Remove executable flag on a few files

Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>

.buildkite/check-wheel-size.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import os
 import sys

.buildkite/generate_index.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import os

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

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 from pathlib import Path
 
 import pytest

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

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 LM eval harness on model to compare vs HF baseline computed offline.
 Configs are found in configs/$MODEL.yaml

.buildkite/nightly-benchmarks/README.md

@@ -113,7 +113,7 @@ WARNING: The benchmarking script will save json results by itself, so please do
 
 ### Visualizing the results
 
-The `convert-results-json-to-markdown.py` helps you put the benchmarking results inside a markdown table, by formatting [descriptions.md](tests/descriptions.md) with real benchmarking results.
+The `convert-results-json-to-markdown.py` helps you put the benchmarking results inside a markdown table, by formatting [descriptions.md](performance-benchmarks-descriptions.md) with real benchmarking results.
 You can find the result presented as a table inside the `buildkite/performance-benchmark` job page.
 If you do not see the table, please wait till the benchmark finish running.
 The json version of the table (together with the json version of the benchmark) will be also attached to the markdown file.

.buildkite/nightly-benchmarks/nightly-annotation.md

@@ -16,7 +16,7 @@ Please download the visualization scripts in the post
   - Download `nightly-benchmarks.zip`.
   - In the same folder, run the following code:
 
-  ```console
+  ```bash
   export HF_TOKEN=<your HF token>
   apt update
   apt install -y git

.buildkite/nightly-benchmarks/scripts/convert-results-json-to-markdown.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import json
 import os

.buildkite/nightly-benchmarks/scripts/download-tokenizer.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 

.buildkite/nightly-benchmarks/scripts/generate-nightly-markdown.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import json

.buildkite/nightly-benchmarks/scripts/get-lmdeploy-modelname.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from lmdeploy.serve.openai.api_client import APIClient
 

.buildkite/nightly-benchmarks/scripts/summary-nightly-results.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import datetime
 import json

.buildkite/release-pipeline.yaml

@@ -1,5 +1,6 @@
 steps:
   - label: "Build wheel - CUDA 12.8"
+    id: build-wheel-cuda-12-8
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -11,6 +12,7 @@ steps:
       DOCKER_BUILDKIT: "1"
 
   - label: "Build wheel - CUDA 12.6"
+    id: build-wheel-cuda-12-6
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -28,6 +30,7 @@ steps:
 
   - label: "Build wheel - CUDA 11.8"
     # depends_on: block-build-cu118-wheel
+    id: build-wheel-cuda-11-8
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -44,6 +47,7 @@ steps:
 
   - label: "Build release image"
     depends_on: block-release-image-build
+    id: build-release-image
     agents:
       queue: cpu_queue_postmerge
     commands:
@@ -51,6 +55,18 @@ steps:
       - "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=12.8.1 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
+  - label: "Annotate release workflow"
+    depends_on:
+      - build-release-image
+      - build-wheel-cuda-12-8
+      - build-wheel-cuda-12-6
+      - build-wheel-cuda-11-8
+    id: annotate-release-workflow
+    agents:
+      queue: cpu_queue_postmerge
+    commands:
+      - "bash .buildkite/scripts/annotate-release.sh"
+
   - label: "Build and publish TPU release image"
     depends_on: ~
     if: build.env("NIGHTLY") == "1"
@@ -70,9 +86,10 @@ steps:
       DOCKER_BUILDKIT: "1"
 
   - input: "Provide Release version here"
+    id: input-release-version
     fields:
       - text: "What is the release version?"
-        key: "release-version"
+        key: release-version
 
   - block: "Build CPU release image"
     key: block-cpu-release-image-build
@@ -85,6 +102,7 @@ steps:
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
       - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest"
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"
@@ -100,6 +118,7 @@ steps:
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
       - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:latest --progress plain -f docker/Dockerfile.neuron ."
+      - "docker push public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:latest"
       - "docker push public.ecr.aws/q9t5s3a7/vllm-neuron-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"

.buildkite/scripts/annotate-release.sh

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+set -ex
+
+# Get release version and strip leading 'v' if present
+RELEASE_VERSION=$(buildkite-agent meta-data get release-version | sed 's/^v//')
+
+if [ -z "$RELEASE_VERSION" ]; then
+  echo "Error: RELEASE_VERSION is empty. 'release-version' metadata might not be set or is invalid."
+  exit 1
+fi
+
+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/${RELEASE_VERSION}+cu126/vllm-${RELEASE_VERSION}+cu126-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:${BUILDKITE_COMMIT}
+docker tag public.ecr.aws/q9t5s3a7/vllm-release-repo:${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/ci-clean-log.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+# Usage: ./ci_clean_log.sh ci.log
+# This script strips timestamps and color codes from CI log files.
+
+# Check if argument is given
+if [ $# -lt 1 ]; then
+    echo "Usage: $0 ci.log"
+    exit 1
+fi
+
+INPUT_FILE="$1"
+
+# Strip timestamps
+sed -i 's/^\[[0-9]\{4\}-[0-9]\{2\}-[0-9]\{2\}T[0-9]\{2\}:[0-9]\{2\}:[0-9]\{2\}Z\] //' "$INPUT_FILE"
+
+# Strip colorization
+sed -i -r 's/\x1B\[[0-9;]*[mK]//g' "$INPUT_FILE"

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

@@ -94,6 +94,10 @@ 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
 
+if [[ $commands == *"pytest -v -s lora"* ]]; then
+  commands=${commands//"pytest -v -s lora"/"VLLM_ROCM_CUSTOM_PAGED_ATTN=0 pytest -v -s lora"}
+fi
+
 #ignore certain kernels tests
 if [[ $commands == *" kernels/core"* ]]; then
   commands="${commands} \

.buildkite/scripts/hardware_ci/run-cpu-test-ppc64le.sh

@@ -7,6 +7,7 @@ set -ex
 # Setup cleanup
 remove_docker_container() {
   if [[ -n "$container_id" ]]; then
+      podman stop --all -t0
       podman rm -f "$container_id" || true
   fi
   podman system prune -f
@@ -37,7 +38,7 @@ function cpu_tests() {
     pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-facebook/opt-125m]
     pytest -v -s tests/models/language/generation/test_common.py::test_models[False-5-32-google/gemma-1.1-2b-it]
     pytest -v -s tests/models/language/pooling/test_classification.py::test_models[float-jason9693/Qwen2.5-1.5B-apeach]
-    pytest -v -s tests/models/language/pooling/test_embedding.py::test_models[half-BAAI/bge-base-en-v1.5]"
+    pytest -v -s tests/models/language/pooling/test_embedding.py -m cpu_model"
 }
 
 # All of CPU tests are expected to be finished less than 40 mins.

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

@@ -6,75 +6,83 @@ set -ex
 
 # allow to bind to different cores
 CORE_RANGE=${CORE_RANGE:-48-95}
+OMP_CORE_RANGE=${OMP_CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 
+export CMAKE_BUILD_PARALLEL_LEVEL=32
+
 # Setup cleanup
 remove_docker_container() { 
     set -e; 
-    docker rm -f cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" || true; 
-    docker image rm cpu-test-"$BUILDKITE_BUILD_NUMBER" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 || true; 
+    docker rm -f cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"-avx2 || true; 
 }
 trap remove_docker_container EXIT
 remove_docker_container
 
 # Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$BUILDKITE_BUILD_NUMBER" --target vllm-test -f docker/Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --tag cpu-test-"$NUMA_NODE" --target vllm-test -f docker/Dockerfile.cpu .
+numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" --tag cpu-test-"$NUMA_NODE"-avx2 --target vllm-test -f docker/Dockerfile.cpu .
 
 # Run the image, setting --shm-size=4g for tensor parallel.
-docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE"  \
- --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"
-docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus="$CORE_RANGE" \
- --cpuset-mems="$NUMA_NODE" --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_OMP_THREADS_BIND="$OMP_CORE_RANGE" --env VLLM_CPU_CI_ENV=1 --shm-size=4g --name cpu-test-"$NUMA_NODE" cpu-test-"$NUMA_NODE"
+docker run -itd --cpuset-cpus="$CORE_RANGE" --cpuset-mems="$NUMA_NODE" --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --privileged=true -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --env VLLM_CPU_OMP_THREADS_BIND="$OMP_CORE_RANGE" --env VLLM_CPU_CI_ENV=1 --shm-size=4g --name cpu-test-"$NUMA_NODE"-avx2 cpu-test-"$NUMA_NODE"-avx2
 
 function cpu_tests() {
   set -e
   export NUMA_NODE=$2
-  export BUILDKITE_BUILD_NUMBER=$3
+
+  # list packages
+  docker exec cpu-test-"$NUMA_NODE"-avx2 bash -c "
+    set -e
+    pip list"
+
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
+    set -e
+    pip list"
 
   # offline inference
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE"-avx2 bash -c "
     set -e
     python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
 
   # Run basic model test
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
-    pytest -v -s tests/kernels/test_cache.py -m cpu_model
-    pytest -v -s tests/kernels/test_mla_decode_cpu.py -m cpu_model
-    pytest -v -s tests/models/decoder_only/language -m cpu_model
-    pytest -v -s tests/models/embedding/language -m cpu_model
-    pytest -v -s tests/models/encoder_decoder/language -m cpu_model
-    pytest -v -s tests/models/decoder_only/audio_language -m cpu_model
-    pytest -v -s tests/models/decoder_only/vision_language -m cpu_model"
+    pytest -v -s tests/kernels/attention/test_cache.py -m cpu_model
+    pytest -v -s tests/kernels/attention/test_mla_decode_cpu.py -m cpu_model
+    pytest -v -s tests/models/language/generation -m cpu_model
+    VLLM_CPU_SGL_KERNEL=1 pytest -v -s tests/models/language/generation -m cpu_model
+    pytest -v -s tests/models/language/pooling -m cpu_model
+    pytest -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"
 
   # Run compressed-tensor test
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
     pytest -s -v \
     tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_static_setup \
     tests/quantization/test_compressed_tensors.py::test_compressed_tensors_w8a8_dynamic_per_token"
 
   # Run AWQ test
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
-    pytest -s -v \
+    VLLM_USE_V1=0 pytest -s -v \
     tests/quantization/test_ipex_quant.py"
 
   # Run chunked-prefill and prefix-cache test
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
     pytest -s -v -k cpu_model \
     tests/basic_correctness/test_chunked_prefill.py"  
 
   # online serving
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
-    export VLLM_CPU_KVCACHE_SPACE=10 
-    export VLLM_CPU_OMP_THREADS_BIND=$1
     python3 -m vllm.entrypoints.openai.api_server --model facebook/opt-125m --dtype half & 
     timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
-    python3 benchmarks/benchmark_serving.py \
+    VLLM_CPU_CI_ENV=0 python3 benchmarks/benchmark_serving.py \
       --backend vllm \
       --dataset-name random \
       --model facebook/opt-125m \
@@ -83,7 +91,7 @@ function cpu_tests() {
       --tokenizer facebook/opt-125m"
 
   # Run multi-lora tests
-  docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
+  docker exec cpu-test-"$NUMA_NODE" bash -c "
     set -e
     pytest -s -v \
     tests/lora/test_qwen2vl.py"
@@ -91,4 +99,4 @@ function cpu_tests() {
 
 # All of CPU tests are expected to be finished less than 40 mins.
 export -f cpu_tests
-timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE $BUILDKITE_BUILD_NUMBER"
+timeout 1.5h bash -c "cpu_tests $CORE_RANGE $NUMA_NODE"

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

@@ -2,10 +2,34 @@
 
 # This script build the CPU docker image and run the offline inference inside the container.
 # It serves a sanity check for compilation and basic model usage.
-set -ex
+set -exuo pipefail
 
 # Try building the docker image
-docker build -t hpu-test-env -f docker/Dockerfile.hpu .
+cat <<EOF | docker build -t hpu-plugin-v1-test-env -f - .
+FROM 1.22-413-pt2.7.1:latest
+
+COPY ./ /workspace/vllm
+
+WORKDIR /workspace/vllm
+
+RUN pip install -v -r requirements/hpu.txt
+RUN pip install git+https://github.com/vllm-project/vllm-gaudi.git
+
+ENV no_proxy=localhost,127.0.0.1
+ENV PT_HPU_ENABLE_LAZY_COLLECTIVES=true
+
+RUN VLLM_TARGET_DEVICE=hpu python3 setup.py install
+
+# install development dependencies (for testing)
+RUN python3 -m pip install -e tests/vllm_test_utils
+
+WORKDIR /workspace/
+
+RUN git clone https://github.com/vllm-project/vllm-gaudi.git
+
+RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
+
+EOF
 
 # Setup cleanup
 # certain versions of HPU software stack have a bug that can
@@ -14,13 +38,21 @@ docker build -t hpu-test-env -f docker/Dockerfile.hpu .
 # functions, while other platforms only need one remove_docker_container
 # function.
 EXITCODE=1
-remove_docker_containers() { docker rm -f hpu-test || true; docker rm -f hpu-test-tp2 || true; }
-remove_docker_containers_and_exit() { remove_docker_containers; exit $EXITCODE; }
-trap remove_docker_containers_and_exit EXIT
+remove_docker_containers() { docker rm -f hpu-plugin-v1-test || true; }
+trap 'remove_docker_containers; exit $EXITCODE;' EXIT
 remove_docker_containers
 
-# Run the image and launch offline inference
-docker run --runtime=habana --name=hpu-test --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
-docker run --runtime=habana --name=hpu-test-tp2 --network=host -e HABANA_VISIBLE_DEVICES=all -e VLLM_SKIP_WARMUP=true --entrypoint="" hpu-test-env python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --tensor-parallel-size 2
+echo "Running HPU plugin v1 test"
+docker run --rm --runtime=habana --name=hpu-plugin-v1-test --network=host \
+  -e HABANA_VISIBLE_DEVICES=all \
+  hpu-plugin-v1-test-env \
+  /bin/bash "/workspace/vllm-gaudi/tests/upstream_tests/ci_tests.sh"
 
 EXITCODE=$?
+if [ $EXITCODE -eq 0 ]; then
+  echo "Test with basic model passed"
+else
+  echo "Test with basic model FAILED with exit code: $EXITCODE" >&2
+fi
+
+# The trap will handle the container removal and final exit.

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

@@ -54,10 +54,11 @@ docker run --rm -it --device=/dev/neuron0 --network bridge \
        --name "${container_name}" \
        ${image_name} \
        /bin/bash -c "
+            set -e; # Exit on first error
             python3 /workspace/vllm/examples/offline_inference/neuron.py;
             python3 -m pytest /workspace/vllm/tests/neuron/1_core/ -v --capture=tee-sys;
             for f in /workspace/vllm/tests/neuron/2_core/*.py; do
-                echo 'Running test file: '$f;
+                echo \"Running test file: \$f\";
                 python3 -m pytest \$f -v --capture=tee-sys;
             done
        "

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

@@ -2,102 +2,186 @@
 
 set -xu
 
+
+remove_docker_container() { 
+    docker rm -f tpu-test || true; 
+    docker rm -f vllm-tpu || true;
+}
+
+trap remove_docker_container EXIT
+
+# Remove the container that might not be cleaned up in the previous run.
+remove_docker_container
+
 # Build the docker image.
 docker build -f docker/Dockerfile.tpu -t vllm-tpu .
 
 # Set up cleanup.
-remove_docker_container() { docker rm -f tpu-test || true; }
-trap remove_docker_container EXIT
-# Remove the container that might not be cleaned up in the previous run.
-remove_docker_container
+cleanup_docker() {
+  # Get Docker's root directory
+  docker_root=$(docker info -f '{{.DockerRootDir}}')
+  if [ -z "$docker_root" ]; then
+    echo "Failed to determine Docker root directory."
+    exit 1
+  fi
+  echo "Docker root directory: $docker_root"
+  # Check disk usage of the filesystem where Docker's root directory is located
+  disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
+  # Define the threshold
+  threshold=70
+  if [ "$disk_usage" -gt "$threshold" ]; then
+    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
+    # Remove dangling images (those that are not tagged and not used by any container)
+    docker image prune -f
+    # Remove unused volumes / force the system prune for old images as well.
+    docker volume prune -f && docker system prune --force --filter "until=72h" --all
+    echo "Docker images and volumes cleanup completed."
+  else
+    echo "Disk usage is below $threshold%. No cleanup needed."
+  fi
+}
+cleanup_docker
 
 # For HF_TOKEN.
 source /etc/environment
-# Run a simple end-to-end example.
+
 docker run --privileged --net host --shm-size=16G -it \
     -e "HF_TOKEN=$HF_TOKEN" --name tpu-test \
-    vllm-tpu /bin/bash -c "python3 -m pip install git+https://github.com/thuml/depyf.git \
-    && python3 -m pip install pytest pytest-asyncio tpu-info \
-    && python3 -m pip install lm_eval[api]==0.4.4 \
-    && export VLLM_XLA_CACHE_PATH= \
-    && export VLLM_USE_V1=1 \
-    && export VLLM_XLA_CHECK_RECOMPILATION=1 \
-    && echo HARDWARE \
-    && tpu-info \
-    && { \
-        echo TEST_0: Running test_perf.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_perf.py; \
-        echo TEST_0_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_1: Running test_compilation.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_compilation.py; \
-        echo TEST_1_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_2: Running test_basic.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_basic.py; \
-        echo TEST_2_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_3: Running test_accuracy.py::test_lm_eval_accuracy_v1_engine; \
-        python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine; \
-        echo TEST_3_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_4: Running test_quantization_accuracy.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py; \
-        echo TEST_4_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_5: Running examples/offline_inference/tpu.py; \
-        python3 /workspace/vllm/examples/offline_inference/tpu.py; \
-        echo TEST_5_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_6: Running test_tpu_model_runner.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/tpu/worker/test_tpu_model_runner.py; \
-        echo TEST_6_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_7: Running test_sampler.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py; \
-        echo TEST_7_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_8: Running test_topk_topp_sampler.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py; \
-        echo TEST_8_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_9: Running test_multimodal.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_multimodal.py; \
-        echo TEST_9_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_10: Running test_pallas.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py; \
-        echo TEST_10_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_11: Running test_struct_output_generate.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py; \
-        echo TEST_11_EXIT_CODE: \$?; \
-    } & \
-    { \
-        echo TEST_12: Running test_moe_pallas.py; \
-        python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py; \
-        echo TEST_12_EXIT_CODE: \$?; \
-    } & \
-    # Disable the TPU LoRA tests until the feature is activated
-    # & { \
-    #     echo TEST_13: Running test_moe_pallas.py; \
-    #     python3 -m pytest -s -v /workspace/vllm/tests/tpu/lora/; \
-    #     echo TEST_13_EXIT_CODE: \$?; \
-    # } & \
-    wait \
-    && echo 'All tests have attempted to run. Check logs for individual test statuses and exit codes.' \
-"
+    vllm-tpu /bin/bash -c '
+set -e # Exit immediately if a command exits with a non-zero status.
+set -u # Treat unset variables as an error.
 
+echo "--- Starting script inside Docker container ---"
+
+# Create results directory
+RESULTS_DIR=$(mktemp -d)
+# If mktemp fails, set -e will cause the script to exit.
+echo "Results will be stored in: $RESULTS_DIR"
+
+# Install dependencies
+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[api]==0.4.4
+echo "--- Python dependencies installed ---"
+export VLLM_USE_V1=1
+export VLLM_XLA_CHECK_RECOMPILATION=1
+export VLLM_XLA_CACHE_PATH=
+echo "Using VLLM V1"
+
+echo "--- Hardware Information ---"
+tpu-info
+echo "--- Starting Tests ---"
+set +e
+overall_script_exit_code=0
+
+# --- Test Definitions ---
+# If a test fails, this function will print logs and will not cause the main script to exit.
+run_test() {
+    local test_num=$1
+    local test_name=$2
+    local test_command=$3
+    local log_file="$RESULTS_DIR/test_${test_num}.log"
+    local actual_exit_code
+
+    echo "--- TEST_$test_num: Running $test_name ---"
+    
+    # Execute the test command.
+    eval "$test_command" > >(tee -a "$log_file") 2> >(tee -a "$log_file" >&2)
+    actual_exit_code=$?
+
+    echo "TEST_${test_num}_COMMAND_EXIT_CODE: $actual_exit_code" # This goes to main log
+    echo "TEST_${test_num}_COMMAND_EXIT_CODE: $actual_exit_code" >> "$log_file" # Also to per-test log
+
+    if [ "$actual_exit_code" -ne 0 ]; then
+        echo "TEST_$test_num ($test_name) FAILED with exit code $actual_exit_code." >&2
+        echo "--- Log for failed TEST_$test_num ($test_name) ---" >&2
+        if [ -f "$log_file" ]; then
+            cat "$log_file" >&2
+        else
+            echo "Log file $log_file not found for TEST_$test_num ($test_name)." >&2
+        fi
+        echo "--- End of log for TEST_$test_num ($test_name) ---" >&2
+        return "$actual_exit_code" # Return the failure code
+    else
+        echo "TEST_$test_num ($test_name) PASSED."
+        return 0 # Return success
+    fi
+}
+
+# Helper function to call run_test and update the overall script exit code
+run_and_track_test() {
+    local test_num_arg="$1"
+    local test_name_arg="$2"
+    local test_command_arg="$3"
+
+    # Run the test
+    run_test "$test_num_arg" "$test_name_arg" "$test_command_arg"
+    local test_specific_exit_code=$?
+
+    # If the test failed, set the overall script exit code to 1
+    if [ "$test_specific_exit_code" -ne 0 ]; then
+        # No need for extra echo here, run_test already logged the failure.
+        overall_script_exit_code=1
+    fi
+}
+
+# --- Actual Test Execution ---
+run_and_track_test 0 "test_perf.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_perf.py"
+run_and_track_test 1 "test_compilation.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_compilation.py"
+run_and_track_test 2 "test_basic.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_basic.py"
+run_and_track_test 3 "test_accuracy.py::test_lm_eval_accuracy_v1_engine" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine"
+run_and_track_test 4 "test_quantization_accuracy.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py"
+run_and_track_test 5 "examples/offline_inference/tpu.py" \
+    "python3 /workspace/vllm/examples/offline_inference/tpu.py"
+run_and_track_test 6 "test_tpu_model_runner.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/worker/test_tpu_model_runner.py"
+run_and_track_test 7 "test_sampler.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py"
+run_and_track_test 8 "test_topk_topp_sampler.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py"
+run_and_track_test 9 "test_multimodal.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_multimodal.py"
+run_and_track_test 10 "test_pallas.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py"
+run_and_track_test 11 "test_struct_output_generate.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/entrypoints/llm/test_struct_output_generate.py -k \"not test_structured_output_with_reasoning_matrices\""
+run_and_track_test 12 "test_moe_pallas.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/tpu/test_moe_pallas.py"
+run_and_track_test 13 "test_lora.py" \
+    "VLLM_XLA_CHECK_RECOMPILATION=0 python3 -m pytest -s -v /workspace/vllm/tests/tpu/lora/test_lora.py"
+run_and_track_test 14 "test_tpu_qkv_linear.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_tpu_qkv_linear.py"
+run_and_track_test 15 "test_spmd_model_weight_loading.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_spmd_model_weight_loading.py"
+run_and_track_test 16 "test_kv_cache_update_kernel.py" \
+    "python3 -m pytest -s -v /workspace/vllm/tests/v1/tpu/test_kv_cache_update_kernel.py"
+
+# After all tests have been attempted, exit with the overall status.
+if [ "$overall_script_exit_code" -ne 0 ]; then
+    echo "--- One or more tests FAILED. Overall script exiting with failure code 1. ---"
+else
+    echo "--- All tests have completed and PASSED. Overall script exiting with success code 0. ---"
+fi
+exit "$overall_script_exit_code"
+' # IMPORTANT: This is the closing single quote for the bash -c "..." command. Ensure it is present and correct.
+
+# Capture the exit code of the docker run command
+DOCKER_RUN_EXIT_CODE=$?
+
+# The trap will run for cleanup.
+# Exit the main script with the Docker run command's exit code.
+if [ "$DOCKER_RUN_EXIT_CODE" -ne 0 ]; then
+    echo "Docker run command failed with exit code $DOCKER_RUN_EXIT_CODE."
+    exit "$DOCKER_RUN_EXIT_CODE"
+else
+    echo "Docker run command completed successfully."
+    exit 0
+fi
 # TODO: This test fails because it uses RANDOM_SEED sampling
-# && VLLM_USE_V1=1 pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
+# pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \

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

@@ -28,4 +28,5 @@ docker run \
     sh -c '
     VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
     VLLM_USE_V1=0 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
+    VLLM_USE_V1=1 python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m --block-size 64 --enforce-eager
 '

.buildkite/scripts/rerun-test.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+# Usage: ./rerun_test.sh path/to/test.py::test_name
+
+# Check if argument is given
+if [ $# -lt 1 ]; then
+    echo "Usage: $0 path/to/test.py::test_name"
+    echo "Example: $0 tests/v1/engine/test_engine_core_client.py::test_kv_cache_events[True-tcp]"
+    exit 1
+fi
+
+TEST=$1
+COUNT=1
+
+while pytest -sv "$TEST"; do
+    COUNT=$((COUNT + 1))
+    echo "RUN NUMBER ${COUNT}"
+done

.buildkite/scripts/tpu/cleanup_docker.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+
+set -euo pipefail
+
+docker_root=$(docker info -f '{{.DockerRootDir}}')
+if [ -z "$docker_root" ]; then
+  echo "Failed to determine Docker root directory."
+  exit 1
+fi
+echo "Docker root directory: $docker_root"
+# Check disk usage of the filesystem where Docker's root directory is located
+disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
+# Define the threshold
+threshold=70
+if [ "$disk_usage" -gt "$threshold" ]; then
+  echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
+  # Remove dangling images (those that are not tagged and not used by any container)
+  docker image prune -f
+  # Remove unused volumes / force the system prune for old images as well.
+  docker volume prune -f && docker system prune --force --filter "until=72h" --all
+  echo "Docker images and volumes cleanup completed."
+else
+  echo "Disk usage is below $threshold%. No cleanup needed."
+fi

.buildkite/scripts/tpu/config_v6e_1.env

@@ -0,0 +1,14 @@
+# Environment config
+TEST_NAME=llama8b
+CONTAINER_NAME=vllm-tpu
+
+# vllm config
+MODEL=meta-llama/Llama-3.1-8B-Instruct
+MAX_NUM_SEQS=256
+MAX_NUM_BATCHED_TOKENS=1024
+TENSOR_PARALLEL_SIZE=1
+MAX_MODEL_LEN=2048
+DOWNLOAD_DIR=/mnt/disks/persist
+EXPECTED_THROUGHPUT=8.0
+INPUT_LEN=1800
+OUTPUT_LEN=128

.buildkite/scripts/tpu/docker_run_bm.sh

@@ -0,0 +1,102 @@
+#!/bin/bash
+
+if [ ! -f "$1" ]; then
+  echo "Error: The env file '$1' does not exist."
+  exit 1  # Exit the script with a non-zero status to indicate an error
+fi
+
+ENV_FILE=$1
+
+# For testing on local vm, use `set -a` to export all variables
+source /etc/environment
+source $ENV_FILE
+
+remove_docker_container() { 
+    docker rm -f tpu-test || true; 
+    docker rm -f vllm-tpu || true;
+    docker rm -f $CONTAINER_NAME || true;
+}
+
+trap remove_docker_container EXIT
+
+# Remove the container that might not be cleaned up in the previous run.
+remove_docker_container
+
+# Build docker image.
+# TODO: build the image outside the script and share the image with other
+# tpu test if building time is too long.
+DOCKER_BUILDKIT=1 docker build \
+  --build-arg max_jobs=16 \
+  --build-arg USE_SCCACHE=1 \
+  --build-arg GIT_REPO_CHECK=0 \
+  --tag vllm/vllm-tpu-bm \
+  --progress plain -f docker/Dockerfile.tpu .
+
+LOG_ROOT=$(mktemp -d)
+# If mktemp fails, set -e will cause the script to exit.
+echo "Results will be stored in: $LOG_ROOT"
+
+if [ -z "$HF_TOKEN" ]; then
+  echo "Error: HF_TOKEN is not set or is empty."  
+  exit 1
+fi
+
+# Make sure mounted disk or dir exists
+if [ ! -d "$DOWNLOAD_DIR" ]; then
+    echo "Error: Folder $DOWNLOAD_DIR does not exist. This is useually a mounted drive. If no mounted drive, just create a folder."
+    exit 1
+fi
+
+echo "Run model $MODEL"
+echo
+
+echo "starting docker...$CONTAINER_NAME"
+echo    
+docker run \
+ -v $DOWNLOAD_DIR:$DOWNLOAD_DIR \
+ --env-file $ENV_FILE \
+ -e HF_TOKEN="$HF_TOKEN" \
+ -e TARGET_COMMIT=$BUILDKITE_COMMIT \
+ -e MODEL=$MODEL \
+ -e WORKSPACE=/workspace \
+ --name $CONTAINER_NAME \
+ -d \
+ --privileged \
+ --network host \
+ -v /dev/shm:/dev/shm \
+ vllm/vllm-tpu-bm tail -f /dev/null
+
+echo "run script..."
+echo
+docker exec "$CONTAINER_NAME" /bin/bash -c ".buildkite/scripts/tpu/run_bm.sh"
+
+echo "copy result back..."
+VLLM_LOG="$LOG_ROOT/$TEST_NAME"_vllm_log.txt
+BM_LOG="$LOG_ROOT/$TEST_NAME"_bm_log.txt
+docker cp "$CONTAINER_NAME:/workspace/vllm_log.txt" "$VLLM_LOG" 
+docker cp "$CONTAINER_NAME:/workspace/bm_log.txt" "$BM_LOG"
+
+throughput=$(grep "Request throughput (req/s):" "$BM_LOG" | sed 's/[^0-9.]//g')
+echo "throughput for $TEST_NAME at $BUILDKITE_COMMIT: $throughput"
+
+if [ "$BUILDKITE" = "true" ]; then
+  echo "Running inside Buildkite"
+  buildkite-agent artifact upload "$VLLM_LOG" 
+  buildkite-agent artifact upload "$BM_LOG"
+else
+  echo "Not running inside Buildkite"
+fi
+
+#
+# compare the throughput with EXPECTED_THROUGHPUT 
+# and assert meeting the expectation
+# 
+if [[ -z "$throughput" || ! "$throughput" =~ ^[0-9]+([.][0-9]+)?$ ]]; then
+  echo "Failed to get the throughput"
+  exit 1
+fi
+
+if (( $(echo "$throughput < $EXPECTED_THROUGHPUT" | bc -l) )); then
+  echo "Error: throughput($throughput) is less than expected($EXPECTED_THROUGHPUT)"
+  exit 1
+fi

.buildkite/scripts/tpu/run_bm.sh

@@ -0,0 +1,94 @@
+#!/bin/bash
+
+set -euo pipefail
+
+VLLM_LOG="$WORKSPACE/vllm_log.txt"
+BM_LOG="$WORKSPACE/bm_log.txt"
+
+if [ -n "$TARGET_COMMIT" ]; then
+  head_hash=$(git rev-parse HEAD)
+  if [ "$TARGET_COMMIT" != "$head_hash" ]; then
+    echo "Error: target commit $TARGET_COMMIT does not match HEAD: $head_hash"
+    exit 1
+  fi
+fi
+
+echo "model: $MODEL"
+echo
+
+#
+# create a log folder
+#
+mkdir "$WORKSPACE/log"
+
+# TODO: Move to image building.
+pip install pandas
+pip install datasets
+
+#
+# create sonnet_4x
+#
+echo "Create sonnet_4x.txt"
+echo "" > benchmarks/sonnet_4x.txt
+for _ in {1..4}
+ do
+  cat benchmarks/sonnet.txt >> benchmarks/sonnet_4x.txt
+done
+
+#
+# start vllm service in backend
+#
+echo "lanching vllm..."
+echo "logging to $VLLM_LOG"
+echo
+
+VLLM_USE_V1=1 vllm serve $MODEL \
+ --seed 42 \
+ --disable-log-requests \
+ --max-num-seqs $MAX_NUM_SEQS \
+ --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
+ --tensor-parallel-size $TENSOR_PARALLEL_SIZE \
+ --no-enable-prefix-caching \
+ --download_dir $DOWNLOAD_DIR \
+ --max-model-len $MAX_MODEL_LEN > "$VLLM_LOG" 2>&1 &
+
+
+echo "wait for 20 minutes.."
+echo
+# sleep 1200
+# wait for 10 minutes...
+for i in {1..120}; do
+    # TODO: detect other type of errors.
+    if grep -Fq "raise RuntimeError" "$VLLM_LOG"; then
+        echo "Detected RuntimeError, exiting."
+        exit 1
+    elif grep -Fq "Application startup complete" "$VLLM_LOG"; then
+        echo "Application started"
+        break
+    else
+        echo "wait for 10 seconds..."
+        sleep 10
+    fi
+done
+
+#
+# run test
+#
+echo "run benchmark test..."
+echo "logging to $BM_LOG"
+echo
+python benchmarks/benchmark_serving.py \
+    --backend vllm \
+    --model $MODEL  \
+    --dataset-name sonnet \
+    --dataset-path benchmarks/sonnet_4x.txt \
+    --sonnet-input-len $INPUT_LEN \
+    --sonnet-output-len $OUTPUT_LEN \
+    --ignore-eos > "$BM_LOG"
+
+echo "completed..."
+echo
+
+throughput=$(grep "Request throughput (req/s):" "$BM_LOG" | sed 's/[^0-9.]//g')
+echo "throughput: $throughput"
+echo

.buildkite/test-pipeline.yaml

@@ -41,6 +41,16 @@ steps:
   # TODO: add `--strict` once warnings in docstrings are fixed
   - mkdocs build
 
+- label: Pytorch Nightly Dependency Override Check # 2min
+  # if this test fails, it means the nightly torch version is not compatible with some
+  # of the dependencies. Please check the error message and add the package to whitelist
+  # in /vllm/tools/generate_nightly_torch_test.py
+  soft_fail: true
+  source_file_dependencies:
+  - requirements/nightly_torch_test.txt
+  commands:
+  - bash standalone_tests/pytorch_nightly_dependency.sh
+
 - label: Async Engine, Inputs, Utils, Worker Test # 24min
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
@@ -89,7 +99,7 @@ steps:
   - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
 
 - label: Chunked Prefill Test
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/
   - tests/basic_correctness/test_chunked_prefill
@@ -145,6 +155,7 @@ steps:
   - examples/offline_inference/rlhf_colocate.py
   - tests/examples/offline_inference/data_parallel.py
   - tests/v1/test_async_llm_dp.py
+  - tests/v1/engine/test_engine_core_client.py
   commands:
   # test with tp=2 and external_dp=2
   - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
@@ -154,6 +165,7 @@ steps:
   # test with internal dp
   - python3 ../examples/offline_inference/data_parallel.py
   - TP_SIZE=2 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
+  - pytest -v -s v1/engine/test_engine_core_client.py::test_kv_cache_events_dp
   - pytest -v -s distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
@@ -166,6 +178,23 @@ steps:
   - VLLM_ALLOW_INSECURE_SERIALIZATION=1 RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
   - popd
 
+- label: EPLB Algorithm Test
+  working_dir: "/vllm-workspace/tests"
+  source_file_dependencies:
+  - vllm/distributed/eplb
+  - tests/distributed/test_eplb_algo.py
+  commands:
+  - pytest -v -s distributed/test_eplb_algo.py
+
+- label: EPLB Execution Test # 5min
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  source_file_dependencies:
+  - vllm/distributed/eplb
+  - tests/distributed/test_eplb_execute.py
+  commands:
+  - pytest -v -s distributed/test_eplb_execute.py
+
 - label: Metrics, Tracing Test # 10min
   mirror_hardwares: [amdexperimental, amdproduction]
   num_gpus: 2
@@ -175,6 +204,11 @@ steps:
   - 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 tracing
 
 ##### fast check tests  #####
@@ -199,8 +233,9 @@ steps:
   - tests/test_sequence
   - tests/test_config
   - tests/test_logger
+  - tests/test_vllm_port
   commands:
-  - pytest -v -s engine test_sequence.py test_config.py test_logger.py
+  - pytest -v -s engine test_sequence.py test_config.py test_logger.py test_vllm_port.py
   # OOM in the CI unless we run this separately
   - pytest -v -s tokenization
 
@@ -263,6 +298,15 @@ steps:
   commands:
     - pytest -v -s prefix_caching
 
+
+- label: Platform Tests (CUDA)
+  mirror_hardwares: [amdexperimental]
+  source_file_dependencies:
+  - vllm/
+  - tests/cuda
+  commands:
+    - pytest -v -s cuda/test_cuda_context.py
+
 - label: Samplers Test # 36min
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
@@ -274,17 +318,6 @@ steps:
     - pytest -v -s samplers
     - VLLM_USE_FLASHINFER_SAMPLER=1 pytest -v -s samplers
 
-- label: LogitsProcessor Test # 5min
-  mirror_hardwares: [amdexperimental, amdproduction]
-  source_file_dependencies:
-  - vllm/model_executor/layers
-  - vllm/model_executor/guided_decoding
-  - tests/test_logits_processor
-  - tests/model_executor/test_guided_processors
-  commands:
-    - pytest -v -s test_logits_processor.py
-    - pytest -v -s model_executor/test_guided_processors.py
-
 - label: Speculative decoding tests # 40min
   mirror_hardwares: [amdexperimental]
   source_file_dependencies:
@@ -297,7 +330,7 @@ steps:
     - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
 
 - label: LoRA Test %N # 15min each
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   source_file_dependencies:
   - vllm/lora
   - tests/lora
@@ -313,6 +346,7 @@ steps:
   commands:
     - pytest -v -s compile/test_pass_manager.py
     - pytest -v -s compile/test_fusion.py
+    - pytest -v -s compile/test_fusion_attn.py
     - pytest -v -s compile/test_silu_mul_quant_fusion.py
     - pytest -v -s compile/test_sequence_parallelism.py
     - pytest -v -s compile/test_async_tp.py
@@ -328,6 +362,7 @@ steps:
   # 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
 
 - label: PyTorch Fullgraph Test # 18min
   mirror_hardwares: [amdexperimental, amdproduction]
@@ -397,6 +432,17 @@ steps:
     - pytest -v -s tensorizer_loader
     - pytest -v -s entrypoints/openai/test_tensorizer_entrypoint.py
 
+- label: Model Executor Test
+  mirror_hardwares: [amdexperimental, amdproduction]
+  soft_fail: true
+  source_file_dependencies:
+  - vllm/model_executor
+  - tests/model_executor
+  commands:
+    - apt-get update && apt-get install -y curl libsodium23
+    - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+    - pytest -v -s model_executor
+
 - label: Benchmarks # 9min
   mirror_hardwares: [amdexperimental, amdproduction]
   working_dir: "/vllm-workspace/.buildkite"
@@ -420,6 +466,9 @@ steps:
   - vllm/model_executor/layers/quantization
   - tests/quantization
   commands:
+  # temporary install here since we need nightly, will move to requirements/test.in
+  # after torchao 0.12 release
+  - pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu126
   - VLLM_TEST_FORCE_LOAD_FORMAT=auto pytest -v -s quantization
 
 - label: LM Eval Small Models # 53min
@@ -487,6 +536,17 @@ steps:
     - pip freeze | grep -E 'torch'
     - pytest -v -s models/language -m core_model
 
+- label: Language Models Test (Hybrid) # 35 min
+  mirror_hardwares: [amdexperimental]
+  torch_nightly: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/language/generation
+  commands:
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
+    - pytest -v -s models/language/generation -m hybrid_model
+
 - label: Language Models Test (Extended Generation) # 1hr20min
   mirror_hardwares: [amdexperimental]
   optional: true
@@ -496,7 +556,7 @@ steps:
   commands:
     # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
     - pip install 'git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8'
-    - pytest -v -s models/language/generation -m 'not core_model'
+    - pytest -v -s models/language/generation -m '(not core_model) and (not hybrid_model)'
 
 - label: Language Models Test (Extended Pooling)  # 36min
   mirror_hardwares: [amdexperimental]
@@ -593,13 +653,18 @@ steps:
   - vllm/executor/
   - vllm/model_executor/models/
   - tests/distributed/
+  - tests/examples/offline_inference/data_parallel.py
   commands:
   - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
     - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed'
+    - 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=0 --master-addr=192.168.10.10 --master-port=12345 --enforce-eager --trust-remote-code
     - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py
     - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
   - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
     - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed'
+    - 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) # 40min
   mirror_hardwares: [amdexperimental]
@@ -617,9 +682,11 @@ steps:
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
   - tests/v1/test_async_llm_dp.py
+  - tests/v1/entrypoints/openai/test_multi_api_servers.py
   - vllm/v1/engine/
   commands:
   - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
+  - DP_SIZE=2 pytest -v -s v1/entrypoints/openai/test_multi_api_servers.py
   - pytest -v -s entrypoints/llm/test_collective_rpc.py
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
@@ -660,7 +727,7 @@ steps:
   - pytest -v -s plugins/lora_resolvers # unit tests for in-tree lora resolver plugins
 
 - label: Multi-step Tests (4 GPUs) # 36min
-  mirror_hardwares: [amdexperimental]
+  mirror_hardwares: [amdexperimental, amdproduction]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   source_file_dependencies:
@@ -721,7 +788,7 @@ steps:
     - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
 
 - label: Weight Loading Multiple GPU Test - Large Models # optional
-  mirror_hardwares: [amdexperimental] 
+  mirror_hardwares: [amdexperimental]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   gpu: a100

.github/CODEOWNERS

@@ -10,15 +10,21 @@
 /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
-/vllm/model_executor/guided_decoding @mgoin @russellb
+/vllm/model_executor/guided_decoding @mgoin @russellb @aarnphm
 /vllm/multimodal @DarkLight1337 @ywang96
 /vllm/vllm_flash_attn @LucasWilkinson
 /vllm/lora @jeejeelee
+/vllm/reasoning @aarnphm
+/vllm/entrypoints @aarnphm
 CMakeLists.txt @tlrmchlsmth
 
+# Any change to the VllmConfig changes can have a large user-facing impact,
+# so spam a lot of people
+/vllm/config.py @simon-mo @WoosukKwon @youkaichao @robertgshaw2-redhat @mgoin @tlrmchlsmth @houseroad @hmellor
+
 # vLLM V1
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
-/vllm/v1/structured_output @mgoin @russellb
+/vllm/v1/structured_output @mgoin @russellb @aarnphm
 
 # Test ownership
 /.buildkite/lm-eval-harness @mgoin @simon-mo
@@ -27,8 +33,8 @@ CMakeLists.txt @tlrmchlsmth
 /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
-/tests/entrypoints/llm/test_guided_generate.py @mgoin @russellb
+/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo @aarnphm
+/tests/entrypoints/llm/test_guided_generate.py @mgoin @russellb @aarnphm
 /tests/kernels @tlrmchlsmth @WoosukKwon
 /tests/model_executor/test_guided_processors.py @mgoin @russellb
 /tests/models @DarkLight1337 @ywang96
@@ -38,11 +44,11 @@ CMakeLists.txt @tlrmchlsmth
 /tests/quantization @mgoin @robertgshaw2-redhat
 /tests/spec_decode @njhill @LiuXiaoxuanPKU
 /tests/test_inputs.py @DarkLight1337 @ywang96
-/tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb
-/tests/v1/structured_output @mgoin @russellb
+/tests/v1/entrypoints/llm/test_struct_output_generate.py @mgoin @russellb @aarnphm
+/tests/v1/structured_output @mgoin @russellb @aarnphm
 /tests/weight_loading @mgoin @youkaichao
 /tests/lora @jeejeelee
 
 # Docs
 /docs @hmellor
-mkdocs.yaml @hmellor
+mkdocs.yaml @hmellor

.github/ISSUE_TEMPLATE/400-bug-report.yml

@@ -8,6 +8,16 @@ body:
   attributes:
     value: >
       #### Before submitting an issue, please make sure the issue hasn't been already addressed by searching through [the existing and past issues](https://github.com/vllm-project/vllm/issues?q=is%3Aissue+sort%3Acreated-desc+).
+- type: markdown
+  attributes:
+    value: |
+      ⚠️ **SECURITY WARNING:** Please review any text you paste to ensure it does not contain sensitive information such as:
+      - API tokens or keys (e.g., Hugging Face tokens, OpenAI API keys)
+      - Passwords or authentication credentials
+      - Private URLs or endpoints
+      - Personal or confidential data
+      
+      Consider redacting or replacing sensitive values with placeholders like `<YOUR_TOKEN_HERE>` when sharing configuration or code examples.
 - type: textarea
   attributes:
     label: Your current environment

.github/PULL_REQUEST_TEMPLATE.md

@@ -1,6 +1,18 @@
-FILL IN THE PR DESCRIPTION HERE
+## Essential Elements of an Effective PR Description Checklist
+- [ ] The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
+- [ ] The test plan, such as providing test command.
+- [ ] The test results, such as pasting the results comparison before and after, or e2e results
+- [ ] (Optional) The necessary documentation update, such as updating `supported_models.md` and `examples` for a new model.
 
-FIX #xxxx (*link existing issues this PR will resolve*)
+PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE BEEN CONSIDERED.
+
+## Purpose
+
+## Test Plan
+
+## Test Result
+
+## (Optional) Documentation Update
 
 <!--- pyml disable-next-line no-emphasis-as-heading -->
 **BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing>** (anything written below this line will be removed by GitHub Actions)

.github/mergify.yml

@@ -27,6 +27,22 @@ pull_request_rules:
       add:
         - ci/build
 
+- name: label-deepseek
+  description: Automatically apply deepseek label
+  conditions:
+    - or:
+      - files~=^examples/.*deepseek.*\.py
+      - files~=^tests/.*deepseek.*\.py
+      - files~=^vllm/entrypoints/openai/tool_parsers/.*deepseek.*\.py
+      - files~=^vllm/model_executor/models/.*deepseek.*\.py
+      - files~=^vllm/reasoning/.*deepseek.*\.py
+      - files~=^vllm/transformers_utils/.*deepseek.*\.py
+      - title~=(?i)DeepSeek
+  actions:
+    label:
+      add:
+        - deepseek
+
 - name: label-frontend
   description: Automatically apply frontend label
   conditions:
@@ -36,6 +52,21 @@ pull_request_rules:
       add:
         - frontend
 
+- name: label-llama
+  description: Automatically apply llama label
+  conditions:
+    - or:
+      - files~=^examples/.*llama.*\.py
+      - files~=^tests/.*llama.*\.py
+      - files~=^vllm/entrypoints/openai/tool_parsers/llama.*\.py
+      - files~=^vllm/model_executor/models/.*llama.*\.py
+      - files~=^vllm/transformers_utils/configs/.*llama.*\.py
+      - title~=(?i)llama
+  actions:
+    label:
+      add:
+        - llama
+
 - name: label-multi-modality
   description: Automatically apply multi-modality label
   conditions:
@@ -51,6 +82,53 @@ pull_request_rules:
       add:
         - multi-modality
 
+- name: label-performance
+  description: Automatically apply performance label
+  conditions:
+    - or:
+      - files~=^benchmarks/
+      - files~=^vllm/benchmarks/
+      - files~=^tests/benchmarks/
+      - files~=^\.buildkite/nightly-benchmarks/
+  actions:
+    label:
+      add:
+        - performance
+
+- name: label-qwen
+  description: Automatically apply qwen label
+  conditions:
+    - or:
+      - files~=^examples/.*qwen.*\.py
+      - files~=^tests/.*qwen.*\.py
+      - files~=^vllm/model_executor/models/.*qwen.*\.py
+      - files~=^vllm/reasoning/.*qwen.*\.py
+      - title~=(?i)Qwen
+  actions:
+    label:
+      add:
+        - qwen
+
+- name: label-rocm
+  description: Automatically apply rocm label
+  conditions:
+    - or:
+      - files~=^csrc/rocm/
+      - files~=^docker/Dockerfile.rocm
+      - files~=^requirements/rocm.*\.txt
+      - files~=^vllm/attention/backends/rocm.*\.py
+      - files~=^vllm/attention/ops/rocm.*\.py
+      - files~=^vllm/model_executor/layers/fused_moe/rocm.*\.py
+      - files~=^vllm/v1/attention/backends/mla/rocm.*\.py
+      - files~=^tests/kernels/.*_rocm.*\.py
+      - files=vllm/platforms/rocm.py
+      - title~=(?i)AMD
+      - title~=(?i)ROCm
+  actions:
+    label:
+      add:
+        - rocm
+
 - name: label-structured-output
   description: Automatically apply structured-output label
   conditions:

.gitignore

@@ -200,5 +200,5 @@ benchmarks/**/*.json
 actionlint
 shellcheck*/
 
-# Ingore moe/marlin_moe gen code
+# Ignore moe/marlin_moe gen code
 csrc/moe/marlin_moe_wna16/kernel_*

.pre-commit-config.yaml

@@ -11,6 +11,8 @@ repos:
   hooks:
   - id: yapf
     args: [--in-place, --verbose]
+    # Keep the same list from yapfignore here to avoid yapf failing without any inputs
+    exclude: '(.buildkite|benchmarks|build|examples)/.*'
 - repo: https://github.com/astral-sh/ruff-pre-commit
   rev: v0.11.7
   hooks:
@@ -18,12 +20,10 @@ repos:
     args: [--output-format, github, --fix]
   - id: ruff-format
     files: ^(.buildkite|benchmarks|examples)/.*
-- repo: https://github.com/codespell-project/codespell
-  rev: v2.4.1
+- repo: https://github.com/crate-ci/typos
+  rev: v1.32.0
   hooks:
-  - id: codespell
-    additional_dependencies: ['tomli']
-    args: ['--toml', 'pyproject.toml']
+  - id: typos
 - repo: https://github.com/PyCQA/isort
   rev: 6.0.1
   hooks:
@@ -53,12 +53,17 @@ repos:
       files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
+  - id: format-torch-nightly-test
+    name: reformat nightly_torch_test.txt to be in sync with test.in
+    language: python
+    entry: python tools/generate_nightly_torch_test.py
+    files: ^requirements/test\.(in|txt)$
   - id: mypy-local
     name: Run mypy for local Python installation
     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]
+    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
   - 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
@@ -115,6 +120,11 @@ repos:
     entry: python tools/check_spdx_header.py
     language: python
     types: [python]
+  - id: check-root-lazy-imports
+    name: Check root lazy imports
+    entry: python tools/check_init_lazy_imports.py
+    language: python
+    types: [python]
   - id: check-filenames
     name: Check for spaces in all filenames
     entry: bash
@@ -143,6 +153,20 @@ repos:
     types: [python]
     pass_filenames: false
     additional_dependencies: [regex]
+  - id: check-pickle-imports
+    name: Prevent new pickle/cloudpickle imports
+    entry: python tools/check_pickle_imports.py
+    language: python
+    types: [python]
+    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
+    types: [python]
+    pass_filenames: true
+    files: vllm/config.py|tests/test_config.py
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion

CMakeLists.txt

@@ -23,6 +23,9 @@ include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)
 # Suppress potential warnings about unused manually-specified variables
 set(ignoreMe "${VLLM_PYTHON_PATH}")
 
+# Prevent installation of dependencies (cutlass) by default.
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
+
 #
 # Supported python versions.  These versions will be searched in order, the
 # first match will be selected.  These should be kept in sync with setup.py.
@@ -179,9 +182,6 @@ include(FetchContent)
 file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
 message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
 
-#
-# Set rocm version dev int.
-#
 if(VLLM_GPU_LANG STREQUAL "HIP")
   #
   # Overriding the default -O set up by cmake, adding ggdb3 for the most verbose devug info
@@ -189,7 +189,6 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
   set(CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG "${CMAKE_${VLLM_GPU_LANG}_FLAGS_DEBUG} -O0 -ggdb3")
   set(CMAKE_CXX_FLAGS_DEBUG "${CMAKE_CXX_FLAGS_DEBUG} -O0 -ggdb3")
 
-
   #
   # Certain HIP functions are marked as [[nodiscard]], yet vllm ignores the result which generates
   # a lot of warnings that always mask real issues. Suppressing until this is properly addressed.
@@ -243,6 +242,7 @@ set(VLLM_EXT_SRC
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
   "csrc/layernorm_quant_kernels.cu"
+  "csrc/sampler.cu"
   "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
@@ -308,7 +308,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # Keep building Marlin for 9.0 as there are some group sizes and shapes that
   # are not supported by Machete yet.
   # 9.0 for latest bf16 atomicAdd PTX
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;9.0+PTX" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
 
     #
@@ -420,9 +420,39 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  # The cutlass_scaled_mm kernels for Blackwell (c3x, i.e. CUTLASS 3.x) require
+
+  # The cutlass_scaled_mm kernels for Geforce Blackwell SM120 (c3x, i.e. CUTLASS 3.x) require
   # CUDA 12.8 or later
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a;12.0a" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "12.0;12.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm120.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm120_fp8.cu"
+    )
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM120=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm120 for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm120 as CUDA Compiler version is "
+                     "not >= 12.8, we recommend upgrading to CUDA 12.8 or "
+                     "later if you intend on running FP8 quantized models on "
+                     "Blackwell.")
+    else()
+      message(STATUS "Not building scaled_mm_c3x_120 as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
+
+
+  # The cutlass_scaled_mm kernels for Blackwell SM100 (c3x, i.e. CUTLASS 3.x)
+  # require CUDA 12.8 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "10.0a;10.1a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
     set(SRCS
       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
@@ -454,7 +484,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # kernels for the remaining archs that are not already built for 3x.
   # (Build 8.9 for FP8)
   cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
-    "7.5;8.0;8.9+PTX" "${CUDA_ARCHS}")
+    "7.5;8.0;8.7;8.9+PTX" "${CUDA_ARCHS}")
   # subtract out the archs that are already built for 3x
   list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})
   if (SCALED_MM_2X_ARCHS)
@@ -513,6 +543,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       CUDA_ARCHS "${FP4_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_NVFP4=1")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM100=1")
     message(STATUS "Building NVFP4 for archs: ${FP4_ARCHS}")
   else()
     message(STATUS "Not building NVFP4 as no compatible archs were found.")
@@ -542,13 +573,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # CUTLASS MoE kernels
 
-  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and only works
-  # on Hopper). get_cutlass_moe_mm_data should only be compiled if it's possible
-  # to compile MoE kernels that use its output.
-  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
+  # The MoE kernel cutlass_moe_mm requires CUDA 12.3 or later (and ONLY works
+  # on Hopper). get_cutlass_(pplx_)moe_mm_data should only be compiled
+  # if it's possible to compile MoE kernels that use its output.
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a" "${CUDA_ARCHS}")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
-    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu"
-             "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/grouped_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${SCALED_MM_ARCHS}")
@@ -562,7 +592,27 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                      "if you intend on running FP8 quantized MoE models on Hopper.")
     else()
       message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
-                     "in CUDA target architectures")
+                     "in CUDA target architectures.")
+    endif()
+  endif()
+
+  # moe_data.cu is used by all CUTLASS MoE kernels.
+  cuda_archs_loose_intersection(CUTLASS_MOE_DATA_ARCHS "9.0a;10.0a" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
+    set(SRCS "csrc/quantization/cutlass_w8a8/moe/moe_data.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${CUTLASS_MOE_DATA_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    message(STATUS "Building moe_data for archs: ${CUTLASS_MOE_DATA_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND CUTLASS_MOE_DATA_ARCHS)
+      message(STATUS "Not building moe_data as CUDA Compiler version is "
+                     "not >= 12.3, we recommend upgrading to CUDA 12.3 or later "
+                     "if you intend on running FP8 quantized MoE models on Hopper or Blackwell.")
+    else()
+      message(STATUS "Not building moe_data as no compatible archs found "
+                     "in CUDA target architectures.")
     endif()
   endif()
 
@@ -638,6 +688,14 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 # if CUDA endif
 endif()
 
+if (VLLM_GPU_LANG STREQUAL "HIP")
+  # Add QuickReduce kernels
+  list(APPEND VLLM_EXT_SRC
+    "csrc/custom_quickreduce.cu"
+  )
+# if ROCM endif
+endif()
+
 message(STATUS "Enabling C extension.")
 define_gpu_extension_target(
   _C
@@ -684,7 +742,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
   # 9.0 for latest bf16 atomicAdd PTX
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;9.0+PTX" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.7;9.0+PTX" "${CUDA_ARCHS}")
   if (MARLIN_MOE_ARCHS)
 
     #
@@ -785,5 +843,7 @@ endif()
 # For CUDA we also build and ship some external projects.
 if (VLLM_GPU_LANG STREQUAL "CUDA")
     include(cmake/external_projects/flashmla.cmake)
+
+    # vllm-flash-attn should be last as it overwrites some CMake functions
     include(cmake/external_projects/vllm_flash_attn.cmake)
 endif ()

README.md

@@ -58,8 +58,8 @@ vLLM is fast with:
 - Efficient management of attention key and value memory with [**PagedAttention**](https://blog.vllm.ai/2023/06/20/vllm.html)
 - Continuous batching of incoming requests
 - Fast model execution with CUDA/HIP graph
-- Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [AutoRound](https://arxiv.org/abs/2309.05516),INT4, INT8, and FP8.
-- Optimized CUDA kernels, including integration with FlashAttention and FlashInfer.
+- Quantizations: [GPTQ](https://arxiv.org/abs/2210.17323), [AWQ](https://arxiv.org/abs/2306.00978), [AutoRound](https://arxiv.org/abs/2309.05516), INT4, INT8, and FP8
+- Optimized CUDA kernels, including integration with FlashAttention and FlashInfer
 - Speculative decoding
 - Chunked prefill
 
@@ -72,14 +72,14 @@ vLLM is flexible and easy to use with:
 - Tensor parallelism and pipeline parallelism support for distributed inference
 - Streaming outputs
 - OpenAI-compatible API server
-- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron.
+- Support NVIDIA GPUs, AMD CPUs and GPUs, Intel CPUs and GPUs, PowerPC CPUs, TPU, and AWS Neuron
 - Prefix caching support
 - Multi-LoRA support
 
 vLLM seamlessly supports most popular open-source models on HuggingFace, including:
 - Transformer-like LLMs (e.g., Llama)
 - Mixture-of-Expert LLMs (e.g., Mixtral, Deepseek-V2 and V3)
-- Embedding Models (e.g. E5-Mistral)
+- Embedding Models (e.g., E5-Mistral)
 - Multi-modal LLMs (e.g., LLaVA)
 
 Find the full list of supported models [here](https://docs.vllm.ai/en/latest/models/supported_models.html).
@@ -154,12 +154,14 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 
 ## Contact Us
 
+<!-- --8<-- [start:contact-us] -->
 - For technical questions and feature requests, please use GitHub [Issues](https://github.com/vllm-project/vllm/issues) or [Discussions](https://github.com/vllm-project/vllm/discussions)
 - For discussing with fellow users, please use the [vLLM Forum](https://discuss.vllm.ai)
-- coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
+- For coordinating contributions and development, please use [Slack](https://slack.vllm.ai)
 - For security disclosures, please use GitHub's [Security Advisories](https://github.com/vllm-project/vllm/security/advisories) feature
 - For collaborations and partnerships, please contact us at [vllm-questions@lists.berkeley.edu](mailto:vllm-questions@lists.berkeley.edu)
+<!-- --8<-- [end:contact-us] -->
 
 ## Media Kit
 
-- If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit).
+- If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit)

SECURITY.md

@@ -8,4 +8,6 @@ Please report security issues privately using [the vulnerability submission form
 
 ---
 
+Please see the [Security Guide in the vLLM documentation](https://docs.vllm.ai/en/latest/usage/security.html) for more information on vLLM's security assumptions and recommendations.
+
 Please see [PyTorch's Security Policy](https://github.com/pytorch/pytorch/blob/main/SECURITY.md) for more information and recommendations on how to securely interact with models.

benchmarks/README.md

@@ -4,7 +4,7 @@ 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.
 
-## Dataset Overview
+**Dataset Overview**
 
 <table style="width:100%; border-collapse: collapse;">
   <thead>
@@ -64,6 +64,12 @@ become available.
       <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>
 
@@ -76,7 +82,10 @@ become available.
 **Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
 
 ---
-## Example - Online Benchmark
+<details>
+<summary><b>🚀 Example - Online Benchmark</b></summary>
+
+<br/>
 
 First start serving your model
 
@@ -124,7 +133,40 @@ P99 ITL (ms):                            8.39
 ==================================================
 ```
 
-### VisionArena Benchmark for Vision Language Models
+**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
+
+```
+{"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 --disable-log-requests
+```
+
+```bash
+# run benchmarking script
+python3 benchmarks/benchmark_serving.py --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
@@ -142,13 +184,13 @@ python3 vllm/benchmarks/benchmark_serving.py \
   --num-prompts 1000
 ```
 
-### InstructCoder Benchmark with Speculative Decoding
+**InstructCoder Benchmark with Speculative Decoding**
 
 ``` bash
 VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
-    --ngram_prompt_lookup_min 2 \
-    --ngram-prompt-lookup-max 5 \
-    --speculative_config '{"model": "[ngram]", "num_speculative_tokens": 5}
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
 ```
 
 ``` bash
@@ -159,7 +201,7 @@ python3 benchmarks/benchmark_serving.py \
     --num-prompts 2048
 ```
 
-### Other HuggingFaceDataset Examples
+**Other HuggingFaceDataset Examples**
 
 ```bash
 vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
@@ -203,7 +245,17 @@ python3 vllm/benchmarks/benchmark_serving.py \
     --seed 42
 ```
 
-### Running With Sampling Parameters
+**`philschmid/mt-bench`**
+
+``` bash
+python3 vllm/benchmarks/benchmark_serving.py \
+    --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:
@@ -221,8 +273,27 @@ python3 vllm/benchmarks/benchmark_serving.py \
   --num-prompts 10
 ```
 
----
-## Example - Offline Throughput Benchmark
+**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>
+
+<details>
+<summary><b>📈 Example - Offline Throughput Benchmark</b></summary>
+
+<br/>
 
 ```bash
 python3 vllm/benchmarks/benchmark_throughput.py \
@@ -240,7 +311,7 @@ Total num prompt tokens:  5014
 Total num output tokens:  1500
 ```
 
-### VisionArena Benchmark for Vision Language Models
+**VisionArena Benchmark for Vision Language Models**
 
 ``` bash
 python3 vllm/benchmarks/benchmark_throughput.py \
@@ -260,7 +331,7 @@ Total num prompt tokens:  14527
 Total num output tokens:  1280
 ```
 
-### InstructCoder Benchmark with Speculative Decoding
+**InstructCoder Benchmark with Speculative Decoding**
 
 ``` bash
 VLLM_WORKER_MULTIPROC_METHOD=spawn \
@@ -273,9 +344,9 @@ python3 vllm/benchmarks/benchmark_throughput.py \
     --output-len=100 \
     --num-prompts=2048 \
     --async-engine \
-    --ngram_prompt_lookup_min=2 \
-    --ngram-prompt-lookup-max=5 \
-    --speculative_config '{"model": "[ngram]", "num_speculative_tokens": 5}
+    --speculative-config $'{"method": "ngram",
+    "num_speculative_tokens": 5, "prompt_lookup_max": 5,
+    "prompt_lookup_min": 2}'
 ```
 
 ```
@@ -284,7 +355,7 @@ Total num prompt tokens:  261136
 Total num output tokens:  204800
 ```
 
-### Other HuggingFaceDataset Examples
+**Other HuggingFaceDataset Examples**
 
 **`lmms-lab/LLaVA-OneVision-Data`**
 
@@ -323,7 +394,7 @@ python3 benchmarks/benchmark_throughput.py \
   --num-prompts 10
 ```
 
-### Benchmark with LoRA Adapters
+**Benchmark with LoRA Adapters**
 
 ``` bash
 # download dataset
@@ -339,3 +410,196 @@ python3 vllm/benchmarks/benchmark_throughput.py \
   --enable-lora \
   --lora-path yard1/llama-2-7b-sql-lora-test
   ```
+
+</details>
+
+<details>
+<summary><b>🛠️ Example - Structured Output Benchmark</b></summary>
+
+<br/>
+
+Benchmark the performance of structured output generation (JSON, grammar, regex).
+
+**Server Setup**
+
+```bash
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
+```
+
+**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>
+
+<details>
+<summary><b>📚 Example - Long Document QA Benchmark</b></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>
+
+<details>
+<summary><b>🗂️ Example - Prefix Caching Benchmark</b></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
+```
+
+</details>
+
+<details>
+<summary><b>⚡ Example - Request Prioritization Benchmark</b></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>

benchmarks/auto_tune.sh

@@ -10,11 +10,16 @@
 # 3. Set variables (ALL REQUIRED)
 #   BASE: your directory for vllm repo
 #   MODEL: the model served by vllm
+#   SYSTEM: the hardware, choice TPU or GPU, for other systems, "get best profile" might not support.
+#   TP: ways of tensor parallelism
 #   DOWNLOAD_DIR: directory to download and load model weights.
 #   INPUT_LEN: request input len
 #   OUTPUT_LEN: request output len
 #   MIN_CACHE_HIT_PCT: prefix cache rate
 #   MAX_LATENCY_ALLOWED_MS: (e2e) latency requirement. If there's no latency requirement, set it to a large number like 1000000000
+#   NUM_SEQS_LIST: a list of `max-num-seqs` you want to loop with.
+#   NUM_BATCHED_TOKENS_LIST: a list of `max-num-batched-tokens` you want to loop with.
+#   Note that the default NUM_SEQS_LIST and NUM_BATCHED_TOKENS_LIST are set for medium size input/output len, for extra short context (such as 20:20), you might need to include larger numbers in NUM_SEQS_LIST.
 # 4. Run the script, it might take a long time, you can use tmux to avoid the script stop if disconnection happens.
 # 5. The final result will be saved in RESULT file. 
 
@@ -30,31 +35,31 @@
 TAG=$(date +"%Y_%m_%d_%H_%M")
 BASE=""
 MODEL="meta-llama/Llama-3.1-8B-Instruct"
+SYSTEM="TPU"
+TP=1
 DOWNLOAD_DIR=""
 INPUT_LEN=4000
 OUTPUT_LEN=16
-MIN_CACHE_HIT_PCT_PCT=0
+MIN_CACHE_HIT_PCT=0
 MAX_LATENCY_ALLOWED_MS=100000000000
+NUM_SEQS_LIST="128 256"
+NUM_BATCHED_TOKENS_LIST="512 1024 2048 4096"
 
 LOG_FOLDER="$BASE/auto-benchmark/$TAG"
 RESULT="$LOG_FOLDER/result.txt"
+PROFILE_PATH="$LOG_FOLDER/profile"
 
-echo "result file$ $RESULT"
+echo "result file: $RESULT"
 echo "model: $MODEL"
-echo
 
 rm -rf $LOG_FOLDER
+rm -rf $PROFILE_PATH
 mkdir -p $LOG_FOLDER
+mkdir -p $PROFILE_PATH
 
 cd "$BASE/vllm"
-# create sonnet-4x.txt so that we can sample 2048 tokens for input
-echo "" > benchmarks/sonnet_4x.txt
-for _ in {1..4}
-do
-cat benchmarks/sonnet.txt >> benchmarks/sonnet_4x.txt
-done
 
-pip install datasets
+pip install -q datasets
 
 current_hash=$(git rev-parse HEAD)
 echo "hash:$current_hash" >> "$RESULT"
@@ -64,53 +69,88 @@ best_throughput=0
 best_max_num_seqs=0
 best_num_batched_tokens=0
 best_goodput=0
-run_benchmark() {
-    local max_num_seqs=$1
-    local max_num_batched_tokens=$2
-    echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
-    local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
-    echo "vllm_log: $vllm_log"
-    echo
-    rm -f $vllm_log
 
-    # start the server
-    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 vllm serve $MODEL \
+start_server() {
+    local gpu_memory_utilization=$1
+    local max_num_seqs=$2
+    local max_num_batched_tokens=$3
+    local vllm_log=$4
+    local profile_dir=$5
+    
+    pkill -f vllm
+
+    VLLM_USE_V1=1 VLLM_SERVER_DEV_MODE=1 VLLM_TORCH_PROFILER_DIR=$profile_dir vllm serve $MODEL \
         --disable-log-requests \
         --port 8004 \
-        --gpu-memory-utilization 0.98 \
+        --gpu-memory-utilization $gpu_memory_utilization \
         --max-num-seqs $max_num_seqs \
         --max-num-batched-tokens $max_num_batched_tokens \
-        --tensor-parallel-size 1 \
+        --tensor-parallel-size $TP \
         --enable-prefix-caching \
         --load-format dummy \
-        --download-dir $DOWNLOAD_DIR \
+        --download-dir "$DOWNLOAD_DIR" \
         --max-model-len $(( INPUT_LEN+OUTPUT_LEN )) > "$vllm_log" 2>&1 &
-    echo "wait for 10 minutes.."
-    echo
+
     # wait for 10 minutes...
     server_started=0
-    for i in {1..60}; do        
-        if grep -Fq "Application startup complete" "$vllm_log"; then
-            echo "Application started"
+    for i in {1..60}; do  
+        RESPONSE=$(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
             server_started=1
             break
         else
-            # echo "wait for 10 seconds..."
             sleep 10
         fi
     done
- 
     if (( ! server_started )); then
-        echo "server did not start within 10 minutes, terminate the benchmarking. Please check server log at $vllm_log"
-        echo "pkill -f vllm"
-        echo
-        pkill vllm
-        sleep 10
+        echo "server did not start within 10 minutes. Please check server log at $vllm_log".
         return 1
+    else
+        return 0
     fi
+}
+
+update_best_profile() {
+    local profile_dir=$1
+    local profile_index=$2
+    sorted_paths=($(find "$profile_dir" -maxdepth 1 -not -path "$profile_dir" | sort))
+    selected_profile_file=
+    if [[ "$SYSTEM" == "TPU" ]]; then
+        selected_profile_file="${sorted_paths[$profile_index]}/*.xplane.pb"
+    fi 
+    if [[ "$SYSTEM" == "GPU" ]]; then
+        selected_profile_file="${sorted_paths[$profile_index]}"
+    fi 
+    rm -f $PROFILE_PATH/*
+    cp $selected_profile_file $PROFILE_PATH
+}
+
+run_benchmark() {
+    local max_num_seqs=$1
+    local max_num_batched_tokens=$2
+    local gpu_memory_utilization=$3
+    echo "max_num_seq: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
+    local vllm_log="$LOG_FOLDER/vllm_log_${max_num_seqs}_${max_num_batched_tokens}.txt"
+    local profile_dir="$LOG_FOLDER/profile_${max_num_seqs}_${max_num_batched_tokens}"
+    echo "vllm_log: $vllm_log"
+    echo
+    rm -f $vllm_log
+    mkdir -p $profile_dir
+    pkill -f vllm
+    local profile_index=0
+
+    echo "starting server..."
+    start_server $gpu_memory_utilization $max_num_seqs $max_num_batched_tokens $vllm_log $profile_dir
+    result=$?
+    if [[ "$result" -eq 1 ]]; then
+        echo "server failed to start. gpu_memory_utilization:$gpu_memory_utilization, max_num_seqs:$max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens"
+    else
+        echo "server started."
+    fi
+    echo
     
     echo "run benchmark test..."
-    echo
     meet_latency_requirement=0
     # get a basic qps by using request-rate inf
     bm_log="$LOG_FOLDER/bm_log_${max_num_seqs}_${max_num_batched_tokens}_requestrate_inf.txt"
@@ -118,30 +158,32 @@ run_benchmark() {
     python benchmarks/benchmark_serving.py \
         --backend vllm \
         --model $MODEL  \
-        --dataset-name sonnet \
-        --dataset-path benchmarks/sonnet_4x.txt \
-        --sonnet-input-len $INPUT_LEN \
-        --sonnet-output-len $OUTPUT_LEN \
+        --dataset-name random \
+        --random-input-len $INPUT_LEN \
+        --random-output-len $OUTPUT_LEN \
         --ignore-eos \
         --disable-tqdm \
         --request-rate inf \
         --percentile-metrics ttft,tpot,itl,e2el \
         --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
-        --num-prompts 100 \
-        --sonnet-prefix-len $prefix_len \
-        --port 8004 > "$bm_log"
-    through_put=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+        --num-prompts 1000 \
+        --random-prefix-len $prefix_len \
+        --port 8004 \
+        --profile &> "$bm_log"
+    throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
     e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
     goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
 
     if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
         meet_latency_requirement=1
+        request_rate=inf
     fi
 
     if (( ! meet_latency_requirement )); then
-    # start from request-rate as int(through_put) + 1
-        request_rate=$((${through_put%.*} + 1))
+    # start from request-rate as int(throughput) + 1
+        request_rate=$((${throughput%.*} + 1))
         while ((request_rate > 0)); do
+            profile_index=$((profile_index+1))
             # clear prefix cache
             curl -X POST http://0.0.0.0:8004/reset_prefix_cache
             sleep 5
@@ -149,19 +191,18 @@ run_benchmark() {
             python benchmarks/benchmark_serving.py \
                 --backend vllm \
                 --model $MODEL  \
-                --dataset-name sonnet \
-                --dataset-path benchmarks/sonnet_4x.txt \
-                --sonnet-input-len $INPUT_LEN \
-                --sonnet-output-len $OUTPUT_LEN \
-                --ignore_eos \
+                --dataset-name random \
+                --random-input-len $INPUT_LEN \
+                --random-output-len $OUTPUT_LEN \
+                --ignore-eos \
                 --disable-tqdm \
                 --request-rate $request_rate \
                 --percentile-metrics ttft,tpot,itl,e2el \
                 --goodput e2el:$MAX_LATENCY_ALLOWED_MS \
                 --num-prompts 100 \
-                --sonnet-prefix-len $prefix_len \
-                --port 8004 > "$bm_log"
-            through_put=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
+                --random-prefix-len $prefix_len \
+                --port 8004 &> "$bm_log"
+            throughput=$(grep "Request throughput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
             e2el=$(grep "P99 E2EL (ms):" "$bm_log" | awk '{print $NF}')
             goodput=$(grep "Request goodput (req/s):" "$bm_log" | sed 's/[^0-9.]//g')
             if (( $(echo "$e2el <= $MAX_LATENCY_ALLOWED_MS" | bc -l) )); then
@@ -173,13 +214,19 @@ run_benchmark() {
     fi
     # write the results and update the best result.
     if ((meet_latency_requirement)); then
-        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, through put: $through_put, goodput: $goodput"
-        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, through put: $through_put, goodput: $goodput" >> "$RESULT"
-        if (( $(echo "$through_put > $best_throughput" | bc -l) )); then
-            best_throughput=$through_put
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput"
+        echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens, request_rate: $request_rate, e2el: $e2el, throughput: $throughput, goodput: $goodput" >> "$RESULT"
+        if (( $(echo "$throughput > $best_throughput" | bc -l) )); then
+            best_throughput=$throughput
             best_max_num_seqs=$max_num_seqs
             best_num_batched_tokens=$max_num_batched_tokens
             best_goodput=$goodput
+            if [[ "$SYSTEM" == "TPU" ]]; then
+                update_best_profile "$profile_dir/plugins/profile" $profile_index
+            fi
+            if [[ "$SYSTEM" == "GPU" ]]; then
+                update_best_profile "$profile_dir" $profile_index
+            fi
         fi
     else
         echo "max_num_seqs: $max_num_seqs, max_num_batched_tokens: $max_num_batched_tokens does not meet latency requirement ${MAX_LATENCY_ALLOWED_MS}"
@@ -188,25 +235,42 @@ run_benchmark() {
 
     echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
 
-    echo "pkill -f vllm"
-    echo
     pkill vllm
     sleep 10
-    rm -f $vllm_log
     printf '=%.0s' $(seq 1 20)
     return 0
 }
 
+read -r -a num_seqs_list <<< "$NUM_SEQS_LIST"
+read -r -a num_batched_tokens_list <<< "$NUM_BATCHED_TOKENS_LIST"
 
-num_seqs_list="128 256"
-num_batched_tokens_list="512 1024 2048 4096"
-for num_seqs in $num_seqs_list; do
-    for num_batched_tokens in $num_batched_tokens_list; do
-        run_benchmark $num_seqs $num_batched_tokens
-        exit 0
+# first find out the max gpu-memory-utilization without HBM OOM.
+gpu_memory_utilization=0.98
+find_gpu_memory_utilization=0
+while (( $(echo "$gpu_memory_utilization >= 0.9" | bc -l) )); do
+    start_server $gpu_memory_utilization "${num_seqs_list[-1]}" "${num_batched_tokens_list[-1]}" "$LOG_FOLDER/vllm_log_gpu_memory_utilization_$gpu_memory_utilization.log"
+    result=$?
+    if [[ "$result" -eq 0 ]]; then
+        find_gpu_memory_utilization=1
+        break
+    else
+        gpu_memory_utilization=$(echo "$gpu_memory_utilization - 0.01" | bc)
+    fi
+done
+
+if [[ "$find_gpu_memory_utilization" -eq 1 ]]; then
+    echo "Using gpu_memory_utilization=$gpu_memory_utilization to serve model."
+else
+    echo "Cannot find a proper gpu_memory_utilization over 0.9 to serve the model, please check logs in $LOG_FOLDER."
+    exit 1
+fi
+
+for num_seqs in "${num_seqs_list[@]}"; do
+    for num_batched_tokens in "${num_batched_tokens_list[@]}"; do
+        run_benchmark $num_seqs $num_batched_tokens $gpu_memory_utilization
     done
 done
 echo "finish permutations"
-echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput"
-echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput" >> "$RESULT"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH"
+echo "best_max_num_seqs: $best_max_num_seqs, best_num_batched_tokens: $best_num_batched_tokens, best_throughput: $best_throughput, profile saved in: $PROFILE_PATH" >> "$RESULT"
 

benchmarks/backend_request_func.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import io
 import json
@@ -324,7 +325,7 @@ async def async_request_openai_completions(
 
                                 most_recent_timestamp = timestamp
                                 generated_text += text or ""
-                            elif usage := data.get("usage"):
+                            if usage := data.get("usage"):
                                 output.output_tokens = usage.get("completion_tokens")
                     if first_chunk_received:
                         output.success = True
@@ -403,8 +404,14 @@ async def async_request_openai_chat_completions(
                         chunk_bytes = chunk_bytes.strip()
                         if not chunk_bytes:
                             continue
+                        chunk_bytes = chunk_bytes.decode("utf-8")
+                        # NOTE: SSE comments (often used as pings) start with a colon.
+                        # These are not JSON data payload and should be skipped.
+                        if chunk_bytes.startswith(":"):
+                            continue
+
+                        chunk = chunk_bytes.removeprefix("data: ")
 
-                        chunk = chunk_bytes.decode("utf-8").removeprefix("data: ")
                         if chunk != "[DONE]":
                             timestamp = time.perf_counter()
                             data = json.loads(chunk)
@@ -611,6 +618,7 @@ ASYNC_REQUEST_FUNCS = {
     "tensorrt-llm": async_request_trt_llm,
     "scalellm": async_request_openai_completions,
     "sglang": async_request_openai_completions,
+    "llama.cpp": async_request_openai_completions,
 }
 
 OPENAI_COMPATIBLE_BACKENDS = [

benchmarks/benchmark_dataset.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 This module defines a framework for sampling benchmark requests from various
 datasets. Each dataset subclass of BenchmarkDataset must implement sample
@@ -9,9 +10,6 @@ generation. Supported dataset types include:
   - BurstGPT
   - HuggingFace
   - VisionArena
-
-TODO: Implement CustomDataset to parse a JSON file and convert its contents into
-SampleRequest instances, similar to the approach used in ShareGPT.
 """
 
 import base64
@@ -351,11 +349,12 @@ class RandomDataset(BenchmarkDataset):
             # [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']
             # To avoid uncontrolled change of the prompt length,
             # the encoded sequence is truncated before being decode again.
+            total_input_len = prefix_len + int(input_lens[i])
             re_encoded_sequence = tokenizer.encode(prompt, add_special_tokens=False)[
-                : input_lens[i]
+                :total_input_len
             ]
             prompt = tokenizer.decode(re_encoded_sequence)
-            total_input_len = prefix_len + int(input_lens[i])
+            total_input_len = len(re_encoded_sequence)
             requests.append(
                 SampleRequest(
                     prompt=prompt,
@@ -442,6 +441,97 @@ class ShareGPTDataset(BenchmarkDataset):
         return samples
 
 
+# -----------------------------------------------------------------------------
+# Custom Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class CustomDataset(BenchmarkDataset):
+    """
+    Implements the Custom dataset.  Loads data from a JSONL file and generates
+    sample requests based on conversation turns. E.g.,
+    ```
+    {"prompt": "What is the capital of India?"}
+    {"prompt": "What is the capital of Iran?"}
+    {"prompt": "What is the capital of China?"}
+    ```
+    """
+
+    def __init__(self, **kwargs) -> None:
+        super().__init__(**kwargs)
+        self.load_data()
+
+    def load_data(self) -> None:
+        if self.dataset_path is None:
+            raise ValueError("dataset_path must be provided for loading data.")
+
+        # self.data will be a list of dictionaries
+        # e.g., [{"prompt": "What is the capital of India?"}, ...]
+        # This will be the standardized format which load_data()
+        # has to convert into depending on the filetype of dataset_path.
+        # sample() will assume this standardized format of self.data
+        self.data = []
+
+        # Load the JSONL file
+        if self.dataset_path.endswith(".jsonl"):
+            jsonl_data = pd.read_json(path_or_buf=self.dataset_path, lines=True)
+
+            # check if the JSONL file has a 'prompt' column
+            if "prompt" not in jsonl_data.columns:
+                raise ValueError("JSONL file must contain a 'prompt' column.")
+
+            # Convert each row to a dictionary and append to self.data
+            # This will convert the DataFrame to a list of dictionaries
+            # where each dictionary corresponds to a row in the DataFrame.
+            # This is the standardized format we want for self.data
+            for _, row in jsonl_data.iterrows():
+                self.data.append(row.to_dict())
+        else:
+            raise NotImplementedError(
+                "Only JSONL format is supported for CustomDataset."
+            )
+
+        random.seed(self.random_seed)
+        random.shuffle(self.data)
+
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        lora_path: Optional[str] = None,
+        max_loras: Optional[int] = None,
+        output_len: Optional[int] = None,
+        enable_multimodal_chat: bool = False,
+        skip_chat_template: bool = False,
+        **kwargs,
+    ) -> list:
+        sampled_requests = []
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            prompt = item["prompt"]
+
+            # apply template
+            if not skip_chat_template:
+                prompt = tokenizer.apply_chat_template(
+                    [{"role": "user", "content": prompt}],
+                    add_generation_prompt=True,
+                    tokenize=False,
+                )
+
+            prompt_len = len(tokenizer(prompt).input_ids)
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                )
+            )
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+
+        return sampled_requests
+
+
 # -----------------------------------------------------------------------------
 # Sonnet Dataset Implementation
 # -----------------------------------------------------------------------------
@@ -776,7 +866,15 @@ class InstructCoderDataset(HuggingFaceDataset):
         for item in self.data:
             if len(sampled_requests) >= num_requests:
                 break
-            prompt = f"{item['instruction']}:\n{item['input']}"
+            prompt = f"{item['input']}\n\n{item['instruction']} Just output \
+            the code, do not include any explanation."
+
+            # apply template
+            prompt = tokenizer.apply_chat_template(
+                [{"role": "user", "content": prompt}],
+                add_generation_prompt=True,
+                tokenize=False,
+            )
             prompt_len = len(tokenizer(prompt).input_ids)
             sampled_requests.append(
                 SampleRequest(

benchmarks/benchmark_latency.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark the latency of processing a single batch of requests."""
 
 import argparse
@@ -6,13 +7,12 @@ import dataclasses
 import json
 import os
 import time
-from pathlib import Path
 from typing import Any, Optional
 
 import numpy as np
-import torch
 from tqdm import tqdm
 
+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
@@ -80,17 +80,9 @@ def main(args: argparse.Namespace):
 
     def run_to_completion(profile_dir: Optional[str] = None):
         if profile_dir:
-            with torch.profiler.profile(
-                activities=[
-                    torch.profiler.ProfilerActivity.CPU,
-                    torch.profiler.ProfilerActivity.CUDA,
-                ],
-                on_trace_ready=torch.profiler.tensorboard_trace_handler(
-                    str(profile_dir)
-                ),
-            ) as p:
-                llm_generate()
-            print(p.key_averages().table(sort_by="self_cuda_time_total"))
+            llm.start_profile()
+            llm_generate()
+            llm.stop_profile()
         else:
             start_time = time.perf_counter()
             llm_generate()
@@ -103,11 +95,7 @@ def main(args: argparse.Namespace):
         run_to_completion(profile_dir=None)
 
     if args.profile:
-        profile_dir = args.profile_result_dir
-        if not profile_dir:
-            profile_dir = (
-                Path(".") / "vllm_benchmark_result" / f"latency_result_{time.time()}"
-            )
+        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
@@ -135,7 +123,7 @@ def main(args: argparse.Namespace):
         save_to_pytorch_benchmark_format(args, results)
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(
         description="Benchmark the latency of processing a single batch of "
         "requests till completion."
@@ -164,15 +152,6 @@ if __name__ == "__main__":
         action="store_true",
         help="profile the generation process of a single batch",
     )
-    parser.add_argument(
-        "--profile-result-dir",
-        type=str,
-        default=None,
-        help=(
-            "path to save the pytorch profiler output. Can be visualized "
-            "with ui.perfetto.dev or Tensorboard."
-        ),
-    )
     parser.add_argument(
         "--output-json",
         type=str,
@@ -192,5 +171,16 @@ if __name__ == "__main__":
     # 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__":
+    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_long_document_qa_throughput.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 Offline benchmark to test the long document QA throughput.
 
@@ -141,7 +142,7 @@ def main(args):
     )
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(
         description="Benchmark the performance with or "
         "without automatic prefix caching."
@@ -191,5 +192,11 @@ if __name__ == "__main__":
     )
 
     parser = EngineArgs.add_cli_args(parser)
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_prefix_caching.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 Benchmark the efficiency of prefix caching.
 
@@ -217,7 +218,7 @@ def main(args):
     )
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(
         description="Benchmark the performance with or without "
         "automatic prefix caching."
@@ -267,5 +268,11 @@ if __name__ == "__main__":
     )
 
     parser = EngineArgs.add_cli_args(parser)
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_prioritization.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark offline prioritization."""
 
 import argparse
@@ -160,7 +161,7 @@ def main(args: argparse.Namespace):
             json.dump(results, f, indent=4)
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(description="Benchmark the throughput.")
     parser.add_argument(
         "--backend", type=str, choices=["vllm", "hf", "mii"], default="vllm"
@@ -203,6 +204,12 @@ if __name__ == "__main__":
     )
 
     parser = EngineArgs.add_cli_args(parser)
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

benchmarks/benchmark_serving.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 r"""Benchmark online serving throughput.
 
 On the server side, run one of the following commands:
@@ -32,7 +33,7 @@ import warnings
 from collections.abc import AsyncGenerator, Iterable
 from dataclasses import dataclass
 from datetime import datetime
-from typing import Any, Optional
+from typing import Any, Literal, Optional
 
 import numpy as np
 from tqdm.asyncio import tqdm
@@ -60,6 +61,7 @@ from benchmark_dataset import (
     ASRDataset,
     BurstGPTDataset,
     ConversationDataset,
+    CustomDataset,
     HuggingFaceDataset,
     InstructCoderDataset,
     MTBenchDataset,
@@ -105,14 +107,42 @@ class BenchmarkMetrics:
     percentiles_e2el_ms: list[tuple[float, float]]
 
 
+def _get_current_request_rate(
+    ramp_up_strategy: Optional[Literal["linear", "exponential"]],
+    ramp_up_start_rps: Optional[int],
+    ramp_up_end_rps: Optional[int],
+    request_index: int,
+    total_requests: int,
+    request_rate: float,
+) -> float:
+    if (
+        ramp_up_strategy
+        and ramp_up_start_rps is not None
+        and ramp_up_end_rps is not None
+    ):
+        progress = request_index / max(total_requests - 1, 1)
+        if ramp_up_strategy == "linear":
+            increase = (ramp_up_end_rps - ramp_up_start_rps) * progress
+            return ramp_up_start_rps + increase
+        elif ramp_up_strategy == "exponential":
+            ratio = ramp_up_end_rps / ramp_up_start_rps
+            return ramp_up_start_rps * (ratio**progress)
+        else:
+            raise ValueError(f"Unknown ramp-up strategy: {ramp_up_strategy}")
+    return request_rate
+
+
 async def get_request(
     input_requests: list[SampleRequest],
     request_rate: float,
     burstiness: float = 1.0,
-) -> AsyncGenerator[SampleRequest, None]:
+    ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None,
+    ramp_up_start_rps: Optional[int] = None,
+    ramp_up_end_rps: Optional[int] = None,
+) -> AsyncGenerator[tuple[SampleRequest, float], None]:
     """
     Asynchronously generates requests at a specified rate
-    with OPTIONAL burstiness.
+    with OPTIONAL burstiness and OPTIONAL ramp-up strategy.
 
     Args:
         input_requests:
@@ -127,22 +157,44 @@ async def get_request(
             A lower burstiness value (0 < burstiness < 1) results
             in more bursty requests, while a higher burstiness value
             (burstiness > 1) results in a more uniform arrival of requests.
+         ramp_up_strategy (optional):
+            The ramp-up strategy. Can be "linear" or "exponential".
+            If None, uses constant request rate (specified by request_rate).
+        ramp_up_start_rps (optional):
+            The starting request rate for ramp-up.
+        ramp_up_end_rps (optional):
+            The ending request rate for ramp-up.
     """
-    input_requests: Iterable[SampleRequest] = iter(input_requests)
-
-    # Calculate scale parameter theta to maintain the desired request_rate.
     assert burstiness > 0, (
         f"A positive burstiness factor is expected, but given {burstiness}."
     )
-    theta = 1.0 / (request_rate * burstiness)
+    # Convert to list to get length for ramp-up calculations
+    if isinstance(input_requests, Iterable) and not isinstance(input_requests, list):
+        input_requests = list(input_requests)
+
+    total_requests = len(input_requests)
+    request_index = 0
 
     for request in input_requests:
-        yield request
+        current_request_rate = _get_current_request_rate(
+            ramp_up_strategy,
+            ramp_up_start_rps,
+            ramp_up_end_rps,
+            request_index,
+            total_requests,
+            request_rate,
+        )
 
-        if request_rate == float("inf"):
+        yield request, current_request_rate
+
+        request_index += 1
+
+        if current_request_rate == float("inf"):
             # If the request rate is infinity, then we don't need to wait.
             continue
 
+        theta = 1.0 / (current_request_rate * burstiness)
+
         # Sample the request interval from the gamma distribution.
         # If burstiness is 1, it follows exponential distribution.
         interval = np.random.gamma(shape=burstiness, scale=theta)
@@ -288,6 +340,9 @@ async def benchmark(
     max_concurrency: Optional[int],
     lora_modules: Optional[Iterable[str]],
     extra_body: Optional[dict],
+    ramp_up_strategy: Optional[Literal["linear", "exponential"]] = None,
+    ramp_up_start_rps: Optional[int] = None,
+    ramp_up_end_rps: Optional[int] = None,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -351,7 +406,15 @@ async def benchmark(
 
     distribution = "Poisson process" if burstiness == 1.0 else "Gamma distribution"
 
-    print(f"Traffic request rate: {request_rate}")
+    if ramp_up_strategy is not None:
+        print(
+            f"Traffic ramp-up strategy: {ramp_up_strategy}. Will increase "
+            f"RPS from {ramp_up_start_rps} to {ramp_up_end_rps} RPS over "
+            "the duration of the benchmark."
+        )
+    else:
+        print(f"Traffic request rate: {request_rate} RPS.")
+
     print(f"Burstiness factor: {burstiness} ({distribution})")
     print(f"Maximum request concurrency: {max_concurrency}")
 
@@ -371,7 +434,34 @@ async def benchmark(
 
     benchmark_start_time = time.perf_counter()
     tasks: list[asyncio.Task] = []
-    async for request in get_request(input_requests, request_rate, burstiness):
+
+    rps_change_events = []
+    last_int_rps = -1
+    if ramp_up_strategy is not None and ramp_up_start_rps is not None:
+        last_int_rps = ramp_up_start_rps
+        rps_change_events.append(
+            {
+                "rps": last_int_rps,
+                "timestamp": datetime.now().isoformat(),
+            }
+        )
+
+    async for request, current_request_rate in get_request(
+        input_requests,
+        request_rate,
+        burstiness,
+        ramp_up_strategy,
+        ramp_up_start_rps,
+        ramp_up_end_rps,
+    ):
+        if ramp_up_strategy is not None:
+            current_int_rps = int(current_request_rate)
+            if current_int_rps > last_int_rps:
+                timestamp = datetime.now().isoformat()
+                for rps_val in range(last_int_rps + 1, current_int_rps + 1):
+                    rps_change_events.append({"rps": rps_val, "timestamp": timestamp})
+                last_int_rps = current_int_rps
+
         prompt, prompt_len, output_len, mm_content = (
             request.prompt,
             request.prompt_len,
@@ -395,11 +485,8 @@ async def benchmark(
             ignore_eos=ignore_eos,
             extra_body=extra_body,
         )
-        tasks.append(
-            asyncio.create_task(
-                limited_request_func(request_func_input=request_func_input, pbar=pbar)
-            )
-        )
+        task = limited_request_func(request_func_input=request_func_input, pbar=pbar)
+        tasks.append(asyncio.create_task(task))
     outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
 
     if profile:
@@ -464,7 +551,7 @@ async def benchmark(
         "total_input_tokens": metrics.total_input,
         "total_output_tokens": metrics.total_output,
         "request_throughput": metrics.request_throughput,
-        "request_goodput:": metrics.request_goodput if goodput_config_dict else None,
+        "request_goodput": metrics.request_goodput if goodput_config_dict else None,
         "output_throughput": metrics.output_throughput,
         "total_token_throughput": metrics.total_token_throughput,
         "input_lens": [output.prompt_len for output in outputs],
@@ -475,6 +562,9 @@ async def benchmark(
         "errors": [output.error for output in outputs],
     }
 
+    if rps_change_events:
+        result["rps_change_events"] = rps_change_events
+
     def process_one_metric(
         # E.g., "ttft"
         metric_attribute_name: str,
@@ -608,6 +698,26 @@ def main(args: argparse.Namespace):
     tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
     tokenizer_mode = args.tokenizer_mode
 
+    # Validate ramp-up arguments
+    if args.ramp_up_strategy is not None:
+        if args.request_rate != float("inf"):
+            raise ValueError(
+                "When using ramp-up, do not specify --request-rate. "
+                "The request rate will be controlled by ramp-up parameters. "
+                "Please remove the --request-rate argument."
+            )
+        if args.ramp_up_start_rps is None or args.ramp_up_end_rps is None:
+            raise ValueError(
+                "When using --ramp-up-strategy, both --ramp-up-start-rps and "
+                "--ramp-up-end-rps must be specified"
+            )
+        if args.ramp_up_start_rps < 0 or args.ramp_up_end_rps < 0:
+            raise ValueError("Ramp-up start and end RPS must be non-negative")
+        if args.ramp_up_start_rps > args.ramp_up_end_rps:
+            raise ValueError("Ramp-up start RPS must be less than end RPS")
+        if args.ramp_up_strategy == "exponential" and args.ramp_up_start_rps == 0:
+            raise ValueError("For exponential ramp-up, the start RPS cannot be 0.")
+
     if args.base_url is not None:
         api_url = f"{args.base_url}{args.endpoint}"
         base_url = f"{args.base_url}"
@@ -627,7 +737,16 @@ def main(args: argparse.Namespace):
             "'--dataset-path' if required."
         )
 
-    if args.dataset_name == "sonnet":
+    if args.dataset_name == "custom":
+        dataset = CustomDataset(dataset_path=args.dataset_path)
+        input_requests = dataset.sample(
+            num_requests=args.num_prompts,
+            tokenizer=tokenizer,
+            output_len=args.custom_output_len,
+            skip_chat_template=args.custom_skip_chat_template,
+        )
+
+    elif args.dataset_name == "sonnet":
         dataset = SonnetDataset(dataset_path=args.dataset_path)
         # For the "sonnet" dataset, formatting depends on the backend.
         if args.backend == "openai-chat":
@@ -762,6 +881,10 @@ def main(args: argparse.Namespace):
     if "temperature" not in sampling_params:
         sampling_params["temperature"] = 0.0  # Default to greedy decoding.
 
+    if args.backend == "llama.cpp":
+        # Disable prompt caching in llama.cpp backend
+        sampling_params["cache_prompt"] = False
+
     # Avoid GC processing "static" data - reduce pause times.
     gc.collect()
     gc.freeze()
@@ -787,6 +910,9 @@ def main(args: argparse.Namespace):
             max_concurrency=args.max_concurrency,
             lora_modules=args.lora_modules,
             extra_body=sampling_params,
+            ramp_up_strategy=args.ramp_up_strategy,
+            ramp_up_start_rps=args.ramp_up_start_rps,
+            ramp_up_end_rps=args.ramp_up_end_rps,
         )
     )
 
@@ -819,6 +945,11 @@ def main(args: argparse.Namespace):
         result_json["burstiness"] = args.burstiness
         result_json["max_concurrency"] = args.max_concurrency
 
+        if args.ramp_up_strategy is not None:
+            result_json["ramp_up_strategy"] = args.ramp_up_strategy
+            result_json["ramp_up_start_rps"] = args.ramp_up_start_rps
+            result_json["ramp_up_end_rps"] = args.ramp_up_end_rps
+
         # Merge with benchmark result
         result_json = {**result_json, **benchmark_result}
 
@@ -834,6 +965,8 @@ def main(args: argparse.Namespace):
             ]:
                 if field in result_json:
                     del result_json[field]
+                if field in benchmark_result:
+                    del benchmark_result[field]
 
         # Save to file
         base_model_id = model_id.split("/")[-1]
@@ -842,10 +975,14 @@ def main(args: argparse.Namespace):
             if args.max_concurrency is not None
             else ""
         )
-        file_name = f"{backend}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json"  # noqa
+        if args.ramp_up_strategy is not None:
+            file_name = f"{backend}-ramp-up-{args.ramp_up_strategy}-{args.ramp_up_start_rps}qps-{args.ramp_up_end_rps}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json"  # noqa
+        else:
+            file_name = f"{backend}-{args.request_rate}qps{max_concurrency_str}-{base_model_id}-{current_dt}.json"  # noqa
         if args.result_filename:
             file_name = args.result_filename
         if args.result_dir:
+            os.makedirs(args.result_dir, exist_ok=True)
             file_name = os.path.join(args.result_dir, file_name)
         with open(
             file_name, mode="a+" if args.append_result else "w", encoding="utf-8"
@@ -857,7 +994,7 @@ def main(args: argparse.Namespace):
         save_to_pytorch_benchmark_format(args, result_json, file_name)
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(
         description="Benchmark the online serving throughput."
     )
@@ -886,7 +1023,7 @@ if __name__ == "__main__":
         "--dataset-name",
         type=str,
         default="sharegpt",
-        choices=["sharegpt", "burstgpt", "sonnet", "random", "hf"],
+        choices=["sharegpt", "burstgpt", "sonnet", "random", "hf", "custom"],
         help="Name of the dataset to benchmark on.",
     )
     parser.add_argument(
@@ -1056,6 +1193,19 @@ if __name__ == "__main__":
     )
 
     # group for dataset specific arguments
+    custom_group = parser.add_argument_group("custom dataset options")
+    custom_group.add_argument(
+        "--custom-output-len",
+        type=int,
+        default=256,
+        help="Number of output tokens per request, used only for custom dataset.",
+    )
+    custom_group.add_argument(
+        "--custom-skip-chat-template",
+        action="store_true",
+        help="Skip applying chat template to prompt, used only for custom dataset.",
+    )
+
     sonnet_group = parser.add_argument_group("sonnet dataset options")
     sonnet_group.add_argument(
         "--sonnet-input-len",
@@ -1194,6 +1344,35 @@ if __name__ == "__main__":
         "script chooses a LoRA module at random.",
     )
 
-    args = parser.parse_args()
+    parser.add_argument(
+        "--ramp-up-strategy",
+        type=str,
+        default=None,
+        choices=["linear", "exponential"],
+        help="The ramp-up strategy. This would be used to "
+        "ramp up the request rate from initial RPS to final "
+        "RPS rate (specified by --ramp-up-start-rps and --ramp-up-end-rps). "
+        "over the duration of the benchmark.",
+    )
+    parser.add_argument(
+        "--ramp-up-start-rps",
+        type=int,
+        default=None,
+        help="The starting request rate for ramp-up (RPS). "
+        "Needs to be specified when --ramp-up-strategy is used.",
+    )
+    parser.add_argument(
+        "--ramp-up-end-rps",
+        type=int,
+        default=None,
+        help="The ending request rate for ramp-up (RPS). "
+        "Needs to be specified when --ramp-up-strategy is used.",
+    )
 
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
+    args = parser.parse_args()
     main(args)

benchmarks/benchmark_serving_structured_output.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 r"""Benchmark online serving throughput with structured outputs.
 
 On the server side, run one of the following commands:
@@ -11,7 +12,6 @@ On the client side, run:
         --model <your_model> \
         --dataset json \
         --structured-output-ratio 1.0 \
-        --structured-output-backend auto \
         --request-rate 10 \
         --num-prompts 1000
 
@@ -850,7 +850,7 @@ def main(args: argparse.Namespace):
             json.dump(results, outfile, indent=4)
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(
         description="Benchmark the online serving throughput."
     )
@@ -1034,5 +1034,10 @@ if __name__ == "__main__":
         help="Ratio of Structured Outputs requests",
     )
 
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Benchmark offline inference throughput."""
 
 import argparse
@@ -96,7 +97,7 @@ def run_vllm(
         assert lora_requests is None, "BeamSearch API does not support LoRA"
         prompts = [request.prompt for request in requests]
         # output_len should be the same for all requests.
-        output_len = requests[0][2]
+        output_len = requests[0].expected_output_len
         for request in requests:
             assert request.expected_output_len == output_len
         start = time.perf_counter()
@@ -594,7 +595,7 @@ def validate_args(args):
         )
 
 
-if __name__ == "__main__":
+def create_argument_parser():
     parser = FlexibleArgumentParser(description="Benchmark the throughput.")
     parser.add_argument(
         "--backend",
@@ -716,6 +717,12 @@ if __name__ == "__main__":
     )
 
     parser = AsyncEngineArgs.add_cli_args(parser)
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = create_argument_parser()
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

benchmarks/benchmark_utils.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import json
@@ -65,4 +66,9 @@ class InfEncoder(json.JSONEncoder):
 
 def write_to_json(filename: str, records: list) -> None:
     with open(filename, "w") as f:
-        json.dump(records, f, cls=InfEncoder)
+        json.dump(
+            records,
+            f,
+            cls=InfEncoder,
+            default=lambda o: f"<{type(o).__name__} object is not JSON serializable>",
+        )

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import copy

benchmarks/cutlass_benchmarks/utils.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 # Cutlass bench utils
 from collections.abc import Iterable

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import copy
@@ -18,7 +19,7 @@ from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.utils.fp8_utils import (
     w8a8_block_fp8_matmul,
 )
-from vllm.utils import FlexibleArgumentParser
+from vllm.utils import FlexibleArgumentParser, cdiv
 
 DEFAULT_MODELS = list(WEIGHT_SHAPES.keys())
 DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
@@ -116,14 +117,9 @@ def bench_fp8(
     scale_a = torch.tensor(1.0, device="cuda", dtype=torch.float32)
     scale_b = torch.tensor(1.0, device="cuda", dtype=torch.float32)
 
-    def ceil_div(x: int, y: int) -> int:
-        return (x + y - 1) // y
-
-    block_scale_a = torch.rand(
-        (m, ceil_div(k, 128)), device="cuda", dtype=torch.float32
-    )
+    block_scale_a = torch.rand((m, cdiv(k, 128)), device="cuda", dtype=torch.float32)
     block_scale_b = torch.rand(
-        ceil_div(k, 128), ceil_div(n, 128), device="cuda", dtype=torch.float32
+        cdiv(k, 128), cdiv(n, 128), device="cuda", dtype=torch.float32
     )
     block_scale_a_M_major = block_scale_a.t().contiguous().t()
     block_scale_b_K_major = block_scale_b.t().contiguous().t()

benchmarks/cutlass_benchmarks/weight_shapes.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)

benchmarks/disagg_benchmarks/disagg_prefill_proxy_server.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import os
 

benchmarks/disagg_benchmarks/round_robin_proxy.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import asyncio
 import itertools

benchmarks/disagg_benchmarks/visualize_benchmark_results.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import json
 

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import pickle as pkl
 import time

benchmarks/kernels/bench_fp8_gemm.py

@@ -0,0 +1,158 @@
+# SPDX-License-Identifier: Apache-2.0
+import argparse
+import copy
+import itertools
+
+import torch
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm
+from vllm._custom_ops import scaled_fp8_quant as vllm_scaled_fp8_quant
+from vllm.triton_utils import triton
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "fp8-tensor-w-token-a": dict(
+        w="tensor", a="token", no_a_quant=False, enabled=False
+    ),
+    "fp8-tensor-w-tensor-a": dict(
+        w="tensor", a="tensor", no_a_quant=False, enabled=True
+    ),
+    "fp8-channel-w-token-a": dict(
+        w="channel", a="token", no_a_quant=False, enabled=True
+    ),
+    "fp8-channel-w-tensor-a": dict(
+        w="channel", a="tensor", no_a_quant=False, enabled=False
+    ),
+    "fp8-tensor-w-token-a-noquant": dict(
+        w="tensor", a="token", no_a_quant=True, enabled=False
+    ),
+    "fp8-tensor-w-tensor-a-noquant": dict(
+        w="tensor", a="tensor", no_a_quant=True, enabled=True
+    ),
+    "fp8-channel-w-token-a-noquant": dict(
+        w="channel", a="token", no_a_quant=True, enabled=True
+    ),
+    "fp8-channel-w-tensor-a-noquant": dict(
+        w="channel", a="tensor", no_a_quant=True, enabled=False
+    ),
+}
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v["enabled"]]
+
+
+def _quant_weight_fp8(b: torch.Tensor, w_type: str, device: str):
+    if w_type == "tensor":
+        scale_b = torch.ones(1, device=device, dtype=torch.float32)
+        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, scale_b)
+    else:
+        b_fp8, scale_b_fp8 = vllm_scaled_fp8_quant(b, use_per_token_if_dynamic=True)
+    return b_fp8.t(), scale_b_fp8
+
+
+def build_fp8_runner(cfg, a, b, dtype, device):
+    b_fp8, scale_b_fp8 = _quant_weight_fp8(b, cfg["w"], device)
+
+    scale_a_const = (
+        torch.ones(1, device=device, dtype=torch.float32)
+        if cfg["a"] == "tensor"
+        else None
+    )
+
+    if cfg["no_a_quant"]:
+        if cfg["a"] == "tensor":
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const)
+        else:
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True)
+
+        def run():
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+        return run
+
+    if cfg["a"] == "tensor":
+
+        def run():
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, scale_a_const)
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+    else:
+
+        def run():
+            a_fp8, scale_a_fp8 = vllm_scaled_fp8_quant(a, use_per_token_if_dynamic=True)
+            return vllm_scaled_mm(a_fp8, b_fp8, scale_a_fp8, scale_b_fp8, dtype)
+
+    return run
+
+
+@triton.testing.perf_report(
+    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=_enabled,
+        line_names=_enabled,
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs FP8 GEMMs",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+
+    a = torch.randn((M, K), device=device, dtype=dtype)
+    b = torch.randn((N, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
+        )
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_fp8_runner(cfg, a, b, dtype, device)
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: run_quant(), quantiles=quantiles
+        )
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    out = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            out.append(KN)
+    return out
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.1-8B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+    )
+    parser.add_argument("--tp-sizes", nargs="+", type=int, default=[1])
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        print(f"{model}, N={N} K={K}, BF16 vs FP8 GEMMs TFLOP/s:")
+        benchmark.run(
+            print_data=True,
+            show_plots=True,
+            save_path=f"bench_fp8_res_n{N}_k{K}",
+            N=N,
+            K=K,
+        )
+
+    print("Benchmark finished!")

benchmarks/kernels/bench_int8_gemm.py

@@ -0,0 +1,169 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import copy
+import itertools
+
+import torch
+from weight_shapes import WEIGHT_SHAPES
+
+from vllm._custom_ops import cutlass_scaled_mm as vllm_scaled_mm
+from vllm._custom_ops import scaled_int8_quant as vllm_scaled_int8_quant
+from vllm.triton_utils import triton
+
+PROVIDER_CFGS = {
+    "torch-bf16": dict(enabled=True),
+    "int8-tensor-w-token-a": dict(
+        w="tensor", a="token", no_a_quant=False, enabled=False
+    ),
+    "int8-tensor-w-tensor-a": dict(
+        w="tensor", a="tensor", no_a_quant=False, enabled=True
+    ),
+    "int8-channel-w-token-a": dict(
+        w="channel", a="token", no_a_quant=False, enabled=True
+    ),
+    "int8-channel-w-tensor-a": dict(
+        w="channel", a="tensor", no_a_quant=False, enabled=False
+    ),
+    "int8-tensor-w-token-a-noquant": dict(
+        w="tensor", a="token", no_a_quant=True, enabled=False
+    ),
+    "int8-tensor-w-tensor-a-noquant": dict(
+        w="tensor", a="tensor", no_a_quant=True, enabled=True
+    ),
+    "int8-channel-w-token-a-noquant": dict(
+        w="channel", a="token", no_a_quant=True, enabled=True
+    ),
+    "int8-channel-w-tensor-a-noquant": dict(
+        w="channel", a="tensor", no_a_quant=True, enabled=False
+    ),
+}
+
+
+def _quant_weight(b, w_type, device):
+    if w_type == "tensor":
+        scale_b = torch.ones(1, device=device, dtype=torch.float32)
+        b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b, scale_b)
+        assert scale_b_int8.numel() == 1
+    else:  # channel
+        b_int8, scale_b_int8, _ = vllm_scaled_int8_quant(b)
+        assert scale_b_int8.numel() == b.shape[0]
+    return b_int8.t(), scale_b_int8
+
+
+def build_int8_runner(cfg, a, b, dtype, device):
+    # quant before running the kernel
+    b_int8, scale_b_int8 = _quant_weight(b, cfg["w"], device)
+
+    scale_a_const = None
+    if cfg["a"] == "tensor":
+        scale_a_const = torch.ones(1, device=device, dtype=torch.float32)
+
+    # no quant, create activation ahead
+    if cfg["no_a_quant"]:
+        if cfg["a"] == "tensor":
+            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const)
+        else:  # token
+            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a)
+
+        def run_quant():
+            return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype)
+
+        return run_quant
+
+    # dynamic quant, create activation inside
+    if cfg["a"] == "tensor":
+
+        def run_quant():
+            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a, scale_a_const)
+            return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype)
+
+    else:  # token
+
+        def run_quant():
+            a_int8, scale_a_int8, _ = vllm_scaled_int8_quant(a)
+            return vllm_scaled_mm(a_int8, b_int8, scale_a_int8, scale_b_int8, dtype)
+
+    return run_quant
+
+
+_enabled = [k for k, v in PROVIDER_CFGS.items() if v.get("enabled")]
+
+
+@triton.testing.perf_report(
+    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=_enabled,
+        line_names=[k for k in _enabled],
+        ylabel="TFLOP/s (larger is better)",
+        plot_name="BF16 vs INT8 GEMMs",
+        args={},
+    )
+)
+def benchmark(batch_size, provider, N, K):
+    M = batch_size
+    device = "cuda"
+    dtype = torch.bfloat16
+    a = torch.randn((M, K), device=device, dtype=dtype)
+    b = torch.randn((N, K), device=device, dtype=dtype)
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "torch-bf16":
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: torch.nn.functional.linear(a, b), quantiles=quantiles
+        )
+    else:
+        cfg = PROVIDER_CFGS[provider]
+        run_quant = build_int8_runner(cfg, a, b, dtype, device)
+        ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(
+            lambda: run_quant(), quantiles=quantiles
+        )
+
+    to_tflops = lambda t_ms: (2 * M * N * K) * 1e-12 / (t_ms * 1e-3)
+    return to_tflops(ms), to_tflops(max_ms), to_tflops(min_ms)
+
+
+def prepare_shapes(args):
+    KN_model_names = []
+    for model, tp_size in itertools.product(args.models, args.tp_sizes):
+        for KN, tp_dim in copy.deepcopy(WEIGHT_SHAPES[model]):
+            KN[tp_dim] //= tp_size
+            KN.append(model)
+            KN_model_names.append(KN)
+    return KN_model_names
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=["meta-llama/Llama-3.1-8B-Instruct"],
+        choices=list(WEIGHT_SHAPES.keys()),
+        help="List of models to benchmark",
+    )
+    parser.add_argument(
+        "--tp-sizes",
+        nargs="+",
+        type=int,
+        default=[1],
+        help="List of tensor parallel sizes",
+    )
+    args = parser.parse_args()
+
+    for K, N, model in prepare_shapes(args):
+        print(f"{model}, N={N} K={K}, BF16 vs INT8 GEMMs TFLOP/s:")
+        benchmark.run(
+            print_data=True,
+            show_plots=True,
+            save_path=f"bench_int8_res_n{N}_k{K}",
+            N=N,
+            K=K,
+        )
+
+    print("Benchmark finished!")

benchmarks/kernels/benchmark_aqlm.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import os
 import sys

benchmarks/kernels/benchmark_bitblas.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 # Copyright (c) Microsoft Corporation.
 # Licensed under the MIT License.
 

benchmarks/kernels/benchmark_cutlass_fp4_moe.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """
 Benchmark the performance of the cutlass_moe_fp4 kernel vs the triton_moe
 kernel. The cutlass_moe_fp4 kernel takes in fp4 quantized weights and 16-bit
@@ -90,7 +91,7 @@ def bench_run(
 
     score = torch.randn((m, num_experts), device=device, dtype=dtype)
 
-    topk_weights, topk_ids = fused_topk(a, score, topk, renormalize=False)
+    topk_weights, topk_ids, _ = fused_topk(a, score, topk, renormalize=False)
 
     quant_blocksize = 16
     w1_blockscale = torch.empty(

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import torch
 import torch.utils.benchmark as benchmark
@@ -6,8 +7,8 @@ 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.cutlass_moe import cutlass_moe_fp8
 from vllm.model_executor.layers.fused_moe.fused_moe import (
-    cutlass_moe_fp8,
     fused_experts,
     fused_topk,
 )
@@ -69,18 +70,9 @@ def bench_run(
     w1_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32)
     w2_scale = torch.empty((num_experts, 1, 1), device="cuda", dtype=torch.float32)
 
-    ab_strides1 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
-    c_strides1 = torch.full((num_experts,), 2 * n, device="cuda", dtype=torch.int64)
-    ab_strides2 = torch.full((num_experts,), n, device="cuda", dtype=torch.int64)
-    c_strides2 = torch.full((num_experts,), k, device="cuda", dtype=torch.int64)
-
     for expert in range(num_experts):
         w1_q[expert], w1_scale[expert] = ops.scaled_fp8_quant(w1[expert])
         w2_q[expert], w2_scale[expert] = ops.scaled_fp8_quant(w2[expert])
-    w1_q_notransp = w1_q.clone()
-    w2_q_notransp = w2_q.clone()
-    w1_q = w1_q.transpose(1, 2)
-    w2_q = w2_q.transpose(1, 2)
 
     score = torch.randn((m, num_experts), device="cuda", dtype=dtype)
 
@@ -121,10 +113,7 @@ def bench_run(
         w2_scale: torch.Tensor,
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
-        ab_strides1: torch.Tensor,
-        c_strides1: torch.Tensor,
-        ab_strides2: torch.Tensor,
-        c_strides2: torch.Tensor,
+        per_act_token: bool,
         num_repeats: int,
     ):
         for _ in range(num_repeats):
@@ -132,15 +121,12 @@ def bench_run(
                 a,
                 w1,
                 w2,
-                w1_scale,
-                w2_scale,
                 topk_weights,
                 topk_ids,
-                ab_strides1,
-                c_strides1,
-                ab_strides2,
-                c_strides2,
-                a1_scale=a_scale,
+                w1_scale,
+                w2_scale,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_cutlass_from_graph(
@@ -152,10 +138,6 @@ def bench_run(
         w2_scale: torch.Tensor,
         topk_weights: torch.Tensor,
         topk_ids: torch.Tensor,
-        ab_strides1: torch.Tensor,
-        c_strides1: torch.Tensor,
-        ab_strides2: torch.Tensor,
-        c_strides2: torch.Tensor,
     ):
         with set_current_vllm_config(
             VllmConfig(parallel_config=ParallelConfig(pipeline_parallel_size=1))
@@ -164,15 +146,12 @@ def bench_run(
                 a,
                 w1_q,
                 w2_q,
-                w1_scale,
-                w2_scale,
                 topk_weights,
                 topk_ids,
-                ab_strides1,
-                c_strides1,
-                ab_strides2,
-                c_strides2,
-                a1_scale=a_scale,
+                w1_scale,
+                w2_scale,
+                per_act_token,
+                a1_scale=None,
             )
 
     def run_triton_from_graph(
@@ -217,10 +196,6 @@ def bench_run(
             w2_scale,
             topk_weights,
             topk_ids,
-            ab_strides1,
-            c_strides1,
-            ab_strides2,
-            c_strides2,
         )
     torch.cuda.synchronize()
 
@@ -229,8 +204,8 @@ def bench_run(
     with torch.cuda.graph(triton_graph, stream=triton_stream):
         run_triton_from_graph(
             a,
-            w1_q_notransp,
-            w2_q_notransp,
+            w1_q,
+            w2_q,
             topk_weights,
             topk_ids,
             w1_scale,
@@ -249,18 +224,13 @@ def bench_run(
         "w2": w2,
         "score": score,
         "topk": topk,
-        "w1_q_notransp": w1_q_notransp,
-        "w2_q_notransp": w2_q_notransp,
         # Cutlass params
         "a_scale": a_scale,
         "w1_q": w1_q,
         "w2_q": w2_q,
         "w1_scale": w1_scale,
         "w2_scale": w2_scale,
-        "ab_strides1": ab_strides1,
-        "c_strides1": c_strides1,
-        "ab_strides2": ab_strides2,
-        "c_strides2": c_strides2,
+        "per_act_token": per_act_token,
         # cuda graph params
         "cutlass_graph": cutlass_graph,
         "triton_graph": triton_graph,
@@ -278,8 +248,8 @@ def bench_run(
     # Warmup
     run_triton_moe(
         a,
-        w1_q_notransp,
-        w2_q_notransp,
+        w1_q,
+        w2_q,
         topk_weights,
         topk_ids,
         w1_scale,
@@ -290,7 +260,7 @@ def bench_run(
 
     results.append(
         benchmark.Timer(
-            stmt="run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)",  # noqa: E501
+            stmt="run_triton_moe(a, w1_q, w2_q, topk_weights, topk_ids, w1_scale, w2_scale, a_scale, num_runs)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
@@ -321,16 +291,13 @@ def bench_run(
         w2_scale,
         topk_weights,
         topk_ids,
-        ab_strides1,
-        c_strides1,
-        ab_strides2,
-        c_strides2,
+        per_act_token,
         num_warmup,
     )
 
     results.append(
         benchmark.Timer(
-            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, ab_strides1, c_strides1, ab_strides2, c_strides2, num_runs)",  # noqa: E501
+            stmt="run_cutlass_moe(a, a_scale, w1_q, w2_q, w1_scale, w2_scale, topk_weights, topk_ids, per_act_token, num_runs)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,

benchmarks/kernels/benchmark_layernorm.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import time
 

benchmarks/kernels/benchmark_lora.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import copy

benchmarks/kernels/benchmark_machete.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import copy
@@ -233,8 +234,10 @@ def marlin_create_bench_fn(bt: BenchmarkTensors) -> Callable:
 
         fn = lambda: ops.gptq_marlin_gemm(
             a=bt.a,
+            c=None,
             b_q_weight=w_q,
             b_scales=w_s,
+            global_scale=None,
             b_zeros=w_zp,
             g_idx=g_idx,
             perm=sort_indices,

benchmarks/kernels/benchmark_marlin.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import torch
 import torch.utils.benchmark as benchmark
@@ -21,8 +22,16 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils import (
     MARLIN_SUPPORTED_GROUP_SIZES,
     query_marlin_supported_quant_types,
 )
+from vllm.model_executor.layers.quantization.utils.marlin_utils_fp4 import (
+    FP4_MARLIN_SUPPORTED_GROUP_SIZES,
+    rand_marlin_weight_fp4_like,
+)
+from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
+    marlin_quant_fp8_torch,
+)
 from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
     MarlinWorkspace,
+    awq_marlin_quantize,
     marlin_quantize,
 )
 from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import (
@@ -34,7 +43,7 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
     quantize_weights,
     sort_weights,
 )
-from vllm.scalar_type import ScalarType
+from vllm.scalar_type import ScalarType, scalar_types
 from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
@@ -56,80 +65,144 @@ def bench_run(
     size_n: int,
 ):
     label = "Quant Matmul"
-
     sub_label = "{}, act={} k_full={}, q={}, g={}, MKN=({}x{}x{})".format(
         model, act_order, is_k_full, str(quant_type), group_size, size_m, size_k, size_n
     )
-
     print(f"Testing: {sub_label}")
 
     a = torch.randn(size_m, size_k).to(torch.half).cuda()
     b = torch.rand(size_k, size_n).to(torch.half).cuda()
+    has_zp = quant_type in [scalar_types.uint4, scalar_types.uint8]
+    if act_order and (group_size == -1 or group_size == size_k or has_zp):
+        return
+    if size_k % group_size != 0:
+        return
 
-    a_tmp = torch.zeros(size_m, size_k).to(torch.half).cuda()
-
-    # Marlin quant
-    (
-        marlin_w_ref,
-        marlin_q_w,
-        marlin_s,
-        marlin_g_idx,
-        marlin_sort_indices,
-        marlin_rand_perm,
-    ) = marlin_quantize(b, quant_type, group_size, act_order)
-
-    # Marlin_24 quant
-    (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) = (
-        marlin_24_quantize(b, quant_type, group_size)
+    marlin_24_supported = (
+        quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES
+        and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES
     )
-
-    marlin_zp = torch.empty(0, dtype=torch.int, device=b.device)
-
-    # GPTQ quant
-    (w_ref, q_w, s, g_idx, rand_perm) = gptq_quantize_weights(
-        b, quant_type, group_size, act_order
+    repack_supported = (
+        quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES
+        and group_size in MARLIN_SUPPORTED_GROUP_SIZES
     )
-    q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n)
-
-    # For act_order, sort the "weights" and "g_idx"
-    # so that group ids are increasing
-    repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device)
-    if act_order:
-        (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx)
-
-    # Prepare
-    marlin_workspace = MarlinWorkspace(
-        size_n, GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL
-    )
-
-    marlin_24_workspace = MarlinWorkspace(
-        size_n, GPTQ_MARLIN_24_MIN_THREAD_N, GPTQ_MARLIN_24_MAX_PARALLEL
-    )
-    marlin_zp = torch.zeros_like(marlin_s, dtype=torch.int)
-
-    # AllSpark W8A16 quant
-    as_supported_case = (
+    allspark_supported = (
         quant_type in ALLSPARK_SUPPORTED_QUANT_TYPES
         and group_size == -1
         and not act_order
         and is_k_full
     )
-    if as_supported_case:
-        properties = torch.cuda.get_device_properties(b.device.index)
-        sm_count = properties.multi_processor_count
-        sm_version = properties.major * 10 + properties.minor
 
-        supported_arch = sm_version >= 80 and sm_version < 90
-        as_supported_case = as_supported_case and supported_arch
-        if supported_arch:
-            has_zp = False
-            w_ref, qw, s, zp = quantize_weights(b, quant_type, group_size, has_zp)
-            qw = qw.to(torch.uint8)
-
-            qw_reorder, s_reorder, zp_reorder = ops.allspark_repack_weight(
-                qw, s, zp, has_zp
+    def gen_marlin_params():
+        # Marlin quant
+        marlin_g_idx = marlin_sort_indices = marlin_zp = marlin_s2 = None
+        if quant_type == scalar_types.float4_e2m1f:
+            if group_size != 16 or act_order:
+                return
+            marlin_w_ref, marlin_q_w, marlin_s, marlin_s2 = rand_marlin_weight_fp4_like(
+                b.T, group_size
             )
-            CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD
+        elif quant_type == scalar_types.float8_e4m3fn:
+            if group_size not in [-1, 128] or act_order:
+                return
+            marlin_w_ref, marlin_q_w, marlin_s = marlin_quant_fp8_torch(b.T, group_size)
+        elif group_size == 16:
+            return
+        elif has_zp:
+            marlin_w_ref, marlin_q_w, marlin_s, marlin_zp = awq_marlin_quantize(
+                b, quant_type, group_size
+            )
+        else:
+            marlin_w_ref, marlin_q_w, marlin_s, marlin_g_idx, marlin_sort_indices, _ = (
+                marlin_quantize(b, quant_type, group_size, act_order)
+            )
+        return (
+            marlin_w_ref,
+            marlin_q_w,
+            marlin_s,
+            marlin_s2,
+            marlin_zp,
+            marlin_g_idx,
+            marlin_sort_indices,
+        )
+
+    def gen_marlin_24_params():
+        marlin_24_w_ref = marlin_24_q_w_comp = marlin_24_meta = marlin_24_s = None
+        if marlin_24_supported:
+            (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s) = (
+                marlin_24_quantize(b, quant_type, group_size)
+            )
+        return (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s)
+
+    def gen_repack_params():
+        q_w_gptq = None
+        repack_sort_indices = None
+        if repack_supported:
+            (w_ref, q_w, s, g_idx, rand_perm) = gptq_quantize_weights(
+                b, quant_type, group_size, act_order
+            )
+            q_w_gptq = gptq_pack(q_w, quant_type.size_bits, size_k, size_n)
+
+            # For act_order, sort the "weights" and "g_idx"
+            # so that group ids are increasing
+            repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device)
+            if act_order:
+                (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx)
+        return q_w_gptq, repack_sort_indices
+
+    def gen_allspark_params():
+        qw_reorder = s_reorder = zp_reorder = sm_count = sm_version = (
+            CUBLAS_M_THRESHOLD
+        ) = None
+        nonlocal allspark_supported
+        if allspark_supported:
+            properties = torch.cuda.get_device_properties(b.device.index)
+            sm_count = properties.multi_processor_count
+            sm_version = properties.major * 10 + properties.minor
+
+            supported_arch = sm_version >= 80 and sm_version < 90
+            allspark_supported = allspark_supported and supported_arch
+            if supported_arch:
+                w_ref, qw, s, zp = quantize_weights(b, quant_type, group_size, has_zp)
+                qw = qw.to(torch.uint8)
+
+                qw_reorder, s_reorder, zp_reorder = ops.allspark_repack_weight(
+                    qw, s, zp, has_zp
+                )
+                CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD
+        return (
+            qw_reorder,
+            s_reorder,
+            zp_reorder,
+            sm_count,
+            sm_version,
+            CUBLAS_M_THRESHOLD,
+        )
+
+    (
+        marlin_w_ref,
+        marlin_q_w,
+        marlin_s,
+        marlin_s2,
+        marlin_zp,
+        marlin_g_idx,
+        marlin_sort_indices,
+    ) = gen_marlin_params()
+    marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s = (
+        gen_marlin_24_params()
+    )
+    q_w_gptq, repack_sort_indices = gen_repack_params()
+    qw_reorder, s_reorder, zp_reorder, sm_count, sm_version, CUBLAS_M_THRESHOLD = (
+        gen_allspark_params()
+    )
+
+    # Prepare
+    marlin_workspace = MarlinWorkspace(
+        size_n, GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_MAX_PARALLEL
+    )
+    marlin_24_workspace = MarlinWorkspace(
+        size_n, GPTQ_MARLIN_24_MIN_THREAD_N, GPTQ_MARLIN_24_MAX_PARALLEL
+    )
 
     globals = {
         # Gen params
@@ -139,15 +212,14 @@ def bench_run(
         "size_n": size_n,
         "size_k": size_k,
         "a": a,
-        "a_tmp": a_tmp,
         # Marlin params
         "marlin_w_ref": marlin_w_ref,
         "marlin_q_w": marlin_q_w,
         "marlin_s": marlin_s,
+        "marlin_s2": marlin_s2,
         "marlin_zp": marlin_zp,
         "marlin_g_idx": marlin_g_idx,
         "marlin_sort_indices": marlin_sort_indices,
-        "marlin_rand_perm": marlin_rand_perm,
         "marlin_workspace": marlin_workspace,
         "is_k_full": is_k_full,
         # Marlin_24 params
@@ -160,12 +232,12 @@ def bench_run(
         "q_w_gptq": q_w_gptq,
         "repack_sort_indices": repack_sort_indices,
         # AllSpark W8A16 params
-        "qw_reorder": qw_reorder if as_supported_case else None,
-        "s_reorder": s_reorder if as_supported_case else None,
-        "zp_reorder": zp_reorder if as_supported_case else None,
-        "sm_count": sm_count if as_supported_case else None,
-        "sm_version": sm_version if as_supported_case else None,
-        "CUBLAS_M_THRESHOLD": CUBLAS_M_THRESHOLD if as_supported_case else None,
+        "qw_reorder": qw_reorder,
+        "s_reorder": s_reorder,
+        "zp_reorder": zp_reorder,
+        "sm_count": sm_count,
+        "sm_version": sm_version,
+        "CUBLAS_M_THRESHOLD": CUBLAS_M_THRESHOLD,
         # Kernels
         "gptq_marlin_gemm": ops.gptq_marlin_gemm,
         "gptq_marlin_24_gemm": ops.gptq_marlin_24_gemm,
@@ -176,7 +248,7 @@ def bench_run(
     min_run_time = 1
 
     # Warmup pytorch
-    for i in range(5):
+    for _ in range(5):
         torch.matmul(a, marlin_w_ref)
 
     results.append(
@@ -191,17 +263,17 @@ def bench_run(
 
     results.append(
         benchmark.Timer(
-            stmt="output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False, False)",  # noqa: E501
+            stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, False, False)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
-            description="gptq_marlin_gemm_fp16",
+            description="gptq_marlin_gemm",
         ).blocked_autorange(min_run_time=min_run_time)
     )
 
     results.append(
         benchmark.Timer(
-            stmt="output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True, False)",  # noqa: E501
+            stmt="output = gptq_marlin_gemm(a, None, marlin_q_w, marlin_s, marlin_s2, marlin_zp, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, quant_type, size_m, size_n, size_k, is_k_full, False, True, False)",  # noqa: E501
             globals=globals,
             label=label,
             sub_label=sub_label,
@@ -209,10 +281,7 @@ def bench_run(
         ).blocked_autorange(min_run_time=min_run_time)
     )
 
-    if (
-        quant_type in GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES
-        and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES
-    ):
+    if marlin_24_supported:
         results.append(
             benchmark.Timer(
                 stmt="output = gptq_marlin_24_gemm(a, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s, marlin_24_workspace.scratch, quant_type, size_m, size_n, size_k)",  # noqa: E501
@@ -223,17 +292,18 @@ def bench_run(
             ).blocked_autorange(min_run_time=min_run_time)
         )
 
-    results.append(
-        benchmark.Timer(
-            stmt="q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, quant_type.size_bits)",  # noqa: E501
-            globals=globals,
-            label=label,
-            sub_label=sub_label,
-            description="gptq_marlin_repack",
-        ).blocked_autorange(min_run_time=min_run_time)
-    )
+    if repack_supported:
+        results.append(
+            benchmark.Timer(
+                stmt="q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, quant_type.size_bits)",  # noqa: E501
+                globals=globals,
+                label=label,
+                sub_label=sub_label,
+                description="gptq_marlin_repack",
+            ).blocked_autorange(min_run_time=min_run_time)
+        )
 
-    if as_supported_case:
+    if allspark_supported:
         results.append(
             benchmark.Timer(
                 stmt="output = allspark_w8a16_gemm(a, qw_reorder, s_reorder, zp_reorder, size_n, group_size, sm_count, sm_version, CUBLAS_M_THRESHOLD, False, True)",  # noqa: E501
@@ -249,7 +319,6 @@ def main(args):
     print("Benchmarking models:")
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")
-
     results: list[benchmark.Measurement] = []
 
     for model in args.models:
@@ -277,14 +346,17 @@ def main(args):
                     ):
                         continue
 
-                    for quant_type in query_marlin_supported_quant_types(False):
+                    for quant_type in query_marlin_supported_quant_types():
                         if (
                             len(args.limit_num_bits) > 0
                             and quant_type.size_bits not in args.limit_num_bits
                         ):
                             continue
 
-                        for group_size in MARLIN_SUPPORTED_GROUP_SIZES:
+                        for group_size in (
+                            MARLIN_SUPPORTED_GROUP_SIZES
+                            + FP4_MARLIN_SUPPORTED_GROUP_SIZES
+                        ):
                             if (
                                 len(args.limit_group_size) > 0
                                 and group_size not in args.limit_group_size

benchmarks/kernels/benchmark_moe.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 import json
@@ -6,7 +7,6 @@ import time
 from contextlib import nullcontext
 from datetime import datetime
 from itertools import product
-from types import SimpleNamespace
 from typing import Any, TypedDict
 
 import ray
@@ -42,7 +42,7 @@ def benchmark_config(
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
     num_iters: int = 100,
-    block_quant_shape: List[int] = None,
+    block_quant_shape: list[int] = None,
     use_deep_gemm: bool = False,
 ) -> float:
     init_dtype = torch.float16 if use_fp8_w8a8 else dtype
@@ -399,7 +399,7 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
-        block_quant_shape: List[int] = None,
+        block_quant_shape: list[int] = None,
         use_deep_gemm: bool = False,
     ) -> tuple[dict[str, int], float]:
         current_platform.seed_everything(self.seed)
@@ -531,7 +531,7 @@ def save_configs(
     dtype: torch.dtype,
     use_fp8_w8a8: bool,
     use_int8_w8a16: bool,
-    block_quant_shape: List[int],
+    block_quant_shape: list[int],
 ) -> None:
     dtype_str = get_config_dtype_str(
         dtype, use_int8_w8a16=use_int8_w8a16, use_fp8_w8a8=use_fp8_w8a8
@@ -562,7 +562,6 @@ def main(args: argparse.Namespace):
     config = get_config(model=args.model, trust_remote_code=args.trust_remote_code)
     if args.model_prefix:
         config = getattr(config, args.model_prefix)
-    config = SimpleNamespace(**config)
 
     if config.architectures[0] == "DbrxForCausalLM":
         E = config.ffn_config.moe_num_experts
@@ -594,11 +593,7 @@ def main(args: argparse.Namespace):
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
 
     hidden_size = config.hidden_size
-    dtype = (
-        torch.float16
-        if current_platform.is_rocm()
-        else getattr(torch, config.torch_dtype)
-    )
+    dtype = torch.float16 if current_platform.is_rocm() else config.torch_dtype
     use_fp8_w8a8 = args.dtype == "fp8_w8a8"
     use_int8_w8a16 = args.dtype == "int8_w8a16"
     block_quant_shape = get_weight_block_size_safety(config)

benchmarks/kernels/benchmark_moe_align_block_size.py

@@ -0,0 +1,159 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+import argparse
+import itertools
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.model_executor.layers.fused_moe.moe_align_block_size import (
+    moe_align_block_size_triton,
+)
+from vllm.triton_utils import triton
+
+
+def get_topk_ids(num_tokens: int, num_experts: int, topk: int) -> torch.Tensor:
+    return torch.stack(
+        [
+            torch.randperm(num_experts, dtype=torch.int32, device="cuda")[:topk]
+            for _ in range(num_tokens)
+        ]
+    )
+
+
+def check_correctness(num_tokens, num_experts=256, block_size=256, topk=8):
+    """
+    Verifies vllm vs. Triton
+    """
+    topk_ids = get_topk_ids(num_tokens, num_experts, topk)
+
+    # 1. malloc space for triton and vllm
+    # malloc enough space (max_num_tokens_padded) for the sorted ids
+    max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
+    sorted_ids_triton = torch.empty(
+        (max_num_tokens_padded,), dtype=torch.int32, device="cuda"
+    )
+    sorted_ids_triton.fill_(topk_ids.numel())  # fill with sentinel value
+    expert_ids_triton = torch.zeros(
+        (max_num_tokens_padded // block_size,), dtype=torch.int32, device="cuda"
+    )
+    num_tokens_post_pad_triton = torch.empty((1,), dtype=torch.int32, device="cuda")
+
+    sorted_ids_vllm = torch.empty_like(sorted_ids_triton)
+    sorted_ids_vllm.fill_(topk_ids.numel())
+    expert_ids_vllm = torch.zeros_like(expert_ids_triton)
+    num_tokens_post_pad_vllm = torch.empty_like(num_tokens_post_pad_triton)
+
+    # 2. run implementations
+    moe_align_block_size_triton(
+        topk_ids,
+        num_experts,
+        block_size,
+        sorted_ids_triton,
+        expert_ids_triton,
+        num_tokens_post_pad_triton,
+    )
+
+    ops.moe_align_block_size(
+        topk_ids,
+        num_experts,
+        block_size,
+        sorted_ids_vllm,
+        expert_ids_vllm,
+        num_tokens_post_pad_vllm,
+    )
+    print(f"✅ VLLM implementation works with {num_experts} experts!")
+
+    # 3. compare results
+    if torch.allclose(expert_ids_triton, expert_ids_vllm) and torch.allclose(
+        num_tokens_post_pad_triton, num_tokens_post_pad_vllm
+    ):
+        print("✅ Triton and VLLM implementations match.")
+    else:
+        print("❌ Triton and VLLM implementations DO NOT match.")
+        print("Triton expert_ids:", expert_ids_triton)
+        print("VLLM expert_ids:", expert_ids_vllm)
+        print("Triton num_tokens_post_pad:", num_tokens_post_pad_triton)
+        print("VLLM num_tokens_post_pad:", num_tokens_post_pad_vllm)
+
+
+# test configurations
+num_tokens_range = [1, 16, 256, 4096]
+num_experts_range = [16, 64, 224, 256, 280, 512]
+topk_range = [1, 2, 8]
+configs = list(itertools.product(num_tokens_range, num_experts_range, topk_range))
+
+
+@triton.testing.perf_report(
+    triton.testing.Benchmark(
+        x_names=["num_tokens", "num_experts", "topk"],
+        x_vals=configs,
+        line_arg="provider",
+        line_vals=["vllm", "triton"],  # "triton"
+        line_names=["VLLM", "Triton"],  # "Triton"
+        plot_name="moe-align-block-size-performance",
+        args={},
+    )
+)
+def benchmark(num_tokens, num_experts, topk, provider):
+    """Benchmark function for Triton."""
+    block_size = 256
+    topk_ids = get_topk_ids(num_tokens, num_experts, topk)
+
+    max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
+    sorted_ids = torch.empty((max_num_tokens_padded,), dtype=torch.int32, device="cuda")
+    sorted_ids.fill_(topk_ids.numel())
+    max_num_m_blocks = max_num_tokens_padded // block_size
+    expert_ids = torch.empty((max_num_m_blocks,), dtype=torch.int32, device="cuda")
+    num_tokens_post_pad = torch.empty((1,), dtype=torch.int32, device="cuda")
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    if provider == "vllm":
+        ms, min_ms, max_ms = triton.testing.do_bench(
+            lambda: ops.moe_align_block_size(
+                topk_ids,
+                num_experts,
+                block_size,
+                sorted_ids.clone(),
+                expert_ids.clone(),
+                num_tokens_post_pad.clone(),
+            ),
+            quantiles=quantiles,
+        )
+    elif provider == "triton":
+        ms, min_ms, max_ms = triton.testing.do_bench(
+            lambda: moe_align_block_size_triton(
+                topk_ids,
+                num_experts,
+                block_size,
+                sorted_ids.clone(),
+                expert_ids.clone(),
+                num_tokens_post_pad.clone(),
+            ),
+            quantiles=quantiles,
+        )
+
+    return 1000 * ms, 1000 * max_ms, 1000 * min_ms
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num_experts",
+        type=int,
+        default=64,
+        choices=[8, 16, 32, 64, 128, 256],
+    )
+    parser.add_argument(
+        "--topk",
+        type=int,
+        default=8,
+        choices=[2, 4, 8],
+        help="Top-k value for correctness check.",
+    )
+    args = parser.parse_args()
+
+    print("Running correctness check...")
+    check_correctness(num_tokens=1024, num_experts=args.num_experts, topk=args.topk)
+    benchmark.run(print_data=True, show_plots=True)

benchmarks/kernels/benchmark_moe_permute_unpermute.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import argparse
 from typing import Any, TypedDict

benchmarks/kernels/benchmark_paged_attention.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import random
 import time

benchmarks/kernels/benchmark_quant.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import time
 

benchmarks/kernels/benchmark_rmsnorm.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import itertools
 from typing import Optional, Union

benchmarks/kernels/benchmark_rope.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from itertools import accumulate
 from typing import Optional
@@ -22,7 +23,7 @@ def benchmark_rope_kernels_multi_lora(
     seed: int,
     device: str,
     max_position: int = 8192,
-    base: int = 10000,
+    base: float = 10000,
 ) -> None:
     current_platform.seed_everything(seed)
     torch.set_default_device(device)

benchmarks/kernels/benchmark_shapes.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 WEIGHT_SHAPES = {
     "ideal": [[4 * 256 * 32, 256 * 32]],

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 # Adapted from sglang quantization/tuning_block_wise_kernel.py
 
 import argparse

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 # fmt: off
 # ruff: noqa: E501
 import time
@@ -84,12 +85,6 @@ def benchmark_shape(m: int,
 
     # === DeepGEMM Implementation ===
     def deepgemm_gemm():
-        # A quantization is inside the loop as it depends on activations
-        # A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A)
-        # A_deepgemm, A_scale_deepgemm = per_token_group_quant_fp8(
-        #     A, block_size[1])
-        # A_scale_aligned = get_col_major_tma_aligned_tensor(A_scale_deepgemm)
-        # C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16)
         deep_gemm.gemm_fp8_fp8_bf16_nt((A_deepgemm, A_scale_deepgemm),
                                        (B_deepgemm, B_scale_deepgemm),
                                        C_deepgemm)
@@ -97,8 +92,6 @@ def benchmark_shape(m: int,
 
     # === vLLM Triton Implementation ===
     def vllm_triton_gemm():
-        # A quantization is inside the loop as it depends on activations
-        # A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1])
         return w8a8_block_fp8_matmul(A_vllm,
                                      B_vllm,
                                      A_scale_vllm,
@@ -108,9 +101,6 @@ def benchmark_shape(m: int,
 
     # === vLLM CUTLASS Implementation ===
     def vllm_cutlass_gemm():
-        # A quantization is inside the loop as it depends on activations
-        # A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8(
-        #     A, block_size[1], column_major_scales=True)
         return ops.cutlass_scaled_mm(A_vllm_cutlass,
                                      B_vllm.T,
                                      scale_a=A_scale_vllm_cutlass,

benchmarks/kernels/graph_machete_bench.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import math
 import pickle

benchmarks/kernels/utils.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import dataclasses
 from collections.abc import Iterable

benchmarks/kernels/weight_shapes.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 # Weight Shapes are in the format
 # ([K, N], TP_SPLIT_DIM)
@@ -48,4 +49,50 @@ WEIGHT_SHAPES = {
         ([16384, 106496], 1),
         ([53248, 16384], 0),
     ],
+    "meta-llama/Llama-3.1-8B-Instruct": [
+        ([4096, 6144], 1),
+        ([4096, 4096], 0),
+        ([4096, 28672], 1),
+        ([14336, 4096], 0),
+    ],
+    "meta-llama/Llama-3.3-70B-Instruct": [
+        ([8192, 10240], 1),
+        ([8192, 8192], 0),
+        ([8192, 57344], 1),
+        ([28672, 8192], 0),
+    ],
+    "mistralai/Mistral-Large-Instruct-2407": [
+        ([12288, 14336], 1),
+        ([12288, 12288], 0),
+        ([12288, 57344], 1),
+        ([28672, 12288], 0),
+    ],
+    "Qwen/Qwen2.5-7B-Instruct": [
+        ([3584, 4608], 1),
+        ([3584, 3584], 0),
+        ([3584, 37888], 1),
+        ([18944, 3584], 0),
+    ],
+    "Qwen/Qwen2.5-32B-Instruct": [
+        ([5120, 7168], 1),
+        ([5120, 5120], 0),
+        ([5120, 55296], 1),
+        ([27648, 5120], 0),
+    ],
+    "Qwen/Qwen2.5-72B-Instruct": [
+        ([8192, 10240], 1),
+        ([8192, 8192], 0),
+        ([8192, 59136], 1),
+        ([29568, 8192], 0),
+    ],
+    "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": [
+        ([2048, 3072], 1),
+        ([2048, 4096], 1),
+        ([2048, 2048], 0),
+        ([2048, 576], 0),
+        ([2048, 21888], 1),
+        ([10944, 2048], 0),
+        ([2048, 2816], 1),
+        ([1408, 2048], 0),
+    ],
 }

benchmarks/overheads/benchmark_hashing.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 import cProfile
 import pstats

cmake/cpu_extension.cmake

@@ -75,6 +75,7 @@ if (MACOSX_FOUND AND CMAKE_SYSTEM_PROCESSOR STREQUAL "arm64")
 else()
     find_isa(${CPUINFO} "avx2" AVX2_FOUND)
     find_isa(${CPUINFO} "avx512f" AVX512_FOUND)
+    find_isa(${CPUINFO} "Power11" POWER11_FOUND)
     find_isa(${CPUINFO} "POWER10" POWER10_FOUND)
     find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
     find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
@@ -95,24 +96,39 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND
             CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3)
             list(APPEND CXX_COMPILE_FLAGS "-mavx512bf16")
+            set(ENABLE_AVX512BF16 ON)
         else()
+            set(ENABLE_AVX512BF16 OFF)
             message(WARNING "Disable AVX512-BF16 ISA support, requires gcc/g++ >= 12.3")
         endif()
     else()
+        set(ENABLE_AVX512BF16 OFF)
         message(WARNING "Disable AVX512-BF16 ISA support, no avx512_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512BF16=1.")
     endif()
+
+    find_isa(${CPUINFO} "avx512_vnni" AVX512VNNI_FOUND)
+    if (AVX512VNNI_FOUND)
+        list(APPEND CXX_COMPILE_FLAGS "-mavx512vnni")
+        set(ENABLE_AVX512VNNI ON) 
+    endif() 
     
 elseif (AVX2_FOUND)
     list(APPEND CXX_COMPILE_FLAGS "-mavx2")
     message(WARNING "vLLM CPU backend using AVX2 ISA")
     
-elseif (POWER9_FOUND OR POWER10_FOUND)
+elseif (POWER9_FOUND OR POWER10_FOUND OR POWER11_FOUND)
     message(STATUS "PowerPC detected")
-    # Check for PowerPC VSX support
-    list(APPEND CXX_COMPILE_FLAGS
-        "-mvsx"
-        "-mcpu=native"
-        "-mtune=native")
+    if (POWER9_FOUND)
+        list(APPEND CXX_COMPILE_FLAGS
+            "-mvsx"
+            "-mcpu=power9"
+            "-mtune=power9")
+    elseif (POWER10_FOUND OR POWER11_FOUND)
+        list(APPEND CXX_COMPILE_FLAGS
+            "-mvsx"
+            "-mcpu=power10"
+            "-mtune=power10")
+    endif()
 
 elseif (ASIMD_FOUND)
     message(STATUS "ARMv8 or later architecture detected")
@@ -224,6 +240,17 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
         "csrc/cpu/quant.cpp"
         "csrc/cpu/shm.cpp"
         ${VLLM_EXT_SRC})
+    if (ENABLE_AVX512BF16 AND ENABLE_AVX512VNNI)
+        set(VLLM_EXT_SRC
+            "csrc/cpu/sgl-kernels/gemm.cpp"
+            "csrc/cpu/sgl-kernels/gemm_int8.cpp"
+            "csrc/cpu/sgl-kernels/gemm_fp8.cpp"
+            "csrc/cpu/sgl-kernels/moe.cpp"
+            "csrc/cpu/sgl-kernels/moe_int8.cpp"
+            "csrc/cpu/sgl-kernels/moe_fp8.cpp"
+            ${VLLM_EXT_SRC})
+        add_compile_definitions(-DCPU_CAPABILITY_AVX512)
+    endif()
 elseif(POWER10_FOUND)
     set(VLLM_EXT_SRC
         "csrc/cpu/quant.cpp"

cmake/external_projects/vllm_flash_attn.cmake

@@ -38,7 +38,7 @@ else()
   FetchContent_Declare(
           vllm-flash-attn
           GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 8798f27777fb57f447070301bf33a9f9c607f491
+          GIT_TAG 1c2624e53c078854e0637ee566c72fe2107e75f4
           GIT_PROGRESS TRUE
           # Don't share the vllm-flash-attn build between build types
           BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
@@ -46,22 +46,38 @@ else()
 endif()
 
 
+# Ensure the vllm/vllm_flash_attn directory exists before installation
+install(CODE "file(MAKE_DIRECTORY \"\${CMAKE_INSTALL_PREFIX}/vllm/vllm_flash_attn\")" ALL_COMPONENTS)
+
+# Make sure vllm-flash-attn install rules are nested under vllm/
+# This is here to support installing all components under the same prefix with cmake --install.
+# setup.py installs every component separately but uses the same prefix for all.
+# ALL_COMPONENTS is used to avoid duplication for FA2 and FA3,
+# and these statements don't hurt when installing neither component.
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY FALSE)" ALL_COMPONENTS)
+install(CODE "set(OLD_CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
+install(CODE "set(CMAKE_INSTALL_PREFIX \"\${CMAKE_INSTALL_PREFIX}/vllm/\")" ALL_COMPONENTS)
+
 # Fetch the vllm-flash-attn library
 FetchContent_MakeAvailable(vllm-flash-attn)
 message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
 
+# Restore the install prefix
+install(CODE "set(CMAKE_INSTALL_PREFIX \"\${OLD_CMAKE_INSTALL_PREFIX}\")" ALL_COMPONENTS)
+install(CODE "set(CMAKE_INSTALL_LOCAL_ONLY TRUE)" ALL_COMPONENTS)
+
 # Copy over the vllm-flash-attn python files (duplicated for fa2 and fa3, in
 # case only one is built, in the case both are built redundant work is done)
 install(
   DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
+  DESTINATION vllm/vllm_flash_attn
   COMPONENT _vllm_fa2_C
   FILES_MATCHING PATTERN "*.py"
 )
 
 install(
   DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
+  DESTINATION vllm/vllm_flash_attn
   COMPONENT _vllm_fa3_C
   FILES_MATCHING PATTERN "*.py"
 )

cmake/hipify.py

@@ -1,5 +1,6 @@
 #!/usr/bin/env python3
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 #
 # A command line tool for running pytorch's hipify preprocessor on CUDA

cmake/utils.cmake

@@ -76,7 +76,7 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
   set(CSRC_BUILD_DIR ${CMAKE_CURRENT_BINARY_DIR}/csrc)
   add_custom_target(
     hipify${NAME}
-    COMMAND ${CMAKE_SOURCE_DIR}/cmake/hipify.py -p ${CMAKE_SOURCE_DIR}/csrc -o ${CSRC_BUILD_DIR} ${SRCS}
+    COMMAND ${Python_EXECUTABLE} ${CMAKE_SOURCE_DIR}/cmake/hipify.py -p ${CMAKE_SOURCE_DIR}/csrc -o ${CSRC_BUILD_DIR} ${SRCS}
     DEPENDS ${CMAKE_SOURCE_DIR}/cmake/hipify.py ${SRCS}
     BYPRODUCTS ${HIP_SRCS}
     COMMENT "Running hipify on ${NAME} extension source files.")
@@ -122,6 +122,7 @@ function (get_torch_gpu_compiler_flags OUT_GPU_FLAGS GPU_LANG)
       "-DENABLE_FP8"
       "-U__HIP_NO_HALF_CONVERSIONS__"
       "-U__HIP_NO_HALF_OPERATORS__"
+      "-Werror=unused-variable"
       "-fno-gpu-rdc")
 
   endif()
@@ -264,8 +265,8 @@ macro(set_gencode_flags_for_srcs)
 endmacro()
 
 #
-# For the given `SRC_CUDA_ARCHS` list of gencode versions in the form 
-#  `<major>.<minor>[letter]` compute the "loose intersection" with the 
+# For the given `SRC_CUDA_ARCHS` list of gencode versions in the form
+#  `<major>.<minor>[letter]` compute the "loose intersection" with the
 #  `TGT_CUDA_ARCHS` list of gencodes. We also support the `+PTX` suffix in
 #  `SRC_CUDA_ARCHS` which indicates that the PTX code should be built when there
 #  is a CUDA_ARCH in `TGT_CUDA_ARCHS` that is equal to or larger than the
@@ -277,7 +278,7 @@ endmacro()
 #  in `SRC_CUDA_ARCHS` that is less or equal to the version in `TGT_CUDA_ARCHS`.
 # We have special handling for x.0a, if x.0a is in `SRC_CUDA_ARCHS` and x.0 is
 #  in `TGT_CUDA_ARCHS` then we should remove x.0a from `SRC_CUDA_ARCHS` and add
-#  x.0a to the result (and remove x.0 from TGT_CUDA_ARCHS). 
+#  x.0a to the result (and remove x.0 from TGT_CUDA_ARCHS).
 # The result is stored in `OUT_CUDA_ARCHS`.
 #
 # Example:
@@ -312,21 +313,16 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   # if x.0a is in SRC_CUDA_ARCHS and x.0 is in CUDA_ARCHS then we should
   # remove x.0a from SRC_CUDA_ARCHS and add x.0a to _CUDA_ARCHS
   set(_CUDA_ARCHS)
-  if ("9.0a" IN_LIST _SRC_CUDA_ARCHS)
-    list(REMOVE_ITEM _SRC_CUDA_ARCHS "9.0a")
-    if ("9.0" IN_LIST TGT_CUDA_ARCHS)
-      list(REMOVE_ITEM _TGT_CUDA_ARCHS "9.0")
-      set(_CUDA_ARCHS "9.0a")
+  foreach(_arch ${_SRC_CUDA_ARCHS})
+    if(_arch MATCHES "\\a$")
+      list(REMOVE_ITEM _SRC_CUDA_ARCHS "${_arch}")
+      string(REPLACE "a" "" _base "${_arch}")
+      if ("${_base}" IN_LIST TGT_CUDA_ARCHS)
+        list(REMOVE_ITEM _TGT_CUDA_ARCHS "${_base}")
+        list(APPEND _CUDA_ARCHS "${_arch}")
+      endif()
     endif()
-  endif()
-
-  if ("10.0a" IN_LIST _SRC_CUDA_ARCHS)
-    list(REMOVE_ITEM _SRC_CUDA_ARCHS "10.0a")
-    if ("10.0" IN_LIST TGT_CUDA_ARCHS)
-      list(REMOVE_ITEM _TGT_CUDA_ARCHS "10.0")
-      set(_CUDA_ARCHS "10.0a")
-    endif()
-  endif()
+  endforeach()
 
   list(SORT _SRC_CUDA_ARCHS COMPARE NATURAL ORDER ASCENDING)
 
@@ -358,7 +354,7 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
   endforeach()
 
   list(REMOVE_DUPLICATES _CUDA_ARCHS)
-  
+
   # reapply +PTX suffix to architectures that requested PTX
   set(_FINAL_ARCHS)
   foreach(_arch ${_CUDA_ARCHS})
@@ -369,7 +365,7 @@ function(cuda_archs_loose_intersection OUT_CUDA_ARCHS SRC_CUDA_ARCHS TGT_CUDA_AR
     endif()
   endforeach()
   set(_CUDA_ARCHS ${_FINAL_ARCHS})
-  
+
   set(${OUT_CUDA_ARCHS} ${_CUDA_ARCHS} PARENT_SCOPE)
 endfunction()
 

csrc/attention/mla/cutlass_mla_kernels.cu

@@ -119,7 +119,7 @@ typename T::Fmha::Arguments args_from_options(
       {static_cast<ElementOut*>(out.data_ptr()), stride_O,
        static_cast<ElementAcc*>(nullptr), stride_LSE},
       hw_info,
-      -1,       // split_kv
+      1,        // split_kv
       nullptr,  // is_var_split_kv
   };
   // TODO(kaixih@nvidia): When split_kv=-1 and is_var_split_kv=false, we compute
@@ -207,7 +207,7 @@ void cutlass_mla_decode_sm100a(torch::Tensor const& out,
               "page_table must be a 32-bit integer tensor");
 
   auto in_dtype = q_nope.dtype();
-  at::cuda::CUDAGuard device_guard{(char)q_nope.get_device()};
+  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) {

csrc/attention/paged_attention_v1.cu

@@ -65,9 +65,6 @@ void paged_attention_v1_launcher(
   int kv_block_stride = key_cache.stride(0);
   int kv_head_stride = key_cache.stride(1);
 
-  [[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
-  assert(head_size % thread_group_size == 0);
-
   // NOTE: alibi_slopes is optional.
   const float* alibi_slopes_ptr =
       alibi_slopes
@@ -193,4 +190,4 @@ void paged_attention_v1(
 #undef WARP_SIZE
 #undef MAX
 #undef MIN
-#undef DIVIDE_ROUND_UP
+#undef DIVIDE_ROUND_UP

csrc/attention/paged_attention_v2.cu

@@ -66,9 +66,6 @@ void paged_attention_v2_launcher(
   int kv_block_stride = key_cache.stride(0);
   int kv_head_stride = key_cache.stride(1);
 
-  [[maybe_unused]] int thread_group_size = MAX(WARP_SIZE / BLOCK_SIZE, 1);
-  assert(head_size % thread_group_size == 0);
-
   // NOTE: alibi_slopes is optional.
   const float* alibi_slopes_ptr =
       alibi_slopes
@@ -203,4 +200,4 @@ void paged_attention_v2(
 #undef WARP_SIZE
 #undef MAX
 #undef MIN
-#undef DIVIDE_ROUND_UP
+#undef DIVIDE_ROUND_UP

csrc/cpu/attention.cpp

@@ -137,8 +137,8 @@ FORCE_INLINE std::pair<T, T> reduceSoftmaxAlibi(T* data, const int size,
 }
 
 template <typename T>
-FORCE_INLINE void reducePartitonSoftmax(const T* max_data, T* sum_data,
-                                        const int size) {
+FORCE_INLINE void reducePartitionSoftmax(const T* max_data, T* sum_data,
+                                         const int size) {
   T max = max_data[0];
   for (int i = 1; i < size; ++i) {
     max = max >= max_data[i] ? max : max_data[i];
@@ -634,7 +634,7 @@ struct paged_attention_v2_impl {
 
         if (partition_num == 1) continue;
 
-        reducePartitonSoftmax(
+        reducePartitionSoftmax(
             max_logits + seq_idx * num_heads * max_num_partitions +
                 head_idx * max_num_partitions,
             exp_sums + seq_idx * num_heads * max_num_partitions +

csrc/cpu/cpu_types_x86.hpp

@@ -83,7 +83,7 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
   explicit FP16Vec16(const void* ptr)
       : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
-  // non-temproal load
+  // non-temporal load
   explicit FP16Vec16(bool, void* ptr)
       : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
 
@@ -120,7 +120,7 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
   explicit BF16Vec16(const void* ptr)
       : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
-  // non-temproal load
+  // non-temporal load
   explicit BF16Vec16(bool, void* ptr)
       : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
 
@@ -327,7 +327,7 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
   // normal load
   explicit FP32Vec16(const float* ptr) : reg(_mm512_loadu_ps(ptr)) {}
 
-  // non-temproal load
+  // non-temporal load
   explicit FP32Vec16(bool, void* ptr)
       : reg((__m512)_mm512_stream_load_si512(ptr)) {}
 
@@ -576,7 +576,7 @@ struct INT8Vec64 : public Vec<INT8Vec64> {
   // normal load
   explicit INT8Vec64(void* ptr) : reg(_mm512_loadu_epi8(ptr)) {}
 
-  // non-temproal load
+  // non-temporal load
   explicit INT8Vec64(bool, void* ptr) : reg(_mm512_stream_load_si512(ptr)) {}
 
   void save(void* ptr) const { _mm512_storeu_epi8(ptr, reg); }
@@ -587,7 +587,7 @@ struct INT8Vec64 : public Vec<INT8Vec64> {
     _mm512_mask_storeu_epi8(ptr, mask, reg);
   }
 
-  // non-temproal save
+  // non-temporal save
   void nt_save(int8_t* ptr) { _mm512_stream_si512((__m512i*)ptr, reg); }
 };
 #endif

csrc/cpu/sgl-kernels/common.h

@@ -0,0 +1,238 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#pragma once
+
+#include <ATen/ATen.h>
+#include <ATen/Parallel.h>
+#include <ATen/record_function.h>
+
+// clang-format off
+
+#if defined(_OPENMP)
+#include <omp.h>
+#endif
+
+namespace {
+
+// dispatch bool
+#define AT_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...)                                 \
+  [&] {                                                                          \
+    if (BOOL_V) {                                                                \
+      constexpr bool BOOL_NAME = true;                                           \
+      return __VA_ARGS__();                                                      \
+    } else {                                                                     \
+      constexpr bool BOOL_NAME = false;                                          \
+      return __VA_ARGS__();                                                      \
+    }                                                                            \
+  }()
+
+// dispatch: bfloat16, float16, int8_t, fp8_e4m3
+#define CPU_DISPATCH_PACKED_TYPES(TYPE, ...)                                    \
+  [&] {                                                                         \
+    switch (TYPE) {                                                             \
+      case at::ScalarType::BFloat16 : {                                         \
+        using packed_t = at::BFloat16;                                          \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      case at::ScalarType::Half: {                                              \
+        using packed_t = at::Half;                                              \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      case at::ScalarType::Char : {                                             \
+        using packed_t = int8_t;                                                \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      case at::ScalarType::Float8_e4m3fn : {                                    \
+        using packed_t = at::Float8_e4m3fn;                                     \
+        return __VA_ARGS__();                                                   \
+      }                                                                         \
+      default:                                                                  \
+        TORCH_CHECK(false, "Unsupported floating data type.\n");                \
+    }                                                                           \
+  }()
+
+#define UNUSED(x) (void)(x)
+
+#define CHECK_CPU(x) TORCH_CHECK(x.device().type() == at::kCPU, #x " must be a CPU tensor")
+
+#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_LAST_DIM_CONTIGUOUS(x) \
+  TORCH_CHECK(x.strides()[x.strides().size() - 1] == 1, #x "must be contiguous at last dimention")
+
+#define CHECK_INPUT(x) \
+  CHECK_CPU(x);        \
+  CHECK_CONTIGUOUS(x)
+#define CHECK_LAST_DIM_CONTIGUOUS_INPUT(x) \
+  CHECK_CPU(x);                            \
+  CHECK_LAST_DIM_CONTIGUOUS(x)
+
+#define CHECK_DIM(d, x) TORCH_CHECK(x.dim() == d, #x " must be a " #d "D tensor")
+
+#define CHECK_EQ(a, b) TORCH_CHECK((a) == (b), "CHECK_EQ(" #a ", " #b ") failed. ", a, " vs ", b)
+
+// parallel routines
+constexpr int GRAIN_SIZE = 1024;
+
+template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
+inline T div_up(T x, T y) { return (x + y - 1) / y; }
+
+template <typename T>
+inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
+#if 0
+    // onednn partition pattern
+    T& n_my = n_end;
+    if (nth <= 1 || n == 0) {
+        n_start = 0;
+        n_my = n;
+    } else {
+        T n1 = div_up(n, nth);
+        T n2 = n1 - 1;
+        T T1 = n - n2 * nth;
+        n_my = ith < T1 ? n1 : n2;
+        n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2;
+    }
+    n_end += n_start;
+#else
+    // pytorch aten partition pattern
+    T n_my = div_up(n, nth);
+    n_start = ith * n_my;
+    n_end = std::min(n_start + n_my, n);
+#endif
+}
+
+template <typename func_t>
+inline void parallel_for(int n, const func_t& f) {
+#if defined(_OPENMP)
+#pragma omp parallel
+{
+    int nth = omp_get_num_threads();
+    int ith = omp_get_thread_num();
+    int tbegin, tend;
+    balance211(n, nth, ith, tbegin, tend);
+    f(tbegin, tend);
+}
+#else
+    f(0, n);
+#endif
+}
+
+// for 1d parallel, use `actual_nth`
+// for 2d parallel, use even nths, e.g. 43->42
+int inline adjust_num_threads(int m) {
+  int actual_nth = at::get_num_threads();
+  if (m == 1) {
+    return actual_nth;
+  }
+  return std::max(1, (actual_nth >> 1) * 2);
+}
+
+template <typename func_t>
+inline void parallel_2d(int m, int n, const func_t& f) {
+
+  // make sure we have even num_threads
+  int nth = adjust_num_threads(m);
+
+  // [NOTE] thread blocking:
+  //
+  //   1) prefer square block per thread
+  //   2) use even number of CPU cores
+  //   3) use all `num_threads` cores
+  //
+  //   we have:
+  //     TM * TN = T
+  //     BM / TM = BN / TN
+  //   then:
+  //     TM = ((BM / BN) * T) ^ 0.5
+  //
+  float r = float(m) / n;
+  int nth_m = std::ceil(std::sqrt(r * nth));
+  int nth_n = 1;
+  for (; nth_m > 0; --nth_m) {
+    nth_n = nth / nth_m;
+    if (nth_m * nth_n == nth) {
+      break;
+    }
+  }
+
+#if defined(_OPENMP)
+#pragma omp parallel num_threads(nth)
+{
+  int ith = omp_get_thread_num();
+  int ith_m = ith / nth_n;
+  int ith_n = ith % nth_n;
+
+  int thread_block_m = div_up(m, nth_m);
+  int thread_block_n = div_up(n, nth_n);
+
+  int begin_m = ith_m * thread_block_m;
+  int end_m = std::min(m, begin_m + thread_block_m);
+  int begin_n = ith_n * thread_block_n;
+  int end_n = std::min(n, begin_n + thread_block_n);
+
+  f(begin_m, end_m, begin_n, end_n);
+}
+#else
+  f(0, m, 0, n);
+#endif
+}
+
+template <typename T>
+int get_cache_blocks(int BLOCK_SIZE, int K) {
+  // L2 2MB and ratio of 50%
+  const int L2_size = 2048 * 1024 >> 1;
+  return std::max(1, int(L2_size / (BLOCK_SIZE * K * sizeof(T))));
+}
+
+// data indexing for dimension collapse
+template <typename T>
+inline T data_index_init(T offset) {
+  return offset;
+}
+
+template <typename T, typename... Args>
+inline T data_index_init(T offset, T& x, const T& X, Args&&... args) {
+  offset = data_index_init(offset, std::forward<Args>(args)...);
+  x = offset % X;
+  return offset / X;
+}
+
+inline bool data_index_step() {
+  return true;
+}
+
+template <typename T, typename... Args>
+inline bool data_index_step(T& x, const T& X, Args&&... args) {
+  if (data_index_step(std::forward<Args>(args)...)) {
+    x = ((x + 1) == X) ? 0 : (x + 1);
+    return x == 0;
+  }
+  return false;
+}
+
+// forced unroll for perf critical path
+
+#if __has_attribute(always_inline)
+#define ALWAYS_INLINE __attribute__((__always_inline__)) inline
+#else
+#define ALWAYS_INLINE inline
+#endif
+
+template <int n>
+struct Unroll {
+  template <typename Func, typename... Args>
+  ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
+    Unroll<n - 1>{}(f, args...);
+    f(std::integral_constant<int, n - 1>{}, args...);
+  }
+};
+
+template <>
+struct Unroll<1> {
+  template <typename Func, typename... Args>
+  ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
+    f(std::integral_constant<int, 0>{}, args...);
+  }
+};
+
+} // anonymous namespace

csrc/cpu/sgl-kernels/gemm.cpp

@@ -0,0 +1,464 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+namespace {
+
+// packed   layout:
+//   quants {N, K}  int8_t
+//   comp   {N}     int32_t
+template <int BLOCK_N>
+inline void s8s8_compensation(int8_t* __restrict__ packed, int K) {
+#if defined(CPU_CAPABILITY_AVX512)
+  constexpr int COLS = BLOCK_N / 16;
+  __m512i vcomp[COLS];
+
+  for (int col = 0; col < COLS; ++col) {
+    vcomp[col] = _mm512_setzero_si512();
+  }
+
+  const int64_t offset = BLOCK_N * K;
+  const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
+  for (int k = 0; k < K / 4; ++k) {
+    for (int col = 0; col < COLS; ++col) {
+      __m512i vb = _mm512_loadu_si512((const __m512i *)(packed + k * BLOCK_N * 4 + col * 64));
+      vcomp[col] = _mm512_dpbusd_epi32(vcomp[col], off, vb);
+    }
+  }
+
+  for (int col = 0; col < COLS; ++col) {
+    _mm512_storeu_si512((__m512i *)(packed + offset + col * 64), vcomp[col]);
+  }
+#else
+  TORCH_CHECK(false, "s8s8_compensation not implemented!");
+#endif
+}
+
+// convert to vnni format
+// from [N, K] to [K/2, N, 2] for bfloat16 and float16
+template <typename packed_t>
+inline void pack_vnni(packed_t* __restrict__ packed, const packed_t* __restrict__ weight, int N, int K) {
+  const int VNNI_BLK = 2;
+  for (int n = 0; n < N; ++n) {
+    for (int k = 0; k < K / VNNI_BLK; ++k) {
+      for (int d = 0; d < VNNI_BLK; ++d) {
+        packed[k * N * VNNI_BLK + n * VNNI_BLK + d] = weight[n * K + k * VNNI_BLK + d];
+      }
+    }
+  }
+}
+
+template <>
+inline void pack_vnni<int8_t>(int8_t* __restrict__ packed, const int8_t* __restrict__ weight, int N, int K) {
+  constexpr int BLOCK_N = block_size_n();
+  TORCH_CHECK(N == BLOCK_N);
+
+  const int VNNI_BLK = 4;
+  for (int n = 0; n < N; ++n) {
+    for (int k = 0; k < K / VNNI_BLK; ++k) {
+      for (int d = 0; d < VNNI_BLK; ++d) {
+        packed[k * N * VNNI_BLK + n * VNNI_BLK + d] = weight[n * K + k * VNNI_BLK + d];
+      }
+    }
+  }
+  s8s8_compensation<BLOCK_N>(packed, K);
+}
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d]);
+  }
+}
+
+template <typename scalar_t>
+inline void copy_add_stub(scalar_t* __restrict__ out, const float* __restrict__ input, const float* __restrict__ bias, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d) + fVec::loadu(bias + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size()) + fVec::loadu(bias + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + bias[d]);
+  }
+}
+
+template <typename scalar_t, bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn {
+  static inline void apply(
+      const scalar_t* __restrict__ A, const scalar_t* __restrict__ B, scalar_t* __restrict__ C,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn<at::BFloat16, has_bias, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const at::BFloat16* __restrict__ A, const at::BFloat16* __restrict__ B, at::BFloat16* __restrict__ C,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+
+    // prefetch distance
+    constexpr int PREFETCH_SIZE_K = 0;
+
+    __m512bh va;
+    __m512bh vb[COLS];
+    __m512 vc[ROWS * COLS];
+
+    auto loadc = [&](auto i) {
+      constexpr int col = i % COLS;
+      if constexpr (has_bias) {
+        vc[i] = _mm512_loadu_ps(bias + col * 16);
+      } else {
+        vc[i] = _mm512_set1_ps(0.f);
+      }
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K2 = K >> 1;
+    const int64_t lda2 = lda >> 1;
+    const int64_t ldb2 = ldb; // ldb * 2 >> 1;
+    const float* a_ptr = reinterpret_cast<const float*>(A);
+    const float* b_ptr = reinterpret_cast<const float*>(B);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = (__m512bh)(_mm512_set1_ps(a_ptr[row * lda2 + k]));
+      }
+      if constexpr (row == 0) {
+        vb[col] = (__m512bh)(_mm512_loadu_si512(b_ptr + k * ldb2 + col * 16));
+        if constexpr (PREFETCH_SIZE_K > 0) {
+          _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2 + col * 16, _MM_HINT_T0);
+        }
+      }
+      vc[i] = _mm512_dpbf16_ps(vc[i], va, vb[col]);
+    };
+    for (int64_t k = 0; k < K2; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+      // for COLS = 2, 4 use 512bit store
+      // for COLS = 1, 3 use 256bit store
+      if constexpr (COLS % 2 == 0) {
+        if constexpr (col % 2 == 0) {
+          _mm512_storeu_si512(
+              reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+              (__m512i)(_mm512_cvtne2ps_pbh(vc[row * COLS + col + 1], vc[row * COLS + col])));
+        }
+      } else {
+        _mm256_storeu_si256(
+            reinterpret_cast<__m256i*>(C + row * ldc + col * 16),
+            (__m256i)(_mm512_cvtneps_pbh(vc[i])));
+      }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE)                          \
+    tinygemm_kernel_nn<scalar_t, has_bias, MB_SIZE, NB_SIZE>::apply(         \
+        A + mb_start * lda, B + nb_start * 2, C + mb_start * ldc + nb_start, \
+        has_bias ? bias + nb_start : nullptr, K, lda, ldb, ldc);
+
+template <typename scalar_t, bool has_bias>
+struct brgemm {
+  static inline void apply(
+      const scalar_t* __restrict__ A, const scalar_t* __restrict__ B, scalar_t* __restrict__ C,
+      float* __restrict__ Ctmp, const float* __restrict__ bias,
+      int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int BLOCK_N = block_size_n();
+    at::native::cpublas::brgemm(
+        M, N, K, lda, ldb, BLOCK_N, /* add_C */false,
+        A, B, Ctmp);
+
+    // copy from Ctmp to C
+    for (int64_t m = 0; m < M; ++m) {
+      if constexpr (has_bias) {
+        copy_add_stub(C + m * ldc, Ctmp + m * BLOCK_N, bias, N);
+      } else {
+        copy_stub(C + m * ldc, Ctmp + m * BLOCK_N, N);
+      }
+    }
+  }
+};
+
+template <typename scalar_t, bool has_bias>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const scalar_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg) {
+
+  if (brg) {
+    brgemm<scalar_t, has_bias>::apply(
+        A, B, C, Ctmp, bias,
+        M, N, K, lda, ldb, ldc);
+    return;
+  }
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        // mb_size = 1
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_NN(1, 32); break;
+        case 0x14: LAUNCH_TINYGEMM_KERNEL_NN(1, 64); break;
+        // mb_size = 2
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_NN(2, 32); break;
+        case 0x24: LAUNCH_TINYGEMM_KERNEL_NN(2, 64); break;
+        // mb_size = 3
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_NN(3, 32); break;
+        case 0x34: LAUNCH_TINYGEMM_KERNEL_NN(3, 64); break;
+        // mb_size = 4
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_NN(4, 32); break;
+        case 0x44: LAUNCH_TINYGEMM_KERNEL_NN(4, 64); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+void weight_packed_linear_kernel_impl(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ mat1,
+    const scalar_t* __restrict__ mat2,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t mat1_strideM,
+    int64_t out_strideM) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  // use avx512-bf16 when a) M is small; b) dtype is bfloat16, otherwise use amx
+  const bool use_brgemm = (M > 4) || (!std::is_same_v<scalar_t, at::BFloat16>);
+
+  // l2 cache block for n
+  int64_t cache_blocks_nb = get_cache_blocks<scalar_t>(BLOCK_N, K);
+
+  // parallel on [MB, NB]
+  AT_DISPATCH_BOOL(bias != nullptr, has_bias, [&] {
+    parallel_2d(MB, NB, [&](int64_t begin_mb, int64_t end_mb, int64_t begin_nb, int64_t end_nb) {
+
+      // for brgemm, use float32 for accumulate
+      alignas(64) float Ctmp[BLOCK_M * BLOCK_N];
+
+      for (int64_t nbb = begin_nb; nbb < end_nb; nbb += cache_blocks_nb) {
+      for (int64_t mb = begin_mb; mb < end_mb; ++mb) {
+      for (int64_t nb = nbb; nb < std::min(nbb + cache_blocks_nb, end_nb); ++nb) {
+
+        int64_t mb_start = mb * BLOCK_M;
+        int64_t mb_size = std::min(M - mb_start, BLOCK_M);
+        int64_t nb_start = nb * BLOCK_N;
+        int64_t nb_size = std::min(N - nb_start, BLOCK_N);
+
+        tinygemm_kernel<scalar_t, has_bias>(
+            /*   A */ mat1 + mb_start * mat1_strideM,
+            /*   B */ mat2 + nb_start * K /* nb * BLOCK_N * K */,
+            /*   C */ out + mb_start * out_strideM + nb_start,
+            /* Ctmp*/ Ctmp,
+            /* bias*/ bias + nb_start,
+            /*   M */ mb_size,
+            /*   N */ nb_size,
+            /*   K */ K,
+            /* lda */ mat1_strideM,
+            /* ldb */ nb_size,
+            /* ldc */ out_strideM,
+            /* brg */ use_brgemm);
+      }}}
+
+      if (use_brgemm) {
+        at::native::cpublas::brgemm_release();
+      }
+    });
+  });
+}
+
+} // anonymous namespace
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(const scalar_t* __restrict__ A, const scalar_t* __restrict__ B, scalar_t* __restrict__ C,
+    float* __restrict__ Ctmp, int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc, bool brg) {
+  tinygemm_kernel<scalar_t, false>(A, B, C, Ctmp, nullptr, M, N, K, lda, ldb, ldc, brg);
+}
+
+#define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE)                                             \
+    template void tinygemm_kernel<TYPE>(                                                \
+        const TYPE* __restrict__ A, const TYPE* __restrict__ B, TYPE* __restrict__ C,   \
+        float* __restrict__ Ctmp, int64_t M, int64_t N, int64_t K, int64_t lda,         \
+        int64_t ldb, int64_t ldc, bool brg)
+
+INSTANTIATE_TINYGEMM_TEMPLATE(at::BFloat16);
+INSTANTIATE_TINYGEMM_TEMPLATE(at::Half);
+
+at::Tensor convert_weight_packed(at::Tensor& weight) {
+  // for 3d moe weights
+  // weight : [E, OC, IC]
+  //     w1 : [E, 2N,  K]
+  //     w2 : [E,  K,  N]
+  CHECK_INPUT(weight);
+
+  const int64_t ndim = weight.ndimension();
+  TORCH_CHECK(ndim == 2 || ndim == 3, "expect weight to be 2d or 3d, got ", ndim, "d tensor.");
+  const auto st = weight.scalar_type();
+  const int64_t E = ndim == 3 ? weight.size(0) : 1;
+  const int64_t OC = ndim == 3 ? weight.size(1) : weight.size(0);
+  const int64_t IC = ndim == 3 ? weight.size(2) : weight.size(1);
+
+  // we handle 2 TILE_N at a time.
+  TORCH_CHECK(OC % TILE_N == 0, "invalid weight out features ", OC);
+  TORCH_CHECK(IC % TILE_K == 0, "invalid weight input features ", IC);
+
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t NB = div_up(OC, BLOCK_N);
+
+  // use phony sizes here [E, OC, IC], for each [E], [OC, IC] -> [IC / 2, OC, 2]
+  auto packed_weight = at::empty({}, weight.options());
+  const int64_t stride = OC * IC;
+
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf || st == at::kChar || st == at::kFloat8_e4m3fn,
+      "expect weight to be bfloat16, float16, int8 or fp8_e4m3.");
+
+  CPU_DISPATCH_PACKED_TYPES(st, [&] {
+    // adjust most inner dimension size
+    const int packed_row_size = get_row_size<packed_t>(IC);
+    auto sizes = weight.sizes().vec();
+    sizes[ndim - 1] = packed_row_size;
+    packed_weight.resize_(sizes);
+
+    const packed_t* w_data = weight.data_ptr<packed_t>();
+    packed_t* packed_data = packed_weight.data_ptr<packed_t>();
+
+    // parallel on {E, NB}
+    at::parallel_for(0, E * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t e{0}, nb{0};
+      data_index_init(begin, e, E, nb, NB);
+
+      for (int64_t i = begin; i < end; ++i) {
+        UNUSED(i);
+
+        int64_t n = nb * BLOCK_N;
+        int64_t n_size = std::min(BLOCK_N, OC - n);
+        pack_vnni<packed_t>(
+            packed_data + e * OC * packed_row_size + n * packed_row_size,
+            w_data + e * stride + n * IC,
+            n_size,
+            IC);
+
+        // move to the next index
+        data_index_step(e, E, nb, NB);
+      }
+    });
+  });
+  return packed_weight;
+}
+
+// mat1 : [M, K]
+// mat2 : [N, K]
+// bias : [N]
+// out  : [M, N]
+//
+at::Tensor weight_packed_linear(at::Tensor& mat1, at::Tensor& mat2,
+    const std::optional<at::Tensor>& bias, bool is_vnni) {
+  RECORD_FUNCTION(
+    "sgl-kernel::weight_packed_linear", std::vector<c10::IValue>({mat1, mat2, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(mat1);
+  CHECK_INPUT(mat2);
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat2.size(1);
+  CHECK_EQ(mat1.size(1), K);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  auto out = at::empty({M, N}, mat1.options());
+
+  // strides
+  int64_t mat1_strideM = mat1.stride(0);
+  int64_t out_strideM = out.stride(0);
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(mat1.scalar_type(), "weight_packed_linear_kernel_impl", [&] {
+    weight_packed_linear_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        mat1.data_ptr<scalar_t>(),
+        packed_w.data_ptr<scalar_t>(),
+        bias_data,
+        M,
+        N,
+        K,
+        mat1_strideM,
+        out_strideM);
+  });
+
+  return out;
+}

csrc/cpu/sgl-kernels/gemm.h

@@ -0,0 +1,266 @@
+#pragma once
+
+#include <ATen/native/CPUBlas.h>
+
+// clang-format off
+
+// amx-bf16
+#define TILE_M 16
+#define TILE_N 16
+#define TILE_K 32
+
+// block size for AMX gemm
+constexpr int block_size_m() { return 2 * TILE_M; }
+constexpr int block_size_n() { return 2 * TILE_N; }
+
+// define threshold using brgemm (intel AMX)
+template <typename T> inline bool can_use_brgemm(int M);
+template <> inline bool can_use_brgemm<at::BFloat16>(int M) { return M > 4; }
+template <> inline bool can_use_brgemm<at::Half>(int M) { return true; }
+// TODO: add u8s8 brgemm, this requires PyTorch 2.7
+template <> inline bool can_use_brgemm<int8_t>(int M) { return false; }
+template <> inline bool can_use_brgemm<at::Float8_e4m3fn>(int M) { return M > 4; }
+template <> inline bool can_use_brgemm<at::quint4x2>(int M) { return M > 4; }
+
+// work around compiler internal error
+#define BLOCK_K 128 // 4 * TILE_K
+
+// adjust leading dimension size for K
+template <typename T>
+inline int64_t get_row_size(int64_t K) {
+  return K;
+}
+
+template <>
+inline int64_t get_row_size<int8_t>(int64_t K) {
+  return K + sizeof(int32_t);
+}
+
+inline int64_t get_row_size(int64_t K, bool use_int8_w8a8) {
+  return use_int8_w8a8 ? K + sizeof(int32_t) : K;
+}
+
+// pack weight to vnni format
+at::Tensor convert_weight_packed(at::Tensor& weight);
+
+// moe implementations for int8 w8a8
+template <typename scalar_t>
+void fused_experts_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    uint8_t* __restrict__ A_tmp,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad);
+
+// moe implementations for fp8 w8a16
+template <typename scalar_t>
+void fused_experts_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    scalar_t* __restrict__ A_tmp,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad);
+
+// moe implementations for int4 w4a16
+template <typename scalar_t>
+void fused_experts_int4_w4a16_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    scalar_t* __restrict__ A_tmp,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::quint4x2* __restrict__ packed_w1,
+    const at::quint4x2* __restrict__ packed_w2,
+    const uint8_t* __restrict__ w1z,
+    const uint8_t* __restrict__ w2z,
+    const scalar_t* __restrict__ w1s,
+    const scalar_t* __restrict__ w2s,
+    int group_size,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad);
+
+// shared expert implememntation for int8 w8a8
+template <typename scalar_t>
+void shared_expert_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K);
+
+template <typename scalar_t>
+void shared_expert_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K);
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const scalar_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    float* __restrict__ Ctmp,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    int32_t* __restrict__ Ctmp,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::Float8_e4m3fn* __restrict__ B,
+    scalar_t* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ scale,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::quint4x2* __restrict__ B,
+    scalar_t* __restrict__ C,
+    const uint8_t* __restrict__ Bz,
+    const scalar_t* __restrict__ Bs,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int group_size,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    int64_t strideBz,
+    int64_t strideBs,
+    bool brg);
+
+// TODO: debug print, remove me later
+inline void print_16x32i(const __m512i x) {
+  int32_t a[16];
+  _mm512_storeu_si512((__m512i *)a, x);
+
+  for (int i = 0; i < 16; i++){
+    std::cout << a[i] << " ";
+  }
+  std::cout << std::endl;
+}
+
+inline void print_16x32(const __m512 x) {
+  float a[16];
+  _mm512_storeu_ps((__m512 *)a, x);
+
+  for (int i = 0; i < 16; i++){
+    std::cout << a[i] << " ";
+  }
+  std::cout << std::endl;
+}
+
+
+inline void print_32x8u(const __m256i x) {
+  uint8_t a[32];
+  _mm256_storeu_si256((__m256i *)a, x);
+
+  for (int i = 0; i < 32; ++i) {
+    std::cout << int32_t(a[i]) << " ";
+  }
+  std::cout << std::endl;
+}

csrc/cpu/sgl-kernels/gemm_fp8.cpp

@@ -0,0 +1,530 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+// we use 4x32 for BLOCK_M
+#define BLOCK_SIZE_M_SCALE 4
+
+namespace {
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const float* __restrict__ input, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d]);
+  }
+}
+
+template <typename scalar_t>
+inline void copy_add_stub(scalar_t* __restrict__ out, const float* __restrict__ input, const float* __restrict__ bias, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d) + fVec::loadu(bias + d);
+    fVec data1 = fVec::loadu(input + d + fVec::size()) + fVec::loadu(bias + d + fVec::size());
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + bias[d]);
+  }
+}
+
+inline void unpack_B(
+    at::BFloat16* __restrict__ Btmp,
+    const at::Float8_e4m3fn* __restrict__ packed_B,
+    int N,
+    int K,
+    int ldb,
+    int ldb_tmp,
+    float scale) {
+#if defined(CPU_CAPABILITY_AVX512)
+  // [K/2, N, 2]
+  const int K2 = K >> 1;
+  const int ldb2 = ldb; // ldb * 2 >> 1;
+  const uint16_t* b_ptr = reinterpret_cast<const uint16_t*>(packed_B);
+  const __m512 vd = _mm512_set1_ps(scale);
+
+  constexpr int BLOCK_N = block_size_n();
+  static_assert(BLOCK_N == 32);
+
+  // prefetch distance
+  constexpr int PREFETCH_SIZE_K = 64;
+
+#pragma GCC unroll 4
+  for (int k = 0; k < K2; ++k) {
+    __m512i b8 = _mm512_loadu_si512(b_ptr + k * ldb2);
+    if constexpr (PREFETCH_SIZE_K > 0) {
+      _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2, _MM_HINT_T0);
+    }
+
+    __m256i b8_0 = _mm512_extracti32x8_epi32(b8, 0);
+    __m256i b8_1 = _mm512_extracti32x8_epi32(b8, 1);
+
+    __m512bh bf16_0 = CVT_FP8_TO_BF16(b8_0);
+    __m512bh bf16_1 = CVT_FP8_TO_BF16(b8_1);
+
+    // Apply scale
+    __m512 f0_lo = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_0, 0));
+    __m512 f0_hi = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_0, 1));
+    __m512 f1_lo = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_1, 0));
+    __m512 f1_hi = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32((__m512i)bf16_1, 1));
+
+    f0_lo = _mm512_mul_ps(f0_lo, vd);
+    f0_hi = _mm512_mul_ps(f0_hi, vd);
+    f1_lo = _mm512_mul_ps(f1_lo, vd);
+    f1_hi = _mm512_mul_ps(f1_hi, vd);
+
+    bf16_0 = _mm512_cvtne2ps_pbh(f0_hi, f0_lo);
+    bf16_1 = _mm512_cvtne2ps_pbh(f1_hi, f1_lo);
+
+    _mm512_storeu_si512(Btmp + k * ldb_tmp * 2 + 0, (__m512i)bf16_0);
+    _mm512_storeu_si512(Btmp + k * ldb_tmp * 2 + 32, (__m512i)bf16_1);
+  }
+#else
+  TORCH_CHECK(false, "unpack_B: scalar path not implemented!");
+#endif
+}
+
+template <typename scalar_t, typename packed_t, bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn {
+  static inline void apply(
+      const scalar_t* __restrict__ A, const packed_t* __restrict__ B, scalar_t* __restrict__ C,
+      const float* __restrict__ bias, const float* __restrict__ scale, int K, int lda, int ldb, int ldc, int64_t block_size_K) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn<at::BFloat16, at::Float8_e4m3fn, has_bias, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const at::BFloat16* __restrict__ A, const at::Float8_e4m3fn* __restrict__ B, at::BFloat16* __restrict__ C,
+      const float* __restrict__ bias, const float* __restrict__ scale, int K, int lda, int ldb, int ldc, int64_t block_size_K) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+
+    const int KB = div_up(K, BLOCK_K);
+
+    // prefetch distance
+    constexpr int PREFETCH_SIZE_K = 64;
+    constexpr int PREFETCH_SIZE_KB = 1;
+
+    __m512bh va;
+    __m512bh vb[COLS];
+    __m512 vc[ROWS * COLS];
+    __m512 vsum[ROWS * COLS];
+
+    // block quant scale
+    __m512 vscale;
+
+    auto loadc = [&](auto i) {
+      constexpr int col = i % COLS;
+      if constexpr (has_bias) {
+        vc[i] = _mm512_loadu_ps(bias + col * 16);
+      } else {
+        vc[i] = _mm512_setzero_ps();
+      }
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int lda2 = lda >> 1;
+    const int ldb2 = ldb; // ldb * 2 >> 1;
+    const float* a_ptr = reinterpret_cast<const float*>(A);
+    const uint16_t* b_ptr = reinterpret_cast<const uint16_t*>(B);
+
+    auto compute = [&](auto i, int k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = (__m512bh)(_mm512_set1_ps(a_ptr[row * lda2 + k]));
+        if constexpr (PREFETCH_SIZE_K > 0) {
+          _mm_prefetch(a_ptr + row * lda2 + k + PREFETCH_SIZE_K, _MM_HINT_T0);
+        }
+      }
+      if constexpr (row == 0) {
+        if constexpr (col % 2 == 0) {
+          __m512i b8 = _mm512_loadu_si512(b_ptr + k * ldb2 + col * 16);
+          if constexpr (PREFETCH_SIZE_K > 0) {
+            _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb2 + col * 16, _MM_HINT_T0);
+          }
+          vb[col + 0] = CVT_FP8_TO_BF16(_mm512_extracti32x8_epi32(b8, 0));
+          vb[col + 1] = CVT_FP8_TO_BF16(_mm512_extracti32x8_epi32(b8, 1));
+        }
+      }
+      vsum[i] = _mm512_dpbf16_ps(vsum[i], va, vb[col]);
+    };
+
+    constexpr int BLOCK_K2 = BLOCK_K >> 1;
+    for (int kb = 0; kb < KB; ++kb) {
+      int kb_start = kb * BLOCK_K2;
+      int kb_end = std::min(K, kb_start + BLOCK_K2);
+      // 1. load scale vector
+      vscale = _mm512_set1_ps(scale[kb]);
+      if constexpr (PREFETCH_SIZE_KB > 0) {
+        _mm_prefetch(scale + kb + PREFETCH_SIZE_KB, _MM_HINT_T0);
+      }
+      // 2. zero vsum for each block
+      Unroll<ROWS * COLS>{}([&](auto i) {
+        vsum[i] = _mm512_setzero_ps();
+      });
+      // 3. accumulate across each block
+      for (int k = kb_start; k < kb_end; ++k) {
+        Unroll<ROWS * COLS>{}(compute, k);
+      }
+      // 4. apply scale
+      Unroll<ROWS * COLS>{}([&](auto i) {
+        vc[i] = _mm512_fmadd_ps(vsum[i], vscale, vc[i]);
+      });
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+      // for COLS = 2,4 use 512bit store
+      if constexpr (col % 2 == 0) {
+        _mm512_storeu_si512(
+            reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+            (__m512i)(_mm512_cvtne2ps_pbh(vc[row * COLS + col + 1], vc[row * COLS + col])));
+      }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE)                          \
+    tinygemm_kernel_nn<scalar_t, at::Float8_e4m3fn, has_bias, MB_SIZE, NB_SIZE>::apply(         \
+        A + mb_start * lda, B + nb_start * 2, C + mb_start * ldc + nb_start, \
+        has_bias ? bias + nb_start : nullptr, scale, K, lda, ldb, ldc, block_size_K);
+
+template <typename scalar_t, typename packed_t, bool has_bias>
+struct brgemm {
+  static inline void apply(
+      const scalar_t* __restrict__ A,
+      const packed_t* __restrict__ B,
+      scalar_t* __restrict__ C,
+      scalar_t* __restrict__ Btmp,
+      float* __restrict__ Ctmp,
+      const float* __restrict__ bias,
+      const float* __restrict__ scale,
+      int M,
+      int N,
+      int K,
+      int lda,
+      int ldb,
+      int ldc) {
+    TORCH_CHECK(false, "struct brgemm: primary template not implemented!");
+  }
+};
+
+template <bool has_bias>
+struct brgemm<at::BFloat16, at::Float8_e4m3fn, has_bias> {
+  static inline void apply(
+      const at::BFloat16* __restrict__ A,
+      const at::Float8_e4m3fn* __restrict__ B,
+      at::BFloat16* __restrict__ C,
+      at::BFloat16* __restrict__ Btmp,
+      float* __restrict__ Ctmp,
+      const float* __restrict__ bias,
+      const float* __restrict__ scale,
+      int M,
+      int N,
+      int K,
+      int lda,
+      int ldb,
+      int ldc) {
+
+    constexpr int BLOCK_N = block_size_n();
+
+    // [K, BLOCK_N] -> [K / 2, BLOCK_N * 2]
+    const int ldb_tmp = BLOCK_N;
+
+    for (int k = 0; k < K; k += BLOCK_K) {
+      int kb_size = std::min(BLOCK_K, K - k);
+
+      int idx = k >> 7; // k / BLOCK_K where BLOCK_K = 128
+      unpack_B(Btmp + k * ldb_tmp, B + k * ldb, N, kb_size, ldb, ldb_tmp, scale[idx]);
+    }
+
+    at::native::cpublas::brgemm(
+        M, N, K, lda, ldb_tmp, BLOCK_N, /* add_C */ false, A, Btmp, Ctmp);
+
+    // copy from Ctmp to C
+    for (int m = 0; m < M; ++m) {
+      if constexpr (has_bias) {
+        copy_add_stub(C + m * ldc, Ctmp + m * BLOCK_N, bias, N);
+      } else {
+        copy_stub(C + m * ldc, Ctmp + m * BLOCK_N, N);
+      }
+    }
+  }
+};
+
+template <typename scalar_t, bool has_bias>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::Float8_e4m3fn* __restrict__ B,
+    scalar_t* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ scale,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K) {
+
+  if (brg) {
+    brgemm<scalar_t, at::Float8_e4m3fn, has_bias>::apply(
+        A, B, C, Btmp, Ctmp, bias, scale, M, N, K, lda, ldb, ldc);
+    return;
+  }
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_NN(1, 32); break;
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_NN(2, 32); break;
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_NN(3, 32); break;
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_NN(4, 32); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+void fp8_scaled_mm_kernel_impl(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ mat1,
+    const at::Float8_e4m3fn* __restrict__ mat2,
+    const float* __restrict__ scales2,
+    const float* __restrict__ bias,
+    scalar_t* __restrict__ buffer,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t mat1_strideM,
+    int64_t out_strideM,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    int64_t buffer_size_per_thread) {
+
+  constexpr int64_t BLOCK_M = block_size_m() * BLOCK_SIZE_M_SCALE;
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  const int64_t scale_size_K = div_up(K, block_size_K);
+  const int64_t blocks_n_per_group = block_size_N / BLOCK_N;
+
+  const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(M);
+
+  // parallel on [MB, NB]
+  AT_DISPATCH_BOOL(bias != nullptr, has_bias, [&] {
+    at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t mb{0}, nb{0};
+      data_index_init(begin, mb, MB, nb, NB);
+
+      int tid = at::get_thread_num();
+      scalar_t* __restrict__ Btmp = buffer + tid * buffer_size_per_thread;
+      float* __restrict__ Ctmp = (float*)((void*)(Btmp + BLOCK_N * K));
+
+      for (int64_t i = begin; i < end; ++i) {
+        UNUSED(i);
+        const float* scale_ptr = scales2 + (nb / blocks_n_per_group) * scale_size_K;
+
+        int64_t mb_start = mb * BLOCK_M;
+        int64_t mb_size = std::min(M - mb_start, BLOCK_M);
+        int64_t nb_start = nb * BLOCK_N;
+        int64_t nb_size = std::min(N - nb_start, BLOCK_N);
+
+        tinygemm_kernel<scalar_t, has_bias>(
+            /*   A            */ mat1 + mb_start * mat1_strideM,
+            /*   B            */ mat2 + nb_start * K, // nb * BLOCK_N * K
+            /*   C            */ out + mb_start * out_strideM + nb_start,
+            /*   Btmp         */ Btmp,
+            /*   Ctmp         */ Ctmp,
+            /*   scale        */ scale_ptr,
+            /*   bias         */ bias + nb_start,
+            /*   M            */ mb_size,
+            /*   N            */ nb_size,
+            /*   K            */ K,
+            /*   lda          */ mat1_strideM,
+            /*   ldb          */ nb_size,
+            /*   ldc          */ out_strideM,
+            /*   brg          */ use_brgemm,
+            /*   block_size_K */ block_size_K);
+
+        // move to the next index
+        data_index_step(mb, MB, nb, NB);
+      }
+
+      if (use_brgemm) {
+        at::native::cpublas::brgemm_release();
+      }
+    });
+  });
+}
+
+} // anonymous namespace
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(
+    const scalar_t* __restrict__ A,
+    const at::Float8_e4m3fn* __restrict__ B,
+    scalar_t* __restrict__ C,
+    scalar_t* __restrict__ Btmp,
+    float* __restrict__ Ctmp,
+    const float* __restrict__ scale,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg,
+    int64_t block_size_K) {
+  tinygemm_kernel<scalar_t, false>(A, B, C, Btmp, Ctmp, scale, nullptr, M, N, K, lda, ldb, ldc, brg, block_size_K);
+}
+
+#define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE)    \
+  template void tinygemm_kernel<TYPE>(         \
+      const TYPE* __restrict__ A,              \
+      const at::Float8_e4m3fn* __restrict__ B, \
+      TYPE* __restrict__ C,                    \
+      TYPE* __restrict__ Btmp,                 \
+      float* __restrict__ Ctmp,                \
+      const float* __restrict__ scale,         \
+      int64_t M,                               \
+      int64_t N,                               \
+      int64_t K,                               \
+      int64_t lda,                             \
+      int64_t ldb,                             \
+      int64_t ldc,                             \
+      bool brg,                                \
+      int64_t block_size_K)
+
+INSTANTIATE_TINYGEMM_TEMPLATE(at::BFloat16);
+INSTANTIATE_TINYGEMM_TEMPLATE(at::Half);
+
+at::Tensor fp8_scaled_mm_cpu(at::Tensor& mat1, at::Tensor& mat2, at::Tensor& scales2,
+    std::vector<int64_t> block_size, std::optional<at::Tensor>& bias,
+    at::ScalarType out_dtype, bool is_vnni) {
+  RECORD_FUNCTION("sgl-kernel::fp8_scaled_mm_cpu", std::vector<c10::IValue>({mat1, mat2, scales2, block_size, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(mat1);
+  CHECK_INPUT(mat2);
+  CHECK_INPUT(scales2);
+  TORCH_CHECK(scales2.scalar_type() == at::kFloat,
+      "fp8_scaled_mm_cpu: expect scales2 to be float32.");
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat2.size(1);
+
+  CHECK_EQ(mat1.size(1), K);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  TORCH_CHECK(block_size.size() == 2,
+      "fp8_scaled_mm_cpu: expect block_size.size() to be 2.");
+
+  int64_t block_size_N = block_size[0];
+  int64_t block_size_K = block_size[1];
+
+  constexpr int64_t BLOCK_M = block_size_m() * BLOCK_SIZE_M_SCALE;
+  constexpr int64_t BLOCK_N = block_size_n();
+  TORCH_CHECK(block_size_N % BLOCK_N == 0, "fp8_scaled_mm_cpu: expect block_size_N to be multiples of BLOCK_N");
+  TORCH_CHECK(block_size_K == BLOCK_K, "fp8_scaled_mm_cpu: expect block_size_K equals to BLOCK_K");
+  CHECK_EQ(scales2.size(0), div_up(N, block_size_N));
+  CHECK_EQ(scales2.size(1), div_up(K, block_size_K));
+
+  const auto st = mat1.scalar_type();
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf,
+      "fp8_scaled_mm_cpu: expect A to be bfloat16 or half.");
+  TORCH_CHECK(st == out_dtype,
+      "fp8_scaled_mm_cpu: expect A has same dtype with out_dtype.");
+  TORCH_CHECK(mat2.scalar_type() == at::kFloat8_e4m3fn,
+      "fp8_scaled_mm_cpu: expect mat2 to be fp8_e4m3.");
+  TORCH_CHECK(scales2.scalar_type() == at::kFloat,
+      "fp8_scaled_mm_cpu: expect scales to be float32.");
+  auto out = at::empty({M, N}, mat1.options().dtype(out_dtype));
+
+  // strides
+  int64_t mat1_strideM = mat1.stride(0);
+  int64_t out_strideM = out.stride(0);
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  // Btmp : [T, BLOCK_N * K]
+  // Ctmp : [T, BLOCK_M * BLOCK_N]
+  int num_threads = at::get_num_threads();
+  int64_t size_per_thread = BLOCK_N * K + BLOCK_M * BLOCK_N * 2;
+  auto buffer = at::empty({num_threads, size_per_thread}, mat1.options());
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "fp8_scaled_mm_kernel_impl", [&] {
+    fp8_scaled_mm_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        mat1.data_ptr<scalar_t>(),
+        packed_w.data_ptr<at::Float8_e4m3fn>(),
+        scales2.data_ptr<float>(),
+        bias_data,
+        buffer.data_ptr<scalar_t>(),
+        M,
+        N,
+        K,
+        mat1_strideM,
+        out_strideM,
+        block_size_N,
+        block_size_K,
+        size_per_thread);
+  });
+
+  return out;
+}

csrc/cpu/sgl-kernels/gemm_int8.cpp

@@ -0,0 +1,440 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+namespace {
+
+template <typename scalar_t, bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, scalar_t* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <bool has_bias, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_nn<at::BFloat16, has_bias, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, at::BFloat16* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      const float* __restrict__ bias, int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+    static_assert(COLS % 2 == 0);
+
+    // prefetch distance
+    constexpr int PREFETCH_SIZE_K = 0;
+
+    __m512i va;
+    __m512i vb[COLS];
+    __m512i vc[ROWS * COLS];
+    __m512i vcomp[COLS];
+    __m512  vd0;
+    __m512  vd1[COLS];
+
+    // oops! 4x4 spills but luckly we use 4x2
+    __m512 vbias[COLS];
+
+    // [NOTE]: s8s8 igemm compensation in avx512-vnni
+    //
+    // avx512-vnni has no s8s8, so we need to change s8s8 to u8s8 with compensate:
+    //
+    //   a * b = (a + 128) * b - 128 * b
+    //   s   s       u       s    u    s
+    //
+    // 1) 128 * b is pre-computed when packing B to vnni formats
+    // 2) a + 128 is fused when dynamically quantize A
+    //
+    auto loadc = [&](auto i) {
+      vc[i] = _mm512_set1_epi32(0);
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K4 = K >> 2;
+    const int64_t lda4 = lda >> 2;
+    const int64_t ldb4 = ldb; // ldb * 4 >> 2;
+    const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A);
+    const int32_t* b_ptr = reinterpret_cast<const int32_t*>(B);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = _mm512_set1_epi32(a_ptr[row * lda4 + k]);
+      }
+      if constexpr (row == 0) {
+        vb[col] = _mm512_loadu_si512(b_ptr + k * ldb4 + col * 16);
+        if constexpr (PREFETCH_SIZE_K > 0) {
+          _mm_prefetch(b_ptr + (k + PREFETCH_SIZE_K) * ldb4 + col * 16, _MM_HINT_T0);
+        }
+      }
+      vc[i] = _mm512_dpbusd_epi32(vc[i], va, vb[col]);
+    };
+    for (int64_t k = 0; k < K4; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      // load a scale
+      if constexpr(col == 0) {
+        vd0 = _mm512_set1_ps(As[row]);
+      }
+      // load b scale and vcomp per 2 vectors
+      // also load bias if any
+      if constexpr (row == 0) {
+        if constexpr (col % 2 == 0) {
+          vd1[col + 0] = _mm512_loadu_ps(Bs + col * 16);
+          vd1[col + 1] = _mm512_loadu_ps(Bs + col * 16 + 16);
+          vcomp[col + 0] = _mm512_loadu_si512(Bcomp + col * 16);
+          vcomp[col + 1] = _mm512_loadu_si512(Bcomp + col * 16 + 16);
+          if constexpr (has_bias) {
+            vbias[col + 0] = _mm512_loadu_ps(bias + col * 16);
+            vbias[col + 1] = _mm512_loadu_ps(bias + col * 16 + 16);
+          }
+        }
+      }
+
+      // for COLS = 2, 4 use 512bit store
+      if constexpr (col % 2 == 0) {
+        __m512 vc0 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[row * COLS + col + 0], vcomp[col + 0]));
+        __m512 vc1 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[row * COLS + col + 1], vcomp[col + 1]));
+        if constexpr (has_bias) {
+          vc0 = _mm512_fmadd_ps(_mm512_mul_ps(vc0, vd0), vd1[col + 0], vbias[col + 0]);
+          vc1 = _mm512_fmadd_ps(_mm512_mul_ps(vc1, vd0), vd1[col + 1], vbias[col + 1]);
+        } else {
+          vc0 = _mm512_mul_ps(_mm512_mul_ps(vc0, vd0), vd1[col + 0]);
+          vc1 = _mm512_mul_ps(_mm512_mul_ps(vc1, vd0), vd1[col + 1]);
+        }
+
+        _mm512_storeu_si512(
+            reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+            (__m512i)(_mm512_cvtne2ps_pbh(vc1, vc0)));
+      }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_NN(MB_SIZE, NB_SIZE)                          \
+    tinygemm_kernel_nn<scalar_t, has_bias, MB_SIZE, NB_SIZE>::apply(         \
+        A + mb_start * lda, B + nb_start * 4, C + mb_start * ldc + nb_start, \
+        As + mb_start, Bs + nb_start, Bcomp + nb_start,                      \
+        has_bias ? bias + nb_start : nullptr, K, lda, ldb, ldc);
+
+template <typename scalar_t, bool has_bias>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B,
+    scalar_t* __restrict__ C,
+    int32_t* __restrict__ Ctmp,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc,
+    bool brg) {
+
+  // B compensation
+  const int32_t* Bcomp = reinterpret_cast<const int32_t*>(B + block_size_n() * K);
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int64_t mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        // mb_size = 1
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_NN(1, 32); break;
+        case 0x14: LAUNCH_TINYGEMM_KERNEL_NN(1, 64); break;
+        // mb_size = 2
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_NN(2, 32); break;
+        case 0x24: LAUNCH_TINYGEMM_KERNEL_NN(2, 64); break;
+        // mb_size = 3
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_NN(3, 32); break;
+        case 0x34: LAUNCH_TINYGEMM_KERNEL_NN(3, 64); break;
+        // mb_size = 4
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_NN(4, 32); break;
+        case 0x44: LAUNCH_TINYGEMM_KERNEL_NN(4, 64); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+template<typename scalar_t>
+void int8_scaled_mm_kernel_impl(
+    scalar_t* __restrict__ out,
+    const uint8_t* __restrict__ mat1,
+    const int8_t* __restrict__ mat2,
+    const float* __restrict__ scales1,
+    const float* __restrict__ scales2,
+    const float* __restrict__ bias,
+    int64_t M,
+    int64_t N,
+    int64_t K) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  // TODO: brgemm u8s8 depends on PyTorch 2.7 release.
+  const bool use_brgemm = false;
+
+  // K + 4 after compensation
+  const int64_t packed_row_size = get_row_size<int8_t>(K);
+
+  AT_DISPATCH_BOOL(bias != nullptr, has_bias, [&] {
+    at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+      int64_t mb{0}, nb{0};
+      data_index_init(begin, mb, MB, nb, NB);
+
+      // for brgemm, use int32_t for accumulate
+      alignas(64) int32_t Ctmp[BLOCK_M * BLOCK_N];
+
+      for (int i = begin; i < end; ++i) {
+        UNUSED(i);
+        int mb_start = mb * BLOCK_M;
+        int mb_size = std::min(M - mb_start, BLOCK_M);
+        int nb_start = nb * BLOCK_N;
+        int nb_size = std::min(N - nb_start, BLOCK_N);
+
+        tinygemm_kernel<scalar_t, has_bias>(
+            /*   A */ mat1 + mb_start * K,
+            /*   B */ mat2 + nb_start * packed_row_size /* nb * BLOCK_N * (K + 4) */,
+            /*   C */ out + mb_start * N + nb_start,
+            /* Ctmp*/ Ctmp,
+            /*  As */ scales1 + mb_start,
+            /*  Bs */ scales2 + nb_start,
+            /* bias*/ bias + nb_start,
+            /*   M */ mb_size,
+            /*   N */ nb_size,
+            /*   K */ K,
+            /* lda */ K,
+            /* ldb */ nb_size,
+            /* ldc */ N,
+            /* brg */ use_brgemm);
+
+        // move to the next index
+        data_index_step(mb, MB, nb, NB);
+      }
+
+      if (use_brgemm) {
+        at::native::cpublas::brgemm_release();
+      }
+    });
+  });
+}
+
+} // anonymous namespace
+
+// tinygemm interface
+template <typename scalar_t>
+void tinygemm_kernel(const uint8_t* __restrict__ A, const int8_t* __restrict__ B, scalar_t* __restrict__ C,
+    int32_t* __restrict__ Ctmp,  const float* __restrict__ As, const float* __restrict__ Bs,
+    int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc, bool brg) {
+  tinygemm_kernel<scalar_t, false>(A, B, C, Ctmp, As, Bs, nullptr, M, N, K, lda, ldb, ldc, brg);
+}
+
+#define INSTANTIATE_TINYGEMM_TEMPLATE(TYPE)                                                     \
+    template void tinygemm_kernel<TYPE>(                                                        \
+        const uint8_t* __restrict__ A, const int8_t* __restrict__ B, TYPE* __restrict__ C,      \
+        int32_t* __restrict__ Ctmp, const float* __restrict__ As, const float* __restrict__ Bs, \
+        int64_t M, int64_t N, int64_t K, int64_t lda, int64_t ldb, int64_t ldc, bool brg)
+
+INSTANTIATE_TINYGEMM_TEMPLATE(at::BFloat16);
+INSTANTIATE_TINYGEMM_TEMPLATE(at::Half);
+
+std::tuple<at::Tensor, at::Tensor> per_token_quant_int8_cpu(at::Tensor& A) {
+  RECORD_FUNCTION("sgl-kernel::per_token_quant_int8_cpu", std::vector<c10::IValue>({A}));
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(A);
+  CHECK_DIM(2, A);
+
+  int64_t M = A.size(0);
+  int64_t K = A.size(1);
+  int64_t lda = A.stride(0);
+
+  const auto st = A.scalar_type();
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf,
+      "per_token_quant_int8: expect A to be bfloat16 or half.");
+
+  auto Aq = at::empty({M, K}, A.options().dtype(at::kByte));
+  auto As = at::empty({M}, A.options().dtype(at::kFloat));
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(st, "per_token_quant_int8", [&] {
+    uint8_t* __restrict__ Aq_data = Aq.data_ptr<uint8_t>();
+    float* __restrict__ As_data = As.data_ptr<float>();
+    const scalar_t* __restrict__ A_data = A.data_ptr<scalar_t>();
+
+    at::parallel_for(0, M, 0, [&] (int64_t begin, int64_t end) {
+      for (int64_t m = begin; m < end; ++m) {
+        quantize_row_int8<scalar_t>(
+            Aq_data + m * K,
+            As_data[m],
+            A_data + m * lda,
+            K);
+      }
+    });
+  });
+  return std::make_tuple(Aq, As);
+}
+
+// weight     :  static, per-channel, symmetric
+// activation : dynamic,   per-token, symmetric
+//
+// mat1    : [M, K]
+// mat2    : [N, K]
+// scales1 : [M]
+// scales2 : [N]
+// bias    : [N]
+// out     : [M, N]
+//
+at::Tensor int8_scaled_mm_cpu(at::Tensor& mat1, at::Tensor& mat2,
+    at::Tensor& scales1, at::Tensor& scales2,
+    std::optional<at::Tensor>& bias, at::ScalarType out_dtype, bool is_vnni) {
+  RECORD_FUNCTION("sgl-kernel::int8_scaled_mm_cpu", std::vector<c10::IValue>({mat1, mat2, scales1, scales2, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_INPUT(mat1);
+  CHECK_INPUT(mat2);
+  CHECK_INPUT(scales1);
+  CHECK_INPUT(scales2);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat1.size(1);
+
+  // see [NOTE]: s8s8 igemm compensation in avx512-vnni
+  CHECK_EQ(mat2.size(1), (int64_t)(is_vnni ? K + sizeof(int32_t) : K));
+  CHECK_EQ(scales1.numel(), M);
+  CHECK_EQ(scales2.numel(), N);
+
+  TORCH_CHECK(mat1.scalar_type() == at::kByte, "int8_scaled_mm: expect mat1 to be uint8.");
+  TORCH_CHECK(mat2.scalar_type() == at::kChar, "int8_scaled_mm: expect mat2 to be int8.");
+  TORCH_CHECK(scales1.scalar_type() == at::kFloat && scales2.scalar_type() == at::kFloat,
+      "int8_scaled_mm: expect scales to be float32.");
+
+  auto out = at::empty({M, N}, mat1.options().dtype(out_dtype));
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "int8_scaled_mm_kernel_impl", [&] {
+    int8_scaled_mm_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        mat1.data_ptr<uint8_t>(),
+        packed_w.data_ptr<int8_t>(),
+        scales1.data_ptr<float>(),
+        scales2.data_ptr<float>(),
+        bias_data,
+        M,
+        N,
+        K);
+  });
+  return out;
+}
+
+// fused `per_token_quant_int8_cpu` and `int8_scaled_mm_cpu`
+at::Tensor int8_scaled_mm_with_quant(at::Tensor& mat1, at::Tensor& mat2, at::Tensor& scales2,
+    const std::optional<at::Tensor>& bias, at::ScalarType out_dtype, bool is_vnni) {
+  RECORD_FUNCTION("sgl-kernel::int8_scaled_mm_cpu", std::vector<c10::IValue>({mat1, mat2, scales2, bias}));
+
+  auto packed_w = is_vnni ? mat2 : convert_weight_packed(mat2);
+
+  CHECK_LAST_DIM_CONTIGUOUS_INPUT(mat1);
+  CHECK_INPUT(mat2);
+  CHECK_INPUT(scales2);
+  CHECK_DIM(2, mat1);
+  CHECK_DIM(2, mat2);
+
+  int64_t M = mat1.size(0);
+  int64_t N = mat2.size(0);
+  int64_t K = mat1.size(1);
+  int64_t lda = mat1.stride(0);
+
+  // see [NOTE]: s8s8 igemm compensation in avx512-vnni
+  CHECK_EQ(mat2.size(1), (int64_t)(is_vnni ? K + sizeof(int32_t) : K));
+  CHECK_EQ(scales2.numel(), N);
+
+  const auto st = mat1.scalar_type();
+  TORCH_CHECK(st == at::kBFloat16 || st == at::kHalf,
+      "int8_scaled_mm_with_quant: expect A to be bfloat16 or half.");
+  TORCH_CHECK(st == out_dtype,
+      "int8_scaled_mm_with_quant: expect A has same dtype with out_dtype.");
+  TORCH_CHECK(mat2.scalar_type() == at::kChar,
+      "int8_scaled_mm_with_quant: expect mat2 to be int8.");
+  TORCH_CHECK(scales2.scalar_type() == at::kFloat,
+      "int8_scaled_mm_with_quant: expect scales to be float32.");
+
+  const int64_t buffer_size = M * K + M * sizeof(float);
+  auto buffer = at::empty({buffer_size}, mat1.options().dtype(at::kByte));
+  auto out = at::empty({M, N}, mat1.options().dtype(out_dtype));
+
+  const bool has_bias = bias.has_value();
+  const float* bias_data = nullptr;
+  if (has_bias) {
+    CHECK_EQ(bias.value().size(0), N);
+    bias_data = bias.value().data_ptr<float>();
+  }
+
+  AT_DISPATCH_REDUCED_FLOATING_TYPES(out_dtype, "int8_scaled_mm_with_quant_kernel_impl", [&] {
+    uint8_t* __restrict__ Aq_data = buffer.data_ptr<uint8_t>();
+    float* __restrict__ As_data = (float*)((void*)(Aq_data + M * K));
+    const scalar_t* __restrict__ A_data = mat1.data_ptr<scalar_t>();
+
+    at::parallel_for(0, M, 0, [&] (int64_t begin, int64_t end) {
+      for (int64_t m = begin; m < end; ++m) {
+        quantize_row_int8<scalar_t>(
+            Aq_data + m * K,
+            As_data[m],
+            A_data + m * lda,
+            K);
+      }
+    });
+
+    int8_scaled_mm_kernel_impl<scalar_t>(
+        out.data_ptr<scalar_t>(),
+        Aq_data,
+        packed_w.data_ptr<int8_t>(),
+        As_data,
+        scales2.data_ptr<float>(),
+        bias_data,
+        M,
+        N,
+        K);
+  });
+  return out;
+}

csrc/cpu/sgl-kernels/moe_fp8.cpp

@@ -0,0 +1,502 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "gemm.h"
+#include "vec.h"
+
+// clang-format off
+
+namespace {
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
+  using Vec = at::vec::Vectorized<scalar_t>;
+  // no remainder
+  #pragma GCC unroll 4
+  for (int64_t d = 0; d < size; d += Vec::size()) {
+    Vec data = Vec::loadu(input + d);
+    data.store(out + d);
+  }
+}
+
+template <typename scalar_t>
+inline void copy_mul_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, float weight, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec weight_vec = fVec(weight);
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    bVec x = bVec::loadu(input + d);
+    fVec x0, x1;
+    std::tie(x0, x1) = at::vec::convert_to_float(x);
+    x0 = x0 * weight_vec;
+    x1 = x1 * weight_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] * weight);
+  }
+}
+
+// acc from [topk, K] to [K]
+template <typename scalar_t>
+inline void sum_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t topk, int64_t K) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  if (topk == 1) {
+    // do copy for topk = 1
+    copy_stub(out, input, K);
+  } else {
+    // do sum for topk != 1
+    int64_t d;
+    #pragma GCC unroll 4
+    for (d = 0; d <= K - kVecSize; d += kVecSize) {
+      fVec sum_fvec0 = fVec(0.f);
+      fVec sum_fvec1 = fVec(0.f);
+      for (int t = 0; t < topk; ++t) {
+        bVec x_bvec = bVec::loadu(input + t * K + d);
+        fVec x_fvec0, x_fvec1;
+        std::tie(x_fvec0, x_fvec1) = at::vec::convert_to_float(x_bvec);
+
+        sum_fvec0 += x_fvec0;
+        sum_fvec1 += x_fvec1;
+      }
+      bVec out_bvec = convert_from_float_ext<scalar_t>(sum_fvec0, sum_fvec1);
+      out_bvec.store(out + d);
+    }
+    for (; d < K; ++d) {
+      float sum_val = 0.f;
+      for (int t = 0; t < topk; ++t) {
+        sum_val += static_cast<float>(input[t * K + d]);
+      }
+      out[d] = static_cast<scalar_t>(sum_val);
+    }
+  }
+}
+
+// out = input + input2 * scale
+template <typename scalar_t>
+inline void add_mul_stub(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ input,
+    const scalar_t* __restrict__ input2,
+    float scale,
+    int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec s_vec = fVec(scale);
+
+  int64_t d;
+#pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    bVec x_bvec = bVec::loadu(input + d);
+    fVec x0, x1;
+    std::tie(x0, x1) = at::vec::convert_to_float(x_bvec);
+
+    bVec y_bvec = bVec::loadu(input2 + d);
+    fVec y0, y1;
+    std::tie(y0, y1) = at::vec::convert_to_float(y_bvec);
+
+    x0 = x0 + y0 * s_vec;
+    x1 = x1 + y1 * s_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + float(input2[d]) * scale);
+  }
+}
+
+template <typename scalar_t>
+inline void silu_and_mul_stub(
+    scalar_t* __restrict__ out,
+    const scalar_t* __restrict__ input,
+    const scalar_t* __restrict__ input2,
+    int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  const fVec one = fVec(1.f);
+
+  // no remainder
+#pragma GCC unroll 4
+  for (int64_t d = 0; d < size; d += bVec::size()) {
+    bVec x = bVec::loadu(input + d);
+    fVec x0, x1;
+    std::tie(x0, x1) = at::vec::convert_to_float(x);
+    bVec y = bVec::loadu(input2 + d);
+    fVec y0, y1;
+    std::tie(y0, y1) = at::vec::convert_to_float(y);
+    x0 = x0 / (one + x0.neg().exp_u20());
+    x1 = x1 / (one + x1.neg().exp_u20());
+    x0 = x0 * y0;
+    x1 = x1 * y1;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+}
+
+} // anonymous namespace
+
+template <typename scalar_t>
+void fused_experts_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    scalar_t* __restrict__ A_tmp,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 1: intermediate_cache0 = hidden_states @ w1
+  const int64_t MB = div_up(num_tokens_post_pad, BLOCK_M);
+  const int64_t NB = div_up(2 * N, BLOCK_N);
+  int64_t scale_size_N = div_up(2 * N, block_size_N);
+  int64_t scale_size_K = div_up(K, block_size_K);
+  int64_t blocks_n_per_group = block_size_N / BLOCK_N;
+
+  const int64_t stride_e = 2 * N * K;
+  const int64_t stride_n = K;
+
+  // here we only parallel on half of 2N to fuse silu_and_mul with gemm
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    scalar_t* __restrict__ A = A_tmp + tid * BLOCK_M * K;
+
+    bool is_brgemm_used = false;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+
+      int64_t n_size = std::min(2 * N - nb * BLOCK_N, BLOCK_N);
+
+      // B shape [K, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const at::Float8_e4m3fn* __restrict__ B = packed_w1 + expert_id * stride_e + nb * BLOCK_N * stride_n;
+      const float* __restrict__ Bs = w1s + expert_id * scale_size_N * scale_size_K + (nb / blocks_n_per_group) * scale_size_K;
+
+      // 1.a load A
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+
+      const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(m_size);
+      is_brgemm_used = is_brgemm_used || use_brgemm;
+
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m] / topk;
+        copy_stub(A + m * K, input + index * K, K);
+      }
+
+      const int64_t offset = offsets[mb];
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ A,
+          /*   B            */ B,
+          /*   C            */ ic0 + offset * 2 * N + nb * BLOCK_N,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ Bs,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ K,
+          /*   lda          */ K,
+          /*   ldb          */ n_size,
+          /*   ldc          */ 2 * N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+    }
+
+    if (is_brgemm_used) {
+      at::native::cpublas::brgemm_release();
+    }
+  });
+
+  // stage 1.5: intermediate_cache1 = silu(intermediate_cache0)
+  at::parallel_for(0, M * topk, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      silu_and_mul_stub(
+          ic1 + m * N,
+          ic0 + m * 2 * N,
+          ic0 + m * 2 * N + N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [E, K, N] as [E, OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(OC, BLOCK_N);
+  scale_size_N = div_up(K, block_size_N);
+  scale_size_K = div_up(N, block_size_K);
+  const int64_t stride_e2 = OC * IC;
+  const int64_t stride_oc = IC;
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    int tid = at::get_thread_num();
+    alignas(64) scalar_t C[BLOCK_M * BLOCK_K];
+
+    bool is_brgemm_used = false;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(m_size);
+      is_brgemm_used = is_brgemm_used || use_brgemm;
+
+      // A ptr from ic1 of [M * topk, N] in sorted order
+      // so as to avoid copy A to tmp buffer again
+      const scalar_t* __restrict__ A = ic1 + offsets[mb] * N;
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+
+      // B shape [IC, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const at::Float8_e4m3fn* __restrict__ B = packed_w2 + expert_id * stride_e2 + nb * BLOCK_N * stride_oc;
+      const float* __restrict__ Bs = w2s + expert_id * scale_size_N * scale_size_K + (nb / blocks_n_per_group) * scale_size_K;
+
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ A,
+          /*   B            */ B,
+          /*   C            */ C,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ Bs,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ IC,
+          /*   lda          */ IC,
+          /*   ldb          */ n_size,
+          /*   ldc          */ BLOCK_N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+
+      // 2.b copy from C to ic2 in original order
+      //   and also mul topk_weights in float32
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m];
+        float weight = topk_weights[index];
+        copy_mul_stub(ic2 + index * K + nb * BLOCK_N, C + m * BLOCK_N, weight, n_size);
+      }
+    }
+
+    if (is_brgemm_used) {
+      at::native::cpublas::brgemm_release();
+    }
+  });
+
+  // stage 3: out = intermediate_cache2.sum(dim=1)
+  //   from [M, topk, K] to [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      sum_stub(output + m * K, ic2 + m * topk * K, topk, K);
+    }
+  });
+}
+
+#define INSTANTIATE_MOE_FP8_TEMPLATE(TYPE)             \
+  template void fused_experts_fp8_kernel_impl<TYPE>(   \
+      TYPE* __restrict__ output,                       \
+      TYPE* __restrict__ ic0,                          \
+      TYPE* __restrict__ ic1,                          \
+      TYPE* __restrict__ ic2,                          \
+      TYPE* __restrict__ A_tmp,                        \
+      TYPE* __restrict__ B_tmp,                        \
+      float* __restrict__ C_tmp,                       \
+      const TYPE* __restrict__ input,                  \
+      const at::Float8_e4m3fn* __restrict__ packed_w1, \
+      const at::Float8_e4m3fn* __restrict__ packed_w2, \
+      const float* __restrict__ w1s,                   \
+      const float* __restrict__ w2s,                   \
+      int64_t block_size_N,                            \
+      int64_t block_size_K,                            \
+      const float* __restrict__ topk_weights,          \
+      const int32_t* __restrict__ sorted_ids,          \
+      const int32_t* __restrict__ expert_ids,          \
+      const int32_t* __restrict__ offsets,             \
+      int64_t M,                                       \
+      int64_t N,                                       \
+      int64_t K,                                       \
+      int64_t E,                                       \
+      int64_t topk,                                    \
+      int64_t num_tokens_post_pad)
+
+INSTANTIATE_MOE_FP8_TEMPLATE(at::BFloat16);
+INSTANTIATE_MOE_FP8_TEMPLATE(at::Half);
+
+template <typename scalar_t>
+void shared_expert_fp8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic0,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ B_tmp,
+    float* __restrict__ C_tmp,
+    const scalar_t* __restrict__ input,
+    const at::Float8_e4m3fn* __restrict__ packed_w1,
+    const at::Float8_e4m3fn* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    int64_t block_size_N,
+    int64_t block_size_K,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K) {
+
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 1: intermediate_cache0 = hidden_states @ w1
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(2 * N, BLOCK_N);
+  int64_t scale_size_K = div_up(K, block_size_K);
+  int64_t blocks_n_per_group = block_size_N / BLOCK_N;
+
+  const bool use_brgemm = can_use_brgemm<at::Float8_e4m3fn>(M);
+
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    int tid = at::get_thread_num();
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+      int64_t n_size = std::min(2 * N - nb * BLOCK_N, BLOCK_N);
+
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ input + mb * BLOCK_M * K,
+          /*   B            */ packed_w1 + nb * BLOCK_N * K,
+          /*   C            */ ic0 + mb * BLOCK_M * 2 * N + nb * BLOCK_N,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ w1s + (nb / blocks_n_per_group) * scale_size_K,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ K,
+          /*   lda          */ K,
+          /*   ldb          */ n_size,
+          /*   ldc          */ 2 * N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+    }
+
+    if (use_brgemm) {
+      at::native::cpublas::brgemm_release();
+    }
+  });
+
+  // stage 1.5: intermediate_cache1 = silu(intermediate_cache0)
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      silu_and_mul_stub(
+          ic1 + m * N,
+          ic0 + m * 2 * N,
+          ic0 + m * 2 * N + N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [K, N] as [OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(K, BLOCK_N);
+  scale_size_K = div_up(N, block_size_K);
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    int tid = at::get_thread_num();
+    alignas(64) scalar_t C[BLOCK_M * BLOCK_K];
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      // 2.a gemm: C = A @ B
+      tinygemm_kernel<scalar_t>(
+          /*   A            */ ic1 + mb * BLOCK_M * N,
+          /*   B            */ packed_w2 + nb * BLOCK_N * N,
+          /*   C            */ C,
+          /*   Btmp         */ B_tmp + tid * BLOCK_N * std::max(K, N),
+          /*   Ctmp         */ C_tmp + tid * 2 * BLOCK_M * BLOCK_N,
+          /*   scale        */ w2s + (nb / blocks_n_per_group) * scale_size_K,
+          /*   M            */ m_size,
+          /*   N            */ n_size,
+          /*   K            */ IC,
+          /*   lda          */ IC,
+          /*   ldb          */ n_size,
+          /*   ldc          */ BLOCK_N,
+          /*   brg          */ use_brgemm,
+          /*   block_size_K */ block_size_K);
+
+      // 2.b copy from C to output and add fused_experts_out
+      scalar_t* __restrict__ out = output + mb * BLOCK_M * K + nb * BLOCK_N;
+      const scalar_t* __restrict__ fused_out = fused_experts_out + mb * BLOCK_M * K + nb * BLOCK_N;
+      for (int64_t m = 0; m < m_size; ++m) {
+        add_mul_stub(out + m * K, C + m * BLOCK_N, fused_out + m * K, routed_scaling_factor, n_size);
+      }
+    }
+  });
+
+  if (use_brgemm) {
+    at::native::cpublas::brgemm_release();
+  }
+}
+
+#define INSTANTIATE_SHARED_EXPERT_FP8_TEMPLATE(TYPE)   \
+  template void shared_expert_fp8_kernel_impl<TYPE>(   \
+      TYPE* __restrict__ output,                       \
+      TYPE* __restrict__ ic0,                          \
+      TYPE* __restrict__ ic1,                          \
+      TYPE* __restrict__ B_tmp,                        \
+      float* __restrict__ C_tmp,                       \
+      const TYPE* __restrict__ input,                  \
+      const at::Float8_e4m3fn* __restrict__ packed_w1, \
+      const at::Float8_e4m3fn* __restrict__ packed_w2, \
+      const float* __restrict__ w1s,                   \
+      const float* __restrict__ w2s,                   \
+      int64_t block_size_N,                            \
+      int64_t block_size_K,                            \
+      const TYPE* __restrict__ fused_experts_out,      \
+      float routed_scaling_factor,                     \
+      int64_t M,                                       \
+      int64_t N,                                       \
+      int64_t K)
+
+INSTANTIATE_SHARED_EXPERT_FP8_TEMPLATE(at::BFloat16);
+INSTANTIATE_SHARED_EXPERT_FP8_TEMPLATE(at::Half);

csrc/cpu/sgl-kernels/moe_int8.cpp

@@ -0,0 +1,769 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#include "common.h"
+#include "vec.h"
+#include "gemm.h"
+
+// clang-format off
+
+namespace {
+
+template <typename scalar_t>
+inline void copy_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t size) {
+  using Vec = at::vec::Vectorized<scalar_t>;
+  // no remainder
+  #pragma GCC unroll 4
+  for (int64_t d = 0; d < size; d += Vec::size()) {
+    Vec data = Vec::loadu(input + d);
+    data.store(out + d);
+  }
+}
+
+template <>
+inline void copy_stub<uint8_t>(uint8_t* __restrict__ out, const uint8_t* __restrict__ input, int64_t size) {
+  // size might be 64x + 32
+  std::memcpy(out, input, size * sizeof(uint8_t));
+}
+
+template <typename scalar_t>
+inline void copy_mul_stub(scalar_t* __restrict__ out, const float* __restrict__ input, float weight, int64_t size) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec weight_vec = fVec(weight);
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec data0 = fVec::loadu(input + d) * weight_vec;
+    fVec data1 = fVec::loadu(input + d + fVec::size()) * weight_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(data0, data1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] * weight);
+  }
+}
+
+// acc from [topk, K] to [K]
+template <typename scalar_t>
+inline void sum_stub(scalar_t* __restrict__ out, const scalar_t* __restrict__ input, int64_t topk, int64_t K) {
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  if (topk == 1) {
+    // do copy for topk = 1
+    copy_stub(out, input, K);
+  } else {
+    // do sum for topk != 1
+    int64_t d;
+    #pragma GCC unroll 4
+    for (d = 0; d <= K - kVecSize; d += kVecSize) {
+      fVec sum_fvec0 = fVec(0.f);
+      fVec sum_fvec1 = fVec(0.f);
+      for (int t = 0; t < topk; ++t) {
+        bVec x_bvec = bVec::loadu(input + t * K + d);
+        fVec x_fvec0, x_fvec1;
+        std::tie(x_fvec0, x_fvec1) = at::vec::convert_to_float(x_bvec);
+
+        sum_fvec0 += x_fvec0;
+        sum_fvec1 += x_fvec1;
+      }
+      bVec out_bvec = convert_from_float_ext<scalar_t>(sum_fvec0, sum_fvec1);
+      out_bvec.store(out + d);
+    }
+    for (; d < K; ++d) {
+      float sum_val = 0.f;
+      for (int t = 0; t < topk; ++t) {
+        sum_val += static_cast<float>(input[t * K + d]);
+      }
+      out[d] = static_cast<scalar_t>(sum_val);
+    }
+  }
+}
+
+// out = input + input2 * scale
+template <typename scalar_t>
+inline void add_mul_stub(scalar_t* __restrict__ out, const float* __restrict__ input,
+    const scalar_t* __restrict__ input2, float scale, int64_t size) {
+
+  using bVec = at::vec::Vectorized<scalar_t>;
+  using fVec = at::vec::Vectorized<float>;
+  constexpr int kVecSize = bVec::size();
+  const fVec s_vec = fVec(scale);
+  int64_t d;
+  #pragma GCC unroll 4
+  for (d = 0; d <= size - kVecSize; d += kVecSize) {
+    fVec x0 = fVec::loadu(input + d);
+    fVec x1 = fVec::loadu(input + d + fVec::size());
+
+    bVec y_bvec = bVec::loadu(input2 + d);
+    fVec y0, y1;
+    std::tie(y0, y1) = at::vec::convert_to_float(y_bvec);
+
+    x0 = x0 + y0 * s_vec;
+    x1 = x1 + y1 * s_vec;
+    bVec out_vec = convert_from_float_ext<scalar_t>(x0, x1);
+    out_vec.store(out + d);
+  }
+  for (; d < size; ++d) {
+    out[d] = static_cast<scalar_t>(input[d] + float(input2[d]) * scale);
+  }
+}
+
+/// gemm for w13
+template <typename scalar_t, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B0, const int8_t* __restrict__ B1, scalar_t* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs0, const float* __restrict__ Bs1,
+      const int32_t* __restrict__ Bcomp0, const int32_t* __restrict__ Bcomp1,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni<at::BFloat16, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B0, const int8_t* __restrict__ B1, at::BFloat16* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs0, const float* __restrict__ Bs1,
+      const int32_t* __restrict__ Bcomp0, const int32_t* __restrict__ Bcomp1,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+    static_assert(COLS % 2 == 0);
+
+    __m512i va;
+    __m512i vb0[COLS];
+    __m512i vb1[COLS];
+    __m512i vc0[ROWS * COLS];
+    __m512i vc1[ROWS * COLS];
+    __m512i vcomp0[COLS];
+    __m512i vcomp1[COLS];
+    __m512  was;
+    __m512  vbs0[COLS];
+    __m512  vbs1[COLS];
+
+    auto loadc = [&](auto i) {
+      vc0[i] = _mm512_set1_epi32(0);
+      vc1[i] = _mm512_set1_epi32(0);
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K4 = K >> 2;
+    const int64_t lda4 = lda >> 2;
+    const int64_t ldb4 = ldb; // ldb * 4 >> 2;
+    const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A);
+    const int32_t* b0_ptr = reinterpret_cast<const int32_t*>(B0);
+    const int32_t* b1_ptr = reinterpret_cast<const int32_t*>(B1);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = _mm512_set1_epi32(a_ptr[row * lda4 + k]);
+      }
+      if constexpr (row == 0) {
+        vb0[col] = _mm512_loadu_si512(b0_ptr + k * ldb4 + col * 16);
+        vb1[col] = _mm512_loadu_si512(b1_ptr + k * ldb4 + col * 16);
+      }
+      vc0[i] = _mm512_dpbusd_epi32(vc0[i], va, vb0[col]);
+      vc1[i] = _mm512_dpbusd_epi32(vc1[i], va, vb1[col]);
+    };
+    for (int64_t k = 0; k < K4; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto scalec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      // load a scale
+      if constexpr(col == 0) {
+        was = _mm512_set1_ps(As[row]);
+      }
+      // load b scale and vcomp
+      if constexpr (row == 0) {
+        vbs0[col] = _mm512_loadu_ps(Bs0 + col * 16);
+        vbs1[col] = _mm512_loadu_ps(Bs1 + col * 16);
+        vcomp0[col] = _mm512_loadu_si512(Bcomp0 + col * 16);
+        vcomp1[col] = _mm512_loadu_si512(Bcomp1 + col * 16);
+      }
+      __m512 c0 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc0[i], vcomp0[col]));
+      __m512 c1 = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc1[i], vcomp1[col]));
+      vc0[i] = _mm512_castps_si512(_mm512_mul_ps(_mm512_mul_ps(c0, was), vbs0[col]));
+      vc1[i] = _mm512_castps_si512(_mm512_mul_ps(_mm512_mul_ps(c1, was), vbs1[col]));
+    };
+    Unroll<ROWS * COLS>{}(scalec);
+
+    using Vec = at::vec::Vectorized<float>;
+    const Vec one = Vec(1.f);
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+      // for COLS = 2, 4 use 512bit store
+      if constexpr (col % 2 == 0) {
+        Vec x0 = _mm512_castsi512_ps(vc0[row * COLS + col + 0]);
+        Vec x1 = _mm512_castsi512_ps(vc0[row * COLS + col + 1]);
+        Vec y0 = _mm512_castsi512_ps(vc1[row * COLS + col + 0]);
+        Vec y1 = _mm512_castsi512_ps(vc1[row * COLS + col + 1]);
+        // silu
+        x0 = x0 / (one + x0.neg().exp_u20());
+        x1 = x1 / (one + x1.neg().exp_u20());
+        // mul
+        x0 = x0 * y0;
+        x1 = x1 * y1;
+
+        _mm512_storeu_si512(
+            reinterpret_cast<__m512i*>((C + row * ldc + col * 16)),
+            (__m512i)(_mm512_cvtne2ps_pbh(__m512(x1), __m512(x0))));
+        }
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_VNNI(MB_SIZE, NB_SIZE)                        \
+    tinygemm_kernel_vnni<scalar_t, MB_SIZE, NB_SIZE>::apply(                 \
+        A + mb_start * lda, B0 + nb_start * 4, B1 + nb_start * 4,            \
+        C + mb_start * ldc + nb_start, As + mb_start,                        \
+        Bs0 + nb_start, Bs1 + nb_start, Bcomp0 + nb_start, Bcomp1 + nb_start,\
+        K, lda, ldb, ldc);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B0,
+    const int8_t* __restrict__ B1,
+    scalar_t* __restrict__ C,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs0,
+    const float* __restrict__ Bs1,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc) {
+
+  const int32_t* Bcomp0 = reinterpret_cast<const int32_t*>(B0 + block_size_n() * K);
+  const int32_t* Bcomp1 = reinterpret_cast<const int32_t*>(B1 + block_size_n() * K);
+
+  // pattern: 1-(2+2)-(8+8)
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 32;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_VNNI(1, 32); break;
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_VNNI(2, 32); break;
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_VNNI(3, 32); break;
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_VNNI(4, 32); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+/// gemm for w2
+template <typename scalar_t, int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni2 {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, float* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+    TORCH_CHECK(false, "tinygemm_kernel_nn: scalar path not implemented!");
+  }
+};
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <int BLOCK_M, int BLOCK_N>
+struct tinygemm_kernel_vnni2<at::BFloat16, BLOCK_M, BLOCK_N> {
+  static inline void apply(
+      const uint8_t* __restrict__ A, const int8_t* __restrict__ B, float* __restrict__ C,
+      const float* __restrict__ As, const float* __restrict__ Bs, const int32_t* __restrict__ Bcomp,
+      int64_t K, int64_t lda, int64_t ldb, int64_t ldc) {
+
+    constexpr int ROWS = BLOCK_M;
+    constexpr int COLS = BLOCK_N / 16;
+    static_assert(COLS % 2 == 0);
+
+    __m512i va;
+    __m512i vb[COLS];
+    __m512i vc[ROWS * COLS];
+    __m512i vcomp[COLS];
+    __m512  was;
+    __m512  vbs[COLS];
+
+    auto loadc = [&](auto i) {
+      vc[i] = _mm512_set1_epi32(0);
+    };
+    Unroll<ROWS * COLS>{}(loadc);
+
+    const int64_t K4 = K >> 2;
+    const int64_t lda4 = lda >> 2;
+    const int64_t ldb4 = ldb; // ldb * 4 >> 2;
+    const int32_t* a_ptr = reinterpret_cast<const int32_t*>(A);
+    const int32_t* b_ptr = reinterpret_cast<const int32_t*>(B);
+
+    auto compute = [&](auto i, int64_t k) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      if constexpr (col == 0) {
+        va = _mm512_set1_epi32(a_ptr[row * lda4 + k]);
+      }
+      if constexpr (row == 0) {
+        vb[col] = _mm512_loadu_si512(b_ptr + k * ldb4 + col * 16);
+      }
+      vc[i] = _mm512_dpbusd_epi32(vc[i], va, vb[col]);
+    };
+    for (int64_t k = 0; k < K4; ++k) {
+      Unroll<ROWS * COLS>{}(compute, k);
+    }
+
+    auto storec = [&](auto i) {
+      constexpr int row = i / COLS;
+      constexpr int col = i % COLS;
+
+      // load a scale
+      if constexpr(col == 0) {
+        was = _mm512_set1_ps(As[row]);
+      }
+      // load b scale and vcomp per 2 vectors
+      // also load bias if any
+      if constexpr (row == 0) {
+        if constexpr (col % 2 == 0) {
+          vbs[col + 0] = _mm512_loadu_ps(Bs + col * 16);
+          vbs[col + 1] = _mm512_loadu_ps(Bs + col * 16 + 16);
+          vcomp[col + 0] = _mm512_loadu_si512(Bcomp + col * 16);
+          vcomp[col + 1] = _mm512_loadu_si512(Bcomp + col * 16 + 16);
+        }
+      }
+      __m512 x = _mm512_cvtepi32_ps(_mm512_sub_epi32(vc[i], vcomp[col]));
+      x = _mm512_mul_ps(_mm512_mul_ps(x, was), vbs[col]);
+      _mm512_storeu_ps(reinterpret_cast<__m512*>(C + row * ldc + col * 16), x);
+    };
+    Unroll<ROWS * COLS>{}(storec);
+  }
+};
+#endif
+
+#define LAUNCH_TINYGEMM_KERNEL_VNNI2(MB_SIZE, NB_SIZE)                       \
+    tinygemm_kernel_vnni2<scalar_t, MB_SIZE, NB_SIZE>::apply(                \
+        A + mb_start * lda, B + nb_start * 4, C + mb_start * ldc + nb_start, \
+        As + mb_start, Bs + nb_start, Bcomp + nb_start,                      \
+        K, lda, ldb, ldc);
+
+template <typename scalar_t>
+void tinygemm_kernel(
+    const uint8_t* __restrict__ A,
+    const int8_t* __restrict__ B,
+    float* __restrict__ C,
+    const float* __restrict__ As,
+    const float* __restrict__ Bs,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t lda,
+    int64_t ldb,
+    int64_t ldc) {
+
+  // B compensation
+  const int32_t* Bcomp = reinterpret_cast<const int32_t*>(B + block_size_n() * K);
+
+  // pattern: 1-4-16
+  constexpr int64_t BLOCK_M = 4;
+  constexpr int64_t BLOCK_N = 64;
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+  for (int64_t mb = 0; mb < MB; ++mb) {
+    int64_t mb_start = mb * BLOCK_M;
+    int64_t mb_size = std::min(BLOCK_M, M - mb_start);
+    for (int64_t nb = 0; nb < NB; ++nb) {
+      int64_t nb_start = nb * BLOCK_N;
+      int64_t nb_size = std::min(BLOCK_N, N - nb_start);
+
+      switch(mb_size << 4 | nb_size >> 4) {
+        case 0x12: LAUNCH_TINYGEMM_KERNEL_VNNI2(1, 32); break;
+        case 0x22: LAUNCH_TINYGEMM_KERNEL_VNNI2(2, 32); break;
+        case 0x32: LAUNCH_TINYGEMM_KERNEL_VNNI2(3, 32); break;
+        case 0x42: LAUNCH_TINYGEMM_KERNEL_VNNI2(4, 32); break;
+        default: TORCH_CHECK(false, "Unexpected block size, ", mb_size, "x", "nb_size");
+      }
+    }
+  }
+}
+
+} // anonymous namespace
+
+template <typename scalar_t>
+void fused_experts_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    scalar_t* __restrict__ ic2,
+    uint8_t* __restrict__ A_tmp,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const float* __restrict__ topk_weights,
+    const int32_t* __restrict__ sorted_ids,
+    const int32_t* __restrict__ expert_ids,
+    const int32_t* __restrict__ offsets,
+    int64_t M,
+    int64_t N,
+    int64_t K,
+    int64_t E,
+    int64_t topk,
+    int64_t num_tokens_post_pad) {
+
+  // handle 2 tiles per block
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 0: quantize input to uint8, [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * K,
+          As_tmp[m],
+          input + m * K,
+          K);
+    }
+  });
+
+  // stage 1: intermediate_cache1 = silu(hidden_states @ w1)
+  const int64_t MB = div_up(num_tokens_post_pad, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  // strides for w1: [E, 2N, K]
+  TORCH_CHECK(N % BLOCK_N == 0, "Fixme when N is not multiples of ", BLOCK_N);
+
+  // K and N are packed for int8
+  const int64_t packed_K = get_row_size<int8_t>(K);
+  const int64_t packed_N = get_row_size<int8_t>(N);
+
+  const int64_t stride_e = 2 * N * packed_K;
+  const int64_t stride_n = packed_K;
+  // here we only parallel on half of 2N to fuse silu_and_mul with gemm
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    uint8_t* __restrict__ A = A_tmp + tid * BLOCK_M * K;
+
+    alignas(64) float As[BLOCK_M];
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+
+      // nb0 from top half and nb1 from bottom half
+      int64_t nb0 = nb, nb1 = nb + NB;
+      int64_t n_size = std::min(N - nb0 * BLOCK_N, BLOCK_N);
+
+      // B shape [K, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const int8_t* __restrict__ B0 = packed_w1 + expert_id * stride_e + nb0 * BLOCK_N * stride_n;
+      const int8_t* __restrict__ B1 = packed_w1 + expert_id * stride_e + nb1 * BLOCK_N * stride_n;
+      const float* __restrict__ Bs0 = w1s + expert_id * 2 * N + nb0 * BLOCK_N;
+      const float* __restrict__ Bs1 = w1s + expert_id * 2 * N + nb1 * BLOCK_N;
+
+      // 1.a load A
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m] / topk;
+        copy_stub(A + m * K, Aq_tmp + index * K, K);
+        As[m] = As_tmp[index];
+      }
+
+      // fused 1.b: silu_and_mul(A @ B0, A @ B1)
+      const int64_t offset = offsets[mb];
+      tinygemm_kernel(
+          /* A     */ A,
+          /* B0    */ B0,
+          /* B1    */ B1,
+          /* C     */ ic1 + offset * N + nb * BLOCK_N,
+          /* As    */ As,
+          /* Bs0   */ Bs0,
+          /* Bs1   */ Bs1,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ K,
+          /* lda   */ K,
+          /* ldb   */ n_size,
+          /* ldc   */ N);
+    }
+  });
+
+  // stage 1.5: quantize ic1 to uint8, [M * topk, N]
+  at::parallel_for(0, M * topk, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * N,
+          As_tmp[m],
+          ic1 + m * N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [E, K, N] as [E, OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(OC, BLOCK_N);
+  const int64_t stride_e2 = OC * packed_N;
+  const int64_t stride_oc = packed_N;
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    // we won't be using C1 for gemm2
+    float* __restrict__ C = C_tmp + tid * 2 * BLOCK_M * BLOCK_N;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+
+      int64_t m_size = offsets[mb + 1] - offsets[mb];
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      // A ptr from ic1 of [M * topk, N] in sorted order
+      // so as to avoid copy A to tmp buffer again
+      const uint8_t* __restrict__ A = Aq_tmp + offsets[mb] * N;
+      const float* __restrict__ As = As_tmp + offsets[mb];
+      const int32_t* A_ids = sorted_ids + mb * BLOCK_M;
+
+      // B shape [IC, n_size] in vnni format
+      int32_t expert_id = expert_ids[mb];
+      const int8_t* __restrict__ B = packed_w2 + expert_id * stride_e2 + nb * BLOCK_N * stride_oc;
+      const float* __restrict__ Bs = w2s + expert_id * K + nb * BLOCK_N;
+
+      // 2.a gemm: C = A @ B
+      tinygemm_kernel<scalar_t>(
+          /* A     */ A,
+          /* B     */ B,
+          /* C     */ C,
+          /* As    */ As,
+          /* Bs    */ Bs,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ IC,
+          /* lda   */ IC,
+          /* ldb   */ n_size,
+          /* ldc   */ BLOCK_N);
+
+      // 2.b copy from C to ic2 in original order
+      //   and also mul topk_weights in float32
+      for (int64_t m = 0; m < m_size; ++m) {
+        int32_t index = A_ids[m];
+        float weight = topk_weights[index];
+        copy_mul_stub(ic2 + index * K + nb * BLOCK_N, C + m * BLOCK_N, weight, n_size);
+      }
+    }
+  });
+
+  // stage 3: out = intermediate_cache2.sum(dim=1)
+  //   from [M, topk, K] to [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      sum_stub(output + m * K, ic2 + m * topk * K, topk, K);
+    }
+  });
+}
+
+#define INSTANTIATE_MOE_INT8_TEMPLATE(TYPE)                                                  \
+  template void fused_experts_int8_kernel_impl<TYPE> (                                       \
+      TYPE* __restrict__ output, TYPE* __restrict__ ic1,                                     \
+      TYPE* __restrict__ ic2, uint8_t* __restrict__ A_tmp,                                   \
+      float* __restrict__ C_tmp, uint8_t* __restrict__ Aq_tmp,                               \
+      float* __restrict__ As_tmp, const TYPE* __restrict__ input,                            \
+      const int8_t* __restrict__ packed_w1, const int8_t* __restrict__ packed_w2,            \
+      const float* __restrict__ w1s, const float* __restrict__ w2s,                          \
+      const float* __restrict__ topk_weights, const int32_t* __restrict__ sorted_ids,        \
+      const int32_t* __restrict__ expert_ids, const int32_t* __restrict__ offsets,           \
+      int64_t M, int64_t N, int64_t K, int64_t E, int64_t topk, int64_t num_tokens_post_pad)
+
+INSTANTIATE_MOE_INT8_TEMPLATE(at::BFloat16);
+INSTANTIATE_MOE_INT8_TEMPLATE(at::Half);
+
+template <typename scalar_t>
+void shared_expert_int8_kernel_impl(
+    scalar_t* __restrict__ output,
+    scalar_t* __restrict__ ic1,
+    float* __restrict__ C_tmp,
+    uint8_t* __restrict__ Aq_tmp,
+    float* __restrict__ As_tmp,
+    const scalar_t* __restrict__ input,
+    const int8_t* __restrict__ packed_w1,
+    const int8_t* __restrict__ packed_w2,
+    const float* __restrict__ w1s,
+    const float* __restrict__ w2s,
+    const scalar_t* __restrict__ fused_experts_out,
+    float routed_scaling_factor,
+    int64_t M,
+    int64_t N,
+    int64_t K) {
+
+  // handle 2 tiles per block
+  constexpr int64_t BLOCK_M = block_size_m();
+  constexpr int64_t BLOCK_N = block_size_n();
+
+  // stage 0: quantize input to uint8, [M, K]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * K,
+          As_tmp[m],
+          input + m * K,
+          K);
+    }
+  });
+
+   // stage 1: intermediate_cache1 = silu(hidden_states @ w1)
+  const int64_t MB = div_up(M, BLOCK_M);
+  const int64_t NB = div_up(N, BLOCK_N);
+
+  TORCH_CHECK(N % BLOCK_N == 0, "Fixme when N is not multiples of ", BLOCK_N);
+
+  // K and N are packed for int8
+  const int64_t packed_K = get_row_size<int8_t>(K);
+  const int64_t packed_N = get_row_size<int8_t>(N);
+  const int64_t stride_n = packed_K;
+
+  // here we only parallel on half of 2N to fuse silu_and_mul with gemm
+  at::parallel_for(0, MB * NB, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB;
+      int64_t nb = i % NB;
+
+      // nb0 from top half and nb1 from bottom half
+      int64_t nb0 = nb, nb1 = nb + NB;
+      int64_t n_size = std::min(N - nb0 * BLOCK_N, BLOCK_N);
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+
+      // A shape [m_size, K]
+      const uint8_t* A = Aq_tmp + mb * BLOCK_M * K;
+      const float* As = As_tmp + mb * BLOCK_M;
+
+      // B shape [K, n_size] in vnni format
+      const int8_t* __restrict__ B0 = packed_w1 + nb0 * BLOCK_N * stride_n;
+      const int8_t* __restrict__ B1 = packed_w1 + nb1 * BLOCK_N * stride_n;
+      const float* __restrict__ Bs0 = w1s + nb0 * BLOCK_N;
+      const float* __restrict__ Bs1 = w1s + nb1 * BLOCK_N;
+
+      // fused 1.b: silu_and_mul(A @ B0, A @ B1)
+      tinygemm_kernel(
+          /* A     */ A,
+          /* B0    */ B0,
+          /* B1    */ B1,
+          /* C     */ ic1 + mb * BLOCK_M * N + nb * BLOCK_N,
+          /* As    */ As,
+          /* Bs0   */ Bs0,
+          /* Bs1   */ Bs1,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ K,
+          /* lda   */ K,
+          /* ldb   */ n_size,
+          /* ldc   */ N);
+    }
+  });
+
+  // stage 1.5: quantize ic1 to uint8, [M * topk, N]
+  at::parallel_for(0, M, 0, [&](int64_t begin, int64_t end) {
+    for (int64_t m = begin; m < end; ++m) {
+      quantize_row_int8<scalar_t>(
+          Aq_tmp + m * N,
+          As_tmp[m],
+          ic1 + m * N,
+          N);
+    }
+  });
+
+  // stage 2: intermediate_cache2 = intermediate_cache1 @ w2
+  //   w2 : [K, N] as [OC, IC]
+  const int64_t OC = K;  // rename K as OC
+  const int64_t IC = N;  // rename N as IC
+  const int64_t MB2 = MB;
+  const int64_t NB2 = div_up(OC, BLOCK_N);
+  const int64_t stride_oc = packed_N;
+
+  // parallel on [MB2, NB2]
+  at::parallel_for(0, MB2 * NB2, 0, [&](int64_t begin, int64_t end) {
+    // get local pointers
+    int tid = at::get_thread_num();
+    // we won't be using C1 for gemm2
+    float* __restrict__ C = C_tmp + tid * 2 * BLOCK_M * BLOCK_N;
+
+    for (int64_t i = begin; i < end; ++i) {
+      int64_t mb = i / NB2;
+      int64_t nb = i % NB2;
+
+      int64_t m_size = std::min(M - mb * BLOCK_M, BLOCK_M);
+      int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
+
+      // A shape [m_size, IC]
+      const uint8_t* __restrict__ A = Aq_tmp + mb * BLOCK_M * N;
+      const float* __restrict__ As = As_tmp + mb * BLOCK_M;
+
+      // B shape [IC, n_size] in vnni format
+      const int8_t* __restrict__ B = packed_w2 + nb * BLOCK_N * stride_oc;
+      const float* __restrict__ Bs = w2s + nb * BLOCK_N;
+
+      // 2.a gemm: C = A @ B
+      tinygemm_kernel<scalar_t>(
+          /* A     */ A,
+          /* B     */ B,
+          /* C     */ C,
+          /* As    */ As,
+          /* Bs    */ Bs,
+          /* M     */ m_size,
+          /* N     */ n_size,
+          /* K     */ IC,
+          /* lda   */ IC,
+          /* ldb   */ n_size,
+          /* ldc   */ BLOCK_N);
+
+      // 2.b copy from C to output and add fused_experts_out
+      scalar_t* __restrict__ out = output + mb * BLOCK_M * K + nb * BLOCK_N;
+      const scalar_t* __restrict__ fused_out = fused_experts_out + mb * BLOCK_M * K + nb * BLOCK_N;
+      for (int64_t m = 0; m < m_size; ++m) {
+        add_mul_stub(out + m * K, C + m * BLOCK_N, fused_out + m * K, routed_scaling_factor, n_size);
+      }
+    }
+  });
+}
+
+#define INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(TYPE)                                        \
+  template void shared_expert_int8_kernel_impl<TYPE> (                                       \
+      TYPE* __restrict__ output, TYPE* __restrict__ ic1,                                     \
+      float* __restrict__ C_tmp, uint8_t* __restrict__ Aq_tmp,                               \
+      float* __restrict__ As_tmp, const TYPE* __restrict__ input,                            \
+      const int8_t* __restrict__ packed_w1, const int8_t* __restrict__ packed_w2,            \
+      const float* __restrict__ w1s, const float* __restrict__ w2s,                          \
+      const TYPE* __restrict__ fused_experts_out, float routed_scaling_factor,               \
+      int64_t M, int64_t N, int64_t K)
+
+INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(at::BFloat16);
+INSTANTIATE_SHARED_EXPERT_INT8_TEMPLATE(at::Half);

csrc/cpu/sgl-kernels/vec.h

@@ -0,0 +1,308 @@
+// Adapted from
+// https://github.com/sgl-project/sglang/tree/main/sgl-kernel/csrc/cpu
+
+#pragma once
+
+// clang-format off
+
+#if defined(__AVX512F__) && defined(__AVX512BF16__) && defined(__AMX_BF16__)
+#define CPU_CAPABILITY_AVX512
+#endif
+
+#include <ATen/cpu/vec/functional.h>
+#include <ATen/cpu/vec/vec.h>
+
+namespace {
+
+using namespace at::vec;
+
+template <typename scalar_t,
+          typename std::enable_if_t<is_reduced_floating_point_v<scalar_t>, int> = 0>
+inline Vectorized<scalar_t> convert_from_float_ext(const Vectorized<float>& a, const Vectorized<float>& b) {
+  return at::vec::convert_from_float<scalar_t>(a, b);
+}
+
+#if defined(CPU_CAPABILITY_AVX512)
+
+// `at::vec::convert_from_float<>` from PyTorch doesn't have avx512-bf16 intrinsics
+// use native instruction for bfloat16->float32 conversion
+template <>
+inline Vectorized<at::BFloat16> convert_from_float_ext<at::BFloat16>(const Vectorized<float>& a, const Vectorized<float>& b) {
+  return (__m512i)(_mm512_cvtne2ps_pbh(__m512(b), __m512(a)));
+}
+
+#define CVT_BF16_TO_FP32(a) \
+    _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(a), 16))
+
+#define CVT_FP16_TO_FP32(a) \
+    _mm512_cvtps_ph(a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC))
+
+// this doesn't hanel NaN.
+inline __m512bh cvt_e4m3_bf16_intrinsic_no_nan(__m256i fp8_vec) {
+  const __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
+
+  const __m512i mant = _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(0x07)), 4);
+  const __m512i raw_exp = _mm512_srli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(0x78)), 3);
+  const __m512i exp = _mm512_slli_epi16(_mm512_add_epi16(raw_exp, _mm512_set1_epi16(120)), 7);
+  const __m512i nonsign = _mm512_or_si512(exp, mant);
+
+  const __m512i sign = _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(0x80)), 8);
+  const __m512i combined = _mm512_or_si512(nonsign, sign);
+
+  const __mmask32 is_nonzero = _mm512_cmpneq_epi16_mask(x, _mm512_setzero_si512());
+  return (__m512bh)_mm512_maskz_mov_epi16(is_nonzero, combined);
+}
+
+inline __m512bh cvt_e4m3_bf16_intrinsic_without_denorm(__m256i fp8_vec) {
+  // The following conversion is without denorm behavior, that is to say,
+  //   Max subnorm   : S.0000.111 = 0.875 ∗ 2**(−6)
+  //   Min subnorm   : S.0000.001 = 2**(−9)
+  // 0.0019 ~ 0.0137 cannot be converted correctly.
+  __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
+  auto mask = _mm512_cmpneq_epi16_mask(
+      _mm512_and_si512(x, _mm512_set1_epi16(127)),
+      _mm512_setzero_si512());  // mask = x & 0x7f
+  auto mask_nan = _mm512_cmpneq_epi16_mask(
+      _mm512_and_si512(x, _mm512_set1_epi16(127)),
+      _mm512_set1_epi16(127));                                                      // mask_nan = x & 0x7f
+  auto mantissa = _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(7)), 4);  // mantissa = (x & 7) << 4
+  auto exponent = _mm512_add_epi16(
+      _mm512_srli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(120)), 3),
+      _mm512_set1_epi16(120));  // exponent = (((x >> 3) & 15) + 120)
+  auto nonsign = _mm512_maskz_mov_epi16(mask, _mm512_or_si512(mantissa, _mm512_slli_epi16(exponent, 7)));
+  nonsign = _mm512_mask_mov_epi16(_mm512_set1_epi16(0x7fff), mask_nan, nonsign);  // deal with Nan
+  return (__m512bh)(_mm512_or_si512(
+      nonsign,
+      _mm512_slli_epi16(
+          _mm512_and_si512(x, _mm512_set1_epi16(128)),
+          8)));  // add sign (x & 128) << 8
+}
+
+inline __m512bh cvt_e4m3_bf16_intrinsic_with_denorm(__m256i fp8_vec) {
+  __m512i x = _mm512_cvtepu8_epi16(fp8_vec);
+  __m512i lg2mant = _mm512_mask_mov_epi16(
+      _mm512_mask_mov_epi16(
+          _mm512_setzero_si512(), _mm512_test_epi16_mask(x, _mm512_set1_epi16(2)), _mm512_set1_epi16(1)),
+      _mm512_test_epi16_mask(x, _mm512_set1_epi16(4)),
+      _mm512_set1_epi16(2));
+  return (__m512bh)(_mm512_or_si512(
+      _mm512_maskz_mov_epi16(
+          _mm512_cmpneq_epi16_mask(_mm512_and_si512(x, _mm512_set1_epi16(127)), _mm512_setzero_si512()),
+          _mm512_mask_blend_epi16(
+              _mm512_test_epi16_mask(x, _mm512_set1_epi16(120)),
+              _mm512_or_si512(
+                  _mm512_and_si512(
+                      _mm512_sllv_epi16(
+                          _mm512_and_si512(x, _mm512_set1_epi16(3)), _mm512_sub_epi16(_mm512_set1_epi16(7), lg2mant)),
+                      _mm512_set1_epi16(0x007f)),
+                  _mm512_slli_epi16(_mm512_add_epi16(lg2mant, _mm512_set1_epi16(118)), 7)),
+              _mm512_or_si512(
+                  _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(7)), 4),
+                  _mm512_slli_epi16(
+                      _mm512_add_epi16(
+                          _mm512_srli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(120)), 3), _mm512_set1_epi16(120)),
+                      7)))),
+      _mm512_slli_epi16(_mm512_and_si512(x, _mm512_set1_epi16(128)), 8)));
+}
+
+inline __m512bh CVT_FP8_TO_BF16(__m256i a) {
+#ifdef SGLANG_CPU_FP8_CVT_FTZ
+  return cvt_e4m3_bf16_intrinsic_no_nan(a);
+#else
+  return cvt_e4m3_bf16_intrinsic_with_denorm(a);
+#endif
+}
+
+#endif
+
+// vector to scalar reduction
+#if defined(CPU_CAPABILITY_AVX512) && 0
+inline float vec_reduce_sum(const Vectorized<float>& a) {
+  return _mm512_reduce_add_ps(__m512(a));
+}
+
+inline float vec_reduce_max(const Vectorized<float>& a) {
+  return _mm512_reduce_max_ps(__m512(a));
+}
+#else
+inline float vec_reduce_sum(const Vectorized<float>& a) {
+  return vec_reduce_all([](Vectorized<float>& x, Vectorized<float>& y) { return x + y; }, a);
+}
+
+inline float vec_reduce_max(const Vectorized<float>& a) {
+  return vec_reduce_all([](Vectorized<float>& x, Vectorized<float>& y) { return maximum(x, y); }, a);
+}
+#endif
+
+// https://github.com/InternLM/lmdeploy/blob/086481ed84b59bee3b8e4274e5fc69620040c048/lmdeploy/pytorch/kernels/cuda/w8a8_triton_kernels.py#L282
+template <typename scalar_t>
+inline void quantize_row_int8(uint8_t* __restrict__ Aq, float& As,
+    const scalar_t* __restrict__ A, int64_t K, float eps = 1e-7) {
+
+  float amax = 0.f; // absolute max
+  for (int64_t k = 0; k < K; ++k) {
+    const float val = static_cast<float>(A[k]);
+    amax = std::max(amax, std::abs(val));
+  }
+
+  amax = std::max(amax, eps);
+  const float scale = amax / 127;
+  const float inv_scale = 127 / amax;
+
+  for (int64_t k = 0; k < K; ++k) {
+    const float val = static_cast<float>(A[k]) * inv_scale;
+    Aq[k] = (uint8_t)(std::round(val)) + 128;
+  }
+  As = scale;
+}
+
+#if defined(CPU_CAPABILITY_AVX512)
+template <>
+inline void quantize_row_int8<at::BFloat16>(uint8_t* __restrict__ Aq, float& As,
+    const at::BFloat16* __restrict__ A, int64_t K, float eps) {
+
+  const __m512 signBit = _mm512_set1_ps(-0.0f);
+  const __m512i off = _mm512_set1_epi32(128);
+
+  // K is 32x, no remainder
+  float amax = 0.f;
+  __m512 vamax0 = _mm512_set1_ps(0.f);
+  __m512 vamax1 = _mm512_set1_ps(0.f);
+  for (int64_t k = 0; k < K; k += 32) {
+    __m512i va = _mm512_loadu_si512((void*)(A + k));
+    __m512 va0 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 0));
+    __m512 va1 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 1));
+    vamax0 = _mm512_max_ps(vamax0, _mm512_andnot_ps(signBit, va0));
+    vamax1 = _mm512_max_ps(vamax1, _mm512_andnot_ps(signBit, va1));
+  }
+  amax = _mm512_reduce_max_ps(_mm512_max_ps(vamax0, vamax1));
+  amax = std::max(amax, eps);
+  const float scale = amax / 127;
+  const float inv_scale = 127 / amax;
+  const __m512 vd = _mm512_set1_ps(inv_scale);
+
+  for (int64_t k = 0; k < K; k += 32) {
+    __m512i va = _mm512_loadu_si512((void*)(A + k));
+    __m512 va0 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 0));
+    __m512 va1 = CVT_BF16_TO_FP32(_mm512_extracti32x8_epi32(va, 1));
+    va0 = _mm512_mul_ps(va0, vd);
+    va1 = _mm512_mul_ps(va1, vd);
+    va0 = _mm512_roundscale_ps(va0, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));
+    va1 = _mm512_roundscale_ps(va1, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));
+    __m128i i0 = _mm512_cvtepi32_epi8(_mm512_add_epi32(_mm512_cvtps_epi32(va0), off));
+    __m128i i1 = _mm512_cvtepi32_epi8(_mm512_add_epi32(_mm512_cvtps_epi32(va1), off));
+    _mm256_storeu_si256(reinterpret_cast<__m256i*>(Aq + k), _mm256_set_m128i(i1, i0));
+  }
+  As = scale;
+}
+#endif
+
+// transpose utils
+// taken from my PR in ggml: https://github.com/ggml-org/llama.cpp/pull/8998
+#if defined(CPU_CAPABILITY_AVX512)
+inline void transpose_16x16_32bit(__m512i * v) {
+  __m512i v1[16];
+  v1[0] = _mm512_unpacklo_epi32(v[0], v[1]);
+  v1[1] = _mm512_unpackhi_epi32(v[0], v[1]);
+  v1[2] = _mm512_unpacklo_epi32(v[2], v[3]);
+  v1[3] = _mm512_unpackhi_epi32(v[2], v[3]);
+  v1[4] = _mm512_unpacklo_epi32(v[4], v[5]);
+  v1[5] = _mm512_unpackhi_epi32(v[4], v[5]);
+  v1[6] = _mm512_unpacklo_epi32(v[6], v[7]);
+  v1[7] = _mm512_unpackhi_epi32(v[6], v[7]);
+  v1[8] = _mm512_unpacklo_epi32(v[8], v[9]);
+  v1[9] = _mm512_unpackhi_epi32(v[8], v[9]);
+  v1[10] = _mm512_unpacklo_epi32(v[10], v[11]);
+  v1[11] = _mm512_unpackhi_epi32(v[10], v[11]);
+  v1[12] = _mm512_unpacklo_epi32(v[12], v[13]);
+  v1[13] = _mm512_unpackhi_epi32(v[12], v[13]);
+  v1[14] = _mm512_unpacklo_epi32(v[14], v[15]);
+  v1[15] = _mm512_unpackhi_epi32(v[14], v[15]);
+
+  v[0] = _mm512_unpacklo_epi64(v1[0], v1[2]);
+  v[1] = _mm512_unpackhi_epi64(v1[0], v1[2]);
+  v[2] = _mm512_unpacklo_epi64(v1[1], v1[3]);
+  v[3] = _mm512_unpackhi_epi64(v1[1], v1[3]);
+  v[4] = _mm512_unpacklo_epi64(v1[4], v1[6]);
+  v[5] = _mm512_unpackhi_epi64(v1[4], v1[6]);
+  v[6] = _mm512_unpacklo_epi64(v1[5], v1[7]);
+  v[7] = _mm512_unpackhi_epi64(v1[5], v1[7]);
+  v[8] = _mm512_unpacklo_epi64(v1[8], v1[10]);
+  v[9] = _mm512_unpackhi_epi64(v1[8], v1[10]);
+  v[10] = _mm512_unpacklo_epi64(v1[9], v1[11]);
+  v[11] = _mm512_unpackhi_epi64(v1[9], v1[11]);
+  v[12] = _mm512_unpacklo_epi64(v1[12], v1[14]);
+  v[13] = _mm512_unpackhi_epi64(v1[12], v1[14]);
+  v[14] = _mm512_unpacklo_epi64(v1[13], v1[15]);
+  v[15] = _mm512_unpackhi_epi64(v1[13], v1[15]);
+
+  v1[0] = _mm512_shuffle_i32x4(v[0], v[4], 0x88);
+  v1[1] = _mm512_shuffle_i32x4(v[1], v[5], 0x88);
+  v1[2] = _mm512_shuffle_i32x4(v[2], v[6], 0x88);
+  v1[3] = _mm512_shuffle_i32x4(v[3], v[7], 0x88);
+  v1[4] = _mm512_shuffle_i32x4(v[0], v[4], 0xdd);
+  v1[5] = _mm512_shuffle_i32x4(v[1], v[5], 0xdd);
+  v1[6] = _mm512_shuffle_i32x4(v[2], v[6], 0xdd);
+  v1[7] = _mm512_shuffle_i32x4(v[3], v[7], 0xdd);
+  v1[8] = _mm512_shuffle_i32x4(v[8], v[12], 0x88);
+  v1[9] = _mm512_shuffle_i32x4(v[9], v[13], 0x88);
+  v1[10] = _mm512_shuffle_i32x4(v[10], v[14], 0x88);
+  v1[11] = _mm512_shuffle_i32x4(v[11], v[15], 0x88);
+  v1[12] = _mm512_shuffle_i32x4(v[8], v[12], 0xdd);
+  v1[13] = _mm512_shuffle_i32x4(v[9], v[13], 0xdd);
+  v1[14] = _mm512_shuffle_i32x4(v[10], v[14], 0xdd);
+  v1[15] = _mm512_shuffle_i32x4(v[11], v[15], 0xdd);
+
+  v[0] = _mm512_shuffle_i32x4(v1[0], v1[8], 0x88);
+  v[1] = _mm512_shuffle_i32x4(v1[1], v1[9], 0x88);
+  v[2] = _mm512_shuffle_i32x4(v1[2], v1[10], 0x88);
+  v[3] = _mm512_shuffle_i32x4(v1[3], v1[11], 0x88);
+  v[4] = _mm512_shuffle_i32x4(v1[4], v1[12], 0x88);
+  v[5] = _mm512_shuffle_i32x4(v1[5], v1[13], 0x88);
+  v[6] = _mm512_shuffle_i32x4(v1[6], v1[14], 0x88);
+  v[7] = _mm512_shuffle_i32x4(v1[7], v1[15], 0x88);
+  v[8] = _mm512_shuffle_i32x4(v1[0], v1[8], 0xdd);
+  v[9] = _mm512_shuffle_i32x4(v1[1], v1[9], 0xdd);
+  v[10] = _mm512_shuffle_i32x4(v1[2], v1[10], 0xdd);
+  v[11] = _mm512_shuffle_i32x4(v1[3], v1[11], 0xdd);
+  v[12] = _mm512_shuffle_i32x4(v1[4], v1[12], 0xdd);
+  v[13] = _mm512_shuffle_i32x4(v1[5], v1[13], 0xdd);
+  v[14] = _mm512_shuffle_i32x4(v1[6], v1[14], 0xdd);
+  v[15] = _mm512_shuffle_i32x4(v1[7], v1[15], 0xdd);
+}
+
+// remove warning : ignoring attributes on template argument ‘__m512i’ [-Wignored-attributes]
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wignored-attributes"
+
+// transpose from [2, 32] to [32, 2]
+inline std::tuple<__m512i, __m512i> transpose_2x32_16bit(__m512i r0, __m512i r1) {
+  // r0: {a0, a1, ..., a31}
+  // r1: {b0, b1, ..., b31}
+  //
+  // d0: {a0,   b0, ..., a15, b15}
+  // d1: {a16, b16, ..., a31, b31}
+  //
+  __m512i d0 = _mm512_unpacklo_epi16(r0, r1);
+  __m512i d1 = _mm512_unpackhi_epi16(r0, r1);
+  r0 = _mm512_shuffle_i32x4(d0, d1, 0x88);
+  r1 = _mm512_shuffle_i32x4(d0, d1, 0xdd);
+  d0 = _mm512_shuffle_i32x4(r0, r1, 0x88);
+  d1 = _mm512_shuffle_i32x4(r0, r1, 0xdd);
+  return std::make_tuple(d0, d1);
+}
+#pragma GCC diagnostic pop
+
+#endif
+
+// TODO: debug print, remove me later
+template<typename scalar_t>
+void print_array(scalar_t* ptr, int size) {
+  for (int d = 0; d < size; ++d) {
+    if (d % 16 == 0) { std::cout << std::endl; }
+    std::cout << ptr[d] << " ";
+  }
+  std::cout << std::endl;
+}
+
+} // anonymous namespace

csrc/cpu/shm.cpp

@@ -7,9 +7,10 @@
 
 namespace {
 #define MAX_SHM_RANK_NUM 8
-#define MAX_THREAD_NUM 12
-#define PER_THREAD_SHM_BUFFER_BYTES (4 * 1024 * 1024)
-#define MIN_THREAD_PROCESS_SIZE (8 * 1024)
+#define PER_THREAD_SHM_BUFFER_BYTES (2 * 1024 * 1024)
+static_assert(PER_THREAD_SHM_BUFFER_BYTES % 2 == 0);
+#define PER_THREAD_SHM_BUFFER_OFFSET (PER_THREAD_SHM_BUFFER_BYTES >> 1)
+#define MIN_THREAD_PROCESS_SIZE (256)
 #define MAX_P2P_SEND_TENSOR_NUM 8
 
 template <typename scalar_t>
@@ -32,10 +33,10 @@ struct KernelVecType<c10::Half> {
   using scalar_vec_t = vec_op::FP16Vec16;
 };
 
-enum class ThreadSHMStat : char { THREAD_READY = 0, SHM_DATA_READY, DONE };
-
 struct ThreadSHMContext {
-  volatile ThreadSHMStat thread_stats[MAX_SHM_RANK_NUM];
+  volatile char _curr_thread_stamp;
+  volatile char _ready_thread_stamp;
+  char _padding1[6];
   int thread_id;
   int thread_num;
   int rank;
@@ -44,14 +45,19 @@ struct ThreadSHMContext {
   int swizzled_ranks[MAX_SHM_RANK_NUM];
   void* thread_shm_ptrs[MAX_SHM_RANK_NUM];
   ThreadSHMContext* shm_contexts[MAX_SHM_RANK_NUM];
+  size_t _thread_buffer_mask;
+  char _padding2[56];
 
   ThreadSHMContext(const int thread_id, const int thread_num, const int rank,
                    const int group_size, void* thread_shm_ptr)
-      : thread_id(thread_id),
+      : _curr_thread_stamp(1),
+        _ready_thread_stamp(0),
+        thread_id(thread_id),
         thread_num(thread_num),
         rank(rank),
         group_size(group_size),
-        _spinning_count(0) {
+        _spinning_count(0),
+        _thread_buffer_mask(0) {
     static_assert(sizeof(ThreadSHMContext) % 64 == 0);
     TORCH_CHECK(group_size <= MAX_SHM_RANK_NUM);
     TORCH_CHECK((size_t)this % 64 == 0);
@@ -60,7 +66,6 @@ struct ThreadSHMContext {
       shm_contexts[i] = nullptr;
       thread_shm_ptrs[i] = nullptr;
       swizzled_ranks[i] = (i + rank) % group_size;
-      thread_stats[i] = ThreadSHMStat::DONE;
     }
     set_context(rank, this, thread_shm_ptr);
   }
@@ -77,59 +82,66 @@ struct ThreadSHMContext {
 
   template <typename T>
   T* get_thread_shm_ptr(int rank) {
-    return reinterpret_cast<T*>(thread_shm_ptrs[rank]);
+    return reinterpret_cast<T*>(
+        reinterpret_cast<int8_t*>(thread_shm_ptrs[rank]) +
+        (PER_THREAD_SHM_BUFFER_OFFSET & _thread_buffer_mask));
+  }
+
+  void next_buffer() { _thread_buffer_mask ^= 0xFFFFFFFFFFFFFFFF; }
+
+  char get_curr_stamp() const { return _curr_thread_stamp; }
+
+  char get_ready_stamp() const { return _ready_thread_stamp; }
+
+  void next_stamp() {
+    _mm_mfence();
+    _curr_thread_stamp += 1;
+  }
+
+  void commit_ready_stamp() {
+    _mm_mfence();
+    _ready_thread_stamp = _curr_thread_stamp;
   }
 
   int get_swizzled_rank(int idx) { return swizzled_ranks[idx]; }
 
-  void wait_for_all(ThreadSHMStat prev_stat) {
-    for (int idx = 0; idx < group_size; ++idx) {
+  template <typename Cond>
+  void wait_for_all(Cond&& cond) {
+    for (int idx = 1; idx < group_size; ++idx) {
       int rank = get_swizzled_rank(idx);
-      while (thread_stats[rank] == prev_stat) {
-        ++_spinning_count;
-        _mm_pause();
-      }
+      wait_for_one(rank, std::forward<Cond>(cond));
     }
-    vec_op::mem_barrier();
   }
 
-  void wait_for_one(int rank, ThreadSHMStat prev_stat) {
-    while (thread_stats[rank] == prev_stat) {
+  template <typename Cond>
+  void wait_for_one(int rank, Cond&& cond) {
+    ThreadSHMContext* rank_ctx = shm_contexts[rank];
+    for (;;) {
+      char local_curr_stamp = get_curr_stamp();
+      char local_ready_stamp = get_ready_stamp();
+      char rank_curr_stamp = rank_ctx->get_curr_stamp();
+      char rank_ready_stamp = rank_ctx->get_ready_stamp();
+      if (cond(local_curr_stamp, local_ready_stamp, rank_curr_stamp,
+               rank_ready_stamp)) {
+        break;
+      }
       ++_spinning_count;
       _mm_pause();
     }
-    vec_op::mem_barrier();
   }
 
-  void set_thread_stat(ThreadSHMStat stat) {
-    for (int idx = 0; idx < group_size; ++idx) {
-      int rank = get_swizzled_rank(idx);
-      shm_contexts[rank]->thread_stats[this->rank] = stat;
-    }
+  static bool check_no_buffer_conflict(char local_curr_stamp,
+                                       char local_ready_stamp,
+                                       char rank_curr_stamp,
+                                       char rank_ready_stamp) {
+    char temp = rank_curr_stamp + 2;
+    return local_curr_stamp != temp;
   }
 
-  void set_thread_stat(int target_rank, ThreadSHMStat stat) {
-    for (int idx = 0; idx < group_size; ++idx) {
-      int rank = get_swizzled_rank(idx);
-      shm_contexts[rank]->thread_stats[target_rank] = stat;
-    }
-  }
-
-  // barrier for all ranks in the group, used for all2all ops
-  // DONE -> THREAD_READY -> SHM_DATA_READY -> DONE -> ...
-  void barrier(ThreadSHMStat next_stat) {
-    if (next_stat == ThreadSHMStat::THREAD_READY) {
-      set_thread_stat(ThreadSHMStat::THREAD_READY);
-      wait_for_all(ThreadSHMStat::DONE);
-    } else if (next_stat == ThreadSHMStat::SHM_DATA_READY) {
-      set_thread_stat(ThreadSHMStat::SHM_DATA_READY);
-      wait_for_all(ThreadSHMStat::THREAD_READY);
-    } else if (next_stat == ThreadSHMStat::DONE) {
-      set_thread_stat(ThreadSHMStat::DONE);
-      wait_for_all(ThreadSHMStat::SHM_DATA_READY);
-    } else {
-      TORCH_CHECK(false, "Invalid next_stat to barrier.");
-    }
+  static bool check_stamp_ready(char local_curr_stamp, char local_ready_stamp,
+                                char rank_curr_stamp, char rank_ready_stamp) {
+    char temp = local_curr_stamp + 1;
+    return (local_curr_stamp == rank_ready_stamp) || (temp == rank_ready_stamp);
   }
 
   std::string to_string() const {
@@ -164,7 +176,7 @@ class SHMManager {
                       const int group_size)
       : _rank(rank),
         _group_size(group_size),
-        _thread_num(std::min(torch::get_num_threads(), MAX_THREAD_NUM)),
+        _thread_num(torch::get_num_threads()),
         _shm_names({""}),
         _shared_mem_ptrs({nullptr}),
         _shm_ctx(nullptr) {
@@ -326,7 +338,8 @@ void shm_cc_loop(ThreadSHMContext* ctx, int64_t elem_num, F&& inner_func) {
       (total_units_num + thread_num - 1) / thread_num;
   int64_t per_unit_elem_num = MIN_THREAD_PROCESS_SIZE / sizeof(scalar_t);
   int64_t max_per_thread_iteration_elem_num =
-      PER_THREAD_SHM_BUFFER_BYTES / sizeof(scalar_t);
+      (PER_THREAD_SHM_BUFFER_BYTES >> 1) /
+      sizeof(scalar_t);  // Note: double buffer
   int64_t per_thread_elem_num = per_unit_elem_num * per_thread_units_num;
 
 #pragma omp parallel for schedule(static, 1)
@@ -336,10 +349,13 @@ void shm_cc_loop(ThreadSHMContext* ctx, int64_t elem_num, F&& inner_func) {
     int64_t curr_elem_num =
         std::min(max_per_thread_iteration_elem_num, end - offset);
     ThreadSHMContext* thread_ctx = ctx + i;
+    bool fast_mode = ((end - offset) <= max_per_thread_iteration_elem_num);
 
     while (curr_elem_num > 0) {
-      inner_func(thread_ctx, offset, curr_elem_num);
+      inner_func(thread_ctx, offset, curr_elem_num, fast_mode);
 
+      thread_ctx->next_stamp();
+      thread_ctx->next_buffer();
       offset += max_per_thread_iteration_elem_num;
       curr_elem_num = std::min(max_per_thread_iteration_elem_num, end - offset);
     }
@@ -397,7 +413,7 @@ void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
   shm_cc_ops::shm_cc_loop<scalar_t>(
       ctx, elem_num,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
+          int64_t data_elem_num, bool fast_mode) {
         int rank = thread_ctx->rank;
         scalar_t* thread_shm_ptr =
             thread_ctx->get_thread_shm_ptr<scalar_t>(rank);
@@ -410,16 +426,17 @@ void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
               thread_ctx->get_swizzled_rank(idx + 1));
         });
 
-        thread_ctx->barrier(ThreadSHMStat::THREAD_READY);
+        if (!fast_mode) {
+          thread_ctx->wait_for_all(ThreadSHMContext::check_no_buffer_conflict);
+        }
 
         shm_cc_ops::memcpy_to_shm(thread_shm_ptr, thread_data_ptr,
                                   thread_data_elem_num);
-
-        thread_ctx->barrier(ThreadSHMStat::SHM_DATA_READY);
-
+        thread_ctx->commit_ready_stamp();
         int64_t aligned_data_elem_num =
             (data_elem_num / vec_elem_num) * vec_elem_num;
         int64_t i = 0;
+        thread_ctx->wait_for_all(ThreadSHMContext::check_stamp_ready);
 #pragma GCC unroll 4
         for (; i < aligned_data_elem_num; i += vec_elem_num) {
           vec_t local_data(thread_data_ptr + i);  // load from cache
@@ -447,8 +464,6 @@ void all_reduce_sum_impl(ThreadSHMContext* ctx, scalar_t* data,
           reduced_data.save(thread_data_ptr + i,
                             data_elem_num - aligned_data_elem_num);
         }
-
-        thread_ctx->barrier(ThreadSHMStat::DONE);
       });
 
   return;
@@ -488,18 +503,18 @@ void shm_gather_impl(ThreadSHMContext* ctx, scalar_t* data, size_t elem_num,
   shm_cc_ops::shm_cc_loop<scalar_t>(
       ctx, elem_num,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
+          int64_t data_elem_num, bool fast_mode) {
         int rank = thread_ctx->rank;
         scalar_t* thread_shm_ptr =
             thread_ctx->get_thread_shm_ptr<scalar_t>(rank);
 
-        thread_ctx->barrier(ThreadSHMStat::THREAD_READY);
-
-        shm_cc_ops::memcpy_to_shm(thread_shm_ptr, data + data_offset,
-                                  data_elem_num * sizeof(scalar_t));
-
-        thread_ctx->barrier(ThreadSHMStat::SHM_DATA_READY);
+        if (!fast_mode) {
+          thread_ctx->wait_for_all(ThreadSHMContext::check_no_buffer_conflict);
+        }
 
+        shm_cc_ops::memcpy(thread_shm_ptr, data + data_offset,
+                           data_elem_num * sizeof(scalar_t));
+        thread_ctx->commit_ready_stamp();
         if (rank == dst) {
           shm_cc_ops::memcpy(outputs[rank] + data_offset, data + data_offset,
                              data_elem_num * sizeof(scalar_t));
@@ -508,12 +523,12 @@ void shm_gather_impl(ThreadSHMContext* ctx, scalar_t* data, size_t elem_num,
             scalar_t* src_ptr =
                 thread_ctx->get_thread_shm_ptr<scalar_t>(src_rank);  // shm
             scalar_t* dst_ptr = outputs[src_rank] + data_offset;
-            shm_cc_ops::memcpy_from_shm(dst_ptr, src_ptr,
-                                        data_elem_num * sizeof(scalar_t));
+            thread_ctx->wait_for_one(src_rank,
+                                     ThreadSHMContext::check_stamp_ready);
+            shm_cc_ops::memcpy(dst_ptr, src_ptr,
+                               data_elem_num * sizeof(scalar_t));
           }
         }
-
-        thread_ctx->barrier(ThreadSHMStat::DONE);
       });
 
   return;
@@ -599,7 +614,7 @@ struct TensorListMeta {
   int8_t _padding[40];
 };
 
-void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
+void shm_send_tensor_list_impl(ThreadSHMContext* ctx, int64_t dst,
                                const std::vector<torch::Tensor>& tensor_list) {
   CPU_KERNEL_GUARD_IN(shm_send_tensor_list_impl)
   std::vector<torch::Tensor> tensor_list_with_metadata;
@@ -620,12 +635,11 @@ void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
   shm_cc_ops::shm_cc_loop<int8_t>(
       ctx, metadata->total_bytes,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
+          int64_t data_elem_num, bool fast_mode) {
         int rank = thread_ctx->rank;
-        // Wait until the receiver set the stat to DONE
-        thread_ctx->wait_for_one(rank, ThreadSHMStat::SHM_DATA_READY);
-
         int64_t curr_shm_offset = 0;
+        thread_ctx->wait_for_one(dst,
+                                 ThreadSHMContext::check_no_buffer_conflict);
         while (curr_shm_offset < data_elem_num) {
           MemPiece frag = metadata->get_data(data_offset + curr_shm_offset);
           frag.size = std::min(frag.size, data_elem_num - curr_shm_offset);
@@ -634,8 +648,7 @@ void shm_send_tensor_list_impl(ThreadSHMContext* ctx,
               frag.ptr, frag.size);
           curr_shm_offset += frag.size;
         }
-
-        thread_ctx->set_thread_stat(rank, ThreadSHMStat::SHM_DATA_READY);
+        thread_ctx->commit_ready_stamp();
       });
 }
 
@@ -646,8 +659,7 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
   torch::Tensor metadata_tensor =
       torch::empty({sizeof(TensorListMeta)}, options);
 
-  // Wait until the sender set the stat of the thread 0 to SHM_DATA_READY
-  ctx->wait_for_one(src, ThreadSHMStat::DONE);
+  ctx->wait_for_one(src, ThreadSHMContext::check_stamp_ready);
   shm_cc_ops::memcpy(metadata_tensor.data_ptr(),
                      ctx->get_thread_shm_ptr<void>(src),
                      sizeof(TensorListMeta));
@@ -664,9 +676,8 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
   shm_cc_ops::shm_cc_loop<int8_t>(
       ctx, metadata.total_bytes,
       [&](ThreadSHMContext* thread_ctx, int64_t data_offset,
-          int64_t data_elem_num) {
-        // Wait until the sender set the stat to SHM_DATA_READY
-        thread_ctx->wait_for_one(src, ThreadSHMStat::DONE);
+          int64_t data_elem_num, bool fast_mode) {
+        ctx->wait_for_one(src, ThreadSHMContext::check_stamp_ready);
         int64_t curr_shm_offset = 0;
         while (curr_shm_offset < data_elem_num) {
           MemPiece frag = metadata.get_data(data_offset + curr_shm_offset);
@@ -677,8 +688,6 @@ std::vector<torch::Tensor> shm_recv_tensor_list_impl(ThreadSHMContext* ctx,
               frag.size);
           curr_shm_offset += frag.size;
         }
-
-        thread_ctx->set_thread_stat(src, ThreadSHMStat::DONE);
       });
 
   std::vector<torch::Tensor> tensor_list;
@@ -756,7 +765,8 @@ void shm_send_tensor_list(int64_t handle,
                           int64_t dst) {
   CPU_KERNEL_GUARD_IN(shm_send_tensor_list)
   shm_send_tensor_list_impl(
-      SHMManager::get_singleton_instance(handle)->get_shm_ctx(), tensor_list);
+      SHMManager::get_singleton_instance(handle)->get_shm_ctx(), dst,
+      tensor_list);
   CPU_KERNEL_GUARD_OUT(shm_send_tensor_list)
 }
 
@@ -778,4 +788,4 @@ std::string join_shm_manager(int64_t handle, const std::string& name) {
   TORCH_CHECK(shm_manager);
   shm_manager->join(name);
   return shm_manager->get_shm_ctx()->to_string();
-}
+}

csrc/cpu/torch_bindings.cpp

@@ -50,6 +50,27 @@ void shm_send_tensor_list(int64_t handle,
 
 std::vector<torch::Tensor> shm_recv_tensor_list(int64_t handle, int64_t src);
 
+at::Tensor weight_packed_linear(at::Tensor& mat1, at::Tensor& mat2,
+                                const std::optional<at::Tensor>& bias,
+                                bool is_vnni);
+
+at::Tensor convert_weight_packed(at::Tensor& weight);
+
+at::Tensor fused_experts_cpu(
+    at::Tensor& hidden_states, at::Tensor& w1, at::Tensor& w2,
+    at::Tensor& topk_weights, at::Tensor& topk_ids, bool inplace,
+    bool use_int8_w8a8, bool use_fp8_w8a16,
+    const std::optional<at::Tensor>& w1_scale,
+    const std::optional<at::Tensor>& w2_scale,
+    const std::optional<std::vector<int64_t>> block_size,
+    const std::optional<at::Tensor>& a1_scale,
+    const std::optional<at::Tensor>& a2_scale, bool is_vnni);
+
+at::Tensor int8_scaled_mm_with_quant(at::Tensor& mat1, at::Tensor& mat2,
+                                     at::Tensor& scales2,
+                                     const std::optional<at::Tensor>& bias,
+                                     at::ScalarType out_dtype, bool is_vnni);
+
 TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // vLLM custom ops
 
@@ -131,16 +152,19 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
 
   // Quantization
 #ifdef __AVX512F__
+  at::Tag stride_tag = at::Tag::needs_fixed_stride_order;
   // Compute int8 quantized tensor for given scaling factor.
   ops.def(
       "static_scaled_int8_quant(Tensor! out, Tensor input, Tensor scale,"
-      "Tensor? azp) -> ()");
+      "Tensor? azp) -> ()",
+      {stride_tag});
   ops.impl("static_scaled_int8_quant", torch::kCPU, &static_scaled_int8_quant);
 
   // Compute int8 quantized tensor and scaling factor
   ops.def(
       "dynamic_scaled_int8_quant(Tensor! out, Tensor input, Tensor! scale, "
-      "Tensor!? azp) -> ()");
+      "Tensor!? azp) -> ()",
+      {stride_tag});
   ops.impl("dynamic_scaled_int8_quant", torch::kCPU,
            &dynamic_scaled_int8_quant);
   // W8A8 GEMM, supporting symmetric per-tensor or per-row/column
@@ -148,7 +172,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def(
       "cutlass_scaled_mm(Tensor! out, Tensor a,"
       "                  Tensor b, Tensor a_scales,"
-      "                  Tensor b_scales, Tensor? bias) -> ()");
+      "                  Tensor b_scales, Tensor? bias) -> ()",
+      {stride_tag});
   ops.impl("cutlass_scaled_mm", torch::kCPU, &int8_scaled_mm);
   // w8a8 GEMM, supporting asymmetric per-tensor or per-row/column
   // quantization.
@@ -156,7 +181,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
       "cutlass_scaled_mm_azp(Tensor! out, Tensor a,"
       "                  Tensor b, Tensor a_scales,"
       "                  Tensor b_scales, Tensor azp_adj,"
-      "                  Tensor? azp, Tensor? bias) -> ()");
+      "                  Tensor? azp, Tensor? bias) -> ()",
+      {stride_tag});
   ops.impl("cutlass_scaled_mm_azp", torch::kCPU, &int8_scaled_mm_azp);
 #elif defined(__powerpc64__)
   // Compute int8 quantized tensor for given scaling factor.
@@ -209,6 +235,28 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   ops.def("shm_recv_tensor_list(int handle, int src) -> Tensor[](a)",
           &shm_recv_tensor_list);
 #endif
+
+  // sgl-kernels
+#if defined(__AVX512BF16__) && defined(__AVX512F__) && defined(__AVX512VNNI__)
+  ops.def(
+      "weight_packed_linear(Tensor(a0!) mat1, Tensor(a1!) mat2, Tensor(a2!)? "
+      "bias, bool is_vnni) -> Tensor");
+  ops.impl("weight_packed_linear", torch::kCPU, &weight_packed_linear);
+  ops.def("convert_weight_packed(Tensor! weight) -> Tensor");
+  ops.impl("convert_weight_packed", torch::kCPU, &convert_weight_packed);
+  ops.def(
+      "fused_experts_cpu(Tensor! hidden_states, Tensor w1, Tensor w2, Tensor "
+      "topk_weights, Tensor topk_ids, bool inplace, bool use_int8_w8a8, bool "
+      "use_fp8_w8a16, Tensor? w1_scale, Tensor? w2_scale, SymInt[]? "
+      "block_size, Tensor? a1_scale, Tensor? a2_scale, bool is_vnni) -> "
+      "Tensor");
+  ops.impl("fused_experts_cpu", torch::kCPU, &fused_experts_cpu);
+  ops.def(
+      "int8_scaled_mm_with_quant(Tensor mat1, Tensor mat2, Tensor scales2, "
+      "Tensor? bias, ScalarType out_dtype, bool is_vnni) -> Tensor");
+  ops.impl("int8_scaled_mm_with_quant", torch::kCPU,
+           &int8_scaled_mm_with_quant);
+#endif
 }
 
 TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {