enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
 
 
-def _quant_weight_nvfp4(b: torch.Tensor, device: str, cfg):
+def _quant_weight_nvfp4(b: torch.Tensor, device: str):
     # Compute global scale for weight
     b_amax = torch.abs(b).max().to(torch.float32)
     b_global_scale = FLOAT8_E4M3_MAX * FLOAT4_E2M1_MAX / b_amax
-    if "fbgemm" in cfg and cfg["fbgemm"]:
-        b_fp4, scale_b_fp4 = triton_scale_nvfp4_quant(b, b_global_scale)
-    else:
-        b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
+    b_fp4, scale_b_fp4 = ops.scaled_fp4_quant(b, b_global_scale)
     return b_fp4, scale_b_fp4, b_global_scale
 
 
 def build_nvfp4_runner(cfg, a, b, dtype, device):
-    b_fp4, scale_b_fp4, b_global_scale = _quant_weight_nvfp4(b, device, cfg)
+    b_fp4, scale_b_fp4, b_global_scale = _quant_weight_nvfp4(b, device)
 
     # Compute global scale for activation
     # NOTE: This is generally provided ahead-of-time by the model checkpoint.
@@ -71,35 +46,6 @@ def build_nvfp4_runner(cfg, a, b, dtype, device):
 
     # Alpha for the GEMM operation
     alpha = 1.0 / (a_global_scale * b_global_scale)
-    if "fbgemm" in cfg and cfg["fbgemm"]:
-        if cfg["no_a_quant"]:
-            a_fp4, scale_a_fp4 = triton_scale_nvfp4_quant(a, a_global_scale)
-
-            def run():
-                return torch.ops.fbgemm.f4f4bf16(
-                    a_fp4,
-                    b_fp4,
-                    scale_a_fp4,
-                    scale_b_fp4,
-                    global_scale=alpha,
-                    use_mx=False,
-                )
-
-            return run
-        else:
-
-            def run():
-                a_fp4, scale_a_fp4 = triton_scale_nvfp4_quant(a, a_global_scale)
-                return torch.ops.fbgemm.f4f4bf16(
-                    a_fp4,
-                    b_fp4,
-                    scale_a_fp4,
-                    scale_b_fp4,
-                    global_scale=alpha,
-                    use_mx=False,
-                )
-
-            return run
 
     if cfg["no_a_quant"]:
         # Pre-quantize activation
@@ -184,13 +130,10 @@ if __name__ == "__main__":
 
     for K, N, model in prepare_shapes(args):
         print(f"{model}, N={N} K={K}, BF16 vs NVFP4 GEMMs TFLOP/s:")
-        save_dir = f"bench_nvfp4_res_n{N}_k{K}"
-        os.makedirs(save_dir, exist_ok=True)
-
         benchmark.run(
             print_data=True,
             show_plots=True,
-            save_path=save_dir,
+            save_path=f"bench_nvfp4_res_n{N}_k{K}",
             N=N,
             K=K,
         )

benchmarks/kernels/bench_per_token_quant_fp8.py

@@ -2,25 +2,14 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import itertools
 from typing import Callable
-from unittest.mock import patch
 
-import pandas as pd
 import torch
 
+from vllm import _custom_ops as ops
+from vllm.config import CompilationConfig, VllmConfig, set_current_vllm_config
 from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
 from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
 from vllm.triton_utils import triton
-from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser
-
-
-def with_triton_mode(fn):
-    """Temporarily force the Triton fallback path"""
-
-    def wrapped(*args, **kwargs):
-        with patch("vllm.platforms.current_platform.is_cuda", return_value=False):
-            return fn(*args, **kwargs)
-
-    return wrapped
 
 
 # TODO(luka): use standalone_compile utility
@@ -32,238 +21,78 @@ def with_dyn_arg(fn: Callable, arg_index: int, dim_index: int):
     return inner
 
 
-def bench_compile(fn: Callable):
-    # recompile for different shapes
-    fwd = torch.compile(fn, fullgraph=True, dynamic=False)
+torch._dynamo.config.recompile_limit = 8888
+compilation_config = CompilationConfig(custom_ops=["none"])
+with set_current_vllm_config(VllmConfig(compilation_config=compilation_config)):
+    torch_per_token_quant_fp8 = torch.compile(
+        QuantFP8(False, GroupShape.PER_TOKEN),
+        fullgraph=True,
+        dynamic=False,  # recompile for different shapes
+    )
 
     # First dim is explicitly dynamic to simulate vLLM usage
-    return with_dyn_arg(fwd, 0, 0)
+    torch_per_token_quant_fp8 = with_dyn_arg(torch_per_token_quant_fp8, 0, 0)
 
 
-torch._dynamo.config.recompile_limit = 8888
+def cuda_per_token_quant_fp8(
+    input: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    return ops.scaled_fp8_quant(input)
 
 
-def calculate_diff(
-    batch_size: int,
-    hidden_size: int,
-    group_shape: GroupShape,
-    dtype: torch.dtype,
-):
-    """Calculate the difference between Inductor and CUDA implementations."""
+def calculate_diff(batch_size: int, seq_len: int):
+    """Calculate difference between Triton and CUDA implementations."""
     device = torch.device("cuda")
-    x = torch.randn((batch_size, hidden_size), dtype=dtype, device=device)
+    x = torch.rand((batch_size * seq_len, 4096), dtype=torch.float16, device=device)
 
-    quant_fp8 = QuantFP8(False, group_shape, column_major_scales=False)
+    torch_out, torch_scale = torch_per_token_quant_fp8(x)
+    cuda_out, cuda_scale = cuda_per_token_quant_fp8(x)
 
-    torch_out, torch_scale = bench_compile(quant_fp8.forward_native)(x)
-    torch_eager_out, torch_eager_scale = quant_fp8.forward_native(x)
-    cuda_out, cuda_scale = quant_fp8.forward_cuda(x)
-
-    try:
-        torch.testing.assert_close(
-            cuda_out.to(torch.float32),
-            torch_out.to(torch.float32),
-            rtol=1e-3,
-            atol=1e-5,
-        )
-        torch.testing.assert_close(cuda_scale, torch_scale, rtol=1e-3, atol=1e-5)
-        torch.testing.assert_close(
-            cuda_out.to(torch.float32),
-            torch_eager_out.to(torch.float32),
-            rtol=1e-3,
-            atol=1e-5,
-        )
-        torch.testing.assert_close(cuda_scale, torch_eager_scale, rtol=1e-3, atol=1e-5)
+    if torch.allclose(
+        cuda_out.to(torch.float32), torch_out.to(torch.float32), rtol=1e-3, atol=1e-5
+    ) and torch.allclose(cuda_scale, torch_scale, rtol=1e-3, atol=1e-5):
         print("✅ All implementations match")
-    except AssertionError as e:
+    else:
         print("❌ Implementations differ")
-        print(e)
 
 
-configs = []
+batch_size_range = [1, 16, 32, 64, 128]
+seq_len_range = [1, 16, 64, 128, 256, 512, 1024, 2048, 4096]
+
+configs = list(itertools.product(batch_size_range, seq_len_range))
 
 
-def benchmark_quantization(
-    batch_size,
-    hidden_size,
-    provider,
-    group_shape: GroupShape,
-    col_major: bool,
-    dtype: torch.dtype,
-):
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["batch_size", "seq_len"],
+        x_vals=configs,
+        line_arg="provider",
+        line_vals=["torch", "cuda"],
+        line_names=["Torch", "CUDA"],
+        styles=[("blue", "-"), ("green", "-")],
+        ylabel="us",
+        plot_name="per-token-dynamic-quant-fp8-performance",
+        args={},
+    )
+)
+def benchmark_quantization(batch_size, seq_len, provider):
+    dtype = torch.float16
     device = torch.device("cuda")
 
-    x = torch.randn(batch_size, hidden_size, device=device, dtype=dtype)
+    x = torch.randn(batch_size * seq_len, 4096, device=device, dtype=dtype)
 
     quantiles = [0.5, 0.2, 0.8]
-    quant_fp8 = QuantFP8(False, group_shape, column_major_scales=col_major)
 
     if provider == "torch":
-        fn = lambda: bench_compile(quant_fp8.forward_native)(x.clone())
+        fn = lambda: torch_per_token_quant_fp8(x.clone())
     elif provider == "cuda":
-        fn = lambda: quant_fp8.forward_cuda(x.clone())
-    elif provider == "triton":
-        if not group_shape.is_per_group():
-            # Triton only supported for per-group
-            return 0, 0, 0
-
-        fn = lambda: with_triton_mode(quant_fp8.forward_cuda)(x.clone())
+        fn = lambda: cuda_per_token_quant_fp8(x.clone())
 
     ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(fn, quantiles=quantiles)
 
     return 1000 * ms, 1000 * max_ms, 1000 * min_ms
 
 
-# TODO(luka) extract to utils
-def compute_geomean_speedups(
-    df: pd.DataFrame,
-    baseline_col: str,
-    speedup_cols: list[str],
-    groupby_cols: list[str] | None = None,
-) -> pd.DataFrame:
-    """
-    Compute geometric mean speedups over a baseline column.
-
-    Args:
-        df: Input dataframe
-        baseline_col: Column to use as baseline
-        speedup_cols: Columns to compute speedups for
-        groupby_cols: Columns to group by. If None, compute over entire df.
-
-    Returns:
-        pd.DataFrame with geometric mean speedups
-    """
-    from scipy.stats import gmean
-
-    def geo_speedup(group: pd.DataFrame) -> pd.Series:
-        ratios = {
-            col: (group[baseline_col] / group[col]).values for col in speedup_cols
-        }
-        return pd.Series({col: gmean(vals) for col, vals in ratios.items()})
-
-    if groupby_cols is None:
-        result = geo_speedup(df).to_frame().T
-    else:
-        result = (
-            df.groupby(groupby_cols)
-            .apply(geo_speedup, include_groups=False)
-            .reset_index()
-        )
-
-    return result
-
-
 if __name__ == "__main__":
-    parser = FlexibleArgumentParser(
-        description="Benchmark the various implementations of QuantFP8 (dynamic-only)"
-    )
-    parser.add_argument("-c", "--check", action="store_true")
-    parser.add_argument(
-        "--dtype", type=str, choices=["half", "bfloat16", "float"], default="bfloat16"
-    )
-    parser.add_argument(
-        "--hidden-sizes",
-        type=int,
-        nargs="+",
-        default=[896, 1024, 2048, 4096, 7168],
-        help="Hidden sizes to benchmark",
-    )
-    parser.add_argument(
-        "--batch-sizes",
-        type=int,
-        nargs="+",
-        default=[1, 16, 128, 512, 1024],
-        help="Batch sizes to benchmark",
-    )
-    parser.add_argument(
-        "--group-sizes",
-        type=int,
-        nargs="+",
-        default=None,
-        help="Group sizes for GroupShape(1,N) to benchmark. "
-        "Use 0 for PER_TENSOR, -1 for PER_TOKEN (default: 0,-1,64,128)",
-    )
-    parser.add_argument(
-        "--no-column-major",
-        action="store_true",
-        help="Disable column-major scales testing",
-    )
-
-    args = parser.parse_args()
-    assert args
-
-    dtype = STR_DTYPE_TO_TORCH_DTYPE[args.dtype]
-
-    hidden_sizes = args.hidden_sizes
-    batch_sizes = args.batch_sizes
-
-    if args.group_sizes is not None:
-        group_shapes = []
-        for size in args.group_sizes:
-            if size == 0:
-                group_shapes.append(GroupShape.PER_TENSOR)
-            elif size == -1:
-                group_shapes.append(GroupShape.PER_TOKEN)
-            else:
-                group_shapes.append(GroupShape(1, size))
-    else:
-        group_shapes = [
-            GroupShape.PER_TENSOR,
-            GroupShape.PER_TOKEN,
-            GroupShape(1, 64),
-            GroupShape(1, 128),
-        ]
-
-    column_major_scales = [False] if args.no_column_major else [True, False]
-
-    config_gen = itertools.product(
-        group_shapes,
-        column_major_scales,
-        batch_sizes,
-        hidden_sizes,
-    )
-
-    # filter out column-major scales for non-group, reverse order
-    configs.extend(c[::-1] for c in config_gen if (c[0].is_per_group() or not c[1]))
-
-    print(f"Running {len(configs)} configurations:")
-    print(f"  Hidden sizes: {hidden_sizes}")
-    print(f"  Batch sizes: {batch_sizes}")
-    print(f"  Group shapes: {[str(g) for g in group_shapes]}")
-    print(f"  Column major scales: {column_major_scales}")
-    print()
-
-    if args.check:
-        for group_shape in group_shapes:
-            group_size = group_shape[1]
-            print(f"{group_size=}")
-            calculate_diff(
-                batch_size=4, hidden_size=4096, group_shape=group_shape, dtype=dtype
-            )
-
-    benchmark = triton.testing.perf_report(
-        triton.testing.Benchmark(
-            x_names=["hidden_size", "batch_size", "col_major", "group_shape"],
-            x_vals=configs,
-            line_arg="provider",
-            line_vals=["torch", "cuda", "triton"],
-            line_names=["Torch (Compiled)", "CUDA", "Triton"],
-            styles=[("blue", "-"), ("green", "-"), ("black", "-")],
-            ylabel="us",
-            plot_name="QuantFP8 performance",
-            args={},
-        )
-    )(benchmark_quantization)
-
-    df = benchmark.run(print_data=True, dtype=dtype, return_df=True)
-
-    # Print geomean speedups
-    geo_table_grouped = compute_geomean_speedups(
-        df,
-        baseline_col="Torch (Compiled)",
-        speedup_cols=["CUDA", "Triton"],
-        groupby_cols=["col_major", "group_shape"],
-    )
-
-    print("Speedup over Torch (Compiled)")
-    print(geo_table_grouped.to_string(index=False))
+    calculate_diff(batch_size=4, seq_len=4096)
+    benchmark_quantization.run(print_data=True)

benchmarks/kernels/benchmark_cutlass_fp4_moe.py

@@ -13,10 +13,6 @@ import torch.utils.benchmark as benchmark
 
 from vllm import _custom_ops as ops
 from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
-from vllm.model_executor.layers.fused_moe.config import (
-    fp8_w8a8_moe_quant_config,
-    nvfp4_moe_quant_config,
-)
 from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp4
 from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
 from vllm.scalar_type import scalar_types
@@ -144,12 +140,6 @@ def bench_run(
         a_fp8_scale: torch.Tensor,
         num_repeats: int,
     ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a_fp8_scale,
-        )
-
         for _ in range(num_repeats):
             fused_experts(
                 a,
@@ -157,7 +147,10 @@ def bench_run(
                 w2,
                 topk_weights,
                 topk_ids,
-                quant_config=quant_config,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_fp8_scale,
             )
 
     def run_cutlass_moe_fp4(
@@ -179,27 +172,25 @@ def bench_run(
         device: torch.device,
         num_repeats: int,
     ):
-        quant_config = nvfp4_moe_quant_config(
-            a1_gscale=a1_gs,
-            a2_gscale=a2_gs,
-            w1_scale=w1_blockscale,
-            w2_scale=w2_blockscale,
-            g1_alphas=w1_gs,
-            g2_alphas=w2_gs,
-        )
         for _ in range(num_repeats):
             with nvtx.annotate("cutlass_moe_fp4", color="green"):
                 cutlass_moe_fp4(
                     a=a,
+                    a1_gscale=a1_gs,
+                    a2_gscale=a2_gs,
                     w1_fp4=w1_fp4,
+                    w1_blockscale=w1_blockscale,
+                    w1_alphas=w1_gs,
                     w2_fp4=w2_fp4,
+                    w2_blockscale=w2_blockscale,
+                    w2_alphas=w2_gs,
                     topk_weights=topk_weights,
                     topk_ids=topk_ids,
                     m=m,
                     n=n,
                     k=k,
                     e=num_experts,
-                    quant_config=quant_config,
+                    device=device,
                 )
 
     def run_cutlass_from_graph(
@@ -220,29 +211,26 @@ def bench_run(
         e: int,
         device: torch.device,
     ):
-        quant_config = nvfp4_moe_quant_config(
-            a1_gscale=a1_gs,
-            a2_gscale=a2_gs,
-            w1_scale=w1_blockscale,
-            w2_scale=w2_blockscale,
-            g1_alphas=w1_gs,
-            g2_alphas=w2_gs,
-        )
-
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
         ):
             return cutlass_moe_fp4(
                 a=a,
+                a1_gscale=a1_gs,
                 w1_fp4=w1_fp4,
+                w1_blockscale=w1_blockscale,
+                w1_alphas=w1_alphas,
+                a2_gscale=a2_gs,
                 w2_fp4=w2_fp4,
+                w2_blockscale=w2_blockscale,
+                w2_alphas=w2_alphas,
                 topk_weights=topk_weights,
                 topk_ids=topk_ids,
                 m=m,
                 n=n,
                 k=k,
                 e=num_experts,
-                quant_config=quant_config,
+                device=device,
             )
 
     def run_triton_from_graph(
@@ -258,18 +246,16 @@ def bench_run(
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
         ):
-            quant_config = fp8_w8a8_moe_quant_config(
-                w1_scale=w1_scale,
-                w2_scale=w2_scale,
-                a1_scale=a_fp8_scale,
-            )
             return fused_experts(
                 a,
                 w1,
                 w2,
                 topk_weights,
                 topk_ids,
-                quant_config=quant_config,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_fp8_scale,
             )
 
     def replay_graph(graph, num_repeats):

benchmarks/kernels/benchmark_cutlass_moe_fp8.py

@@ -1,406 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-"""
-Benchmark the performance of the cutlass_moe_fp8 kernel vs the triton_moe
-kernel. Both kernels take in fp8 quantized weights and 16-bit activations,
-but use different quantization strategies and backends.
-"""
-
-import nvtx
-import torch
-
-from vllm import _custom_ops as ops
-from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
-from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8
-from vllm.model_executor.layers.fused_moe.fused_moe import fused_experts, fused_topk
-from vllm.platforms import current_platform
-from vllm.utils import FlexibleArgumentParser
-
-# Weight shapes for different models: [num_experts, topk, hidden_size,
-# intermediate_size]
-WEIGHT_SHAPES_MOE = {
-    "mixtral-8x7b": [
-        [8, 2, 4096, 14336],
-    ],
-    "deepseek-v2": [
-        [160, 6, 5120, 12288],
-    ],
-    "custom-small": [
-        [8, 2, 2048, 7168],
-    ],
-    "glm45-fp8": [
-        [128, 8, 4096, 1408],
-    ],
-    "Llama-4-Maverick-17B-128E-Instruct-FP8": [
-        [128, 1, 5120, 8192],
-    ],
-}
-
-DEFAULT_MODELS = [
-    "mixtral-8x7b",
-]
-
-DEFAULT_BATCH_SIZES = [4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048]
-DEFAULT_TP_SIZES = [1]
-
-PER_ACT_TOKEN_OPTS = [False, True]
-PER_OUT_CH_OPTS = [False, True]
-
-FP8_DTYPE = current_platform.fp8_dtype()
-
-
-def bench_run(
-    results: list,
-    model: str,
-    num_experts: int,
-    topk: int,
-    per_act_token: bool,
-    per_out_ch: bool,
-    mkn: tuple[int, int, int],
-):
-    (m, k, n) = mkn
-
-    dtype = torch.half
-    device = "cuda"
-
-    # Create input activations
-    a = torch.randn((m, k), device=device, dtype=dtype) / 10
-
-    # Create weights
-    w1 = torch.randn((num_experts, 2 * n, k), device=device, dtype=dtype) / 10
-    w2 = torch.randn((num_experts, k, n), device=device, dtype=dtype) / 10
-
-    # Create FP8 quantized weights and scales for both kernels
-    w1_fp8q = torch.empty((num_experts, 2 * n, k), device=device, dtype=FP8_DTYPE)
-    w2_fp8q = torch.empty((num_experts, k, n), device=device, dtype=FP8_DTYPE)
-
-    # Create scales based on quantization strategy
-    if per_out_ch:
-        # Per-channel quantization
-        w1_scale = torch.empty(
-            (num_experts, 2 * n, 1), device=device, dtype=torch.float32
-        )
-        w2_scale = torch.empty((num_experts, k, 1), device=device, dtype=torch.float32)
-    else:
-        # Per-tensor quantization
-        w1_scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
-        w2_scale = torch.empty((num_experts, 1, 1), device=device, dtype=torch.float32)
-
-    # Quantize weights
-    for expert in range(num_experts):
-        if per_out_ch:
-            # Per-channel quantization - not yet implemented properly
-            # For now, fall back to per-tensor quantization
-            w1_fp8q[expert], w1_scale_temp = ops.scaled_fp8_quant(w1[expert])
-            w2_fp8q[expert], w2_scale_temp = ops.scaled_fp8_quant(w2[expert])
-            # Expand scalar scales to the expected per-channel shape
-            w1_scale[expert] = w1_scale_temp.expand(2 * n, 1)
-            w2_scale[expert] = w2_scale_temp.expand(k, 1)
-        else:
-            # Per-tensor quantization
-            w1_fp8q[expert], w1_scale_temp = ops.scaled_fp8_quant(w1[expert])
-            w2_fp8q[expert], w2_scale_temp = ops.scaled_fp8_quant(w2[expert])
-            # Store scalar scales in [1, 1] tensors
-            w1_scale[expert, 0, 0] = w1_scale_temp
-            w2_scale[expert, 0, 0] = w2_scale_temp
-
-    # Prepare weights for CUTLASS (no transpose needed)
-    w1_fp8q_cutlass = w1_fp8q  # Keep original [E, 2N, K]
-    w2_fp8q_cutlass = w2_fp8q  # Keep original [E, K, N]
-
-    # Create router scores and get topk
-    score = torch.randn((m, num_experts), device=device, dtype=dtype)
-    topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False)
-
-    # WORKAROUND: CUTLASS MoE FP8 has issues with per-token quantization
-    # Force per-tensor quantization for all cases to match working e2e setup
-    a1_scale = torch.full((), 1e-2, device=device, dtype=torch.float32)
-    a2_scale = torch.full((), 1e-2, device=device, dtype=torch.float32)
-
-    # Force per-tensor quantization for all cases
-    per_act_token = False
-
-    # Create stride tensors for CUTLASS
-    ab_strides1 = torch.full((num_experts,), k, dtype=torch.int64, device=device)
-    ab_strides2 = torch.full((num_experts,), n, dtype=torch.int64, device=device)
-    c_strides1 = torch.full((num_experts,), 2 * n, dtype=torch.int64, device=device)
-    c_strides2 = torch.full((num_experts,), k, dtype=torch.int64, device=device)
-
-    def run_triton_moe(
-        a: torch.Tensor,
-        w1: torch.Tensor,
-        w2: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        w1_scale: torch.Tensor,
-        w2_scale: torch.Tensor,
-        a1_scale: torch.Tensor,
-        a2_scale: torch.Tensor,
-        num_repeats: int,
-    ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            per_act_token_quant=per_act_token,
-            per_out_ch_quant=per_out_ch,
-        )
-
-        for _ in range(num_repeats):
-            fused_experts(
-                a,
-                w1,
-                w2,
-                topk_weights,
-                topk_ids,
-                quant_config=quant_config,
-            )
-
-    def run_cutlass_moe_fp8(
-        a: torch.Tensor,
-        w1: torch.Tensor,
-        w2: torch.Tensor,
-        topk_weights: torch.Tensor,
-        topk_ids: torch.Tensor,
-        ab_strides1: torch.Tensor,
-        ab_strides2: torch.Tensor,
-        c_strides1: torch.Tensor,
-        c_strides2: torch.Tensor,
-        w1_scale: torch.Tensor,
-        w2_scale: torch.Tensor,
-        a1_scale: torch.Tensor,
-        a2_scale: torch.Tensor,
-        num_repeats: int,
-    ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            per_act_token_quant=per_act_token,
-            per_out_ch_quant=per_out_ch,
-        )
-
-        for _ in range(num_repeats):
-            with nvtx.annotate("cutlass_moe_fp8", color="blue"):
-                cutlass_moe_fp8(
-                    a=a,
-                    w1_q=w1,
-                    w2_q=w2,
-                    topk_weights=topk_weights,
-                    topk_ids=topk_ids,
-                    ab_strides1=ab_strides1,
-                    ab_strides2=ab_strides2,
-                    c_strides1=c_strides1,
-                    c_strides2=c_strides2,
-                    quant_config=quant_config,
-                    activation="silu",
-                    global_num_experts=num_experts,
-                )
-
-    # Pre-create quantization config to avoid creating it inside CUDA graph
-    quant_config = fp8_w8a8_moe_quant_config(
-        w1_scale=w1_scale,
-        w2_scale=w2_scale,
-        a1_scale=a1_scale,
-        a2_scale=a2_scale,
-        per_act_token_quant=per_act_token,
-        per_out_ch_quant=per_out_ch,
-    )
-
-    # Create CUDA graphs for CUTLASS (match benchmark_moe.py pattern exactly)
-    cutlass_stream = torch.cuda.Stream()
-    cutlass_graph = torch.cuda.CUDAGraph()
-    with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
-        # Capture 10 invocations like benchmark_moe.py
-        for _ in range(10):
-            cutlass_moe_fp8(
-                a=a,
-                w1_q=w1_fp8q_cutlass,
-                w2_q=w2_fp8q_cutlass,
-                topk_weights=topk_weights,
-                topk_ids=topk_ids,
-                ab_strides1=ab_strides1,
-                ab_strides2=ab_strides2,
-                c_strides1=c_strides1,
-                c_strides2=c_strides2,
-                quant_config=quant_config,
-                activation="silu",
-                global_num_experts=num_experts,
-            )
-    torch.cuda.synchronize()
-
-    # Create CUDA graphs for Triton (match benchmark_moe.py pattern exactly)
-    triton_stream = torch.cuda.Stream()
-    triton_graph = torch.cuda.CUDAGraph()
-    with torch.cuda.graph(triton_graph, stream=triton_stream):
-        # Capture 10 invocations like benchmark_moe.py
-        for _ in range(10):
-            fused_experts(
-                a,
-                w1_fp8q,
-                w2_fp8q,
-                topk_weights,
-                topk_ids,
-                quant_config=quant_config,
-            )
-    torch.cuda.synchronize()
-
-    def bench_cuda_graph(graph, num_warmup=5, num_iters=100):
-        """Benchmark CUDA graph using events like benchmark_moe.py"""
-        # Warmup
-        for _ in range(num_warmup):
-            graph.replay()
-        torch.cuda.synchronize()
-
-        # Timing
-        start_event = torch.cuda.Event(enable_timing=True)
-        end_event = torch.cuda.Event(enable_timing=True)
-
-        latencies = []
-        for _ in range(num_iters):
-            torch.cuda.synchronize()
-            start_event.record()
-            graph.replay()
-            end_event.record()
-            end_event.synchronize()
-            latencies.append(start_event.elapsed_time(end_event))
-
-        # Divide by 10 since graph contains 10 calls
-        return sum(latencies) / (num_iters * 10)
-
-    # Benchmark parameters
-    num_warmup = 5
-    num_iters = 100
-
-    # Benchmark only CUDA graphs (more reliable and faster)
-    # Benchmark Triton MoE with CUDA graphs
-    triton_graph_time = bench_cuda_graph(
-        triton_graph, num_warmup=num_warmup, num_iters=num_iters
-    )
-
-    # Benchmark CUTLASS MoE with CUDA graphs
-    cutlass_graph_time = bench_cuda_graph(
-        cutlass_graph, num_warmup=num_warmup, num_iters=num_iters
-    )
-
-    # Convert ms to us and return results
-    triton_time_us = triton_graph_time * 1000
-    cutlass_time_us = cutlass_graph_time * 1000
-
-    return {
-        "batch_size": m,
-        "triton_time_us": triton_time_us,
-        "cutlass_time_us": cutlass_time_us,
-    }
-
-
-def main(args):
-    print("Benchmarking models:")
-    for i, model in enumerate(args.models):
-        print(f"[{i}]  {model}")
-
-    all_results = []
-
-    for model in args.models:
-        for tp in args.tp_sizes:
-            for layer in WEIGHT_SHAPES_MOE[model]:
-                num_experts = layer[0]
-                topk = layer[1]
-                size_k = layer[2]
-                size_n = layer[3] // tp
-
-                if len(args.limit_k) > 0 and size_k not in args.limit_k:
-                    continue
-
-                if len(args.limit_n) > 0 and size_n not in args.limit_n:
-                    continue
-
-                for per_act_token in args.per_act_token_opts:
-                    for per_out_ch in args.per_out_ch_opts:
-                        print(
-                            f"\n=== {model}, experts={num_experts}, topk={topk},"
-                            f"per_act={per_act_token}, per_out_ch={per_out_ch} ==="
-                        )
-
-                        config_results = []
-                        for size_m in args.batch_sizes:
-                            mkn = (size_m, size_k, size_n)
-                            result = bench_run(
-                                [],  # Not used anymore
-                                model,
-                                num_experts,
-                                topk,
-                                per_act_token,
-                                per_out_ch,
-                                mkn,
-                            )
-                            if result:
-                                config_results.append(result)
-
-                        # Print results table for this configuration
-                        if config_results:
-                            print(
-                                f"\n{'Batch Size':<12}"
-                                f"{'Triton (us)':<15}"
-                                f"{'CUTLASS (us)':<15}"
-                            )
-                            print("-" * 45)
-                            for result in config_results:
-                                print(
-                                    f"{result['batch_size']:<12}"
-                                    f"{result['triton_time_us']:<15.2f}"
-                                    f"{result['cutlass_time_us']:<15.2f}"
-                                )
-
-                            all_results.extend(config_results)
-
-    print(f"\nTotal benchmarks completed: {len(all_results)}")
-
-
-if __name__ == "__main__":
-    parser = FlexibleArgumentParser(
-        description="""Benchmark CUTLASS FP8 MOE vs Triton FP8 FUSED MOE
-         across specified models/shapes/batches
-
-        Example usage:
-        python benchmark_cutlass_moe_fp8.py  \
-            --model "Llama-4-Maverick-17B-128E-Instruct-FP8"  \
-            --tp-sizes 8 \
-            --batch-size 2 4 8  \
-            --per-act-token-opts false \
-            --per-out-ch-opts false
-
-        """
-    )
-    parser.add_argument(
-        "--models",
-        nargs="+",
-        type=str,
-        default=DEFAULT_MODELS,
-        choices=WEIGHT_SHAPES_MOE.keys(),
-    )
-    parser.add_argument("--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES)
-    parser.add_argument(
-        "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES
-    )
-    parser.add_argument("--limit-k", nargs="+", type=int, default=[])
-    parser.add_argument("--limit-n", nargs="+", type=int, default=[])
-    parser.add_argument(
-        "--per-act-token-opts",
-        nargs="+",
-        type=lambda x: x.lower() == "true",
-        default=[False, True],
-        help="Per-activation token quantization options (true/false)",
-    )
-    parser.add_argument(
-        "--per-out-ch-opts",
-        nargs="+",
-        type=lambda x: x.lower() == "true",
-        default=[False, True],
-        help="Per-output channel quantization options (true/false)",
-    )
-
-    args = parser.parse_args()
-    main(args)

benchmarks/kernels/benchmark_device_communicators.py

@@ -1,508 +0,0 @@
-#!/usr/bin/env python3
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-"""
-Benchmark script for device communicators:
-CustomAllreduce (oneshot, twoshot), PyNcclCommunicator,
-and SymmMemCommunicator (multimem, two-shot).
-
-for NCCL symmetric memory you need to set the environment variables
-NCCL_NVLS_ENABLE=1 NCCL_CUMEM_ENABLE=1 VLLM_USE_NCCL_SYMM_MEM=1, otherwise NCCL does
-not use fast NVLS implementation for all reduce.
-
-Usage:
-    torchrun --nproc_per_node=<N> benchmark_device_communicators.py [options]
-
-Example:
-    torchrun --nproc_per_node=2 benchmark_device_communicators.py
-    --sequence-lengths 512 1024 2048 --num-warmup 10 --num-trials 100
-"""
-
-import json
-import os
-import time
-from contextlib import nullcontext
-from typing import Callable, Optional
-
-import torch
-import torch.distributed as dist
-from torch.distributed import ProcessGroup
-
-from vllm.distributed.device_communicators.custom_all_reduce import CustomAllreduce
-from vllm.distributed.device_communicators.pynccl import (
-    PyNcclCommunicator,
-    register_nccl_symmetric_ops,
-)
-from vllm.distributed.device_communicators.pynccl_allocator import (
-    set_graph_pool_id,
-)
-from vllm.distributed.device_communicators.symm_mem import SymmMemCommunicator
-from vllm.logger import init_logger
-from vllm.utils import FlexibleArgumentParser
-
-logger = init_logger(__name__)
-
-# Default sequence lengths to benchmark
-DEFAULT_SEQUENCE_LENGTHS = [128, 512, 1024, 2048, 4096, 8192]
-
-# Fixed hidden size and dtype for all benchmarks
-HIDDEN_SIZE = 8192
-BENCHMARK_DTYPE = torch.bfloat16
-
-# CUDA graph settings
-CUDA_GRAPH_CAPTURE_CYCLES = 10
-
-
-class CommunicatorBenchmark:
-    """Benchmark class for testing device communicators."""
-
-    def __init__(
-        self,
-        rank: int,
-        world_size: int,
-        device: torch.device,
-        cpu_group: ProcessGroup,
-        sequence_lengths: list[int],
-    ):
-        self.rank = rank
-        self.world_size = world_size
-        self.device = device
-        self.cpu_group = cpu_group
-
-        # Calculate max_size_override based on largest sequence length
-        max_seq_len = max(sequence_lengths)
-        max_tensor_elements = max_seq_len * HIDDEN_SIZE
-        self.max_size_override = max_tensor_elements * BENCHMARK_DTYPE.itemsize + 1
-
-        # Initialize communicators
-        self.custom_allreduce = None
-        self.pynccl_comm = None
-        self.symm_mem_comm = None
-        self.symm_mem_comm_multimem = None
-        self.symm_mem_comm_two_shot = None
-
-        self._init_communicators()
-
-    def _init_communicators(self):
-        """Initialize all available communicators."""
-        try:
-            self.custom_allreduce = CustomAllreduce(
-                group=self.cpu_group,
-                device=self.device,
-                max_size=self.max_size_override,
-            )
-            if not self.custom_allreduce.disabled:
-                logger.info("Rank %s: CustomAllreduce initialized", self.rank)
-            else:
-                logger.info("Rank %s: CustomAllreduce disabled", self.rank)
-        except Exception as e:
-            logger.warning(
-                "Rank %s: Failed to initialize CustomAllreduce: %s", self.rank, e
-            )
-            self.custom_allreduce = None
-
-        try:
-            self.pynccl_comm = PyNcclCommunicator(
-                group=self.cpu_group, device=self.device
-            )
-            if not self.pynccl_comm.disabled:
-                logger.info("Rank %s: PyNcclCommunicator initialized", self.rank)
-                register_nccl_symmetric_ops(self.pynccl_comm)
-            else:
-                logger.info("Rank %s: PyNcclCommunicator disabled", self.rank)
-                self.pynccl_comm = None
-        except Exception as e:
-            logger.warning(
-                "Rank %s: Failed to initialize PyNcclCommunicator: %s", self.rank, e
-            )
-            self.pynccl_comm = None
-
-        # Initialize variants for SymmMemCommunicator
-        try:
-            self.symm_mem_comm_multimem = SymmMemCommunicator(
-                group=self.cpu_group,
-                device=self.device,
-                force_multimem=True,
-                max_size_override=self.max_size_override,
-            )
-            if not self.symm_mem_comm_multimem.disabled:
-                logger.info(
-                    "Rank %s: SymmMemCommunicator (multimem) initialized", self.rank
-                )
-            else:
-                self.symm_mem_comm_multimem = None
-        except Exception as e:
-            logger.warning(
-                "Rank %s: Failed to initialize SymmMemCommunicator (multimem): %s",
-                self.rank,
-                e,
-            )
-            self.symm_mem_comm_multimem = None
-
-        try:
-            self.symm_mem_comm_two_shot = SymmMemCommunicator(
-                group=self.cpu_group,
-                device=self.device,
-                force_multimem=False,
-                max_size_override=self.max_size_override,
-            )
-            if not self.symm_mem_comm_two_shot.disabled:
-                logger.info(
-                    "Rank %s: SymmMemCommunicator (two_shot) initialized", self.rank
-                )
-            else:
-                self.symm_mem_comm_two_shot = None
-        except Exception as e:
-            logger.warning(
-                "Rank %s: Failed to initialize SymmMemCommunicator (two_shot): %s",
-                self.rank,
-                e,
-            )
-            self.symm_mem_comm_two_shot = None
-
-    def benchmark_allreduce(
-        self, sequence_length: int, num_warmup: int, num_trials: int
-    ) -> dict[str, float]:
-        """Benchmark allreduce operations for all available communicators."""
-
-        results = {}
-
-        # Define communicators with their benchmark functions
-        communicators = []
-
-        if self.custom_allreduce is not None:
-            comm = self.custom_allreduce
-            # CustomAllreduce one-shot
-            communicators.append(
-                (
-                    "ca_1stage",
-                    lambda t, c=comm: c.custom_all_reduce(t),
-                    lambda t, c=comm: c.should_custom_ar(t),
-                    comm.capture(),
-                    "1stage",  # env variable value
-                )
-            )
-            # CustomAllreduce two-shot
-            communicators.append(
-                (
-                    "ca_2stage",
-                    lambda t, c=comm: c.custom_all_reduce(t),
-                    lambda t, c=comm: c.should_custom_ar(t),
-                    comm.capture(),
-                    "2stage",  # env variable value
-                )
-            )
-
-        if self.pynccl_comm is not None:
-            comm = self.pynccl_comm
-            communicators.append(
-                (
-                    "pynccl",
-                    lambda t, c=comm: c.all_reduce(t),
-                    lambda t: True,  # Always available if initialized
-                    nullcontext(),
-                    None,  # no env variable needed
-                )
-            )
-            communicators.append(
-                (
-                    "pynccl-symm",
-                    lambda t: torch.ops.vllm.all_reduce_symmetric_with_copy(t),
-                    lambda t: True,  # Always available if initialized
-                    nullcontext(),
-                    None,  # no env variable needed
-                )
-            )
-
-        if self.symm_mem_comm_multimem is not None:
-            comm = self.symm_mem_comm_multimem
-            communicators.append(
-                (
-                    "symm_mem_multimem",
-                    lambda t, c=comm: c.all_reduce(t),
-                    lambda t, c=comm: c.should_use_symm_mem(t),
-                    nullcontext(),
-                    None,  # no env variable needed
-                )
-            )
-
-        if self.symm_mem_comm_two_shot is not None:
-            comm = self.symm_mem_comm_two_shot
-            communicators.append(
-                (
-                    "symm_mem_two_shot",
-                    lambda t, c=comm: c.all_reduce(t),
-                    lambda t, c=comm: c.should_use_symm_mem(t),
-                    nullcontext(),
-                    None,  # no env variable needed
-                )
-            )
-
-        # Benchmark each communicator
-        for name, allreduce_fn, should_use_fn, context, env_var in communicators:
-            # Set environment variable if needed
-            if env_var is not None:
-                os.environ["VLLM_CUSTOM_ALLREDUCE_ALGO"] = env_var
-            else:
-                # Clear the environment variable to avoid interference
-                os.environ.pop("VLLM_CUSTOM_ALLREDUCE_ALGO", None)
-
-            latency = self.benchmark_allreduce_single(
-                sequence_length,
-                allreduce_fn,
-                should_use_fn,
-                context,
-                num_warmup,
-                num_trials,
-            )
-            if latency is not None:
-                results[name] = latency
-
-        return results
-
-    def benchmark_allreduce_single(
-        self,
-        sequence_length: int,
-        allreduce_fn: Callable[[torch.Tensor], Optional[torch.Tensor]],
-        should_use_fn: Callable[[torch.Tensor], bool],
-        context,
-        num_warmup: int,
-        num_trials: int,
-    ) -> Optional[float]:
-        """Benchmark method with CUDA graph optimization."""
-        try:
-            # Create test tensor (2D: sequence_length x hidden_size)
-            tensor = torch.randn(
-                sequence_length, HIDDEN_SIZE, dtype=BENCHMARK_DTYPE, device=self.device
-            )
-            if not should_use_fn(tensor):
-                return None
-
-            torch.cuda.synchronize()
-            stream = torch.cuda.Stream()
-            with torch.cuda.stream(stream):
-                graph_input = tensor.clone()
-
-                # Warmup before capture
-                for _ in range(3):
-                    allreduce_fn(graph_input)
-
-                # Capture the graph using context manager
-                with context:
-                    graph = torch.cuda.CUDAGraph()
-                    graph_pool = torch.cuda.graph_pool_handle()
-                    set_graph_pool_id(graph_pool)
-                    with torch.cuda.graph(graph, pool=graph_pool):
-                        for _ in range(CUDA_GRAPH_CAPTURE_CYCLES):
-                            allreduce_fn(graph_input)
-
-            torch.cuda.synchronize()
-            for _ in range(num_warmup):
-                graph.replay()
-            torch.cuda.synchronize()
-
-            torch.cuda.synchronize()
-            start_time = time.perf_counter()
-
-            for _ in range(num_trials):
-                graph.replay()
-            torch.cuda.synchronize()
-
-            end_time = time.perf_counter()
-
-            # Convert to ms and divide by CUDA_GRAPH_CAPTURE_CYCLES
-            return (
-                (end_time - start_time) / num_trials / CUDA_GRAPH_CAPTURE_CYCLES * 1000
-            )
-
-        except Exception as e:
-            logger.error("CUDA graph benchmark failed: %s", e)
-            raise RuntimeError(
-                f"CUDA graph benchmark failed for communicator: {e}"
-            ) from e
-
-
-def _calculate_speedup_info(comm_results: dict[str, float]) -> str:
-    """Calculate speedup information for a single tensor size."""
-    if not comm_results:
-        return "N/A"
-
-    # Find the fastest communicator
-    fastest_comm = min(comm_results.keys(), key=lambda k: comm_results[k])
-    fastest_time = comm_results[fastest_comm]
-
-    # Calculate speedup vs PyNccl if available
-    if "pynccl" in comm_results:
-        pynccl_time = comm_results["pynccl"]
-        speedup = pynccl_time / fastest_time
-        return f"{fastest_comm} ({speedup:.2f}x)"
-    else:
-        return f"{fastest_comm} (N/A)"
-
-
-def print_results(
-    results: dict[str, dict[str, float]], sequence_lengths: list[int], world_size: int
-):
-    """Print benchmark results in a formatted table."""
-
-    print(f"\n{'=' * 130}")
-    print("Device Communicator Benchmark Results")
-    print(
-        f"World Size: {world_size}, Data Type: {BENCHMARK_DTYPE}, "
-        f"Hidden Size: {HIDDEN_SIZE}"
-    )
-    print(f"{'=' * 130}")
-
-    # Get all communicator names
-    all_comms = set()
-    for size_results in results.values():
-        all_comms.update(size_results.keys())
-
-    all_comms = sorted(list(all_comms))
-
-    # Print header
-    header = f"{'Tensor Shape':<20}{'Tensor Size':<15}"
-    for comm in all_comms:
-        header += f"{comm:<20}"
-    header += f"{'Best (Speedup vs PyNccl)':<30}"
-    print(header)
-    print("-" * len(header))
-
-    # Print results for each sequence length
-    for seq_len in sequence_lengths:
-        if seq_len in results:
-            # Calculate tensor size in elements and bytes
-            tensor_elements = seq_len * HIDDEN_SIZE
-            tensor_bytes = tensor_elements * BENCHMARK_DTYPE.itemsize
-
-            # Format tensor size (MB)
-            tensor_size_mb = tensor_bytes / (1024 * 1024)
-            tensor_size_str = f"{tensor_size_mb:.2f} MB"
-
-            # Format tensor shape
-            tensor_shape = f"({seq_len}, {HIDDEN_SIZE})"
-
-            row = f"{tensor_shape:<20}{tensor_size_str:<15}"
-            for comm in all_comms:
-                if comm in results[seq_len]:
-                    row += f"{results[seq_len][comm]:<20.3f}"
-                else:
-                    row += f"{'N/A':<20}"
-
-            # Calculate speedup information
-            speedup_info = _calculate_speedup_info(results[seq_len])
-            row += f"{speedup_info:<30}"
-
-            print(row)
-
-    print(f"{'=' * 130}")
-    print("All times are in milliseconds (ms) per allreduce operation")
-    print("Speedup column shows: fastest_algorithm (speedup_vs_pynccl)")
-
-
-def main():
-    parser = FlexibleArgumentParser(description="Benchmark device communicators")
-
-    parser.add_argument(
-        "--sequence-lengths",
-        type=int,
-        nargs="+",
-        default=DEFAULT_SEQUENCE_LENGTHS,
-        help="Sequence lengths to benchmark (tensor shape: seq_len x hidden_size)",
-    )
-
-    parser.add_argument(
-        "--num-warmup", type=int, default=5, help="Number of warmup iterations"
-    )
-
-    parser.add_argument(
-        "--num-trials", type=int, default=50, help="Number of benchmark trials"
-    )
-
-    parser.add_argument("--output-json", type=str, help="Output results to JSON file")
-
-    args = parser.parse_args()
-
-    # Initialize distributed
-    if not dist.is_initialized():
-        dist.init_process_group(backend="gloo")
-    rank = dist.get_rank()
-    world_size = dist.get_world_size()
-
-    # Set device
-    device = torch.device(f"cuda:{rank}")
-    torch.cuda.set_device(device)
-
-    # Get CPU process group
-    cpu_group = dist.new_group(backend="gloo")
-
-    # Disable USE_SYMM_MEM to avoid affecting the max_sizes
-    # in symm_mem and custom_all_reduce for benchmark
-    os.environ["VLLM_ALLREDUCE_USE_SYMM_MEM"] = "0"
-
-    # Initialize benchmark
-    benchmark = CommunicatorBenchmark(
-        rank, world_size, device, cpu_group, args.sequence_lengths
-    )
-
-    # Run benchmarks
-    all_results = {}
-
-    for seq_len in args.sequence_lengths:
-        if rank == 0:
-            logger.info(
-                "Benchmarking sequence length: %s (tensor shape: %s x %s)",
-                seq_len,
-                seq_len,
-                HIDDEN_SIZE,
-            )
-
-        results = benchmark.benchmark_allreduce(
-            sequence_length=seq_len,
-            num_warmup=args.num_warmup,
-            num_trials=args.num_trials,
-        )
-
-        all_results[seq_len] = results
-
-        # Synchronize between ranks
-        dist.barrier()
-
-    # Print results (only rank 0)
-    if rank == 0:
-        print_results(all_results, args.sequence_lengths, world_size)
-
-        # Save to JSON if requested
-        if args.output_json:
-            # Add speedup information to results
-            enhanced_results = {}
-            for seq_len, comm_results in all_results.items():
-                enhanced_results[seq_len] = {
-                    "timings": comm_results,
-                    "speedup_info": _calculate_speedup_info(comm_results),
-                }
-
-            output_data = {
-                "world_size": world_size,
-                "dtype": str(BENCHMARK_DTYPE),
-                "hidden_size": HIDDEN_SIZE,
-                "sequence_lengths": args.sequence_lengths,
-                "num_warmup": args.num_warmup,
-                "num_trials": args.num_trials,
-                "cuda_graph_capture_cycles": CUDA_GRAPH_CAPTURE_CYCLES,
-                "results": enhanced_results,
-            }
-
-            with open(args.output_json, "w") as f:
-                json.dump(output_data, f, indent=2)
-
-            logger.info("Results saved to %s", args.output_json)
-
-    # Cleanup
-    if cpu_group != dist.group.WORLD:
-        dist.destroy_process_group(cpu_group)
-
-
-if __name__ == "__main__":
-    main()

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -7,7 +7,6 @@ from benchmark_shapes import WEIGHT_SHAPES_MOE
 
 from vllm import _custom_ops as ops
 from vllm.config import ParallelConfig, VllmConfig, set_current_vllm_config
-from vllm.model_executor.layers.fused_moe.config import fp8_w8a8_moe_quant_config
 from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp8
 from vllm.model_executor.layers.fused_moe.fused_moe import (
     fused_experts,
@@ -97,11 +96,6 @@ def bench_run(
         a_scale: torch.Tensor,
         num_repeats: int,
     ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a_scale,
-        )
         for _ in range(num_repeats):
             fused_experts(
                 a,
@@ -109,7 +103,10 @@ def bench_run(
                 w2,
                 topk_weights,
                 topk_ids,
-                quant_config=quant_config,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_scale,
             )
 
     def run_cutlass_moe(
@@ -128,12 +125,6 @@ def bench_run(
         per_act_token: bool,
         num_repeats: int,
     ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            per_act_token_quant=per_act_token,
-        )
-
         for _ in range(num_repeats):
             cutlass_moe_fp8(
                 a,
@@ -141,11 +132,14 @@ def bench_run(
                 w2,
                 topk_weights,
                 topk_ids,
+                w1_scale,
+                w2_scale,
                 ab_strides1,
                 ab_strides2,
                 c_strides1,
                 c_strides2,
-                quant_config=quant_config,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_cutlass_from_graph(
@@ -162,12 +156,6 @@ def bench_run(
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
     ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            per_act_token_quant=per_act_token,
-        )
-
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
         ):
@@ -177,11 +165,14 @@ def bench_run(
                 w2_q,
                 topk_weights,
                 topk_ids,
+                w1_scale,
+                w2_scale,
                 ab_strides1,
                 ab_strides2,
                 c_strides1,
                 c_strides2,
-                quant_config=quant_config,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_triton_from_graph(
@@ -194,11 +185,6 @@ def bench_run(
         w2_scale: torch.Tensor,
         a_scale: torch.Tensor,
     ):
-        quant_config = fp8_w8a8_moe_quant_config(
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a_scale,
-        )
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
         ):
@@ -208,7 +194,10 @@ def bench_run(
                 w2,
                 topk_weights,
                 topk_ids,
-                quant_config=quant_config,
+                use_fp8_w8a8=True,
+                w1_scale=w1_scale,
+                w2_scale=w2_scale,
+                a1_scale=a_scale,
             )
 
     def replay_graph(graph, num_repeats):

benchmarks/kernels/benchmark_lora.py

@@ -79,9 +79,9 @@ def make_rand_lora_weight_tensor(
 
 
 def make_rand_tensors(
-    a_shape: tuple[int, ...],
-    b_shape: tuple[int, ...],
-    c_shape: tuple[int, ...],
+    a_shape: tuple[int],
+    b_shape: tuple[int],
+    c_shape: tuple[int],
     a_dtype: torch.dtype,
     b_dtype: torch.dtype,
     c_dtype: torch.dtype,
@@ -243,7 +243,7 @@ class OpType(Enum):
         lora_rank: int,
         num_loras: int,
         num_slices: int,
-    ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]:
+    ) -> tuple[tuple[int], tuple[int], tuple[int]]:
         """
         Given num_slices, return the shapes of the A, B, and C matrices
         in A x B = C, for the op_type
@@ -464,11 +464,7 @@ class BenchmarkTensors:
         for field_name in LoRAKernelMeta.__dataclass_fields__:
             field = getattr(self.lora_kernel_meta, field_name)
             assert isinstance(field, torch.Tensor)
-            setattr(
-                self.lora_kernel_meta,
-                field_name,
-                to_device(field) if field_name != "no_lora_flag_cpu" else field,
-            )
+            setattr(self.lora_kernel_meta, field_name, to_device(field))
 
     def metadata(self) -> tuple[int, int, int]:
         """
@@ -516,7 +512,6 @@ class BenchmarkTensors:
             "lora_token_start_loc": self.lora_kernel_meta.lora_token_start_loc,
             "lora_ids": self.lora_kernel_meta.active_lora_ids,
             "scaling": 1.0,
-            "no_lora_flag_cpu": self.lora_kernel_meta.no_lora_flag_cpu,
         }
 
     def as_lora_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]:
@@ -557,7 +552,6 @@ class BenchmarkTensors:
             "lora_ids": self.lora_kernel_meta.active_lora_ids,
             "offset_start": 0,
             "add_inputs": add_inputs,
-            "no_lora_flag_cpu": self.lora_kernel_meta.no_lora_flag_cpu,
         }
 
     def bench_fn_kwargs(

benchmarks/kernels/benchmark_moe.py

@@ -14,10 +14,6 @@ import ray
 import torch
 from ray.experimental.tqdm_ray import tqdm
 
-from vllm.model_executor.layers.fused_moe.config import (
-    FusedMoEQuantConfig,
-    _get_config_dtype_str,
-)
 from vllm.model_executor.layers.fused_moe.fused_moe import *
 from vllm.platforms import current_platform
 from vllm.transformers_utils.config import get_config
@@ -138,36 +134,43 @@ def benchmark_config(
     def run():
         from vllm.model_executor.layers.fused_moe import override_config
 
-        if use_fp8_w8a8:
-            quant_dtype = torch.float8_e4m3fn
-        elif use_int8_w8a16:
-            quant_dtype = torch.int8
-        else:
-            quant_dtype = None
-
-        quant_config = FusedMoEQuantConfig.make(
-            quant_dtype=quant_dtype,
-            w1_scale=w1_scale,
-            w2_scale=w2_scale,
-            a1_scale=a1_scale,
-            a2_scale=a2_scale,
-            block_shape=block_quant_shape,
-        )
-
         with override_config(config):
-            topk_weights, topk_ids, token_expert_indices = fused_topk(
-                x, input_gating, topk, renormalize=not use_deep_gemm
-            )
-            return fused_experts(
-                x,
-                w1,
-                w2,
-                topk_weights,
-                topk_ids,
-                inplace=True,
-                quant_config=quant_config,
-                allow_deep_gemm=use_deep_gemm,
-            )
+            if use_deep_gemm:
+                topk_weights, topk_ids, token_expert_indices = fused_topk(
+                    x, input_gating, topk, False
+                )
+                return fused_experts(
+                    x,
+                    w1,
+                    w2,
+                    topk_weights,
+                    topk_ids,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                    allow_deep_gemm=True,
+                )
+            else:
+                fused_moe(
+                    x,
+                    w1,
+                    w2,
+                    input_gating,
+                    topk,
+                    renormalize=True,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    use_int8_w8a16=use_int8_w8a16,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                )
 
     # JIT compilation & warmup
     run()
@@ -411,7 +414,7 @@ class BenchmarkWorker:
         use_deep_gemm: bool = False,
     ) -> tuple[dict[str, int], float]:
         current_platform.seed_everything(self.seed)
-        dtype_str = _get_config_dtype_str(
+        dtype_str = get_config_dtype_str(
             dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
         )
         # NOTE(woosuk): The current naming convention uses w2.shape[2], which
@@ -544,7 +547,7 @@ def save_configs(
     block_quant_shape: list[int],
     save_dir: str,
 ) -> None:
-    dtype_str = _get_config_dtype_str(
+    dtype_str = get_config_dtype_str(
         dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
     )
 
@@ -557,7 +560,7 @@ def save_configs(
     filename = os.path.join(save_dir, filename)
     print(f"Writing best config to {filename}...")
     with open(filename, "w") as f:
-        json.dump({"triton_version": triton.__version__, **configs}, f, indent=4)
+        json.dump(configs, f, indent=4)
         f.write("\n")
 
 
@@ -584,19 +587,14 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
     elif config.architectures[0] in (
-        "DeepseekV2ForCausalLM",
         "DeepseekV3ForCausalLM",
-        "DeepseekV32ForCausalLM",
+        "DeepseekV2ForCausalLM",
         "Glm4MoeForCausalLM",
     ):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
-    elif config.architectures[0] in (
-        "Qwen2MoeForCausalLM",
-        "Qwen3MoeForCausalLM",
-        "Qwen3NextForCausalLM",
-    ):
+    elif config.architectures[0] in ("Qwen2MoeForCausalLM", "Qwen3MoeForCausalLM"):
         E = config.num_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size

benchmarks/kernels/benchmark_polynorm.py

@@ -1,155 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import itertools
-
-import torch
-
-from vllm import _custom_ops as vllm_ops
-from vllm.triton_utils import triton
-
-
-def polynorm_naive(
-    x: torch.Tensor,
-    weight: torch.Tensor,
-    bias: torch.Tensor,
-    eps: float = 1e-6,
-):
-    orig_shape = x.shape
-    x = x.view(-1, x.shape[-1])
-
-    def norm(x, eps: float):
-        return x / torch.sqrt(x.pow(2).mean(-1, keepdim=True) + eps)
-
-    x = x.float()
-    return (
-        (
-            weight[0] * norm(x**3, eps)
-            + weight[1] * norm(x**2, eps)
-            + weight[2] * norm(x, eps)
-            + bias
-        )
-        .to(weight.dtype)
-        .view(orig_shape)
-    )
-
-
-def polynorm_vllm(
-    x: torch.Tensor,
-    weight: torch.Tensor,
-    bias: torch.Tensor,
-    eps: float = 1e-6,
-):
-    orig_shape = x.shape
-    x = x.view(-1, x.shape[-1])
-
-    out = torch.empty_like(x)
-    vllm_ops.poly_norm(out, x, weight, bias, eps)
-    output = out
-
-    output = output.view(orig_shape)
-    return output
-
-
-def calculate_diff(batch_size, seq_len, hidden_dim):
-    dtype = torch.bfloat16
-    x = torch.randn(batch_size, seq_len, hidden_dim, dtype=dtype, device="cuda")
-    weight = torch.ones(3, dtype=dtype, device="cuda")
-    bias = torch.ones(1, dtype=dtype, device="cuda")
-
-    output_naive = polynorm_naive(x, weight, bias)
-    output_vllm = polynorm_vllm(x, weight, bias)
-
-    if torch.allclose(output_naive, output_vllm, atol=1e-2, rtol=1e-2):
-        print("✅ All implementations match")
-    else:
-        print("❌ Implementations differ")
-
-
-batch_size_range = [2**i for i in range(0, 7, 2)]
-seq_length_range = [2**i for i in range(6, 11, 1)]
-dim_range = [2048, 4096]
-configs = list(itertools.product(dim_range, batch_size_range, seq_length_range))
-
-
-def get_benchmark():
-    @triton.testing.perf_report(
-        triton.testing.Benchmark(
-            x_names=["dim", "batch_size", "seq_len"],
-            x_vals=[list(_) for _ in configs],
-            line_arg="provider",
-            line_vals=["naive", "vllm"],
-            line_names=["Naive", "vLLM"],
-            styles=[("blue", "-"), ("red", "-")],
-            ylabel="us",
-            plot_name="polynorm-perf",
-            args={},
-        )
-    )
-    def benchmark(dim, batch_size, seq_len, provider):
-        dtype = torch.bfloat16
-        hidden_dim = dim * 4
-
-        x = torch.randn(batch_size, seq_len, hidden_dim, dtype=dtype, device="cuda")
-        weight = torch.ones(3, dtype=dtype, device="cuda")
-        bias = torch.ones(1, dtype=dtype, device="cuda")
-
-        quantiles = [0.5, 0.2, 0.8]
-
-        if provider == "naive":
-            ms, min_ms, max_ms = triton.testing.do_bench(
-                lambda: polynorm_naive(x, weight, bias),
-                quantiles=quantiles,
-            )
-        else:
-            ms, min_ms, max_ms = triton.testing.do_bench(
-                lambda: polynorm_vllm(x, weight, bias),
-                quantiles=quantiles,
-            )
-
-        return 1000 * ms, 1000 * max_ms, 1000 * min_ms
-
-    return benchmark
-
-
-if __name__ == "__main__":
-    import argparse
-
-    parser = argparse.ArgumentParser()
-    parser.add_argument(
-        "--batch-size",
-        type=int,
-        default=4,
-        help="Batch size",
-    )
-    parser.add_argument(
-        "--seq-len",
-        type=int,
-        default=128,
-        help="Sequence length",
-    )
-    parser.add_argument(
-        "--hidden-dim",
-        type=int,
-        default=8192,
-        help="Intermediate size of MLP",
-    )
-    parser.add_argument(
-        "--save-path",
-        type=str,
-        default="./configs/polnorm/",
-        help="Path to save polnorm benchmark results",
-    )
-
-    args = parser.parse_args()
-
-    # Run correctness test
-    calculate_diff(
-        batch_size=args.batch_size,
-        seq_len=args.seq_len,
-        hidden_dim=args.hidden_dim,
-    )
-
-    benchmark = get_benchmark()
-    # Run performance benchmark
-    benchmark.run(print_data=True, save_path=args.save_path)

benchmarks/kernels/benchmark_reshape_and_cache_flash.py

@@ -9,9 +9,6 @@ import torch
 from tabulate import tabulate
 
 from vllm import _custom_ops as ops
-from vllm.attention.ops.triton_reshape_and_cache_flash import (
-    triton_reshape_and_cache_flash,
-)
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.utils import (
@@ -34,8 +31,6 @@ def run_benchmark(
     kv_cache_dtype: str,
     kv_cache_layout: str,
     num_iters: int,
-    implementation: str,
-    benchmark_mode: str,
     device: str = "cuda",
 ) -> float:
     """Return latency (seconds) for given num_tokens."""
@@ -43,14 +38,6 @@ def run_benchmark(
     if kv_cache_dtype == "fp8" and head_size % 16:
         raise ValueError("fp8 kv-cache requires head_size to be a multiple of 16.")
 
-    if implementation not in ("cuda", "triton"):
-        raise ValueError(
-            f"Unsupported implementation: {implementation}. "
-            "Only 'cuda' and 'triton' are supported."
-        )
-    if implementation == "triton" and kv_cache_layout == "HND":
-        return float("nan")  # Triton does not support HND layout yet.
-
     current_platform.seed_everything(42)
     torch.set_default_device(device)
 
@@ -78,49 +65,27 @@ def run_benchmark(
         cache_layout=kv_cache_layout,
     )
     key_cache, value_cache = key_caches[0], value_caches[0]
-    # to free unused memory
-    del key_caches, value_caches
 
     # compute per-kernel scaling factors for fp8 conversion (if used).
     k_scale = (key.amax() / 64.0).to(torch.float32)
     v_scale = (value.amax() / 64.0).to(torch.float32)
 
-    if implementation == "cuda":
-        function_under_test = lambda: ops.reshape_and_cache_flash(
-            key,  # noqa: F821
-            value,  # noqa: F821
-            key_cache,  # noqa: F821
-            value_cache,  # noqa: F821
-            slot_mapping,  # noqa: F821
-            kv_cache_dtype,
-            k_scale,
-            v_scale,
-        )
-    else:
-        function_under_test = lambda: triton_reshape_and_cache_flash(
-            key,  # noqa: F821
-            value,  # noqa: F821
-            key_cache,  # noqa: F821
-            value_cache,  # noqa: F821
-            slot_mapping,  # noqa: F821
-            kv_cache_dtype,
-            k_scale,
-            v_scale,
-        )
-    if benchmark_mode == "cudagraph":
-        g = torch.cuda.CUDAGraph()
-        with torch.cuda.graph(g):
-            function_under_test()
-        torch.cuda.synchronize()
-        function_under_test = lambda: g.replay()
-
     def run_cuda_benchmark(n_iters: int) -> float:
         nonlocal key, value, key_cache, value_cache, slot_mapping
         torch.cuda.synchronize()
         start = time.perf_counter()
         for _ in range(n_iters):
-            function_under_test()
-            torch.cuda.synchronize()
+            ops.reshape_and_cache_flash(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                slot_mapping,
+                kv_cache_dtype,
+                k_scale,
+                v_scale,
+            )
+        torch.cuda.synchronize()
         end = time.perf_counter()
         return (end - start) / n_iters
 
@@ -151,16 +116,10 @@ def main(args):
                 kv_cache_dtype=args.kv_cache_dtype,
                 kv_cache_layout=layout,
                 num_iters=args.iters,
-                implementation=args.implementation,
-                benchmark_mode=args.mode,
                 device="cuda",
             )
             rows.append([n_tok, layout, f"{lat * 1e6:.3f}"])
 
-    print(
-        f"Benchmark results for implementation {args.implementation}"
-        f" (measuring with {args.mode}):"
-    )
     print(tabulate(rows, headers=["num_tokens", "layout", "latency (µs)"]))
 
 
@@ -192,21 +151,6 @@ if __name__ == "__main__":
     )
 
     parser.add_argument("--iters", type=int, default=100)
-
-    parser.add_argument(
-        "--implementation",
-        type=str,
-        choices=["cuda", "triton"],
-        default="cuda",
-    )
-
-    parser.add_argument(
-        "--mode",
-        type=str,
-        choices=["cudagraph", "no_graph"],
-        default="cudagraph",
-    )
-
     args = parser.parse_args()
 
     main(args)

benchmarks/kernels/benchmark_silu_mul_fp8_quant.py

@@ -1,675 +1,77 @@
+#!/usr/bin/env python3
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-from collections.abc import Callable
+import time
 
-import matplotlib.pyplot as plt
-import numpy as np
 import torch
 
 from vllm.model_executor.layers.fused_moe.batched_deep_gemm_moe import (
-    silu_mul_fp8_quant_deep_gemm_cuda,
+    silu_mul_fp8_quant_deep_gemm,
 )
 from vllm.platforms import current_platform
-from vllm.triton_utils import tl, triton
-from vllm.utils.deep_gemm import is_deep_gemm_e8m0_used
 
 
-@triton.jit
-def _silu_mul_fp8_quant_deep_gemm(
-    # Pointers ------------------------------------------------------------
-    input_ptr,  # 16-bit activations (E, T, 2*H)
-    y_q_ptr,  # fp8 quantized activations (E, T, H)
-    y_s_ptr,  # 16-bit scales (E, T, G)
-    counts_ptr,  # int32 num tokens per expert (E)
-    # Sizes ---------------------------------------------------------------
-    H: tl.constexpr,  # hidden dimension (per output)
-    GROUP_SIZE: tl.constexpr,  # elements per group (usually 128)
-    # Strides for input (elements) ---------------------------------------
-    stride_i_e,
-    stride_i_t,
-    stride_i_h,
-    # Strides for y_q (elements) -----------------------------------------
-    stride_yq_e,
-    stride_yq_t,
-    stride_yq_h,
-    # Strides for y_s (elements) -----------------------------------------
-    stride_ys_e,
-    stride_ys_t,
-    stride_ys_g,
-    # Stride for counts (elements)
-    stride_counts_e,
-    # Numeric params ------------------------------------------------------
-    eps: tl.constexpr,
-    fp8_min: tl.constexpr,
-    fp8_max: tl.constexpr,
-    use_ue8m0: tl.constexpr,
-    # Meta ---------------------------------------------------------------
-    BLOCK: tl.constexpr,
-    NUM_STAGES: tl.constexpr,
-):
-    G = H // GROUP_SIZE
-
-    # map program id -> (e, g)
-    pid = tl.program_id(0)
-    e = pid // G
-    g = pid % G
-
-    e = e.to(tl.int64)
-    g = g.to(tl.int64)
-
-    # number of valid tokens for this expert
-    n_tokens = tl.load(counts_ptr + e * stride_counts_e).to(tl.int64)
-
-    cols = tl.arange(0, BLOCK).to(tl.int64)
-    mask = cols < BLOCK
-
-    base_input_offset = e * stride_i_e + g * GROUP_SIZE * stride_i_h
-    base_gate_offset = base_input_offset + cols * stride_i_h
-    base_up_offset = base_input_offset + H * stride_i_h + cols * stride_i_h
-    base_yq_offset = e * stride_yq_e + g * GROUP_SIZE * stride_yq_h + cols * stride_yq_h
-    base_ys_offset = e * stride_ys_e + g * stride_ys_g
-
-    for t in tl.range(0, n_tokens, num_stages=NUM_STAGES):
-        gate = tl.load(
-            input_ptr + base_gate_offset + t * stride_i_t, mask=mask, other=0.0
-        ).to(tl.float32)
-        up = tl.load(input_ptr + base_up_offset + t * stride_i_t, mask=mask, other=0.0)
-
-        gate = gate * (1.0 / (1.0 + tl.exp(-gate)))
-        y = gate * up
-
-        y_s = tl.maximum(tl.max(tl.abs(y)), eps) / fp8_max
-        if use_ue8m0:
-            y_s = tl.exp2(tl.ceil(tl.log2(y_s)))
-
-        y_q = tl.clamp(y / y_s, fp8_min, fp8_max).to(y_q_ptr.dtype.element_ty)
-
-        tl.store(y_q_ptr + base_yq_offset + t * stride_yq_t, y_q, mask=mask)
-        tl.store(y_s_ptr + base_ys_offset + t * stride_ys_t, y_s)
-
-
-def silu_mul_fp8_quant_deep_gemm_triton(
-    y: torch.Tensor,  # (E, T, 2*H)
-    tokens_per_expert: torch.Tensor,  # (E,) number of valid tokens per expert
-    num_parallel_tokens,
-    group_size: int = 128,
-    eps: float = 1e-10,
-) -> tuple[torch.Tensor, torch.Tensor]:
-    """Quantize silu(y[..., :H]) * y[..., H:] to FP8 with group per-token scales
-
-    y has shape (E, T, 2*H). The first half of the last dimension is
-    silu-activated, multiplied by the second half, then quantized into FP8.
-
-    Returns `(y_q, y_s)` where
-    * `y_q`: FP8 tensor, shape (E, T, H), same layout as y[..., :H]
-    * `y_s`: FP32 tensor, shape (E, T, H // group_size), strides (T*G, 1, T)
-    """
-    assert y.ndim == 3, "y must be (E, T, 2*H)"
-    E, T, H2 = y.shape
-    assert H2 % 2 == 0, "last dim of y must be even (2*H)"
-    H = H2 // 2
-    G = (H + group_size - 1) // group_size
-    assert H % group_size == 0, "H must be divisible by group_size"
-    assert tokens_per_expert.ndim == 1 and tokens_per_expert.shape[0] == E, (
-        "tokens_per_expert must be shape (E,)"
+def benchmark(E, T, H, G=128, runs=50):
+    current_platform.seed_everything(42)
+    y = torch.randn((E, T, 2 * H), dtype=torch.bfloat16, device="cuda")
+    tokens_per_expert = torch.randint(
+        T // 2, T, size=(E,), dtype=torch.int32, device="cuda"
     )
-    tokens_per_expert = tokens_per_expert.to(device=y.device, dtype=torch.int32)
-
-    # allocate outputs
-    fp8_dtype = torch.float8_e4m3fn
-    y_q = torch.empty((E, T, H), dtype=fp8_dtype, device=y.device)
-
-    # strides (elements)
-    stride_i_e, stride_i_t, stride_i_h = y.stride()
-    stride_yq_e, stride_yq_t, stride_yq_h = y_q.stride()
-
-    # desired scale strides (elements): (T*G, 1, T)
-    stride_ys_e = T * G
-    stride_ys_t = 1
-    stride_ys_g = T
-    y_s = torch.empty_strided(
-        (E, T, G),
-        (stride_ys_e, stride_ys_t, stride_ys_g),
-        dtype=torch.float32,
-        device=y.device,
-    )
-
-    stride_cnt_e = tokens_per_expert.stride()[0]
-
-    # Static grid over experts and H-groups.
-    # A loop inside the kernel handles the token dim
-    grid = (E * G,)
-
-    f_info = torch.finfo(fp8_dtype)
-    fp8_max = f_info.max
-    fp8_min = f_info.min
-
-    _silu_mul_fp8_quant_deep_gemm[grid](
-        y,
-        y_q,
-        y_s,
-        tokens_per_expert,
-        H,
-        group_size,
-        stride_i_e,
-        stride_i_t,
-        stride_i_h,
-        stride_yq_e,
-        stride_yq_t,
-        stride_yq_h,
-        stride_ys_e,
-        stride_ys_t,
-        stride_ys_g,
-        stride_cnt_e,
-        eps,
-        fp8_min,
-        fp8_max,
-        is_deep_gemm_e8m0_used(),
-        BLOCK=group_size,
-        NUM_STAGES=4,
-        num_warps=1,
-    )
-
-    return y_q, y_s
-
-
-# Parse generation strategies
-strategies = ["uniform", "max_t", "first_t"]
-
-
-def benchmark(
-    kernel: Callable,
-    E: int,
-    T: int,
-    H: int,
-    total_tokens: int,
-    num_parallel_tokens: int = 64,
-    G: int = 128,
-    runs: int = 200,
-    num_warmups: int = 20,
-    gen_strategy: str = "default",
-    iterations_per_run: int = 20,
-):
-    def generate_data(seed_offset=0):
-        """Generate input data with given seed offset"""
-        current_platform.seed_everything(42 + seed_offset)
-        y = torch.rand((E, T, 2 * H), dtype=torch.bfloat16, device="cuda").contiguous()
-
-        if gen_strategy == "uniform":
-            r = torch.rand(size=(E,), device="cuda")
-            r /= r.sum()
-            r *= total_tokens
-            tokens_per_expert = r.int()
-            tokens_per_expert = torch.minimum(
-                tokens_per_expert,
-                torch.ones((E,), device=r.device, dtype=torch.int) * T,
-            )
-        elif gen_strategy == "max_t":
-            tokens_per_expert = torch.empty(size=(E,), dtype=torch.int32, device="cuda")
-            tokens_per_expert.fill_(total_tokens / E)
-        elif gen_strategy == "first_t":
-            tokens_per_expert = torch.zeros(size=(E,), dtype=torch.int32, device="cuda")
-            tokens_per_expert[0] = min(T, total_tokens)
-        else:
-            raise ValueError(f"Unknown generation strategy: {gen_strategy}")
-        return y, tokens_per_expert
-
-    dataset_count = 4
-    # Pre-generate different input matrices for each iteration to avoid cache effects
-    data_sets = [generate_data(i) for i in range(dataset_count)]
 
     # Warmup
-    y, tokens_per_expert = data_sets[0]
-    for _ in range(num_warmups):
-        kernel(
-            y, tokens_per_expert, num_parallel_tokens=num_parallel_tokens, group_size=G
-        )
-    torch.cuda.synchronize()
-
-    start_event = torch.cuda.Event(enable_timing=True)
-    end_event = torch.cuda.Event(enable_timing=True)
-
-    # Benchmark
-    latencies: list[float] = []
-    for _ in range(runs):
+    for _ in range(10):
+        silu_mul_fp8_quant_deep_gemm(y, tokens_per_expert, group_size=G)
         torch.cuda.synchronize()
 
-        start_event.record()
-        for i in range(iterations_per_run):
-            y, tokens_per_expert = data_sets[i % dataset_count]
-            kernel(
-                y,
-                tokens_per_expert,
-                num_parallel_tokens=num_parallel_tokens,
-                group_size=G,
-            )
-        end_event.record()
-        end_event.synchronize()
+    # Benchmark
+    torch.cuda.synchronize()
+    start = time.perf_counter()
+    for _ in range(runs):
+        silu_mul_fp8_quant_deep_gemm(y, tokens_per_expert, group_size=G)
+    torch.cuda.synchronize()
 
-        total_time_ms = start_event.elapsed_time(end_event)
-        per_iter_time_ms = total_time_ms / iterations_per_run
-        latencies.append(per_iter_time_ms)
+    avg_time = (time.perf_counter() - start) / runs * 1000
 
-    # Use median instead of average for better outlier handling
-    median_time_ms = np.median(latencies)
-    median_time_s = median_time_ms / 1000
-
-    # Calculate actual work done (using first dataset for consistency)
-    _, tokens_per_expert = data_sets[0]
+    # Calculate actual work done (only count valid tokens)
     actual_tokens = tokens_per_expert.sum().item()
     actual_elements = actual_tokens * H
 
     # GFLOPS: operations per element = exp + 3 muls + 1 div + quantization ops ≈ 8 ops
     ops_per_element = 8
     total_ops = actual_elements * ops_per_element
-    gflops = total_ops / median_time_s / 1e9
+    gflops = total_ops / (avg_time / 1000) / 1e9
 
     # Memory bandwidth: bfloat16 inputs (2 bytes), fp8 output (1 byte), scales (4 bytes)
     input_bytes = actual_tokens * 2 * H * 2  # 2*H bfloat16 inputs
     output_bytes = actual_tokens * H * 1  # H fp8 outputs
     scale_bytes = actual_tokens * (H // G) * 4  # scales in float32
     total_bytes = input_bytes + output_bytes + scale_bytes
-    memory_bw = total_bytes / median_time_s / 1e9
+    memory_bw = total_bytes / (avg_time / 1000) / 1e9
 
-    HOPPER_BANDWIDTH_TBPS = 3.35
-    return (
-        median_time_ms,
-        gflops,
-        memory_bw,
-        (memory_bw / (HOPPER_BANDWIDTH_TBPS * 1024)) * 100,
-    )
+    return avg_time, gflops, memory_bw
 
 
-def create_comparison_plot(
-    ratio, cuda_times, baseline_times, config_labels, strategy_name, id
-):
-    """Create a comparison plot for a specific generation strategy"""
-    fig, ax = plt.subplots(1, 1, figsize=(16, 6))
-
-    # Configure x-axis positions
-    x = np.arange(len(config_labels))
-    width = 0.35
-
-    # Execution Time plot (lower is better)
-    ax.bar(
-        x - width / 2, cuda_times, width, label="CUDA Kernel", alpha=0.8, color="blue"
-    )
-    ax.bar(
-        x + width / 2,
-        baseline_times,
-        width,
-        label="Baseline",
-        alpha=0.8,
-        color="orange",
-    )
-
-    # Add speedup labels over each bar pair
-    for i in range(len(x)):
-        speedup = ratio[i]
-        max_height = max(cuda_times[i], baseline_times[i])
-        ax.text(
-            x[i],
-            max_height + max_height * 0.02,
-            f"{speedup:.2f}x",
-            ha="center",
-            va="bottom",
-            fontweight="bold",
-            fontsize=9,
-        )
-
-    ax.set_xlabel("Configuration")
-    ax.set_ylabel("% Utilization")
-    ax.set_title(
-        f"Memory Bandwidth Utilization (%) - {strategy_name}\n(Higher is Better)"
-    )
-    ax.set_xticks(x)
-    ax.set_xticklabels(config_labels, rotation=45, ha="right")
-    ax.legend()
-    ax.grid(True, alpha=0.3)
-
-    plt.tight_layout()
-    return fig, ax
-
-
-def create_combined_plot(all_results):
-    """Create a combined plot with all strategies in one PNG"""
-    num_strategies = len(all_results)
-    fig, axes = plt.subplots(num_strategies, 1, figsize=(20, 6 * num_strategies))
-
-    if num_strategies == 1:
-        axes = [axes]
-
-    for idx, (
-        strategy_name,
-        ratio,
-        cuda_times,
-        baseline_times,
-        config_labels,
-    ) in enumerate(all_results):
-        ax = axes[idx]
-
-        # Configure x-axis positions
-        x = np.arange(len(config_labels))
-        width = 0.35
-
-        # Execution Time plot (lower is better)
-        ax.bar(
-            x - width / 2,
-            cuda_times,
-            width,
-            label="CUDA Kernel",
-            alpha=0.8,
-            color="blue",
-        )
-        ax.bar(
-            x + width / 2,
-            baseline_times,
-            width,
-            label="Baseline",
-            alpha=0.8,
-            color="orange",
-        )
-
-        # Add speedup labels over each bar pair
-        for i in range(len(x)):
-            speedup = ratio[i]
-            max_height = max(cuda_times[i], baseline_times[i])
-            ax.text(
-                x[i],
-                max_height + max_height * 0.02,
-                f"{speedup:.2f}x",
-                ha="center",
-                va="bottom",
-                fontweight="bold",
-                fontsize=9,
-            )
-
-        ax.set_xlabel("Configuration")
-        ax.set_ylabel("% Utilization")
-        ax.set_title(
-            f"Memory Bandwidth Utilization (%) - {strategy_name}\n(Higher is Better)"
-        )
-        ax.set_xticks(x)
-        ax.set_xticklabels(config_labels, rotation=45, ha="right")
-        ax.legend()
-        ax.grid(True, alpha=0.3)
-
-    plt.tight_layout()
-    filename = "../../silu_bench/silu_benchmark_combined.png"
-    plt.savefig(filename, dpi=300, bbox_inches="tight")
-    plt.show()
-
-    return filename
-
-
-outer_dim = 7168
 configs = [
+    (8, 32, 1024),
+    (16, 64, 2048),
+    (32, 128, 4096),
     # DeepSeekV3 Configs
-    (8, 1024, 7168),
-    # DeepSeekV3 Configs
-    (32, 1024, 7168),
-    # DeepSeekV3 Configs
+    (256, 16, 7168),
+    (256, 32, 7168),
+    (256, 64, 7168),
+    (256, 128, 7168),
+    (256, 256, 7168),
+    (256, 512, 7168),
     (256, 1024, 7168),
 ]
 
-runs = 100
-num_warmups = 20
-
-strategy_descriptions = {
-    "uniform": "Uniform Random",
-    "max_t": "Even Assignment",
-    "first_t": "experts[0] = T, experts[1:] = 0",
-}
-
 print(f"GPU: {torch.cuda.get_device_name()}")
-print(f"Testing strategies: {', '.join(strategies)}")
-print(f"Configurations: {len(configs)} configs")
+print(f"{'Config':<20} {'Time(ms)':<10} {'GFLOPS':<10} {'GB/s':<10}")
+print("-" * 50)
 
-all_results = []
-
-# Run benchmarks for each strategy
-for id, strategy in enumerate(strategies):
-    print(f"\n{'=' * 60}")
-    print(f"Testing strategy: {strategy_descriptions[strategy]}")
-    print(f"{'=' * 60}")
-
-    # Collect benchmark data for both algorithms
-    config_labels = []
-    config_x_axis = []
-    all_cuda_results = []
-    all_baseline_results = []
-    all_ratios = []
-
-    for E, T, H in configs:
-        total_tokens_config = [8 * E, 16 * E, 32 * E, 64 * E, 128 * E, 256 * E]
-        config_x_axis.append(total_tokens_config)
-
-        cuda_results = []
-        baseline_results = []
-        ratios = []
-
-        for total_tokens in total_tokens_config:
-            config_label = f"E={E},T={T},H={H},TT={total_tokens}"
-            config_labels.append(config_label)
-
-            # CUDA kernel results
-            time_ms_cuda, gflops, gbps, perc = benchmark(
-                silu_mul_fp8_quant_deep_gemm_cuda,
-                E,
-                T,
-                H,
-                total_tokens,
-                runs=runs,
-                num_warmups=num_warmups,
-                gen_strategy=strategy,
-            )
-            cuda_results.append((time_ms_cuda, gflops, gbps, perc))
-
-            # Baseline results
-            time_ms_triton, gflops, gbps, perc = benchmark(
-                silu_mul_fp8_quant_deep_gemm_triton,
-                E,
-                T,
-                H,
-                total_tokens,
-                runs=runs,
-                num_warmups=num_warmups,
-                gen_strategy=strategy,
-            )
-            baseline_results.append((time_ms_triton, gflops, gbps, perc))
-            ratios.append(time_ms_triton / time_ms_cuda)
-
-            print(f"Completed: {config_label}")
-        all_cuda_results.append(cuda_results)
-        all_baseline_results.append(baseline_results)
-        all_ratios.append(ratios)
-
-    # Store results for combined plotting
-    all_results.append(
-        (
-            strategy_descriptions[strategy],
-            all_ratios,
-            all_cuda_results,
-            all_baseline_results,
-            config_labels,
-            config_x_axis,
-        )
-    )
-
-    # Print summary table for this strategy
-    print(f"\nSummary Table - {strategy_descriptions[strategy]}:")
-    print(f"{'Config':<20} {'CUDA Time(ms)':<12} {'Base Time(ms)':<12} {'Speedup':<8}")
-    print("-" * 60)
-
-    for i, (E, T, H) in enumerate(configs):
-        speedup = baseline_results[i][0] / cuda_results[i][0]
-        config_label = f"E={E:3d},T={T:4d},H={H:4d}"
-        print(
-            f"{config_label:<20} {cuda_results[i][0]:8.5f} "
-            f"{baseline_results[i][0]:8.5f} {speedup:6.2f}x"
-        )
-
-
-def create_total_tokens_plot(all_results):
-    num_strategies = len(all_results)
-    num_configs = len(configs)
-
-    # Create side-by-side subplots: 2 columns for speedup and bandwidth percentage
-    fig, axs = plt.subplots(
-        num_strategies, num_configs * 2, figsize=(28, 6 * num_strategies)
-    )
-
-    # Add main title to the entire figure
-    fig.suptitle(
-        "Performance Analysis: Speedup vs Bandwidth Utilization (Triton & CUDA)",
-        fontsize=16,
-        fontweight="bold",
-        y=0.98,
-    )
-
-    # Handle single strategy case
-    if num_strategies == 1:
-        axs = axs.reshape(1, -1)
-
-    # Handle single config case
-    if num_configs == 1:
-        axs = axs.reshape(-1, 2)
-
-    for strategy_idx, result in enumerate(all_results):
-        (
-            strategy_name,
-            all_ratios,
-            all_cuda_results,
-            all_baseline_results,
-            config_labels,
-            config_x_axis,
-        ) = result
-
-        for config_idx in range(num_configs):
-            # Speedup plot (left column)
-            ax_speedup = axs[strategy_idx, config_idx * 2]
-            # Bandwidth plot (right column)
-            ax_bandwidth = axs[strategy_idx, config_idx * 2 + 1]
-
-            E, T, H = configs[config_idx]
-            ratios = all_ratios[config_idx]
-            total_tokens_values = config_x_axis[config_idx]
-
-            # Extract CUDA and Triton bandwidth percentages
-            cuda_bandwidth_percentages = [
-                result[3] for result in all_cuda_results[config_idx]
-            ]
-            triton_bandwidth_percentages = [
-                result[3] for result in all_baseline_results[config_idx]
-            ]
-
-            # Plot speedup ratios vs total tokens (left plot)
-            ax_speedup.plot(
-                total_tokens_values, ratios, "bo-", linewidth=3, markersize=8
-            )
-            ax_speedup.set_title(
-                f"{strategy_name}\nSpeedup (CUDA/Triton)\nE={E}, T={T}, H={H}",
-                fontsize=12,
-                fontweight="bold",
-            )
-            ax_speedup.set_xlabel("Total Tokens", fontweight="bold", fontsize=11)
-            ax_speedup.set_ylabel("Speedup Ratio", fontweight="bold", fontsize=11)
-            ax_speedup.grid(True, alpha=0.3)
-
-            ax_bandwidth.plot(
-                total_tokens_values,
-                cuda_bandwidth_percentages,
-                "ro-",
-                linewidth=3,
-                markersize=8,
-                label="CUDA",
-            )
-            ax_bandwidth.plot(
-                total_tokens_values,
-                triton_bandwidth_percentages,
-                "go-",
-                linewidth=3,
-                markersize=8,
-                label="Triton",
-            )
-            ax_bandwidth.set_title(
-                f"{strategy_name}\nBandwidth Utilization (Hopper)\nE={E}, T={T}, H={H}",
-                fontsize=12,
-                fontweight="bold",
-            )
-            ax_bandwidth.set_xlabel("Total Tokens", fontweight="bold", fontsize=11)
-            ax_bandwidth.set_ylabel(
-                "% of Peak Bandwidth", fontweight="bold", fontsize=11
-            )
-            ax_bandwidth.legend(prop={"weight": "bold"})
-            ax_bandwidth.grid(True, alpha=0.3)
-
-            # Format x-axis labels for both plots
-            for ax in [ax_speedup, ax_bandwidth]:
-                ax.set_xticks(total_tokens_values)
-                ax.set_xticklabels(
-                    [
-                        f"{tt // 1000}K" if tt >= 1000 else str(tt)
-                        for tt in total_tokens_values
-                    ],
-                    fontweight="bold",
-                )
-                # Make tick labels bold
-                for label in ax.get_xticklabels() + ax.get_yticklabels():
-                    label.set_fontweight("bold")
-
-            # Add value labels on speedup points
-            for x, y in zip(total_tokens_values, ratios):
-                ax_speedup.annotate(
-                    f"{y:.2f}x",
-                    (x, y),
-                    textcoords="offset points",
-                    xytext=(0, 12),
-                    ha="center",
-                    fontsize=10,
-                    fontweight="bold",
-                    bbox=dict(boxstyle="round,pad=0.3", facecolor="white", alpha=0.7),
-                )
-
-            # Add value labels on CUDA bandwidth points
-            for x, y in zip(total_tokens_values, cuda_bandwidth_percentages):
-                ax_bandwidth.annotate(
-                    f"{y:.1f}%",
-                    (x, y),
-                    textcoords="offset points",
-                    xytext=(0, 12),
-                    ha="center",
-                    fontsize=9,
-                    fontweight="bold",
-                    bbox=dict(boxstyle="round,pad=0.2", facecolor="red", alpha=0.3),
-                )
-
-            # Add value labels on Triton bandwidth points
-            for x, y in zip(total_tokens_values, triton_bandwidth_percentages):
-                ax_bandwidth.annotate(
-                    f"{y:.1f}%",
-                    (x, y),
-                    textcoords="offset points",
-                    xytext=(0, -15),
-                    ha="center",
-                    fontsize=9,
-                    fontweight="bold",
-                    bbox=dict(boxstyle="round,pad=0.2", facecolor="green", alpha=0.3),
-                )
-
-    plt.tight_layout()
-    plt.subplots_adjust(top=0.93)  # Make room for main title
-    filename = "silu_benchmark_total_tokens.png"
-    plt.savefig(filename, dpi=300, bbox_inches="tight")
-    plt.show()
-
-    return filename
-
-
-# Create combined plot with all strategies
-combined_plot_filename = create_total_tokens_plot(all_results)
-
-print(f"\n{'=' * 60}")
-print("Benchmark Complete!")
-print(f"Generated combined plot: {combined_plot_filename}")
-print(f"{'=' * 60}")
+for E, T, H in configs:
+    try:
+        time_ms, gflops, gbps = benchmark(E, T, H)
+        print(f"E={E:3d},T={T:4d},H={H:4d} {time_ms:8.3f} {gflops:8.1f} {gbps:8.1f}")
+    except Exception:
+        print(f"E={E:3d},T={T:4d},H={H:4d} FAILED")

benchmarks/kernels/benchmark_trtllm_decode_attention.py

@@ -259,7 +259,6 @@ if __name__ == "__main__":
         # (q_quant_dtype, kv_quant_dtype, o_quant_dtype)
         (None, None, None),
         (None, FP8_DTYPE, None),
-        (FP8_DTYPE, FP8_DTYPE, None),
         (FP8_DTYPE, FP8_DTYPE, FP8_DTYPE),
         (FP8_DTYPE, FP8_DTYPE, FP4_DTYPE),
     ]

benchmarks/kernels/benchmark_trtllm_prefill_attention.py

@@ -274,7 +274,6 @@ if __name__ == "__main__":
     quant_dtypes = [
         # (q_quant_dtype, kv_quant_dtype, o_quant_dtype)
         (None, None, None),
-        (FP8_DTYPE, FP8_DTYPE, None),
         (FP8_DTYPE, FP8_DTYPE, FP8_DTYPE),
         (FP8_DTYPE, FP8_DTYPE, FP4_DTYPE),
     ]

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -11,13 +11,13 @@ from datetime import datetime
 from typing import Any
 
 import torch
+import triton
 from tqdm import tqdm
 
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     _w8a8_block_fp8_matmul,
 )
 from vllm.platforms import current_platform
-from vllm.triton_utils import triton
 from vllm.utils import FlexibleArgumentParser
 
 mp.set_start_method("spawn", force=True)
@@ -56,7 +56,7 @@ def w8a8_block_matmul(
         Bs: The per-block quantization scale for `B`.
         block_size: The block size for per-block quantization.
                     It should be 2-dim, e.g., [128, 128].
-        output_dtype: The dtype of the returned tensor.
+        output_dytpe: The dtype of the returned tensor.
 
     Returns:
         torch.Tensor: The result of matmul.

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -8,16 +8,12 @@ import torch
 
 from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    get_col_major_tma_aligned_tensor,
     per_token_group_quant_fp8,
     w8a8_block_fp8_matmul,
 )
 from vllm.triton_utils import triton
-from vllm.utils.deep_gemm import (
-    calc_diff,
-    fp8_gemm_nt,
-    get_col_major_tma_aligned_tensor,
-    per_block_cast_to_fp8,
-)
+from vllm.utils.deep_gemm import calc_diff, fp8_gemm_nt, per_block_cast_to_fp8
 
 
 def benchmark_shape(m: int,

benchmarks/multi_turn/README.md

@@ -55,107 +55,6 @@ output_num_chunks  166.0    99.01   11.80    79.00    90.00    98.00   108.75
 ----------------------------------------------------------------------------------------------------
 ```
 
-### JSON configuration file for synthetic conversations generation
-
-The input flag `--input-file` is used to determine the input conversations for the benchmark.<br/>
-When the input is a JSON file with the field `"filetype": "generate_conversations"` the tool will generate synthetic multi-turn (questions and answers) conversations.
-
-The file `generate_multi_turn.json` is an example file.
-
-The file must contain the sections `prompt_input` and `prompt_output`.
-
-The `prompt_input` section must contain `num_turns`, `prefix_num_tokens` and `num_tokens`:
-
-* `num_turns` - Number of total turns in the conversation (both user & assistant).<br/>
-The final value will always be rounded to an even number so each user turn has a reply.
-* `prefix_num_tokens` - Tokens added at the start of only the **first user turn** in a conversation (unique per conversation).
-* `num_tokens` - Total token length of each **user** message (one turn).
-
-The `prompt_output` section must contain `num_tokens`:
-
-* `num_tokens` - Total token length of each **assistant** message (one turn).
-
-### Random distributions for synthetic conversations generation
-
-When creating an input JSON file (such as `generate_multi_turn.json`),<br/>
-every numeric field (such as `num_turns` or `num_tokens`) requires a distribution.<br/>
-The distribution determines how to randomly sample values for the field.
-
-The available distributions are listed below.
-
-**Note:** The optional `max` field (for lognormal, zipf, and poisson) can be used to cap sampled values at an upper bound.</br>
-Can be used to make sure that the total number of tokens in every request does not exceed `--max-model-len`.
-
-#### constant
-
-```json
-{
-    "distribution": "constant",
-    "value": 500
-}
-```
-
-* `value` - the fixed integer value (always returns the same number).
-
-#### uniform
-
-```json
-{
-    "distribution": "uniform",
-    "min": 12,
-    "max": 18
-}
-```
-
-* `min` - minimum value (inclusive).
-* `max` - maximum value (inclusive), should be equal or larger than min.
-
-#### lognormal
-
-```json
-{
-    "distribution": "lognormal",
-    "average": 1000,
-    "max": 5000
-}
-```
-
-You can parameterize the lognormal distribution in one of two ways:
-
-Using the average and optional median ratio:
-
-* `average` - target average value of the distribution.
-* `median_ratio` - the ratio of the median to the average; controls the skewness. Must be in the range (0, 1).
-
-Using the parameters of the underlying normal distribution:
-
-* `mean` - mean of the underlying normal distribution.
-* `sigma` - standard deviation of the underlying normal distribution.
-
-#### zipf
-
-```json
-{
-    "distribution": "zipf",
-    "alpha": 1.2,
-    "max": 100
-}
-```
-
-* `alpha` - skew parameter (> 1). Larger values produce stronger skew toward smaller integers.
-
-#### poisson
-
-```json
-{
-    "distribution": "poisson",
-    "alpha": 10,
-    "max": 50
-}
-```
-
-* `alpha` - expected value (λ). Also the variance of the distribution.
-
 ## ShareGPT Conversations
 
 To run with the ShareGPT data, download the following ShareGPT dataset:

benchmarks/multi_turn/bench_dataset.py

@@ -99,105 +99,21 @@ class PoissonDistribution(Distribution):
 
 class LognormalDistribution(Distribution):
     def __init__(
-        self,
-        mean: Optional[float] = None,
-        sigma: Optional[float] = None,
-        average: Optional[int] = None,
-        median_ratio: Optional[float] = None,
-        max_val: Optional[int] = None,
+        self, mean: float, sigma: float, max_val: Optional[int] = None
     ) -> None:
-        self.average = average
-        self.median_ratio = median_ratio
-        self.max_val = max_val
-
-        if average is not None:
-            if average < 1:
-                raise ValueError("Lognormal average must be positive")
-
-            if mean or sigma:
-                raise ValueError(
-                    "When using lognormal average, you can't provide mean/sigma"
-                )
-
-            if self.median_ratio is None:
-                # Default value that provides relatively wide range of values
-                self.median_ratio = 0.85
-
-            # Calculate mean/sigma of np.random.lognormal based on the average
-            mean, sigma = self._generate_lognormal_by_median(
-                target_average=self.average, median_ratio=self.median_ratio
-            )
-        else:
-            if mean is None or sigma is None:
-                raise ValueError(
-                    "Must provide both mean and sigma if average is not used"
-                )
-
-            if mean <= 0 or sigma < 0:
-                raise ValueError(
-                    "Lognormal mean must be positive and sigma must be non-negative"
-                )
-
-        # Mean and standard deviation of the underlying normal distribution
-        # Based on numpy.random.lognormal
         self.mean = mean
         self.sigma = sigma
-
-    @staticmethod
-    def _generate_lognormal_by_median(
-        target_average: int, median_ratio: float
-    ) -> tuple[float, float]:
-        """
-        Compute (mu, sigma) for a lognormal distribution given:
-        - a target average (mean of the distribution)
-        - a ratio of median / mean (controls skewness), assume mean > median
-
-        Background:
-        If Z ~ Normal(mu, sigma^2), then X = exp(Z) ~ LogNormal(mu, sigma).
-        * mean(X)   = exp(mu + sigma^2 / 2)
-        * median(X) = exp(mu)
-
-        So:
-        median / mean = exp(mu) / exp(mu + sigma^2 / 2)
-                      = exp(-sigma^2 / 2)
-
-        Rearranging:
-        sigma^2 = 2 * ln(mean / median)
-        mu      = ln(median)
-
-        This gives a unique (mu, sigma) for any valid mean and median.
-        """
-        # Check input validity: median must be smaller than mean
-        if median_ratio <= 0 or median_ratio >= 1:
-            raise ValueError("median_ratio must be in range (0, 1)")
-
-        target_median = target_average * median_ratio
-
-        # Solve sigma^2 = 2 * ln(mean / median)
-        sigma = np.sqrt(2 * np.log(target_average / target_median))
-        mu = np.log(target_median)
-
-        return mu, sigma
+        self.max_val = max_val
 
     def sample(self, size: int = 1) -> np.ndarray:
         samples = np.random.lognormal(mean=self.mean, sigma=self.sigma, size=size)
-
-        if self.average is not None:
-            # Scale to average
-            samples *= self.average / samples.mean()
-
         if self.max_val:
             samples = np.minimum(samples, self.max_val)
 
         return np.round(samples).astype(int)
 
     def __repr__(self) -> str:
-        if self.average:
-            return (
-                f"LognormalDistribution[{self.average}, "
-                f"{self.median_ratio}, {self.max_val}]"
-            )
-        return f"LognormalDistribution[{self.mean}, {self.sigma}, {self.max_val}]"
+        return f"LognormalDistribution[{self.mean}, {self.sigma}]"
 
 
 class GenConvArgs(NamedTuple):
@@ -257,21 +173,10 @@ def get_random_distribution(
         return PoissonDistribution(conf["alpha"], max_val=max_val)
 
     elif distribution == "lognormal":
-        max_val = conf.get("max", None)
-
-        if "average" in conf:
-            # Infer lognormal mean/sigma (numpy) from input average
-            median_ratio = conf.get("median_ratio", None)
-            return LognormalDistribution(
-                average=conf["average"], median_ratio=median_ratio, max_val=max_val
-            )
-
-        # Use mean/sigma directly (for full control over the distribution)
         verify_field_exists(conf, "mean", section, subsection)
         verify_field_exists(conf, "sigma", section, subsection)
-        return LognormalDistribution(
-            mean=conf["mean"], sigma=conf["sigma"], max_val=max_val
-        )
+        max_val = conf.get("max", None)
+        return LognormalDistribution(conf["mean"], conf["sigma"], max_val=max_val)
 
     elif distribution == "uniform":
         verify_field_exists(conf, "min", section, subsection)

benchmarks/multi_turn/generate_multi_turn.json

@@ -15,8 +15,9 @@
         },
         "prefix_num_tokens": {
             "distribution": "lognormal",
-            "average": 1000,
-            "max": 5000
+            "mean": 6,
+            "sigma": 4,
+            "max": 1500
         },
         "num_tokens": {
             "distribution": "uniform",

cmake/cpu_extension.cmake

@@ -101,7 +101,6 @@ else()
     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(${CPUINFO} "S390" S390_FOUND)
-    find_isa(${CPUINFO} "v" RVV_FOUND) # Check for RISC-V RVV support
 endif()
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
@@ -178,14 +177,8 @@ elseif (S390_FOUND)
         "-mzvector"
         "-march=native"
         "-mtune=native")
-elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
-    if(RVV_FOUND)
-	    message(FAIL_ERROR "Can't support rvv now.")
-    else()
-        list(APPEND CXX_COMPILE_FLAGS "-march=rv64gc")
-    endif()
 else()
-    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA, S390X ISA, ARMv8 or RISC-V support.")
+    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA, S390X ISA or ARMv8 support.")
 endif()
 
 #
@@ -265,8 +258,7 @@ set(VLLM_EXT_SRC
     "csrc/cpu/layernorm.cpp"
     "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
-    "csrc/cpu/torch_bindings.cpp"
-    "csrc/moe/dynamic_4bit_int_moe_cpu.cpp")
+    "csrc/cpu/torch_bindings.cpp")
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     set(VLLM_EXT_SRC

cmake/external_projects/flashmla.cmake

@@ -18,8 +18,8 @@ if(FLASH_MLA_SRC_DIR)
 else()
   FetchContent_Declare(
         flashmla
-        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA
-        GIT_TAG 5f65b85703c7ed75fda01e06495077caad207c3f
+        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA.git
+        GIT_TAG a757314c04eedd166e329e846c820eb1bdd702de
         GIT_PROGRESS TRUE
         CONFIGURE_COMMAND ""
         BUILD_COMMAND ""
@@ -33,64 +33,23 @@ message(STATUS "FlashMLA is available at ${flashmla_SOURCE_DIR}")
 # The FlashMLA kernels only work on hopper and require CUDA 12.3 or later.
 # Only build FlashMLA kernels if we are building for something compatible with 
 # sm90a
-
-set(SUPPORT_ARCHS)
-if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3)
-    list(APPEND SUPPORT_ARCHS 9.0a)
-endif()
-if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8)
-    list(APPEND SUPPORT_ARCHS 10.0a)
-endif()
-
-
-cuda_archs_loose_intersection(FLASH_MLA_ARCHS "${SUPPORT_ARCHS}" "${CUDA_ARCHS}")
-if(FLASH_MLA_ARCHS)
-    set(VLLM_FLASHMLA_GPU_FLAGS ${VLLM_GPU_FLAGS})
-    list(APPEND VLLM_FLASHMLA_GPU_FLAGS "--expt-relaxed-constexpr" "--expt-extended-lambda" "--use_fast_math")
-
+cuda_archs_loose_intersection(FLASH_MLA_ARCHS "9.0a" "${CUDA_ARCHS}")
+if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3 AND FLASH_MLA_ARCHS)
     set(FlashMLA_SOURCES
-        ${flashmla_SOURCE_DIR}/csrc/torch_api.cpp
-        ${flashmla_SOURCE_DIR}/csrc/pybind.cpp
-        ${flashmla_SOURCE_DIR}/csrc/smxx/get_mla_metadata.cu
-        ${flashmla_SOURCE_DIR}/csrc/smxx/mla_combine.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm90/decode/dense/splitkv_mla.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm90/decode/sparse_fp8/splitkv_mla.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm90/prefill/sparse/fwd.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm100/decode/sparse_fp8/splitkv_mla.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm100/prefill/dense/fmha_cutlass_fwd_sm100.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm100/prefill/dense/fmha_cutlass_bwd_sm100.cu
-        ${flashmla_SOURCE_DIR}/csrc/sm100/prefill/sparse/fwd.cu
-    )
-
-    set(FlashMLA_Extension_SOURCES
-        ${flashmla_SOURCE_DIR}/csrc/extension/torch_api.cpp
-        ${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/pybind.cpp
-        ${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/flash_fwd_mla_fp8_sm90.cu
-    )
+        ${flashmla_SOURCE_DIR}/csrc/flash_api.cpp
+        ${flashmla_SOURCE_DIR}/csrc/kernels/get_mla_metadata.cu
+        ${flashmla_SOURCE_DIR}/csrc/kernels/mla_combine.cu
+        ${flashmla_SOURCE_DIR}/csrc/kernels/splitkv_mla.cu
+        ${flashmla_SOURCE_DIR}/csrc/kernels_fp8/flash_fwd_mla_fp8_sm90.cu)
 
     set(FlashMLA_INCLUDES
-        ${flashmla_SOURCE_DIR}/csrc
-        ${flashmla_SOURCE_DIR}/csrc/sm90
         ${flashmla_SOURCE_DIR}/csrc/cutlass/include
-        ${flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
-    )
-
-    set(FlashMLA_Extension_INCLUDES
-        ${flashmla_SOURCE_DIR}/csrc
-        ${flashmla_SOURCE_DIR}/csrc/sm90
-        ${flashmla_SOURCE_DIR}/csrc/extension/sm90/dense_fp8/
-        ${flashmla_SOURCE_DIR}/csrc/cutlass/include
-        ${flashmla_SOURCE_DIR}/csrc/cutlass/tools/util/include
-    )
+        ${flashmla_SOURCE_DIR}/csrc)
 
     set_gencode_flags_for_srcs(
         SRCS "${FlashMLA_SOURCES}"
         CUDA_ARCHS "${FLASH_MLA_ARCHS}")
 
-    set_gencode_flags_for_srcs(
-        SRCS "${FlashMLA_Extension_SOURCES}"
-        CUDA_ARCHS "${FLASH_MLA_ARCHS}")
-
     define_gpu_extension_target(
         _flashmla_C
         DESTINATION vllm
@@ -101,32 +60,8 @@ if(FLASH_MLA_ARCHS)
         INCLUDE_DIRECTORIES ${FlashMLA_INCLUDES}
         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_C PRIVATE
-        $<$<COMPILE_LANGUAGE:CUDA>:-UPy_LIMITED_API>
-        $<$<COMPILE_LANGUAGE:CXX>:-UPy_LIMITED_API>)
-
-    define_gpu_extension_target(
-        _flashmla_extension_C
-        DESTINATION vllm
-        LANGUAGE ${VLLM_GPU_LANG}
-        SOURCES ${FlashMLA_Extension_SOURCES}
-        COMPILE_FLAGS ${VLLM_FLASHMLA_GPU_FLAGS}
-        ARCHITECTURES ${VLLM_GPU_ARCHES}
-        INCLUDE_DIRECTORIES ${FlashMLA_Extension_INCLUDES}
-        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>)
 else()
-    # Create empty targets for setup.py when not targeting sm90a systems
+    # Create an empty target for setup.py when not targeting sm90a systems
     add_custom_target(_flashmla_C)
-    add_custom_target(_flashmla_extension_C)
 endif()
 

cmake/utils.cmake

@@ -480,6 +480,7 @@ function (define_gpu_extension_target GPU_MOD_NAME)
       ${GPU_LANGUAGE}_ARCHITECTURES "${GPU_ARCHITECTURES}")
   endif()
 
+  set_property(TARGET ${GPU_MOD_NAME} PROPERTY CXX_STANDARD 17)
 
   target_compile_options(${GPU_MOD_NAME} PRIVATE
     $<$<COMPILE_LANGUAGE:${GPU_LANGUAGE}>:${GPU_COMPILE_FLAGS}>)

csrc/attention/mla/cutlass_mla_entry.cu

@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <torch/all.h>
+
+#if defined ENABLE_CUTLASS_MLA && ENABLE_CUTLASS_MLA
+void cutlass_mla_decode_sm100a(torch::Tensor const& out,
+                               torch::Tensor const& q_nope,
+                               torch::Tensor const& q_pe,
+                               torch::Tensor const& kv_c_and_k_pe_cache,
+                               torch::Tensor const& seq_lens,
+                               torch::Tensor const& page_table, double scale);
+#endif
+
+void cutlass_mla_decode(torch::Tensor const& out, torch::Tensor const& q_nope,
+                        torch::Tensor const& q_pe,
+                        torch::Tensor const& kv_c_and_k_pe_cache,
+                        torch::Tensor const& seq_lens,
+                        torch::Tensor const& page_table, double scale) {
+#if defined ENABLE_CUTLASS_MLA && ENABLE_CUTLASS_MLA
+  return cutlass_mla_decode_sm100a(out, q_nope, q_pe, kv_c_and_k_pe_cache,
+                                   seq_lens, page_table, scale);
+#endif
+  TORCH_CHECK_NOT_IMPLEMENTED(false, "No compiled cutlass MLA");
+}

csrc/attention/mla/cutlass_mla_kernels.cu

@@ -0,0 +1,225 @@
+/*
+ * Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <torch/all.h>
+
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include "cute/tensor.hpp"
+
+#include "cutlass/cutlass.h"
+#include "cutlass/kernel_hardware_info.h"
+
+#include "cutlass_extensions/common.hpp"
+
+#include "device/sm100_mla.hpp"
+#include "kernel/sm100_mla_tile_scheduler.hpp"
+
+using namespace cute;
+using namespace cutlass::fmha::kernel;
+
+template <typename T, bool PersistenceOption = true>
+struct MlaSm100 {
+  using Element = T;
+  using ElementAcc = float;
+  using ElementOut = T;
+
+  using TileShape = Shape<_128, _128, Shape<_512, _64>>;
+  using TileShapeH = cute::tuple_element_t<0, TileShape>;
+  using TileShapeD = cute::tuple_element_t<2, TileShape>;
+
+  // H K (D_latent D_rope) B
+  using ProblemShape = cute::tuple<TileShapeH, int, TileShapeD, int>;
+
+  using StrideQ = cute::tuple<int64_t, _1, int64_t>;  // H D B
+  using StrideK = cute::tuple<int64_t, _1, int64_t>;  // K D B
+  using StrideO = StrideK;                            // H D B
+  using StrideLSE = cute::tuple<_1, int>;             // H B
+
+  using TileScheduler =
+      std::conditional_t<PersistenceOption, Sm100MlaPersistentTileScheduler,
+                         Sm100MlaIndividualTileScheduler>;
+
+  using FmhaKernel =
+      cutlass::fmha::kernel::Sm100FmhaMlaKernelTmaWarpspecialized<
+          TileShape, Element, ElementAcc, ElementOut, ElementAcc, TileScheduler,
+          /*kIsCpAsync=*/true>;
+  using Fmha = cutlass::fmha::device::MLA<FmhaKernel>;
+};
+
+template <typename T>
+typename T::Fmha::Arguments args_from_options(
+    at::Tensor const& out, at::Tensor const& q_nope, at::Tensor const& q_pe,
+    at::Tensor const& kv_c_and_k_pe_cache, at::Tensor const& seq_lens,
+    at::Tensor const& page_table, double scale) {
+  cutlass::KernelHardwareInfo hw_info;
+  hw_info.device_id = q_nope.device().index();
+  hw_info.sm_count =
+      cutlass::KernelHardwareInfo::query_device_multiprocessor_count(
+          hw_info.device_id);
+
+  int batches = q_nope.sizes()[0];
+  int page_count_per_seq = page_table.sizes()[1];
+  int page_count_total = kv_c_and_k_pe_cache.sizes()[0];
+  int page_size = kv_c_and_k_pe_cache.sizes()[1];
+  int max_seq_len = page_size * page_count_per_seq;
+  using TileShapeH = typename T::TileShapeH;
+  using TileShapeD = typename T::TileShapeD;
+  auto problem_shape =
+      cute::make_tuple(TileShapeH{}, max_seq_len, TileShapeD{}, batches);
+
+  auto [H, K, D, B] = problem_shape;
+  auto [D_latent, D_rope] = D;
+
+  using StrideQ = typename T::StrideQ;
+  using StrideK = typename T::StrideK;
+  using StrideO = typename T::StrideO;
+  using StrideLSE = typename T::StrideLSE;
+
+  StrideQ stride_Q_latent = cute::make_tuple(
+      static_cast<int64_t>(D_latent), _1{}, static_cast<int64_t>(H * D_latent));
+  StrideQ stride_Q_rope = cute::make_tuple(static_cast<int64_t>(D_rope), _1{},
+                                           static_cast<int64_t>(H * D_rope));
+  StrideK stride_C =
+      cute::make_tuple(static_cast<int64_t>(D_latent + D_rope), _1{},
+                       static_cast<int64_t>(page_size * (D_latent + D_rope)));
+  StrideLSE stride_PT = cute::make_stride(_1{}, page_count_per_seq);
+  StrideLSE stride_LSE = cute::make_tuple(_1{}, static_cast<int>(H));
+  StrideO stride_O = cute::make_tuple(static_cast<int64_t>(D_latent), _1{},
+                                      static_cast<int64_t>(H * D_latent));
+
+  using Element = typename T::Element;
+  using ElementOut = typename T::ElementOut;
+  using ElementAcc = typename T::ElementAcc;
+  auto Q_latent_ptr = static_cast<Element*>(q_nope.data_ptr());
+  auto Q_rope_ptr = static_cast<Element*>(q_pe.data_ptr());
+  auto C_ptr = static_cast<Element*>(kv_c_and_k_pe_cache.data_ptr());
+  auto scale_f = static_cast<float>(scale);
+  typename T::Fmha::Arguments arguments{
+      problem_shape,
+      {scale_f, Q_latent_ptr, stride_Q_latent, Q_rope_ptr, stride_Q_rope, C_ptr,
+       stride_C, C_ptr + D_latent, stride_C,
+       static_cast<int*>(seq_lens.data_ptr()),
+       static_cast<int*>(page_table.data_ptr()), stride_PT, page_count_total,
+       page_size},
+      {static_cast<ElementOut*>(out.data_ptr()), stride_O,
+       static_cast<ElementAcc*>(nullptr), stride_LSE},
+      hw_info,
+      1,        // split_kv
+      nullptr,  // is_var_split_kv
+  };
+  // TODO(kaixih@nvidia): When split_kv=-1 and is_var_split_kv=false, we compute
+  // split_kv automatically based on batch size and sequence length to balance
+  // workload across available SMs. Consider using var_split_kv for manual
+  // control if needed.
+  T::Fmha::set_split_kv(arguments);
+  return arguments;
+}
+
+template <typename Element>
+void runMla(at::Tensor const& out, at::Tensor const& q_nope,
+            at::Tensor const& q_pe, at::Tensor const& kv_c_and_k_pe_cache,
+            at::Tensor const& seq_lens, at::Tensor const& page_table,
+            float scale, cudaStream_t stream) {
+  using MlaSm100Type = MlaSm100<Element>;
+  typename MlaSm100Type::Fmha fmha;
+  auto arguments = args_from_options<MlaSm100Type>(
+      out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens, page_table, scale);
+  size_t workspace_size = MlaSm100Type::Fmha::get_workspace_size(arguments);
+  auto const workspace_options =
+      torch::TensorOptions().dtype(torch::kUInt8).device(q_nope.device());
+  auto workspace = torch::empty(workspace_size, workspace_options);
+
+  CUTLASS_CHECK(fmha.can_implement(arguments));
+
+  CUTLASS_CHECK(fmha.initialize(arguments, workspace.data_ptr(), stream));
+
+  CUTLASS_CHECK(fmha.run(arguments, workspace.data_ptr(), stream));
+}
+
+void cutlass_mla_decode_sm100a(torch::Tensor const& out,
+                               torch::Tensor const& q_nope,
+                               torch::Tensor const& q_pe,
+                               torch::Tensor const& kv_c_and_k_pe_cache,
+                               torch::Tensor const& seq_lens,
+                               torch::Tensor const& page_table, double scale) {
+  TORCH_CHECK(q_nope.device().is_cuda(), "q_nope must be on CUDA");
+  TORCH_CHECK(q_nope.dim() == 3, "q_nope must be a 3D tensor");
+  TORCH_CHECK(q_pe.dim() == 3, "q_pe must be a 3D tensor");
+  TORCH_CHECK(kv_c_and_k_pe_cache.dim() == 3,
+              "kv_c_and_k_pe_cache must be a 3D tensor");
+  TORCH_CHECK(seq_lens.dim() == 1, "seq_lens must be a 1D tensor");
+  TORCH_CHECK(page_table.dim() == 2, "page_table must be a 2D tensor");
+  TORCH_CHECK(out.dim() == 3, "out must be a 3D tensor");
+
+  auto B_q_nope = q_nope.size(0);
+  auto H_q_nope = q_nope.size(1);
+  auto D_q_nope = q_nope.size(2);
+  auto B_q_pe = q_pe.size(0);
+  auto H_q_pe = q_pe.size(1);
+  auto D_q_pe = q_pe.size(2);
+  auto B_pt = page_table.size(0);
+  auto PAGE_NUM = page_table.size(1);
+  auto PAGE_SIZE = kv_c_and_k_pe_cache.size(1);
+  auto D_ckv = kv_c_and_k_pe_cache.size(2);
+  auto B_o = out.size(0);
+  auto H_o = out.size(1);
+  auto D_o = out.size(2);
+
+  TORCH_CHECK(D_q_nope == 512, "D_q_nope must be equal to 512");
+  TORCH_CHECK(D_q_pe == 64, "D_q_pe must be equal to 64");
+  TORCH_CHECK(D_ckv == 576, "D_ckv must be equal to 576");
+  TORCH_CHECK(H_q_nope == H_q_pe && H_q_nope == H_o && H_o == 128,
+              "H_q_nope, H_q_pe, and H_o must be equal to 128");
+  TORCH_CHECK(PAGE_SIZE > 0 && (PAGE_SIZE & (PAGE_SIZE - 1)) == 0,
+              "PAGE_SIZE must be a power of 2");
+  TORCH_CHECK(
+      B_q_nope == B_q_pe && B_q_nope == B_pt && B_q_nope == B_o,
+      "Batch dims must be same for page_table, q_nope and q_pe, and out");
+  TORCH_CHECK(PAGE_NUM % (128 / PAGE_SIZE) == 0,
+              "PAGE_NUM must be divisible by 128 / PAGE_SIZE");
+  TORCH_CHECK(D_o == 512, "D_o must be equal to 512");
+
+  TORCH_CHECK(q_nope.dtype() == at::ScalarType::Half ||
+                  q_nope.dtype() == at::ScalarType::BFloat16 ||
+                  q_nope.dtype() == at::ScalarType::Float8_e4m3fn,
+              "q_nope must be a half, bfloat16, or float8_e4m3fn tensor");
+  TORCH_CHECK(kv_c_and_k_pe_cache.dtype() == q_nope.dtype() &&
+                  q_nope.dtype() == q_pe.dtype(),
+              "kv_c_and_k_pe_cache, q_nope, and q_pe must be the same type");
+  TORCH_CHECK(seq_lens.dtype() == torch::kInt32,
+              "seq_lens must be a 32-bit integer tensor");
+  TORCH_CHECK(page_table.dtype() == torch::kInt32,
+              "page_table must be a 32-bit integer tensor");
+
+  auto in_dtype = q_nope.dtype();
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(q_nope));
+  const cudaStream_t stream =
+      at::cuda::getCurrentCUDAStream(q_nope.get_device());
+  if (in_dtype == at::ScalarType::Half) {
+    runMla<cutlass::half_t>(out, q_nope, q_pe, kv_c_and_k_pe_cache, seq_lens,
+                            page_table, scale, stream);
+  } else if (in_dtype == at::ScalarType::BFloat16) {
+    runMla<cutlass::bfloat16_t>(out, q_nope, q_pe, kv_c_and_k_pe_cache,
+                                seq_lens, page_table, scale, stream);
+  } else if (in_dtype == at::ScalarType::Float8_e4m3fn) {
+    runMla<cutlass::float_e4m3_t>(out, q_nope, q_pe, kv_c_and_k_pe_cache,
+                                  seq_lens, page_table, scale, stream);
+  } else {
+    TORCH_CHECK(false, "Unsupported input data type of MLA");
+  }
+}

csrc/attention/mla/cutlass_sm100_mla/device/sm100_mla.hpp

@@ -133,14 +133,6 @@ public:
     // printf("    sm_count = %d\n", sm_count);
     int max_splits = ceil_div(K, 128);
     max_splits = min(16, max_splits);
-
-    // TODO: This avoids a hang when the batch size larger than 1 and 
-    // there is more than 1 kv_splits. 
-    // Discuss with NVIDIA how this can be fixed.
-    if (B > 1) {
-      max_splits = min(1, max_splits);
-    }
-    
     // printf("    max_splits = %d\n", max_splits);
     int sms_per_batch = max(1, sm_count / B);
     // printf("    sms_per_batch = %d\n", sms_per_batch);

csrc/attention/mla/cutlass_sm100_mla/kernel/sm100_fmha_mla_tma_warpspecialized.hpp

@@ -580,22 +580,22 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
       for (; tile_scheduler.is_valid(); ++tile_scheduler) {
         auto blk_coord = tile_scheduler.get_block_coord();
         auto problem_shape = params.problem_shape;
-        auto local_split_kv = params.split_kv;
+	auto local_split_kv = params.split_kv;
         if (params.mainloop.ptr_seq != nullptr) {
           get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
-          if (params.ptr_split_kv != nullptr) {
+	  if (params.ptr_split_kv != nullptr) {
             local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
           }
         }
-        if (local_split_kv <= get<3>(blk_coord))
-          continue;
+	if (local_split_kv <= get<3>(blk_coord))
+	  continue;
         load_page_table(
           blk_coord,
           problem_shape,
           params.mainloop,
           shared_storage.tensors,
           pipeline_page_table, pipeline_pt_producer_state,
-          local_split_kv
+	  local_split_kv
         );
       }
     }
@@ -604,15 +604,15 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
         CUTLASS_PRAGMA_NO_UNROLL
         for (; tile_scheduler.is_valid(); ++tile_scheduler) {
           auto blk_coord = tile_scheduler.get_block_coord();
-          auto problem_shape = params.problem_shape;
-          auto local_split_kv = params.split_kv;
+	  auto problem_shape = params.problem_shape;
+	  auto local_split_kv = params.split_kv;
           if (params.mainloop.ptr_seq != nullptr) {
             get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
-            if (params.ptr_split_kv != nullptr) {
+	    if (params.ptr_split_kv != nullptr) {
               local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
             }
           }
-          if (local_split_kv <= get<3>(blk_coord))
+	  if (local_split_kv <= get<3>(blk_coord))
             continue;
           load_cpasync(
             blk_coord,
@@ -621,7 +621,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
             params.mainloop_params,
             shared_storage.tensors,
             pipeline_load_qk, pipeline_load_qk_producer_state,
-            local_split_kv,
+	    local_split_kv,
             /* must be shared pipe */
             pipeline_page_table, pipeline_pt_consumer_state
           );
@@ -633,15 +633,15 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
           CUTLASS_PRAGMA_NO_UNROLL
           for (; tile_scheduler.is_valid(); ++tile_scheduler) {
             auto blk_coord = tile_scheduler.get_block_coord();
-            auto problem_shape = params.problem_shape;
-            auto local_split_kv = params.split_kv;
+	    auto problem_shape = params.problem_shape;
+	    auto local_split_kv = params.split_kv;
             if (params.mainloop.ptr_seq != nullptr) {
               get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
-              if (params.ptr_split_kv != nullptr) {
-                local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
-              }
+	      if (params.ptr_split_kv != nullptr) {
+	        local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
+	      }
             }
-            if (local_split_kv <= get<3>(blk_coord))
+	    if (local_split_kv <= get<3>(blk_coord))
               continue;
             load_tma</* paged= */ true>(
               blk_coord,
@@ -651,7 +651,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
               shared_storage.tensors,
               pipeline_load_qk, pipeline_load_qk_producer_state,
               pipeline_load_qk, pipeline_load_qk_producer_state,
-              local_split_kv
+	      local_split_kv
             );
             cutlass::arch::NamedBarrier((kNumComputeWarps + kNumLoadWarps) * NumThreadsPerWarp, kNamedBarrierEpilogue).arrive_and_wait();
           }
@@ -660,15 +660,15 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
           CUTLASS_PRAGMA_NO_UNROLL
           for (; tile_scheduler.is_valid(); ++tile_scheduler) {
             auto blk_coord = tile_scheduler.get_block_coord();
-            auto problem_shape = params.problem_shape;
-            auto local_split_kv = params.split_kv;
+	    auto problem_shape = params.problem_shape;
+	    auto local_split_kv = params.split_kv;
             if (params.mainloop.ptr_seq != nullptr) {
               get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
-              if (params.ptr_split_kv != nullptr) {
+	      if (params.ptr_split_kv != nullptr) {
                 local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
-              }
+	      }
             }
-            if (local_split_kv <= get<3>(blk_coord))
+	    if (local_split_kv <= get<3>(blk_coord))
               continue;
             load_tma<false>(
               blk_coord,
@@ -678,7 +678,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
               shared_storage.tensors,
               pipeline_load_qk, pipeline_load_qk_producer_state,
               pipeline_load_qk, pipeline_load_qk_producer_state,
-              local_split_kv
+	      local_split_kv
             );
             cutlass::arch::NamedBarrier((kNumComputeWarps + kNumLoadWarps) * NumThreadsPerWarp, kNamedBarrierEpilogue).arrive_and_wait();
           }
@@ -694,14 +694,14 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
         for (; tile_scheduler.is_valid(); ++tile_scheduler) {
           auto blk_coord = tile_scheduler.get_block_coord();
           auto problem_shape = params.problem_shape;
-          auto local_split_kv = params.split_kv;
+	  auto local_split_kv = params.split_kv;
           if (params.mainloop.ptr_seq != nullptr) {
             get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
             if (params.ptr_split_kv != nullptr) {
                 local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
             }
           }
-          if (local_split_kv <= get<3>(blk_coord))
+	  if (local_split_kv <= get<3>(blk_coord))
             continue;
           mma(blk_coord,
             problem_shape,
@@ -711,7 +711,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
             pipeline_mma_s, pipeline_mma_s_producer_state,
             pipeline_p_mma, pipeline_p_mma_consumer_state,
             pipeline_mma_o, pipeline_mma_o_producer_state,
-            local_split_kv
+	    local_split_kv
           );
         }
       }
@@ -726,15 +726,15 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
       for (; tile_scheduler.is_valid(); ++tile_scheduler) {
         auto blk_coord = tile_scheduler.get_block_coord();
         auto problem_shape = params.problem_shape;
-        auto split_kv = params.split_kv;
-        auto local_split_kv = split_kv;
+	auto split_kv = params.split_kv;
+	auto local_split_kv = split_kv;
         if (params.mainloop.ptr_seq != nullptr) {
           get<1>(problem_shape) = params.mainloop.ptr_seq[get<2>(blk_coord)];
-          if (params.ptr_split_kv != nullptr) {
+	  if (params.ptr_split_kv != nullptr) {
             local_split_kv = params.ptr_split_kv[get<2>(blk_coord)];
           }
         }
-        if (local_split_kv <= get<3>(blk_coord))
+	if (local_split_kv <= get<3>(blk_coord))
           continue;
         compute(
           blk_coord,
@@ -745,7 +745,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
           pipeline_mma_s, pipeline_mma_s_consumer_state,
           pipeline_p_mma, pipeline_p_mma_producer_state,
           pipeline_mma_o, pipeline_mma_o_consumer_state,
-          local_split_kv
+	  local_split_kv
         );
       }
 
@@ -1900,7 +1900,7 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized {
       cutlass::arch::NamedBarrier(
           (kNumComputeWarps + kNumLoadWarps) * NumThreadsPerWarp,
           kNamedBarrierEpilogue
-      ).arrive_and_wait();
+      ).arrive();
 
       return;
     }

csrc/attention/mla/sm100_cutlass_mla_kernel.cu

@@ -43,7 +43,6 @@ void sm100_cutlass_mla_decode(
     torch::Tensor const& seq_lens,
     torch::Tensor const& page_table,
     torch::Tensor const& workspace,
-    double sm_scale,
     int64_t num_kv_splits) {
   TORCH_CHECK(false, "CUDA version must be >= 12.4 for cutlass_mla_decode");
 }

csrc/cache.h

@@ -56,11 +56,3 @@ void cp_gather_cache(
     torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
     torch::Tensor const& cu_seq_lens,  // [BATCH+1]
     int64_t batch_size, std::optional<torch::Tensor> seq_starts = std::nullopt);
-
-// Indexer K quantization and cache function
-void indexer_k_quant_and_cache(
-    torch::Tensor& k,             // [num_tokens, head_dim]
-    torch::Tensor& kv_cache,      // [num_blocks, block_size, cache_stride]
-    torch::Tensor& slot_mapping,  // [num_tokens]
-    int64_t quant_block_size,     // quantization block size
-    const std::string& scale_fmt);

csrc/cache_kernels.cu

@@ -16,7 +16,6 @@
 
 #include <algorithm>
 #include <cassert>
-#include <cfloat>  // FLT_MIN
 #include <map>
 #include <vector>
 
@@ -397,180 +396,6 @@ __global__ void concat_and_cache_mla_kernel(
   copy(k_pe, kv_cache, k_pe_stride, block_stride, pe_dim, kv_lora_rank);
 }
 
-template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
-__global__ void concat_and_cache_ds_mla_kernel(
-    const scalar_t* __restrict__ kv_c,  // [num_tokens, kv_lora_rank]
-    const scalar_t* __restrict__ k_pe,  // [num_tokens, pe_dim]
-    cache_t* __restrict__ kv_cache,  // [num_blocks, block_size, (kv_lora_rank
-                                     // + pe_dim)]
-    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
-    const int block_stride,                    //
-    const int entry_stride,                    //
-    const int kv_c_stride,                     //
-    const int k_pe_stride,                     //
-    const int kv_lora_rank,                    //
-    const int pe_dim,                          //
-    const int block_size,                      //
-    const float* scale                         //
-) {
-  const int64_t token_idx = blockIdx.x;
-  const int64_t slot_idx = slot_mapping[token_idx];
-  // NOTE: slot_idx can be -1 if the token is padded
-  if (slot_idx < 0) {
-    return;
-  }
-  const int64_t block_idx = slot_idx / block_size;
-  const int64_t block_offset = slot_idx % block_size;
-  const int64_t dst_idx_start =
-      block_idx * block_stride + block_offset * entry_stride;
-
-  // Create 4 tile scales in shared memory
-  __shared__ float smem[20];
-  float* shard_abs_max = smem;
-  float* tile_scales = smem + 16;
-
-  // For the NoPE part, each tile of 128 elements is handled by 4 warps
-  // (128 threads). There are 4 total tiles, so 16 warps (512 threads).
-  // The first thread of the first warp in each tile writes the scale
-  // value for the tile. The RoPE part (last 64 elements) is handled
-  // by another 2 warps (64 threads).
-  // So in total, we use 18 warps (576 threads) per block.
-
-  // Cast kv_cache to 16_bit for RoPE values
-  scalar_t* kv_cache_16bit =
-      reinterpret_cast<scalar_t*>(&kv_cache[dst_idx_start]);
-
-  // The last 64 threads handle the RoPE part
-  if (threadIdx.x >= kv_lora_rank) {
-    const int8_t pe_idx = threadIdx.x - kv_lora_rank;
-    const int64_t src_idx = token_idx * k_pe_stride + pe_idx;
-    // RoPE values start after the packed 8-bit NoPE values and the
-    // 32-bit scales
-    const int64_t dst_idx = kv_lora_rank / 2 + 8 + pe_idx;
-    kv_cache_16bit[dst_idx] = k_pe[src_idx];
-    return;
-  }
-
-  // Determine the scale for each chunk of NoPE
-  const int16_t tile_idx = threadIdx.x >> 7;
-  const int16_t warp_idx = (threadIdx.x & 127) >> 5;
-  const int16_t lane_idx = threadIdx.x & 31;
-
-  // Load the NoPE element for this thread into registers
-  const int64_t src_idx = token_idx * kv_c_stride + threadIdx.x;
-  const scalar_t src_val = kv_c[src_idx];
-
-  // Warp-level reduction to find the max absolute value in the warp
-  float max_abs = fabsf(src_val);
-#pragma unroll
-  for (int offset = 16; offset > 0; offset /= 2) {
-#ifdef USE_ROCM
-    max_abs = fmaxf(max_abs, __shfl_down_sync(UINT64_MAX, max_abs, offset));
-#else
-    max_abs = fmaxf(max_abs, __shfl_down_sync(0xFFFFFFFF, max_abs, offset));
-#endif
-  }
-
-  // The first lane of each warp in each tile writes the max_abs of this part
-  // of the tile to shared memory
-  if (lane_idx == 0) {
-    shard_abs_max[tile_idx * 4 + warp_idx] = max_abs;
-  }
-  __syncthreads();
-
-  // The first lane of the first warp in each tile computes the scale for the
-  // tile and writes it to shared memory and to kv_cache
-  if (warp_idx == 0 && lane_idx == 0) {
-    float4 shard_abs_max_vec =
-        reinterpret_cast<float4*>(shard_abs_max)[tile_idx];
-    float tile_scale = fmaxf(fmaxf(shard_abs_max_vec.x, shard_abs_max_vec.y),
-                             fmaxf(shard_abs_max_vec.z, shard_abs_max_vec.w)) /
-                       448.f;
-
-    // Avoid division by zero in `scaled_convert`
-    tile_scales[tile_idx] = fmaxf(tile_scale, FLT_MIN);
-    float* kv_cache_32bit = reinterpret_cast<float*>(&kv_cache[dst_idx_start]);
-    const uint64_t dst_idx = kv_lora_rank / 4 + tile_idx;
-    kv_cache_32bit[dst_idx] = tile_scales[tile_idx];
-  }
-
-  __syncthreads();
-
-  // Now all threads in the block scale and write their element
-  const float scale_val = tile_scales[tile_idx];
-  const int64_t dst_idx = dst_idx_start + threadIdx.x;
-  kv_cache[dst_idx] =
-      fp8::scaled_convert<uint8_t, scalar_t, Fp8KVCacheDataType::kFp8E4M3>(
-          src_val, scale_val);
-}
-
-template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
-__global__ void indexer_k_quant_and_cache_kernel(
-    const scalar_t* __restrict__ k,  // [num_tokens, head_dim]
-    cache_t* __restrict__ kv_cache,  // [num_blocks, block_size, cache_stride]
-    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
-    const int head_dim,                        // dimension of each head
-    const int quant_block_size,                // quantization block size
-    const int cache_block_size,                // cache block size
-    const int cache_stride,  // stride for each token in kv_cache
-    const bool use_ue8m0     // use ue8m0 scale format
-) {
-  constexpr int VEC_SIZE = 4;
-  const int64_t token_idx = blockIdx.x;
-  const int64_t head_dim_idx = (blockIdx.y * blockDim.y * blockDim.x +
-                                threadIdx.y * blockDim.x + threadIdx.x) *
-                               VEC_SIZE;
-  const int64_t slot_idx = slot_mapping[token_idx];
-  const int64_t block_idx = slot_idx / cache_block_size;
-  const int64_t block_offset = slot_idx % cache_block_size;
-
-  // NOTE: slot_idx can be -1 if the token is padded
-  if (slot_idx < 0 || (head_dim_idx >= head_dim)) {
-    return;
-  }
-
-  float2 k_val = (reinterpret_cast<const float2*>(
-      k))[(token_idx * head_dim + head_dim_idx) / VEC_SIZE];
-  scalar_t* k_val_ptr = reinterpret_cast<scalar_t*>(&k_val);
-  float amax = 0.0f;
-  for (int i = 0; i < VEC_SIZE; i++) {
-    amax = fmaxf(amax, fabsf(float(k_val_ptr[i])));
-  }
-#ifndef USE_ROCM
-  __syncwarp();
-#endif
-
-  // Reduced amax
-  for (int mask = 16; mask > 0; mask /= 2) {
-#ifdef USE_ROCM
-    amax = fmaxf(amax, __shfl_xor_sync(uint64_t(-1), amax, mask));
-#else
-    amax = fmaxf(amax, __shfl_xor_sync(unsigned(-1), amax, mask));
-#endif
-  }
-#ifndef USE_ROCM
-  __syncwarp();
-#endif
-  float scale = fmaxf(amax, 1e-4) / 448.0f;
-  if (use_ue8m0) {
-    scale = exp2f(ceilf(log2f(scale)));
-  }
-
-  const int64_t dst_offset = block_idx * cache_block_size * cache_stride +
-                             block_offset * head_dim + head_dim_idx;
-  for (int i = 0; i < VEC_SIZE; i++) {
-    kv_cache[dst_offset + i] =
-        fp8::scaled_convert<cache_t, scalar_t, kv_dt>(k_val_ptr[i], scale);
-  }
-  if (threadIdx.x == 0) {
-    const int64_t dst_scale_idx =
-        block_idx * cache_block_size * cache_stride +
-        cache_block_size * head_dim +
-        (block_offset * head_dim + head_dim_idx) * 4 / quant_block_size;
-    reinterpret_cast<float*>(kv_cache)[dst_scale_idx / 4] = scale;
-  }
-}
-
 }  // namespace vllm
 
 // KV_T is the data type of key and value tensors.
@@ -613,7 +438,7 @@ void reshape_and_cache(
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
   DISPATCH_BY_KV_CACHE_DTYPE(key.dtype(), kv_cache_dtype,
-                             CALL_RESHAPE_AND_CACHE);
+                             CALL_RESHAPE_AND_CACHE)
 }
 
 // KV_T is the data type of key and value tensors.
@@ -684,18 +509,6 @@ void reshape_and_cache_flash(
           kv_c_stride, k_pe_stride, kv_lora_rank, pe_dim, block_size,   \
           reinterpret_cast<const float*>(scale.data_ptr()));
 
-// KV_T is the data type of key and value tensors.
-// CACHE_T is the stored data type of kv-cache.
-#define CALL_CONCAT_AND_CACHE_DS_MLA(KV_T, CACHE_T, KV_DTYPE)           \
-  vllm::concat_and_cache_ds_mla_kernel<KV_T, CACHE_T, KV_DTYPE>         \
-      <<<grid, block, 0, stream>>>(                                     \
-          reinterpret_cast<KV_T*>(kv_c.data_ptr()),                     \
-          reinterpret_cast<KV_T*>(k_pe.data_ptr()),                     \
-          reinterpret_cast<CACHE_T*>(kv_cache.data_ptr()),              \
-          slot_mapping.data_ptr<int64_t>(), block_stride, entry_stride, \
-          kv_c_stride, k_pe_stride, kv_lora_rank, pe_dim, block_size,   \
-          reinterpret_cast<const float*>(scale.data_ptr()));
-
 void concat_and_cache_mla(
     torch::Tensor& kv_c,          // [num_tokens, kv_lora_rank]
     torch::Tensor& k_pe,          // [num_tokens, pe_dim]
@@ -718,44 +531,20 @@ void concat_and_cache_mla(
   int pe_dim = k_pe.size(1);
   int block_size = kv_cache.size(1);
 
-  if (kv_cache_dtype == "fp8_ds_mla") {
-    TORCH_CHECK(kv_lora_rank == 512, "kv_lora_rank must be 512 for fp8_ds_mla");
-    TORCH_CHECK(pe_dim == 64, "pe_dim must be 64 for fp8_ds_mla");
-    TORCH_CHECK(kv_cache.size(2) == 656 / kv_cache.itemsize(),
-                "kv_cache.size(2) must be 656 bytes for fp8_ds_mla");
-    TORCH_CHECK(kv_c.itemsize() == 2,
-                "kv_c.itemsize() must be 2 for fp8_ds_mla");
-    TORCH_CHECK(k_pe.itemsize() == 2,
-                "k_pe.itemsize() must be 2 for fp8_ds_mla");
-  } else {
-    TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
-  }
+  TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
 
   int kv_c_stride = kv_c.stride(0);
   int k_pe_stride = k_pe.stride(0);
   int block_stride = kv_cache.stride(0);
   int entry_stride = kv_cache.stride(1);
 
+  dim3 grid(num_tokens);
+  dim3 block(std::min(kv_lora_rank, 512));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(kv_c));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  if (kv_cache_dtype == "fp8_ds_mla") {
-    dim3 grid(num_tokens);
-    // For the NoPE part, each tile of 128 elements is handled by 4 warps
-    // (128 threads). There are 4 total tiles, so 16 warps (512 threads).
-    // The first thread of the first warp in each tile writes the scale
-    // value for the tile. The RoPE part (last 64 elements) is handled
-    // by another 2 warps (64 threads).
-    // So in total, we use 18 warps (576 threads) per block.
-    dim3 block(576);
-    DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
-                               CALL_CONCAT_AND_CACHE_DS_MLA);
-  } else {
-    dim3 grid(num_tokens);
-    dim3 block(std::min(kv_lora_rank, 512));
-    DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
-                               CALL_CONCAT_AND_CACHE_MLA);
-  }
+  DISPATCH_BY_KV_CACHE_DTYPE(kv_c.dtype(), kv_cache_dtype,
+                             CALL_CONCAT_AND_CACHE_MLA);
 }
 
 namespace vllm {
@@ -1133,42 +922,3 @@ void cp_gather_cache(
     TORCH_CHECK(false, "Unsupported data type width: ", dtype_bits);
   }
 }
-
-// Macro to dispatch the kernel based on the data type.
-#define CALL_INDEXER_K_QUANT_AND_CACHE(KV_T, CACHE_T, KV_DTYPE)         \
-  vllm::indexer_k_quant_and_cache_kernel<KV_T, CACHE_T, KV_DTYPE>       \
-      <<<grid, block, 0, stream>>>(                                     \
-          reinterpret_cast<KV_T*>(k.data_ptr()),                        \
-          reinterpret_cast<CACHE_T*>(kv_cache.data_ptr()),              \
-          slot_mapping.data_ptr<int64_t>(), head_dim, quant_block_size, \
-          cache_block_size, cache_stride, use_ue8m0);
-
-void indexer_k_quant_and_cache(
-    torch::Tensor& k,             // [num_tokens, head_dim]
-    torch::Tensor& kv_cache,      // [num_blocks, block_size, cache_stride]
-    torch::Tensor& slot_mapping,  // [num_tokens]
-    int64_t quant_block_size,     // quantization block size
-    const std::string& scale_fmt) {
-  int num_tokens = k.size(0);
-  int head_dim = k.size(1);
-  int cache_block_size = kv_cache.size(1);
-  int cache_stride = kv_cache.size(2);
-  bool use_ue8m0 = scale_fmt == "ue8m0";
-
-  TORCH_CHECK(k.device() == kv_cache.device(),
-              "k and kv_cache must be on the same device");
-  TORCH_CHECK(k.device() == slot_mapping.device(),
-              "k and slot_mapping must be on the same device");
-  TORCH_CHECK(head_dim % quant_block_size == 0,
-              "head_dim must be divisible by quant_block_size");
-
-  constexpr int vec_size = 4;
-  dim3 grid(num_tokens, (head_dim + quant_block_size * vec_size - 1) /
-                            (quant_block_size * vec_size));
-  dim3 block(32, vec_size);
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(k));
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-
-  DISPATCH_BY_KV_CACHE_DTYPE(k.dtype(), "fp8_e4m3",
-                             CALL_INDEXER_K_QUANT_AND_CACHE);
-}

csrc/cpu/cpu_types.hpp

@@ -14,12 +14,7 @@
   // arm implementation
   #include "cpu_types_arm.hpp"
 #else
-  #warning "unsupported vLLM cpu implementation, vLLM will compile with scalar"
-  #include "cpu_types_scalar.hpp"
-#endif
-
-#ifdef _OPENMP
-  #include <omp.h>
+  #warning "unsupported vLLM cpu implementation"
 #endif
 
 #endif

csrc/cpu/cpu_types_scalar.hpp

@@ -1,513 +0,0 @@
-#include <cmath>
-#include <cstdint>
-#include <cstring>
-#include <torch/all.h>
-#include "float_convert.hpp"
-
-namespace vec_op {
-
-#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)            \
-  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)    \
-  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
-  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
-
-#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
-  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
-
-#ifndef CPU_OP_GUARD
-  #define CPU_KERNEL_GUARD_IN(NAME)
-  #define CPU_KERNEL_GUARD_OUT(NAME)
-#else
-  #define CPU_KERNEL_GUARD_IN(NAME) \
-    std::cout << #NAME << " invoked." << std::endl;
-  #define CPU_KERNEL_GUARD_OUT(NAME) \
-    std::cout << #NAME << " exit." << std::endl;
-#endif
-
-#define FORCE_INLINE __attribute__((always_inline)) inline
-
-#define __max(a, b) ((a) > (b) ? (a) : (b))
-#define __min(a, b) ((a) < (b) ? (a) : (b))
-#define __abs(a) ((a) < (0) ? (0 - a) : (a))
-
-typedef struct f16x8_t {
-  uint16_t val[8];
-} f16x8_t;
-
-typedef struct f16x16_t {
-  uint16_t val[16];
-} f16x16_t;
-
-typedef struct f16x32_t {
-  uint16_t val[32];
-} f16x32_t;
-
-typedef struct f32x4_t {
-  float val[4];
-} f32x4_t;
-
-typedef struct f32x8_t {
-  float val[8];
-} f32x8_t;
-
-typedef struct f32x16_t {
-  float val[16];
-} f32x16_t;
-
-namespace {
-template <typename T, T... indexes, typename F>
-constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
-  (f(std::integral_constant<T, indexes>{}), ...);
-};
-};  // namespace
-
-template <typename T, T count, typename F,
-          typename = std::enable_if_t<std::is_invocable_v<F, T> > >
-constexpr void unroll_loop(F&& f) {
-  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
-}
-
-template <typename T>
-struct Vec {
-  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; }
-};
-
-struct FP32Vec8;
-struct FP32Vec16;
-
-struct FP16Vec8 : public Vec<FP16Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-  f16x8_t reg;
-
-  explicit FP16Vec8(const void* ptr)
-      : reg(*reinterpret_cast<const f16x8_t*>(ptr)) {};
-
-  explicit FP16Vec8(const FP32Vec8&);
-
-  void save(void* ptr) const { *reinterpret_cast<f16x8_t*>(ptr) = reg; }
-};
-
-struct FP16Vec16 : public Vec<FP16Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  f16x16_t reg;
-
-  explicit FP16Vec16(const void* ptr)
-      : reg(*reinterpret_cast<const f16x16_t*>(ptr)) {};
-
-  explicit FP16Vec16(const FP32Vec16&);
-
-  void save(void* ptr) const { *reinterpret_cast<f16x16_t*>(ptr) = reg; }
-
-  void save(void* ptr, const int elem_num) const {
-    int num = __min(elem_num, VEC_ELEM_NUM);
-    std::memcpy(ptr, &(reg.val[0]), num * sizeof(uint16_t));
-  }
-};
-
-struct BF16Vec8 : public Vec<BF16Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-  f16x8_t reg;
-
-  explicit BF16Vec8(const void* ptr)
-      : reg(*reinterpret_cast<const f16x8_t*>(ptr)) {};
-
-  explicit BF16Vec8(const FP32Vec8&);
-
-  void save(void* ptr) const { *reinterpret_cast<f16x8_t*>(ptr) = reg; }
-};
-
-struct BF16Vec16 : public Vec<BF16Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  f16x16_t reg;
-
-  explicit BF16Vec16(const void* ptr)
-      : reg(*reinterpret_cast<const f16x16_t*>(ptr)) {};
-
-  explicit BF16Vec16(const FP32Vec16&);
-
-  void save(void* ptr) const { *reinterpret_cast<f16x16_t*>(ptr) = reg; }
-
-  void save(void* ptr, const int elem_num) const {
-    int num = __min(elem_num, VEC_ELEM_NUM);
-    std::memcpy(ptr, &(reg.val[0]), num * sizeof(uint16_t));
-  }
-};
-
-struct BF16Vec32 : public Vec<BF16Vec32> {
-  constexpr static int VEC_ELEM_NUM = 32;
-  f16x32_t reg;
-
-  explicit BF16Vec32(const void* ptr)
-      : reg(*reinterpret_cast<const f16x32_t*>(ptr)) {};
-
-  explicit BF16Vec32(f16x32_t data) : reg(data) {};
-
-  explicit BF16Vec32(BF16Vec8& vec8_data) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = vec8_data.reg.val[i % BF16Vec8::VEC_ELEM_NUM];
-    }
-  }
-
-  void save(void* ptr) const { *reinterpret_cast<f16x32_t*>(ptr) = reg; }
-};
-
-struct FP32Vec4 : public Vec<FP32Vec4> {
-  constexpr static int VEC_ELEM_NUM = 4;
-
-  f32x4_t reg;
-
-  explicit FP32Vec4(float v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = v;
-    }
-  }
-
-  explicit FP32Vec4() {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = 0.0f;
-    }
-  }
-
-  explicit FP32Vec4(const float* ptr)
-      : reg(*reinterpret_cast<const f32x4_t*>(ptr)) {};
-
-  explicit FP32Vec4(f32x4_t data) : reg(data) {};
-
-  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {};
-};
-
-struct FP32Vec8 : public Vec<FP32Vec8> {
-  constexpr static int VEC_ELEM_NUM = 8;
-
-  f32x8_t reg;
-
-  explicit FP32Vec8(float v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = v;
-    }
-  }
-
-  explicit FP32Vec8() {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = 0.0f;
-    }
-  }
-
-  explicit FP32Vec8(const float* ptr)
-      : reg(*reinterpret_cast<const f32x8_t*>(ptr)) {};
-
-  explicit FP32Vec8(f32x8_t data) : reg(data) {};
-
-  explicit FP32Vec8(const FP32Vec8& data) : reg(data.reg) {};
-
-  explicit FP32Vec8(const FP16Vec8& v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = fp16_to_float(v.reg.val[i]);
-    }
-  }
-
-  FP32Vec8(const BF16Vec8& v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = bf16_to_float(v.reg.val[i]);
-    }
-  }
-
-  float reduce_sum() const {
-    float result = 0;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result += reg.val[i];
-    }
-    return result;
-  }
-
-  FP32Vec8 exp() const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = expf(reg.val[i]);
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 tanh() const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = tanhf(reg.val[i]);
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 er() const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = erf(reg.val[i]);
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 operator*(const FP32Vec8& b) const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = reg.val[i] * b.reg.val[i];
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 operator+(const FP32Vec8& b) const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = reg.val[i] + b.reg.val[i];
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 operator-(const FP32Vec8& b) const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = reg.val[i] - b.reg.val[i];
-    }
-    return FP32Vec8(ret);
-  }
-
-  FP32Vec8 operator/(const FP32Vec8& b) const {
-    f32x8_t ret;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      ret.val[i] = reg.val[i] / b.reg.val[i];
-    }
-    return FP32Vec8(ret);
-  }
-
-  void save(void* ptr) const { *reinterpret_cast<f32x8_t*>(ptr) = reg; }
-};
-
-struct FP32Vec16 : public Vec<FP32Vec16> {
-  constexpr static int VEC_ELEM_NUM = 16;
-  f32x16_t reg;
-
-  explicit FP32Vec16(float v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = v;
-    }
-  }
-
-  explicit FP32Vec16() {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = 0.0f;
-    }
-  }
-
-  explicit FP32Vec16(const float* ptr)
-      : reg(*reinterpret_cast<const f32x16_t*>(ptr)) {};
-
-  explicit FP32Vec16(f32x16_t data) : reg(data) {};
-
-  FP32Vec16(const FP32Vec4& data) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = data.reg.val[i % FP32Vec4::VEC_ELEM_NUM];
-    }
-  }
-
-  FP32Vec16(const FP32Vec8& data) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = data.reg.val[i % FP32Vec8::VEC_ELEM_NUM];
-    }
-  }
-
-  FP32Vec16(const FP32Vec16& data) : reg(data.reg) {};
-
-  explicit FP32Vec16(const FP16Vec16& v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = fp16_to_float(v.reg.val[i]);
-    }
-  }
-
-  explicit FP32Vec16(const BF16Vec16& v) {
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      reg.val[i] = bf16_to_float(v.reg.val[i]);
-    }
-  }
-
-  explicit FP32Vec16(const FP16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
-
-  FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {};
-
-  FP32Vec16 operator*(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = reg.val[i] * b.reg.val[i];
-    }
-    return result;
-  }
-
-  FP32Vec16 operator+(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = reg.val[i] + b.reg.val[i];
-    }
-    return result;
-  }
-
-  FP32Vec16 operator-(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = reg.val[i] - b.reg.val[i];
-    }
-    return result;
-  }
-
-  FP32Vec16 operator/(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = reg.val[i] / b.reg.val[i];
-    }
-    return result;
-  }
-
-  FP32Vec16 max(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = __max(reg.val[i], b.reg.val[i]);
-    }
-    return result;
-  }
-
-  FP32Vec16 min(const FP32Vec16& b) const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = __min(reg.val[i], b.reg.val[i]);
-    }
-    return result;
-  }
-
-  FP32Vec16 abs() const {
-    FP32Vec16 result(0.0f);
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result.reg.val[i] = __abs(reg.val[i]);
-    }
-    return result;
-  }
-
-  float reduce_sum() const {
-    float result = 0.0f;
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result += reg.val[i];
-    }
-    return result;
-  }
-
-  float reduce_max() const {
-    float result = reg.val[0];
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result = __max(reg.val[i], result);
-    }
-    return result;
-  }
-
-  float reduce_min() const {
-    float result = reg.val[0];
-    for (int i = 0; i < VEC_ELEM_NUM; ++i) {
-      result = __min(reg.val[i], result);
-    }
-    return result;
-  }
-
-  template <int group_size>
-  float reduce_sub_sum(int idx) {
-    static_assert(VEC_ELEM_NUM % group_size == 0);
-    float sum = 0.0;
-    int start = idx * group_size;
-    int end = (idx + 1) * group_size;
-
-    for (; (start < VEC_ELEM_NUM) && (start < end); ++start) {
-      sum += reg.val[start];
-    }
-
-    return sum;
-  }
-
-  void save(void* ptr) const { *reinterpret_cast<f32x16_t*>(ptr) = reg; }
-};
-
-template <typename T>
-struct VecType {
-  using vec_type = void;
-};
-
-template <typename T>
-using vec_t = typename VecType<T>::vec_type;
-
-template <>
-struct VecType<float> {
-  using vec_type = FP32Vec8;
-};
-
-template <>
-struct VecType<c10::Half> {
-  using vec_type = FP16Vec8;
-};
-
-template <>
-struct VecType<c10::BFloat16> {
-  using vec_type = BF16Vec8;
-};
-
-template <typename T>
-void storeFP32(float v, T* ptr) {
-  *ptr = v;
-}
-
-/*
-template <> inline void storeFP32<c10::Half>(float v, c10::Half *ptr) {
-  c10::Half __attribute__((__may_alias__)) *v_ptr =
-      reinterpret_cast<c10::Half *>(&v);
-  *ptr = *(v_ptr + 1);
-}
-*/
-
-template <>
-inline void storeFP32<c10::Half>(float v, c10::Half* ptr) {
-  uint16_t fp16 = float_to_fp16(v);
-  *reinterpret_cast<uint16_t*>(ptr) = fp16;
-}
-
-template <>
-inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
-  c10::BFloat16 __attribute__((__may_alias__))* v_ptr =
-      reinterpret_cast<c10::BFloat16*>(&v);
-  *ptr = *(v_ptr + 1);
-}
-
-inline FP16Vec16::FP16Vec16(const FP32Vec16& v) {
-  int i = 0;
-  for (i = 0; i < FP16Vec16::VEC_ELEM_NUM; ++i) {
-    reg.val[i] = float_to_fp16(v.reg.val[i]);
-  }
-}
-
-inline FP16Vec8 ::FP16Vec8(const FP32Vec8& v) {
-  int i = 0;
-  for (i = 0; i < FP16Vec8::VEC_ELEM_NUM; ++i) {
-    reg.val[i] = float_to_fp16(v.reg.val[i]);
-  }
-}
-
-inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) {
-  acc = acc + a * b;
-}
-
-inline BF16Vec8::BF16Vec8(const FP32Vec8& v) {
-  int i = 0;
-  for (i = 0; i < BF16Vec8::VEC_ELEM_NUM; ++i) {
-    reg.val[i] = float_to_bf16(v.reg.val[i]);
-  }
-}
-
-inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
-  int i = 0;
-  for (i = 0; i < BF16Vec16::VEC_ELEM_NUM; ++i) {
-    reg.val[i] = float_to_bf16(v.reg.val[i]);
-  }
-}
-
-inline void prefetch(const void* addr) { __builtin_prefetch(addr, 0, 3); }
-
-};  // namespace vec_op

csrc/cpu/cpu_types_vxe.hpp

@@ -12,7 +12,7 @@ namespace vec_op {
 #define vec_sub(a, b) ((a) - (b))
 #define vec_mul(a, b) ((a) * (b))
 #define vec_div(a, b) ((a) / (b))
-#define vec_sr(a, b) ((a) >> (b))  // Vector Shift Right Algebraic
+#define vec_sr(a, b) ((a) >> (b))  // Vector Shift Right Algebaic
 #define vec_sl(a, b) ((a) << (b))  // Vector Shift Left
 
 // FIXME: FP16 is not fully supported in Torch-CPU

csrc/cpu/dnnl_kernels.cpp

@@ -523,7 +523,7 @@ void onednn_mm(torch::Tensor& c,        // [M, OC], row-major
   CPU_KERNEL_GUARD_IN(onednn_mm)
   TORCH_CHECK(a.dim() == 2);
   TORCH_CHECK(a.stride(-1) == 1);
-  TORCH_CHECK(c.stride(-1) == 1);
+  TORCH_CHECK(c.is_contiguous());
   MatMulPrimitiveHandler* ptr =
       reinterpret_cast<MatMulPrimitiveHandler*>(handler);
 

csrc/cpu/float_convert.hpp

@@ -1,106 +0,0 @@
-
-static float bf16_to_float(uint16_t bf16) {
-  uint32_t bits = static_cast<uint32_t>(bf16) << 16;
-  float fp32;
-  std::memcpy(&fp32, &bits, sizeof(fp32));
-  return fp32;
-}
-
-static uint16_t float_to_bf16(float fp32) {
-  uint32_t bits;
-  std::memcpy(&bits, &fp32, sizeof(fp32));
-  return static_cast<uint16_t>(bits >> 16);
-}
-
-/************************************************
- * Copyright (c) 2015 Princeton Vision Group
- * Licensed under the MIT license.
- * Codes below copied from
- * https://github.com/PrincetonVision/marvin/tree/master/tools/tensorIO_matlab
- *************************************************/
-static uint16_t float_to_fp16(float fp32) {
-  uint16_t fp16;
-
-  unsigned x;
-  unsigned u, remainder, shift, lsb, lsb_s1, lsb_m1;
-  unsigned sign, exponent, mantissa;
-
-  std::memcpy(&x, &fp32, sizeof(fp32));
-  u = (x & 0x7fffffff);
-
-  // Get rid of +NaN/-NaN case first.
-  if (u > 0x7f800000) {
-    fp16 = 0x7fffU;
-    return fp16;
-  }
-
-  sign = ((x >> 16) & 0x8000);
-
-  // Get rid of +Inf/-Inf, +0/-0.
-  if (u > 0x477fefff) {
-    fp16 = sign | 0x7c00U;
-    return fp16;
-  }
-  if (u < 0x33000001) {
-    fp16 = (sign | 0x0000);
-    return fp16;
-  }
-
-  exponent = ((u >> 23) & 0xff);
-  mantissa = (u & 0x7fffff);
-
-  if (exponent > 0x70) {
-    shift = 13;
-    exponent -= 0x70;
-  } else {
-    shift = 0x7e - exponent;
-    exponent = 0;
-    mantissa |= 0x800000;
-  }
-  lsb = (1 << shift);
-  lsb_s1 = (lsb >> 1);
-  lsb_m1 = (lsb - 1);
-
-  // Round to nearest even.
-  remainder = (mantissa & lsb_m1);
-  mantissa >>= shift;
-  if (remainder > lsb_s1 || (remainder == lsb_s1 && (mantissa & 0x1))) {
-    ++mantissa;
-    if (!(mantissa & 0x3ff)) {
-      ++exponent;
-      mantissa = 0;
-    }
-  }
-
-  fp16 = (sign | (exponent << 10) | mantissa);
-
-  return fp16;
-}
-
-static float fp16_to_float(uint16_t fp16) {
-  unsigned sign = ((fp16 >> 15) & 1);
-  unsigned exponent = ((fp16 >> 10) & 0x1f);
-  unsigned mantissa = ((fp16 & 0x3ff) << 13);
-  int temp;
-  float fp32;
-  if (exponent == 0x1f) { /* NaN or Inf */
-    mantissa = (mantissa ? (sign = 0, 0x7fffff) : 0);
-    exponent = 0xff;
-  } else if (!exponent) { /* Denorm or Zero */
-    if (mantissa) {
-      unsigned int msb;
-      exponent = 0x71;
-      do {
-        msb = (mantissa & 0x400000);
-        mantissa <<= 1; /* normalize */
-        --exponent;
-      } while (!msb);
-      mantissa &= 0x7fffff; /* 1.mantissa is implicit */
-    }
-  } else {
-    exponent += 0x70;
-  }
-  temp = ((sign << 31) | (exponent << 23) | mantissa);
-  std::memcpy(&fp32, &temp, sizeof(temp));
-  return fp32;
-}

csrc/cpu/sgl-kernels/moe.cpp

@@ -215,7 +215,7 @@ int moe_align_block_size(
       offsets[mb + 1] = sorted_id_size(sorted_ids + mb * BLOCK_M);
     }
   });
-  // TODO: do we need to vectorize this ?
+  // TODO: do we need to vecterize this ?
   for (int mb = 0; mb < num_token_blocks; ++mb) {
     offsets[mb + 1] += offsets[mb];
   }

csrc/cpu/torch_bindings.cpp

@@ -88,18 +88,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "    int tp_rank, int blocksparse_local_blocks,"
       "    int blocksparse_vert_stride, int blocksparse_block_size,"
       "    int blocksparse_head_sliding_step) -> ()");
-
   ops.impl("paged_attention_v1", torch::kCPU, &paged_attention_v1);
 
-  ops.def(
-      "dynamic_4bit_int_moe("
-      "Tensor x, Tensor topk_ids, Tensor topk_weights,"
-      "Tensor w13_packed, Tensor w2_packed, int H, int I, int I2,"
-      "int group_size, bool apply_router_weight_on_input, int activation_kind"
-      ") -> Tensor");
-
-  ops.impl("dynamic_4bit_int_moe", torch::kCPU, &dynamic_4bit_int_moe_cpu);
-
   // PagedAttention V2.
   ops.def(
       "paged_attention_v2("

csrc/cub_helpers.h

@@ -1,17 +0,0 @@
-#pragma once
-
-#ifndef USE_ROCM
-  #include <cub/cub.cuh>
-  #if CUB_VERSION >= 200800
-    #include <cuda/std/functional>
-using CubAddOp = cuda::std::plus<>;
-using CubMaxOp = cuda::maximum<>;
-  #else   // if CUB_VERSION < 200800
-using CubAddOp = cub::Sum;
-using CubMaxOp = cub::Max;
-  #endif  // CUB_VERSION
-#else
-  #include <hipcub/hipcub.hpp>
-using CubAddOp = cub::Sum;
-using CubMaxOp = cub::Max;
-#endif  // USE_ROCM

csrc/custom_all_reduce.cuh

@@ -15,8 +15,6 @@ typedef __hip_bfloat16 nv_bfloat16;
 #include <map>
 #include <unordered_map>
 #include <vector>
-#include <cstdlib>
-#include <cstring>
 
 namespace vllm {
 #define CUDACHECK(cmd)                                              \
@@ -557,47 +555,22 @@ class CustomAllreduce {
     size /= d;
     auto bytes = size * sizeof(typename packed_t<T>::P);
     int blocks = std::min(block_limit, (size + threads - 1) / threads);
-
-    // Check environment variable once
-    const char* env_algo = std::getenv("VLLM_CUSTOM_ALLREDUCE_ALGO");
-    bool force_1stage = false;
-    bool force_2stage = false;
-    if (env_algo != nullptr) {
-      if (std::strcmp(env_algo, "1stage") == 0 ||
-          std::strcmp(env_algo, "oneshot") == 0) {
-        force_1stage = true;
-      } else if (std::strcmp(env_algo, "2stage") == 0 ||
-                 std::strcmp(env_algo, "twoshot") == 0) {
-        force_2stage = true;
-      } else {
-        throw std::runtime_error(
-            "Invalid VLLM_CUSTOM_ALLREDUCE_ALGO: " + std::string(env_algo) +
-            ". Valid values: 1stage, oneshot, 2stage, twoshot");
-      }
-    }
-
 #define KL(ngpus, name)                                                       \
   name<T, ngpus><<<blocks, threads, 0, stream>>>(ptrs, sg_, self_sg_, output, \
                                                  rank_, size);
-#define REDUCE_CASE(ngpus)                              \
-  case ngpus: {                                         \
-    if (force_1stage) {                                 \
-      KL(ngpus, cross_device_reduce_1stage);            \
-    } else if (force_2stage) {                          \
-      KL(ngpus, cross_device_reduce_2stage);            \
-    } else {                                            \
-      if (world_size_ == 2) {                           \
-        KL(ngpus, cross_device_reduce_1stage);          \
-      } else if (fully_connected_) {                    \
-        if ((world_size_ <= 4 && bytes < 512 * 1024) || \
-            (world_size_ <= 8 && bytes < 256 * 1024)) { \
-          KL(ngpus, cross_device_reduce_1stage);        \
-        } else {                                        \
-          KL(ngpus, cross_device_reduce_2stage);        \
-        }                                               \
-      }                                                 \
-    }                                                   \
-    break;                                              \
+#define REDUCE_CASE(ngpus)                            \
+  case ngpus: {                                       \
+    if (world_size_ == 2) {                           \
+      KL(ngpus, cross_device_reduce_1stage);          \
+    } else if (fully_connected_) {                    \
+      if ((world_size_ <= 4 && bytes < 512 * 1024) || \
+          (world_size_ <= 8 && bytes < 256 * 1024)) { \
+        KL(ngpus, cross_device_reduce_1stage);        \
+      } else {                                        \
+        KL(ngpus, cross_device_reduce_2stage);        \
+      }                                               \
+    }                                                 \
+    break;                                            \
   }
 
     switch (world_size_) {

csrc/cutlass_extensions/gemm/collective/collective_builder.hpp

@@ -0,0 +1,123 @@
+// Modified from: cutlass/gemm/collective/builders/sm90_gmma_builder.inl
+// clang-format off
+#pragma once
+
+#include "cutlass/gemm/collective/builders/sm90_gmma_builder.inl"
+
+#include "cutlass_extensions/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp"
+
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+namespace cutlass::gemm::collective {
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+// GMMA_TMA_WS_SS (BlockScaled Builders)
+template <
+  class ElementA,
+  class GmemLayoutATag,
+  int AlignmentA,
+  class ElementB,
+  class GmemLayoutBTag,
+  int AlignmentB,
+  class ElementAccumulator,
+  class TileShape_MNK,
+  class ClusterShape_MNK,
+  class StageCountType,
+  int ScaleGranularityM
+>
+struct CollectiveBuilder<
+    arch::Sm90,
+    arch::OpClassTensorOp,
+    ElementA,
+    GmemLayoutATag,
+    AlignmentA,
+    ElementB,
+    GmemLayoutBTag,
+    AlignmentB,
+    ElementAccumulator,
+    TileShape_MNK,
+    ClusterShape_MNK,
+    StageCountType,
+    KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum<ScaleGranularityM>,
+    cute::enable_if_t<
+      not detail::is_use_rmem_A<ElementA, GmemLayoutATag, ElementB, GmemLayoutBTag>()>
+> {
+  using KernelScheduleType = KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum<ScaleGranularityM>;
+
+  static_assert(is_static<TileShape_MNK>::value);
+  static_assert(is_static<ClusterShape_MNK>::value);
+#ifndef CUTLASS_SM90_COLLECTIVE_BUILDER_SUPPORTED
+  static_assert(cutlass::detail::dependent_false<ElementA>, "Unsupported Toolkit for SM90 Collective Builder\n");
+#endif
+  static_assert(detail::is_aligned<ElementA, AlignmentA, ElementB, AlignmentB, detail::tma_alignment_bytes>(),
+                "Should meet TMA alignment requirement\n");
+
+  static constexpr bool IsArrayOfPointersGemm = (cute::is_any_of_v<KernelScheduleType,
+                                                                   KernelPtrArrayTmaWarpSpecializedCooperative,
+                                                                   KernelPtrArrayTmaWarpSpecializedPingpong>);
+  static constexpr bool IsFP8Input = detail::is_input_fp8<ElementA, ElementB>();
+  static_assert((!IsFP8Input || !IsArrayOfPointersGemm),
+                "KernelTmaWarpSpecializedCooperativeFP8BlockScaledAccum is only compatible with FP8 Blocked Scaled version right now.");
+
+  // For fp32 types, map to tf32 MMA value type
+  using ElementAMma = cute::conditional_t<cute::is_same_v<ElementA, float>, tfloat32_t, ElementA>;
+  using ElementBMma = cute::conditional_t<cute::is_same_v<ElementB, float>, tfloat32_t, ElementB>;
+
+  static constexpr cute::GMMA::Major GmmaMajorA = detail::gmma_ss_tag_to_major_A<ElementAMma, GmemLayoutATag>();
+  static constexpr cute::GMMA::Major GmmaMajorB = detail::gmma_ss_tag_to_major_B<ElementBMma, GmemLayoutBTag>();
+
+  static constexpr bool IsCooperative = cute::is_any_of_v<KernelScheduleType,
+                                                          KernelTmaWarpSpecializedCooperative,
+                                                          KernelPtrArrayTmaWarpSpecializedCooperative,
+                                                          KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum<ScaleGranularityM>>;
+  using AtomLayoutMNK = cute::conditional_t<IsCooperative,
+      Layout<Shape<_2,_1,_1>>, Layout<Shape<_1,_1,_1>>>;
+
+  using TiledMma = decltype(cute::make_tiled_mma(cute::GMMA::ss_op_selector<
+      ElementAMma, ElementBMma, ElementAccumulator, TileShape_MNK, GmmaMajorA, GmmaMajorB>(), AtomLayoutMNK{}));
+
+  using GmemTiledCopyA = decltype(detail::sm90_cluster_shape_to_tma_atom(shape<1>(ClusterShape_MNK{})));
+  using GmemTiledCopyB = decltype(detail::sm90_cluster_shape_to_tma_atom(shape<0>(ClusterShape_MNK{})));
+
+  using SmemLayoutAtomA = decltype(detail::ss_smem_selector<
+      GmmaMajorA, ElementAMma, decltype(cute::get<0>(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>());
+  using SmemLayoutAtomB = decltype(detail::ss_smem_selector<
+      GmmaMajorB, ElementBMma, decltype(cute::get<1>(TileShape_MNK{})), decltype(cute::get<2>(TileShape_MNK{}))>());
+
+  static constexpr size_t TensorMapStorage = IsArrayOfPointersGemm ? sizeof(cute::TmaDescriptor) * 2 /* for A and B */ : 0;
+  static constexpr int KernelSmemCarveout = static_cast<int>(TensorMapStorage);
+
+  static constexpr int PipelineStages = detail::compute_stage_count_or_override<detail::sm90_smem_capacity_bytes - KernelSmemCarveout,
+      ElementAMma, ElementBMma, TileShape_MNK>(StageCountType{});
+  using DispatchPolicy = MainloopSm90TmaGmmaWarpSpecializedBlockScalingSubGroupMFP8<PipelineStages, ClusterShape_MNK, KernelScheduleType, ScaleGranularityM>;
+
+  using SmemCopyAtomA = void;
+  using SmemCopyAtomB = void;
+
+  using CollectiveOp = CollectiveMma<
+      DispatchPolicy,
+      TileShape_MNK,
+      ElementA,
+      TagToStrideA_t<GmemLayoutATag>,
+      ElementB,
+      TagToStrideB_t<GmemLayoutBTag>,
+      TiledMma,
+      GmemTiledCopyA,
+      SmemLayoutAtomA,
+      SmemCopyAtomA,
+      cute::identity,
+      GmemTiledCopyB,
+      SmemLayoutAtomB,
+      SmemCopyAtomB,
+      cute::identity
+    >;
+};
+
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass::gemm::collective
+
+/////////////////////////////////////////////////////////////////////////////////////////////////

csrc/cutlass_extensions/gemm/collective/fp8_accumulation.hpp

@@ -0,0 +1,183 @@
+// clang-format off
+// adapted from: https://github.com/soundOfDestiny/cutlass/blob/a4208aa6958864923505cade9c63eb2a6daf16e5/include/cutlass/gemm/collective/fp8_accumulation.hpp
+
+/***************************************************************************************************
+ * Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice, this
+ * list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ * this list of conditions and the following disclaimer in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ * 3. Neither the name of the copyright holder nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+ * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+ * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+#pragma once
+
+#include "cute/algorithm/clear.hpp"
+#include "cute/tensor.hpp"
+
+//////////////////////////////////////////////////////////////////////////////
+///////////////////////////////////FP8 Accumulation///////////////////////////
+//////////////////////////////////////////////////////////////////////////////
+/// This class provides API to promote (add) or scale (multiply_add) the results
+/// from the tensor core accumulators to the main accumulators when the number 
+/// of MMAs reaches the max number of MMA interval specified by user, after that
+/// the tensor core accumulators are zeroed.
+//////////////////////////////////////////////////////////////////////////////
+
+namespace cutlass::gemm::collective {
+
+template <
+    class EngineAccum,
+    class LayoutAccum>
+struct GmmaFP8AccumulationWithScale {  
+  using TensorAccum = cute::Tensor<EngineAccum, LayoutAccum>;
+  using ElementAccumulator = typename EngineAccum::value_type;
+
+  static_assert(is_static<LayoutAccum>::value, "Accumulator Layout should be static");
+  static_assert(is_rmem<TensorAccum>::value , "Accumulator tensor must be rmem resident.");
+
+private:
+  TensorAccum& accum_;
+  TensorAccum accum_temp_;
+
+  uint32_t accum_promotion_interval_;         // defines the max num of executed MMAs after which accum should be promoted.
+  uint32_t mma_count_per_mainloop_iteration_; // num of MMAs per k_tile of mainloop
+  uint32_t mma_count_;                        // current executed MMAs
+  uint32_t reset_accum_flag_;                 // accum needs to be zeroed or not. 
+
+  // promote or `add` the partial accumulators to main accumulator (FADD).
+  CUTLASS_DEVICE
+  void promote_core() {
+    warpgroup_wait<0>();
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0; i < size(accum_); ++i) {
+      accum_(i) += accum_temp_(i);
+    }
+  }
+
+  // `multiply` scale the partial accumulators and `add` to main accumulator (FFMA).
+  template <
+    class EngineScale,
+    class LayoutScale>
+  CUTLASS_DEVICE
+  void scale_core(const cute::Tensor<EngineScale, LayoutScale> &scale) {
+    using TensorScale = cute::Tensor<EngineScale, LayoutScale>;
+
+    static_assert(is_static<LayoutScale>::value, "Scale Layout should be static");
+    static_assert(is_rmem<TensorScale>::value , "Scale tensor must be rmem resident.");
+
+    static_assert(LayoutAccum{}.shape() == LayoutScale{}.shape(), "Accumulator and scale must have same shape.");
+
+    warpgroup_wait<0>();
+    CUTLASS_PRAGMA_UNROLL
+    for (int i = 0; i < size(accum_); ++i) {
+      accum_(i) += accum_temp_(i) * scale(i);
+    }
+  }
+
+public:
+  CUTLASS_DEVICE
+  GmmaFP8AccumulationWithScale(
+      TensorAccum &accum,
+      uint32_t accum_promotion_interval,
+      uint32_t mma_count_per_mainloop_iteration)
+      : accum_(accum), 
+        accum_promotion_interval_(accum_promotion_interval),
+        mma_count_per_mainloop_iteration_(mma_count_per_mainloop_iteration),
+        mma_count_(0), 
+        reset_accum_flag_(0) 
+  {
+    accum_temp_ = cute::make_fragment_like(accum);
+  }
+
+  //
+  // Methods (Common)
+  //
+
+  CUTLASS_DEVICE 
+  TensorAccum& operator()() {
+    return accum_temp_;
+  }
+
+  /// prepare the MMA accumulators when initialization or zeroing is required.
+  CUTLASS_DEVICE
+  bool prepare_if_needed() { 
+    return reset_accum_flag_;
+  }
+
+  //
+  // Methods (for FADD version)
+  //
+
+  /// promote (add) the results from the MMA accumulators to main accumulator if needed.
+  CUTLASS_DEVICE
+  void promote_if_needed() {
+    mma_count_ += mma_count_per_mainloop_iteration_;
+    reset_accum_flag_ = __shfl_sync(0xffffffff, mma_count_ == accum_promotion_interval_, 0);
+    if (reset_accum_flag_) {
+      promote_core();
+      mma_count_ = 0;
+    }
+  }
+
+  /// promote (add) the residue results from the MMA accumulators to main accumulator if needed.
+  CUTLASS_DEVICE
+  void promote_residue_if_needed() {
+    if (__shfl_sync(0xffffffff, mma_count_ > 0, 0)) {
+      promote_core();
+    }
+  }
+
+  //
+  // Methods (for FFMA version)
+  //
+
+  /// scale (multiply_add) the results from the MMA accumulators to main accumulator if needed.
+  template <
+    class EngineScale,
+    class LayoutScale>
+  CUTLASS_DEVICE
+  void scale_if_needed(const cute::Tensor<EngineScale, LayoutScale> &scale) {
+    mma_count_ += mma_count_per_mainloop_iteration_;
+    reset_accum_flag_ = __shfl_sync(0xffffffff, mma_count_ == accum_promotion_interval_, 0);
+    if (reset_accum_flag_) {
+      scale_core(scale);
+      mma_count_ = 0;
+    }
+  }
+
+  /// scale (multiply_add) the residue results from the MMA accumulators to main accumulator if needed.
+  template <
+    class EngineScale,
+    class LayoutScale>
+  CUTLASS_DEVICE
+  void scale_residue_if_needed(const cute::Tensor<EngineScale, LayoutScale> &scale) {
+    if (__shfl_sync(0xffffffff, mma_count_ > 0, 0)) {
+      scale_core(scale);
+    }
+  }
+};
+
+} // namespace cutlass::gemm::collective

csrc/cutlass_extensions/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp

@@ -0,0 +1,729 @@
+// clang-format off
+// Adapted (Heavily) from: https://github.com/soundOfDestiny/cutlass/blob/9d997ce0dea4c5fa1a617db6b7ff29aa9235822c/include/cutlass/gemm/collective/sm90_mma_tma_gmma_ss_warpspecialized_fp8_blockwise_scaling.hpp
+
+/***************************************************************************************************
+ * Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: BSD-3-Clause
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice, this
+ * list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ * this list of conditions and the following disclaimer in the documentation
+ * and/or other materials provided with the distribution.
+ *
+ * 3. Neither the name of the copyright holder nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+ * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+ * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+ * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+ * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+ * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+ * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *
+ **************************************************************************************************/
+
+#pragma once
+
+#include "cutlass/cutlass.h"
+#include "cutlass/gemm/dispatch_policy.hpp"
+#include "cutlass/trace.h"
+#include "cutlass/numeric_types.h"
+
+#include "cute/arch/cluster_sm90.hpp"
+#include "cute/arch/copy_sm80.hpp"
+#include "cute/arch/copy_sm90.hpp"
+#include "cute/algorithm/functional.hpp"
+#include "cute/atom/mma_atom.hpp"
+#include "cute/algorithm/gemm.hpp"
+#include "cute/numeric/arithmetic_tuple.hpp"
+
+#include "cutlass_extensions/gemm/dispatch_policy.hpp"
+#include "cutlass_extensions/gemm/collective/fp8_accumulation.hpp"
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+namespace cutlass::gemm::collective {
+using namespace cute;
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+// WarpSpecialized Mainloop
+template <
+  int Stages,
+  class ClusterShape,
+  class KernelSchedule,
+  int ScaleGranularityM_,
+  class TileShape_,
+  class ElementA_,
+  class StrideA_,
+  class ElementB_,
+  class StrideB_,
+  class TiledMma_,
+  class GmemTiledCopyA_,
+  class SmemLayoutAtomA_,
+  class SmemCopyAtomA_,
+  class TransformA_,
+  class GmemTiledCopyB_,
+  class SmemLayoutAtomB_,
+  class SmemCopyAtomB_,
+  class TransformB_>
+struct CollectiveMma<
+    MainloopSm90TmaGmmaWarpSpecializedBlockScalingSubGroupMFP8<Stages, ClusterShape, KernelSchedule, ScaleGranularityM_>,
+    TileShape_,
+    ElementA_,
+    StrideA_,
+    ElementB_,
+    StrideB_,
+    TiledMma_,
+    GmemTiledCopyA_,
+    SmemLayoutAtomA_,
+    SmemCopyAtomA_,
+    TransformA_,
+    GmemTiledCopyB_,
+    SmemLayoutAtomB_,
+    SmemCopyAtomB_,
+    TransformB_>
+{
+  //
+  // Type Aliases
+  //
+  using DispatchPolicy = MainloopSm90TmaGmmaWarpSpecializedBlockScalingSubGroupMFP8<Stages, ClusterShape, KernelSchedule, ScaleGranularityM_>;
+  using TileShape = TileShape_;
+  using ElementA = ElementA_;
+  using StrideA = StrideA_;
+  using ElementB = ElementB_;
+  using StrideB = StrideB_;
+  using TiledMma = TiledMma_;
+  using ElementAccumulator = typename TiledMma::ValTypeC;
+  using ElementBlockScale = ElementAccumulator;
+  using GmemTiledCopyA = GmemTiledCopyA_;
+  using GmemTiledCopyB = GmemTiledCopyB_;
+  using SmemLayoutAtomA = SmemLayoutAtomA_;
+  using SmemLayoutAtomB = SmemLayoutAtomB_;
+  using SmemCopyAtomA = SmemCopyAtomA_;
+  using SmemCopyAtomB = SmemCopyAtomB_;
+  using TransformA = TransformA_;
+  using TransformB = TransformB_;
+  using ArchTag = typename DispatchPolicy::ArchTag;
+
+  using CtaShape_MNK = decltype(shape_div(TileShape{}, ClusterShape{}));
+  using MainloopPipeline = cutlass::PipelineTmaAsync<DispatchPolicy::Stages>;
+  using PipelineState = cutlass::PipelineState<DispatchPolicy::Stages>;
+  using PipelineParams = typename MainloopPipeline::Params;
+
+  // Two threads per CTA are producers (1 for operand tile and 32 for scales)
+  static constexpr int NumProducerThreadEvents = 33; 
+
+  static constexpr int ScaleGranularityM = ScaleGranularityM_ == 0 ? size<0>(TileShape{}) : ScaleGranularityM_;
+  static constexpr int ScaleMsPerTile = size<0>(TileShape{}) / ScaleGranularityM;
+
+  static_assert(cute::rank(SmemLayoutAtomA{}) == 2, "SmemLayoutAtom must be rank 2 (M/N, K)");
+  static_assert((size<0>(TileShape{}) % size<0>(SmemLayoutAtomA{})) == 0, "SmemLayoutAtom must evenly divide tile shape.");
+  static_assert((size<2>(TileShape{}) % size<1>(SmemLayoutAtomA{})) == 0, "SmemLayoutAtom must evenly divide tile shape.");
+
+  static_assert(cute::rank(SmemLayoutAtomB{}) == 2, "SmemLayoutAtom must be rank 2 (M/N, K)");
+  static_assert((size<1>(TileShape{}) % size<0>(SmemLayoutAtomB{})) == 0, "SmemLayoutAtom must evenly divide tile shape.");
+  static_assert((size<2>(TileShape{}) % size<1>(SmemLayoutAtomB{})) == 0, "SmemLayoutAtom must evenly divide tile shape.");
+
+  static_assert((size<0>(TileShape{}) % ScaleGranularityM) == 0, "FP8 scaling granularity must evenly divide tile shape along M.");
+
+  // Tile along modes in a way that maximizes the TMA box size.
+  using SmemLayoutA = decltype(tile_to_shape(
+      SmemLayoutAtomA{},
+      make_shape(shape<0>(TileShape{}), shape<2>(TileShape{}), Int<DispatchPolicy::Stages>{}),
+      cute::conditional_t< ::cutlass::gemm::detail::is_major<0,StrideA>(), Step<_2,_1,_3>, Step<_1,_2,_3>>{}));
+  using SmemLayoutB = decltype(tile_to_shape(
+      SmemLayoutAtomB{},
+      make_shape(shape<1>(TileShape{}), shape<2>(TileShape{}), Int<DispatchPolicy::Stages>{}),
+      cute::conditional_t< ::cutlass::gemm::detail::is_major<0,StrideB>(), Step<_2,_1,_3>, Step<_1,_2,_3>>{}));
+  
+  // Block scaling gmem-to-smem copy atom 
+  using SmemBlockScalingCopyAtomA = Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<ElementBlockScale>, ElementBlockScale>;
+  using SmemBlockScalingCopyAtomB = Copy_Atom<SM80_CP_ASYNC_CACHEALWAYS<ElementBlockScale>, ElementBlockScale>;
+  
+  // Block scaling smem layout
+  using SmemLayoutScaleA = Layout<Shape<Int<ScaleMsPerTile>, Int<DispatchPolicy::Stages>>>;
+  using SmemLayoutScaleB = Layout<Shape<Int<DispatchPolicy::Stages>>, Stride<_1>>; // `ScaleNsPerTile` is always 1.
+
+  static_assert(DispatchPolicy::Stages >= 2, "Specialization requires Stages set to value 1 or more.");
+  static_assert(cute::is_base_of<cute::GMMA::DescriptorIterator, typename TiledMma::FrgTypeA>::value &&
+                cute::is_base_of<cute::GMMA::DescriptorIterator, typename TiledMma::FrgTypeB>::value,
+                "MMA atom must source both A and B operand from smem_desc for this mainloop.");
+  static_assert(cute::is_same_v<GmemTiledCopyA, SM90_TMA_LOAD> || cute::is_same_v<GmemTiledCopyA, SM90_TMA_LOAD_MULTICAST>,
+      "GmemTiledCopy - invalid SM90 TMA copy atom specified.");
+  static_assert(cute::is_same_v<GmemTiledCopyB, SM90_TMA_LOAD> || cute::is_same_v<GmemTiledCopyB, SM90_TMA_LOAD_MULTICAST>,
+      "GmemTiledCopy - invalid SM90 TMA copy atom specified.");
+  static_assert(cute::is_same_v<ElementAccumulator, ElementBlockScale>,
+             "ElementAccumulator and ElementBlockScale should be same datatype");
+
+  struct SharedStorage
+  {
+    struct TensorStorage : cute::aligned_struct<128> {
+      cute::array_aligned<typename TiledMma::ValTypeA, cute::cosize_v<SmemLayoutA>> smem_A;  // mxk
+      cute::array_aligned<typename TiledMma::ValTypeB, cute::cosize_v<SmemLayoutB>> smem_B;  // nxk
+      cute::array_aligned<ElementBlockScale, cute::cosize_v<SmemLayoutScaleA>> smem_scale_A; // ScaleMsPerTile x k
+      cute::array_aligned<ElementBlockScale, cute::cosize_v<SmemLayoutScaleB>> smem_scale_B; // 1xk
+    } tensors;
+
+    using PipelineStorage = typename MainloopPipeline::SharedStorage;
+    PipelineStorage pipeline;
+  };
+  using TensorStorage = typename SharedStorage::TensorStorage;
+  using PipelineStorage = typename SharedStorage::PipelineStorage;
+
+  // Host side kernel arguments
+  struct Arguments {
+    ElementA const* ptr_A;
+    StrideA dA;
+    ElementB const* ptr_B;
+    StrideB dB;
+    ElementBlockScale const* ptr_scale_A; 
+    ElementBlockScale const* ptr_scale_B;
+  };
+
+  // Device side kernel params
+  struct Params {
+    // Assumption: StrideA is congruent with Problem_MK
+    using TMA_A = decltype(make_tma_copy_A_sm90(
+        GmemTiledCopyA{},
+        make_tensor(static_cast<ElementA const*>(nullptr), repeat_like(StrideA{}, int32_t(0)), StrideA{}),
+        SmemLayoutA{}(_,_,0),
+        TileShape{},
+        ClusterShape{}));
+    // Assumption: StrideB is congruent with Problem_NK
+    using TMA_B = decltype(make_tma_copy_B_sm90(
+        GmemTiledCopyB{},
+        make_tensor(static_cast<ElementB const*>(nullptr), repeat_like(StrideB{}, int32_t(0)), StrideB{}),
+        SmemLayoutB{}(_,_,0),
+        TileShape{},
+        ClusterShape{}));
+    TMA_A tma_load_a;
+    TMA_B tma_load_b;
+    uint32_t tma_transaction_bytes = TmaTransactionBytes;
+    uint32_t tma_transaction_bytes_mk = TmaTransactionBytesMK;
+    uint32_t tma_transaction_bytes_nk = TmaTransactionBytesNK;
+    // Block scaling factors for A and B
+    ElementBlockScale const* ptr_scale_A; 
+    ElementBlockScale const* ptr_scale_B;
+  };
+
+  //
+  // Methods
+  //
+
+  template <class ProblemShape>
+  static constexpr Params
+  to_underlying_arguments(ProblemShape const& problem_shape, Arguments const& args, void* workspace) {
+    (void) workspace;
+
+    // Optionally append 1s until problem shape is rank-4 (MNKL), in case it is only rank-3 (MNK)
+    auto problem_shape_MNKL = append<4>(problem_shape, 1);
+    auto [M,N,K,L] = problem_shape_MNKL;
+
+    auto ptr_A = reinterpret_cast<ElementA const*>(args.ptr_A);
+    auto ptr_B = reinterpret_cast<ElementB const*>(args.ptr_B);
+
+    Tensor tensor_a = make_tensor(ptr_A, make_layout(make_shape(M,K,L), args.dA));
+    Tensor tensor_b = make_tensor(ptr_B, make_layout(make_shape(N,K,L), args.dB));
+    typename Params::TMA_A tma_load_a = make_tma_copy_A_sm90(
+        GmemTiledCopyA{},
+        tensor_a,
+        SmemLayoutA{}(_,_,cute::Int<0>{}),
+        TileShape{},
+        ClusterShape{});
+    typename Params::TMA_B tma_load_b = make_tma_copy_B_sm90(
+        GmemTiledCopyB{},
+        tensor_b,
+        SmemLayoutB{}(_,_,cute::Int<0>{}),
+        TileShape{},
+        ClusterShape{});
+    uint32_t transaction_bytes_mk = TmaTransactionBytesMK;
+    uint32_t transaction_bytes_nk = TmaTransactionBytesNK;
+    uint32_t transaction_bytes = transaction_bytes_mk + transaction_bytes_nk;
+
+    return {
+      tma_load_a,
+      tma_load_b,
+      transaction_bytes,
+      transaction_bytes_mk,
+      transaction_bytes_nk,
+      args.ptr_scale_A,
+      args.ptr_scale_B
+    };
+  }
+
+  template<class ProblemShape>
+  static bool
+  can_implement(
+      ProblemShape const& problem_shape,
+      [[maybe_unused]] Arguments const& args) {
+    constexpr int tma_alignment_bits = 128;
+    auto problem_shape_MNKL = append<4>(problem_shape, 1);
+    auto [M,N,K,L] = problem_shape_MNKL;
+    
+    bool implementable = true;
+    constexpr int min_tma_aligned_elements_A = tma_alignment_bits / cutlass::sizeof_bits<ElementA>::value;
+    implementable = implementable && cutlass::detail::check_alignment<min_tma_aligned_elements_A>(cute::make_shape(M,K,L), StrideA{});
+    constexpr int min_tma_aligned_elements_B = tma_alignment_bits / cutlass::sizeof_bits<ElementB>::value;
+    implementable = implementable && cutlass::detail::check_alignment<min_tma_aligned_elements_B>(cute::make_shape(N,K,L), StrideB{});
+
+    if (!implementable) {
+      CUTLASS_TRACE_HOST("  CAN IMPLEMENT: Problem Size doesn't meet the minimum alignment requirements for TMA.\n");
+    }
+    return implementable;
+  }
+
+  static constexpr int K_PIPE_MAX = DispatchPolicy::Stages;
+  static constexpr int K_PIPE_MMAS = 1;
+  static constexpr uint32_t TmaTransactionBytesMK =
+        cutlass::bits_to_bytes(size<0>(SmemLayoutA{}) * size<1>(SmemLayoutA{}) * static_cast<uint32_t>(sizeof_bits<ElementA>::value));
+  static constexpr uint32_t TmaTransactionBytesNK =
+        cutlass::bits_to_bytes(size<0>(SmemLayoutB{}) * size<1>(SmemLayoutB{}) * static_cast<uint32_t>(sizeof_bits<ElementB>::value));
+  static constexpr uint32_t TmaTransactionBytes = TmaTransactionBytesMK + TmaTransactionBytesNK;
+
+  /// Issue Tma Descriptor Prefetch -- ideally from a single thread for best performance
+  CUTLASS_DEVICE
+  static void prefetch_tma_descriptors(Params const& mainloop_params)
+  {
+    cute::prefetch_tma_descriptor(mainloop_params.tma_load_a.get_tma_descriptor());
+    cute::prefetch_tma_descriptor(mainloop_params.tma_load_b.get_tma_descriptor());
+  }
+
+  /// Set up the data needed by this collective for load and mma.
+  /// Returns a tuple of tensors. The collective and the kernel layer have the contract
+  /// Returned tuple must contain at least two elements, with the first two elements being:
+  /// gA_mkl - The tma tensor, A after a local tile so it has shape  (BLK_M,BLK_K,m,k,l)
+  /// gB_nkl - The tma tensor, B after a local tile so it has shape  (BLK_N,BLK_K,n,k,l)
+  template <class ProblemShape_MNKL>
+  CUTLASS_DEVICE auto
+  load_init(ProblemShape_MNKL const& problem_shape_MNKL, Params const& mainloop_params) const {
+    using X = Underscore;
+    // Separate out problem shape for convenience
+    auto [M,N,K,L] = problem_shape_MNKL;
+
+    // TMA requires special handling of strides to deal with coord codomain mapping
+    // Represent the full tensors -- get these from TMA
+    Tensor mA_mkl = mainloop_params.tma_load_a.get_tma_tensor(make_shape(M,K,L));                            // (m,k,l)
+    Tensor mB_nkl = mainloop_params.tma_load_b.get_tma_tensor(make_shape(N,K,L));                            // (n,k,l)
+
+    // Make tiled views, defer the slice
+    Tensor gA_mkl = local_tile(mA_mkl, TileShape{}, make_coord(_,_,_), Step<_1, X,_1>{});        // (BLK_M,BLK_K,m,k,l)
+    Tensor gB_nkl = local_tile(mB_nkl, TileShape{}, make_coord(_,_,_), Step< X,_1,_1>{});        // (BLK_N,BLK_K,n,k,l)
+
+    constexpr auto scales_m = Int<ScaleMsPerTile>{};
+    auto tM = get<2>(gA_mkl.shape());
+    auto tN = get<2>(gB_nkl.shape());
+    auto tK = get<3>(gA_mkl.shape());
+
+    // Make the tiled views of scale tensors
+    auto scaleA_shape = make_shape(M / ScaleGranularityM, tK, L); // (scale_m,k,l)
+    auto scaleA_layout = make_ordered_layout(scaleA_shape,  Step<_0, _1, _2>{});
+    auto scaleB_shape = make_shape(tN, tK, L); // (n,k,l)
+    auto scaleB_layout = make_ordered_layout(scaleB_shape, Step<_1, _0, _2>{});
+
+    // Note that mScaleA_mkl and mScaleB_nkl are already blocked tiled in the `m` host and 
+    // gScaleA_mkl and gScaleB_nkl in `g` global memory are same as mScaleA_mkl and mScaleB_nkl.
+    Tensor mScaleA_mkl = make_tensor(make_gmem_ptr(mainloop_params.ptr_scale_A), scaleA_layout); // (scale_m,k,l)
+    Tensor mScaleB_nkl = make_tensor(make_gmem_ptr(mainloop_params.ptr_scale_B), scaleB_layout); // (n,k,l)
+
+    return cute::make_tuple(gA_mkl, gB_nkl, mScaleA_mkl, mScaleB_nkl);
+  }
+
+  /// Perform a collective-scoped matrix multiply-accumulate
+  /// Producer Perspective
+  template <
+    class TensorA, class TensorB,
+    class TensorScaleA, class TensorScaleB,
+    class KTileIterator, class BlockCoord
+  >
+  CUTLASS_DEVICE void
+  load(
+      Params const& mainloop_params,
+      MainloopPipeline pipeline,
+      PipelineState smem_pipe_write,
+      cute::tuple<TensorA, TensorB, TensorScaleA, TensorScaleB> const& load_inputs,
+      BlockCoord const& blk_coord,
+      KTileIterator k_tile_iter, int k_tile_count,
+      int thread_idx,
+      uint32_t block_rank_in_cluster,
+      TensorStorage& shared_tensors) {
+    int lane_predicate = cute::elect_one_sync();
+
+    // Blockscaling: Tma loads for load_input and CpAsync for load_scale
+    Tensor sA = make_tensor(make_smem_ptr(shared_tensors.smem_A.data()), SmemLayoutA{});        // (BLK_M,BLK_K,PIPE)
+    Tensor sB = make_tensor(make_smem_ptr(shared_tensors.smem_B.data()), SmemLayoutB{});        // (BLK_N,BLK_K,PIPE)
+    Tensor sScaleA = make_tensor(cute::make_smem_ptr(shared_tensors.smem_scale_A.data()), SmemLayoutScaleA{}); // (ScaleMsPerTile,k)
+    Tensor sScaleB = make_tensor(cute::make_smem_ptr(shared_tensors.smem_scale_B.data()), SmemLayoutScaleB{}); // (k)
+
+    //
+    // Prepare the TMA loads for A and B
+    //
+
+    constexpr uint32_t cluster_shape_x = get<0>(ClusterShape());
+    uint2 cluster_local_block_id = {block_rank_in_cluster % cluster_shape_x, block_rank_in_cluster / cluster_shape_x};
+
+    Tensor gA_mkl = get<0>(load_inputs);
+    Tensor gB_nkl = get<1>(load_inputs);
+
+    auto block_tma_a = mainloop_params.tma_load_a.get_slice(cluster_local_block_id.y);
+    auto block_tma_b = mainloop_params.tma_load_b.get_slice(cluster_local_block_id.x);
+
+    // Partition the inputs based on the current block coordinates.
+    auto [m_coord, n_coord, k_coord, l_coord] = blk_coord;
+    Tensor gA = gA_mkl(_,_,m_coord,_,l_coord);                                                     // (BLK_M,BLK_K,k)
+    Tensor gB = gB_nkl(_,_,n_coord,_,l_coord);                                                     // (BLK_N,BLK_K,k)
+
+
+    // Block scaling: load_scale has scaling tensors in global memory which are not tiled
+    Tensor mScaleA_mkl = get<2>(load_inputs);
+    Tensor mScaleB_nkl = get<3>(load_inputs);
+    auto scales_m = get<0>(mScaleA_mkl.shape());
+
+    Tensor cScaleA_mkl = make_identity_tensor(mScaleA_mkl.shape());
+
+    Tensor gScaleA = local_tile( 
+      mScaleA_mkl, make_tile(Int<ScaleMsPerTile>{}), 
+      make_coord(m_coord,_,l_coord));                   // (ScaleMsPerTile,k,1)
+    Tensor cScaleA = local_tile( 
+      cScaleA_mkl, make_tile(Int<ScaleMsPerTile>{}), 
+      make_coord(m_coord,_,l_coord));
+    Tensor gScaleB = mScaleB_nkl(n_coord,_,l_coord);                                           // (1,k,1)
+
+    // TODO: test `scale_copy_a` with `ScaleMsPerTile` < 128
+    TiledCopy scale_copy_a = make_tiled_copy(SmemBlockScalingCopyAtomA{}, 
+      Layout<Shape<_32>>{}, Layout<Shape<_1>>{}); // (1,1,1)
+    TiledCopy scale_copy_b = make_tiled_copy(SmemBlockScalingCopyAtomB{}, 
+      Layout<Shape<_1>>{}, Layout<Shape<_1>>{}); // (1,1,1)
+    ThrCopy thr_scale_copy_a = scale_copy_a.get_slice(threadIdx.x);
+    ThrCopy thr_scale_copy_b = scale_copy_b.get_slice(threadIdx.x);
+    
+    Tensor tAgA_ScaleA = thr_scale_copy_a.partition_S(gScaleA);
+    Tensor tAcA_ScaleA = thr_scale_copy_a.partition_S(cScaleA);
+    Tensor tAsA_ScaleA = thr_scale_copy_a.partition_D(sScaleA);
+    
+    Tensor tBgB_ScaleB = thr_scale_copy_b.partition_S(gScaleB);
+    Tensor tBsB_ScaleB = thr_scale_copy_b.partition_D(sScaleB);
+
+    // Applies the mapping from block_tma_a
+    Tensor tAgA = block_tma_a.partition_S(gA);                                              // (TMA,TMA_M,TMA_K,k)
+    Tensor tAsA = block_tma_a.partition_D(sA);                                              // (TMA,TMA_M,TMA_K,PIPE)
+
+    Tensor tBgB = block_tma_b.partition_S(gB);                                              // (TMA,TMA_N,TMA_K,k)
+    Tensor tBsB = block_tma_b.partition_D(sB);                                              // (TMA,TMA_N,TMA_K,PIPE)
+
+    uint16_t mcast_mask_a = 0;
+    uint16_t mcast_mask_b = 0;
+
+    // Issue TmaLoads for GEMM operands A/B and CpAsync for scale tensors
+    // Maps the tile -> block, value
+    if constexpr (cute::is_same_v<GmemTiledCopyA, SM90_TMA_LOAD_MULTICAST>) {
+      auto block_layout = Layout<typename DispatchPolicy::ClusterShape>{};                       // (m,n) -> block_id
+      for (int n = 0; n < size<1>(block_layout); ++n) {
+        mcast_mask_a |= (uint16_t(1) << block_layout(cluster_local_block_id.x,n,Int<0>{}));
+      }
+    }
+
+    if constexpr (cute::is_same_v<GmemTiledCopyB, SM90_TMA_LOAD_MULTICAST>) {
+      auto block_layout = Layout<typename DispatchPolicy::ClusterShape>{};                       // (m,n) -> block_id
+      for (int m = 0; m < size<0>(block_layout); ++m) {
+        mcast_mask_b |= (uint16_t(1) << block_layout(m,cluster_local_block_id.y,Int<0>{}));
+      }
+    }
+
+    // Allocate predicate tensors for a_scales (since we can't guarantee that 
+    // all scales are valid, since we could have a partial tiles along M)
+    Tensor tApA_ScaleA = make_tensor<bool>(shape(tAsA_ScaleA(_,_,0)));
+    #pragma unroll
+    for (int i = 0; i < size(tApA_ScaleA); ++i) {
+      tApA_ScaleA(i) = get<0>(tAcA_ScaleA(i)) < scales_m;
+    }
+
+    // Mainloop
+    CUTLASS_PRAGMA_NO_UNROLL
+    for ( ; k_tile_count > 0; --k_tile_count) {
+      // LOCK smem_pipe_write for _writing_
+      pipeline.producer_acquire(smem_pipe_write);
+
+      //
+      // Copy gmem to smem for *k_tile_iter
+      //
+      int write_stage = smem_pipe_write.index();
+      using BarrierType = typename MainloopPipeline::ProducerBarrierType;
+      BarrierType* tma_barrier = pipeline.producer_get_barrier(smem_pipe_write);
+
+      // Copy operands A and B from global memory to shared memory
+      if (lane_predicate) copy(mainloop_params.tma_load_a.with(*tma_barrier, mcast_mask_a), tAgA(_,_,_,*k_tile_iter), tAsA(_,_,_,write_stage));
+      if (lane_predicate) copy(mainloop_params.tma_load_b.with(*tma_barrier, mcast_mask_b), tBgB(_,_,_,*k_tile_iter), tBsB(_,_,_,write_stage));
+
+      // Copy scale tensors from global memory to shared memory
+      copy_if(scale_copy_a, tApA_ScaleA, tAgA_ScaleA(_,_,*k_tile_iter), tAsA_ScaleA(_,_,write_stage));
+      copy(scale_copy_b, tBgB_ScaleB(_,*k_tile_iter), tBsB_ScaleB(_,write_stage));
+      pipeline.producer_commit(smem_pipe_write, cutlass::arch::cpasync_barrier_arrive_noinc);
+
+      ++k_tile_iter;
+
+      // Advance smem_pipe_write
+      ++smem_pipe_write;
+    }
+  }
+
+  /// Perform a Producer Epilogue to prevent early exit of blocks in a Cluster
+  CUTLASS_DEVICE void
+  load_tail(
+      MainloopPipeline pipeline,
+      PipelineState smem_pipe_write) {
+    int lane_predicate = cute::elect_one_sync();
+
+    // Issue the epilogue waits
+    if (lane_predicate) {
+      /* This helps avoid early exit of blocks in Cluster
+       * Waits for all stages to either be released (all
+       * Consumer UNLOCKs), or if the stage was never used
+       * then would just be acquired since the phase was
+       * still inverted from make_producer_start_state
+       */
+      pipeline.producer_tail(smem_pipe_write);
+    }
+  }
+
+  /// Perform a collective-scoped matrix multiply-accumulate
+  /// Consumer Perspective
+  template <
+    class FrgTensorC
+  >
+  CUTLASS_DEVICE void
+  mma(MainloopPipeline pipeline,
+      PipelineState smem_pipe_read,
+      FrgTensorC& accum,
+      int k_tile_count,
+      int thread_idx,
+      TensorStorage& shared_tensors,
+      Params const& mainloop_params) {
+
+
+    static_assert(is_rmem<FrgTensorC>::value, "C tensor must be rmem resident.");
+    static_assert(cute::rank(SmemLayoutA{}) == 3, "Smem layout must be rank 3.");
+    static_assert(cute::rank(SmemLayoutB{}) == 3, "Smem layout must be rank 3.");
+    static_assert(cute::is_void_v<SmemCopyAtomA>,
+      "SM90 GMMA mainloops cannot have a non-void copy atom for smem sourced instructions.");
+    static_assert(cute::is_void_v<SmemCopyAtomB>,
+      "SM90 GMMA mainloops cannot have a non-void copy atom for smem sourced instructions.");
+
+    Tensor sA = make_tensor(make_smem_ptr(shared_tensors.smem_A.data()), SmemLayoutA{});          // (BLK_M,BLK_K,PIPE)
+    Tensor sB = make_tensor(make_smem_ptr(shared_tensors.smem_B.data()), SmemLayoutB{});          // (BLK_N,BLK_K,PIPE)
+    
+    // Block scaling
+    Tensor sScaleAViewAsC = make_tensor(cute::make_smem_ptr(shared_tensors.smem_scale_A.data()),
+      Layout<
+        Shape<Shape<Int<ScaleGranularityM>, Int<ScaleMsPerTile>>, cute::tuple_element_t<1, TileShape>, Int<DispatchPolicy::Stages>>,
+        Stride<Stride<_0, _1>, _0, Int<ScaleMsPerTile>>
+      >{}); // ((ScaleGranularityM,ScaleMsPerTile),n,k)
+    Tensor sScaleB = make_tensor(cute::make_smem_ptr(shared_tensors.smem_scale_B.data()), SmemLayoutScaleB{}); // (k)
+
+    //
+    // Define C accumulators and A/B partitioning
+    //
+    
+    // Layout of warp group to thread mapping
+
+    static_assert(stride<0>(typename TiledMma::ALayout{}) == 0 and 
+                  stride<0>(typename TiledMma::BLayout{}) == 0 and
+                  size<0>(typename TiledMma::ALayout{}) == NumThreadsPerWarpGroup and
+                  size<0>(typename TiledMma::BLayout{}) == NumThreadsPerWarpGroup, 
+                  "Stride of the first mode must be 0 and the size of the mode must be NumThreadsPerWarpGroup");
+
+    constexpr int MmaWarpGroups = size(TiledMma{}) / NumThreadsPerWarpGroup;
+    Layout warp_group_thread_layout = make_layout(Int<MmaWarpGroups>{}, 
+                                                  Int<NumThreadsPerWarpGroup>{});
+
+    int warp_group_idx = __shfl_sync(0xFFFFFFFF, thread_idx / NumThreadsPerWarpGroup, 0);
+
+    TiledMma tiled_mma;
+    auto thread_mma = tiled_mma.get_slice(warp_group_thread_layout(warp_group_idx));
+
+    Tensor tCsScaleAViewAsC = tiled_mma.get_slice(thread_idx).partition_C(sScaleAViewAsC);    // (MMA,MMA_M,MMA_N,PIPE), `thread_mma` above is correct when partitioning A and B, but it is not correct when partitioning C.
+
+    Tensor tCsA = thread_mma.partition_A(sA);                                                 // (MMA,MMA_M,MMA_K,PIPE)
+    Tensor tCsB = thread_mma.partition_B(sB);                                                 // (MMA,MMA_N,MMA_K,PIPE)
+
+    // Allocate "fragments/descriptors"
+    Tensor tCrA = thread_mma.make_fragment_A(tCsA);                                           // (MMA,MMA_M,MMA_K,PIPE)
+    Tensor tCrB = thread_mma.make_fragment_B(tCsB);                                           // (MMA,MMA_N,MMA_K,PIPE)
+
+    CUTE_STATIC_ASSERT_V(size<1>(tCsA) == size<1>(accum));                                                         // M
+    CUTE_STATIC_ASSERT_V(size<1>(tCsB) == size<2>(accum));                                                         // N
+    CUTE_STATIC_ASSERT_V(size<2>(tCsA) == size<2>(tCsB));                                                          // K
+    CUTE_STATIC_ASSERT_V(size<3>(tCsA) == size<3>(tCsB));                                                       // PIPE
+    CUTE_STATIC_ASSERT_V(Int<DispatchPolicy::Stages>{} == size<2>(sA));                                         // PIPE
+    CUTE_STATIC_ASSERT_V(Int<DispatchPolicy::Stages>{} == size<2>(sB));                                         // PIPE
+
+    //
+    // PIPELINED MAIN LOOP
+    //
+    static_assert((0 <= K_PIPE_MMAS) && (K_PIPE_MMAS <  K_PIPE_MAX),
+        "ERROR : Incorrect number of MMAs in flight");
+
+    // We release buffers to producer warps(dma load) with some mmas in flight
+    PipelineState smem_pipe_release = smem_pipe_read;
+    
+    // Per block scale values for operand A and B
+
+    using RegLayoutScaleAViewAsC = decltype(make_layout_like(tCsScaleAViewAsC(_, _, _, 0).layout())); // `make_layout_like` makes a compact layout.
+    using RegLayoutScaleAEssential = decltype(filter_zeros(RegLayoutScaleAViewAsC{}.stride(), RegLayoutScaleAViewAsC{}.shape())); // an interface to traverse the underlying storage for the compact layout mentioned above
+
+    Tensor tCrScaleAViewAsC = make_tensor<ElementBlockScale>(RegLayoutScaleAViewAsC{});              // (MMA,MMA_M,MMA_N)
+    ElementBlockScale scale_b;
+
+    // Prologue GMMAs
+    int prologue_mma_count = min(K_PIPE_MMAS, k_tile_count);
+
+    tiled_mma.accumulate_ = GMMA::ScaleOut::Zero;
+
+    GmmaFP8AccumulationWithScale accumulation(accum, size<2>(TileShape{}) / size<2>(typename TiledMma::AtomShape_MNK{}), size<2>(tCrA));
+    warpgroup_fence_operand(accumulation());
+    CUTLASS_PRAGMA_UNROLL
+    for (int k_tile_prologue = prologue_mma_count; k_tile_prologue > 0; --k_tile_prologue)
+    {
+      // WAIT on smem_pipe_read until its data are available (phase bit flips from rdPhaseBit value)
+      auto barrier_token = pipeline.consumer_try_wait(smem_pipe_read);
+      pipeline.consumer_wait(smem_pipe_read, barrier_token);
+
+      if (accumulation.prepare_if_needed()) {
+        tiled_mma.accumulate_ = GMMA::ScaleOut::Zero;
+      }
+
+      int read_stage = smem_pipe_read.index();
+      
+      // Load per block scale values from shared memory to registers.
+      scale_b = sScaleB[read_stage];
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < size(RegLayoutScaleAEssential{}); i++) {
+        tCrScaleAViewAsC.data()[i] = tCsScaleAViewAsC(_, _, _, read_stage)(idx2crd(i, RegLayoutScaleAEssential{}));
+      }
+      if constexpr (ScaleMsPerTile == 1) {
+        static_assert(size(RegLayoutScaleAEssential{}) == 1);
+        tCrScaleAViewAsC.data()[0] = __shfl_sync(0xffffffff, tCrScaleAViewAsC.data()[0] * scale_b, 0); // `tCrScaleAViewAsC.data()[0]` are all same in a warp group when `ScaleMsPerTile == 1`.
+      } else {
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < size(RegLayoutScaleAEssential{}); i++) {
+          tCrScaleAViewAsC.data()[i] = tCrScaleAViewAsC.data()[i] * scale_b;
+        }
+      }
+
+      warpgroup_arrive();
+      // Unroll the K mode manually to set scale D to 1
+      CUTLASS_PRAGMA_UNROLL
+      for (int k_block = 0; k_block < size<2>(tCrA); ++k_block) {
+        // (V,M,K) x (V,N,K) => (V,M,N)
+        cute::gemm(tiled_mma, tCrA(_,_,k_block,read_stage), tCrB(_,_,k_block,read_stage), accumulation());
+        tiled_mma.accumulate_ = GMMA::ScaleOut::One;
+      }
+      warpgroup_commit_batch();
+
+      // Block scale the accumulators with reg tensor `tCrScaleAViewAsC`
+      accumulation.scale_if_needed(tCrScaleAViewAsC);
+
+      ++smem_pipe_read;
+    }
+
+    warpgroup_fence_operand(accumulation());
+    // Mainloop GMMAs
+    k_tile_count -= prologue_mma_count;
+
+    CUTLASS_PRAGMA_NO_UNROLL
+    for ( ; k_tile_count > 0; --k_tile_count)
+    {
+      // WAIT on smem_pipe_read until its data are available (phase bit flips from rdPhaseBit value)
+      auto barrier_token = pipeline.consumer_try_wait(smem_pipe_read);
+      pipeline.consumer_wait(smem_pipe_read, barrier_token);
+
+      //
+      // Compute on k_tile
+      //
+
+      int read_stage = smem_pipe_read.index();
+
+      // Load per block scale values from shared memory to registers (at most twice per block along M and exactly once per block along N) 
+      scale_b = sScaleB[read_stage];
+      CUTLASS_PRAGMA_UNROLL
+      for (int i = 0; i < size(RegLayoutScaleAEssential{}); i++) {
+        tCrScaleAViewAsC.data()[i] = tCsScaleAViewAsC(_, _, _, read_stage)(idx2crd(i, RegLayoutScaleAEssential{}));
+      }
+      if constexpr (ScaleMsPerTile == 1) {
+        static_assert(size(RegLayoutScaleAEssential{}) == 1);
+        tCrScaleAViewAsC.data()[0] = __shfl_sync(0xffffffff, tCrScaleAViewAsC.data()[0] * scale_b, 0); // `tCrScaleAViewAsC.data()[0]` are all same in a warp group when `ScaleMsPerTile == 1`.
+      } else {
+        CUTLASS_PRAGMA_UNROLL
+        for (int i = 0; i < size(RegLayoutScaleAEssential{}); i++) {
+          tCrScaleAViewAsC.data()[i] = tCrScaleAViewAsC.data()[i] * scale_b;
+        }
+      }
+
+      if (accumulation.prepare_if_needed()) {
+        tiled_mma.accumulate_ = GMMA::ScaleOut::Zero;
+      }
+
+      warpgroup_fence_operand(accumulation());
+      warpgroup_arrive();
+      // Unroll the K mode manually to set scale D to 1
+      CUTLASS_PRAGMA_UNROLL
+      for (int k_block = 0; k_block < size<2>(tCrA); ++k_block) {
+        // (V,M,K) x (V,N,K) => (V,M,N)
+        cute::gemm(tiled_mma, tCrA(_,_,k_block,read_stage), tCrB(_,_,k_block,read_stage), accumulation());
+        tiled_mma.accumulate_ = GMMA::ScaleOut::One;
+      }
+      warpgroup_commit_batch();
+
+      /// Wait on the GMMA barrier for K_PIPE_MMAS (or fewer) outstanding to ensure smem_pipe_write is consumed
+      warpgroup_wait<K_PIPE_MMAS>();
+      warpgroup_fence_operand(accumulation());
+
+      // Block scale the accumulators with reg tensor `tCrScaleAViewAsC`
+      accumulation.scale_if_needed(tCrScaleAViewAsC);
+
+      pipeline.consumer_release(smem_pipe_release);                 // UNLOCK smem_pipe_release, done _computing_ on it
+
+      // Advance smem_pipe_read and smem_pipe_release
+      ++smem_pipe_read;
+      ++smem_pipe_release;
+    }
+    
+    accumulation.scale_residue_if_needed(tCrScaleAViewAsC);
+
+    warpgroup_fence_operand(accumulation());
+  }
+
+  /// Perform a Consumer Epilogue to release all buffers
+  CUTLASS_DEVICE void
+  mma_tail(MainloopPipeline pipeline, PipelineState smem_pipe_release, int k_tile_count) {
+    // Prologue GMMAs
+    int prologue_mma_count = min(K_PIPE_MMAS, k_tile_count);
+    k_tile_count -= prologue_mma_count;
+
+    smem_pipe_release.advance(k_tile_count);
+
+    // Wait on all GMMAs to complete
+    warpgroup_wait<0>();
+
+    for (int count = 0; count < prologue_mma_count; ++count) {
+      pipeline.consumer_release(smem_pipe_release);                 // UNLOCK smem_pipe_release, done _computing_ on it
+      ++smem_pipe_release;
+    }
+  }
+};
+
+/////////////////////////////////////////////////////////////////////////////////////////////////
+
+} // namespace cutlass::gemm::collective
+
+/////////////////////////////////////////////////////////////////////////////////////////////////

csrc/cutlass_extensions/gemm/dispatch_policy.hpp

@@ -0,0 +1,39 @@
+#pragma once
+
+#include "cutlass/gemm/dispatch_policy.hpp"
+
+namespace cutlass::gemm {
+
+//////////////////////////////////////////////////////////////////////////////
+
+// FP8 related policies (including Blocked Scaled Accumulation)
+//  `ScaleGranularityM` specifies scaling granularity along M, while zero-value
+//  `ScaleGranularityM` indicates that scaling granularity is
+//  `size<0>(TileShape_MNK{})` along M.
+template <int ScaleGranularityM = 0>
+struct KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum
+    : KernelTmaWarpSpecializedCooperative {};
+
+// n-buffer in smem (Hopper TMA), pipelined with Hopper GMMA and TMA, Warp
+// specialized dynamic schedule For FP8 kernels with Block Scaling
+template <int Stages_, class ClusterShape_ = Shape<_1, _1, _1>,
+          class KernelSchedule = KernelTmaWarpSpecialized,
+          int ScaleGranularityM =
+              0  // `ScaleGranularityM` specifies scaling granularity along M,
+                 // while zero-value `ScaleGranularityM` indicates that scaling
+                 // granularity is `size<0>(TileShape_MNK{})` along M.
+          >
+struct MainloopSm90TmaGmmaWarpSpecializedBlockScalingSubGroupMFP8
+    : MainloopSm90TmaGmmaWarpSpecialized<Stages_, ClusterShape_,
+                                         KernelSchedule> {
+  static_assert(
+      cute::is_same_v<
+          KernelSchedule,
+          KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum<
+              ScaleGranularityM>>,
+      "KernelSchedule must be one of the warp specialized policies");
+};
+
+//////////////////////////////////////////////////////////////////////////////
+
+}  // namespace cutlass::gemm

csrc/cutlass_extensions/vllm_collective_builder.cuh

@@ -1,6 +1,6 @@
 #pragma once
 
-#include "cutlass/gemm/collective/collective_builder.hpp"
+#include "cutlass_extensions/gemm/collective/collective_builder.hpp"
 
 namespace cutlass::gemm::collective {
 using namespace cute;

csrc/launch_bounds_utils.h

@@ -1,38 +0,0 @@
-#pragma once
-
-#include <cuda_runtime_api.h>
-#include <algorithm>
-
-// maximum blocks per SM cap
-#ifndef VLLM_LAUNCH_BLOCKS_CAP
-  #define VLLM_LAUNCH_BLOCKS_CAP 4
-#endif
-
-// compile-time estimate of max threads per SM for launch bounds.
-#ifndef VLLM_MAX_THREADS_PER_SM
-  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 300
-    #define VLLM_MAX_THREADS_PER_SM 1536
-  #else
-    #define VLLM_MAX_THREADS_PER_SM 2048
-  #endif
-#endif
-
-// compute the number of blocks per SM to request in __launch_bounds__
-#define VLLM_BLOCKS_DIV(VAL) (VLLM_MAX_THREADS_PER_SM / (VAL))
-#define VLLM_CLAMP_BLOCKS_PER_SM(VAL) \
-  (((VAL) <= 0)                       \
-       ? 1                            \
-       : (((VAL) < VLLM_LAUNCH_BLOCKS_CAP) ? (VAL) : VLLM_LAUNCH_BLOCKS_CAP))
-#define VLLM_BLOCKS_PER_SM(BLOCK_THREADS) \
-  VLLM_CLAMP_BLOCKS_PER_SM(VLLM_BLOCKS_DIV(BLOCK_THREADS))
-
-// runtime-time helper to compute blocks/SM
-static inline int vllm_runtime_blocks_per_sm(int block_threads) {
-  int device = -1;
-  cudaGetDevice(&device);
-  int max_threads_per_sm = VLLM_MAX_THREADS_PER_SM;
-  cudaDeviceGetAttribute(&max_threads_per_sm,
-                         cudaDevAttrMaxThreadsPerMultiProcessor, device);
-  int blocks = (block_threads > 0) ? (max_threads_per_sm / block_threads) : 1;
-  return VLLM_CLAMP_BLOCKS_PER_SM(blocks);
-}

csrc/layernorm_kernels.cu

@@ -1,10 +1,15 @@
 #include "type_convert.cuh"
 #include "dispatch_utils.h"
-#include "cub_helpers.h"
 
 #include <torch/cuda.h>
 #include <c10/cuda/CUDAGuard.h>
 
+#ifndef USE_ROCM
+  #include <cub/cub.cuh>
+#else
+  #include <hipcub/hipcub.hpp>
+#endif
+
 namespace vllm {
 
 // TODO(woosuk): Further optimize this kernel.
@@ -25,7 +30,7 @@ __global__ void rms_norm_kernel(
 
   using BlockReduce = cub::BlockReduce<float, 1024>;
   __shared__ typename BlockReduce::TempStorage reduceStore;
-  variance = BlockReduce(reduceStore).Reduce(variance, CubAddOp{}, blockDim.x);
+  variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
 
   if (threadIdx.x == 0) {
     s_variance = rsqrtf(variance / hidden_size + epsilon);
@@ -80,7 +85,7 @@ fused_add_rms_norm_kernel(
 
   using BlockReduce = cub::BlockReduce<float, 1024>;
   __shared__ typename BlockReduce::TempStorage reduceStore;
-  variance = BlockReduce(reduceStore).Reduce(variance, CubAddOp{}, blockDim.x);
+  variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
 
   if (threadIdx.x == 0) {
     s_variance = rsqrtf(variance / hidden_size + epsilon);
@@ -121,7 +126,7 @@ fused_add_rms_norm_kernel(
 
   using BlockReduce = cub::BlockReduce<float, 1024>;
   __shared__ typename BlockReduce::TempStorage reduceStore;
-  variance = BlockReduce(reduceStore).Reduce(variance, CubAddOp{}, blockDim.x);
+  variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
 
   if (threadIdx.x == 0) {
     s_variance = rsqrtf(variance / hidden_size + epsilon);
@@ -135,211 +140,6 @@ fused_add_rms_norm_kernel(
   }
 }
 
-/* Function specialization in the case of FP16/BF16 tensors.
-   Additional optimizations we can make in this case are
-   packed and vectorized operations, which help with the
-   memory latency bottleneck.
-
-   _f16VecPN struct extends _f16Vec to add operations specifically required for
-   polynomial normalization (poly norm).
-   The original _f16Vec does not include the sum-of-powers computation or
-   in-place polynomial normalization logic. */
-template <typename scalar_t, int width>
-struct alignas(16) _f16VecPN : _f16Vec<scalar_t, width> {
-  using Base = _f16Vec<scalar_t, width>;
-  using Converter = typename Base::Converter;
-  using T1 = typename Base::T1;
-  using T2 = typename Base::T2;
-  using Base::data;
-
-  __device__ auto sum_pows() const {
-    float s2 = 0.0f, s4 = 0.0f, s6 = 0.0f;
-
-#pragma unroll
-    for (int i = 0; i < width; i += 2) {
-      float2 z = Converter::convert(T2{data[i], data[i + 1]});
-      float x2 = z.x * z.x;
-      float x4 = x2 * x2;
-      float x6 = x4 * x2;
-
-      float y2 = z.y * z.y;
-      float y4 = y2 * y2;
-      float y6 = y4 * y2;
-
-      s2 += x2 + y2;
-      s4 += x4 + y4;
-      s6 += x6 + y6;
-    }
-    return std::make_tuple(s2, s4, s6);
-  }
-
-  __device__ void poly_norm_inplace(const float w2_inv_std,
-                                    const float w1_inv_std2,
-                                    const float w0_inv_std3, const float bias) {
-#pragma unroll
-    for (int i = 0; i < width; i += 2) {
-      float2 z = Converter::convert(T2{data[i], data[i + 1]});
-
-      float x2 = z.x * z.x;
-      float x3 = x2 * z.x;
-      z.x = w2_inv_std * z.x + w1_inv_std2 * x2 + w0_inv_std3 * x3 + bias;
-
-      float y2 = z.y * z.y;
-      float y3 = y2 * z.y;
-      z.y = w2_inv_std * z.y + w1_inv_std2 * y2 + w0_inv_std3 * y3 + bias;
-
-      auto out = Converter::convert(z);
-      data[i] = out.x;
-      data[i + 1] = out.y;
-    }
-  }
-};
-
-template <typename scalar_t, int width>
-__global__ std::enable_if_t<(width > 0) && _typeConvert<scalar_t>::exists>
-poly_norm_kernel(scalar_t* __restrict__ out,           // [..., hidden_size]
-                 const scalar_t* __restrict__ input,   // [..., hidden_size]
-                 const scalar_t* __restrict__ weight,  // [3]
-                 const scalar_t* __restrict__ bias,    // [1]
-                 const float epsilon, const int hidden_size) {
-  // Sanity checks on our vector struct and type-punned pointer arithmetic
-  static_assert(std::is_pod_v<_f16VecPN<scalar_t, width>>);
-  static_assert(sizeof(_f16VecPN<scalar_t, width>) == sizeof(scalar_t) * width);
-
-  /* These and the argument pointers are all declared `restrict` as they are
-     not aliased in practice. Argument pointers should not be dereferenced
-     in this kernel as that would be undefined behavior */
-  auto* __restrict__ input_v =
-      reinterpret_cast<const _f16VecPN<scalar_t, width>*>(input);
-  const int vec_hidden_size = hidden_size / width;
-  float variance = 0.0f;
-  float variance2 = 0.0f;
-  float variance3 = 0.0f;
-
-  for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
-    int id = blockIdx.x * vec_hidden_size + idx;
-    _f16VecPN<scalar_t, width> temp = input_v[id];
-    auto [x2, x4, x6] = temp.sum_pows();
-
-    variance += x2;
-    variance2 += x4;
-    variance3 += x6;
-  }
-
-  float3 thread_variances = make_float3(variance, variance2, variance3);
-
-  struct SumOp {
-    __device__ float3 operator()(const float3& a, const float3& b) const {
-      return make_float3(a.x + b.x, a.y + b.y, a.z + b.z);
-    }
-  };
-
-  using BlockReduce = cub::BlockReduce<float3, 1024>;
-  __shared__ typename BlockReduce::TempStorage reduceStore;
-  float3 block_variances =
-      BlockReduce(reduceStore).Reduce(thread_variances, SumOp{}, blockDim.x);
-
-  variance = block_variances.x;
-  variance2 = block_variances.y;
-  variance3 = block_variances.z;
-
-  __shared__ float s_w2_inv_std;
-  __shared__ float s_w1_inv_std2;
-  __shared__ float s_w0_inv_std3;
-  __shared__ float s_bias;
-
-  if (threadIdx.x == 0) {
-    float w0 = (float)weight[0];
-    float w1 = (float)weight[1];
-    float w2 = (float)weight[2];
-    s_bias = (float)bias[0];
-
-    s_w2_inv_std = w2 * rsqrtf(variance / hidden_size + epsilon);
-    s_w1_inv_std2 = w1 * rsqrtf(variance2 / hidden_size + epsilon);
-    s_w0_inv_std3 = w0 * rsqrtf(variance3 / hidden_size + epsilon);
-  }
-  __syncthreads();
-
-  auto* __restrict__ out_v = reinterpret_cast<_f16VecPN<scalar_t, width>*>(out);
-
-  for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
-    int id = blockIdx.x * vec_hidden_size + idx;
-    _f16VecPN<scalar_t, width> temp = input_v[id];
-    temp.poly_norm_inplace(s_w2_inv_std, s_w1_inv_std2, s_w0_inv_std3, s_bias);
-    out_v[id] = temp;
-  }
-}
-
-/* Generic poly_norm_kernel
-   The width field is not used here but necessary for other specializations.
- */
-template <typename scalar_t, int width>
-__global__ std::enable_if_t<(width == 0) || !_typeConvert<scalar_t>::exists>
-poly_norm_kernel(scalar_t* __restrict__ out,           // [..., hidden_size]
-                 const scalar_t* __restrict__ input,   // [..., hidden_size]
-                 const scalar_t* __restrict__ weight,  // [3]
-                 const scalar_t* __restrict__ bias,    // [1]
-                 const float epsilon, const int hidden_size) {
-  float variance = 0.0f;
-  float variance2 = 0.0f;
-  float variance3 = 0.0f;
-
-  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float)input[blockIdx.x * hidden_size + idx];
-    float x2 = x * x;
-    float x4 = x2 * x2;
-    float x6 = x4 * x2;
-
-    variance += x2;
-    variance2 += x4;
-    variance3 += x6;
-  }
-
-  float3 thread_variances = make_float3(variance, variance2, variance3);
-
-  struct SumOp {
-    __device__ float3 operator()(const float3& a, const float3& b) const {
-      return make_float3(a.x + b.x, a.y + b.y, a.z + b.z);
-    }
-  };
-
-  using BlockReduce = cub::BlockReduce<float3, 1024>;
-  __shared__ typename BlockReduce::TempStorage reduceStore;
-  float3 block_variances =
-      BlockReduce(reduceStore).Reduce(thread_variances, SumOp{}, blockDim.x);
-
-  variance = block_variances.x;
-  variance2 = block_variances.y;
-  variance3 = block_variances.z;
-
-  __shared__ float s_w2_inv_std;
-  __shared__ float s_w1_inv_std2;
-  __shared__ float s_w0_inv_std3;
-  __shared__ float s_bias;
-
-  if (threadIdx.x == 0) {
-    float w0 = (float)weight[0];
-    float w1 = (float)weight[1];
-    float w2 = (float)weight[2];
-    s_bias = (float)bias[0];
-
-    s_w2_inv_std = w2 * rsqrtf(variance / hidden_size + epsilon);
-    s_w1_inv_std2 = w1 * rsqrtf(variance2 / hidden_size + epsilon);
-    s_w0_inv_std3 = w0 * rsqrtf(variance3 / hidden_size + epsilon);
-  }
-  __syncthreads();
-
-  for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float)input[blockIdx.x * hidden_size + idx];
-    float x2 = x * x;
-    float x3 = x2 * x;
-
-    out[blockIdx.x * hidden_size + idx] =
-        (scalar_t)(x * s_w2_inv_std + x2 * s_w1_inv_std2 + x3 * s_w0_inv_std3 +
-                   s_bias);
-  }
-}
-
 }  // namespace vllm
 
 void rms_norm(torch::Tensor& out,     // [..., hidden_size]
@@ -419,49 +219,3 @@ void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
     LAUNCH_FUSED_ADD_RMS_NORM(0);
   }
 }
-
-#define LAUNCH_FUSED_POLY_NORM(width)                                         \
-  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "poly_norm_kernel", [&] { \
-    vllm::poly_norm_kernel<scalar_t, width><<<grid, block, 0, stream>>>(      \
-        out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),                 \
-        weight.data_ptr<scalar_t>(), bias.data_ptr<scalar_t>(), epsilon,      \
-        hidden_size);                                                         \
-  });
-
-void poly_norm(torch::Tensor& out,     // [..., hidden_size]
-               torch::Tensor& input,   // [..., hidden_size]
-               torch::Tensor& weight,  // [3]
-               torch::Tensor& bias,    // [1]
-               double epsilon) {
-  TORCH_CHECK(out.is_contiguous());
-  TORCH_CHECK(input.is_contiguous());
-  TORCH_CHECK(out.data_ptr() != input.data_ptr());
-
-  int hidden_size = input.size(-1);
-  int num_tokens = input.numel() / hidden_size;
-
-  dim3 grid(num_tokens);
-  /* This kernel is memory-latency bound in many scenarios.
-     When num_tokens is large, a smaller block size allows
-     for increased block occupancy on CUs and better latency
-     hiding on global mem ops. */
-  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
-  dim3 block(std::min(hidden_size, max_block_size));
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  /*If the tensor types are FP16/BF16, try to use the optimized kernel
-    with packed + vectorized ops.
-    Max optimization is achieved with a width-8 vector of FP16/BF16s
-    since we can load at most 128 bits at once in a global memory op.
-    However, this requires each tensor's data to be aligned to 16
-    bytes.
-   */
-  auto inp_ptr = reinterpret_cast<std::uintptr_t>(input.data_ptr());
-  auto out_ptr = reinterpret_cast<std::uintptr_t>(out.data_ptr());
-  bool ptrs_are_aligned = inp_ptr % 16 == 0 && out_ptr % 16 == 0;
-  if (ptrs_are_aligned && hidden_size % 8 == 0) {
-    LAUNCH_FUSED_POLY_NORM(8);
-  } else {
-    LAUNCH_FUSED_POLY_NORM(0);
-  }
-}

csrc/layernorm_quant_kernels.cu

@@ -8,11 +8,16 @@
 #include "type_convert.cuh"
 #include "quantization/fp8/common.cuh"
 #include "dispatch_utils.h"
-#include "cub_helpers.h"
 
 #include <torch/cuda.h>
 #include <c10/cuda/CUDAGuard.h>
 
+#ifndef USE_ROCM
+  #include <cub/cub.cuh>
+#else
+  #include <hipcub/hipcub.hpp>
+#endif
+
 namespace vllm {
 
 // TODO(woosuk): Further optimize this kernel.
@@ -34,7 +39,7 @@ __global__ void rms_norm_static_fp8_quant_kernel(
 
   using BlockReduce = cub::BlockReduce<float, 1024>;
   __shared__ typename BlockReduce::TempStorage reduceStore;
-  variance = BlockReduce(reduceStore).Reduce(variance, CubAddOp{}, blockDim.x);
+  variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
 
   if (threadIdx.x == 0) {
     s_variance = rsqrtf(variance / hidden_size + epsilon);
@@ -95,7 +100,7 @@ fused_add_rms_norm_static_fp8_quant_kernel(
 
   using BlockReduce = cub::BlockReduce<float, 1024>;
   __shared__ typename BlockReduce::TempStorage reduceStore;
-  variance = BlockReduce(reduceStore).Reduce(variance, CubAddOp{}, blockDim.x);
+  variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
 
   if (threadIdx.x == 0) {
     s_variance = rsqrtf(variance / hidden_size + epsilon);
@@ -144,7 +149,7 @@ fused_add_rms_norm_static_fp8_quant_kernel(
 
   using BlockReduce = cub::BlockReduce<float, 1024>;
   __shared__ typename BlockReduce::TempStorage reduceStore;
-  variance = BlockReduce(reduceStore).Reduce(variance, CubAddOp{}, blockDim.x);
+  variance = BlockReduce(reduceStore).Reduce(variance, cub::Sum{}, blockDim.x);
 
   if (threadIdx.x == 0) {
     s_variance = rsqrtf(variance / hidden_size + epsilon);

csrc/moe/dynamic_4bit_int_moe_cpu.cpp

@@ -1,156 +0,0 @@
-#include <ATen/ATen.h>
-#include <ATen/Parallel.h>
-#include <torch/all.h>
-
-// _dyn_quant_matmul_4bit is only available on AArch64.
-#if defined(__aarch64__)
-  #include <ATen/ops/_dyn_quant_matmul_4bit.h>
-#endif
-
-inline torch::Tensor mm(const torch::Tensor& a, const torch::Tensor& packed_w,
-                        int64_t group_size_eff, int64_t in_features,
-                        int64_t out_features) {
-#if defined(__aarch64__)
-  return at::_ops::_dyn_quant_matmul_4bit::call(a, packed_w, group_size_eff,
-                                                in_features, out_features);
-#else
-  TORCH_CHECK(false,
-              "dynamic 4-bit int MoE path requires AArch64 (ARM64); "
-              "_dyn_quant_matmul_4bit is unavailable on this architecture");
-  return {};
-#endif
-}
-
-enum ActivationKind : int64_t {
-  SwiGLU_Gu = 0,  // act = SiLU(g) * u
-  SwiGLUOAI = 1,  // act = SiLU(u) * g
-  SiLU = 2        // SiLU
-};
-
-torch::Tensor dynamic_4bit_int_moe_cpu(
-    torch::Tensor x, torch::Tensor topk_ids, torch::Tensor topk_weights,
-    torch::Tensor w13_packed, torch::Tensor w2_packed, int64_t H, int64_t I,
-    int64_t I2, int64_t group_size, bool apply_router_weight_on_input,
-    int64_t activation_kind) {
-  TORCH_CHECK(x.dim() == 2, "x must be 2D");
-  TORCH_CHECK(topk_ids.dim() == 2 && topk_weights.dim() == 2,
-              "topk tensors must be [T, K]");
-  TORCH_CHECK(
-      w13_packed.size(0) == w2_packed.size(0),
-      "w13_packed and w2_packed must have same number of experts in dim 0");
-  TORCH_CHECK(I2 == 2 * I, "I2 must equal 2*I");
-
-  const int64_t T = x.size(0);
-  const int64_t K = topk_ids.size(1);
-  const int64_t E = w13_packed.size(0);
-  const int64_t N = T * K;
-
-  auto x_c = x.contiguous();
-  auto ids_c = topk_ids.contiguous();
-  auto gates_c = topk_weights.to(at::kFloat).contiguous();
-
-  // bucketing tokens -> experts
-  c10::SmallVector<int64_t, 64> counts(
-      E, 0);  // Small vector uses stack allocation
-  {
-    const auto* ids_ptr = ids_c.data_ptr<int64_t>();
-    for (int64_t i = 0; i < N; ++i) {
-      const int64_t e_id = ids_ptr[i];
-      TORCH_CHECK(0 <= e_id && e_id < E, "expert id out of range");
-      counts[e_id]++;
-    }
-  }
-  c10::SmallVector<int64_t, 65> offsets(E + 1, 0);  // ( E +1 )
-  for (int64_t e = 0; e < E; ++e) offsets[e + 1] = offsets[e] + counts[e];
-
-  auto expert_tokens = at::empty({offsets[E]}, ids_c.options());
-  auto expert_gates = at::empty({offsets[E]}, gates_c.options());
-  {
-    c10::SmallVector<int64_t, 64> cursor(E, 0);
-    const auto* ids_ptr = ids_c.data_ptr<int64_t>();
-    const auto* gts_ptr = gates_c.data_ptr<float>();
-    auto* tok_ptr = expert_tokens.data_ptr<int64_t>();
-    auto* gate_ptr = expert_gates.data_ptr<float>();
-
-    for (int64_t t = 0; t < T; ++t) {
-      const int64_t base = t * K;
-      for (int64_t k = 0; k < K; ++k) {
-        const int64_t idx = base + k;
-        const int64_t e = ids_ptr[idx];
-        const int64_t p = offsets[e] + (cursor[e]++);
-        tok_ptr[p] = t;
-        gate_ptr[p] = gts_ptr[idx];
-      }
-    }
-  }
-
-  const int64_t g_eff_13 = (group_size != -1) ? group_size : H;
-  const int64_t g_eff_2 = (group_size != -1) ? group_size : I;
-
-  // Per-expert outputs filled in parallel
-  std::vector<torch::Tensor> y_list(E);
-  y_list.resize(E);
-
-  at::parallel_for(0, E, 1, [&](int64_t e_begin, int64_t e_end) {
-    for (int64_t e = e_begin; e < e_end; ++e) {
-      const int64_t te = counts[e];
-      if (te == 0) {
-        y_list[e] = at::empty({0, H}, x_c.options());
-        continue;
-      }
-
-      const int64_t start = offsets[e];
-
-      auto sel_tokens =
-          expert_tokens.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);
-      auto gates_e =
-          expert_gates.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);
-
-      auto x_e = x_c.index_select(/*dim=*/0, sel_tokens);
-
-      if (apply_router_weight_on_input) {
-        x_e = x_e.mul(gates_e.unsqueeze(1));
-      }
-
-      auto w13_e = w13_packed.select(/*dim=*/0, e);
-      auto w2_e = w2_packed.select(/*dim=*/0, e);
-
-      // W13
-      auto y13 =
-          mm(x_e, w13_e, g_eff_13, /*in_features=*/H, /*out_features=*/I2);
-
-      auto g_part = y13.narrow(/*dim=*/1, /*start=*/0, /*length=*/I);
-      auto u_part = y13.narrow(/*dim=*/1, /*start=*/I, /*length=*/I);
-
-      torch::Tensor act;
-      if (activation_kind == ActivationKind::SwiGLUOAI) {  // SwiGLUOAI
-        constexpr double kAlpha = 1.702;                   // GPT-OSS default
-        constexpr double kLimit = 7.0;                     // GPT-OSS default
-        auto gate_c = at::clamp_max(g_part, kLimit);
-        auto up_c = at::clamp(u_part, -kLimit, kLimit);
-        auto glu = gate_c.mul(at::sigmoid(gate_c.mul(kAlpha)));
-        act = up_c.add(1.0).mul(glu);
-      } else {  // SiLU , SwiGLU_GU, vLLM maps silu to SiluAndMul()
-        act = at::silu(g_part).mul(u_part);
-      }
-
-      // W2
-      auto y = mm(act, w2_e, g_eff_2, /*in_features=*/I, /*out_features=*/H);
-
-      if (!apply_router_weight_on_input) {
-        y = y.mul(gates_e.unsqueeze(1));
-      }
-
-      // Store per-expert result
-      y_list[e] = y;
-    }
-  });
-
-  // Concatenate all expert outputs to match expert_tokens order
-  auto Y_all = at::cat(y_list, /*dim=*/0);
-  auto out = at::zeros({T, H}, x.options());
-  out =
-      at::index_add(out, /*dim=*/0, /*index=*/expert_tokens, /*source=*/Y_all);
-
-  return out;
-}

csrc/moe/grouped_topk_kernels.cu

@@ -21,7 +21,6 @@
 #include <torch/all.h>
 #include <cuda_fp16.h>
 #include <cuda_bf16.h>
-#include <cuda/std/limits>
 #include <cooperative_groups.h>
 #include <cooperative_groups/reduce.h>
 namespace cg = cooperative_groups;
@@ -29,6 +28,7 @@ namespace cg = cooperative_groups;
 namespace vllm {
 namespace moe {
 
+constexpr float kNegInfinity = INFINITY * -1;
 constexpr unsigned FULL_WARP_MASK = 0xffffffff;
 constexpr int32_t WARP_SIZE = 32;
 constexpr int32_t BLOCK_SIZE = 512;
@@ -411,30 +411,14 @@ __device__ inline float cuda_cast<float, __nv_bfloat16>(__nv_bfloat16 val) {
   return __bfloat162float(val);
 }
 
-template <typename T>
-__device__ inline T neg_inf() {
-  // cuda::std::numeric_limits<T>::infinity() returns `0` for [T=bf16 or fp16]
-  // so we need to cast from fp32
-  return cuda_cast<T, float>(-cuda::std::numeric_limits<float>::infinity());
-}
-
-template <typename T>
-__device__ inline bool is_finite(const T val) {
-#if (__CUDACC_VER_MAJOR__ * 10000 + __CUDACC_VER_MINOR__ * 100 >= 120800)
-  return cuda::std::isfinite(val);
-#else
-  return isfinite(cuda_cast<float, T>(val));
-#endif
-}
-
 template <typename T>
 __device__ void topk_with_k2(T* output, T const* input,
                              cg::thread_block_tile<32> const& tile,
                              int32_t const lane_id,
                              int const num_experts_per_group) {
   // Get the top2 per thread
-  T largest = neg_inf<T>();
-  T second_largest = neg_inf<T>();
+  T largest = -INFINITY;
+  T second_largest = -INFINITY;
 
   if (num_experts_per_group > WARP_SIZE) {
     for (int i = lane_id; i < num_experts_per_group; i += WARP_SIZE) {
@@ -529,8 +513,8 @@ __global__ void group_idx_and_topk_idx_kernel(
       warp_id * topk;
   s_topk_idx += warp_id * topk;
 
-  T value = neg_inf<T>();
-  T topk_group_value = neg_inf<T>();
+  T value = kNegInfinity;
+  T topk_group_value = kNegInfinity;
   int32_t num_equalto_topkth_group;
 
 #if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
@@ -541,8 +525,11 @@ __global__ void group_idx_and_topk_idx_kernel(
   if (case_id < num_tokens) {
     // calculate group_idx
     int32_t target_num_min = WARP_SIZE - n_group + topk_group;
-    // The check is necessary to avoid abnormal input
-    if (lane_id < n_group && is_finite(group_scores[lane_id])) {
+    if (lane_id < n_group &&
+        (isfinite(cuda_cast<float, T>(
+            group_scores[lane_id]))))  // The check is necessary to avoid
+                                       // abnormal input
+    {
       value = group_scores[lane_id];
     }
 
@@ -553,11 +540,11 @@ __global__ void group_idx_and_topk_idx_kernel(
       __syncwarp();  // Ensure all threads have valid data before reduction
       topk_group_value = cg::reduce(tile, value, cg::greater<T>());
       if (value == topk_group_value) {
-        value = neg_inf<T>();
+        value = kNegInfinity;
       }
       pre_count_equal_to_top_value = count_equal_to_top_value;
-      count_equal_to_top_value =
-          __popc(__ballot_sync(FULL_WARP_MASK, (value == neg_inf<T>())));
+      count_equal_to_top_value = __popc(__ballot_sync(
+          FULL_WARP_MASK, (value == cuda_cast<T, float>(kNegInfinity))));
     }
     num_equalto_topkth_group = target_num_min - pre_count_equal_to_top_value;
   }
@@ -565,10 +552,11 @@ __global__ void group_idx_and_topk_idx_kernel(
 
   warp_topk::WarpSelect</*capability*/ WARP_SIZE, /*greater*/ true, T, int32_t,
                         /* is_stable */ true>
-      queue((int32_t)topk, neg_inf<T>());
+      queue((int32_t)topk, -INFINITY);
 
   int count_equalto_topkth_group = 0;
-  bool if_proceed_next_topk = topk_group_value != neg_inf<T>();
+  bool if_proceed_next_topk =
+      (topk_group_value != cuda_cast<T, float>(kNegInfinity));
   if (case_id < num_tokens && if_proceed_next_topk) {
     for (int i_group = 0; i_group < n_group; i_group++) {
       if ((group_scores[i_group] > topk_group_value) ||
@@ -577,10 +565,11 @@ __global__ void group_idx_and_topk_idx_kernel(
         int32_t offset = i_group * num_experts_per_group;
         for (int32_t i = lane_id; i < align_num_experts_per_group;
              i += WARP_SIZE) {
-          T candidates = (i < num_experts_per_group) &&
-                                 is_finite(scores_with_bias[offset + i])
-                             ? scores_with_bias[offset + i]
-                             : neg_inf<T>();
+          T candidates =
+              (i < num_experts_per_group) && isfinite(cuda_cast<float, T>(
+                                                 scores_with_bias[offset + i]))
+                  ? scores_with_bias[offset + i]
+                  : cuda_cast<T, float>(kNegInfinity);
           queue.add(candidates, offset + i);
         }
         if (group_scores[i_group] == topk_group_value) {
@@ -609,8 +598,7 @@ __global__ void group_idx_and_topk_idx_kernel(
       if (i < topk) {
         s_topk_value[i] = value;
       }
-      topk_sum +=
-          cg::reduce(tile, cuda_cast<float, T>(value), cg::plus<float>());
+      topk_sum += reduce(tile, cuda_cast<float, T>(value), cg::plus<float>());
     }
   }
 

csrc/moe/moe_align_sum_kernels.cu

@@ -44,9 +44,6 @@ __global__ void moe_align_block_size_kernel(
 
   for (size_t i = tid; i < numel; i += stride) {
     int expert_id = topk_ids[i];
-    if (expert_id >= num_experts) {
-      continue;
-    }
     int warp_idx = expert_id / experts_per_warp;
     int expert_offset = expert_id % experts_per_warp;
     atomicAdd(&shared_counts[warp_idx * experts_per_warp + expert_offset], 1);
@@ -98,15 +95,12 @@ template <typename scalar_t>
 __global__ void count_and_sort_expert_tokens_kernel(
     const scalar_t* __restrict__ topk_ids,
     int32_t* __restrict__ sorted_token_ids, int32_t* __restrict__ cumsum_buffer,
-    size_t numel, int32_t num_experts) {
+    size_t numel) {
   const size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
   const size_t stride = blockDim.x * gridDim.x;
 
   for (size_t i = tid; i < numel; i += stride) {
     int32_t expert_id = topk_ids[i];
-    if (expert_id >= num_experts) {
-      continue;
-    }
     int32_t rank_post_pad = atomicAdd(&cumsum_buffer[expert_id], 1);
     sorted_token_ids[rank_post_pad] = i;
   }
@@ -275,7 +269,7 @@ void moe_align_block_size(torch::Tensor topk_ids, int64_t num_experts,
           sort_kernel<<<actual_blocks, block_threads, 0, stream>>>(
               topk_ids.data_ptr<scalar_t>(),
               sorted_token_ids.data_ptr<int32_t>(),
-              cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel(), num_experts);
+              cumsum_buffer.data_ptr<int32_t>(), topk_ids.numel());
         }
       });
 }

csrc/moe/topk_softmax_kernels.cu

@@ -20,7 +20,17 @@
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
 #include "../cuda_compat.h"
-#include "../cub_helpers.h"
+
+#ifndef USE_ROCM
+    #include <cub/util_type.cuh>
+    #include <cub/cub.cuh>
+    #include <cuda/std/functional>
+    using AddOp = cuda::std::plus<float>;
+#else
+    #include <hipcub/util_type.hpp>
+    #include <hipcub/hipcub.hpp>
+    using AddOp = cub::Sum; 
+#endif
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
@@ -69,7 +79,7 @@ __launch_bounds__(TPB) __global__
         threadData = max(static_cast<float>(input[idx]), threadData);
     }
 
-    const float maxElem = BlockReduce(tmpStorage).Reduce(threadData, CubMaxOp());
+    const float maxElem = BlockReduce(tmpStorage).Reduce(threadData, cub::Max());
     if (threadIdx.x == 0)
     {
         float_max = maxElem;
@@ -84,7 +94,7 @@ __launch_bounds__(TPB) __global__
         threadData += exp((static_cast<float>(input[idx]) - float_max));
     }
 
-    const auto Z = BlockReduce(tmpStorage).Reduce(threadData, CubAddOp());
+    const auto Z = BlockReduce(tmpStorage).Reduce(threadData, AddOp());
 
     if (threadIdx.x == 0)
     {

csrc/ops.h

@@ -92,9 +92,6 @@ void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
 void fused_add_rms_norm(torch::Tensor& input, torch::Tensor& residual,
                         torch::Tensor& weight, double epsilon);
 
-void poly_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
-               torch::Tensor& bias, double epsilon);
-
 void apply_repetition_penalties_(torch::Tensor& logits,
                                  const torch::Tensor& prompt_mask,
                                  const torch::Tensor& output_mask,
@@ -122,6 +119,12 @@ void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
                       std::optional<torch::Tensor> key, int64_t head_size,
                       torch::Tensor& cos_sin_cache, bool is_neox);
 
+void batched_rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
+                              std::optional<torch::Tensor> key,
+                              int64_t head_size, torch::Tensor& cos_sin_cache,
+                              bool is_neox, int64_t rot_dim,
+                              torch::Tensor& cos_sin_cache_offsets);
+
 void silu_and_mul(torch::Tensor& out, torch::Tensor& input);
 
 void silu_and_mul_quant(torch::Tensor& out, torch::Tensor& input,
@@ -133,12 +136,6 @@ void silu_and_mul_nvfp4_quant(torch::Tensor& out,
                               torch::Tensor& input,
                               torch::Tensor& input_global_scale);
 #endif
-void silu_mul_fp8_quant_deep_gemm_cuda(
-    const at::Tensor& input,   // (E, T, 2*H)
-    const at::Tensor& counts,  // (E)
-    at::Tensor& y_q,           // (E, T, H) [OUT]
-    at::Tensor& y_s,           // (E, T, H//group_size) [OUT]
-    int64_t group_size, bool use_ue8m0, int64_t num_parallel_tokens);
 
 void mul_and_silu(torch::Tensor& out, torch::Tensor& input);
 
@@ -328,12 +325,6 @@ void selective_scan_fwd(const torch::Tensor& u, const torch::Tensor& delta,
                         const std::optional<torch::Tensor>& has_initial_state,
                         const torch::Tensor& ssm_states, int64_t pad_slot_id);
 
-torch::Tensor dynamic_4bit_int_moe_cpu(
-    torch::Tensor x, torch::Tensor topk_ids, torch::Tensor topk_weights,
-    torch::Tensor w13_packed, torch::Tensor w2_packed, int64_t H, int64_t I,
-    int64_t I2, int64_t group_size, bool apply_router_weight_on_input,
-    int64_t activation_kind);
-
 using fptr_t = int64_t;
 fptr_t init_custom_ar(const std::vector<int64_t>& fake_ipc_ptrs,
                       torch::Tensor& rank_data, int64_t rank,
@@ -353,8 +344,6 @@ std::tuple<int64_t, torch::Tensor> allocate_shared_buffer_and_handle(
 int64_t open_mem_handle(torch::Tensor& mem_handle);
 void free_shared_buffer(int64_t buffer);
 
-torch::Tensor hadacore_transform(torch::Tensor& x, bool inplace);
-
 #ifdef USE_ROCM
 fptr_t init_custom_qr(int64_t rank, int64_t world_size,
                       std::optional<int64_t> qr_max_size = std::nullopt);
@@ -364,4 +353,4 @@ void qr_open_handles(fptr_t _fa, const std::vector<torch::Tensor>& handles);
 void qr_all_reduce(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out,
                    int64_t quant_level, bool cast_bf2half = false);
 int64_t qr_max_size();
-#endif
+#endif

csrc/pos_encoding_kernels.cu

@@ -99,6 +99,35 @@ __global__ void rotary_embedding_kernel(
       token_idx, query_stride, key_stride, head_stride);
 }
 
+template <typename scalar_t, bool IS_NEOX>
+__global__ void batched_rotary_embedding_kernel(
+    const int64_t* __restrict__ positions,  // [batch_size, seq_len] or
+                                            // [num_tokens]
+    scalar_t* __restrict__ query,           // [batch_size, seq_len, num_heads,
+                                   // head_size] or [num_tokens, num_heads,
+                                   // head_size]
+    scalar_t* __restrict__ key,  // nullptr or
+                                 // [batch_size, seq_len, num_kv_heads,
+                                 // head_size] or [num_tokens, num_kv_heads,
+                                 // head_size]
+    const scalar_t* __restrict__ cos_sin_cache,  // [max_position, 2, rot_dim //
+                                                 // 2]
+    const int64_t* __restrict__ cos_sin_cache_offsets,  // [batch_size, seq_len]
+    const int rot_dim, const int64_t query_stride, const int64_t key_stride,
+    const int64_t head_stride, const int num_heads, const int num_kv_heads,
+    const int head_size) {
+  // Each thread block is responsible for one token.
+  const int token_idx = blockIdx.x;
+  int64_t pos = positions[token_idx];
+  int64_t cos_sin_cache_offset = cos_sin_cache_offsets[token_idx];
+  const scalar_t* cache_ptr =
+      cos_sin_cache + (cos_sin_cache_offset + pos) * rot_dim;
+
+  apply_rotary_embedding<scalar_t, IS_NEOX>(
+      query, key, cache_ptr, head_size, num_heads, num_kv_heads, rot_dim,
+      token_idx, query_stride, key_stride, head_stride);
+}
+
 }  // namespace vllm
 
 void rotary_embedding(
@@ -182,3 +211,96 @@ void rotary_embedding(
     }
   });
 }
+
+/*
+Batched version of rotary embedding, pack multiple LoRAs together
+and process in batched manner.
+*/
+void batched_rotary_embedding(
+    torch::Tensor& positions,  // [batch_size, seq_len] or [num_tokens]
+    torch::Tensor& query,  // [batch_size, seq_len, num_heads * head_size] or
+                           // [num_tokens, num_heads * head_size] or
+                           // [batch_size, seq_len, num_heads, head_size] or
+                           // [num_tokens, num_heads, head_size]
+    std::optional<torch::Tensor>
+        key,  // null or
+              // [batch_size, seq_len, num_kv_heads * head_size] or
+              // [num_tokens, num_kv_heads * head_size] or
+              // [batch_size, seq_len, num_heads, head_size] or
+              // [num_tokens, num_heads, head_size]
+    int64_t head_size,
+    torch::Tensor& cos_sin_cache,  // [max_position, rot_dim]
+    bool is_neox, int64_t rot_dim,
+    torch::Tensor& cos_sin_cache_offsets  // [num_tokens] or [batch_size]
+) {
+  // num_tokens = batch_size * seq_len
+  int64_t num_tokens = cos_sin_cache_offsets.size(0);
+  TORCH_CHECK(
+      positions.size(0) == num_tokens || positions.numel() == num_tokens,
+      "positions must have the same num_tokens or batch_size as "
+      "cos_sin_cache_offsets");
+
+  int positions_ndim = positions.dim();
+  // Make sure num_tokens dim is consistent across positions, query, and key
+  TORCH_CHECK(
+      positions_ndim == 1 || positions_ndim == 2,
+      "positions must have shape [num_tokens] or [batch_size, seq_len]");
+  if (positions_ndim == 1) {
+    TORCH_CHECK(query.size(0) == positions.size(0) &&
+                    (!key.has_value() || key->size(0) == positions.size(0)),
+                "query, key and positions must have the same number of tokens");
+  }
+  if (positions_ndim == 2) {
+    TORCH_CHECK(
+        query.size(0) == positions.size(0) &&
+            (!key.has_value() || key->size(0) == positions.size(0)) &&
+            query.size(1) == positions.size(1) &&
+            (!key.has_value() || key->size(1) == positions.size(1)),
+        "query, key and positions must have the same batch_size and seq_len");
+  }
+
+  // Make sure head_size is valid for query and key
+  int query_hidden_size = query.numel() / num_tokens;
+  int key_hidden_size = key.has_value() ? key->numel() / num_tokens : 0;
+  TORCH_CHECK(query_hidden_size % head_size == 0);
+  TORCH_CHECK(key_hidden_size % head_size == 0);
+
+  // Make sure query and key have concistent number of heads
+  int num_heads = query_hidden_size / head_size;
+  int num_kv_heads = key.has_value() ? key_hidden_size / head_size : num_heads;
+  TORCH_CHECK(num_heads % num_kv_heads == 0);
+
+  int seq_dim_idx = positions_ndim - 1;
+  int64_t query_stride = query.stride(seq_dim_idx);
+  int64_t key_stride = key.has_value() ? key->stride(seq_dim_idx) : 0;
+  // Determine head stride: for [*, heads, head_size] use stride of last dim;
+  // for flat [*, heads*head_size], heads blocks are contiguous of size
+  // head_size
+  int query_ndim = query.dim();
+  int64_t head_stride =
+      (query_ndim == positions_ndim + 2) ? query.stride(-2) : head_size;
+
+  dim3 grid(num_tokens);
+  dim3 block(std::min<int64_t>(num_heads * rot_dim / 2, 512));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "rotary_embedding", [&] {
+    if (is_neox) {
+      vllm::batched_rotary_embedding_kernel<scalar_t, true>
+          <<<grid, block, 0, stream>>>(
+              positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
+              key.has_value() ? key->data_ptr<scalar_t>() : nullptr,
+              cos_sin_cache.data_ptr<scalar_t>(),
+              cos_sin_cache_offsets.data_ptr<int64_t>(), rot_dim, query_stride,
+              key_stride, head_stride, num_heads, num_kv_heads, head_size);
+    } else {
+      vllm::batched_rotary_embedding_kernel<scalar_t, false>
+          <<<grid, block, 0, stream>>>(
+              positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
+              key.has_value() ? key->data_ptr<scalar_t>() : nullptr,
+              cos_sin_cache.data_ptr<scalar_t>(),
+              cos_sin_cache_offsets.data_ptr<int64_t>(), rot_dim, query_stride,
+              key_stride, head_stride, num_heads, num_kv_heads, head_size);
+    }
+  });
+}

csrc/quantization/activation_kernels.cu

@@ -9,31 +9,6 @@
 
 #include "quantization/fp8/common.cuh"
 
-#include <c10/util/Float8_e4m3fn.h>
-
-#ifndef USE_ROCM
-  #include <cuda_bf16.h>
-  #include <cuda_fp16.h>
-  #include <cuda_fp8.h>
-#else
-  #include <hip/hip_bf16.h>
-  #include <hip/hip_fp16.h>
-  #include <hip/hip_fp8.h>
-
-typedef __hip_bfloat162 __nv_bfloat162;
-typedef __hip_bfloat16 __nv_bfloat16;
-typedef __hip_bfloat16_raw __nv_bfloat16_raw;
-  #if defined(HIP_FP8_TYPE_OCP)
-typedef __hip_fp8_e4m3 __nv_fp8_e4m3;
-typedef __hip_fp8x4_e4m3 __nv_fp8x4_e4m3;
-  #else
-// ROCm 6.2 fallback: only *_fnuz types exist
-typedef __hip_fp8_e4m3_fnuz __nv_fp8_e4m3;
-typedef __hip_fp8x4_e4m3_fnuz __nv_fp8x4_e4m3;
-  #endif
-#endif
-
-#include "core/registration.h"
 namespace vllm {
 
 template <typename T>
@@ -112,336 +87,6 @@ __global__ void act_and_mul_quant_kernel(
     }
   }
 }
-
-__device__ __forceinline__ float silu(float x) {
-  return (__fdividef(x, (1.f + expf(-x))));
-}
-
-__device__ __forceinline__ float2 silu2(float2 x) {
-  return make_float2(silu(x.x), silu(x.y));
-}
-
-#ifndef USE_ROCM
-__device__ __forceinline__ float warp_max(float v) {
-  static constexpr unsigned FULL_MASK = 0xffffffffu;
-  for (int offset = 1; offset < WARP_SIZE; offset *= 2) {
-    v = fmaxf(v, __shfl_xor_sync(FULL_MASK, v, offset));
-  }
-  return v;
-}
-
-__device__ __forceinline__ __nv_bfloat16 warp_max(__nv_bfloat16 v) {
-  static constexpr unsigned FULL_MASK = 0xffffffffu;
-  for (int offset = 1; offset < WARP_SIZE; offset *= 2) {
-    v = __hmax(v, __shfl_xor_sync(FULL_MASK, v, offset));
-  }
-  return v;
-}
-#endif
-
-template <typename T, typename U>
-__device__ __forceinline__ void cp_async4(T* _smem_ptr, const U* _glob_ptr) {
-#if __CUDACC_VER_MAJOR__ >= 11 && __CUDA_ARCH__ >= 800
-  auto smem_ptr = reinterpret_cast<void*>(_smem_ptr);
-  auto glob_ptr = reinterpret_cast<const void*>(_glob_ptr);
-  const int BYTES = 16;
-  uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
-  asm volatile(
-      "{\n"
-      "   cp.async.cg.shared.global [%0], [%1], %2;\n"
-      "}\n" ::"r"(smem),
-      "l"(glob_ptr), "n"(BYTES));
-#else
-  _smem_ptr[0] = _glob_ptr[0];
-#endif
-}
-
-__device__ __forceinline__ void cp_async_fence() {
-#if __CUDACC_VER_MAJOR__ >= 11 && __CUDA_ARCH__ >= 800
-  asm volatile("cp.async.commit_group;\n" ::);
-#else
-#endif
-}
-
-template <int N>
-__device__ __forceinline__ void cp_async_wait() {
-#if __CUDACC_VER_MAJOR__ >= 11 && __CUDA_ARCH__ >= 800
-  asm volatile("cp.async.wait_group %0;\n" ::"n"(N));
-#else
-#endif
-}
-
-template <>
-__device__ __forceinline__ void cp_async_wait<0>() {
-#if __CUDACC_VER_MAJOR__ >= 11 && __CUDA_ARCH__ >= 800
-  asm volatile("cp.async.wait_all;\n" ::);
-#else
-#endif
-}
-
-__device__ __forceinline__ float clip(float v, float mmin, float mmax) {
-#if __CUDACC_VER_MAJOR__ >= 11 && __CUDA_ARCH__ >= 800
-  return fminf(mmax, fmaxf(v, mmin));
-#else
-#endif
-}
-
-__device__ __forceinline__ __nv_bfloat16 clip(__nv_bfloat16 v,
-                                              __nv_bfloat16 mmin,
-                                              __nv_bfloat16 mmax) {
-  return __hmin(mmax, __hmax(v, mmin));
-}
-
-__device__ __forceinline__ __nv_bfloat162 clip(__nv_bfloat162 v,
-                                               __nv_bfloat162 mmin,
-                                               __nv_bfloat162 mmax) {
-  return __hmin2(mmax, __hmax2(v, mmin));
-}
-
-// We use the following values for fp8 min/max:
-//  __nv_fp8_e4m3 = (-448, +448)
-//  __nv_fp8_e4m3uz = (-240.0, +240.0)
-// It is currently assumed that only
-template <class T>
-constexpr __nv_bfloat16 get_fp8_max() {
-  static_assert(std::is_same_v<T, c10::Float8_e4m3fn> ||
-                std::is_same_v<T, c10::Float8_e4m3fnuz>);
-  if constexpr (std::is_same_v<T, c10::Float8_e4m3fn>) {
-    return __nv_bfloat16(__nv_bfloat16_raw{.x = 17376});
-  } else {
-    return __nv_bfloat16(__nv_bfloat16_raw{.x = 17264});
-  }
-}
-
-template <class T>
-constexpr __nv_bfloat16 get_fp8_min() {
-  static_assert(std::is_same_v<T, c10::Float8_e4m3fn> ||
-                std::is_same_v<T, c10::Float8_e4m3fnuz>);
-  if constexpr (std::is_same_v<T, c10::Float8_e4m3fn>) {
-    return __nv_bfloat16(__nv_bfloat16_raw{.x = 50144});
-  } else {
-    return __nv_bfloat16(__nv_bfloat16_raw{.x = 50032});
-  }
-}
-#ifndef USE_ROCM
-template <typename fp8_type, int32_t NUM_WARPS, typename Idx_t,
-          int NUM_PARALLEL_TOKENS, bool USE_UE8M0, int GROUP_SIZE = 128,
-          int NUM_STAGES = 3>
-__global__ void silu_mul_fp8_quant_deep_gemm_kernel(
-    const __nv_bfloat16* __restrict__ _input, fp8_type* __restrict__ _y_q,
-    float* __restrict__ _y_s, const int32_t* __restrict__ counts,
-
-    // sizes
-    int H, int G,
-
-    // strides (in elements)
-    Idx_t stride_i_e, Idx_t stride_i_t, Idx_t stride_i_h, Idx_t stride_yq_e,
-    Idx_t stride_yq_t, Idx_t stride_yq_h, Idx_t stride_ys_e, Idx_t stride_ys_t,
-    Idx_t stride_ys_g, Idx_t stride_counts_e) {
-  static constexpr __nv_bfloat16 fp8_min = get_fp8_min<fp8_type>();
-  static constexpr __nv_bfloat16 fp8_max = get_fp8_max<fp8_type>();
-  // We assign EPS with its 16-bit unsigned counterpart to allow constexpr.
-  static constexpr __nv_bfloat16 EPS = (__nv_bfloat16_raw{.x = 11996});
-
-  // We pack 8 16-bit bfloat16 values into a 128-bit __int128_t.
-  static constexpr int32_t BFLOAT16_PER_GROUP = 8;
-
-  // We split the shared memory in half, corresponding to gate and up matrices:
-  // [...gate_i, ...up_i]  where 0 <= i < stages.
-  static constexpr int32_t S_NUM_128 =
-      2u * (GROUP_SIZE / BFLOAT16_PER_GROUP) * NUM_WARPS * NUM_STAGES;
-  static constexpr auto THREAD_COUNT = NUM_WARPS * WARP_SIZE;
-  static constexpr int HALF_THREAD_COUNT = THREAD_COUNT / 2;
-  static constexpr int32_t S_NUM_64 = S_NUM_128 * 2;
-  __shared__ __int128_t __align__(16) s_buff_128[S_NUM_128];
-
-  const int32_t tid = threadIdx.x;
-  const int32_t warp_id = tid / WARP_SIZE;
-  const int32_t lane_id = tid % WARP_SIZE;
-
-  auto s_buff_compute_32 = reinterpret_cast<__nv_bfloat162*>(s_buff_128);
-
-  // block handles one (expert e, group g)
-  int32_t pid = blockIdx.x;
-  int32_t e = pid / G;
-  int32_t g = pid % G;
-
-  const int32_t n_tokens = counts[e * stride_counts_e];
-
-  if (!n_tokens) {
-    return;  // Exit ASAP.
-  }
-
-  const Idx_t stride_i_t_128 = stride_i_t / 8u;
-
-  int32_t n_tokens_lower, n_tokens_upper;
-
-  // Each block i iterates over tokens of a slice of n_tokens =
-  // expert_counts[i], with the size of chunk being
-  // (n_tokens / NUM_PARALLEL_TOKENS) + residual, instead of
-  // updiv(n_tokens, NUM_PARALLEL_TOKENS) for better scheduling.
-  if (n_tokens < NUM_PARALLEL_TOKENS && blockIdx.y < n_tokens) {
-    // Specialize this, but can be likely fused.
-    if (blockIdx.y >= NUM_PARALLEL_TOKENS) {
-      return;
-    }
-    n_tokens_lower = blockIdx.y;
-    n_tokens_upper = blockIdx.y + 1;
-  } else {
-    auto chunk_size = n_tokens / NUM_PARALLEL_TOKENS;
-    auto residual = n_tokens - chunk_size * NUM_PARALLEL_TOKENS;
-    auto calc_id = [&](int32_t id) {
-      if (id < residual) {
-        return min(n_tokens, id * (chunk_size + 1));
-      } else {
-        return min(n_tokens, id * chunk_size + residual);
-      }
-    };
-    n_tokens_lower = calc_id(blockIdx.y);
-    n_tokens_upper = calc_id(blockIdx.y + 1);
-  }
-
-  if (n_tokens_lower >= n_tokens_upper) {
-    return;
-  }
-
-  // We do calculations here, using constexpr wherever possible.
-  const Idx_t base_i = e * stride_i_e + NUM_WARPS * g * GROUP_SIZE * stride_i_h;
-  const Idx_t base_ys = e * stride_ys_e + NUM_WARPS * g * stride_ys_g;
-  const Idx_t base_yq =
-      e * stride_yq_e + NUM_WARPS * g * GROUP_SIZE * stride_yq_h;
-  Idx_t gate_off_128 = (base_i / static_cast<Idx_t>(8u));
-  auto input_128_ptr = reinterpret_cast<const __int128_t*>(_input);
-  auto gate_128_ptr = input_128_ptr + gate_off_128 + (tid % HALF_THREAD_COUNT) +
-                      stride_i_t_128 * n_tokens_lower;
-  auto up_128_ptr = gate_128_ptr + (H * stride_i_h) / 8u;
-  auto y_s_ptr =
-      _y_s + base_ys + warp_id * stride_ys_g + n_tokens_lower * stride_ys_t;
-  auto y_q_ptr = _y_q + base_yq + warp_id * GROUP_SIZE +
-                 stride_yq_t * n_tokens_lower + 4 * lane_id;
-  int32_t t_load = n_tokens_lower, load_stage_id = 0;
-  auto s_buff_gate_load_128 = s_buff_128 + (tid % HALF_THREAD_COUNT);
-  auto s_buff_up_load_128 = s_buff_gate_load_128 + S_NUM_128 / 2u;
-  int32_t stage_offset{};
-
-  static constexpr int32_t LOAD_STAGE_SIZE = (NUM_WARPS * WARP_SIZE / 2);
-  static constexpr int32_t LOAD_STAGE_MOD =
-      NUM_STAGES * (NUM_WARPS * WARP_SIZE / 2);
-
-  // Two halves of all threads in a block conduct global loads for gate and up,
-  // repsectively.
-  auto load_and_advance_y_pred = [&] {
-    if (t_load < n_tokens_upper) {
-      auto s_gate_stage_128_staged_ptr = s_buff_gate_load_128 + stage_offset;
-      auto s_up_stage_128_staged_ptr = s_buff_up_load_128 + stage_offset;
-
-      // It is very important that LOAD_STAGE_SIZE is constexpr to avoid
-      // unnecessary ALU ops.
-      stage_offset += LOAD_STAGE_SIZE;
-      stage_offset %= LOAD_STAGE_MOD;
-
-      if (tid < HALF_THREAD_COUNT) {
-        cp_async4(s_gate_stage_128_staged_ptr, gate_128_ptr);
-        gate_128_ptr += stride_i_t_128;
-      } else {
-        cp_async4(s_up_stage_128_staged_ptr, up_128_ptr);
-        up_128_ptr += stride_i_t_128;
-      }
-      ++t_load;
-      ++load_stage_id;
-    }
-    // We fence even if there is nothing to load to simplify pipelining.
-    cp_async_fence();
-  };
-
-  #pragma unroll
-  for (int i = 0; i < NUM_STAGES - 1; i++) {
-    load_and_advance_y_pred();
-  }
-
-  __int64_t* s_gate_ptr = reinterpret_cast<__int64_t*>(
-                              s_buff_compute_32 + warp_id * (GROUP_SIZE / 2)) +
-                          lane_id;
-  __int64_t* s_up_ptr = s_gate_ptr + S_NUM_64 / 2;
-
-  static constexpr int32_t STAGE_SIZE = (GROUP_SIZE * NUM_WARPS) / 4u;
-  static constexpr int32_t STAGE_MOD = STAGE_SIZE * NUM_STAGES;
-
-  int32_t compute_pipeline_offset_64 = 0;
-
-  for (int32_t t = n_tokens_lower; t < n_tokens_upper; ++t) {
-    __nv_bfloat162 results_bf162[2];
-
-    cp_async_wait<NUM_STAGES - 2>();
-    __syncthreads();
-
-    // We double-buffer pipelined loads so that the next load will
-    // concurrently run with compute without overwrites.
-    load_and_advance_y_pred();
-
-    auto s_gate_compute_64 = s_gate_ptr + compute_pipeline_offset_64;
-    auto s_up_compute_64 = s_up_ptr + compute_pipeline_offset_64;
-
-    // STAGE_SIZE must also be constexpr!
-    compute_pipeline_offset_64 += STAGE_SIZE;
-    compute_pipeline_offset_64 %= STAGE_MOD;
-
-    // Each thread loads (gate/up) 2X 4X bfloat16 values into registers.
-    __int64_t gate64 = *s_gate_compute_64;
-    __nv_bfloat162* s_gate_compute_32 =
-        reinterpret_cast<__nv_bfloat162*>(&gate64);
-
-    __int64_t up64 = *s_up_compute_64;
-    __nv_bfloat162* s_up_compute_32 = reinterpret_cast<__nv_bfloat162*>(&up64);
-
-  #pragma unroll
-    for (int i = 0; i < 2; i++) {
-      // For silu, we make sure that div is emitted.
-      float2 gate = silu2(__bfloat1622float2(s_gate_compute_32[i]));
-      results_bf162[i] = __float22bfloat162_rn(gate);
-    }
-
-  #pragma unroll
-    for (int i = 0; i < 2; i++) {
-      results_bf162[i] = __hmul2(results_bf162[i], s_up_compute_32[i]);
-    }
-
-    auto _y_max2 =
-        __hmax2(__habs2(results_bf162[0]), __habs2(results_bf162[1]));
-
-    __nv_bfloat16 y_max_bf16 = __hmax(EPS, __hmax(_y_max2.x, _y_max2.y));
-
-    // An entire group is assigned to a single warp, so a simple warp reduce
-    // is used.
-    __nv_bfloat16 y_s = warp_max(y_max_bf16) / fp8_max;
-
-    if constexpr (USE_UE8M0) {
-      y_s = hexp2(hceil(hlog2(y_s)));
-    }
-
-    auto inv_y = __float2bfloat16_rn(1.f) / y_s;
-
-    auto y_s2 = make_bfloat162(inv_y, inv_y);
-
-  #pragma unroll
-    for (int32_t i = 0; i < 2; ++i) {
-      results_bf162[i] =
-          clip(__hmul2(results_bf162[i], y_s2), __bfloat162bfloat162(fp8_min),
-               __bfloat162bfloat162(fp8_max));
-    }
-
-    auto fp8x4 = __nv_fp8x4_e4m3(results_bf162[0], results_bf162[1]);
-    *reinterpret_cast<__nv_fp8x4_e4m3*>(y_q_ptr) = fp8x4;
-    y_q_ptr += stride_yq_t;
-
-    if (lane_id == 0) {
-      *y_s_ptr = y_s;
-      y_s_ptr += stride_ys_t;
-    }
-  }
-}
-#endif
-
 }  // namespace vllm
 
 // Launch activation, gating, and quantize kernel.
@@ -474,117 +119,3 @@ void silu_and_mul_quant(torch::Tensor& out,    // [..., d]
   TORCH_CHECK(input.size(-1) % 2 == 0);
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel);
 }
-
-void silu_mul_fp8_quant_deep_gemm_cuda(
-    const at::Tensor& input,   // (E, T, 2*H)
-    const at::Tensor& counts,  // (E)
-    at::Tensor& y_q,           // (E, T, H) [OUT]
-    at::Tensor& y_s,           // (E, T, H//group_size) [OUT]
-    int64_t group_size, bool use_ue8m0, int64_t num_parallel_tokens) {
-#ifndef USE_ROCM
-  // This kernel relies heavily on cp.async and fp8 support.
-  // This kernel currently only supports H % 128 == 0 and assumes a
-  // fixed GROUP_SIZE of 128.
-  TORCH_CHECK(input.dtype() == torch::kBFloat16);
-  TORCH_CHECK(y_q.dtype() == torch::kFloat8_e4m3fn ||
-              y_q.dtype() == torch::kFloat8_e4m3fnuz);
-  TORCH_CHECK(y_s.dtype() == torch::kFloat32);
-  TORCH_CHECK(input.size(-1) % 256 == 0);
-
-  // Check that num_parallel_tokens is of power of 2 and between 1 and 64.
-  TORCH_CHECK(1 <= num_parallel_tokens && num_parallel_tokens <= 64);
-  TORCH_CHECK(!(num_parallel_tokens & (num_parallel_tokens - 1)));
-
-  using Idx_t = int64_t;
-
-  Idx_t E = input.size(0);
-  Idx_t T = input.size(1);
-  Idx_t H = input.size(2) / 2;
-  Idx_t stride_i_e = input.stride(0);
-  Idx_t stride_i_t = input.stride(1);
-  Idx_t stride_i_h = input.stride(2);
-  Idx_t stride_yq_e = y_q.stride(0);
-  Idx_t stride_yq_t = y_q.stride(1);
-  Idx_t stride_yq_h = y_q.stride(2);
-  Idx_t stride_ys_e = y_s.stride(0);
-  Idx_t stride_ys_t = y_s.stride(1);
-  Idx_t stride_ys_g = y_s.stride(2);
-
-  Idx_t stride_counts_e = counts.stride(0);
-
-  static constexpr int GROUP_SIZE = 128;
-
-  #define KERNEL_FN                                                         \
-    if (use_ue8m0) {                                                        \
-      vllm::silu_mul_fp8_quant_deep_gemm_kernel<fp8_t, NUM_WARPS, Idx_t,    \
-                                                NUM_PARALLEL_TOKENS, true>  \
-          <<<grid, block, 0, stream>>>(                                     \
-              reinterpret_cast<__nv_bfloat16*>(input.data_ptr()),           \
-              (fp8_t*)y_q.data_ptr(), y_s.data_ptr<float>(),                \
-              reinterpret_cast<int32_t*>(counts.data_ptr<int>()), H, G,     \
-              stride_i_e, stride_i_t, stride_i_h, stride_yq_e, stride_yq_t, \
-              stride_yq_h, stride_ys_e, stride_ys_t, stride_ys_g,           \
-              stride_counts_e);                                             \
-    } else {                                                                \
-      vllm::silu_mul_fp8_quant_deep_gemm_kernel<fp8_t, NUM_WARPS, Idx_t,    \
-                                                NUM_PARALLEL_TOKENS, false> \
-          <<<grid, block, 0, stream>>>(                                     \
-              reinterpret_cast<__nv_bfloat16*>(input.data_ptr()),           \
-              (fp8_t*)y_q.data_ptr(), y_s.data_ptr<float>(),                \
-              reinterpret_cast<int32_t*>(counts.data_ptr<int>()), H, G,     \
-              stride_i_e, stride_i_t, stride_i_h, stride_yq_e, stride_yq_t, \
-              stride_yq_h, stride_ys_e, stride_ys_t, stride_ys_g,           \
-              stride_counts_e);                                             \
-    }
-
-  #define KERNEL_CALL_H                                       \
-    if (H % (4 * GROUP_SIZE) == 0) {                          \
-      static constexpr int NUM_WARPS = 4;                     \
-      populate_launch_params(NUM_WARPS, NUM_PARALLEL_TOKENS); \
-      KERNEL_FN                                               \
-    } else {                                                  \
-      static constexpr int NUM_WARPS = 1;                     \
-      populate_launch_params(NUM_WARPS, NUM_PARALLEL_TOKENS); \
-      KERNEL_FN                                               \
-    }
-
-  #define KERNEL_CALL_TOP_LEVEL                      \
-    if (num_parallel_tokens == 1) {                  \
-      static constexpr int NUM_PARALLEL_TOKENS = 1;  \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 2) {           \
-      static constexpr int NUM_PARALLEL_TOKENS = 2;  \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 4) {           \
-      static constexpr int NUM_PARALLEL_TOKENS = 4;  \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 8) {           \
-      static constexpr int NUM_PARALLEL_TOKENS = 8;  \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 16) {          \
-      static constexpr int NUM_PARALLEL_TOKENS = 16; \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 32) {          \
-      static constexpr int NUM_PARALLEL_TOKENS = 32; \
-      KERNEL_CALL_H                                  \
-    } else if (num_parallel_tokens == 64) {          \
-      static constexpr int NUM_PARALLEL_TOKENS = 64; \
-      KERNEL_CALL_H                                  \
-    }
-
-  Idx_t G;
-  dim3 block, grid;
-  auto populate_launch_params = [&](int num_warps, int _num_parallel_tokens) {
-    G = H / Idx_t(group_size * num_warps);
-    grid = dim3(E * G, _num_parallel_tokens);
-    block = dim3(num_warps * WARP_SIZE);
-  };
-
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
-  VLLM_DISPATCH_FP8_TYPES(y_q.scalar_type(),
-                          "silu_mul_fp8_quant_deep_gemm_kernel",
-                          [&] { KERNEL_CALL_TOP_LEVEL });
-
-#endif
-}

csrc/quantization/compressed_tensors/int8_quant_kernels.cu

@@ -7,10 +7,17 @@
 
 #include <cmath>
 
-#include "../../cub_helpers.h"
 #include "../../dispatch_utils.h"
 #include "../vectorization_utils.cuh"
 
+#ifndef USE_ROCM
+  #include <cub/cub.cuh>
+  #include <cub/util_type.cuh>
+#else
+  #include <hipcub/hipcub.hpp>
+  #include <hipcub/util_type.hpp>
+#endif
+
 static inline __device__ int8_t float_to_int8_rn(float x) {
 #ifdef USE_ROCM
   static constexpr auto i8_min =
@@ -166,7 +173,7 @@ __global__ void dynamic_scaled_int8_quant_kernel(
       });
   using BlockReduce = cub::BlockReduce<float, 256>;
   __shared__ typename BlockReduce::TempStorage tmp;
-  float block_max = BlockReduce(tmp).Reduce(thread_max, CubMaxOp{}, blockDim.x);
+  float block_max = BlockReduce(tmp).Reduce(thread_max, cub::Max{}, blockDim.x);
   __shared__ float absmax;
   if (tid == 0) {
     absmax = block_max;

csrc/quantization/cutlass_w4a8/w4a8_mm_entry.cu

@@ -25,8 +25,6 @@
 #include "cutlass_extensions/common.hpp"
 #include "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp"
 
-#include <cuda_runtime.h>
-
 namespace vllm::cutlass_w4a8 {
 
 using namespace cute;
@@ -395,71 +393,6 @@ torch::Tensor pack_scale_fp8(torch::Tensor const& scales) {
   return packed_scales;
 }
 
-/*
-  GPU-accelerated implementation of cutlass::unified_encode_int4b.
-  Constructs a lookup table in constant memory to map 8 bits
-  (two 4-bit values) at a time. Assumes memory is contiguous
-  and pointers are 16-byte aligned.
-*/
-__constant__ uint8_t kNibbleLUT[256];
-
-__global__ void unified_encode_int4b_device(const uint8_t* in, uint8_t* out,
-                                            size_t nbytes) {
-  constexpr size_t V = sizeof(uint4);  // 16 bytes
-  const size_t tid = blockIdx.x * blockDim.x + threadIdx.x;
-  const size_t nthreads = size_t(gridDim.x) * blockDim.x;
-  const size_t nvec = nbytes / V;
-
-  // 1-D grid-stride loop over 16-byte chunks
-  for (size_t vec = tid; vec < nvec; vec += nthreads) {
-    uint4 v = reinterpret_cast<const uint4*>(in)[vec];
-    uint8_t* b = reinterpret_cast<uint8_t*>(&v);
-#pragma unroll
-    for (int i = 0; i < int(V); ++i) b[i] = kNibbleLUT[b[i]];
-    reinterpret_cast<uint4*>(out)[vec] = v;
-  }
-}
-
-static bool upload_lut() {
-  std::array<uint8_t, 256> lut{};
-  auto map_nib = [](uint8_t v) -> uint8_t {
-    // 1..7 -> (8 - v); keep 0 and 8..15
-    return (v == 0 || (v & 0x8)) ? v : uint8_t(8 - v);
-  };
-  for (int b = 0; b < 256; ++b) {
-    uint8_t lo = b & 0xF;
-    uint8_t hi = (b >> 4) & 0xF;
-    lut[b] = uint8_t((map_nib(hi) << 4) | map_nib(lo));
-  }
-  cudaError_t e = cudaMemcpyToSymbol(kNibbleLUT, lut.data(), lut.size(),
-                                     /*offset=*/0, cudaMemcpyHostToDevice);
-
-  return (e == cudaSuccess);
-}
-
-static bool unified_encode_int4b(cutlass::int4b_t const* in,
-                                 cutlass::int4b_t* out, size_t num_int4_elems) {
-  // Build/upload LUT
-  if (!upload_lut()) return false;
-
-  static_assert(sizeof(typename cutlass::int4b_t::Storage) == 1,
-                "int4 storage must be 1 byte");
-  const size_t nbytes = num_int4_elems >> 1;
-
-  auto* in_bytes = reinterpret_cast<uint8_t const*>(in);
-  auto* out_bytes = reinterpret_cast<uint8_t*>(out);
-
-  // kernel launch params
-  constexpr int block = 256;
-  const size_t nvec = nbytes / sizeof(uint4);  // # of 16B vectors
-  int grid = int((nvec + block - 1) / block);
-  if (grid == 0) grid = 1;  // ensure we still cover the tail in the kernel
-
-  unified_encode_int4b_device<<<grid, block>>>(in_bytes, out_bytes, nbytes);
-  cudaError_t err = cudaGetLastError();
-  return (err == cudaSuccess);
-}
-
 torch::Tensor encode_and_reorder_int4b(torch::Tensor const& B) {
   TORCH_CHECK(B.dtype() == torch::kInt32);
   TORCH_CHECK(B.dim() == 2);
@@ -468,7 +401,6 @@ torch::Tensor encode_and_reorder_int4b(torch::Tensor const& B) {
 
   int k = B.size(0) * PackFactor;  // logical k
   int n = B.size(1);
-  TORCH_CHECK((n * k) % 32 == 0, "need multiples of 32 int4s for 16B chunks");
 
   auto B_ptr = static_cast<QuantType const*>(B.const_data_ptr());
   auto B_packed_ptr = static_cast<QuantType*>(B_packed.data_ptr());
@@ -477,9 +409,7 @@ torch::Tensor encode_and_reorder_int4b(torch::Tensor const& B) {
   LayoutB_Reordered layout_B_reordered =
       cute::tile_to_shape(LayoutAtomQuant{}, shape_B);
 
-  bool ok =
-      vllm::cutlass_w4a8::unified_encode_int4b(B_ptr, B_packed_ptr, n * k);
-  TORCH_CHECK(ok, "unified_encode_int4b failed");
+  cutlass::unified_encode_int4b(B_ptr, B_packed_ptr, n * k);
   cutlass::reorder_tensor(B_packed_ptr, layout_B, layout_B_reordered);
 
   return B_packed;

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm100_fp8_dispatch.cuh

@@ -14,6 +14,9 @@
 #include "cutlass/epilogue/dispatch_policy.hpp"
 #include "cutlass/epilogue/collective/collective_builder.hpp"
 
+#include "cutlass_extensions/gemm/dispatch_policy.hpp"
+#include "cutlass_extensions/gemm/collective/collective_builder.hpp"
+
 #include "cutlass_gemm_caller.cuh"
 
 namespace vllm {
@@ -146,7 +149,6 @@ void cutlass_gemm_caller_blockwise(torch::Tensor& out, torch::Tensor const& a,
 
   using ElementAB = typename Gemm::ElementAB;
   using ElementD = typename Gemm::ElementD;
-  using ElementBlockScale = typename Gemm::ElementBlockScale;
 
   int32_t m = a.size(0), n = b.size(1), k = a.size(1);
 
@@ -167,29 +169,26 @@ void cutlass_gemm_caller_blockwise(torch::Tensor& out, torch::Tensor const& a,
       ScaleConfig::tile_atom_to_shape_SFB(make_shape(n, m, k, 1)) :
       ScaleConfig::tile_atom_to_shape_SFB(make_shape(m, n, k, 1));
 
-  auto a_ptr = static_cast<ElementAB const*>(a.data_ptr());
-  auto b_ptr = static_cast<ElementAB const*>(b.data_ptr());
-  auto a_scales_ptr = static_cast<ElementBlockScale const*>(a_scales.data_ptr());
-  auto b_scales_ptr = static_cast<ElementBlockScale const*>(b_scales.data_ptr());
+  auto a_ptr = static_cast<ElementAB*>(a.data_ptr());
+  auto b_ptr = static_cast<ElementAB*>(b.data_ptr());
+  auto a_scales_ptr = static_cast<float*>(a_scales.data_ptr());
+  auto b_scales_ptr = static_cast<float*>(b_scales.data_ptr());
 
-  typename GemmKernel::MainloopArguments mainloop_args{};
-  mainloop_args.layout_SFA = layout_SFA;
-  mainloop_args.layout_SFB = layout_SFB;
-  if (swap_ab) {
-    mainloop_args.ptr_A = b_ptr;
-    mainloop_args.dA = b_stride;
-    mainloop_args.ptr_B = a_ptr;
-    mainloop_args.dB = a_stride;
-    mainloop_args.ptr_SFA = b_scales_ptr;
-    mainloop_args.ptr_SFB = a_scales_ptr;
-  } else {
-    mainloop_args.ptr_A = a_ptr;
-    mainloop_args.dA = a_stride;
-    mainloop_args.ptr_B = b_ptr;
-    mainloop_args.dB = b_stride;
-    mainloop_args.ptr_SFA = a_scales_ptr;
-    mainloop_args.ptr_SFB = b_scales_ptr;
-  }
+  auto mainloop_args = [&](){
+    // layout_SFA and layout_SFB cannot be swapped since they are deduced.
+    if (swap_ab) {
+      return typename GemmKernel::MainloopArguments{
+          b_ptr,        b_stride,   a_ptr,        a_stride,
+          b_scales_ptr, layout_SFA, a_scales_ptr, layout_SFB
+      };
+    }
+    else {
+      return typename GemmKernel::MainloopArguments{
+          a_ptr,        a_stride,   b_ptr,        b_stride,
+          a_scales_ptr, layout_SFA, b_scales_ptr, layout_SFB
+      };
+    }
+  }();
   auto prob_shape = swap_ab ? cute::make_shape(n, m, k, 1) : cute::make_shape(m, n, k, 1);
 
   auto c_ptr = static_cast<ElementD*>(out.data_ptr());

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm120_fp8_dispatch.cuh

@@ -14,6 +14,9 @@
 #include "cutlass/epilogue/dispatch_policy.hpp"
 #include "cutlass/epilogue/collective/collective_builder.hpp"
 
+#include "cutlass_extensions/gemm/dispatch_policy.hpp"
+#include "cutlass_extensions/gemm/collective/collective_builder.hpp"
+
 #include "cutlass_gemm_caller.cuh"
 
 namespace vllm {
@@ -125,7 +128,6 @@ void cutlass_gemm_caller_blockwise(torch::Tensor& out, torch::Tensor const& a,
 
   using ElementAB = typename Gemm::ElementAB;
   using ElementD = typename Gemm::ElementD;
-  using ElementBlockScale = typename Gemm::ElementBlockScale;
 
   int32_t m = a.size(0), n = b.size(1), k = a.size(1);
 
@@ -144,20 +146,17 @@ void cutlass_gemm_caller_blockwise(torch::Tensor& out, torch::Tensor const& a,
   LayoutSFB layout_SFB = 
       ScaleConfig::tile_atom_to_shape_SFB(make_shape(m, n, k, 1));
 
-  auto a_ptr = static_cast<ElementAB const*>(a.data_ptr());
-  auto b_ptr = static_cast<ElementAB const*>(b.data_ptr());
-  auto a_scales_ptr = static_cast<ElementBlockScale const*>(a_scales.data_ptr());
-  auto b_scales_ptr = static_cast<ElementBlockScale const*>(b_scales.data_ptr());
+  auto a_ptr = static_cast<ElementAB*>(a.data_ptr());
+  auto b_ptr = static_cast<ElementAB*>(b.data_ptr());
+  auto a_scales_ptr = static_cast<float*>(a_scales.data_ptr());
+  auto b_scales_ptr = static_cast<float*>(b_scales.data_ptr());
 
-  typename GemmKernel::MainloopArguments mainloop_args{};
-  mainloop_args.ptr_A = a_ptr;
-  mainloop_args.dA = a_stride;
-  mainloop_args.ptr_B = b_ptr;
-  mainloop_args.dB = b_stride;
-  mainloop_args.ptr_SFA = a_scales_ptr;
-  mainloop_args.layout_SFA = layout_SFA;
-  mainloop_args.ptr_SFB = b_scales_ptr;
-  mainloop_args.layout_SFB = layout_SFB;
+  auto mainloop_args = [&](){
+    return typename GemmKernel::MainloopArguments{
+        a_ptr,        a_stride,   b_ptr,        b_stride,
+        a_scales_ptr, layout_SFA, b_scales_ptr, layout_SFB
+    };
+  }();
   auto prob_shape = cute::make_shape(m, n, k, 1);
 
   auto c_ptr = static_cast<ElementD*>(out.data_ptr());

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8_dispatch.cuh

@@ -13,18 +13,27 @@
 #include "cutlass/epilogue/dispatch_policy.hpp"
 #include "cutlass/epilogue/collective/collective_builder.hpp"
 
+#include "cutlass_extensions/gemm/dispatch_policy.hpp"
+#include "cutlass_extensions/gemm/collective/collective_builder.hpp"
+
 #include "cutlass_gemm_caller.cuh"
 
 namespace vllm {
 
 using namespace cute;
 
-// clang-format off
-template <class OutType, int ScaleGranularityM,
-          int ScaleGranularityN, int ScaleGranularityK,
-          class MmaTileShape, class ClusterShape,
-          class EpilogueScheduler, class MainloopScheduler>
+template <typename SchedulerType, typename OutType, int GroupSizeM_,
+          int GroupSizeN_, int GroupSizeK_, int TileSizeM_ = 128,
+          class ClusterShape = Shape<_1, _2, _1>>
 struct cutlass_3x_gemm_fp8_blockwise {
+  using GroupSizeM = Int<GroupSizeM_>;
+  using GroupSizeN = Int<GroupSizeN_>;
+  using GroupSizeK = Int<GroupSizeK_>;
+  using TileSizeM = Int<TileSizeM_>;
+
+  static_assert(TileSizeM_ % GroupSizeM_ == 0,
+                "TileSizeM must be a multiple of GroupSizeM");
+
   using ElementAB = cutlass::float_e4m3_t;
 
   using ElementA = ElementAB;
@@ -36,67 +45,52 @@ struct cutlass_3x_gemm_fp8_blockwise {
   static constexpr int AlignmentB = 128 / cutlass::sizeof_bits<ElementB>::value;
 
   using ElementD = OutType;
-  using LayoutD = cutlass::layout::RowMajor;
+  using StrideD = Stride<int64_t, Int<1>, Int<0>>;
   static constexpr int AlignmentD = 128 / cutlass::sizeof_bits<ElementD>::value;
 
-  using ElementC = void; // TODO: support bias
-  using LayoutC = LayoutD;
+  using ElementC = void;
+  using StrideC = StrideD;
   static constexpr int AlignmentC = AlignmentD;
 
   using ElementAccumulator = float;
-  using ElementCompute = float;
   using ElementBlockScale = float;
-
-  using ScaleConfig = cutlass::detail::Sm90BlockwiseScaleConfig<
-        ScaleGranularityM, ScaleGranularityN, ScaleGranularityK>;
-
-  using LayoutSFA = decltype(ScaleConfig::deduce_layoutSFA());
-  using LayoutSFB = decltype(ScaleConfig::deduce_layoutSFB());
-
+  using ElementCompute = float;
   using ArchTag = cutlass::arch::Sm90;
   using OperatorClass = cutlass::arch::OpClassTensorOp;
+  using TileShape = Shape<TileSizeM, GroupSizeN, GroupSizeK>;
 
-  static constexpr auto RoundStyle = cutlass::FloatRoundStyle::round_to_nearest;
-  using ElementScalar = float;
-  using DefaultOperation = cutlass::epilogue::fusion::LinearCombination<ElementD, ElementCompute, ElementC, ElementScalar, RoundStyle>;
-  using CollectiveEpilogue = typename cutlass::epilogue::collective::CollectiveBuilder<
-      ArchTag,
-      OperatorClass,
-      MmaTileShape,
-      ClusterShape,
-      cutlass::epilogue::collective::EpilogueTileAuto,
-      ElementAccumulator,
-      ElementCompute,
-      ElementC,
-      LayoutC,
-      AlignmentC,
-      ElementD,
-      LayoutD,
-      AlignmentD,
-      EpilogueScheduler,
-      DefaultOperation
-  >::CollectiveOp;
+  using KernelSchedule = cutlass::gemm::
+      KernelTmaWarpSpecializedCooperativeFP8BlockScaledSubGroupMAccum<
+          GroupSizeM_>;
+  using EpilogueSchedule = cutlass::epilogue::TmaWarpSpecializedCooperative;
+  using EpilogueTileType = cutlass::epilogue::collective::EpilogueTileAuto;
 
-  using CollectiveMainloop = typename cutlass::gemm::collective::CollectiveBuilder<
-      ArchTag,
-      OperatorClass,
-      ElementA,
-      cute::tuple<LayoutA, LayoutSFA>,
-      AlignmentA,
-      ElementB,
-      cute::tuple<LayoutB, LayoutSFB>,
-      AlignmentB,
-      ElementAccumulator,
-      MmaTileShape,
-      ClusterShape,
-      cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(sizeof(typename CollectiveEpilogue::SharedStorage))>,
-      MainloopScheduler
-  >::CollectiveOp;
+  using StoreEpilogueCompute = typename cutlass::epilogue::fusion::Sm90EVT<
+      cutlass::epilogue::fusion::Sm90AccFetch>;
+
+  using CollectiveEpilogue =
+      typename cutlass::epilogue::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, TileShape, ClusterShape, EpilogueTileType,
+          ElementAccumulator, ElementCompute, ElementC, StrideC, AlignmentC,
+          ElementD, StrideD, AlignmentD, EpilogueSchedule,
+          StoreEpilogueCompute>::CollectiveOp;
+
+  using CollectiveMainloop =
+      typename cutlass::gemm::collective::CollectiveBuilder<
+          ArchTag, OperatorClass, ElementA, LayoutA, AlignmentA, ElementB,
+          LayoutB, AlignmentB, ElementAccumulator, TileShape, ClusterShape,
+          cutlass::gemm::collective::StageCountAutoCarveout<static_cast<int>(
+              sizeof(typename CollectiveEpilogue::SharedStorage))>,
+          KernelSchedule>::CollectiveOp;
 
   using KernelType = enable_sm90_or_later<cutlass::gemm::kernel::GemmUniversal<
-      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue>>;
+      Shape<int, int, int, int>, CollectiveMainloop, CollectiveEpilogue,
+      SchedulerType>>;
 
   struct GemmKernel : public KernelType {};
+
+  using StrideA = typename GemmKernel::StrideA;
+  using StrideB = typename GemmKernel::StrideB;
 };
 
 template <typename Gemm>
@@ -105,58 +99,76 @@ void cutlass_gemm_caller_blockwise(torch::Tensor& out, torch::Tensor const& a,
                                    torch::Tensor const& a_scales,
                                    torch::Tensor const& b_scales) {
   using GemmKernel = typename Gemm::GemmKernel;
-  using StrideA = typename Gemm::GemmKernel::StrideA;
-  using StrideB = typename Gemm::GemmKernel::StrideB;
-  using StrideD = typename Gemm::GemmKernel::StrideD;
-  using StrideC = typename Gemm::GemmKernel::StrideC;
-  using LayoutSFA = typename Gemm::LayoutSFA;
-  using LayoutSFB = typename Gemm::LayoutSFB;
-  using ScaleConfig = typename Gemm::ScaleConfig;
 
   using ElementAB = typename Gemm::ElementAB;
   using ElementD = typename Gemm::ElementD;
-  using ElementBlockScale = typename Gemm::ElementBlockScale;
 
-  int32_t m = a.size(0), n = b.size(1), k = a.size(1);
+  auto prob_shape = c3x::get_problem_shape(a, b);
+  int32_t m = get<0>(prob_shape), n = get<1>(prob_shape),
+          k = get<2>(prob_shape);
 
-  TORCH_CHECK(m % 4 == 0, "m must be divisible by 4");
+  int64_t lda = a.stride(0);
+  int64_t ldb = b.stride(1);
+  int64_t ldc = out.stride(0);
 
-  StrideA a_stride;
-  StrideB b_stride;
-  StrideC c_stride;
-  a_stride =
-      cutlass::make_cute_packed_stride(StrideA{}, cute::make_shape(m, k, 1));
-  b_stride =
-      cutlass::make_cute_packed_stride(StrideB{}, cute::make_shape(n, k, 1));
-  c_stride =
-      cutlass::make_cute_packed_stride(StrideC{}, cute::make_shape(m, n, 1));
+  using StrideA = Stride<int64_t, Int<1>, int64_t>;
+  using StrideB = Stride<int64_t, Int<1>, int64_t>;
+  using StrideC = typename Gemm::StrideC;
 
-  LayoutSFA layout_SFA = 
-      ScaleConfig::tile_atom_to_shape_SFA(make_shape(m, n, k, 1));
-  LayoutSFB layout_SFB = 
-      ScaleConfig::tile_atom_to_shape_SFB(make_shape(m, n, k, 1));
+  StrideA a_stride{lda, Int<1>{}, 0};
+  StrideB b_stride{ldb, Int<1>{}, 0};
+  StrideC c_stride{ldc, Int<1>{}, Int<0>{}};
 
-  auto a_ptr = static_cast<ElementAB const*>(a.data_ptr());
-  auto b_ptr = static_cast<ElementAB const*>(b.data_ptr());
-  auto a_scales_ptr = static_cast<ElementBlockScale const*>(a_scales.data_ptr());
-  auto b_scales_ptr = static_cast<ElementBlockScale const*>(b_scales.data_ptr());
+  auto a_ptr = static_cast<ElementAB*>(a.data_ptr());
+  auto b_ptr = static_cast<ElementAB*>(b.data_ptr());
+  auto a_scales_ptr = static_cast<float*>(a_scales.data_ptr());
+  auto b_scales_ptr = static_cast<float*>(b_scales.data_ptr());
 
-  typename GemmKernel::MainloopArguments mainloop_args{};
-  mainloop_args.ptr_A = a_ptr;
-  mainloop_args.dA = a_stride;
-  mainloop_args.ptr_B = b_ptr;
-  mainloop_args.dB = b_stride;
-  mainloop_args.ptr_SFA = a_scales_ptr;
-  mainloop_args.layout_SFA = layout_SFA;
-  mainloop_args.ptr_SFB = b_scales_ptr;
-  mainloop_args.layout_SFB = layout_SFB;
-  auto prob_shape = cute::make_shape(m, n, k, 1);
+  // Check is the t is contiguous and is 1D or 2D with one of the dimensions
+  // being 1 (i.e. a row or column vector)
+  auto is_contiguous_vector = [](const torch::Tensor& t) {
+    auto t_sizes = t.sizes();
+    return t.is_contiguous() &&
+           (t.dim() == 1 ||
+            (t.dim() == 2 &&
+             *std::min_element(t_sizes.begin(), t_sizes.end()) == 1));
+  };
+
+  // TODO(lucas): lets clean-up the kernel so that we pass in Strides so
+  //  we don't have to deal with enforcing implicit layouts
+  TORCH_CHECK(a_scales.size(0) == m / Gemm::GroupSizeM::value);
+  TORCH_CHECK(a_scales.size(1) == k / Gemm::GroupSizeK::value);
+  TORCH_CHECK(a_scales.stride(0) == 1 || is_contiguous_vector(a_scales),
+              "a_scales must be M major");
+  TORCH_CHECK(b_scales.size(0) == k / Gemm::GroupSizeK::value);
+  TORCH_CHECK(b_scales.size(1) == n / Gemm::GroupSizeN::value);
+  TORCH_CHECK(b_scales.stride(0) == 1 || is_contiguous_vector(b_scales),
+              "b_scales must be K major");
+  typename GemmKernel::MainloopArguments mainloop_args{
+      a_ptr, a_stride, b_ptr, b_stride, a_scales_ptr, b_scales_ptr};
 
   auto c_ptr = static_cast<ElementD*>(out.data_ptr());
   typename GemmKernel::EpilogueArguments epilogue_args{
       {}, c_ptr, c_stride, c_ptr, c_stride};
+
+  typename GemmKernel::TileSchedulerArguments scheduler;
+
+  static constexpr bool UsesStreamKScheduler =
+      cute::is_same_v<typename GemmKernel::TileSchedulerTag,
+                      cutlass::gemm::StreamKScheduler>;
+
+  if constexpr (UsesStreamKScheduler) {
+    using DecompositionMode = typename cutlass::gemm::kernel::detail::
+        PersistentTileSchedulerSm90StreamKParams::DecompositionMode;
+    using ReductionMode = typename cutlass::gemm::kernel::detail::
+        PersistentTileSchedulerSm90StreamKParams::ReductionMode;
+
+    scheduler.decomposition_mode = DecompositionMode::StreamK;
+    scheduler.reduction_mode = ReductionMode::Nondeterministic;
+  }
+
   c3x::cutlass_gemm_caller<GemmKernel>(a.device(), prob_shape, mainloop_args,
-                                       epilogue_args);
+                                       epilogue_args, scheduler);
 }
 
 template <typename OutType>
@@ -165,12 +177,18 @@ void cutlass_gemm_blockwise_sm90_fp8_dispatch(torch::Tensor& out,
                                               torch::Tensor const& b,
                                               torch::Tensor const& a_scales,
                                               torch::Tensor const& b_scales) {
-  // TODO: better heuristics
-  cutlass_gemm_caller_blockwise<cutlass_3x_gemm_fp8_blockwise<
-      OutType, 1, 128, 128, Shape<_128, _128, _128>,
-      Shape<_1, _2, _1>, cutlass::epilogue::TmaWarpSpecializedCooperative,
-      cutlass::gemm::KernelTmaWarpSpecializedCooperativeFP8BlockScaledAccum>>(
-      out, a, b, a_scales, b_scales);
+  auto k = a.size(1);
+  auto n = b.size(1);
+
+  if (k > 3 * n) {
+    cutlass_gemm_caller_blockwise<cutlass_3x_gemm_fp8_blockwise<
+        cutlass::gemm::StreamKScheduler, OutType, 1, 128, 128>>(
+        out, a, b, a_scales, b_scales);
+  } else {
+    cutlass_gemm_caller_blockwise<cutlass_3x_gemm_fp8_blockwise<
+        cutlass::gemm::PersistentScheduler, OutType, 1, 128, 128>>(
+        out, a, b, a_scales, b_scales);
+  }
 }
 
 }  // namespace vllm

csrc/quantization/cutlass_w8a8/c3x/scaled_mm_helper.hpp

@@ -32,7 +32,7 @@ void dispatch_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
     TORCH_CHECK(a_scales.dim() == 2, "a scale must be 2d tensor.");
     TORCH_CHECK(b_scales.dim() == 2, "b scale must be 2d tensor.");
     int32_t version_num = get_sm_version_num();
-    if (version_num >= 90) {
+    if (version_num >= 100) {
       TORCH_CHECK(
           a.size(0) == a_scales.size(0) &&
               cuda_utils::ceil_div(a.size(1), int64_t(128)) == a_scales.size(1),
@@ -41,6 +41,32 @@ void dispatch_scaled_mm(torch::Tensor& c, torch::Tensor const& a,
           cuda_utils::ceil_div(b.size(0), int64_t(128)) == b_scales.size(0) &&
               cuda_utils::ceil_div(b.size(1), int64_t(128)) == b_scales.size(1),
           "b_scale_group_shape must be [128, 128].");
+    } else {
+      // TODO: Remove this after using cutlass sm90 blockwise scaling gemm
+      // kernel, or introducing ceil_div to the load_init() of mainloop.
+      using GroupShape = std::array<int64_t, 2>;
+      auto make_group_shape = [](torch::Tensor const& x,
+                                 torch::Tensor const& s) -> GroupShape {
+        TORCH_CHECK(s.dim() == 2, "cutlass_scaled_mm group scales must be 2D");
+        return {cuda_utils::ceil_div(x.size(0), s.size(0)),
+                cuda_utils::ceil_div(x.size(1), s.size(1))};
+      };
+
+      GroupShape a_scale_group_shape = make_group_shape(a, a_scales);
+      GroupShape b_scale_group_shape = make_group_shape(b, b_scales);
+
+      // 1x128 per-token group scales for activations
+      // 128x128 blockwise scales for weights
+      TORCH_CHECK((a_scale_group_shape == GroupShape{1, 128} &&
+                   b_scale_group_shape == GroupShape{128, 128} &&
+                   a.dtype() == torch::kFloat8_e4m3fn &&
+                   b.dtype() == torch::kFloat8_e4m3fn),
+                  "cutlass_scaled_mm only supports datatype float8_e4m3fn.\n"
+                  "a_scale_group_shape must be [1, 128]. Got: [",
+                  a_scale_group_shape[0], ", ", a_scale_group_shape[1],
+                  "]\n"
+                  "b_scale_group_shape must be [128, 128]. Got: [",
+                  b_scale_group_shape[0], ", ", b_scale_group_shape[1], "]");
     }
 
     TORCH_CHECK(!bias, "Bias not yet supported blockwise scaled_mm");

csrc/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu

@@ -26,46 +26,113 @@
 #include "dispatch_utils.h"
 
 #include "cuda_utils.h"
-#include "launch_bounds_utils.h"
 #include "nvfp4_utils.cuh"
 
 namespace vllm {
 
-// silu in float32
-__device__ __forceinline__ float silu(float x) {
-  return __fdividef(x, (1.f + __expf(-x)));
-}
-
-__device__ __forceinline__ float2 silu2(float2 x) {
-  return make_float2(silu(x.x), silu(x.y));
-}
-
 template <class Type>
-__inline__ __device__ PackedVec<Type> compute_silu_mul(PackedVec<Type>& vec,
-                                                       PackedVec<Type>& vec2) {
+__inline__ __device__ PackedVec<Type> compute_silu(PackedVec<Type>& vec,
+                                                   PackedVec<Type>& vec2) {
   PackedVec<Type> result;
-  using packed_type = typename TypeConverter<Type>::Type;
-
 #pragma unroll
   for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 2; ++i) {
-    // silu_mul in float32
     if constexpr (std::is_same_v<Type, half>) {
-      float2 silu_vec = silu2(__half22float2(vec.elts[i]));
-      result.elts[i] =
-          __float22half2_rn(__fmul2_rn(silu_vec, __half22float2(vec2.elts[i])));
+      half2 val(0.5f, 0.5f);
+      half2 t0 = __hmul2(vec.elts[i], val);
+      half2 t1 = __hfma2(h2tanh(t0), val, val);
+      half2 t2 = __hmul2(vec.elts[i], t1);
+      result.elts[i] = __hmul2(t2, vec2.elts[i]);
     } else {
-      float2 silu_vec = silu2(__bfloat1622float2(vec.elts[i]));
-      result.elts[i] = __float22bfloat162_rn(
-          __fmul2_rn(silu_vec, __bfloat1622float2(vec2.elts[i])));
+      __nv_bfloat162 val(0.5f, 0.5f);
+      __nv_bfloat162 t0 = __hmul2(vec.elts[i], val);
+      __nv_bfloat162 t1 = __hfma2(h2tanh(t0), val, val);
+      __nv_bfloat162 t2 = __hmul2(vec.elts[i], t1);
+      result.elts[i] = __hmul2(t2, vec2.elts[i]);
     }
   }
   return result;
 }
 
+// Quantizes the provided PackedVec into the uint32_t output
+template <class Type, bool UE8M0_SF = false>
+__device__ uint32_t silu_and_cvt_warp_fp16_to_fp4(PackedVec<Type>& vec,
+                                                  PackedVec<Type>& vec2,
+                                                  float SFScaleVal,
+                                                  uint8_t* SFout) {
+  PackedVec<Type> out_silu = compute_silu(vec, vec2);
+  // Get absolute maximum values among the local 8 values.
+  auto localMax = __habs2(out_silu.elts[0]);
+
+// Local maximum value.
+#pragma unroll
+  for (int i = 1; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) {
+    localMax = __hmax2(localMax, __habs2(out_silu.elts[i]));
+  }
+
+  // Get the absolute maximum among all 16 values (two threads).
+  localMax = __hmax2(__shfl_xor_sync(uint32_t(-1), localMax, 1), localMax);
+  // Get the final absolute maximum values.
+  float vecMax = float(__hmax(localMax.x, localMax.y));
+
+  // Get the SF (max value of the vector / max value of e2m1).
+  // maximum value of e2m1 = 6.0.
+  // TODO: use half as compute data type.
+  float SFValue = SFScaleVal * (vecMax * reciprocal_approximate_ftz(6.0f));
+  // 8 bits representation of the SF.
+  uint8_t fp8SFVal;
+  // Write the SF to global memory (STG.8).
+  if constexpr (UE8M0_SF) {
+    // Extract the 8 exponent bits from float32.
+    // float 32bits = 1 sign bit + 8 exponent bits + 23 mantissa bits.
+    uint32_t tmp = reinterpret_cast<uint32_t&>(SFValue) >> 23;
+    fp8SFVal = tmp & 0xff;
+    // Convert back to fp32.
+    reinterpret_cast<uint32_t&>(SFValue) = tmp << 23;
+  } else {
+    // Here SFValue is always positive, so E4M3 is the same as UE4M3.
+    __nv_fp8_e4m3 tmp = __nv_fp8_e4m3(SFValue);
+    reinterpret_cast<__nv_fp8_e4m3&>(fp8SFVal) = tmp;
+    // Convert back to fp32.
+    SFValue = float(tmp);
+  }
+  // Get the output scale.
+  // Recipe: final_scale = reciprocal(fp32(fp8(SFValue * SFScaleVal))) *
+  //                       reciprocal(SFScaleVal))
+  float outputScale =
+      SFValue != 0 ? reciprocal_approximate_ftz(
+                         SFValue * reciprocal_approximate_ftz(SFScaleVal))
+                   : 0.0f;
+
+  if (SFout) {
+    // Write the SF to global memory (STG.8).
+    *SFout = fp8SFVal;
+  }
+
+  // Convert the input to float.
+  float2 fp2Vals[CVT_FP4_ELTS_PER_THREAD / 2];
+
+#pragma unroll
+  for (int i = 0; i < CVT_FP4_ELTS_PER_THREAD / 2; i++) {
+    if constexpr (std::is_same_v<Type, half>) {
+      fp2Vals[i] = __half22float2(out_silu.elts[i]);
+    } else {
+      fp2Vals[i] = __bfloat1622float2(out_silu.elts[i]);
+    }
+    fp2Vals[i].x *= outputScale;
+    fp2Vals[i].y *= outputScale;
+  }
+
+  // Convert to e2m1 values.
+  uint32_t e2m1Vec = fp32_vec_to_e2m1(fp2Vals);
+
+  // Write the e2m1 values to global memory.
+  return e2m1Vec;
+}
+
 // Use UE4M3 by default.
 template <class Type, bool UE8M0_SF = false>
-__global__ void __launch_bounds__(1024, VLLM_BLOCKS_PER_SM(1024))
-    silu_mul_cvt_fp16_to_fp4(int32_t numRows, int32_t numCols, Type const* in,
+__global__ void __launch_bounds__(1024, 4)
+    silu_and_cvt_fp16_to_fp4(int32_t numRows, int32_t numCols, Type const* in,
                              float const* SFScale, uint32_t* out,
                              uint32_t* SFout) {
   using PackedVec = PackedVec<Type>;
@@ -93,18 +160,16 @@ __global__ void __launch_bounds__(1024, VLLM_BLOCKS_PER_SM(1024))
       // Get the output tensor offset.
       // Same as inOffset because 8 elements are packed into one uint32_t.
       int64_t outOffset = rowIdx * (numCols / CVT_FP4_ELTS_PER_THREAD) + colIdx;
+      ;
       auto& out_pos = out[outOffset];
 
-      // Compute silu and mul
-      PackedVec out_silu_mul = compute_silu_mul(in_vec, in_vec2);
-
       auto sf_out =
           cvt_quant_to_fp4_get_sf_out_offset<uint32_t,
                                              CVT_FP4_NUM_THREADS_PER_SF>(
               rowIdx, colIdx, numCols, SFout);
 
-      out_pos = cvt_warp_fp16_to_fp4<Type, UE8M0_SF>(out_silu_mul, SFScaleVal,
-                                                     sf_out);
+      out_pos = silu_and_cvt_warp_fp16_to_fp4<Type, UE8M0_SF>(
+          in_vec, in_vec2, SFScaleVal, sf_out);
     }
   }
 }
@@ -132,15 +197,14 @@ void silu_and_mul_nvfp4_quant_sm1xxa(torch::Tensor& output,  // [..., d]
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   auto stream = at::cuda::getCurrentCUDAStream(input.get_device());
   dim3 block(std::min(int(n / ELTS_PER_THREAD), 1024));
-  int const numBlocksPerSM =
-      vllm_runtime_blocks_per_sm(static_cast<int>(block.x));
+  int const numBlocksPerSM = 2048 / block.x;
   dim3 grid(std::min(int(m), multiProcessorCount * numBlocksPerSM));
 
   VLLM_DISPATCH_HALF_TYPES(
       input.scalar_type(), "silu_and_mul_nvfp4_quant_kernel", [&] {
         using cuda_type = vllm::CUDATypeConverter<scalar_t>::Type;
         auto input_ptr = static_cast<cuda_type const*>(input.data_ptr());
-        vllm::silu_mul_cvt_fp16_to_fp4<cuda_type><<<grid, block, 0, stream>>>(
+        vllm::silu_and_cvt_fp16_to_fp4<cuda_type><<<grid, block, 0, stream>>>(
             m, n, input_ptr, input_sf_ptr,
             reinterpret_cast<uint32_t*>(output_ptr),
             reinterpret_cast<uint32_t*>(sf_out));

csrc/quantization/fp4/nvfp4_experts_quant.cu

@@ -26,13 +26,12 @@
 #include "dispatch_utils.h"
 
 #include "nvfp4_utils.cuh"
-#include "launch_bounds_utils.h"
 
 namespace vllm {
 
 // Use UE4M3 by default.
 template <class Type, bool UE8M0_SF = false, bool SMALL_NUM_EXPERTS = false>
-__global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
+__global__ void __launch_bounds__(512, 4)
     cvt_fp16_to_fp4(int32_t numRows, int32_t numCols, Type const* in,
                     float const* SFScale, uint32_t* out, uint32_t* SFout,
                     uint32_t* input_offset_by_experts,
@@ -130,7 +129,7 @@ __global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
 
 // Kernel for LARGE_M_TOPK = true (large m_topk optimized version)
 template <class Type, bool UE8M0_SF = false, bool SMALL_NUM_EXPERTS = false>
-__global__ void __launch_bounds__(1024, VLLM_BLOCKS_PER_SM(1024))
+__global__ void __launch_bounds__(1024, 4)
     cvt_fp16_to_fp4(int32_t numRows, int32_t numCols, Type const* in,
                     float const* SFScale, uint32_t* out, uint32_t* SFout,
                     uint32_t* input_offset_by_experts,
@@ -234,9 +233,8 @@ void quant_impl(void* output, void* output_scale, void* input,
   int const workSizePerRow = k / ELTS_PER_THREAD;
   int const totalWorkSize = m_topk * workSizePerRow;
   dim3 block(std::min(workSizePerRow, 512));
-  // Get number of blocks per SM
-  int const numBlocksPerSM =
-      vllm_runtime_blocks_per_sm(static_cast<int>(block.x));
+  // Get number of blocks per SM (assume we can fully utilize the SM).
+  int const numBlocksPerSM = 2048 / block.x;
   dim3 grid(std::min(static_cast<int>((totalWorkSize + block.x - 1) / block.x),
                      multiProcessorCount * numBlocksPerSM));
   while (grid.x <= multiProcessorCount && block.x > 64) {

csrc/quantization/fp4/nvfp4_quant_kernels.cu

@@ -26,14 +26,13 @@
 #include "dispatch_utils.h"
 
 #include "cuda_utils.h"
-#include "launch_bounds_utils.h"
 #include "nvfp4_utils.cuh"
 
 namespace vllm {
 
 // Use UE4M3 by default.
 template <class Type, bool UE8M0_SF = false>
-__global__ void __launch_bounds__(512, VLLM_BLOCKS_PER_SM(512))
+__global__ void __launch_bounds__(512, 4)
     cvt_fp16_to_fp4(int32_t numRows, int32_t numCols, Type const* in,
                     float const* SFScale, uint32_t* out, uint32_t* SFout) {
   using PackedVec = PackedVec<Type>;
@@ -76,9 +75,8 @@ void invokeFP4Quantization(int m, int n, T const* input, float const* SFScale,
   // Grid, Block size.
   // Each thread converts 8 values.
   dim3 block(std::min(int(n / ELTS_PER_THREAD), 512));
-  // Get number of blocks per SM
-  int const numBlocksPerSM =
-      vllm_runtime_blocks_per_sm(static_cast<int>(block.x));
+  // Get number of blocks per SM (assume we can fully utilize the SM).
+  int const numBlocksPerSM = 2048 / block.x;
   dim3 grid(std::min(int(m), multiProcessorCount * numBlocksPerSM));
 
   // Launch the cvt kernel.

csrc/quantization/fp8/common.cu

@@ -1,10 +1,15 @@
 #include "common.cuh"
 #include "dispatch_utils.h"
-#include "../../cub_helpers.h"
 #include "../vectorization_utils.cuh"
 #include <c10/cuda/CUDAGuard.h>
 #include <ATen/cuda/Exceptions.h>
 
+#ifndef USE_ROCM
+  #include <cub/cub.cuh>
+#else
+  #include <hipcub/hipcub.hpp>
+#endif
+
 namespace vllm {
 
 template <typename scalar_t, typename fp8_type>
@@ -111,7 +116,7 @@ __global__ void dynamic_per_token_scaled_fp8_quant_kernel_strided(
   using BlockReduce = cub::BlockReduce<float, 256>;
   __shared__ typename BlockReduce::TempStorage tmp;
   const float block_max =
-      BlockReduce(tmp).Reduce(absmax_val, CubMaxOp{}, blockDim.x);
+      BlockReduce(tmp).Reduce(absmax_val, cub::Max{}, blockDim.x);
 
   __shared__ float token_scale;
   if (tid == 0) {

csrc/quantization/fp8/common.cuh

@@ -5,9 +5,7 @@
 
 #include <cmath>
 
-#ifndef USE_ROCM
-  #include "nvidia/quant_utils.cuh"
-#else
+#ifdef USE_ROCM
   #include "amd/quant_utils.cuh"
 #endif
 
@@ -50,9 +48,7 @@ __device__ __forceinline__ fp8_type scaled_fp8_conversion(float const val,
   float r =
       fmaxf(-quant_type_max_v<fp8_type>, fminf(x, quant_type_max_v<fp8_type>));
 #ifndef USE_ROCM
-  // Use hardware cvt instruction for fp8 on nvidia
-  // Currently only support fp8_type = c10::Float8_e4m3fn
-  return fp8::vec_conversion<fp8_type, float>(r);
+  return static_cast<fp8_type>(r);
 #else
   // Use hardware cvt instruction for fp8 on rocm
   return fp8::cvt_c10<fp8_type>(r);