Initial stubs for P/D scheduling changes

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

.buildkite/lm-eval-harness/configs/Minitron-4B-Base-FP8.yaml

@@ -4,8 +4,8 @@ tasks:
 - name: "gsm8k"
   metrics:
   - name: "exact_match,strict-match"
-    value: 0.233
+    value: 0.231
   - name: "exact_match,flexible-extract"
-    value: 0.236
+    value: 0.22
 limit: 1000
 num_fewshot: 5

.buildkite/lm-eval-harness/configs/Qwen1.5-MoE-W4A16-compressed-tensors.yaml

@@ -0,0 +1,11 @@
+# bash .buildkite/lm-eval-harness/run-lm-eval-gsm-vllm-baseline.sh -m nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16 -b auto -l 1319 -f 5 -t 1
+model_name: "nm-testing/Qwen1.5-MoE-A2.7B-Chat-quantized.w4a16"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.31
+  - name: "exact_match,flexible-extract"
+    value: 0.47
+limit: 1319
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/models-small.txt

@@ -4,7 +4,7 @@ Meta-Llama-3.2-1B-Instruct-INT8-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-INT8-compressed-tensors-asym.yaml
 Meta-Llama-3-8B-Instruct-nonuniform-compressed-tensors.yaml
 Meta-Llama-3-8B-Instruct-Channelwise-compressed-tensors.yaml
-Minitron-4B-Base-FP8.yaml
+Qwen1.5-MoE-W4A16-compressed-tensors.yaml
 Qwen2-1.5B-Instruct-INT8-compressed-tensors.yaml
 Qwen2-1.5B-Instruct-FP8W8.yaml
 Meta-Llama-3-8B-QQQ.yaml

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

@@ -13,6 +13,7 @@ from pathlib import Path
 
 import lm_eval
 import numpy
+import pytest
 import yaml
 
 RTOL = 0.05
@@ -46,6 +47,10 @@ def test_lm_eval_correctness():
     eval_config = yaml.safe_load(
         Path(TEST_DATA_FILE).read_text(encoding="utf-8"))
 
+    if eval_config[
+            "model_name"] == "nm-testing/Meta-Llama-3-70B-Instruct-FBGEMM-nonuniform":  #noqa: E501
+        pytest.skip("FBGEMM is currently failing on main.")
+
     # Launch eval requests.
     results = launch_lm_eval(eval_config)
 

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

@@ -84,8 +84,13 @@ if __name__ == "__main__":
             # this result is generated via `benchmark_serving.py`
 
             # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands")) as f:
-                command = json.loads(f.read())
+            try:
+                with open(test_file.with_suffix(".commands")) as f:
+                    command = json.loads(f.read())
+            except OSError as e:
+                print(e)
+                continue
+
             raw_result.update(command)
 
             # update the test name of this result
@@ -99,8 +104,13 @@ if __name__ == "__main__":
             # this result is generated via `benchmark_latency.py`
 
             # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands")) as f:
-                command = json.loads(f.read())
+            try:
+                with open(test_file.with_suffix(".commands")) as f:
+                    command = json.loads(f.read())
+            except OSError as e:
+                print(e)
+                continue
+
             raw_result.update(command)
 
             # update the test name of this result
@@ -121,8 +131,13 @@ if __name__ == "__main__":
             # this result is generated via `benchmark_throughput.py`
 
             # attach the benchmarking command to raw_result
-            with open(test_file.with_suffix(".commands")) as f:
-                command = json.loads(f.read())
+            try:
+                with open(test_file.with_suffix(".commands")) as f:
+                    command = json.loads(f.read())
+            except OSError as e:
+                print(e)
+                continue
+
             raw_result.update(command)
 
             # update the test name of this result

.buildkite/nightly-benchmarks/scripts/run-nightly-benchmarks.sh

@@ -426,7 +426,7 @@ main() {
 
   pip install -U transformers
 
-  pip install -r requirements-dev.txt
+  pip install -r requirements/dev.txt
   which genai-perf
 
   # check storage

.buildkite/nightly-benchmarks/scripts/run-performance-benchmarks.sh

@@ -10,15 +10,24 @@ set -x
 set -o pipefail
 
 check_gpus() {
-  # check the number of GPUs and GPU type.
-  declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+  if command -v nvidia-smi; then
+    # check the number of GPUs and GPU type.
+    declare -g gpu_count=$(nvidia-smi --list-gpus | wc -l)
+  elif command -v amd-smi; then
+    declare -g gpu_count=$(amd-smi list | grep 'GPU' | wc -l)
+  fi
+
   if [[ $gpu_count -gt 0 ]]; then
     echo "GPU found."
   else
     echo "Need at least 1 GPU to run benchmarking."
     exit 1
   fi
-  declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
+  if command -v nvidia-smi; then
+    declare -g gpu_type=$(nvidia-smi --query-gpu=name --format=csv,noheader | awk '{print $2}')
+  elif command -v amd-smi; then
+    declare -g gpu_type=$(amd-smi static -g 0 -a | grep 'MARKET_NAME' | awk '{print $2}')
+  fi
   echo "GPU type is $gpu_type"
 }
 
@@ -90,9 +99,15 @@ kill_gpu_processes() {
 
 
   # wait until GPU memory usage smaller than 1GB
-  while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
-    sleep 1
-  done
+  if command -v nvidia-smi; then
+    while [ "$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits | head -n 1)" -ge 1000 ]; do
+      sleep 1
+    done
+  elif command -v amd-smi; then
+    while [ "$(amd-smi metric -g 0 | grep 'USED_VRAM' | awk '{print $2}')" -ge 1000 ]; do
+      sleep 1
+    done
+  fi
 
   # remove vllm config file
   rm -rf ~/.config/vllm
@@ -309,11 +324,14 @@ run_serving_tests() {
 
       new_test_name=$test_name"_qps_"$qps
 
+      # pass the tensor parallel size to the client so that it can be displayed
+      # on the benchmark dashboard
       client_command="python3 benchmark_serving.py \
         --save-result \
         --result-dir $RESULTS_FOLDER \
         --result-filename ${new_test_name}.json \
         --request-rate $qps \
+        --metadata "tensor_parallel_size=$tp" \
         $client_args"
 
       echo "Running test case $test_name with qps $qps"
@@ -358,7 +376,7 @@ main() {
   # get the current IP address, required by benchmark_serving.py
   export VLLM_HOST_IP=$(hostname -I | awk '{print $1}')
   # turn of the reporting of the status of each request, to clean up the terminal output
-  export VLLM_LOG_LEVEL="WARNING"
+  export VLLM_LOGGING_LEVEL="WARNING"
 
   # prepare for benchmarking
   cd benchmarks || exit 1

.buildkite/nightly-benchmarks/tests/serving-tests.json

@@ -64,9 +64,11 @@
             "disable_log_requests": "", 
             "tensor_parallel_size": 4,
             "swap_space": 16,
-            "speculative_model": "turboderp/Qwama-0.5B-Instruct",
-            "num_speculative_tokens": 4,
-            "speculative_draft_tensor_parallel_size": 1
+            "speculative_config": {
+                "model": "turboderp/Qwama-0.5B-Instruct",
+                "num_speculative_tokens": 4,
+                "draft_tensor_parallel_size": 1
+            }
         },
         "client_parameters": {
             "model": "meta-llama/Meta-Llama-3.1-70B-Instruct",

.buildkite/release-pipeline.yaml

@@ -1,12 +1,23 @@
 steps:
+  - label: "Build wheel - CUDA 12.4"
+    agents:
+      queue: cpu_queue_postmerge
+    commands:
+      - "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=12.4.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
+      - "mkdir artifacts"
+      - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
+      - "bash .buildkite/scripts/upload-wheels.sh"
+    env:
+      DOCKER_BUILDKIT: "1"
+
   - label: "Build wheel - CUDA 12.1"
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "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=12.1.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "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=12.1.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
@@ -20,10 +31,10 @@ steps:
     agents:
       queue: cpu_queue_postmerge
     commands:
-      - "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=11.8.0 --tag vllm-ci:build-image --target build --progress plain ."
+      - "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=11.8.0 --tag vllm-ci:build-image --target build --progress plain -f docker/Dockerfile ."
       - "mkdir artifacts"
       - "docker run --rm -v $(pwd)/artifacts:/artifacts_host vllm-ci:build-image bash -c 'cp -r dist /artifacts_host && chmod -R a+rw /artifacts_host'"
-      - "bash .buildkite/upload-wheels.sh"
+      - "bash .buildkite/scripts/upload-wheels.sh"
     env:
       DOCKER_BUILDKIT: "1"
 
@@ -37,7 +48,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.1.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --build-arg CUDA_VERSION=12.4.0 --tag public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT --target vllm-openai --progress plain -f docker/Dockerfile ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-release-repo:$BUILDKITE_COMMIT"
 
   - label: "Build and publish TPU release image"
@@ -46,7 +57,7 @@ steps:
     agents:
       queue: tpu_queue_postmerge
     commands:
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f Dockerfile.tpu ."
+      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg USE_SCCACHE=1 --build-arg GIT_REPO_CHECK=1 --tag vllm/vllm-tpu:nightly --tag vllm/vllm-tpu:$BUILDKITE_COMMIT --progress plain -f docker/Dockerfile.tpu ."
       - "docker push vllm/vllm-tpu:nightly"
       - "docker push vllm/vllm-tpu:$BUILDKITE_COMMIT"
     plugins:
@@ -71,7 +82,7 @@ steps:
       queue: cpu_queue_postmerge
     commands:
       - "aws ecr-public get-login-password --region us-east-1 | docker login --username AWS --password-stdin public.ecr.aws/q9t5s3a7"
-      - "DOCKER_BUILDKIT=1 docker build --build-arg max_jobs=16 --build-arg GIT_REPO_CHECK=1 --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version) --progress plain -f Dockerfile.cpu ."
+      - "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:$(buildkite-agent meta-data get release-version) --tag public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:latest --progress plain --target vllm-openai -f docker/Dockerfile.cpu ."
       - "docker push public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:$(buildkite-agent meta-data get release-version)"
     env:
       DOCKER_BUILDKIT: "1"

.buildkite/run-openvino-test.sh

@@ -1,16 +0,0 @@
-#!/bin/bash
-
-# This script build the OpenVINO docker image and run the offline inference inside the container.
-# It serves a sanity check for compilation and basic model usage.
-set -ex
-
-# Try building the docker image
-docker build -t openvino-test -f Dockerfile.openvino .
-
-# Setup cleanup
-remove_docker_container() { docker rm -f openvino-test || true; }
-trap remove_docker_container EXIT
-remove_docker_container
-
-# Run the image and launch offline inference
-docker run --network host --env VLLM_OPENVINO_KVCACHE_SPACE=1 --name openvino-test openvino-test python3 /workspace/examples/offline_inference/basic/generate.py --model facebook/opt-125m

.buildkite/run-tpu-test.sh

@@ -1,26 +0,0 @@
-#!/bin/bash
-
-set -e
-
-# Build the docker image.
-docker build -f 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
-
-# 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 \
-    && python3 -m pip install lm_eval[api]==0.4.4 \
-    && pytest -v -s /workspace/vllm/tests/entrypoints/openai/test_accuracy.py \
-    && pytest -v -s /workspace/vllm/tests/tpu/test_custom_dispatcher.py \
-    && python3 /workspace/vllm/tests/tpu/test_compilation.py \
-    && python3 /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
-    && python3 /workspace/vllm/examples/offline_inference/tpu.py"

.buildkite/run-xpu-test.sh

@@ -1,19 +0,0 @@
-#!/bin/bash
-
-# 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
-
-# Try building the docker image
-docker build -t xpu-test -f Dockerfile.xpu .
-
-# Setup cleanup
-remove_docker_container() { docker rm -f xpu-test || true; }
-trap remove_docker_container EXIT
-remove_docker_container
-
-# Run the image and test offline inference/tensor parallel
-docker run --name xpu-test --device /dev/dri -v /dev/dri/by-path:/dev/dri/by-path --entrypoint="" xpu-test sh -c '
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m
-    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m -tp 2
-'

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

@@ -77,7 +77,6 @@ echo "Commands:$commands"
 #ignore certain kernels tests
 if [[ $commands == *" kernels "* ]]; then
   commands="${commands} \
-  --ignore=kernels/test_attention.py \
   --ignore=kernels/test_attention_selector.py \
   --ignore=kernels/test_blocksparse_attention.py \
   --ignore=kernels/test_causal_conv1d.py \
@@ -92,19 +91,54 @@ if [[ $commands == *" kernels "* ]]; then
   --ignore=kernels/test_moe.py \
   --ignore=kernels/test_prefix_prefill.py \
   --ignore=kernels/test_rand.py \
-  --ignore=kernels/test_sampler.py"
+  --ignore=kernels/test_sampler.py \
+  --ignore=kernels/test_cascade_flash_attn.py \
+  --ignore=kernels/test_mamba_mixer2.py \
+  --ignore=kernels/test_aqlm.py \
+  --ignore=kernels/test_machete_mm.py \
+  --ignore=kernels/test_mha_attn.py \
+  --ignore=kernels/test_block_fp8.py \
+  --ignore=kernels/test_permute_cols.py"
 fi
 
-#ignore certain Entrypoints tests
+#ignore certain Entrypoints/openai tests
 if [[ $commands == *" entrypoints/openai "* ]]; then
   commands=${commands//" entrypoints/openai "/" entrypoints/openai \
-  --ignore=entrypoints/openai/test_accuracy.py \
   --ignore=entrypoints/openai/test_audio.py \
-  --ignore=entrypoints/openai/test_encoder_decoder.py \
-  --ignore=entrypoints/openai/test_embedding.py \
-  --ignore=entrypoints/openai/test_oot_registration.py "}
+  --ignore=entrypoints/openai/test_shutdown.py \
+  --ignore=entrypoints/openai/test_completion.py \
+  --ignore=entrypoints/openai/test_sleep.py \
+  --ignore=entrypoints/openai/test_models.py \
+  --ignore=entrypoints/openai/test_lora_adapters.py \
+  --ignore=entrypoints/openai/test_return_tokens_as_ids.py \
+  --ignore=entrypoints/openai/test_root_path.py \
+  --ignore=entrypoints/openai/test_tokenization.py \
+  --ignore=entrypoints/openai/test_prompt_validation.py "}
 fi
 
+#ignore certain Entrypoints/llm tests
+if [[ $commands == *" entrypoints/llm "* ]]; then
+  commands=${commands//" entrypoints/llm "/" entrypoints/llm \
+  --ignore=entrypoints/llm/test_chat.py \
+  --ignore=entrypoints/llm/test_accuracy.py \
+  --ignore=entrypoints/llm/test_init.py \
+  --ignore=entrypoints/llm/test_generate_multiple_loras.py \
+  --ignore=entrypoints/llm/test_prompt_validation.py "}
+fi
+
+#Obsolete currently
+##ignore certain Entrypoints/llm tests
+#if [[ $commands == *" && pytest -v -s entrypoints/llm/test_guided_generate.py"* ]]; then
+#  commands=${commands//" && pytest -v -s entrypoints/llm/test_guided_generate.py"/" "}
+#fi
+
+# --ignore=entrypoints/openai/test_encoder_decoder.py \
+# --ignore=entrypoints/openai/test_embedding.py \
+# --ignore=entrypoints/openai/test_oot_registration.py
+# --ignore=entrypoints/openai/test_accuracy.py \
+# --ignore=entrypoints/openai/test_models.py <= Fails on MI250 but passes on MI300 as of 2025-03-13
+
+
 PARALLEL_JOB_COUNT=8
 # check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs. 
 if [[ $commands == *"--shard-id="* ]]; then
@@ -114,9 +148,10 @@ if [[ $commands == *"--shard-id="* ]]; then
     # assign shard-id for each shard
     commands_gpu=${commands//"--shard-id= "/"--shard-id=${GPU} "}
     echo "Shard ${GPU} commands:$commands_gpu"
+    echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
     docker run \
-        --device /dev/kfd --device /dev/dri \
-        --network host \
+        --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+        --network=host \
         --shm-size=16gb \
         --rm \
         -e HIP_VISIBLE_DEVICES="${GPU}" \
@@ -143,9 +178,10 @@ if [[ $commands == *"--shard-id="* ]]; then
     fi
   done
 else
+  echo "Render devices: $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES"
   docker run \
-          --device /dev/kfd --device /dev/dri \
-          --network host \
+          --device /dev/kfd $BUILDKITE_AGENT_META_DATA_RENDER_DEVICES \
+          --network=host \
           --shm-size=16gb \
           --rm \
           -e HIP_VISIBLE_DEVICES=0 \

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

@@ -0,0 +1,38 @@
+#!/bin/bash
+
+# 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
+
+# Setup cleanup
+remove_docker_container() { podman rm -f cpu-test-ubi9-ppc || true; podman system prune -f; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Try building the docker image
+podman build -t cpu-test-ubi9-ppc -f docker/Dockerfile.ppc64le .
+
+# Run the image
+podman run -itd --entrypoint /bin/bash -v /tmp/:/root/.cache/huggingface --privileged=true --network host -e HF_TOKEN --name cpu-test-ubi9-ppc cpu-test-ubi9-ppc
+
+function cpu_tests() {
+
+  # offline inference
+  podman exec cpu-test-ubi9-ppc bash -c "
+    set -e
+    python3 examples/offline_inference/basic/generate.py --model facebook/opt-125m"
+
+  # Run basic model test
+  podman exec cpu-test-ubi9-ppc bash -c "
+    set -e
+    pip install pytest pytest-asyncio einops peft Pillow soundfile transformers_stream_generator matplotlib
+    pip install sentence-transformers datamodel_code_generator
+    pytest -v -s tests/models/embedding/language/test_cls_models.py::test_classification_models[float-jason9693/Qwen2.5-1.5B-apeach]
+    pytest -v -s tests/models/embedding/language/test_embedding.py::test_models[half-BAAI/bge-base-en-v1.5]
+    pytest -v -s tests/models/encoder_decoder/language -m cpu_model"
+}
+
+# All of CPU tests are expected to be finished less than 40 mins.
+export -f cpu_tests
+timeout 40m bash -c cpu_tests
+

.buildkite/scripts/hardware_ci/run-cpu-test-s390x.sh

@@ -10,5 +10,4 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Try building the docker image
-docker build -t cpu-test -f Dockerfile.ppc64le .
-
+docker build -t cpu-test -f docker/Dockerfile.s390x .

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

@@ -8,24 +8,29 @@ set -ex
 CORE_RANGE=${CORE_RANGE:-48-95}
 NUMA_NODE=${NUMA_NODE:-1}
 
-# Try building the docker image
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build -t cpu-test-"$BUILDKITE_BUILD_NUMBER" -f Dockerfile.cpu .
-numactl -C "$CORE_RANGE" -N "$NUMA_NODE" docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" -t cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2 -f Dockerfile.cpu .
-
 # 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; }
+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; 
+}
 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 .
+
 # 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 --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"
+ --cpuset-mems="$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="$NUMA_NODE" --privileged=true --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --shm-size=4g --name cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2
+ --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
 
 function cpu_tests() {
   set -e
   export NUMA_NODE=$2
+  export BUILDKITE_BUILD_NUMBER=$3
 
   # offline inference
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-avx2-"$NUMA_NODE" bash -c "
@@ -35,7 +40,8 @@ function cpu_tests() {
   # Run basic model test
   docker exec cpu-test-"$BUILDKITE_BUILD_NUMBER"-"$NUMA_NODE" bash -c "
     set -e
-    pip install -r vllm/requirements-test.txt
+    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
@@ -85,4 +91,4 @@ 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"
+timeout 40m bash -c "cpu_tests $CORE_RANGE $NUMA_NODE $BUILDKITE_BUILD_NUMBER"

.buildkite/scripts/hardware_ci/run-gh200-test.sh

@@ -9,11 +9,13 @@ python3 use_existing_torch.py
 
 # Try building the docker image
 DOCKER_BUILDKIT=1 docker build . \
+  --file docker/Dockerfile \
   --target vllm-openai \
   --platform "linux/arm64" \
   -t gh200-test \
   --build-arg max_jobs=66 \
   --build-arg nvcc_threads=2 \
+  --build-arg RUN_WHEEL_CHECK=false \
   --build-arg torch_cuda_arch_list="9.0+PTX" \
   --build-arg vllm_fa_cmake_gpu_arches="90-real"
 
@@ -23,6 +25,6 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image and test offline inference
-docker run -e HF_TOKEN -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
+docker run -e HF_TOKEN -e VLLM_WORKER_MULTIPROC_METHOD=spawn -v /root/.cache/huggingface:/root/.cache/huggingface --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
     python3 examples/offline_inference/basic/generate.py --model meta-llama/Llama-3.2-1B
 '

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

@@ -5,7 +5,7 @@
 set -ex
 
 # Try building the docker image
-docker build -t hpu-test-env -f Dockerfile.hpu .
+docker build -t hpu-test-env -f docker/Dockerfile.hpu .
 
 # Setup cleanup
 # certain versions of HPU software stack have a bug that can

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

@@ -35,7 +35,7 @@ else
     date "+%s" > /tmp/neuron-docker-build-timestamp
 fi
 
-docker build -t "${image_name}" -f Dockerfile.neuron .
+docker build -t "${image_name}" -f docker/Dockerfile.neuron .
 
 # Setup cleanup
 remove_docker_container() {
@@ -44,11 +44,11 @@ remove_docker_container() {
 trap remove_docker_container EXIT
 
 # Run the image
-docker run --rm -it --device=/dev/neuron0 --device=/dev/neuron1 --network host \
+docker run --rm -it --device=/dev/neuron0 --network bridge \
        -v "${HF_CACHE}:${HF_MOUNT}" \
        -e "HF_HOME=${HF_MOUNT}" \
        -v "${NEURON_COMPILE_CACHE_URL}:${NEURON_COMPILE_CACHE_MOUNT}" \
        -e "NEURON_COMPILE_CACHE_URL=${NEURON_COMPILE_CACHE_MOUNT}" \
        --name "${container_name}" \
        ${image_name} \
-       /bin/bash -c "python3 /workspace/vllm/examples/offline_inference/neuron.py && python3 -m pytest /workspace/vllm/tests/neuron/ -v --capture=tee-sys"
+       /bin/bash -c "python3 /workspace/vllm/examples/offline_inference/neuron.py && python3 -m pytest /workspace/vllm/tests/neuron/1_core/ -v --capture=tee-sys && python3 -m pytest /workspace/vllm/tests/neuron/2_core/ -v --capture=tee-sys"

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

@@ -0,0 +1,49 @@
+#!/bin/bash
+
+set -xue
+
+# 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
+
+# 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 tpu-info \
+    && python3 -m pip install lm_eval[api]==0.4.4 \
+    && export VLLM_USE_V1=1 \
+    && export VLLM_XLA_CHECK_RECOMPILATION=1 \
+    && echo HARDWARE \
+    && tpu-info \
+    && echo TEST_0 \
+    && pytest -v -s /workspace/vllm/tests/v1/tpu/test_perf.py \
+    && echo TEST_1 \
+    && pytest -v -s /workspace/vllm/tests/tpu/test_compilation.py \
+    && echo TEST_2 \
+    && pytest -v -s /workspace/vllm/tests/v1/tpu/test_basic.py \
+    && echo TEST_3 \
+    && pytest -v -s /workspace/vllm/tests/entrypoints/llm/test_accuracy.py::test_lm_eval_accuracy_v1_engine \
+    && echo TEST_4 \
+    && pytest -s -v /workspace/vllm/tests/tpu/test_quantization_accuracy.py \
+    && echo TEST_5 \
+    && python3 /workspace/vllm/examples/offline_inference/tpu.py \
+    && echo TEST_6 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/worker/test_tpu_model_runner.py \
+    && echo TEST_7 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_sampler.py \
+    && echo TEST_8 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_topk_topp_sampler.py \
+    && echo TEST_9 \
+    && pytest -s -v /workspace/vllm/tests/v1/tpu/test_pallas.py" \
+
+
+# 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 \

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

@@ -0,0 +1,31 @@
+#!/bin/bash
+
+# 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
+
+image_name="xpu/vllm-ci:${BUILDKITE_COMMIT}"
+container_name="xpu_${BUILDKITE_COMMIT}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)"
+
+# Try building the docker image
+docker build -t ${image_name} -f docker/Dockerfile.xpu .
+
+# Setup cleanup
+remove_docker_container() { 
+  docker rm -f "${container_name}" || true; 
+  docker image rm -f "${image_name}" || true;
+  docker system prune -f || true;
+}
+trap remove_docker_container EXIT
+
+# Run the image and test offline inference/tensor parallel
+docker run \
+    --device /dev/dri \
+    -v /dev/dri/by-path:/dev/dri/by-path \
+    --entrypoint="" \
+    --name "${container_name}" \
+    "${image_name}" \
+    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
+'

.buildkite/scripts/run-benchmarks.sh

@@ -5,8 +5,8 @@
 set -ex
 set -o pipefail
 
-# cd into parent directory of this file
-cd "$(dirname "${BASH_SOURCE[0]}")/.."
+# cd 2 levels into the working directory
+cd "$(dirname "${BASH_SOURCE[0]}")/../.."
 
 (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
 

.buildkite/scripts/run-multi-node-test.sh

@@ -3,7 +3,7 @@
 set -euox pipefail
 
 if [[ $# -lt 4 ]]; then
-    echo "Usage: .buildkite/run-multi-node-test.sh WORKING_DIR NUM_NODES NUM_GPUS DOCKER_IMAGE COMMAND1 COMMAND2 ... COMMANDN"
+    echo "Usage: .buildkite/scripts/run-multi-node-test.sh WORKING_DIR NUM_NODES NUM_GPUS DOCKER_IMAGE COMMAND1 COMMAND2 ... COMMANDN"
     exit 1
 fi
 

.buildkite/scripts/upload-wheels.sh

@@ -50,8 +50,11 @@ aws s3 cp "$normal_wheel" "s3://vllm-wheels/$BUILDKITE_COMMIT/"
 if [[ $normal_wheel == *"cu118"* ]]; then
     # if $normal_wheel matches cu118, do not upload the index.html
     echo "Skipping index files for cu118 wheels"
+elif [[ $normal_wheel == *"cu121"* ]]; then
+    # if $normal_wheel matches cu121, do not upload the index.html
+    echo "Skipping index files for cu121 wheels"
 else
-    # only upload index.html for cu12 wheels (default wheels)
+    # only upload index.html for cu124 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/$BUILDKITE_COMMIT/vllm/index.html"
     aws s3 cp "s3://vllm-wheels/nightly/index.html" "s3://vllm-wheels/$BUILDKITE_COMMIT/index.html"
 fi
@@ -63,8 +66,11 @@ aws s3 cp "$normal_wheel" "s3://vllm-wheels/nightly/"
 if [[ $normal_wheel == *"cu118"* ]]; then
     # if $normal_wheel matches cu118, do not upload the index.html
     echo "Skipping index files for cu118 wheels"
+elif [[ $normal_wheel == *"cu121"* ]]; then
+    # if $normal_wheel matches cu121, do not upload the index.html
+    echo "Skipping index files for cu121 wheels"
 else
-    # only upload index.html for cu12 wheels (default wheels)
+    # only upload index.html for cu124 wheels (default wheels)
     aws s3 cp index.html "s3://vllm-wheels/nightly/vllm/index.html"
 fi
 

.buildkite/test-pipeline.yaml

@@ -35,13 +35,12 @@ steps:
   fast_check: true
   no_gpu: True
   commands:
-  - pip install -r requirements-docs.txt
+  - pip install -r ../../requirements/docs.txt
   - SPHINXOPTS=\"-W\" make html
   # Check API reference (if it fails, you may have missing mock imports)
   - grep \"sig sig-object py\" build/html/api/inference_params.html
 
 - label: Async Engine, Inputs, Utils, Worker Test # 24min
-  fast_check: true
   source_file_dependencies:
   - vllm/
   - tests/mq_llm_engine
@@ -78,6 +77,7 @@ steps:
   - tests/basic_correctness/test_preemption
   - tests/basic_correctness/test_cumem.py
   commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s basic_correctness/test_cumem.py
   - pytest -v -s basic_correctness/test_basic_correctness.py
   - pytest -v -s basic_correctness/test_cpu_offload.py
@@ -104,7 +104,7 @@ steps:
 - label: Entrypoints Test # 40min
   working_dir: "/vllm-workspace/tests"
   fast_check: true
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/entrypoints/llm
@@ -112,19 +112,19 @@ steps:
   - tests/entrypoints/test_chat_utils
   - tests/entrypoints/offline_mode
   commands:
+  - export VLLM_WORKER_MULTIPROC_METHOD=spawn
   - pytest -v -s entrypoints/llm --ignore=entrypoints/llm/test_lazy_outlines.py --ignore=entrypoints/llm/test_generate.py --ignore=entrypoints/llm/test_generate_multiple_loras.py --ignore=entrypoints/llm/test_guided_generate.py --ignore=entrypoints/llm/test_collective_rpc.py
   - pytest -v -s entrypoints/llm/test_lazy_outlines.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate.py # it needs a clean process
   - pytest -v -s entrypoints/llm/test_generate_multiple_loras.py # it needs a clean process
-  - pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
-  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py --ignore=entrypoints/openai/correctness/
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/llm/test_guided_generate.py # it needs a clean process
+  - pytest -v -s entrypoints/openai --ignore=entrypoints/openai/test_oot_registration.py  --ignore=entrypoints/openai/test_chat_with_tool_reasoning.py --ignore=entrypoints/openai/correctness/ --ignore=entrypoints/openai/test_openai_schema.py
   - pytest -v -s entrypoints/test_chat_utils.py
-  - pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
+  - VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode # Needs to avoid interference with other tests
 
 - label: Distributed Tests (4 GPUs) # 10min
   working_dir: "/vllm-workspace/tests"
   num_gpus: 4
-  fast_check: true
   source_file_dependencies:
   - vllm/distributed/
   - vllm/core/
@@ -134,37 +134,42 @@ steps:
   - tests/compile/test_basic_correctness
   - examples/offline_inference/rlhf.py
   - examples/offline_inference/rlhf_colocate.py
+  - tests/examples/offline_inference/data_parallel.py
+  - tests/v1/test_async_llm_dp.py
   commands:
+  # test with tp=2 and external_dp=2
+  - VLLM_USE_V1=0 torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  - torchrun --nproc-per-node=4 distributed/test_torchrun_example.py
+  # test with 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 distributed/test_utils.py
   - pytest -v -s compile/test_basic_correctness.py
   - pytest -v -s distributed/test_pynccl.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
   # TODO: create a dedicated test section for multi-GPU example tests
   # when we have multiple distributed example tests
-  - python3 ../examples/offline_inference/rlhf.py
-  - RAY_DEDUP_LOGS=0 python3 ../examples/offline_inference/rlhf_colocate.py
+  - pushd ../examples/offline_inference
+  - python3 rlhf.py
+  - RAY_DEDUP_LOGS=0 python3 rlhf_colocate.py
+  - popd
 
 - label: Metrics, Tracing Test # 10min
+  mirror_hardwares: [amd]
   num_gpus: 2
-  fast_check: true
   source_file_dependencies:
   - vllm/
   - tests/metrics
   - tests/tracing
   commands:
   - pytest -v -s metrics
-  - "pip install \
-      'opentelemetry-sdk>=1.26.0,<1.27.0' \
-      'opentelemetry-api>=1.26.0,<1.27.0' \
-      'opentelemetry-exporter-otlp>=1.26.0,<1.27.0' \
-      'opentelemetry-semantic-conventions-ai>=0.4.1,<0.5.0'"
   - pytest -v -s tracing
 
 ##### fast check tests  #####
 #####  1 GPU test  #####
 
 - label: Regression Test # 5min
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/test_regression
@@ -194,15 +199,18 @@ steps:
     - tests/v1
   commands:
     # split the test to avoid interference
-    - VLLM_USE_V1=1 pytest -v -s v1/core
-    - VLLM_USE_V1=1 pytest -v -s v1/engine
-    - VLLM_USE_V1=1 pytest -v -s v1/sample
-    - VLLM_USE_V1=1 pytest -v -s v1/worker
-    - VLLM_USE_V1=1 pytest -v -s v1/test_stats.py
-    - VLLM_USE_V1=1 pytest -v -s v1/test_utils.py
+    - pytest -v -s v1/core
+    - pytest -v -s v1/engine
+    - pytest -v -s v1/entrypoints
+    - pytest -v -s v1/sample
+    - pytest -v -s v1/worker
+    - pytest -v -s v1/structured_output
+    - pytest -v -s v1/test_stats.py
+    - pytest -v -s v1/test_utils.py
+    - pytest -v -s v1/test_oracle.py
     # TODO: accuracy does not match, whether setting
     # VLLM_USE_FLASHINFER_SAMPLER or not on H100.
-    - VLLM_USE_V1=1 pytest -v -s v1/e2e
+    - pytest -v -s v1/e2e
     # Integration test for streaming correctness (requires special branch).
     - pip install -U git+https://github.com/robertgshaw2-neuralmagic/lm-evaluation-harness.git@streaming-api
     - pytest -v -s entrypoints/openai/correctness/test_lmeval.py::test_lm_eval_accuracy_v1_engine
@@ -220,14 +228,17 @@ steps:
     - python3 offline_inference/basic/chat.py
     - python3 offline_inference/prefix_caching.py
     - python3 offline_inference/llm_engine_example.py
-    - python3 offline_inference/vision_language.py
-    - python3 offline_inference/vision_language_multi_image.py
-    - python3 other/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 other/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
+    - python3 offline_inference/audio_language.py --seed 0
+    - python3 offline_inference/vision_language.py --seed 0
+    - python3 offline_inference/vision_language_embedding.py --seed 0
+    - python3 offline_inference/vision_language_multi_image.py --seed 0
+    - VLLM_USE_V1=0 python3 other/tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 other/tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
     - python3 offline_inference/encoder_decoder.py
+    - python3 offline_inference/encoder_decoder_multimodal.py --model-type whisper --seed 0
     - python3 offline_inference/basic/classify.py
     - python3 offline_inference/basic/embed.py
     - python3 offline_inference/basic/score.py
-    - python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
+    - VLLM_USE_V1=0 python3 offline_inference/profiling.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Prefix Caching Test # 9min
   mirror_hardwares: [amd]
@@ -269,15 +280,22 @@ steps:
     - pytest -v -s spec_decode/e2e/test_eagle_correctness.py
 
 - label: LoRA Test %N # 15min each
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/lora
   - tests/lora
-  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py --ignore=lora/test_minicpmv_tp.py
+  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_chatglm3_tp.py --ignore=lora/test_llama_tp.py
   parallelism: 4
 
-- label: "PyTorch Fullgraph Smoke Test" # 9min
-  fast_check: true
+- label: PyTorch Compilation Unit Tests
+  source_file_dependencies:
+    - vllm/
+    - tests/compile
+  commands:
+    - pytest -v -s compile/test_pass_manager.py
+    - pytest -v -s compile/test_fusion.py
+
+- label: PyTorch Fullgraph Smoke Test # 9min
   source_file_dependencies:
   - vllm/
   - tests/compile
@@ -287,7 +305,7 @@ steps:
   - pytest -v -s compile/piecewise/test_simple.py
   - pytest -v -s compile/piecewise/test_toy_llama.py
 
-- label: "PyTorch Fullgraph Test" # 18min
+- label: PyTorch Fullgraph Test # 18min
   source_file_dependencies:
   - vllm/
   - tests/compile
@@ -295,7 +313,7 @@ steps:
   - pytest -v -s compile/test_full_graph.py
 
 - label: Kernels Test %N # 1h each
-  mirror_hardwares: [amd]
+  # mirror_hardwares: [amd]
   source_file_dependencies:
   - csrc/
   - vllm/attention
@@ -305,7 +323,7 @@ steps:
   parallelism: 4
 
 - label: Tensorizer Test # 11min
-  mirror_hardwares: [amd]
+  # mirror_hardwares: [amd]
   soft_fail: true
   source_file_dependencies:
   - vllm/model_executor/model_loader
@@ -321,7 +339,14 @@ steps:
   source_file_dependencies:
   - benchmarks/
   commands:
-  - bash run-benchmarks.sh
+  - bash scripts/run-benchmarks.sh
+
+- label: Benchmarks CLI Test # 10min
+  source_file_dependencies:
+  - vllm/
+  - tests/benchmarks/
+  commands:
+  - pytest -v -s benchmarks/
 
 - label: Quantization Test # 33min
   source_file_dependencies:
@@ -356,12 +381,14 @@ steps:
 
 - label: OpenAI-Compatible Tool Use # 20 min
   fast_check: false
-  mirror_hardwares: [ amd ]
+  #mirror_hardwares: [ amd ]
   source_file_dependencies:
     - vllm/
     - tests/tool_use
+    - tests/mistral_tool_use
   commands:
     - pytest -v -s tool_use
+    - pytest -v -s mistral_tool_use
 
 #####  models test  #####
 
@@ -372,7 +399,10 @@ steps:
   commands:
     - pytest -v -s models/test_transformers.py
     - pytest -v -s models/test_registry.py
-    - pytest -v -s models/test_initialization.py
+    # V1 Test: https://github.com/vllm-project/vllm/issues/14531
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'not llama4 and not plamo2'
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'llama4'
+    - VLLM_USE_V1=0 pytest -v -s models/test_initialization.py -k 'plamo2'
 
 - label: Language Models Test (Standard) # 32min
   #mirror_hardwares: [amd]
@@ -382,6 +412,8 @@ steps:
   - tests/models/embedding/language
   - tests/models/encoder_decoder/language
   commands:
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install causal-conv1d
     - pytest -v -s models/decoder_only/language -m 'core_model or quant_model'
     - pytest -v -s models/embedding/language -m core_model
 
@@ -393,6 +425,8 @@ steps:
   - tests/models/embedding/language
   - tests/models/encoder_decoder/language
   commands:
+    # Install causal-conv1d for plamo2 models here, as it is not compatible with pip-compile.
+    - pip install causal-conv1d
     - pytest -v -s models/decoder_only/language -m 'not core_model and not quant_model'
     - pytest -v -s models/embedding/language -m 'not core_model'
 
@@ -409,11 +443,12 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/multimodal
     - pytest -v -s models/decoder_only/audio_language -m 'core_model or quant_model'
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'core_model or quant_model'
+    - pytest -v -s models/decoder_only/vision_language -m 'core_model or quant_model'
     - pytest -v -s models/embedding/vision_language -m core_model
     - pytest -v -s models/encoder_decoder/audio_language -m core_model
     - pytest -v -s models/encoder_decoder/language -m core_model
     - pytest -v -s models/encoder_decoder/vision_language -m core_model
+    - pytest -v -s models/decoder_only/vision_language/test_interleaved.py
 
 - label: Multi-Modal Models Test (Extended) 1 # 48m
   optional: true
@@ -427,10 +462,7 @@ steps:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/decoder_only/audio_language -m 'not core_model and not quant_model'
     - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=0) and not core_model and not quant_model'
-    # HACK - run phi3v tests separately to sidestep this transformers bug
-    # https://github.com/huggingface/transformers/issues/34307
-    - pytest -v -s models/decoder_only/vision_language/test_phi3v.py
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_models.py --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
+    - pytest -v -s --ignore models/decoder_only/vision_language/test_models.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
     - pytest -v -s models/embedding/vision_language -m 'not core_model'
     - pytest -v -s models/encoder_decoder/language -m 'not core_model'
     - pytest -v -s models/encoder_decoder/vision_language -m 'not core_model'
@@ -446,6 +478,7 @@ steps:
 
 # This test is used only in PR development phase to test individual models and should never run on main
 - label: Custom Models Test
+  mirror_hardwares: [amd]
   optional: true
   commands:
     - echo 'Testing custom models...'
@@ -457,6 +490,7 @@ steps:
 #####  multi gpus test  #####
 
 - label: Distributed Comm Ops Test # 7min
+  mirror_hardwares: [amd]
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   source_file_dependencies:
@@ -499,9 +533,11 @@ steps:
   - vllm/worker/worker.py
   - vllm/worker/model_runner.py
   - entrypoints/llm/test_collective_rpc.py
+  - tests/v1/test_async_llm_dp.py
+  - vllm/v1/engine/
   commands:
+  - TP_SIZE=1 DP_SIZE=2 pytest -v -s v1/test_async_llm_dp.py
   - pytest -v -s entrypoints/llm/test_collective_rpc.py
-  - torchrun --nproc-per-node=2 distributed/test_torchrun_example.py
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
   - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
@@ -514,13 +550,13 @@ steps:
   # this test fails consistently.
   # TODO: investigate and fix
   # - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
-  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/disagg_test.py
+  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
+  - VLLM_USE_V1=0 CUDA_VISIBLE_DEVICES=0,1 pytest -v -s kv_transfer/test_disagg.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s v1/shutdown
 
 - label: Plugin Tests (2 GPUs) # 40min
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
-  fast_check: true
   source_file_dependencies:
   - vllm/plugins/
   - tests/plugins/
@@ -579,13 +615,10 @@ steps:
     # FIXIT: find out which code initialize cuda before running the test
     # before the fix, we need to use spawn to test it
     - export VLLM_WORKER_MULTIPROC_METHOD=spawn
-    # This test runs llama 13B, so it is required to run on 4 GPUs.
-    - pytest -v -s -x lora/test_long_context.py
     # There is some Tensor Parallelism related processing logic in LoRA that
     # requires multi-GPU testing for validation.
     - pytest -v -s -x lora/test_chatglm3_tp.py
     - pytest -v -s -x lora/test_llama_tp.py
-    - pytest -v -s -x lora/test_minicpmv_tp.py
 
 
 - label: Weight Loading Multiple GPU Test  # 33min

.github/CODEOWNERS

@@ -10,27 +10,32 @@
 /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
+/vllm/model_executor/guided_decoding @mgoin @russellb
 /vllm/multimodal @DarkLight1337 @ywang96
 CMakeLists.txt @tlrmchlsmth
 
 # vLLM V1
 /vllm/v1 @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @comaniac @alexm-redhat
+/vllm/v1/structured_output @mgoin @russellb
 
 # Test ownership
-/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
-/tests/test_inputs.py @DarkLight1337 @ywang96
-/tests/entrypoints @DarkLight1337 @robertgshaw2-redhat @simon-mo
-/tests/models @DarkLight1337 @ywang96
-/tests/multimodal @DarkLight1337 @ywang96
-/tests/prefix_caching @comaniac @KuntaiDu
-/tests/spec_decode @njhill @LiuXiaoxuanPKU
-/tests/kernels @tlrmchlsmth @WoosukKwon
-/tests/quantization @mgoin @robertgshaw2-redhat
 /.buildkite/lm-eval-harness @mgoin @simon-mo
+/tests/async_engine @njhill @robertgshaw2-redhat @simon-mo
+/tests/basic_correctness/test_chunked_prefill @rkooo567 @comaniac
 /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/kernels @tlrmchlsmth @WoosukKwon
+/tests/model_executor/test_guided_processors.py @mgoin @russellb
+/tests/models @DarkLight1337 @ywang96
 /tests/multi_step @alexm-redhat @comaniac
+/tests/multimodal @DarkLight1337 @ywang96
+/tests/prefix_caching @comaniac @KuntaiDu
+/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/weight_loading @mgoin @youkaichao
-/tests/basic_correctness/test_chunked_prefill @rkooo567 @comaniac

.github/ISSUE_TEMPLATE/200-installation.yml

@@ -14,7 +14,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

.github/ISSUE_TEMPLATE/300-usage.yml

@@ -14,7 +14,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

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

@@ -14,7 +14,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

.github/ISSUE_TEMPLATE/600-new-model.yml

@@ -9,7 +9,7 @@ body:
     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+).
 
-      #### We also highly recommend you read https://docs.vllm.ai/en/latest/contributing/model/adding_model.html first to understand how to add a new model.
+      #### We also highly recommend you read https://docs.vllm.ai/en/latest/contributing/model/index.html first to understand how to add a new model.
 - type: textarea
   attributes:
     label: The model to consider.

.github/ISSUE_TEMPLATE/700-performance-discussion.yml

@@ -35,7 +35,7 @@ body:
     description: |
       Please run the following and paste the output below.
       ```sh
-      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/vllm/collect_env.py
       # For security purposes, please feel free to check the contents of collect_env.py before running it.
       python collect_env.py
       ```

.github/ISSUE_TEMPLATE/800-misc-discussion.yml

@@ -1,28 +0,0 @@
-name: 🎲 Misc/random discussions that do not fit into the above categories.
-description: Submit a discussion as you like. Note that developers are heavily overloaded and we mainly rely on community users to answer these issues.
-title: "[Misc]: "
-labels: ["misc"]
-
-body:
-- type: markdown
-  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: textarea
-  attributes:
-    label: Anything you want to discuss about vllm.
-    description: >
-      Anything you want to discuss about vllm.
-  validations:
-    required: true
-- type: markdown
-  attributes:
-    value: >
-      Thanks for contributing 🎉!
-- type: checkboxes
-  id: askllm
-  attributes:
-    label: Before submitting a new issue...
-    options:
-      - label: Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the [documentation page](https://docs.vllm.ai/en/latest/), which can answer lots of frequently asked questions.
-        required: true

.github/ISSUE_TEMPLATE/config.yml

@@ -1 +1,5 @@
 blank_issues_enabled: false
+contact_links:
+  - name: Questions
+    url: https://discuss.vllm.ai
+    about: Ask questions and discuss with other vLLM community members

.github/PULL_REQUEST_TEMPLATE.md

@@ -3,4 +3,4 @@ FILL IN THE PR DESCRIPTION HERE
 FIX #xxxx (*link existing issues this PR will resolve*)
 
 <!--- pyml disable-next-line no-emphasis-as-heading -->
-**BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing/overview.html>**
+**BEFORE SUBMITTING, PLEASE READ <https://docs.vllm.ai/en/latest/contributing/overview.html>** (anything written below this line will be removed by GitHub Actions)

.github/dependabot.yml

@@ -23,7 +23,7 @@ updates:
       - dependency-name: "lm-format-enforcer"
       - dependency-name: "gguf"
       - dependency-name: "compressed-tensors"
-      - dependency-name: "ray[adag]"
+      - dependency-name: "ray[cgraph]" # Ray Compiled Graph
       - dependency-name: "lm-eval"
     groups:
       minor-update:

.github/mergify.yml

@@ -5,6 +5,7 @@ pull_request_rules:
     - or:
       - files~=^[^/]+\.md$
       - files~=^docs/
+      - files~=^examples/
   actions:
     label:
       add:
@@ -18,7 +19,7 @@ pull_request_rules:
       - files~=\.buildkite/
       - files~=^cmake/
       - files=CMakeLists.txt
-      - files~=^Dockerfile
+      - files~=^docker/Dockerfile
       - files~=^requirements.*\.txt
       - files=setup.py
   actions:
@@ -35,6 +36,21 @@ pull_request_rules:
       add:
         - frontend
 
+- name: label-multi-modality
+  description: Automatically apply multi-modality label
+  conditions:
+    - or:
+      - files~=^vllm/multimodal/
+      - files~=^tests/multimodal/
+      - files~=^tests/models/multimodal/
+      - files~=^tests/models/*/audio_language/
+      - files~=^tests/models/*/vision_language/
+      - files=tests/models/test_vision.py
+  actions:
+    label:
+      add:
+        - multi-modality
+
 - name: label-structured-output
   description: Automatically apply structured-output label
   conditions:
@@ -72,6 +88,36 @@ pull_request_rules:
       add:
         - v1
 
+- name: label-tpu
+  description: Automatically apply tpu label
+  # Keep this list in sync with `label-tpu-remove` conditions
+  conditions:
+    - or:
+      - files~=tpu.py
+      - files~=_tpu
+      - files~=tpu_
+      - files~=/tpu/
+      - files~=pallas
+  actions:
+    label:
+      add:
+        - tpu
+
+- name: label-tpu-remove
+  description: Automatically remove tpu label
+  # Keep this list in sync with `label-tpu` conditions
+  conditions:
+    - and:
+      - -files~=tpu.py
+      - -files~=_tpu
+      - -files~=tpu_
+      - -files~=/tpu/
+      - -files~=pallas
+  actions:
+    label:
+      remove:
+        - tpu
+
 - name: ping author on conflicts and add 'needs-rebase' label
   conditions:
       - conflict

.github/workflows/lint-and-deploy.yaml

@@ -12,7 +12,7 @@ jobs:
           fetch-depth: 0
 
       - name: Set up Helm
-        uses: azure/setup-helm@fe7b79cd5ee1e45176fcad797de68ecaf3ca4814 # v4.2.0
+        uses: azure/setup-helm@b9e51907a09c216f16ebe8536097933489208112 # v4.3.0
         with:
           version: v3.14.4
 
@@ -50,7 +50,7 @@ jobs:
         uses: helm/kind-action@a1b0e391336a6ee6713a0583f8c6240d70863de3 # v1.12.0
 
       - name: Build the Docker image vllm cpu
-        run: docker buildx build -f Dockerfile.cpu -t vllm-cpu-env .
+        run: docker buildx build -f docker/Dockerfile.cpu -t vllm-cpu-env .
 
       - name: Configuration of docker images, network and namespace for the kind cluster
         run: |

.github/workflows/publish.yml

@@ -39,7 +39,7 @@ jobs:
             const script = require('.github/workflows/scripts/create_release.js')
             await script(github, context, core)
 
-  # NOTE(simon): No longer build wheel using Github Actions. See buildkite's release workflow. 
+  # NOTE(simon): No longer build wheel using GitHub Actions. See buildkite's release workflow. 
   # wheel:
   #   name: Build Wheel
   #   runs-on: ${{ matrix.os }}
@@ -50,7 +50,7 @@ jobs:
   #     matrix:
   #         os: ['ubuntu-20.04']
   #         python-version: ['3.9', '3.10', '3.11', '3.12']
-  #         pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements-cuda.txt.
+  #         pytorch-version: ['2.4.0']  # Must be the most recent version that meets requirements/cuda.txt.
   #         cuda-version: ['11.8', '12.1']
 
   #   steps:

.github/workflows/scripts/build.sh

@@ -9,7 +9,7 @@ PATH=${cuda_home}/bin:$PATH
 LD_LIBRARY_PATH=${cuda_home}/lib64:$LD_LIBRARY_PATH
 
 # Install requirements
-$python_executable -m pip install -r requirements-build.txt -r requirements-cuda.txt
+$python_executable -m pip install -r requirements/build.txt -r requirements/cuda.txt
 
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1

.github/workflows/scripts/create_release.js

@@ -1,4 +1,4 @@
-// Uses Github's API to create the release and wait for result.
+// Uses GitHub's API to create the release and wait for result.
 // We use a JS script since github CLI doesn't provide a way to wait for the release's creation and returns immediately.
 
 module.exports = async (github, context, core) => {

.gitignore

@@ -2,7 +2,8 @@
 /vllm/_version.py
 
 # vllm-flash-attn built from source
-vllm/vllm_flash_attn/
+vllm/vllm_flash_attn/*
+!vllm/vllm_flash_attn/fa_utils.py
 
 # Byte-compiled / optimized / DLL files
 __pycache__/
@@ -197,8 +198,11 @@ _build/
 hip_compat.h
 
 # Benchmark dataset
-benchmarks/*.json
+benchmarks/**/*.json
 
 # Linting
 actionlint
 shellcheck*/
+
+# Ingore moe/marlin_moe gen code
+csrc/moe/marlin_moe_wna16/kernel_*

.pre-commit-config.yaml

@@ -1,20 +1,21 @@
+default_install_hook_types:
+  - pre-commit
+  - commit-msg
 default_stages:
   - pre-commit # Run locally
   - manual # Run in CI
+exclude: 'vllm/third_party/.*'
 repos:
 - repo: https://github.com/google/yapf
   rev: v0.43.0
   hooks:
   - id: yapf
     args: [--in-place, --verbose]
-    additional_dependencies: [toml] # TODO: Remove when yapf is upgraded
-    exclude: 'vllm/third_party/.*'
 - repo: https://github.com/astral-sh/ruff-pre-commit
   rev: v0.9.3
   hooks:
   - id: ruff
     args: [--output-format, github, --fix]
-    exclude: 'vllm/third_party/.*'
 - repo: https://github.com/codespell-project/codespell
   rev: v2.4.0
   hooks:
@@ -22,10 +23,9 @@ repos:
     additional_dependencies: ['tomli']
     args: ['--toml', 'pyproject.toml']
 - repo: https://github.com/PyCQA/isort
-  rev: 5.13.2
+  rev: 0a0b7a830386ba6a31c2ec8316849ae4d1b8240d # 6.0.0
   hooks:
   - id: isort
-    exclude: 'vllm/third_party/.*'
 - repo: https://github.com/pre-commit/mirrors-clang-format
   rev: v19.1.7
   hooks:
@@ -38,12 +38,16 @@ repos:
   hooks:
   - id: pymarkdown
     args: [fix]
-    exclude: 'vllm/third_party/.*'
 - repo: https://github.com/rhysd/actionlint
   rev: v1.7.7
   hooks:
   - id: actionlint
-    exclude: 'vllm/third_party/.*'
+- repo: https://github.com/astral-sh/uv-pre-commit
+  rev: 0.6.2
+  hooks:
+    - id: pip-compile
+      args: [requirements/test.in, -o, requirements/test.txt]
+      files: ^requirements/test\.(in|txt)$
 - repo: local
   hooks:
   - id: mypy-local
@@ -51,9 +55,8 @@ repos:
     entry: tools/mypy.sh 0 "local"
     language: python
     types: [python]
-    additional_dependencies: &mypy_deps [mypy==1.11.1, types-setuptools, types-PyYAML, types-requests]
+    additional_dependencies: &mypy_deps [mypy==1.11.1, types-cachetools, types-setuptools, types-PyYAML, types-requests]
     stages: [pre-commit] # Don't run in CI
-    exclude: 'vllm/third_party/.*'
   - id: mypy-3.9 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.9
     entry: tools/mypy.sh 1 "3.9"
@@ -61,7 +64,6 @@ repos:
     types: [python]
     additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
-    exclude: 'vllm/third_party/.*'
   - id: mypy-3.10 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.10
     entry: tools/mypy.sh 1 "3.10"
@@ -69,7 +71,6 @@ repos:
     types: [python]
     additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
-    exclude: 'vllm/third_party/.*'
   - id: mypy-3.11 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.11
     entry: tools/mypy.sh 1 "3.11"
@@ -77,7 +78,6 @@ repos:
     types: [python]
     additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
-    exclude: 'vllm/third_party/.*'
   - id: mypy-3.12 # TODO: Use https://github.com/pre-commit/mirrors-mypy when mypy setup is less awkward
     name: Run mypy for Python 3.12
     entry: tools/mypy.sh 1 "3.12"
@@ -85,19 +85,16 @@ repos:
     types: [python]
     additional_dependencies: *mypy_deps
     stages: [manual] # Only run in CI
-    exclude: 'vllm/third_party/.*'
   - id: shellcheck
     name: Lint shell scripts
     entry: tools/shellcheck.sh
     language: script
     types: [shell]
-    exclude: 'vllm/third_party/.*'
   - id: png-lint
     name: Lint PNG exports from excalidraw
     entry: tools/png-lint.sh
     language: script
     types: [png]
-    exclude: 'vllm/third_party/.*'
   - id: signoff-commit
     name: Sign-off Commit
     entry: bash
@@ -110,13 +107,11 @@ repos:
     language: system
     verbose: true
     stages: [commit-msg]
-    exclude: 'vllm/third_party/.*'
   - id: check-spdx-header
     name: Check SPDX headers
     entry: python tools/check_spdx_header.py
     language: python
     types: [python]
-    exclude: 'vllm/third_party/.*'
   - id: check-filenames
     name: Check for spaces in all filenames
     entry: bash
@@ -126,7 +121,12 @@ repos:
     language: system
     always_run: true
     pass_filenames: false
-    exclude: 'vllm/third_party/.*'
+  - id: update-dockerfile-graph
+    name: Update Dockerfile dependency graph
+    entry: tools/update-dockerfile-graph.sh
+    language: script
+    files: ^docker/Dockerfile$
+    pass_filenames: false
   # Keep `suggestion` last
   - id: suggestion
     name: Suggestion
@@ -134,5 +134,4 @@ repos:
     language: system
     verbose: true
     pass_filenames: false
-    exclude: 'vllm/third_party/.*'
   # Insert new entries above the `suggestion` entry

.readthedocs.yaml

@@ -18,4 +18,4 @@ formats: []
 # Optionally declare the Python requirements required to build your docs
 python:
   install:
-    - requirements: docs/requirements-docs.txt
+    - requirements: requirements/docs.txt

CMakeLists.txt

@@ -31,10 +31,10 @@ set(ignoreMe "${VLLM_PYTHON_PATH}")
 set(PYTHON_SUPPORTED_VERSIONS "3.9" "3.10" "3.11" "3.12")
 
 # Supported NVIDIA architectures.
-set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0")
+set(CUDA_SUPPORTED_ARCHS "7.0;7.2;7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -44,10 +44,10 @@ set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101")
 #
 # Note: the CUDA torch version is derived from pyproject.toml and various
 # requirements.txt files and should be kept consistent.  The ROCm torch
-# versions are derived from Dockerfile.rocm
+# versions are derived from docker/Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.5.1")
-set(TORCH_SUPPORTED_VERSION_ROCM "2.5.1")
+set(TORCH_SUPPORTED_VERSION_CUDA "2.6.0")
+set(TORCH_SUPPORTED_VERSION_ROCM "2.6.0")
 
 #
 # Try to find python package with an executable that exactly matches
@@ -174,6 +174,25 @@ 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
+  #
+  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.
+  #
+  set(CMAKE_${VLLM_GPU_LANG}_FLAGS "${CMAKE_${VLLM_GPU_LANG}_FLAGS} -Wno-unused-result")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-unused-result")
+endif()
+
 #
 # Define other extension targets
 #
@@ -211,10 +230,12 @@ set(VLLM_EXT_SRC
   "csrc/cache_kernels.cu"
   "csrc/attention/paged_attention_v1.cu"
   "csrc/attention/paged_attention_v2.cu"
+  "csrc/attention/merge_attn_states.cu"
   "csrc/pos_encoding_kernels.cu"
   "csrc/activation_kernels.cu"
   "csrc/layernorm_kernels.cu"
   "csrc/layernorm_quant_kernels.cu"
+  "csrc/cuda_view.cu"
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
   "csrc/quantization/fp8/common.cu"
@@ -222,6 +243,7 @@ set(VLLM_EXT_SRC
   "csrc/quantization/gguf/gguf_kernel.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/prepare_inputs/advance_step.cu"
+  "csrc/custom_all_reduce.cu"
   "csrc/torch_bindings.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
@@ -229,7 +251,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
 
   # Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
   # Please keep this in sync with FetchContent_Declare line below.
-  set(CUTLASS_REVISION "v3.7.0" CACHE STRING "CUTLASS revision to use")
+  set(CUTLASS_REVISION "v3.8.0" CACHE STRING "CUTLASS revision to use")
 
   # Use the specified CUTLASS source directory for compilation if VLLM_CUTLASS_SRC_DIR is provided
   if (DEFINED ENV{VLLM_CUTLASS_SRC_DIR})
@@ -247,7 +269,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
         cutlass
         GIT_REPOSITORY https://github.com/nvidia/cutlass.git
         # Please keep this in sync with CUTLASS_REVISION line above.
-        GIT_TAG v3.7.0
+        GIT_TAG v3.8.0
         GIT_PROGRESS TRUE
 
         # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
@@ -263,10 +285,10 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/mamba/causal_conv1d/causal_conv1d.cu"
     "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
-    "csrc/custom_all_reduce.cu"
     "csrc/permute_cols.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
     "csrc/quantization/fp4/nvfp4_quant_entry.cu"
+    "csrc/quantization/fp4/nvfp4_scaled_mm_entry.cu"
     "csrc/sparse/cutlass/sparse_scaled_mm_entry.cu"
     "csrc/cutlass_extensions/common.cpp")
 
@@ -277,7 +299,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # Only build Marlin kernels if we are building for at least some compatible archs.
   # Keep building Marlin for 9.0 as there are some group sizes and shapes that
   # are not supported by Machete yet.
-  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
+  cuda_archs_loose_intersection(MARLIN_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
   if (MARLIN_ARCHS)
     set(MARLIN_SRCS
        "csrc/quantization/fp8/fp8_marlin.cu"
@@ -297,43 +319,87 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
                    " in CUDA target architectures")
   endif()
 
+  # Only build AllSpark kernels if we are building for at least some compatible archs.
+  cuda_archs_loose_intersection(ALLSPARK_ARCHS "8.0;8.6;8.7;8.9" "${CUDA_ARCHS}")
+  if (ALLSPARK_ARCHS)
+    set(ALLSPARK_SRCS
+       "csrc/quantization/gptq_allspark/allspark_repack.cu"
+       "csrc/quantization/gptq_allspark/allspark_qgemm_w8a16.cu")
+    set_gencode_flags_for_srcs(
+      SRCS "${ALLSPARK_SRCS}"
+      CUDA_ARCHS "${ALLSPARK_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${ALLSPARK_SRCS}")
+    message(STATUS "Building AllSpark kernels for archs: ${ALLSPARK_ARCHS}")
+  else()
+    message(STATUS "Not building AllSpark kernels as no compatible archs found"
+                   " in CUDA target architectures")
+  endif()
+
+
+  set(SCALED_MM_3X_ARCHS)
   # The cutlass_scaled_mm kernels for Hopper (c3x, i.e. CUTLASS 3.x) require
-  # CUDA 12.0 or later (and only work on Hopper, 9.0a for now).
-  cuda_archs_loose_intersection(SCALED_MM_3X_ARCHS "9.0a" "${CUDA_ARCHS}")
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
+  # CUDA 12.0 or later
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
     set(SRCS
-       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x.cu"
+       "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm90.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_fp8.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm90_int8.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_azp_sm90_int8.cu"
        "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_blockwise_sm90_fp8.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
-    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_C3X=1")
-    message(STATUS "Building scaled_mm_c3x for archs: ${SCALED_MM_3X_ARCHS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM90=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm90 for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_3X_ARCHS)
-      message(STATUS "Not building scaled_mm_c3x as CUDA Compiler version is "
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building scaled_mm_c3x_sm90 as CUDA Compiler version is "
                      "not >= 12.0, we recommend upgrading to CUDA 12.0 or "
                      "later if you intend on running FP8 quantized models on "
                      "Hopper.")
     else()
-      message(STATUS "Not building scaled_mm_c3x as no compatible archs found "
+      message(STATUS "Not building scaled_mm_c3x_sm90 as no compatible archs found "
                      "in CUDA target architectures")
     endif()
+  endif()
 
-    # clear SCALED_MM_3X_ARCHS so the scaled_mm_c2x kernels know we didn't
-    # build any 3x kernels
-    set(SCALED_MM_3X_ARCHS)
+  # The cutlass_scaled_mm kernels for Blackwell (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" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND SCALED_MM_ARCHS)
+    set(SRCS
+      "csrc/quantization/cutlass_w8a8/scaled_mm_c3x_sm100.cu"
+      "csrc/quantization/cutlass_w8a8/c3x/scaled_mm_sm100_fp8.cu"
+    )
+    set_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_SCALED_MM_SM100=1")
+    # Let scaled_mm_c2x know it doesn't need to build these arches
+    list(APPEND SCALED_MM_3X_ARCHS "${SCALED_MM_ARCHS}")
+    message(STATUS "Building scaled_mm_c3x_sm100 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_sm100 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_100 as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
   endif()
 
   #
   # For the cutlass_scaled_mm kernels we want to build the c2x (CUTLASS 2.x)
   # kernels for the remaining archs that are not already built for 3x.
   cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
-    "7.5;8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
+    "7.5;8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${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)
@@ -358,17 +424,18 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # 2:4 Sparse Kernels
 
   # The 2:4 sparse kernels cutlass_scaled_sparse_mm and cutlass_compressor
-  # require CUDA 12.2 or later (and only work on Hopper, 9.0a for now).
-  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
+  # require CUDA 12.2 or later (and only work on Hopper).
+  cuda_archs_loose_intersection(SCALED_MM_ARCHS "9.0a;" "${CUDA_ARCHS}")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
     set(SRCS "csrc/sparse/cutlass/sparse_scaled_mm_c3x.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
-      CUDA_ARCHS "${SCALED_MM_3X_ARCHS}")
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
     list(APPEND VLLM_EXT_SRC "${SRCS}")
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_SPARSE_SCALED_MM_C3X=1")
-    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_3X_ARCHS}")
+    message(STATUS "Building sparse_scaled_mm_c3x for archs: ${SCALED_MM_ARCHS}")
   else()
-    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_3X_ARCHS)
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.2 AND SCALED_MM_ARCHS)
       message(STATUS "Not building sparse_scaled_mm_c3x kernels as CUDA Compiler version is "
                      "not >= 12.2, we recommend upgrading to CUDA 12.2 or later "
                      "if you intend on running FP8 sparse quantized models on Hopper.")
@@ -383,7 +450,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.8 AND FP4_ARCHS)
     set(SRCS
       "csrc/quantization/fp4/nvfp4_quant_kernels.cu"
-    )
+      "csrc/quantization/fp4/nvfp4_scaled_mm_kernels.cu")
     set_gencode_flags_for_srcs(
       SRCS "${SRCS}"
       CUDA_ARCHS "${FP4_ARCHS}")
@@ -396,6 +463,33 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     set(FP4_ARCHS)
   endif()
 
+  #
+  # 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;" "${CUDA_ARCHS}")
+  if(${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_gencode_flags_for_srcs(
+      SRCS "${SRCS}"
+      CUDA_ARCHS "${SCALED_MM_ARCHS}")
+    list(APPEND VLLM_EXT_SRC "${SRCS}")
+    list(APPEND VLLM_GPU_FLAGS "-DENABLE_CUTLASS_MOE_SM90=1")
+    message(STATUS "Building grouped_mm_c3x for archs: ${SCALED_MM_ARCHS}")
+  else()
+    if (NOT ${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.3 AND SCALED_MM_ARCHS)
+      message(STATUS "Not building grouped_mm_c3x kernels 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.")
+    else()
+      message(STATUS "Not building grouped_mm_c3x as no compatible archs found "
+                     "in CUDA target architectures")
+    endif()
+  endif()
+
   #
   # Machete kernels
 
@@ -477,6 +571,7 @@ define_gpu_extension_target(
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
   INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR}
+  INCLUDE_DIRECTORIES ${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
   USE_SABI 3
   WITH_SOABI)
 
@@ -495,28 +590,70 @@ set(VLLM_MOE_EXT_SRC
   "csrc/moe/moe_align_sum_kernels.cu"
   "csrc/moe/topk_softmax_kernels.cu")
 
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  list(APPEND VLLM_MOE_EXT_SRC "csrc/moe/moe_wna16.cu")
+endif()
+
 set_gencode_flags_for_srcs(
   SRCS "${VLLM_MOE_EXT_SRC}"
   CUDA_ARCHS "${CUDA_ARCHS}")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
-  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
-  if (MARLIN_MOE_ARCHS)
-    set(MARLIN_MOE_SRC
-        "csrc/moe/marlin_kernels/marlin_moe_kernel.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4b8.cu"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku8b128.cu"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.h"
-        "csrc/moe/marlin_kernels/marlin_moe_kernel_ku4.cu"
-        "csrc/moe/marlin_moe_ops.cu")
+  set(VLLM_MOE_WNA16_SRC
+    "csrc/moe/moe_wna16.cu")
 
+  set_gencode_flags_for_srcs(
+    SRCS "${VLLM_MOE_WNA16_SRC}"
+    CUDA_ARCHS "${CUDA_ARCHS}")
+
+  list(APPEND VLLM_MOE_EXT_SRC "${VLLM_MOE_WNA16_SRC}")
+  cuda_archs_loose_intersection(MARLIN_MOE_ARCHS "8.0;8.6;8.7;8.9;9.0;10.0;10.1;12.0" "${CUDA_ARCHS}")
+  if (MARLIN_MOE_ARCHS)
+
+    #
+    # For the Marlin MOE kernels we automatically generate sources for various
+    # preselected input type pairs and schedules.
+    # Generate sources:
+    set(MOE_MARLIN_GEN_SCRIPT
+      ${CMAKE_CURRENT_SOURCE_DIR}/csrc/moe/marlin_moe_wna16/generate_kernels.py)
+    file(MD5 ${MOE_MARLIN_GEN_SCRIPT} MOE_MARLIN_GEN_SCRIPT_HASH)
+
+    message(STATUS "Marlin MOE generation script hash: ${MOE_MARLIN_GEN_SCRIPT_HASH}")
+    message(STATUS "Last run Marlin MOE generate script hash: $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}")
+
+    if (NOT DEFINED CACHE{MOE_MARLIN_GEN_SCRIPT_HASH}
+        OR NOT $CACHE{MOE_MARLIN_GEN_SCRIPT_HASH} STREQUAL ${MOE_MARLIN_GEN_SCRIPT_HASH})
+      execute_process(
+        COMMAND ${CMAKE_COMMAND} -E env
+        PYTHONPATH=${CMAKE_CURRENT_SOURCE_DIR}/csrc/cutlass_extensions/:${CUTLASS_DIR}/python/:${VLLM_PYTHON_PATH}:$PYTHONPATH
+          ${Python_EXECUTABLE} ${MOE_MARLIN_GEN_SCRIPT}
+        RESULT_VARIABLE moe_marlin_generation_result
+        OUTPUT_VARIABLE moe_marlin_generation_output
+        OUTPUT_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
+        ERROR_FILE ${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log
+      )
+
+      if (NOT moe_marlin_generation_result EQUAL 0)
+        message(FATAL_ERROR "Marlin MOE generation failed."
+                            " Result: \"${moe_marlin_generation_result}\""
+                            "\nCheck the log for details: "
+                            "${CMAKE_CURRENT_BINARY_DIR}/moe_marlin_generation.log")
+      else()
+        set(MOE_MARLIN_GEN_SCRIPT_HASH ${MOE_MARLIN_GEN_SCRIPT_HASH}
+            CACHE STRING "Last run Marlin MOE generate script hash" FORCE)
+        message(STATUS "Marlin MOE generation completed successfully.")
+      endif()
+    else()
+      message(STATUS "Marlin MOE generation script has not changed, skipping generation.")
+    endif()
+
+    file(GLOB MOE_WNAA16_MARLIN_SRC "csrc/moe/marlin_moe_wna16/*.cu")
     set_gencode_flags_for_srcs(
-      SRCS "${MARLIN_MOE_SRC}"
+      SRCS "${MOE_WNAA16_MARLIN_SRC}"
       CUDA_ARCHS "${MARLIN_MOE_ARCHS}")
 
-    list(APPEND VLLM_MOE_EXT_SRC "${MARLIN_MOE_SRC}")
+    list(APPEND VLLM_MOE_EXT_SRC ${MOE_WNAA16_MARLIN_SRC})
+
     message(STATUS "Building Marlin MOE kernels for archs: ${MARLIN_MOE_ARCHS}")
   else()
     message(STATUS "Not building Marlin MOE kernels as no compatible archs found"
@@ -554,77 +691,8 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
     WITH_SOABI)
 endif()
 
-# vllm-flash-attn currently only supported on CUDA
-if (NOT VLLM_GPU_LANG STREQUAL "CUDA")
-  return()
+# For CUDA we also build and ship some external projects.
+if (VLLM_GPU_LANG STREQUAL "CUDA")
+    include(cmake/external_projects/flashmla.cmake)
+    include(cmake/external_projects/vllm_flash_attn.cmake)
 endif ()
-
-# vLLM flash attention requires VLLM_GPU_ARCHES to contain the set of target
-# arches in the CMake syntax (75-real, 89-virtual, etc), since we clear the
-# arches in the CUDA case (and instead set the gencodes on a per file basis)
-# we need to manually set VLLM_GPU_ARCHES here.
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  foreach(_ARCH ${CUDA_ARCHS})
-    string(REPLACE "." "" _ARCH "${_ARCH}")
-    list(APPEND VLLM_GPU_ARCHES "${_ARCH}-real")
-  endforeach()
-endif()
-
-#
-# Build vLLM flash attention from source
-#
-# IMPORTANT: This has to be the last thing we do, because vllm-flash-attn uses the same macros/functions as vLLM.
-# Because functions all belong to the global scope, vllm-flash-attn's functions overwrite vLLMs.
-# They should be identical but if they aren't, this is a massive footgun.
-#
-# The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
-# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3).
-# If no component is specified, vllm-flash-attn is still installed.
-
-# If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
-# This is to enable local development of vllm-flash-attn within vLLM.
-# It can be set as an environment variable or passed as a cmake argument.
-# The environment variable takes precedence.
-if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
-  set(VLLM_FLASH_ATTN_SRC_DIR $ENV{VLLM_FLASH_ATTN_SRC_DIR})
-endif()
-
-if(VLLM_FLASH_ATTN_SRC_DIR)
-  FetchContent_Declare(
-          vllm-flash-attn SOURCE_DIR 
-          ${VLLM_FLASH_ATTN_SRC_DIR}
-          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
-  )
-else()
-  FetchContent_Declare(
-          vllm-flash-attn
-          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
-          GIT_TAG 720c94869cf2e0ff5a706e9c7f1dce0939686ade
-          GIT_PROGRESS TRUE
-          # Don't share the vllm-flash-attn build between build types
-          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
-  )
-endif()
-
-
-# Fetch the vllm-flash-attn library
-FetchContent_MakeAvailable(vllm-flash-attn)
-message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
-
-# 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
-  COMPONENT _vllm_fa2_C
-  FILES_MATCHING PATTERN "*.py"
-)
-
-install(
-  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
-  DESTINATION vllm_flash_attn
-  COMPONENT _vllm_fa3_C
-  FILES_MATCHING PATTERN "*.py"
-)
-
-# Nothing after vllm-flash-attn, see comment about macros above

Dockerfile.cpu

@@ -1,69 +0,0 @@
-# This vLLM Dockerfile is used to construct image that can build and run vLLM on x86 CPU platform.
-
-FROM ubuntu:22.04 AS cpu-test-1
-
-ENV CCACHE_DIR=/root/.cache/ccache
-
-ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
-
-RUN --mount=type=cache,target=/var/cache/apt \
-    apt-get update -y \
-    && apt-get install -y curl ccache git wget vim numactl gcc-12 g++-12 python3 python3-pip libtcmalloc-minimal4 libnuma-dev \
-    && apt-get install -y ffmpeg libsm6 libxext6 libgl1 \
-    && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
-
-# https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/performance_tuning/tuning_guide.html
-# intel-openmp provides additional performance improvement vs. openmp
-# tcmalloc provides better memory allocation efficiency, e.g, holding memory in caches to speed up access of commonly-used objects.
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install intel-openmp==2025.0.1
-
-ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/usr/local/lib/libiomp5.so"
-
-RUN echo 'ulimit -c 0' >> ~/.bashrc
-
-RUN pip install intel_extension_for_pytorch==2.5.0
-
-WORKDIR /workspace
-
-ARG PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu"
-ENV PIP_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements-build.txt,target=requirements-build.txt \
-    pip install --upgrade pip && \
-    pip install -r requirements-build.txt
-
-FROM cpu-test-1 AS build
-
-WORKDIR /workspace/vllm
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,src=requirements-common.txt,target=requirements-common.txt \
-    --mount=type=bind,src=requirements-cpu.txt,target=requirements-cpu.txt \
-    pip install -v -r requirements-cpu.txt
-
-COPY . .
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
-
-# Support for building with non-AVX512 vLLM: docker build --build-arg VLLM_CPU_DISABLE_AVX512="true" ...
-ARG VLLM_CPU_DISABLE_AVX512
-ENV VLLM_CPU_DISABLE_AVX512=${VLLM_CPU_DISABLE_AVX512}
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=cache,target=/root/.cache/ccache \
-    --mount=type=bind,source=.git,target=.git \
-    VLLM_TARGET_DEVICE=cpu python3 setup.py bdist_wheel && \
-    pip install dist/*.whl && \
-    rm -rf dist
-
-WORKDIR /workspace/
-
-RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
-
-# install development dependencies (for testing)
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -e tests/vllm_test_utils
-
-ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.openvino

@@ -1,29 +0,0 @@
-# The vLLM Dockerfile is used to construct vLLM image that can be directly used
-# to run the OpenAI compatible server.
-
-FROM ubuntu:22.04 AS dev
-
-RUN apt-get update -y && \
-    apt-get install -y \
-        git python3-pip \
-        ffmpeg libsm6 libxext6 libgl1
-WORKDIR /workspace
-
-COPY . .
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh ; fi
-
-RUN python3 -m pip install -U pip
-# install build requirements
-RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" python3 -m pip install -r /workspace/requirements-build.txt
-# build vLLM with OpenVINO backend
-RUN PIP_EXTRA_INDEX_URL="https://download.pytorch.org/whl/cpu" VLLM_TARGET_DEVICE="openvino" python3 -m pip install /workspace
-
-COPY examples/ /workspace/examples
-COPY benchmarks/ /workspace/benchmarks
-
-# install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
-
-CMD ["/bin/bash"]

Dockerfile.ppc64le

@@ -1,37 +0,0 @@
-FROM mambaorg/micromamba
-ARG MAMBA_DOCKERFILE_ACTIVATE=1
-USER root
-
-ENV PATH="/usr/local/cargo/bin:$PATH:/opt/conda/bin/"
-
-RUN apt-get update -y && apt-get install -y git wget kmod curl vim libnuma-dev libsndfile-dev libprotobuf-dev build-essential ffmpeg libsm6 libxext6 libgl1 libssl-dev 
-
-# Some packages in requirements-cpu are installed here
-# IBM provides optimized packages for ppc64le processors in the open-ce project for mamba
-# Currently these may not be available for venv or pip directly
-RUN micromamba install -y -n base -c https://ftp.osuosl.org/pub/open-ce/1.11.0-p10/ -c defaults python=3.10 rust && micromamba clean --all --yes
-
-COPY ./ /workspace/vllm
-
-WORKDIR /workspace/vllm
-ARG GIT_REPO_CHECK=0
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
-
-RUN --mount=type=cache,target=/root/.cache/pip  \
-    RUSTFLAGS='-L /opt/conda/lib' pip install -v --prefer-binary --extra-index-url https://repo.fury.io/mgiessing \
-        'cmake>=3.26' ninja packaging 'setuptools-scm>=8' wheel jinja2 \
-        -r requirements-cpu.txt \
-        xformers uvloop==0.20.0
-
-RUN --mount=type=bind,source=.git,target=.git \
-    VLLM_TARGET_DEVICE=cpu python3 setup.py install
-
-# install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
-
-WORKDIR /workspace/
-
-RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
-
-ENTRYPOINT ["/opt/conda/bin/python3", "-m", "vllm.entrypoints.openai.api_server"]

Dockerfile.xpu

@@ -1,69 +0,0 @@
-FROM intel/oneapi-basekit:2024.2.1-0-devel-ubuntu22.04 AS vllm-base
-
-RUN wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | tee /usr/share/keyrings/intel-oneapi-archive-keyring.gpg > /dev/null && \
-    echo "deb [signed-by=/usr/share/keyrings/intel-oneapi-archive-keyring.gpg] https://apt.repos.intel.com/oneapi all main " | tee /etc/apt/sources.list.d/oneAPI.list && \
-    chmod 644 /usr/share/keyrings/intel-oneapi-archive-keyring.gpg && \
-    wget -O- https://repositories.intel.com/graphics/intel-graphics.key | gpg --dearmor | tee /usr/share/keyrings/intel-graphics.gpg > /dev/null && \
-    echo "deb [arch=amd64,i386 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/graphics/ubuntu jammy arc" | tee /etc/apt/sources.list.d/intel.gpu.jammy.list && \
-    chmod 644 /usr/share/keyrings/intel-graphics.gpg
-
-RUN apt-get update -y && \
-    apt-get install -y --no-install-recommends --fix-missing \
-    curl \
-    ffmpeg \
-    git \
-    libsndfile1 \
-    libsm6 \
-    libxext6 \
-    libgl1 \
-    lsb-release \
-    numactl \
-    python3 \
-    python3-dev \
-    python3-pip \
-    # vim \
-    wget
-
-WORKDIR /workspace/vllm
-COPY requirements-xpu.txt /workspace/vllm/requirements-xpu.txt
-COPY requirements-common.txt /workspace/vllm/requirements-common.txt
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install --no-cache-dir \
-    -r requirements-xpu.txt
-
-RUN git clone https://github.com/intel/pti-gpu && \
-    cd pti-gpu/sdk && \
-    git checkout 6c491f07a777ed872c2654ca9942f1d0dde0a082 && \
-    mkdir build && \
-    cd build && \
-    cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_TOOLCHAIN_FILE=../cmake/toolchains/icpx_toolchain.cmake -DBUILD_TESTING=OFF .. && \
-    make -j && \
-    cmake --install . --config Release --prefix "/usr/local"
-
-ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/lib/"
-
-COPY . .
-ARG GIT_REPO_CHECK
-RUN --mount=type=bind,source=.git,target=.git \
-    if [ "$GIT_REPO_CHECK" != 0 ]; then bash tools/check_repo.sh; fi
-
-ENV VLLM_TARGET_DEVICE=xpu
-
-RUN --mount=type=cache,target=/root/.cache/pip \
-    --mount=type=bind,source=.git,target=.git \
-    python3 setup.py install
-
-CMD ["/bin/bash"]
-
-FROM vllm-base AS vllm-openai
-
-# install additional dependencies for openai api server
-RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install accelerate hf_transfer 'modelscope!=1.15.0'
-
-ENV VLLM_USAGE_SOURCE production-docker-image \
-    TRITON_XPU_PROFILE 1
-# install development dependencies (for testing)
-RUN python3 -m pip install -e tests/vllm_test_utils
-ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

MANIFEST.in

@@ -1,9 +1,9 @@
 include LICENSE
-include requirements-common.txt
-include requirements-cuda.txt
-include requirements-rocm.txt
-include requirements-neuron.txt
-include requirements-cpu.txt
+include requirements/common.txt
+include requirements/cuda.txt
+include requirements/rocm.txt
+include requirements/neuron.txt
+include requirements/cpu.txt
 include CMakeLists.txt
 
 recursive-include cmake *

README.md

@@ -10,20 +10,24 @@ Easy, fast, and cheap LLM serving for everyone
 </h3>
 
 <p align="center">
-| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://vllm.ai"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
+| <a href="https://docs.vllm.ai"><b>Documentation</b></a> | <a href="https://blog.vllm.ai/"><b>Blog</b></a> | <a href="https://arxiv.org/abs/2309.06180"><b>Paper</b></a> | <a href="https://x.com/vllm_project"><b>Twitter/X</b></a> | <a href="https://discuss.vllm.ai"><b>User Forum</b></a> | <a href="https://slack.vllm.ai"><b>Developer Slack</b></a> |
 </p>
 
 ---
 
-We are excited to invite you to our Menlo Park meetup with Meta, evening of Thursday, February 27! Meta engineers will discuss the improvements on top of vLLM, and vLLM contributors will share updates from the v0.7.x series of releases. [Register Now](https://lu.ma/h7g3kuj9)
-
----
-
 *Latest News* 🔥
-
+- [2025/04] We hosted [Asia Developer Day](https://www.sginnovate.com/event/limited-availability-morning-evening-slots-remaining-inaugural-vllm-asia-developer-day)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/19cp6Qu8u48ihB91A064XfaXruNYiBOUKrBxAmDOllOo/edit?usp=sharing).
+- [2025/03] We hosted [vLLM x Ollama Inference Night](https://lu.ma/vllm-ollama)! Please find the meetup slides from the vLLM team [here](https://docs.google.com/presentation/d/16T2PDD1YwRnZ4Tu8Q5r6n53c5Lr5c73UV9Vd2_eBo4U/edit?usp=sharing).
+- [2025/03] We hosted [the first vLLM China Meetup](https://mp.weixin.qq.com/s/n77GibL2corAtQHtVEAzfg)! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1REHvfQMKGnvz6p3Fd23HhSO4c8j5WPGZV0bKYLwnHyQ/edit?usp=sharing).
+- [2025/03] We hosted [the East Coast vLLM Meetup](https://lu.ma/7mu4k4xx)! Please find the meetup slides [here](https://docs.google.com/presentation/d/1NHiv8EUFF1NLd3fEYODm56nDmL26lEeXCaDgyDlTsRs/edit#slide=id.g31441846c39_0_0).
+- [2025/02] We hosted [the ninth vLLM meetup](https://lu.ma/h7g3kuj9) with Meta! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1jzC_PZVXrVNSFVCW-V4cFXb6pn7zZ2CyP_Flwo05aqg/edit?usp=sharing) and AMD [here](https://drive.google.com/file/d/1Zk5qEJIkTmlQ2eQcXQZlljAx3m9s7nwn/view?usp=sharing). The slides from Meta will not be posted.
 - [2025/01] We are excited to announce the alpha release of vLLM V1: A major architectural upgrade with 1.7x speedup! Clean code, optimized execution loop, zero-overhead prefix caching, enhanced multimodal support, and more. Please check out our blog post [here](https://blog.vllm.ai/2025/01/27/v1-alpha-release.html).
 - [2025/01] We hosted [the eighth vLLM meetup](https://lu.ma/zep56hui) with Google Cloud! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1epVkt4Zu8Jz_S5OhEHPc798emsYh2BwYfRuDDVEF7u4/edit?usp=sharing), and Google Cloud team [here](https://drive.google.com/file/d/1h24pHewANyRL11xy5dXUbvRC9F9Kkjix/view?usp=sharing).
 - [2024/12] vLLM joins [pytorch ecosystem](https://pytorch.org/blog/vllm-joins-pytorch)! Easy, Fast, and Cheap LLM Serving for Everyone!
+
+<details>
+<summary>Previous News</summary>
+
 - [2024/11] We hosted [the seventh vLLM meetup](https://lu.ma/h0qvrajz) with Snowflake! Please find the meetup slides from vLLM team [here](https://docs.google.com/presentation/d/1e3CxQBV3JsfGp30SwyvS3eM_tW-ghOhJ9PAJGK6KR54/edit?usp=sharing), and Snowflake team [here](https://docs.google.com/presentation/d/1qF3RkDAbOULwz9WK5TOltt2fE9t6uIc_hVNLFAaQX6A/edit?usp=sharing).
 - [2024/10] We have just created a developer slack ([slack.vllm.ai](https://slack.vllm.ai)) focusing on coordinating contributions and discussing features. Please feel free to join us there!
 - [2024/10] Ray Summit 2024 held a special track for vLLM! Please find the opening talk slides from the vLLM team [here](https://docs.google.com/presentation/d/1B_KQxpHBTRa_mDF-tR6i8rWdOU5QoTZNcEg2MKZxEHM/edit?usp=sharing). Learn more from the [talks](https://www.youtube.com/playlist?list=PLzTswPQNepXl6AQwifuwUImLPFRVpksjR) from other vLLM contributors and users!
@@ -37,8 +41,9 @@ We are excited to invite you to our Menlo Park meetup with Meta, evening of Thur
 - [2023/08] We would like to express our sincere gratitude to [Andreessen Horowitz](https://a16z.com/2023/08/30/supporting-the-open-source-ai-community/) (a16z) for providing a generous grant to support the open-source development and research of vLLM.
 - [2023/06] We officially released vLLM! FastChat-vLLM integration has powered [LMSYS Vicuna and Chatbot Arena](https://chat.lmsys.org) since mid-April. Check out our [blog post](https://vllm.ai).
 
----
+</details>
 
+---
 ## About
 
 vLLM is a fast and easy-to-use library for LLM inference and serving.
@@ -86,14 +91,14 @@ pip install vllm
 ```
 
 Visit our [documentation](https://docs.vllm.ai/en/latest/) to learn more.
-- [Installation](https://docs.vllm.ai/en/latest/getting_started/installation/index.html)
+- [Installation](https://docs.vllm.ai/en/latest/getting_started/installation.html)
 - [Quickstart](https://docs.vllm.ai/en/latest/getting_started/quickstart.html)
 - [List of Supported Models](https://docs.vllm.ai/en/latest/models/supported_models.html)
 
 ## Contributing
 
 We welcome and value any contributions and collaborations.
-Please check out [CONTRIBUTING.md](./CONTRIBUTING.md) for how to get involved.
+Please check out [Contributing to vLLM](https://docs.vllm.ai/en/stable/contributing/overview.html) for how to get involved.
 
 ## Sponsors
 
@@ -116,6 +121,7 @@ Compute Resources:
 - Databricks
 - DeepInfra
 - Google Cloud
+- Intel
 - Lambda Lab
 - Nebius
 - Novita AI
@@ -146,10 +152,11 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 
 ## Contact Us
 
-- For technical questions and feature requests, please use Github issues or discussions.
-- For discussing with fellow users and coordinating contributions and development, please use Slack.
-- For security disclosures, please use Github's security advisory feature.
-- For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.
+- 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 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)
 
 ## Media Kit
 

RELEASE.md

@@ -0,0 +1,54 @@
+# Releasing vLLM
+
+vLLM releases offer a reliable version of the code base, packaged into a binary format that can be conveniently accessed via PyPI. These releases also serve as key milestones for the development team to communicate with the community about newly available features, improvements, and upcoming changes that could affect users, including potential breaking changes.
+
+## Release Versioning
+
+vLLM uses a “right-shifted” versioning scheme where a new patch release is out every 2 weeks. And patch releases contain features and bug fixes (as opposed to semver where patch release contains only backwards-compatible bug fixes). When critical fixes need to be made, special release post1 is released.
+
+* _major_ major architectural milestone and when incompatible API changes are made, similar to PyTorch 2.0.
+* _minor_ major features
+* _patch_ features and backwards-compatible bug fixes
+* _post1_ or _patch-1_ backwards-compatible bug fixes, either explicit or implicit post release
+
+## Release Cadence
+
+Patch release is released on bi-weekly basis. Post release 1-3 days after patch release and uses same branch as patch release.
+Following is the release cadence for year 2025. All future release dates below are tentative. Please note: Post releases are optional.
+
+| Release Date | Patch release versions | Post Release versions |
+| --- | --- | --- |
+| Jan 2025 | 0.7.0 | --- |
+| Feb 2025 | 0.7.1, 0.7.2, 0.7.3  | --- |
+| Mar 2025 | 0.7.4, 0.7.5 | --- |
+| Apr 2025 | 0.7.6, 0.7.7 | --- |
+| May 2025 | 0.7.8, 0.7.9 | --- |
+| Jun 2025 | 0.7.10, 0.7.11 | --- |
+| Jul 2025 | 0.7.12, 0.7.13 | --- |
+| Aug 2025 | 0.7.14, 0.7.15 | --- |
+| Sep 2025 | 0.7.16, 0.7.17 | --- |
+| Oct 2025 | 0.7.18, 0.7.19 | --- |
+| Nov 2025 | 0.7.20, 0.7.21 | --- |
+| Dec 2025 | 0.7.22, 0.7.23 | --- |
+
+## Release branch
+
+Each release is built from a dedicated release branch.
+
+* For _major_, _minor_, _patch_ releases, the release branch cut is performed 1-2 days before release is live.
+* For post releases, previously cut release branch is reused
+* Release builds are triggered via push to RC tag like vX.Y.Z-rc1 . This enables us to build and test multiple RCs for each release.
+* Final tag : vX.Y.Z does not trigger the build but used for Release notes and assets.
+* After branch cut is created we monitor the main branch for any reverts and apply these reverts to a release branch.
+
+## Release Cherry-Pick Criteria
+
+After branch cut, we approach finalizing the release branch with clear criteria on what cherry picks are allowed in. Note: a cherry pick is a process to land a PR in the release branch after branch cut. These are typically limited to ensure that the team has sufficient time to complete a thorough round of testing on a stable code base.
+
+* Regression fixes - that address functional/performance regression against the most recent release (e.g. 0.7.0 for 0.7.1 release)
+* Critical fixes - critical fixes for severe issue such as silent incorrectness, backwards compatibility, crashes, deadlocks, (large) memory leaks
+* Fixes to new features introduced in the most recent release (e.g. 0.7.0 for 0.7.1 release)
+* Documentation improvements
+* Release branch specific changes (e.g. change version identifiers or CI fixes)
+
+Please note: **No feature work allowed for cherry picks**. All PRs that are considered for cherry-picks need to be merged on trunk, the only exception are Release branch specific changes.

benchmarks/README.md

@@ -1,29 +1,343 @@
 # Benchmarking vLLM
 
-## Downloading the ShareGPT dataset
+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.
 
-You can download the dataset by running:
+## Dataset Overview
+
+<table style="width:100%; border-collapse: collapse;">
+  <thead>
+    <tr>
+      <th style="width:15%; text-align: left;">Dataset</th>
+      <th style="width:10%; text-align: center;">Online</th>
+      <th style="width:10%; text-align: center;">Offline</th>
+      <th style="width:65%; text-align: left;">Data Path</th>
+    </tr>
+  </thead>
+  <tbody>
+    <tr>
+      <td><strong>ShareGPT</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json</code></td>
+    </tr>
+    <tr>
+      <td><strong>BurstGPT</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv</code></td>
+    </tr>
+    <tr>
+      <td><strong>Sonnet</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td>Local file: <code>benchmarks/sonnet.txt</code></td>
+    </tr>
+    <tr>
+      <td><strong>Random</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>synthetic</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-VisionArena</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmarena-ai/VisionArena-Chat</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-InstructCoder</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>likaixin/InstructCoder</code></td>
+    </tr>
+      <tr>
+      <td><strong>HuggingFace-AIMO</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>AI-MO/aimo-validation-aime</code> , <code>AI-MO/NuminaMath-1.5</code>, <code>AI-MO/NuminaMath-CoT</code></td>
+    </tr>
+    <tr>
+      <td><strong>HuggingFace-Other</strong></td>
+      <td style="text-align: center;">✅</td>
+      <td style="text-align: center;">✅</td>
+      <td><code>lmms-lab/LLaVA-OneVision-Data</code>, <code>Aeala/ShareGPT_Vicuna_unfiltered</code></td>
+    </tr>
+  </tbody>
+</table>
+
+✅: supported
+
+🟡: Partial support
+
+🚧: to be supported
+
+**Note**: HuggingFace dataset's `dataset-name` should be set to `hf`
+
+---
+## Example - Online Benchmark
+
+First start serving your model
 
 ```bash
-wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+vllm serve NousResearch/Hermes-3-Llama-3.1-8B --disable-log-requests
 ```
 
-## Downloading the ShareGPT4V dataset
-
-The json file refers to several image datasets (coco, llava, etc.). The benchmark scripts
-will ignore a datapoint if the referred image is missing.
+Then run the benchmarking script
 
 ```bash
-wget https://huggingface.co/datasets/Lin-Chen/ShareGPT4V/resolve/main/sharegpt4v_instruct_gpt4-vision_cap100k.json
-mkdir coco -p
-wget http://images.cocodataset.org/zips/train2017.zip -O coco/train2017.zip
-unzip coco/train2017.zip -d coco/
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --num-prompts 10
 ```
 
-# Downloading the BurstGPT dataset
+If successful, you will see the following output
 
-You can download the BurstGPT v1.1 dataset by running:
+```
+============ Serving Benchmark Result ============
+Successful requests:                     10        
+Benchmark duration (s):                  5.78      
+Total input tokens:                      1369      
+Total generated tokens:                  2212      
+Request throughput (req/s):              1.73      
+Output token throughput (tok/s):         382.89    
+Total Token throughput (tok/s):          619.85    
+---------------Time to First Token----------------
+Mean TTFT (ms):                          71.54     
+Median TTFT (ms):                        73.88     
+P99 TTFT (ms):                           79.49     
+-----Time per Output Token (excl. 1st token)------
+Mean TPOT (ms):                          7.91      
+Median TPOT (ms):                        7.96      
+P99 TPOT (ms):                           8.03      
+---------------Inter-token Latency----------------
+Mean ITL (ms):                           7.74      
+Median ITL (ms):                         7.70      
+P99 ITL (ms):                            8.39      
+==================================================
+```
+
+### VisionArena Benchmark for Vision Language Models
 
 ```bash
-wget https://github.com/HPMLL/BurstGPT/releases/download/v1.1/BurstGPT_without_fails_2.csv
+# need a model with vision capability here
+vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
 ```
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --hf-split train \
+  --num-prompts 1000
+```
+
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_USE_V1=1 vllm serve meta-llama/Meta-Llama-3-8B-Instruct \
+    --speculative-model "[ngram]" \
+    --ngram_prompt_lookup_min 2 \
+    --ngram-prompt-lookup-max 5 \
+    --num_speculative_tokens 5
+```
+
+``` bash
+python3 benchmarks/benchmark_serving.py \
+    --model meta-llama/Meta-Llama-3-8B-Instruct \
+    --dataset-name hf \
+    --dataset-path likaixin/InstructCoder \
+    --num-prompts 2048
+```
+
+### Other HuggingFaceDataset Examples
+
+```bash
+vllm serve Qwen/Qwen2-VL-7B-Instruct --disable-log-requests
+```
+
+**`lmms-lab/LLaVA-OneVision-Data`**
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+**`Aeala/ShareGPT_Vicuna_unfiltered`**
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend openai-chat \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --endpoint /v1/chat/completions \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+**`AI-MO/aimo-validation-aime`**
+
+``` bash
+python3 vllm/benchmarks/benchmark_serving.py \
+    --model Qwen/QwQ-32B \
+    --dataset-name hf \
+    --dataset-path AI-MO/aimo-validation-aime \
+    --num-prompts 10 \
+    --seed 42
+```
+
+### Running With Sampling Parameters
+
+When using OpenAI-compatible backends such as `vllm`, optional sampling
+parameters can be specified. Example client command:
+
+```bash
+python3 vllm/benchmarks/benchmark_serving.py \
+  --backend vllm \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --endpoint /v1/completions \
+  --dataset-name sharegpt \
+  --dataset-path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --top-k 10 \
+  --top-p 0.9 \
+  --temperature 0.5 \
+  --num-prompts 10
+```
+
+---
+## Example - Offline Throughput Benchmark
+
+```bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model NousResearch/Hermes-3-Llama-3.1-8B \
+  --dataset-name sonnet \
+  --dataset-path vllm/benchmarks/sonnet.txt \
+  --num-prompts 10
+```
+
+If successful, you will see the following output
+
+```
+Throughput: 7.15 requests/s, 4656.00 total tokens/s, 1072.15 output tokens/s
+Total num prompt tokens:  5014
+Total num output tokens:  1500
+```
+
+### VisionArena Benchmark for Vision Language Models
+
+``` bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmarena-ai/VisionArena-Chat \
+  --num-prompts 1000 \
+  --hf-split train
+```
+
+The `num prompt tokens` now includes image token counts
+
+```
+Throughput: 2.55 requests/s, 4036.92 total tokens/s, 326.90 output tokens/s
+Total num prompt tokens:  14527
+Total num output tokens:  1280
+```
+
+### InstructCoder Benchmark with Speculative Decoding
+
+``` bash
+VLLM_WORKER_MULTIPROC_METHOD=spawn \
+VLLM_USE_V1=1 \
+python3 vllm/benchmarks/benchmark_throughput.py \
+    --dataset-name=hf \
+    --dataset-path=likaixin/InstructCoder \
+    --model=meta-llama/Meta-Llama-3-8B-Instruct \
+    --input-len=1000 \
+    --output-len=100 \
+    --num-prompts=2048 \
+    --async-engine \
+    --speculative-model="[ngram]" \
+    --ngram_prompt_lookup_min=2 \
+    --ngram-prompt-lookup-max=5 \
+    --num_speculative_tokens=5
+```
+
+```
+Throughput: 104.77 requests/s, 23836.22 total tokens/s, 10477.10 output tokens/s
+Total num prompt tokens:  261136
+Total num output tokens:  204800
+```
+
+### Other HuggingFaceDataset Examples
+
+**`lmms-lab/LLaVA-OneVision-Data`**
+
+```bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path lmms-lab/LLaVA-OneVision-Data \
+  --hf-split train \
+  --hf-subset "chart2text(cauldron)" \
+  --num-prompts 10
+```
+
+**`Aeala/ShareGPT_Vicuna_unfiltered`**
+
+```bash
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model Qwen/Qwen2-VL-7B-Instruct \
+  --backend vllm-chat \
+  --dataset-name hf \
+  --dataset-path Aeala/ShareGPT_Vicuna_unfiltered \
+  --hf-split train \
+  --num-prompts 10
+```
+
+**`AI-MO/aimo-validation-aime`**
+
+```bash
+python3 benchmarks/benchmark_throughput.py \
+  --model Qwen/QwQ-32B \
+  --backend vllm \
+  --dataset-name hf \
+  --dataset-path AI-MO/aimo-validation-aime \
+  --hf-split train \
+  --num-prompts 10
+```
+
+### Benchmark with LoRA Adapters
+
+``` bash
+# download dataset
+# wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+python3 vllm/benchmarks/benchmark_throughput.py \
+  --model meta-llama/Llama-2-7b-hf \
+  --backend vllm \
+  --dataset_path <your data path>/ShareGPT_V3_unfiltered_cleaned_split.json \
+  --dataset_name sharegpt \
+  --num-prompts 10 \
+  --max-loras 2 \
+  --max-lora-rank 8 \
+  --enable-lora \
+  --lora-path yard1/llama-2-7b-sql-lora-test
+  ```

benchmarks/backend_request_func.py

@@ -6,7 +6,7 @@ import sys
 import time
 import traceback
 from dataclasses import dataclass, field
-from typing import List, Optional, Union
+from typing import Optional, Union
 
 import aiohttp
 import huggingface_hub.constants
@@ -14,6 +14,9 @@ from tqdm.asyncio import tqdm
 from transformers import (AutoTokenizer, PreTrainedTokenizer,
                           PreTrainedTokenizerFast)
 
+# NOTE(simon): do not import vLLM here so the benchmark script
+# can run without vLLM installed.
+
 AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
 
 
@@ -25,7 +28,6 @@ class RequestFuncInput:
     output_len: int
     model: str
     model_name: Optional[str] = None
-    best_of: int = 1
     logprobs: Optional[int] = None
     extra_body: Optional[dict] = None
     multi_modal_content: Optional[dict] = None
@@ -39,8 +41,8 @@ class RequestFuncOutput:
     latency: float = 0.0
     output_tokens: int = 0
     ttft: float = 0.0  # Time to first token
-    itl: List[float] = field(
-        default_factory=list)  # List of inter-token latencies
+    itl: list[float] = field(
+        default_factory=list)  # list of inter-token latencies
     tpot: float = 0.0  # avg next-token latencies
     prompt_len: int = 0
     error: str = ""
@@ -56,13 +58,12 @@ async def async_request_tgi(
     async with aiohttp.ClientSession(trust_env=True,
                                      timeout=AIOHTTP_TIMEOUT) as session:
         params = {
-            "best_of": request_func_input.best_of,
             "max_new_tokens": request_func_input.output_len,
             "do_sample": True,
             "temperature": 0.01,  # TGI does not accept 0.0 temperature.
             "top_p": 0.99,  # TGI does not accept 1.0 top_p.
             "truncate": request_func_input.prompt_len,
-            # TGI does not accept ignore_eos flag.
+            "ignore_eos_token": request_func_input.ignore_eos,
         }
         payload = {
             "inputs": request_func_input.prompt,
@@ -70,6 +71,10 @@ async def async_request_tgi(
         }
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
+        if request_func_input.ignore_eos:
+            output.output_tokens = request_func_input.output_len
+        else:
+            output.output_tokens = None
 
         ttft = 0.0
         st = time.perf_counter()
@@ -128,7 +133,6 @@ async def async_request_trt_llm(
 
     async with aiohttp.ClientSession(trust_env=True,
                                      timeout=AIOHTTP_TIMEOUT) as session:
-        assert request_func_input.best_of == 1
         payload = {
             "accumulate_tokens": True,
             "text_input": request_func_input.prompt,
@@ -193,7 +197,6 @@ async def async_request_deepspeed_mii(
 ) -> RequestFuncOutput:
     async with aiohttp.ClientSession(trust_env=True,
                                      timeout=AIOHTTP_TIMEOUT) as session:
-        assert request_func_input.best_of == 1
 
         payload = {
             "prompt": request_func_input.prompt,
@@ -216,7 +219,15 @@ async def async_request_deepspeed_mii(
                 if response.status == 200:
                     parsed_resp = await response.json()
                     output.latency = time.perf_counter() - st
-                    output.generated_text = parsed_resp["text"][0]
+                    if "choices" in parsed_resp:
+                        output.generated_text = parsed_resp["choices"][0][
+                            "text"]
+                    elif "text" in parsed_resp:
+                        output.generated_text = parsed_resp["text"][0]
+                    else:
+                        output.error = ("Unexpected response format: "
+                                        "neither 'choices' nor 'text' found")
+                        output.success = False
                     output.success = True
                 else:
                     output.error = response.reason or ""
@@ -247,7 +258,6 @@ async def async_request_openai_completions(
                 if request_func_input.model_name else request_func_input.model,
             "prompt": request_func_input.prompt,
             "temperature": 0.0,
-            "best_of": request_func_input.best_of,
             "max_tokens": request_func_input.output_len,
             "logprobs": request_func_input.logprobs,
             "stream": True,
@@ -336,7 +346,7 @@ async def async_request_openai_chat_completions(
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        "chat/completions"
+        ("chat/completions", "profile")
     ), "OpenAI Chat Completions API URL must end with 'chat/completions'."
 
     async with aiohttp.ClientSession(trust_env=True,
@@ -430,12 +440,17 @@ def get_model(pretrained_model_name_or_path: str) -> str:
     if os.getenv('VLLM_USE_MODELSCOPE', 'False').lower() == 'true':
         from modelscope import snapshot_download
 
-        model_path = snapshot_download(
-            model_id=pretrained_model_name_or_path,
-            local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
-            ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"])
+        from vllm.model_executor.model_loader.weight_utils import get_lock
 
-        return model_path
+        # Use file lock to prevent multiple processes from
+        # downloading the same model weights at the same time.
+        with get_lock(pretrained_model_name_or_path):
+            model_path = snapshot_download(
+                model_id=pretrained_model_name_or_path,
+                local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
+                ignore_file_pattern=[".*.pt", ".*.safetensors", ".*.bin"])
+
+            return model_path
     return pretrained_model_name_or_path
 
 
@@ -482,3 +497,9 @@ ASYNC_REQUEST_FUNCS = {
     "scalellm": async_request_openai_completions,
     "sglang": async_request_openai_completions,
 }
+
+OPENAI_COMPATIBLE_BACKENDS = [
+    k for k, v in ASYNC_REQUEST_FUNCS.items()
+    if v in (async_request_openai_completions,
+             async_request_openai_chat_completions)
+]

benchmarks/benchmark_dataset.py

@@ -0,0 +1,817 @@
+# SPDX-License-Identifier: Apache-2.0
+"""
+This module defines a framework for sampling benchmark requests from various
+datasets. Each dataset subclass of BenchmarkDataset must implement sample
+generation. Supported dataset types include:
+  - ShareGPT
+  - Random (synthetic)
+  - Sonnet
+  - 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
+import io
+import json
+import logging
+import random
+from abc import ABC, abstractmethod
+from collections.abc import Mapping
+from dataclasses import dataclass
+from functools import cache
+from io import BytesIO
+from typing import Any, Callable, Optional, Union
+
+import numpy as np
+import pandas as pd
+from datasets import load_dataset
+from PIL import Image
+from transformers import PreTrainedTokenizerBase
+
+from vllm.lora.request import LoRARequest
+from vllm.lora.utils import get_adapter_absolute_path
+from vllm.multimodal import MultiModalDataDict
+from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer
+
+logger = logging.getLogger(__name__)
+
+# -----------------------------------------------------------------------------
+# Data Classes
+# -----------------------------------------------------------------------------
+
+
+@dataclass
+class SampleRequest:
+    """
+    Represents a single inference request for benchmarking.
+    """
+
+    prompt: Union[str, Any]
+    prompt_len: int
+    expected_output_len: int
+    multi_modal_data: Optional[Union[MultiModalDataDict, dict]] = None
+    lora_request: Optional[LoRARequest] = None
+
+
+# -----------------------------------------------------------------------------
+# Benchmark Dataset Base Class
+# -----------------------------------------------------------------------------
+
+
+class BenchmarkDataset(ABC):
+    DEFAULT_SEED = 0
+
+    def __init__(
+        self,
+        dataset_path: Optional[str] = None,
+        random_seed: int = DEFAULT_SEED,
+    ) -> None:
+        """
+        Initialize the BenchmarkDataset with an optional dataset path and random
+        seed.  Args:
+            dataset_path (Optional[str]): Path to the dataset. If None, it
+            indicates that a default or random dataset might be used.
+            random_seed (int): Seed value for reproducible shuffling or
+            sampling. Defaults to DEFAULT_SEED.
+        """
+        self.dataset_path = dataset_path
+        # Set the random seed, ensuring that a None value is replaced with the
+        # default seed.
+        self.random_seed = (random_seed
+                            if random_seed is not None else self.DEFAULT_SEED)
+        self.data = None
+
+    def apply_multimodal_chat_transformation(
+            self,
+            prompt: str,
+            mm_content: Optional[MultiModalDataDict] = None) -> list[dict]:
+        """
+        Transform a prompt and optional multimodal content into a chat format.
+        This method is used for chat models that expect a specific conversation
+        format.
+        """
+        content = [{"text": prompt, "type": "text"}]
+        if mm_content is not None:
+            content.append(mm_content)
+        return [{"role": "user", "content": content}]
+
+    def load_data(self) -> None:
+        """
+        Load data from the dataset path into self.data.
+
+        This method must be overridden by subclasses since the method to load
+        data will vary depending on the dataset format and source.
+
+        Raises:
+            NotImplementedError: If a subclass does not implement this method.
+        """
+        # TODO (jenniferzhao): add support for downloading data
+        raise NotImplementedError(
+            "load_data must be implemented in subclasses.")
+
+    def get_random_lora_request(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        max_loras: Optional[int] = None,
+        lora_path: Optional[str] = None,
+    ) -> tuple[Optional[LoRARequest], AnyTokenizer]:
+        """
+        Optionally select a random LoRA request and return its associated
+        tokenizer.
+
+        This method is used when LoRA parameters are provided.  It randomly
+        selects a LoRA based on max_loras and retrieves a cached tokenizer for
+        that LoRA if available. Otherwise, it returns the base tokenizer.
+
+        Args:
+            tokenizer (PreTrainedTokenizerBase): The base tokenizer to use if no
+            LoRA is selected.  max_loras (Optional[int]): The maximum number of
+            LoRAs available. If None, LoRA is not used.  lora_path
+            (Optional[str]): Path to the LoRA parameters on disk. If None, LoRA
+            is not used.
+
+        Returns:
+            tuple[Optional[LoRARequest], AnyTokenizer]: A tuple where the first
+            element is a LoRARequest (or None if not applicable) and the second
+            element is the tokenizer associated with the LoRA request (or the
+            base tokenizer).
+        """
+        if max_loras is None or lora_path is None:
+            return None, tokenizer
+
+        # Generate a random LoRA ID in the range [1, max_loras].
+        lora_id = random.randint(1, max_loras)
+        lora_request = LoRARequest(
+            lora_name=str(lora_id),
+            lora_int_id=lora_id,
+            lora_path=lora_path_on_disk(lora_path),
+        )
+        if lora_id not in lora_tokenizer_cache:
+            lora_tokenizer_cache[lora_id] = get_lora_tokenizer(lora_request)
+        # Return lora_request and the cached tokenizer if available; otherwise,
+        # return the base tokenizer
+        return lora_request, lora_tokenizer_cache[lora_id] or tokenizer
+
+    @abstractmethod
+    def sample(self, tokenizer: PreTrainedTokenizerBase,
+               num_requests: int) -> list[SampleRequest]:
+        """
+        Abstract method to generate sample requests from the dataset.
+
+        Subclasses must override this method to implement dataset-specific logic
+        for generating a list of SampleRequest objects.
+
+        Args:
+            tokenizer (PreTrainedTokenizerBase): The tokenizer to be used
+             for processing the dataset's text.
+            num_requests (int): The number of sample requests to generate.
+
+        Returns:
+            list[SampleRequest]: A list of sample requests generated from the
+            dataset.
+        """
+        raise NotImplementedError("sample must be implemented in subclasses.")
+
+    def maybe_oversample_requests(self, requests: list[SampleRequest],
+                                  num_requests: int) -> None:
+        """
+        Oversamples the list of requests if its size is less than the desired
+        number.
+
+        Args:
+            requests (List[SampleRequest]): The current list of sampled
+            requests.  num_requests (int): The target number of requests.
+        """
+        if len(requests) < num_requests:
+            random.seed(self.random_seed)
+            additional = random.choices(requests,
+                                        k=num_requests - len(requests))
+            requests.extend(additional)
+            logger.info("Oversampled requests to reach %d total samples.",
+                        num_requests)
+
+
+# -----------------------------------------------------------------------------
+# Utility Functions and Global Caches
+# -----------------------------------------------------------------------------
+
+
+def is_valid_sequence(
+    prompt_len: int,
+    output_len: int,
+    min_len: int = 4,
+    max_prompt_len: int = 1024,
+    max_total_len: int = 2048,
+    skip_min_output_len_check: bool = False,
+) -> bool:
+    """
+    Validate a sequence based on prompt and output lengths.
+
+    Default pruning criteria are copied from the original `sample_hf_requests`
+    and `sample_sharegpt_requests` functions in benchmark_serving.py, as well as
+    from `sample_requests` in benchmark_throughput.py.
+    """
+    # Check for invalid conditions
+    prompt_too_short = prompt_len < min_len
+    output_too_short = (not skip_min_output_len_check) and (output_len
+                                                            < min_len)
+    prompt_too_long = prompt_len > max_prompt_len
+    combined_too_long = (prompt_len + output_len) > max_total_len
+
+    # Return True if none of the invalid conditions are met
+    return not (prompt_too_short or output_too_short or prompt_too_long
+                or combined_too_long)
+
+
+@cache
+def lora_path_on_disk(lora_path: str) -> str:
+    return get_adapter_absolute_path(lora_path)
+
+
+# Global cache for LoRA tokenizers.
+lora_tokenizer_cache: dict[int, AnyTokenizer] = {}
+
+
+def process_image(image: Any) -> Mapping[str, Any]:
+    """
+    Process a single image input and return a multimedia content dictionary.
+
+    Supports three input types:
+
+    1. Dictionary with raw image bytes: - Expects a dict with a 'bytes' key
+       containing raw image data.  - Loads the bytes as a PIL.Image.Image.
+
+    2. PIL.Image.Image input: - Converts the image to RGB.  - Saves the image as
+       a JPEG in memory.  - Encodes the JPEG data as a base64 string.  - Returns
+       a dictionary with the image as a base64 data URL.
+
+    3. String input: - Treats the string as a URL or local file path.  -
+       Prepends "file://" if the string doesn't start with "http://" or
+       "file://".  - Returns a dictionary with the image URL.
+
+    Raises:
+        ValueError: If the input is not a supported type.
+    """
+    if isinstance(image, dict) and 'bytes' in image:
+        image = Image.open(BytesIO(image['bytes']))
+    if isinstance(image, Image.Image):
+        image = image.convert("RGB")
+        with io.BytesIO() as image_data:
+            image.save(image_data, format="JPEG")
+            image_base64 = base64.b64encode(
+                image_data.getvalue()).decode("utf-8")
+        return {
+            "type": "image_url",
+            "image_url": {
+                "url": f"data:image/jpeg;base64,{image_base64}"
+            },
+        }
+
+    if isinstance(image, str):
+        image_url = (image if image.startswith(
+            ("http://", "file://")) else f"file://{image}")
+        return {"type": "image_url", "image_url": {"url": image_url}}
+
+    raise ValueError(f"Invalid image input {image}. Must be a PIL.Image.Image"
+                     " or str or dictionary with raw image bytes.")
+
+
+# -----------------------------------------------------------------------------
+# Random Dataset Implementation (Synthetic Data)
+# -----------------------------------------------------------------------------
+
+
+class RandomDataset(BenchmarkDataset):
+    # Default values copied from benchmark_serving.py for the random dataset.
+    DEFAULT_PREFIX_LEN = 0
+    DEFAULT_RANGE_RATIO = 0.0
+    DEFAULT_INPUT_LEN = 1024
+    DEFAULT_OUTPUT_LEN = 128
+
+    def __init__(
+        self,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        prefix_len: int = DEFAULT_PREFIX_LEN,
+        range_ratio: float = DEFAULT_RANGE_RATIO,
+        input_len: int = DEFAULT_INPUT_LEN,
+        output_len: int = DEFAULT_OUTPUT_LEN,
+        **kwargs,
+    ) -> list[SampleRequest]:
+        # Enforce range_ratio < 1
+        assert range_ratio < 1.0, (
+            "random_range_ratio must be < 1.0 to ensure a valid sampling range"
+        )
+
+        vocab_size = tokenizer.vocab_size
+
+        prefix_token_ids = (np.random.randint(
+            0, vocab_size, size=prefix_len).tolist() if prefix_len > 0 else [])
+
+        # New sampling logic: [X * (1 - b), X * (1 + b)]
+        input_low = int(input_len * (1 - range_ratio))
+        input_high = int(input_len * (1 + range_ratio))
+        output_low = int(output_len * (1 - range_ratio))
+        output_high = int(output_len * (1 + range_ratio))
+
+        # Add logging for debugging
+        logger.info("Sampling input_len from [%s, %s]", input_low, input_high)
+        logger.info("Sampling output_len from [%s, %s]", output_low,
+                    output_high)
+
+        input_lens = np.random.randint(input_low,
+                                       input_high + 1,
+                                       size=num_requests)
+        output_lens = np.random.randint(output_low,
+                                        output_high + 1,
+                                        size=num_requests)
+        offsets = np.random.randint(0, vocab_size, size=num_requests)
+
+        requests = []
+        for i in range(num_requests):
+            inner_seq = ((offsets[i] + i + np.arange(input_lens[i])) %
+                         vocab_size).tolist()
+            token_sequence = prefix_token_ids + inner_seq
+            prompt = tokenizer.decode(token_sequence)
+            total_input_len = prefix_len + int(input_lens[i])
+            requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=total_input_len,
+                    expected_output_len=int(output_lens[i]),
+                ))
+        return requests
+
+
+# -----------------------------------------------------------------------------
+# ShareGPT Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class ShareGPTDataset(BenchmarkDataset):
+    """
+    Implements the ShareGPT dataset.  Loads data from a JSON file and generates
+    sample requests based on conversation turns.
+    """
+
+    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.")
+
+        with open(self.dataset_path, encoding="utf-8") as f:
+            self.data = json.load(f)
+        # Filter entries with at least two conversation turns.
+        self.data = [
+            entry for entry in self.data
+            if "conversations" in entry and len(entry["conversations"]) >= 2
+        ]
+        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,
+        **kwargs,
+    ) -> list:
+        samples: list = []
+        for entry in self.data:
+            if len(samples) >= num_requests:
+                break
+            prompt, completion = (
+                entry["conversations"][0]["value"],
+                entry["conversations"][1]["value"],
+            )
+
+            lora_request, tokenizer = self.get_random_lora_request(
+                tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path)
+            prompt_ids = tokenizer(prompt).input_ids
+            completion_ids = tokenizer(completion).input_ids
+            prompt_len = len(prompt_ids)
+            new_output_len = (len(completion_ids)
+                              if output_len is None else output_len)
+            if not is_valid_sequence(prompt_len,
+                                     new_output_len,
+                                     skip_min_output_len_check=output_len
+                                     is not None):
+                continue
+            if enable_multimodal_chat:
+                prompt = self.apply_multimodal_chat_transformation(
+                    prompt, None)
+            samples.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=new_output_len,
+                    lora_request=lora_request,
+                ))
+        self.maybe_oversample_requests(samples, num_requests)
+        return samples
+
+
+# -----------------------------------------------------------------------------
+# Sonnet Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class SonnetDataset(BenchmarkDataset):
+    """
+    Simplified implementation of the Sonnet dataset.  Loads poem lines from a
+    text file and generates sample requests.  Default values here copied from
+    `benchmark_serving.py` for the sonnet dataset.
+    """
+
+    DEFAULT_PREFIX_LEN = 200
+    DEFAULT_INPUT_LEN = 550
+    DEFAULT_OUTPUT_LEN = 150
+
+    def __init__(
+        self,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.load_data()
+
+    def load_data(self) -> None:
+        if not self.dataset_path:
+            raise ValueError("dataset_path must be provided.")
+        with open(self.dataset_path, encoding="utf-8") as f:
+            self.data = f.readlines()
+
+    def sample(
+        self,
+        tokenizer,
+        num_requests: int,
+        prefix_len: int = DEFAULT_PREFIX_LEN,
+        input_len: int = DEFAULT_INPUT_LEN,
+        output_len: int = DEFAULT_OUTPUT_LEN,
+        return_prompt_formatted: bool = False,
+        **kwargs,
+    ) -> list:
+        # Calculate average token length for a poem line.
+        tokenized_lines = [tokenizer(line).input_ids for line in self.data]
+        avg_len = sum(len(tokens)
+                      for tokens in tokenized_lines) / len(tokenized_lines)
+
+        # Build the base prompt.
+        base_prompt = "Pick as many lines as you can from these poem lines:\n"
+        base_msg = [{"role": "user", "content": base_prompt}]
+        base_fmt = tokenizer.apply_chat_template(base_msg,
+                                                 add_generation_prompt=True,
+                                                 tokenize=False)
+        base_offset = len(tokenizer(base_fmt).input_ids)
+        if input_len <= base_offset:
+            raise ValueError(
+                f"'input_len' must be higher than the base prompt length "
+                f"({base_offset}).")
+
+        # Determine how many poem lines to use.
+        num_input_lines = round((input_len - base_offset) / avg_len)
+        num_prefix_lines = max(round((prefix_len - base_offset) / avg_len), 0)
+        prefix_lines = self.data[:num_prefix_lines]
+
+        samples = []
+        while len(samples) < num_requests:
+            extra_lines = random.choices(self.data,
+                                         k=num_input_lines - num_prefix_lines)
+            prompt = f"{base_prompt}{''.join(prefix_lines + extra_lines)}"
+            msg = [{"role": "user", "content": prompt}]
+            prompt_formatted = tokenizer.apply_chat_template(
+                msg, add_generation_prompt=True, tokenize=False)
+            prompt_len = len(tokenizer(prompt_formatted).input_ids)
+            if prompt_len <= input_len:
+                samples.append(
+                    SampleRequest(
+                        prompt=prompt_formatted
+                        if return_prompt_formatted else prompt,
+                        prompt_len=prompt_len,
+                        expected_output_len=output_len,
+                    ))
+        return samples
+
+
+# -----------------------------------------------------------------------------
+# BurstGPT Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class BurstGPTDataset(BenchmarkDataset):
+    """
+    Implements the BurstGPT dataset.  Loads data from a CSV file and generates
+    sample requests based on synthetic prompt generation. Only rows with Model
+    "GPT-4" and positive response tokens are used.
+    """
+
+    def __init__(self, **kwargs) -> None:
+        super().__init__(**kwargs)
+        self.load_data()
+
+    def load_data(self, ):
+        if self.dataset_path is None:
+            raise ValueError("dataset_path must be provided for loading data.")
+
+        df = pd.read_csv(self.dataset_path)
+        # Filter to keep only GPT-4 rows.
+        gpt4_df = df[df["Model"] == "GPT-4"]
+        # Remove failed requests (where Response tokens is 0 or less).
+        gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0]
+        # Sample the desired number of rows.
+        self.data = gpt4_df
+
+    def _sample_loaded_data(self, num_requests: int) -> list:
+        if num_requests <= len(self.data):
+            data = self.data.sample(n=num_requests,
+                                    random_state=self.random_seed)
+        else:
+            data = self.data.sample(
+                n=num_requests,
+                random_state=self.random_seed,
+                replace=True,
+            )
+        # Convert the dataframe to a list of lists.
+        return data.values.tolist()
+
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        max_loras: Optional[int] = None,
+        lora_path: Optional[str] = None,
+        **kwargs,
+    ) -> list[SampleRequest]:
+        samples = []
+        data = self._sample_loaded_data(num_requests=num_requests)
+        for i in range(num_requests):
+            input_len = int(data[i][2])
+            output_len = int(data[i][3])
+            lora_req, tokenizer = self.get_random_lora_request(
+                tokenizer=tokenizer, max_loras=max_loras, lora_path=lora_path)
+            vocab_size = tokenizer.vocab_size
+            # Generate a synthetic prompt: a list of token IDs computed as (i +
+            # j) modulo vocab_size.
+            token_ids = [(i + j) % vocab_size for j in range(input_len)]
+            prompt = tokenizer.decode(token_ids)
+            samples.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=input_len,
+                    expected_output_len=output_len,
+                    lora_request=lora_req,
+                ))
+        return samples
+
+
+# -----------------------------------------------------------------------------
+# HuggingFace Dataset Base Implementation
+# -----------------------------------------------------------------------------
+class HuggingFaceDataset(BenchmarkDataset):
+    """Base class for datasets hosted on HuggingFace."""
+
+    SUPPORTED_DATASET_PATHS: Union[set[str], dict[str, Callable]] = set()
+
+    def __init__(
+        self,
+        dataset_path: str,
+        dataset_split: str,
+        dataset_subset: Optional[str] = None,
+        **kwargs,
+    ) -> None:
+        super().__init__(dataset_path=dataset_path, **kwargs)
+
+        self.dataset_split = dataset_split
+        self.dataset_subset = dataset_subset
+        self.load_data()
+
+    def load_data(self) -> None:
+        """Load data from HuggingFace datasets."""
+        self.data = load_dataset(
+            self.dataset_path,
+            name=self.dataset_subset,
+            split=self.dataset_split,
+            streaming=True,
+        )
+        self.data = self.data.shuffle(seed=self.random_seed)
+
+
+# -----------------------------------------------------------------------------
+# Conversation Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class ConversationDataset(HuggingFaceDataset):
+    """Dataset for conversation data with multimodal support."""
+    SUPPORTED_DATASET_PATHS = {
+        'lmms-lab/LLaVA-OneVision-Data', 'Aeala/ShareGPT_Vicuna_unfiltered'
+    }
+
+    def sample(self,
+               tokenizer: PreTrainedTokenizerBase,
+               num_requests: int,
+               output_len: Optional[int] = None,
+               enable_multimodal_chat: bool = False,
+               **kwargs) -> list:
+        # Filter examples with at least 2 conversations
+        filtered_data = self.data.filter(
+            lambda x: len(x["conversations"]) >= 2)
+        sampled_requests = []
+        dynamic_output = output_len is None
+
+        for item in filtered_data:
+            if len(sampled_requests) >= num_requests:
+                break
+            conv = item["conversations"]
+            prompt, completion = conv[0]["value"], conv[1]["value"]
+
+            prompt_ids = tokenizer(prompt).input_ids
+            completion_ids = tokenizer(completion).input_ids
+            prompt_len = len(prompt_ids)
+            completion_len = len(completion_ids)
+            output_len = completion_len if dynamic_output else output_len
+            assert isinstance(output_len, int) and output_len > 0
+            if dynamic_output and not is_valid_sequence(
+                    prompt_len, completion_len):
+                continue
+            mm_content = process_image(
+                item["image"]) if "image" in item else None
+            if enable_multimodal_chat:
+                # Note: when chat is enabled the request prompt_len is no longer
+                # accurate and we will be using request output to count the
+                # actual prompt len and output len
+                prompt = self.apply_multimodal_chat_transformation(
+                    prompt, mm_content)
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                    multi_modal_data=mm_content,
+                ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+        return sampled_requests
+
+
+# -----------------------------------------------------------------------------
+# Vision Arena Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class VisionArenaDataset(HuggingFaceDataset):
+    """
+    Vision Arena Dataset.
+    """
+
+    DEFAULT_OUTPUT_LEN = 128
+    SUPPORTED_DATASET_PATHS = {
+        "lmarena-ai/VisionArena-Chat":
+        lambda x: x["conversation"][0][0]["content"],
+        "lmarena-ai/vision-arena-bench-v0.1":
+        lambda x: x["turns"][0][0]["content"]
+    }
+
+    def sample(
+        self,
+        tokenizer: PreTrainedTokenizerBase,
+        num_requests: int,
+        output_len: Optional[int] = None,
+        enable_multimodal_chat: bool = False,
+        **kwargs,
+    ) -> list:
+        output_len = (output_len
+                      if output_len is not None else self.DEFAULT_OUTPUT_LEN)
+        sampled_requests = []
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            parser_fn = self.SUPPORTED_DATASET_PATHS.get(self.dataset_path)
+            if parser_fn is None:
+                raise ValueError(
+                    f"Unsupported dataset path: {self.dataset_path}")
+            prompt = parser_fn(item)
+            mm_content = process_image(item["images"][0])
+            prompt_len = len(tokenizer(prompt).input_ids)
+            if enable_multimodal_chat:
+                # Note: when chat is enabled the request prompt_len is no longer
+                # accurate and we will be using request output to count the
+                # actual prompt len
+                prompt = self.apply_multimodal_chat_transformation(
+                    prompt, mm_content)
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                    multi_modal_data=mm_content,
+                ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+        return sampled_requests
+
+
+# -----------------------------------------------------------------------------
+# Instruct Coder Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class InstructCoderDataset(HuggingFaceDataset):
+    """
+    InstructCoder Dataset.
+    https://huggingface.co/datasets/likaixin/InstructCoder
+
+    InstructCoder is the dataset designed for general code editing.  It consists
+    of 114,239 instruction-input-output triplets, and covers multiple distinct
+    code editing scenario.
+    """
+
+    DEFAULT_OUTPUT_LEN = 200  # this is the average default output length
+    SUPPORTED_DATASET_PATHS = {
+        "likaixin/InstructCoder",
+    }
+
+    def sample(self,
+               tokenizer: PreTrainedTokenizerBase,
+               num_requests: int,
+               output_len: Optional[int] = None,
+               enable_multimodal_chat: bool = False,
+               **kwargs) -> list:
+        output_len = (output_len
+                      if output_len is not None else self.DEFAULT_OUTPUT_LEN)
+        sampled_requests = []
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            prompt = f"{item['instruction']}:\n{item['input']}"
+            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
+
+
+# -----------------------------------------------------------------------------
+# AIMO Dataset Implementation
+# -----------------------------------------------------------------------------
+
+
+class AIMODataset(HuggingFaceDataset):
+    """
+    Dataset class for processing a AIMO dataset with reasoning questions.
+    """
+    SUPPORTED_DATASET_PATHS = {
+        "AI-MO/aimo-validation-aime", "AI-MO/NuminaMath-1.5",
+        "AI-MO/NuminaMath-CoT"
+    }
+
+    def sample(self,
+               tokenizer: PreTrainedTokenizerBase,
+               num_requests: int,
+               output_len: Optional[int] = None,
+               **kwargs) -> list:
+        sampled_requests = []
+        dynamic_output = output_len is None
+
+        for item in self.data:
+            if len(sampled_requests) >= num_requests:
+                break
+            prompt, completion = item['problem'], item["solution"]
+
+            prompt_ids = tokenizer(prompt).input_ids
+            completion_ids = tokenizer(completion).input_ids
+            prompt_len = len(prompt_ids)
+            completion_len = len(completion_ids)
+            output_len = completion_len if dynamic_output else output_len
+            assert isinstance(output_len, int) and output_len > 0
+            if dynamic_output and not is_valid_sequence(prompt_len,
+                                                        completion_len,
+                                                        max_prompt_len=2048,
+                                                        max_total_len=32000):
+                continue
+            sampled_requests.append(
+                SampleRequest(
+                    prompt=prompt,
+                    prompt_len=prompt_len,
+                    expected_output_len=output_len,
+                    multi_modal_data=None,
+                ))
+        self.maybe_oversample_requests(sampled_requests, num_requests)
+        return sampled_requests

benchmarks/benchmark_guided.py

@@ -1,495 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-"""Benchmark guided decoding throughput."""
-import argparse
-import dataclasses
-import json
-import os
-import random
-import time
-from typing import List
-
-import datasets
-import pandas as pd
-import uvloop
-from transformers import AutoTokenizer, PreTrainedTokenizerBase
-
-from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
-from vllm.entrypoints.openai.api_server import (
-    build_async_engine_client_from_engine_args)
-from vllm.sampling_params import GuidedDecodingParams
-from vllm.utils import FlexibleArgumentParser, merge_async_iterators
-
-
-@dataclasses.dataclass
-class SampleRequest:
-    """A class representing a single inference request for benchmarking.
-
-    Attributes:
-        prompt: The input text prompt for the model.
-        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
-            images).
-        prompt_len: The length of the prompt in tokens.
-        expected_output_len: The expected length of the output in tokens.
-    """
-    prompt: str
-    prompt_len: int
-    expected_output_len: int
-    schema: dict
-    structure_type: str = 'json'
-    completion: str = None
-
-
-def run_vllm(requests: List[SampleRequest],
-             engine_args: EngineArgs,
-             n: int,
-             guided_decoding_rate: float = 1.0,
-             warmup: bool = False) -> float:
-    from vllm import LLM, SamplingParams
-    llm = LLM(**vars(engine_args))
-
-    # Add the requests to the engine.
-    prompts: List[str] = []
-    sampling_params: List[SamplingParams] = []
-    # create a list containing random selected true or false
-    guided_decoding_req_idx = random.sample(
-        range(len(requests)), int(len(requests) * guided_decoding_rate))
-
-    if warmup:
-        print(">>>>> Running warmup prompt, for the first 5")
-        # We setup the first 5 requests to warmup FSM
-        # if using xgrammar dataset, we will skip warmup
-        warmup_requests = requests[:5]
-        for i, request in enumerate(warmup_requests):
-            prompts.append(request.prompt)
-            sampling_params.append(
-                SamplingParams(
-                    n=n,
-                    temperature=1.0,
-                    top_p=1.0,
-                    ignore_eos=True,
-                    max_tokens=request.expected_output_len,
-                    guided_decoding=GuidedDecodingParams(json=request.schema)
-                    if guided_decoding_rate > 0 else None,
-                ))
-        llm.generate(prompts, sampling_params, use_tqdm=False)
-
-    print(">>>>> Benchmark started...")
-    prompts = []
-    sampling_params = []
-    for i, request in enumerate(requests):
-        prompts.append(request.prompt)
-        sampling_params.append(
-            SamplingParams(
-                n=n,
-                temperature=1.0,
-                top_p=1.0,
-                ignore_eos=True,
-                max_tokens=request.expected_output_len,
-                guided_decoding=GuidedDecodingParams(
-                    **{request.structure_type: request.schema})
-                if i in guided_decoding_req_idx else None,
-            ))
-
-    start = time.perf_counter()
-    outputs = llm.generate(prompts, sampling_params, use_tqdm=False)
-    ret = []
-    for output, request in zip(outputs, requests):
-        generated_text = output.outputs[0].text
-        ret.append({
-            "generated": generated_text,
-            "expected": request.completion
-        })
-    end = time.perf_counter()
-    return end - start, ret
-
-
-async def run_vllm_async(
-        requests: List[SampleRequest],
-        engine_args: AsyncEngineArgs,
-        n: int,
-        guided_decoding_rate: float = 1.0,
-        warmup: bool = False,
-        disable_frontend_multiprocessing: bool = False) -> float:
-    from vllm import SamplingParams
-
-    async with build_async_engine_client_from_engine_args(
-            engine_args, disable_frontend_multiprocessing) as llm:
-
-        # Add the requests to the engine.
-        prompts: List[str] = []
-        sampling_params: List[SamplingParams] = []
-        guided_decoding_req_idx = random.sample(
-            range(len(requests)), int(len(requests) * guided_decoding_rate))
-
-        if warmup:
-            print(">>>>>> Running warmup prompt, for the first 5")
-            # We setup the first 5 requests to warmup FSM
-            # if using xgrammar dataset, we will skip warmup
-            warmup_requests = requests[:5]
-            for i, request in enumerate(warmup_requests):
-                prompts.append(request.prompt)
-                sampling_params.append(
-                    SamplingParams(
-                        n=n,
-                        temperature=1.0,
-                        top_p=1.0,
-                        ignore_eos=True,
-                        max_tokens=request.expected_output_len,
-                        guided_decoding=GuidedDecodingParams(
-                            json=request.schema)
-                        if guided_decoding_rate > 0 else None,
-                    ))
-            generators = []
-            for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
-                generator = llm.generate(prompt, sp, request_id=f"test{i}")
-                generators.append(generator)
-            all_gens = merge_async_iterators(*generators)
-            async for i, res in all_gens:
-                pass
-
-        print(">>>>> Benchmark started...")
-        prompts = []
-        sampling_params = []
-        for i, request in enumerate(requests):
-            prompts.append(request.prompt)
-            sampling_params.append(
-                SamplingParams(
-                    n=n,
-                    temperature=1.0,
-                    top_p=1.0,
-                    ignore_eos=True,
-                    max_tokens=request.expected_output_len,
-                    guided_decoding=GuidedDecodingParams(json=request.schema)
-                    if i in guided_decoding_req_idx else None,
-                ))
-
-        generators = []
-        start_time = []
-        latencies = []
-        start = time.perf_counter()
-        for i, (prompt, sp) in enumerate(zip(prompts, sampling_params)):
-            generator = llm.generate(prompt, sp, request_id=f"test{i}")
-            generators.append(generator)
-            start_time.append(time.perf_counter())
-            latencies.append([])
-        all_gens = merge_async_iterators(*generators)
-        generated_texts = [''] * len(requests)
-        async for i, res in all_gens:
-            generated_texts[i] = res.outputs[0].text
-            lat = time.perf_counter() - start_time[i]
-            latencies[i].append(lat)
-        ret = [{
-            'generated': gt,
-            'expected': req.completion
-        } for gt, req in zip(generated_texts, requests)]
-        end = time.perf_counter()
-        first_latency = pd.Series([lat[0] * 1000 for lat in latencies])
-        next_latency = pd.Series([(lat[-1] - lat[0]) / len(lat[1:]) * 1000
-                                  for lat in latencies])
-        return end - start, ret, (first_latency, next_latency)
-
-
-def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    args: argparse.Namespace) -> List[SampleRequest]:
-    if args.dataset == 'json':
-        if args.json_schema_path is None:
-            dir_path = os.path.dirname(os.path.realpath(__file__))
-            args.json_schema_path = os.path.join(dir_path,
-                                                 "structured_schemas",
-                                                 "structured_schema_1.json")
-        with open(args.json_schema_path) as f:
-            schema = json.load(f)
-        prompt = f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=schema,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "grammar":
-        schema = """
-            ?start: select_statement
-
-            ?select_statement: "SELECT " column_list " FROM " table_name
-
-            ?column_list: column_name ("," column_name)*
-
-            ?table_name: identifier
-
-            ?column_name: identifier
-
-            ?identifier: /[a-zA-Z_][a-zA-Z0-9_]*/
-        """
-        prompt = "Generate an SQL query to show the 'username' \
-            and 'email' from the 'users' table."
-
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=schema,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "regex":
-        regex = r"\w+@\w+\.com\n"
-        args.regex = regex
-        prompt = "Generate an email address for Alan Turing, \
-            who works in Enigma. End in .com and new line. \
-                Example result: alan.turing@enigma.com\n"
-
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=regex,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "choice":
-        choice = ["Positive", "Negative"]
-        args.choice = choice
-        prompt = "Classify this sentiment: vLLM is wonderful!"
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
-        requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
-                          expected_output_len=args.output_len,
-                          schema=choice,
-                          structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
-        ]
-
-    elif args.dataset == "xgrammar_bench":
-        args.warmup = False
-        requests: List[SampleRequest] = []
-        dataset = datasets.load_dataset("NousResearch/json-mode-eval",
-                                        split="train")
-        print(f"dataset has {len(dataset)} entries")
-        len_dataset = len(dataset)
-        for data_point_idx in range(args.num_prompts):
-            idx = data_point_idx
-            while idx >= len_dataset:
-                idx -= len_dataset
-            schema = dataset["schema"][idx]
-            prompt = tokenizer.apply_chat_template(dataset["prompt"][idx],
-                                                   tokenize=False)
-            input_len = len(tokenizer(prompt).input_ids)
-            completion = dataset["completion"][idx]
-
-            requests.append(
-                SampleRequest(prompt=prompt,
-                              prompt_len=input_len,
-                              expected_output_len=args.output_len,
-                              schema=schema,
-                              completion=completion))
-
-    return requests
-
-
-def evaluate(ret, args):
-
-    def _eval_correctness_json(expected, actual):
-        # extract json string from string using regex
-        import re
-        actual = actual.replace('\n', '').replace(' ', '').strip()
-        try:
-            actual = re.search(r'\{.*\}', actual).group()
-            actual = json.loads(actual)
-        except Exception:
-            return False
-
-        return True
-
-    def _eval_correctness_choice(expected, actual):
-        return actual in args.choice
-
-    def _eval_correctness_regex(expected, actual):
-        import re
-        return re.match(args.regex, actual) is not None
-
-    def _eval_correctness(expected, actual):
-        if args.structure_type == 'json':
-            return _eval_correctness_json(expected, actual)
-        elif args.structure_type == 'regex':
-            return _eval_correctness_regex(expected, actual)
-        elif args.structure_type == 'choice':
-            return _eval_correctness_choice(expected, actual)
-        else:
-            return None
-
-    scores = []
-    for res in ret:
-        score = _eval_correctness(res['expected'], res['generated'])
-        res['correctness'] = score
-        scores.append(score)
-
-    not_none_scores = [score for score in scores if score is not None]
-
-    return (sum(not_none_scores) / len(not_none_scores) *
-            100) if len(not_none_scores) > 0 else None
-
-
-def main(args: argparse.Namespace):
-    print(args)
-    random.seed(args.seed)
-
-    # async engine is working for 'regex', 'choice' and 'grammar'
-    if args.dataset == 'grammar':
-        args.structure_type = 'grammar'
-        args.async_engine = False
-    elif args.dataset == 'regex':
-        args.structure_type = 'regex'
-        args.async_engine = False
-    elif args.dataset == 'choice':
-        args.structure_type = 'choice'
-        args.async_engine = False
-    else:
-        args.structure_type = 'json'
-
-    if args.no_guided_decoding:
-        args.guided_decoding_ratio = 0
-    if args.save_results:
-        result_file_name = f'{args.guided_decoding_ratio}guided'
-        result_file_name += f"_{args.model.split('/')[-1]}"
-        result_file_name += f"_{args.dataset}"
-        result_file_name += f"_{args.num_prompts}"
-        result_file_name += f"_out{args.output_len}"
-        result_file_name += f"_async{args.async_engine}"
-        result_file_name += f"_warmup{args.warmup}"
-        result_file_name += f"_chunkedprefill{args.enable_chunked_prefill}"
-        result_file_name += ".txt"
-    else:
-        result_file_name = None
-
-    # Synthesize a prompt with the given input length.
-    tokenizer = AutoTokenizer.from_pretrained(
-        args.tokenizer, trust_remote_code=args.trust_remote_code)
-    requests = sample_requests(tokenizer, args)
-
-    if args.async_engine:
-        engine_args = AsyncEngineArgs.from_cli_args(args)
-        elapsed_time, ret, (first_latency, next_latency) = uvloop.run(
-            run_vllm_async(requests, engine_args, args.n,
-                           args.guided_decoding_ratio, args.warmup,
-                           args.disable_frontend_multiprocessing))
-    else:
-        engine_args = EngineArgs.from_cli_args(args)
-        elapsed_time, ret = run_vllm(requests, engine_args, args.n,
-                                     args.guided_decoding_ratio, args.warmup)
-        first_latency, next_latency = None, None
-
-    score = evaluate(ret, args)
-    total_num_tokens = sum(request.prompt_len + request.expected_output_len
-                           for request in requests)
-    total_output_tokens = sum(request.expected_output_len
-                              for request in requests)
-    if first_latency is not None:
-        latency_breakdown = "\nFirst token latency(msecs):\n"
-        latency_breakdown += f"{first_latency.describe()}"
-        latency_breakdown += "\nNext token latency(msecs):\n"
-        latency_breakdown += f"{next_latency.describe()}"
-    print(
-        f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
-        f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
-        f"{total_output_tokens / elapsed_time:.2f} output tokens/s",
-        f"Correct rate is {score} %",
-        f"{latency_breakdown if first_latency is not None else ''}")
-
-    # Output JSON results if specified
-    if args.output_json or result_file_name:
-        results = {
-            "elapsed_time": elapsed_time,
-            "num_requests": len(requests),
-            "total_num_tokens": total_num_tokens,
-            "total_output_tokens": total_output_tokens,
-            "requests_per_second": len(requests) / elapsed_time,
-            "tokens_per_second": f"{total_num_tokens / elapsed_time:.2f}",
-            "output_tokens_per_second":
-            f"{total_output_tokens / elapsed_time:.2f}",
-            "correct_rate(%)": score
-        }
-        results = {"outputs": ret, **results}
-        if first_latency is not None:
-            results["first_token_latency(msecs)"] = first_latency.describe(
-            ).to_dict()
-            results["next_token_latency(msecs)"] = next_latency.describe(
-            ).to_dict()
-        if args.output_json:
-            with open(args.output_json, "w") as f:
-                json.dump(results, f, indent=4)
-        elif result_file_name:
-            with open(result_file_name, "w") as f:
-                json.dump(results, f, indent=4)
-
-
-if __name__ == "__main__":
-    parser = FlexibleArgumentParser(description="Benchmark guided decoding.")
-    parser = AsyncEngineArgs.add_cli_args(parser)
-
-    parser.add_argument("--output-len",
-                        type=int,
-                        default=512,
-                        help="Output length for each request. Overrides the "
-                        "output length from the dataset.")
-    parser.add_argument(
-        "--dataset",
-        default='json',
-        choices=['json', 'grammar', 'regex', 'choice', 'xgrammar_bench'])
-    parser.add_argument("--json_schema_path",
-                        type=str,
-                        default=None,
-                        help="Path to json schema.")
-    parser.add_argument("--n",
-                        type=int,
-                        default=1,
-                        help="Number of generated sequences per prompt.")
-    parser.add_argument("--num-prompts",
-                        type=int,
-                        default=10,
-                        help="Number of prompts to process.")
-    parser.add_argument(
-        '--output-json',
-        type=str,
-        default=None,
-        help='Path to save the throughput results in JSON format.')
-    parser.add_argument("--async-engine",
-                        action='store_true',
-                        default=False,
-                        help="Use vLLM async engine rather than LLM class.")
-    parser.add_argument("--no-guided-decoding",
-                        action='store_true',
-                        default=False,
-                        help="Whether to disable JSON decoding or not.")
-    parser.add_argument("--guided-decoding-ratio",
-                        type=float,
-                        default=1.0,
-                        help="Ratio of Guided Decoding requests")
-    parser.add_argument("--disable-frontend-multiprocessing",
-                        action='store_true',
-                        default=False,
-                        help="Disable decoupled async engine frontend.")
-    parser.add_argument("--warmup",
-                        action="store_true",
-                        default=False,
-                        help="Run warmup prompts before benchmark.")
-    parser.add_argument("--save-results",
-                        action="store_true",
-                        default=False,
-                        help="save output results.")
-    args = parser.parse_args()
-    if args.tokenizer is None:
-        args.tokenizer = args.model
-    main(args)

benchmarks/benchmark_latency.py

@@ -7,11 +7,11 @@ import json
 import os
 import time
 from pathlib import Path
-from typing import Any, Dict, List, Optional
+from typing import Any, Optional
 
 import numpy as np
 import torch
-from benchmark_utils import convert_to_pytorch_benchmark_format
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from tqdm import tqdm
 
 from vllm import LLM, SamplingParams
@@ -22,7 +22,7 @@ from vllm.utils import FlexibleArgumentParser
 
 
 def save_to_pytorch_benchmark_format(args: argparse.Namespace,
-                                     results: Dict[str, Any]) -> None:
+                                     results: dict[str, Any]) -> None:
     pt_records = convert_to_pytorch_benchmark_format(
         args=args,
         metrics={"latency": results["latencies"]},
@@ -30,8 +30,7 @@ def save_to_pytorch_benchmark_format(args: argparse.Namespace,
                     for k in ["avg_latency", "percentiles"]})
     if pt_records:
         pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
-        with open(pt_file, "w") as f:
-            json.dump(pt_records, f)
+        write_to_json(pt_file, pt_records)
 
 
 def main(args: argparse.Namespace):
@@ -42,6 +41,10 @@ def main(args: argparse.Namespace):
     # NOTE(woosuk): If the request cannot be processed in a single batch,
     # the engine will automatically process the request in multiple batches.
     llm = LLM(**dataclasses.asdict(engine_args))
+    assert llm.llm_engine.model_config.max_model_len >= (
+        args.input_len +
+        args.output_len), ("Please ensure that max_model_len is greater than"
+                           " the sum of input_len and output_len.")
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -49,12 +52,13 @@ def main(args: argparse.Namespace):
         top_p=1.0,
         ignore_eos=True,
         max_tokens=args.output_len,
+        detokenize=not args.disable_detokenize,
     )
     print(sampling_params)
     dummy_prompt_token_ids = np.random.randint(10000,
                                                size=(args.batch_size,
                                                      args.input_len))
-    dummy_prompts: List[PromptType] = [{
+    dummy_prompts: list[PromptType] = [{
         "prompt_token_ids": batch
     } for batch in dummy_prompt_token_ids.tolist()]
 
@@ -170,6 +174,12 @@ if __name__ == "__main__":
         default=None,
         help="Path to save the latency results in JSON format.",
     )
+    parser.add_argument(
+        "--disable-detokenize",
+        action="store_true",
+        help=("Do not detokenize responses (i.e. do not include "
+              "detokenization time in the latency measurement)"),
+    )
 
     parser = EngineArgs.add_cli_args(parser)
     args = parser.parse_args()

benchmarks/benchmark_prefix_caching.py

@@ -31,7 +31,7 @@ import dataclasses
 import json
 import random
 import time
-from typing import List, Optional, Tuple
+from typing import Optional
 
 from transformers import PreTrainedTokenizerBase
 
@@ -77,9 +77,9 @@ def sample_requests_from_dataset(
     dataset_path: str,
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
-    input_length_range: Tuple[int, int],
+    input_length_range: tuple[int, int],
     fixed_output_len: Optional[int],
-) -> List[Request]:
+) -> list[Request]:
     if fixed_output_len is not None and fixed_output_len < 4:
         raise ValueError("output_len too small")
 
@@ -99,7 +99,7 @@ def sample_requests_from_dataset(
     assert min_len >= 0 and max_len >= min_len, "input_length_range too small"
 
     # Filter out sequences that are too long or too short
-    filtered_requests: List[Request] = []
+    filtered_requests: list[Request] = []
 
     for i in range(len(dataset)):
         if len(filtered_requests) == num_requests:
@@ -122,10 +122,10 @@ def sample_requests_from_dataset(
 def sample_requests_from_random(
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
-    input_length_range: Tuple[int, int],
+    input_length_range: tuple[int, int],
     fixed_output_len: Optional[int],
     prefix_len: int,
-) -> List[Request]:
+) -> list[Request]:
 
     requests = []
     prefix_token_ids = sample_tokens(tokenizer, prefix_len)
@@ -144,9 +144,9 @@ def sample_requests_from_random(
     return requests
 
 
-def repeat_and_sort_requests(requests: List[Request],
+def repeat_and_sort_requests(requests: list[Request],
                              repeat_count: int,
-                             sort: bool = False) -> List[str]:
+                             sort: bool = False) -> list[str]:
     repeated_requests = requests * repeat_count
     if sort:
         repeated_requests.sort(key=lambda x: x[1])
@@ -194,7 +194,9 @@ def main(args):
 
     llm = LLM(**dataclasses.asdict(engine_args))
 
-    sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
+    sampling_params = SamplingParams(temperature=0,
+                                     max_tokens=args.output_len,
+                                     detokenize=not args.disable_detokenize)
 
     print("Testing filtered requests")
     prompts = repeat_and_sort_requests(filtered_requests,
@@ -243,6 +245,12 @@ if __name__ == "__main__":
         "subtract this length when filtering prompts. Only used "
         "when dataset-path is not provided.",
     )
+    parser.add_argument(
+        '--disable-detokenize',
+        action='store_true',
+        help=("Do not detokenize responses (i.e. do not include "
+              "detokenization time in the latency measurement)"),
+    )
 
     parser = EngineArgs.add_cli_args(parser)
     args = parser.parse_args()

benchmarks/benchmark_prioritization.py

@@ -5,7 +5,7 @@ import dataclasses
 import json
 import random
 import time
-from typing import List, Optional, Tuple
+from typing import Optional
 
 from transformers import AutoTokenizer, PreTrainedTokenizerBase
 
@@ -13,12 +13,17 @@ from vllm.engine.arg_utils import EngineArgs
 from vllm.utils import FlexibleArgumentParser
 
 
+#Select a equi-probable random priority
+def get_random_flag():
+    return 0 if random.random() < 0.5 else 1
+
+
 def sample_requests(
     dataset_path: str,
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
     fixed_output_len: Optional[int],
-) -> List[Tuple[str, int, int]]:
+) -> list[tuple[str, int, int, int]]:
     if fixed_output_len is not None and fixed_output_len < 4:
         raise ValueError("output_len too small")
 
@@ -35,7 +40,7 @@ def sample_requests(
     random.shuffle(dataset)
 
     # Filter out sequences that are too long or too short
-    filtered_dataset: List[Tuple[str, int, int]] = []
+    filtered_dataset: list[tuple[str, int, int]] = []
     for i in range(len(dataset)):
         if len(filtered_dataset) == num_requests:
             break
@@ -55,8 +60,7 @@ def sample_requests(
             # Prune too long sequences.
             continue
 
-        #Select a equi-probable random priority
-        priority = 0 if random.random() < 0.5 else 1
+        priority = get_random_flag()
 
         filtered_dataset.append((prompt, prompt_len, output_len, priority))
 
@@ -64,13 +68,20 @@ def sample_requests(
 
 
 def run_vllm(
-    requests: List[Tuple[str, int, int]],
+    requests: list[tuple[str, int, int]],
     n: int,
     engine_args: EngineArgs,
+    disable_detokenize: bool = False,
 ) -> float:
     from vllm import LLM, SamplingParams
     llm = LLM(**dataclasses.asdict(engine_args))
 
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (request[1] + request[2])
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " input_len and output_len for all requests.")
+
     # Add the requests to the engine.
     prompts = []
     sampling_params = []
@@ -85,6 +96,7 @@ def run_vllm(
                 top_p=1.0,
                 ignore_eos=True,
                 max_tokens=output_len,
+                detokenize=not disable_detokenize,
             ))
 
     start = time.perf_counter()
@@ -103,15 +115,16 @@ def main(args: argparse.Namespace):
     if args.dataset is None:
         # Synthesize a prompt with the given input length.
         prompt = "hi" * (args.input_len - 1)
-        requests = [(prompt, args.input_len, args.output_len)
-                    for _ in range(args.num_prompts)]
+        requests = [(prompt, args.input_len, args.output_len,
+                     get_random_flag()) for _ in range(args.num_prompts)]
     else:
         requests = sample_requests(args.dataset, args.num_prompts, tokenizer,
                                    args.output_len)
 
     if args.backend == "vllm":
         elapsed_time = run_vllm(requests, args.n,
-                                EngineArgs.from_cli_args(args))
+                                EngineArgs.from_cli_args(args),
+                                args.disable_detokenize)
     else:
         raise ValueError(f"Unknown backend: {args.backend}")
     total_num_tokens = sum(prompt_len + output_len
@@ -164,6 +177,12 @@ if __name__ == "__main__":
         type=str,
         default=None,
         help='Path to save the throughput results in JSON format.')
+    parser.add_argument(
+        '--disable-detokenize',
+        action='store_true',
+        help=("Do not detokenize responses (i.e. do not include "
+              "detokenization time in the latency measurement)"),
+    )
 
     parser = EngineArgs.add_cli_args(parser)
     args = parser.parse_args()

benchmarks/benchmark_serving.py

@@ -7,9 +7,6 @@ On the server side, run one of the following commands:
         --swap-space 16 \
         --disable-log-requests
 
-    (TGI backend)
-    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
-
 On the client side, run:
     python benchmarks/benchmark_serving.py \
         --backend <backend> \
@@ -25,24 +22,21 @@ On the client side, run:
 """
 import argparse
 import asyncio
-import base64
 import gc
-import io
 import json
 import os
 import random
 import time
 import warnings
+from collections.abc import AsyncGenerator, Iterable
 from dataclasses import dataclass
 from datetime import datetime
-from typing import Any, AsyncGenerator, Collection, Dict, List, Optional, Tuple
+from typing import Any, Optional
 
 import numpy as np
-import pandas as pd
-from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
+from backend_request_func import (ASYNC_REQUEST_FUNCS,
+                                  OPENAI_COMPATIBLE_BACKENDS, RequestFuncInput,
                                   RequestFuncOutput)
-from datasets import load_dataset
-from PIL.Image import Image
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
 
@@ -56,7 +50,12 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
-from benchmark_utils import convert_to_pytorch_benchmark_format
+from benchmark_dataset import (AIMODataset, BurstGPTDataset,
+                               ConversationDataset, HuggingFaceDataset,
+                               InstructCoderDataset, RandomDataset,
+                               SampleRequest, ShareGPTDataset, SonnetDataset,
+                               VisionArenaDataset)
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
 
@@ -73,343 +72,36 @@ class BenchmarkMetrics:
     mean_ttft_ms: float
     median_ttft_ms: float
     std_ttft_ms: float
-    percentiles_ttft_ms: List[Tuple[float, float]]
+    percentiles_ttft_ms: list[tuple[float, float]]
     mean_tpot_ms: float
     median_tpot_ms: float
     std_tpot_ms: float
-    percentiles_tpot_ms: List[Tuple[float, float]]
+    percentiles_tpot_ms: list[tuple[float, float]]
     mean_itl_ms: float
     median_itl_ms: float
     std_itl_ms: float
-    percentiles_itl_ms: List[Tuple[float, float]]
+    percentiles_itl_ms: list[tuple[float, float]]
     # E2EL stands for end-to-end latency per request.
     # It is the time taken on the client side from sending
     # a request to receiving a complete response.
     mean_e2el_ms: float
     median_e2el_ms: float
     std_e2el_ms: float
-    percentiles_e2el_ms: List[Tuple[float, float]]
-
-
-def sample_sharegpt_requests(
-    dataset_path: str,
-    num_requests: int,
-    tokenizer: PreTrainedTokenizerBase,
-    fixed_output_len: Optional[int] = None,
-) -> List[Tuple[str, int, int, None]]:
-    # Load the dataset.
-    with open(dataset_path, encoding='utf-8') as f:
-        dataset = json.load(f)
-    # Filter out the conversations with less than 2 turns.
-    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
-    # Only keep the first two turns of each conversation.
-    dataset = [(data["conversations"][0]["value"],
-                data["conversations"][1]["value"]) for data in dataset]
-
-    # Shuffle the dataset.
-    random.shuffle(dataset)
-
-    # Filter out sequences that are too long or too short
-    filtered_dataset: List[Tuple[str, int, int]] = []
-    for i in range(len(dataset)):
-        if len(filtered_dataset) == num_requests:
-            break
-
-        # Tokenize the prompts and completions.
-        prompt = dataset[i][0]
-        prompt_token_ids = tokenizer(prompt).input_ids
-        completion = dataset[i][1]
-        completion_token_ids = tokenizer(completion).input_ids
-        prompt_len = len(prompt_token_ids)
-        output_len = len(completion_token_ids
-                         ) if fixed_output_len is None else fixed_output_len
-        if prompt_len < 4 or (fixed_output_len is None and output_len < 4):
-            # Prune too short sequences.
-            continue
-        if prompt_len > 1024 or prompt_len + output_len > 2048:
-            # Prune too long sequences.
-            continue
-        filtered_dataset.append((prompt, prompt_len, output_len, None))
-
-    return filtered_dataset
-
-
-def sample_burstgpt_requests(
-    dataset_path: str,
-    num_requests: int,
-    random_seed: int,
-    tokenizer: PreTrainedTokenizerBase,
-) -> List[Tuple[str, int, int, None]]:
-    df = pd.read_csv(dataset_path)
-    gpt4_df = df[df["Model"] == "GPT-4"]
-    # Remove the failed requests (i.e., response length is 0)
-    gpt4_df = gpt4_df[gpt4_df["Response tokens"] > 0]
-    # Randomly sample num_requests from the dataset
-    if num_requests <= len(gpt4_df):
-        gpt4_df = gpt4_df.sample(n=num_requests, random_state=random_seed)
-    else:
-        gpt4_df = gpt4_df.sample(n=num_requests,
-                                 random_state=random_seed,
-                                 replace=True)
-    # Convert the dataframe to a list of tuples
-    dataset = gpt4_df.values.tolist()
-    input_requests = []
-    for i in range(num_requests):
-        input_len = int(dataset[i][2])
-        output_len = int(dataset[i][3])
-        prompt = tokenizer.decode([(i + j) % tokenizer.vocab_size
-                                   for j in range(input_len)])
-        input_requests.append((prompt, input_len, output_len, None))
-    return input_requests
-
-
-def sample_sonnet_requests(
-    dataset_path: str,
-    num_requests: int,
-    input_len: int,
-    output_len: int,
-    prefix_len: int,
-    tokenizer: PreTrainedTokenizerBase,
-) -> List[Tuple[str, str, int, int, None]]:
-    assert (
-        input_len > prefix_len
-    ), "'args.sonnet-input-len' must be greater than 'args.prefix-input-len'."
-
-    # Load the dataset.
-    with open(dataset_path, encoding='utf-8') as f:
-        poem_lines = f.readlines()
-
-    # Tokenize the poem lines.
-    poem_token_ids = tokenizer(poem_lines).input_ids
-    average_poem_len = sum(
-        len(token_ids) for token_ids in poem_token_ids) / len(poem_token_ids)
-
-    # Base prefix for all requests.
-    base_prompt = "Pick as many lines as you can from these poem lines:\n"
-    base_message = [{
-        "role": "user",
-        "content": base_prompt,
-    }]
-    base_prompt_formatted = tokenizer.apply_chat_template(
-        base_message, add_generation_prompt=True, tokenize=False)
-    base_prompt_offset = len(tokenizer(base_prompt_formatted).input_ids)
-
-    assert (
-        input_len > base_prompt_offset
-    ), f"Please set 'args.sonnet-input-len' higher than {base_prompt_offset}."
-    num_input_lines = round(
-        (input_len - base_prompt_offset) / average_poem_len)
-
-    # First approximately `prefix_len` number of tokens in the
-    # prompt are fixed poem lines.
-    assert (
-        prefix_len > base_prompt_offset
-    ), f"Please set 'args.sonnet-prefix-len' higher than {base_prompt_offset}."
-
-    num_prefix_lines = round(
-        (prefix_len - base_prompt_offset) / average_poem_len)
-    prefix_lines = poem_lines[:num_prefix_lines]
-
-    # Sample the rest of lines per request.
-    sampled_requests: List[Tuple[str, int, int]] = []
-    for _ in range(num_requests):
-        num_lines_needed = num_input_lines - num_prefix_lines
-        sampled_lines = "".join(prefix_lines +
-                                random.choices(poem_lines, k=num_lines_needed))
-
-        prompt = f"{base_prompt}{sampled_lines}"
-        message = [
-            {
-                "role": "user",
-                "content": prompt,
-            },
-        ]
-        prompt_formatted = tokenizer.apply_chat_template(
-            message, add_generation_prompt=True, tokenize=False)
-        prompt_len = len(tokenizer(prompt_formatted).input_ids)
-        sampled_requests.append(
-            (prompt, prompt_formatted, prompt_len, output_len, None))
-
-    return sampled_requests
-
-
-def sample_vision_arena_requests(
-    dataset,
-    num_requests: int,
-    tokenizer: PreTrainedTokenizerBase,
-    fixed_output_len: Optional[int] = None,
-) -> List[Tuple[str, str, int, Optional[Dict[str, Collection[str]]]]]:
-    sampled_requests: List[Tuple[str, int, int, Dict[str,
-                                                     Collection[str]]]] = []
-    for data in dataset:
-        if len(sampled_requests) == num_requests:
-            break
-
-        prompt = data["turns"][0][0]['content']
-
-        prompt_token_ids = tokenizer(prompt).input_ids
-        if fixed_output_len is None:
-            # Default max output len is set to 128
-            print("--hf-output-len is not provided. Using default value 128.")
-            fixed_output_len = 128
-
-        prompt_len = len(prompt_token_ids)
-        output_len = fixed_output_len
-
-        assert isinstance(
-            data["images"][0],
-            Image), ("Input image format must be `PIL.Image.Image`, "
-                     f"given {type(data['image'])}.")
-        image: Image = data["images"][0]
-        image = image.convert("RGB")
-        image_data = io.BytesIO()
-        image.save(image_data, format='JPEG')
-        image_base64 = base64.b64encode(image_data.getvalue()).decode("utf-8")
-        mm_content = {
-            "type": "image_url",
-            "image_url": {
-                "url": f"data:image/jpeg;base64,{image_base64}"
-            },
-        }
-
-        sampled_requests.append((prompt, prompt_len, output_len, mm_content))
-
-    return sampled_requests
-
-
-def sample_hf_requests(
-    dataset_path: str,
-    dataset_subset: Optional[str],
-    dataset_split: str,
-    num_requests: int,
-    tokenizer: PreTrainedTokenizerBase,
-    random_seed: int,
-    fixed_output_len: Optional[int] = None,
-) -> List[Tuple[str, str, int, Optional[Dict[str, Collection[str]]]]]:
-
-    # Special case for vision_arena dataset
-    if dataset_path == 'lmarena-ai/vision-arena-bench-v0.1' \
-        and dataset_subset is None:
-        assert dataset_split == "train"
-        dataset = load_dataset(dataset_path,
-                               name=dataset_subset,
-                               split=dataset_split,
-                               streaming=True)
-        dataset = dataset.shuffle(seed=random_seed)
-        return sample_vision_arena_requests(dataset, num_requests, tokenizer,
-                                            fixed_output_len)
-
-    dataset = load_dataset(dataset_path,
-                           name=dataset_subset,
-                           split=dataset_split,
-                           streaming=True)
-    assert "conversations" in dataset.features, (
-        "HF Dataset must have 'conversations' column.")
-    filter_func = lambda x: len(x["conversations"]) >= 2
-    filtered_dataset = dataset.shuffle(seed=random_seed).filter(filter_func)
-    sampled_requests: List[Tuple[str, int, int, Dict[str,
-                                                     Collection[str]]]] = []
-    for data in filtered_dataset:
-        if len(sampled_requests) == num_requests:
-            break
-
-        # Tokenize the prompts and completions.
-        prompt = data["conversations"][0]["value"]
-        prompt_token_ids = tokenizer(prompt).input_ids
-        completion = data["conversations"][1]["value"]
-        completion_token_ids = tokenizer(completion).input_ids
-        prompt_len = len(prompt_token_ids)
-        output_len = len(completion_token_ids
-                         ) if fixed_output_len is None else fixed_output_len
-        if fixed_output_len is None and (prompt_len < 4 or output_len < 4):
-            # Prune too short sequences.
-            continue
-        if fixed_output_len is None and \
-            (prompt_len > 1024 or prompt_len + output_len > 2048):
-            # Prune too long sequences.
-            continue
-
-        if "image" in data and isinstance(data["image"], Image):
-            image: Image = data["image"]
-            image = image.convert("RGB")
-            image_data = io.BytesIO()
-            image.save(image_data, format='JPEG')
-            image_base64 = base64.b64encode(
-                image_data.getvalue()).decode("utf-8")
-            mm_content = {
-                "type": "image_url",
-                "image_url": {
-                    "url": f"data:image/jpeg;base64,{image_base64}"
-                },
-            }
-        elif "image" in data and isinstance(data["image"], str):
-            if (data["image"].startswith("http://") or \
-                data["image"].startswith("file://")):
-                image_url = data["image"]
-            else:
-                image_url = f"file://{data['image']}"
-
-            mm_content = {
-                "type": "image_url",
-                "image_url": {
-                    "url": image_url
-                },
-            }
-        else:
-            mm_content = None
-
-        sampled_requests.append((prompt, prompt_len, output_len, mm_content))
-
-    return sampled_requests
-
-
-def sample_random_requests(
-    prefix_len: int,
-    input_len: int,
-    output_len: int,
-    num_prompts: int,
-    range_ratio: float,
-    tokenizer: PreTrainedTokenizerBase,
-) -> List[Tuple[str, int, int]]:
-    prefix_token_ids = np.random.randint(0,
-                                         tokenizer.vocab_size,
-                                         size=prefix_len).tolist()
-
-    input_lens = np.random.randint(
-        int(input_len * range_ratio),
-        input_len + 1,
-        size=num_prompts,
-    )
-    output_lens = np.random.randint(
-        int(output_len * range_ratio),
-        output_len + 1,
-        size=num_prompts,
-    )
-    offsets = np.random.randint(0, tokenizer.vocab_size, size=num_prompts)
-    input_requests = []
-    for i in range(num_prompts):
-        prompt = tokenizer.decode(prefix_token_ids +
-                                  [(offsets[i] + i + j) % tokenizer.vocab_size
-                                   for j in range(input_lens[i])])
-
-        input_requests.append((prompt, int(prefix_len + input_lens[i]),
-                               int(output_lens[i]), None))
-
-    return input_requests
+    percentiles_e2el_ms: list[tuple[float, float]]
 
 
 async def get_request(
-    input_requests: List[Tuple[str, int, int]],
+    input_requests: list[SampleRequest],
     request_rate: float,
     burstiness: float = 1.0,
-) -> AsyncGenerator[Tuple[str, int, int], None]:
+) -> AsyncGenerator[SampleRequest, None]:
     """
     Asynchronously generates requests at a specified rate
     with OPTIONAL burstiness.
 
     Args:
         input_requests:
-            A list of input requests, each represented as a tuple.
+            A list of input requests, each represented as a SampleRequest.
         request_rate:
             The rate at which requests are generated (requests/s).
         burstiness (optional):
@@ -421,7 +113,7 @@ async def get_request(
             in more bursty requests, while a higher burstiness value
             (burstiness > 1) results in a more uniform arrival of requests.
     """
-    input_requests = iter(input_requests)
+    input_requests: Iterable[SampleRequest] = iter(input_requests)
 
     # Calculate scale parameter theta to maintain the desired request_rate.
     assert burstiness > 0, (
@@ -443,28 +135,28 @@ async def get_request(
 
 
 def calculate_metrics(
-    input_requests: List[Tuple[str, int, int]],
-    outputs: List[RequestFuncOutput],
+    input_requests: list[SampleRequest],
+    outputs: list[RequestFuncOutput],
     dur_s: float,
     tokenizer: PreTrainedTokenizerBase,
-    selected_percentile_metrics: List[str],
-    selected_percentiles: List[float],
-    goodput_config_dict: Dict[str, float],
-) -> Tuple[BenchmarkMetrics, List[int]]:
-    actual_output_lens: List[int] = []
+    selected_percentile_metrics: list[str],
+    selected_percentiles: list[float],
+    goodput_config_dict: dict[str, float],
+) -> tuple[BenchmarkMetrics, list[int]]:
+    actual_output_lens: list[int] = []
     total_input = 0
     completed = 0
     good_completed = 0
-    itls: List[float] = []
-    tpots: List[float] = []
-    all_tpots: List[float] = []
-    ttfts: List[float] = []
-    e2els: List[float] = []
+    itls: list[float] = []
+    tpots: list[float] = []
+    all_tpots: list[float] = []
+    ttfts: list[float] = []
+    e2els: list[float] = []
     for i in range(len(outputs)):
         if outputs[i].success:
             output_len = outputs[i].output_tokens
 
-            if output_len is None:
+            if not output_len:
                 # We use the tokenizer to count the number of output tokens
                 # for some serving backends instead of looking at
                 # len(outputs[i].itl) since multiple output tokens may be
@@ -474,7 +166,7 @@ def calculate_metrics(
                     tokenizer(outputs[i].generated_text,
                               add_special_tokens=False).input_ids)
             actual_output_lens.append(output_len)
-            total_input += input_requests[i][1]
+            total_input += input_requests[i].prompt_len
             tpot = 0
             if output_len > 1:
                 latency_minus_ttft = outputs[i].latency - outputs[i].ttft
@@ -557,19 +249,19 @@ async def benchmark(
     model_id: str,
     model_name: str,
     tokenizer: PreTrainedTokenizerBase,
-    input_requests: List[Tuple[str, int, int]],
+    input_requests: list[SampleRequest],
     logprobs: Optional[int],
-    best_of: int,
     request_rate: float,
     burstiness: float,
     disable_tqdm: bool,
     profile: bool,
-    selected_percentile_metrics: List[str],
-    selected_percentiles: List[str],
+    selected_percentile_metrics: list[str],
+    selected_percentiles: list[float],
     ignore_eos: bool,
-    goodput_config_dict: Dict[str, float],
+    goodput_config_dict: dict[str, float],
     max_concurrency: Optional[int],
-    lora_modules: Optional[List[str]],
+    lora_modules: Optional[Iterable[str]],
+    extra_body: Optional[dict],
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -577,12 +269,16 @@ async def benchmark(
         raise ValueError(f"Unknown backend: {backend}")
 
     print("Starting initial single prompt test run...")
-    test_prompt, test_prompt_len, test_output_len, test_mm_content = (
-        input_requests[0])
+    test_prompt, test_prompt_len, test_output_len, test_mm_content = \
+        input_requests[0].prompt, input_requests[0].prompt_len, \
+        input_requests[0].expected_output_len, \
+            input_requests[0].multi_modal_data
+
     if backend != "openai-chat" and test_mm_content is not None:
         # multi-modal benchmark is only available on OpenAI Chat backend.
         raise ValueError(
             "Multi-modal content is only supported on 'openai-chat' backend.")
+    assert test_mm_content is None or isinstance(test_mm_content, dict)
     test_input = RequestFuncInput(
         model=model_id,
         model_name=model_name,
@@ -591,9 +287,9 @@ async def benchmark(
         prompt_len=test_prompt_len,
         output_len=test_output_len,
         logprobs=logprobs,
-        best_of=best_of,
         multi_modal_content=test_mm_content,
         ignore_eos=ignore_eos,
+        extra_body=extra_body,
     )
 
     test_output = await request_func(request_func_input=test_input)
@@ -607,7 +303,8 @@ async def benchmark(
     if lora_modules:
         # For each input request, choose a LoRA module at random.
         lora_modules = iter(
-            [random.choice(lora_modules) for _ in range(len(input_requests))])
+            [random.choice(lora_modules) \
+                for _ in range(len(input_requests))])
 
     if profile:
         print("Starting profiler...")
@@ -618,9 +315,9 @@ async def benchmark(
                                          prompt_len=test_prompt_len,
                                          output_len=test_output_len,
                                          logprobs=logprobs,
-                                         best_of=best_of,
                                          multi_modal_content=test_mm_content,
-                                         ignore_eos=ignore_eos)
+                                         ignore_eos=ignore_eos,
+                                         extra_body=extra_body)
         profile_output = await request_func(request_func_input=profile_input)
         if profile_output.success:
             print("Profiler started")
@@ -652,9 +349,11 @@ async def benchmark(
                                       pbar=pbar)
 
     benchmark_start_time = time.perf_counter()
-    tasks: List[asyncio.Task] = []
+    tasks: list[asyncio.Task] = []
     async for request in get_request(input_requests, request_rate, burstiness):
-        prompt, prompt_len, output_len, mm_content = request
+        prompt, prompt_len, output_len, mm_content = request.prompt, \
+            request.prompt_len, request.expected_output_len, \
+                request.multi_modal_data
         req_model_id, req_model_name = model_id, model_name
         if lora_modules:
             req_lora_module = next(lora_modules)
@@ -667,14 +366,14 @@ async def benchmark(
                                               prompt_len=prompt_len,
                                               output_len=output_len,
                                               logprobs=logprobs,
-                                              best_of=best_of,
                                               multi_modal_content=mm_content,
-                                              ignore_eos=ignore_eos)
+                                              ignore_eos=ignore_eos,
+                                              extra_body=extra_body)
         tasks.append(
             asyncio.create_task(
                 limited_request_func(request_func_input=request_func_input,
                                      pbar=pbar)))
-    outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
+    outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
 
     if profile:
         print("Stopping profiler...")
@@ -685,7 +384,6 @@ async def benchmark(
             prompt_len=test_prompt_len,
             output_len=test_output_len,
             logprobs=logprobs,
-            best_of=best_of,
         )
         profile_output = await request_func(request_func_input=profile_input)
         if profile_output.success:
@@ -820,7 +518,7 @@ def parse_goodput(slo_pairs):
 
 
 def save_to_pytorch_benchmark_format(args: argparse.Namespace,
-                                     results: Dict[str, Any],
+                                     results: dict[str, Any],
                                      file_name: str) -> None:
     metrics = [
         "median_ttft_ms", "mean_ttft_ms", "std_ttft_ms", "p99_ttft_ms",
@@ -841,8 +539,7 @@ def save_to_pytorch_benchmark_format(args: argparse.Namespace,
     if pt_records:
         # Don't use json suffix here as we don't want CI to pick it up
         pt_file = f"{os.path.splitext(file_name)[0]}.pytorch.json"
-        with open(pt_file, "w") as f:
-            json.dump(pt_records, f)
+        write_to_json(pt_file, pt_records)
 
 
 def main(args: argparse.Namespace):
@@ -867,91 +564,119 @@ def main(args: argparse.Namespace):
                               tokenizer_mode=tokenizer_mode,
                               trust_remote_code=args.trust_remote_code)
 
-    if args.dataset is not None:
-        warnings.warn(
-            "The '--dataset' argument will be deprecated in the next "
-            "release. Please use '--dataset-name' and "
-            "'--dataset-path' in the future runs.",
-            stacklevel=2)
-        input_requests = sample_sharegpt_requests(
-            dataset_path=args.dataset,
-            num_requests=args.num_prompts,
-            tokenizer=tokenizer,
-            fixed_output_len=args.sharegpt_output_len,
-        )
+    if args.dataset_name is None:
+        raise ValueError(
+            "Please specify '--dataset-name' and the corresponding "
+            "'--dataset-path' if required.")
 
-    elif args.dataset_name == "sharegpt":
-        input_requests = sample_sharegpt_requests(
-            dataset_path=args.dataset_path,
-            num_requests=args.num_prompts,
-            tokenizer=tokenizer,
-            fixed_output_len=args.sharegpt_output_len,
-        )
-
-    elif args.dataset_name == "burstgpt":
-        input_requests = sample_burstgpt_requests(
-            dataset_path=args.dataset_path,
-            num_requests=args.num_prompts,
-            random_seed=args.seed,
-            tokenizer=tokenizer,
-        )
-
-    elif args.dataset_name == "sonnet":
-        # Do not format the prompt, pass to message directly
+    if 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":
-            input_requests = sample_sonnet_requests(
-                dataset_path=args.dataset_path,
-                num_requests=args.num_prompts,
-                input_len=args.sonnet_input_len,
-                output_len=args.sonnet_output_len,
-                prefix_len=args.sonnet_prefix_len,
-                tokenizer=tokenizer,
-            )
-            input_requests = [(prompt, prompt_len, output_len, None)
-                              for prompt, prompt_formatted, prompt_len,
-                              output_len, _ in input_requests]
+            input_requests = dataset.sample(num_requests=args.num_prompts,
+                                            input_len=args.sonnet_input_len,
+                                            output_len=args.sonnet_output_len,
+                                            prefix_len=args.sonnet_prefix_len,
+                                            tokenizer=tokenizer,
+                                            return_prompt_formatted=False)
         else:
-            assert (
-                tokenizer.chat_template or tokenizer.default_chat_template
-            ), "Tokenizer/model must have chat template for sonnet dataset."
-            input_requests = sample_sonnet_requests(
-                dataset_path=args.dataset_path,
-                num_requests=args.num_prompts,
-                input_len=args.sonnet_input_len,
-                output_len=args.sonnet_output_len,
-                prefix_len=args.sonnet_prefix_len,
-                tokenizer=tokenizer,
-            )
-            input_requests = [(prompt_formatted, prompt_len, output_len, None)
-                              for prompt, prompt_formatted, prompt_len,
-                              output_len, _ in input_requests]
+            assert tokenizer.chat_template or tokenizer.default_chat_template, (
+                "Tokenizer/model must have chat template for sonnet dataset.")
+            input_requests = dataset.sample(num_requests=args.num_prompts,
+                                            input_len=args.sonnet_input_len,
+                                            output_len=args.sonnet_output_len,
+                                            prefix_len=args.sonnet_prefix_len,
+                                            tokenizer=tokenizer,
+                                            return_prompt_formatted=True)
 
     elif args.dataset_name == "hf":
-        input_requests = sample_hf_requests(
+        # all following datasets are implemented from the
+        # HuggingFaceDataset base class
+        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = VisionArenaDataset
+            args.hf_split = "train"
+            args.hf_subset = None
+        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = InstructCoderDataset
+            args.hf_split = "train"
+        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = ConversationDataset
+        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
+            dataset_class = AIMODataset
+            args.hf_split = "train"
+        else:
+            supported_datasets = set([
+                dataset_name for cls in HuggingFaceDataset.__subclasses__()
+                for dataset_name in cls.SUPPORTED_DATASET_PATHS
+            ])
+            raise ValueError(
+                f"Unsupported dataset path: {args.dataset_path}. "
+                "Huggingface dataset only supports dataset_path"
+                f" from one of following: {supported_datasets}. "
+                "Please consider contributing if you would "
+                "like to add support for additional dataset formats.")
+        input_requests = dataset_class(
             dataset_path=args.dataset_path,
             dataset_subset=args.hf_subset,
             dataset_split=args.hf_split,
+            random_seed=args.seed,
+        ).sample(
             num_requests=args.num_prompts,
             tokenizer=tokenizer,
-            random_seed=args.seed,
-            fixed_output_len=args.hf_output_len,
-        )
-
-    elif args.dataset_name == "random":
-        input_requests = sample_random_requests(
-            prefix_len=args.random_prefix_len,
-            input_len=args.random_input_len,
-            output_len=args.random_output_len,
-            num_prompts=args.num_prompts,
-            range_ratio=args.random_range_ratio,
-            tokenizer=tokenizer,
+            output_len=args.hf_output_len,
         )
 
     else:
-        raise ValueError(f"Unknown dataset: {args.dataset_name}")
+        # For datasets that follow a similar structure, use a mapping.
+        dataset_mapping = {
+            "sharegpt":
+            lambda: ShareGPTDataset(random_seed=args.seed,
+                                    dataset_path=args.dataset_path).sample(
+                                        tokenizer=tokenizer,
+                                        num_requests=args.num_prompts,
+                                        output_len=args.sharegpt_output_len,
+                                    ),
+            "burstgpt":
+            lambda: BurstGPTDataset(random_seed=args.seed,
+                                    dataset_path=args.dataset_path).
+            sample(tokenizer=tokenizer, num_requests=args.num_prompts),
+            "random":
+            lambda: RandomDataset(dataset_path=args.dataset_path).sample(
+                tokenizer=tokenizer,
+                num_requests=args.num_prompts,
+                prefix_len=args.random_prefix_len,
+                input_len=args.random_input_len,
+                output_len=args.random_output_len,
+                range_ratio=args.random_range_ratio,
+            )
+        }
 
+        try:
+            input_requests = dataset_mapping[args.dataset_name]()
+        except KeyError as err:
+            raise ValueError(f"Unknown dataset: {args.dataset_name}") from err
     goodput_config_dict = check_goodput_args(args)
 
+    # Collect the sampling parameters.
+    sampling_params = {
+        k: v
+        for k, v in {
+            "top_p": args.top_p,
+            "top_k": args.top_k,
+            "min_p": args.min_p,
+            "temperature": args.temperature
+        }.items() if v is not None
+    }
+
+    # Sampling parameters are only supported by openai-compatible backend.
+    if sampling_params and args.backend not in OPENAI_COMPATIBLE_BACKENDS:
+        raise ValueError(
+            "Sampling parameters are only supported by openai-compatible "
+            "backends.")
+
+    if "temperature" not in sampling_params:
+        sampling_params["temperature"] = 0.0  # Default to greedy decoding.
+
     # Avoid GC processing "static" data - reduce pause times.
     gc.collect()
     gc.freeze()
@@ -966,7 +691,6 @@ def main(args: argparse.Namespace):
             tokenizer=tokenizer,
             input_requests=input_requests,
             logprobs=args.logprobs,
-            best_of=args.best_of,
             request_rate=args.request_rate,
             burstiness=args.burstiness,
             disable_tqdm=args.disable_tqdm,
@@ -979,11 +703,12 @@ def main(args: argparse.Namespace):
             goodput_config_dict=goodput_config_dict,
             max_concurrency=args.max_concurrency,
             lora_modules=args.lora_modules,
+            extra_body=sampling_params,
         ))
 
     # Save config and results to json
     if args.save_result:
-        result_json: Dict[str, Any] = {}
+        result_json: dict[str, Any] = {}
 
         # Setup
         current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
@@ -991,7 +716,6 @@ def main(args: argparse.Namespace):
         result_json["backend"] = backend
         result_json["model_id"] = model_id
         result_json["tokenizer_id"] = tokenizer_id
-        result_json["best_of"] = args.best_of
         result_json["num_prompts"] = args.num_prompts
 
         # Metadata
@@ -1005,6 +729,15 @@ def main(args: argparse.Namespace):
                         "Invalid metadata format. Please use KEY=VALUE format."
                     )
 
+        if not args.save_detailed:
+            # Remove fields with too many data points
+            for field in [
+                    "input_lens", "output_lens", "ttfts", "itls",
+                    "generated_texts", "errors"
+            ]:
+                if field in result_json:
+                    del result_json[field]
+
         # Traffic
         result_json["request_rate"] = (args.request_rate if args.request_rate
                                        < float("inf") else "inf")
@@ -1052,13 +785,6 @@ if __name__ == "__main__":
         default="/v1/completions",
         help="API endpoint.",
     )
-    parser.add_argument(
-        "--dataset",
-        type=str,
-        default=None,
-        help="Path to the ShareGPT dataset, will be deprecated in the "
-        "next release.",
-    )
     parser.add_argument(
         "--dataset-name",
         type=str,
@@ -1096,13 +822,6 @@ if __name__ == "__main__":
         help=
         "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
-    parser.add_argument(
-        "--best-of",
-        type=int,
-        default=1,
-        help="Generates `best_of` sequences per prompt and "
-        "returns the best one.",
-    )
     parser.add_argument("--use-beam-search", action="store_true")
     parser.add_argument(
         "--num-prompts",
@@ -1163,6 +882,12 @@ if __name__ == "__main__":
         action="store_true",
         help="Specify to save benchmark results to a json file",
     )
+    parser.add_argument(
+        "--save-detailed",
+        action="store_true",
+        help="When saving the results, whether to include per request "
+        "information such as response, error, ttfs, tpots, etc.",
+    )
     parser.add_argument(
         "--metadata",
         metavar="KEY=VALUE",
@@ -1196,7 +921,7 @@ if __name__ == "__main__":
         "--percentile-metrics",
         type=str,
         default="ttft,tpot,itl",
-        help="Comma-seperated list of selected metrics to report percentils. "
+        help="Comma-separated list of selected metrics to report percentils. "
         "This argument specifies the metrics to report percentiles. "
         "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
         "Default value is \"ttft,tpot,itl\".")
@@ -1204,7 +929,7 @@ if __name__ == "__main__":
         "--metric-percentiles",
         type=str,
         default="99",
-        help="Comma-seperated list of percentiles for selected metrics. "
+        help="Comma-separated list of percentiles for selected metrics. "
         "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
         "Default value is \"99\". "
         "Use \"--percentile-metrics\" to select metrics.",
@@ -1271,18 +996,23 @@ if __name__ == "__main__":
     random_group.add_argument(
         "--random-range-ratio",
         type=float,
-        default=1.0,
-        help="Range of sampled ratio of input/output length, "
-        "used only for random sampling.",
+        default=0.0,
+        help="Range ratio for sampling input/output length, "
+        "used only for random sampling. Must be in the range [0, 1) to define "
+        "a symmetric sampling range"
+        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
     )
     random_group.add_argument(
         "--random-prefix-len",
         type=int,
         default=0,
-        help="Number of fixed prefix tokens before random "
-        " context. The length range of context in a random "
-        " request is [random-prefix-len, "
-        " random-prefix-len + random-prefix-len * random-range-ratio).")
+        help=("Number of fixed prefix tokens before the random context "
+              "in a request. "
+              "The total input length is the sum of `random-prefix-len` and "
+              "a random "
+              "context length sampled from [input_len * (1 - range_ratio), "
+              "input_len * (1 + range_ratio)]."),
+    )
 
     hf_group = parser.add_argument_group("hf dataset options")
     hf_group.add_argument("--hf-subset",
@@ -1301,6 +1031,33 @@ if __name__ == "__main__":
         "from the sampled HF dataset.",
     )
 
+    sampling_group = parser.add_argument_group("sampling parameters")
+    sampling_group.add_argument(
+        "--top-p",
+        type=float,
+        default=None,
+        help="Top-p sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--top-k",
+        type=int,
+        default=None,
+        help="Top-k sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--min-p",
+        type=float,
+        default=None,
+        help="Min-p sampling parameter. Only has effect on openai-compatible "
+        "backends.")
+    sampling_group.add_argument(
+        "--temperature",
+        type=float,
+        default=None,
+        help="Temperature sampling parameter. Only has effect on "
+        "openai-compatible backends. If not specified, default to greedy "
+        "decoding (i.e. temperature==0.0).")
+
     parser.add_argument(
         '--tokenizer-mode',
         type=str,
@@ -1327,4 +1084,5 @@ if __name__ == "__main__":
                         "script chooses a LoRA module at random.")
 
     args = parser.parse_args()
+
     main(args)

benchmarks/benchmark_serving_structured_output.py

@@ -1,20 +1,17 @@
 # SPDX-License-Identifier: Apache-2.0
-r"""Benchmark online serving throughput with guided decoding.
+r"""Benchmark online serving throughput with structured outputs.
 
 On the server side, run one of the following commands:
     (vLLM OpenAI API server)
     vllm serve <your_model> --disable-log-requests
 
-    (TGI backend)
-    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
-
 On the client side, run:
-    python benchmarks/benchmark_serving.py \
+    python benchmarks/benchmark_serving_structured_output.py \
         --backend <backend> \
         --model <your_model> \
         --dataset json \
-        --guided-decoding-ratio 1.0 \
-        --guided-decoding-backend xgrammar \
+        --structured-output-ratio 1.0 \
+        --structured-output-backend auto \
         --request-rate 10 \
         --num-prompts 1000
 
@@ -24,14 +21,17 @@ On the client side, run:
 """
 import argparse
 import asyncio
+import copy
 import dataclasses
 import json
 import os
 import random
 import time
+import uuid
 import warnings
+from collections.abc import AsyncGenerator
 from dataclasses import dataclass
-from typing import AsyncGenerator, List, Optional, Tuple
+from typing import Optional
 
 import datasets
 import numpy as np
@@ -51,6 +51,9 @@ try:
 except ImportError:
     from argparse import ArgumentParser as FlexibleArgumentParser
 
+from vllm.v1.structured_output.utils import (
+    has_xgrammar_unsupported_json_features)
+
 MILLISECONDS_TO_SECONDS_CONVERSION = 1000
 
 
@@ -66,22 +69,22 @@ class BenchmarkMetrics:
     mean_ttft_ms: float
     median_ttft_ms: float
     std_ttft_ms: float
-    percentiles_ttft_ms: List[Tuple[float, float]]
+    percentiles_ttft_ms: list[tuple[float, float]]
     mean_tpot_ms: float
     median_tpot_ms: float
     std_tpot_ms: float
-    percentiles_tpot_ms: List[Tuple[float, float]]
+    percentiles_tpot_ms: list[tuple[float, float]]
     mean_itl_ms: float
     median_itl_ms: float
     std_itl_ms: float
-    percentiles_itl_ms: List[Tuple[float, float]]
+    percentiles_itl_ms: list[tuple[float, float]]
     # E2EL stands for end-to-end latency per request.
     # It is the time taken on the client side from sending
     # a request to receiving a complete response.
     mean_e2el_ms: float
     median_e2el_ms: float
     std_e2el_ms: float
-    percentiles_e2el_ms: List[Tuple[float, float]]
+    percentiles_e2el_ms: list[tuple[float, float]]
 
 
 @dataclasses.dataclass
@@ -104,25 +107,45 @@ class SampleRequest:
 
 
 def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    args: argparse.Namespace) -> List[SampleRequest]:
-    if args.dataset == 'json':
+                    args: argparse.Namespace) -> list[SampleRequest]:
+    if args.dataset == 'json' or args.dataset == 'json-unique':
         if args.json_schema_path is None:
             dir_path = os.path.dirname(os.path.realpath(__file__))
             args.json_schema_path = os.path.join(dir_path,
                                                  "structured_schemas",
                                                  "structured_schema_1.json")
+        json_schemas = []
         with open(args.json_schema_path) as f:
             schema = json.load(f)
-        prompt = f"Generate an example of a user profile given the following schema: {json.dumps(schema)}"  # noqa: E501
-        input_len = len(tokenizer(prompt).input_ids)
-        print(f"Input length of the prompt: {input_len} tokens")
+
+        if args.dataset == 'json-unique':
+            json_schemas = [
+                copy.deepcopy(schema) for _ in range(args.num_prompts)
+            ]
+            for i in range(len(json_schemas)):
+                json_schemas[i]["properties"][
+                    f"__optional_field_{uuid.uuid4()}"] = {
+                        "type":
+                        "string",
+                        "description":
+                        "An unique optional field to avoid cached schemas"
+                    }
+        else:
+            json_schemas = [schema] * args.num_prompts
+
+        def gen_prompt(index: int):
+            return f"Generate an example of a user profile given the following schema: {json.dumps(get_schema(index))}"  # noqa: E501
+
+        def get_schema(index: int):
+            return json_schemas[index % len(json_schemas)]
+
         requests = [
-            SampleRequest(prompt=prompt,
-                          prompt_len=input_len,
+            SampleRequest(prompt=gen_prompt(i),
+                          prompt_len=len(tokenizer(gen_prompt(i)).input_ids),
                           expected_output_len=args.output_len,
-                          schema=schema,
+                          schema=get_schema(i),
                           structure_type=args.structure_type)
-            for _ in range(args.num_prompts)
+            for i in range(args.num_prompts)
         ]
 
     elif args.dataset == "grammar":
@@ -187,10 +210,20 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
         ]
 
     elif args.dataset == "xgrammar_bench":
-        requests: List[SampleRequest] = []
+        requests: list[SampleRequest] = []
         dataset = datasets.load_dataset("NousResearch/json-mode-eval",
                                         split="train")
-        print(f"dataset has {len(dataset)} entries")
+        full_dataset_len = len(dataset)
+
+        def _filter_func(item):
+            import json
+            schema = json.loads(item["schema"])
+            return not has_xgrammar_unsupported_json_features(schema)
+
+        dataset = dataset.filter(_filter_func)
+        num_filtered_out = full_dataset_len - len(dataset)
+        print(f"dataset has {len(dataset)} entries after filtering "
+              f"out {num_filtered_out} entries with unsupported features")
         len_dataset = len(dataset)
         for data_point_idx in range(args.num_prompts):
             idx = data_point_idx
@@ -214,10 +247,10 @@ def sample_requests(tokenizer: PreTrainedTokenizerBase,
 
 
 async def get_request(
-    input_requests: List[SampleRequest],
+    input_requests: list[SampleRequest],
     request_rate: float,
     burstiness: float = 1.0,
-) -> AsyncGenerator[Tuple[int, SampleRequest], None]:
+) -> AsyncGenerator[tuple[int, SampleRequest], None]:
     """
     Asynchronously generates requests at a specified rate
     with OPTIONAL burstiness.
@@ -258,22 +291,23 @@ async def get_request(
 
 
 def calculate_metrics(
-    input_requests: List[Tuple[str, int, int]],
-    outputs: List[RequestFuncOutput],
+    input_requests: list[tuple[str, int, int]],
+    outputs: list[RequestFuncOutput],
     dur_s: float,
     tokenizer: PreTrainedTokenizerBase,
-    selected_percentile_metrics: List[str],
-    selected_percentiles: List[float],
-) -> Tuple[BenchmarkMetrics, List[int]]:
-    actual_output_lens: List[int] = []
+    selected_percentile_metrics: list[str],
+    selected_percentiles: list[float],
+    goodput_config_dict: Optional[dict[str, float]] = None,
+) -> tuple[BenchmarkMetrics, list[int]]:
+    actual_output_lens: list[int] = []
     total_input = 0
     completed = 0
     good_completed = 0
-    itls: List[float] = []
-    tpots: List[float] = []
-    all_tpots: List[float] = []
-    ttfts: List[float] = []
-    e2els: List[float] = []
+    itls: list[float] = []
+    tpots: list[float] = []
+    all_tpots: list[float] = []
+    ttfts: list[float] = []
+    e2els: list[float] = []
     for i in range(len(outputs)):
         if outputs[i].success:
             # We use the tokenizer to count the number of output tokens for all
@@ -287,10 +321,10 @@ def calculate_metrics(
             total_input += input_requests[i].prompt_len
             tpot = 0
             if output_len > 1:
-                tpot = (outputs[i].latency - outputs[i].ttft) / (output_len -
-                                                                 1)
+                latency_minus_ttft = outputs[i].latency - outputs[i].ttft
+                tpot = latency_minus_ttft / (output_len - 1)
                 tpots.append(tpot)
-            outputs[i].tpot = sum(tpots) / len(tpots) if len(tpots) else 0
+            outputs[i].tpot = tpot
             # Note: if output_len <= 1, we regard tpot as 0 for goodput
             all_tpots.append(tpot)
             itls += outputs[i].itl
@@ -300,6 +334,28 @@ def calculate_metrics(
         else:
             actual_output_lens.append(0)
 
+    if goodput_config_dict:
+        valid_metrics = []
+        slo_values = []
+
+        if "ttft" in goodput_config_dict:
+            valid_metrics.append(ttfts)
+            slo_values.append(goodput_config_dict["ttft"] /
+                              MILLISECONDS_TO_SECONDS_CONVERSION)
+        if "tpot" in goodput_config_dict:
+            valid_metrics.append(all_tpots)
+            slo_values.append(goodput_config_dict["tpot"] /
+                              MILLISECONDS_TO_SECONDS_CONVERSION)
+        if "e2el" in goodput_config_dict:
+            valid_metrics.append(e2els)
+            slo_values.append(goodput_config_dict["e2el"] /
+                              MILLISECONDS_TO_SECONDS_CONVERSION)
+
+        for req_metric in zip(*valid_metrics):
+            is_good_req = all([s >= r for s, r in zip(slo_values, req_metric)])
+            if is_good_req:
+                good_completed += 1
+
     if completed == 0:
         warnings.warn(
             "All requests failed. This is likely due to a misconfiguration "
@@ -345,17 +401,18 @@ async def benchmark(
     base_url: str,
     model_id: str,
     tokenizer: PreTrainedTokenizerBase,
-    input_requests: List[SampleRequest],
+    input_requests: list[SampleRequest],
     request_rate: float,
     burstiness: float,
     disable_tqdm: bool,
     profile: bool,
-    selected_percentile_metrics: List[str],
-    selected_percentiles: List[str],
+    selected_percentile_metrics: list[str],
+    selected_percentiles: list[str],
     ignore_eos: bool,
     max_concurrency: Optional[int],
-    guided_decoding_ratio: float,
-    guided_decoding_backend: str,
+    structured_output_ratio: float,
+    structured_output_backend: str,
+    goodput_config_dict: Optional[dict[str, float]] = None,
 ):
     if backend in ASYNC_REQUEST_FUNCS:
         request_func = ASYNC_REQUEST_FUNCS[backend]
@@ -366,16 +423,18 @@ async def benchmark(
         extra_body = {}
         # Add the schema to the extra_body
         extra_body[request.structure_type] = request.schema
-        # Add the specific guided_decoding_backend
-        extra_body["guided_decoding_backend"] = guided_decoding_backend
+        # Add the specific structured_output_backend
+        extra_body["guided_decoding_backend"] = structured_output_backend
         return extra_body
 
     print("Starting initial single prompt test run...")
-    guided_decoding_req_idx = random.sample(
+    structured_output_req_idx = random.sample(
         range(len(input_requests)),
-        int(len(input_requests) * guided_decoding_ratio))
+        int(len(input_requests) * structured_output_ratio))
 
     test_request = input_requests[0]
+    test_req_extra_body = (prepare_extra_body(test_request)
+                           if 0 in structured_output_req_idx else None)
     test_input = RequestFuncInput(
         model=model_id,
         prompt=test_request.prompt,
@@ -383,7 +442,7 @@ async def benchmark(
         prompt_len=test_request.prompt_len,
         output_len=test_request.expected_output_len,
         ignore_eos=ignore_eos,
-        extra_body=prepare_extra_body(test_request),
+        extra_body=test_req_extra_body,
     )
     test_output = await request_func(request_func_input=test_input)
     if not test_output.success:
@@ -402,7 +461,7 @@ async def benchmark(
             prompt_len=test_request.prompt_len,
             output_len=test_request.expected_output_len,
             ignore_eos=ignore_eos,
-            extra_body=prepare_extra_body(test_request),
+            extra_body=test_req_extra_body,
         )
         profile_output = await request_func(request_func_input=profile_input)
         if profile_output.success:
@@ -435,12 +494,12 @@ async def benchmark(
                                       pbar=pbar)
 
     benchmark_start_time = time.perf_counter()
-    tasks: List[asyncio.Task] = []
-    expected: List[str] = []
+    tasks: list[asyncio.Task] = []
+    expected: list[str] = []
     async for i, request in get_request(input_requests, request_rate,
                                         burstiness):
         extra_body = prepare_extra_body(
-            request) if i in guided_decoding_req_idx else None
+            request) if i in structured_output_req_idx else None
         request_func_input = RequestFuncInput(
             model=model_id,
             prompt=request.prompt,
@@ -455,7 +514,7 @@ async def benchmark(
             asyncio.create_task(
                 limited_request_func(request_func_input=request_func_input,
                                      pbar=pbar)))
-    outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
+    outputs: list[RequestFuncOutput] = await asyncio.gather(*tasks)
 
     if profile:
         print("Stopping profiler...")
@@ -483,6 +542,7 @@ async def benchmark(
         tokenizer=tokenizer,
         selected_percentile_metrics=selected_percentile_metrics,
         selected_percentiles=selected_percentiles,
+        goodput_config_dict=goodput_config_dict,
     )
 
     print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
@@ -494,6 +554,9 @@ async def benchmark(
                                  metrics.total_output))
     print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
                                     metrics.request_throughput))
+    if goodput_config_dict:
+        print("{:<40} {:<10.2f}".format("Request goodput (req/s):",
+                                        metrics.request_goodput))
     print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
                                     metrics.output_throughput))
     print("{:<40} {:<10.2f}".format("Total Token throughput (tok/s):",
@@ -617,6 +680,40 @@ def evaluate(ret, args):
             100) if len(not_none_scores) > 0 else None
 
 
+def parse_goodput(slo_pairs):
+    goodput_config_dict = {}
+    try:
+        for slo_pair in slo_pairs:
+            slo_name, slo_val = slo_pair.split(":")
+            goodput_config_dict[slo_name] = float(slo_val)
+    except ValueError as err:
+        raise argparse.ArgumentTypeError(
+            "Invalid format found for service level objectives. "
+            "Specify service level objectives for goodput as \"KEY:VALUE\" "
+            "pairs, where the key is a metric name, and the value is a "
+            "number in milliseconds.") from err
+    return goodput_config_dict
+
+
+def check_goodput_args(args):
+    goodput_config_dict = {}
+    VALID_NAMES = ["ttft", "tpot", "e2el"]
+    if args.goodput:
+        goodput_config_dict = parse_goodput(args.goodput)
+        for slo_name, slo_val in goodput_config_dict.items():
+            if slo_name not in VALID_NAMES:
+                raise ValueError(
+                    f"Invalid metric name found, {slo_name}: {slo_val}. "
+                    "The service level objective name should be one of "
+                    f"{str(VALID_NAMES)}. ")
+            if slo_val < 0:
+                raise ValueError(
+                    f"Invalid value found, {slo_name}: {slo_val}. "
+                    "The service level objective value should be "
+                    "non-negative.")
+    return goodput_config_dict
+
+
 def main(args: argparse.Namespace):
     print(args)
     random.seed(args.seed)
@@ -633,8 +730,11 @@ def main(args: argparse.Namespace):
         api_url = f"http://{args.host}:{args.port}{args.endpoint}"
         base_url = f"http://{args.host}:{args.port}"
 
-    tokenizer = get_tokenizer(tokenizer_id,
-                              trust_remote_code=args.trust_remote_code)
+    tokenizer = get_tokenizer(
+        tokenizer_id,
+        trust_remote_code=args.trust_remote_code,
+        tokenizer_mode=args.tokenizer_mode,
+    )
 
     if args.dataset == 'grammar':
         args.structure_type = 'guided_grammar'
@@ -645,10 +745,10 @@ def main(args: argparse.Namespace):
     else:
         args.structure_type = 'guided_json'
 
-    if args.no_guided_decoding:
-        args.guided_decoding_ratio = 0
+    if args.no_structured_output:
+        args.structured_output_ratio = 0
     if args.save_results:
-        result_file_name = f'{args.guided_decoding_ratio}guided'
+        result_file_name = f'{args.structured_output_ratio}guided'
         result_file_name += f"_{backend}"
         result_file_name += f"_{args.request_rate}qps"
         result_file_name += f"_{args.model.split('/')[-1]}"
@@ -661,6 +761,8 @@ def main(args: argparse.Namespace):
 
     input_requests = sample_requests(tokenizer, args)
 
+    goodput_config_dict = check_goodput_args(args)
+
     benchmark_result, ret = asyncio.run(
         benchmark(
             backend=backend,
@@ -679,8 +781,9 @@ def main(args: argparse.Namespace):
             ],
             ignore_eos=args.ignore_eos,
             max_concurrency=args.max_concurrency,
-            guided_decoding_ratio=args.guided_decoding_ratio,
-            guided_decoding_backend=args.guided_decoding_backend,
+            structured_output_ratio=args.structured_output_ratio,
+            structured_output_backend=args.structured_output_backend,
+            goodput_config_dict=goodput_config_dict,
         ))
 
     # Save config and results to json
@@ -740,10 +843,12 @@ if __name__ == "__main__":
         default="/v1/completions",
         help="API endpoint.",
     )
-    parser.add_argument(
-        "--dataset",
-        default='json',
-        choices=['json', 'grammar', 'regex', 'choice', 'xgrammar_bench'])
+    parser.add_argument("--dataset",
+                        default='json',
+                        choices=[
+                            'json', 'json-unique', 'grammar', 'regex',
+                            'choice', 'xgrammar_bench'
+                        ])
     parser.add_argument("--json_schema_path",
                         type=str,
                         default=None,
@@ -772,6 +877,13 @@ if __name__ == "__main__":
         help=
         "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
     )
+    parser.add_argument(
+        "--tokenizer-mode",
+        type=str,
+        default="auto",
+        help=
+        "Name or path of the tokenizer, if not using the default tokenizer.",  # noqa: E501
+    )
     parser.add_argument(
         "--num-prompts",
         type=int,
@@ -852,7 +964,7 @@ if __name__ == "__main__":
         "--percentile-metrics",
         type=str,
         default="ttft,tpot,itl",
-        help="Comma-seperated list of selected metrics to report percentils. "
+        help="Comma-separated list of selected metrics to report percentils. "
         "This argument specifies the metrics to report percentiles. "
         "Allowed metric names are \"ttft\", \"tpot\", \"itl\", \"e2el\". "
         "Default value is \"ttft,tpot,itl\".")
@@ -860,24 +972,39 @@ if __name__ == "__main__":
         "--metric-percentiles",
         type=str,
         default="99",
-        help="Comma-seperated list of percentiles for selected metrics. "
+        help="Comma-separated list of percentiles for selected metrics. "
         "To report 25-th, 50-th, and 75-th percentiles, use \"25,50,75\". "
         "Default value is \"99\". "
         "Use \"--percentile-metrics\" to select metrics.",
     )
-    parser.add_argument("--no-guided-decoding",
+    parser.add_argument(
+        "--goodput",
+        nargs="+",
+        required=False,
+        help="Specify service level objectives for goodput as \"KEY:VALUE\" "
+        "pairs, where the key is a metric name, and the value is in "
+        "milliseconds. Multiple \"KEY:VALUE\" pairs can be provided, "
+        "separated by spaces. Allowed request level metric names are "
+        "\"ttft\", \"tpot\", \"e2el\". For more context on the definition of "
+        "goodput, refer to DistServe paper: https://arxiv.org/pdf/2401.09670 "
+        "and the blog: https://hao-ai-lab.github.io/blogs/distserve")
+
+    parser.add_argument("--no-structured-output",
                         action='store_true',
                         default=False,
                         help="Whether to disable JSON decoding or not.")
-    parser.add_argument("--guided-decoding-ratio",
+    parser.add_argument("--structured-output-ratio",
                         type=float,
                         default=1.0,
-                        help="Ratio of Guided Decoding requests")
-    parser.add_argument("--guided-decoding-backend",
+                        help="Ratio of Structured Outputs requests")
+    parser.add_argument("--structured-output-backend",
                         type=str,
-                        choices=["outlines", "lm-format-enforcer", "xgrammar"],
-                        default="xgrammar",
-                        help="Backend to use for guided decoding")
+                        choices=[
+                            "outlines", "lm-format-enforcer", "xgrammar",
+                            "guidance", "auto"
+                        ],
+                        default="auto",
+                        help="Backend to use for structured outputs")
 
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -6,13 +6,16 @@ import json
 import os
 import random
 import time
-from functools import cache
-from typing import Any, Dict, List, Optional, Tuple
+import warnings
+from typing import Any, Optional, Union
 
 import torch
 import uvloop
-from benchmark_utils import convert_to_pytorch_benchmark_format
-from PIL import Image
+from benchmark_dataset import (AIMODataset, BurstGPTDataset,
+                               ConversationDataset, InstructCoderDataset,
+                               RandomDataset, SampleRequest, ShareGPTDataset,
+                               SonnetDataset, VisionArenaDataset)
+from benchmark_utils import convert_to_pytorch_benchmark_format, write_to_json
 from tqdm import tqdm
 from transformers import (AutoModelForCausalLM, AutoTokenizer,
                           PreTrainedTokenizerBase)
@@ -20,163 +23,35 @@ from transformers import (AutoModelForCausalLM, AutoTokenizer,
 from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.entrypoints.openai.api_server import (
     build_async_engine_client_from_engine_args)
-from vllm.inputs import TextPrompt
+from vllm.inputs import TextPrompt, TokensPrompt
 from vllm.lora.request import LoRARequest
-from vllm.lora.utils import get_adapter_absolute_path
-from vllm.multimodal import MultiModalDataDict
+from vllm.outputs import RequestOutput
 from vllm.sampling_params import BeamSearchParams
-from vllm.transformers_utils.tokenizer import AnyTokenizer, get_lora_tokenizer
 from vllm.utils import FlexibleArgumentParser, merge_async_iterators
 
 
-@dataclasses.dataclass
-class SampleRequest:
-    """A class representing a single inference request for benchmarking.
-
-    Attributes:
-        prompt: The input text prompt for the model.
-        prompt_len: The length of the prompt in tokens.
-        expected_output_len: The expected length of the output in tokens.
-        multi_modal_data: Optional dictionary containing multi-modal data (e.g.
-            images).
-        lora_request: Optional LoRARequest specifying the LoRA to use. 
-    """
-    prompt: str
-    prompt_len: int
-    expected_output_len: int
-    multi_modal_data: Optional[MultiModalDataDict] = None
-    lora_request: Optional[LoRARequest] = None
-
-
-def _get_prompt_for_image_model(question: str, *, model: str) -> str:
-    """Prepend and append special tokens around the question to form a prompt.
-
-    Args:
-        question: The input question text to wrap with special tokens
-        model: The name of the model being used, to determine which special
-            tokens to add
-
-    Returns:
-        The formatted prompt string with appropriate special tokens for the
-            model
-
-    Raises:
-        ValueError: If an unsupported model name is provided
-    """
-    model = model.lower()
-    if "pixtral" in model:
-        return f"<s>[INST]{question}\n[IMG][/INST]"
-    raise ValueError(f"Unsupported model {model}")
-
-
-@cache
-def lora_path_on_disk(lora_path: str) -> str:
-    return get_adapter_absolute_path(lora_path)
-
-
-lora_tokenizer_cache: Dict[int, AnyTokenizer] = {}
-
-
-def get_random_lora_request(
-        args: argparse.Namespace
-) -> Tuple[LoRARequest, Optional[AnyTokenizer]]:
-    global lora_tokenizer_cache
-    lora_id = random.randint(1, args.max_loras)
-    lora_request = LoRARequest(lora_name=str(lora_id),
-                               lora_int_id=lora_id,
-                               lora_path=lora_path_on_disk(args.lora_path))
-    if lora_id not in lora_tokenizer_cache:
-        lora_tokenizer_cache[lora_id] = get_lora_tokenizer(lora_request)
-    return lora_request, lora_tokenizer_cache[lora_id]
-
-
-def sample_requests(tokenizer: PreTrainedTokenizerBase,
-                    args: argparse.Namespace) -> List[SampleRequest]:
-
-    dataset_path: str = args.dataset
-    num_requests: int = args.num_prompts
-    fixed_output_len: Optional[int] = args.output_len
-    model: str = args.model
-    if fixed_output_len is not None and fixed_output_len < 4:
-        raise ValueError("output_len too small")
-
-    # Load the dataset.
-    with open(dataset_path) as f:
-        dataset = json.load(f)
-    # Filter out the conversations with less than 2 turns.
-    dataset = [data for data in dataset if len(data["conversations"]) >= 2]
-    # Shuffle the dataset.
-    random.shuffle(dataset)
-
-    # Filter out sequences that are too long or too short
-    filtered_dataset: List[SampleRequest] = []
-    for data in tqdm(dataset,
-                     total=len(filtered_dataset),
-                     desc="sampling requests"):
-        if len(filtered_dataset) == num_requests:
-            break
-
-        # Only keep the first two turns of each conversation.
-        prompt = data["conversations"][0]["value"]
-        completion = data["conversations"][1]["value"]
-
-        multi_modal_data: Optional[MultiModalDataDict] = None
-        if "image" in data:
-            multi_modal_data = multi_modal_data or {}
-            image_path = data["image"]
-            # TODO(vllm-project/vllm/issues/9778): Support multiple images.
-            assert isinstance(image_path,
-                              str), "Only support single image input"
-            try:
-                multi_modal_data["image"] = Image.open(image_path).convert(
-                    "RGB")
-            except FileNotFoundError:
-                # Ignore datapoint where asset is missing
-                continue
-            prompt = _get_prompt_for_image_model(question=prompt, model=model)
-
-        request_tokenizer = tokenizer
-        lora_request: Optional[LoRARequest] = None
-        if args.enable_lora:
-            lora_request, lora_tokenizer = get_random_lora_request(args)
-            if lora_tokenizer:
-                request_tokenizer = lora_tokenizer
-
-        # Tokenize the prompts and completions.
-        prompt_token_ids = request_tokenizer(prompt).input_ids
-        completion_token_ids = request_tokenizer(completion).input_ids
-        prompt_len = len(prompt_token_ids)
-        output_len = len(completion_token_ids
-                         ) if fixed_output_len is None else fixed_output_len
-        if prompt_len < 4 or output_len < 4:
-            # Prune too short sequences.
-            continue
-        if prompt_len > 1024 or prompt_len + output_len > 2048:
-            # Prune too long sequences.
-            continue
-        filtered_dataset.append(
-            SampleRequest(prompt=prompt,
-                          prompt_len=prompt_len,
-                          expected_output_len=output_len,
-                          multi_modal_data=multi_modal_data,
-                          lora_request=lora_request))
-
-    return filtered_dataset
-
-
 def run_vllm(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
     n: int,
     engine_args: EngineArgs,
-) -> float:
+    disable_detokenize: bool = False,
+) -> tuple[float, Optional[list[RequestOutput]]]:
     from vllm import LLM, SamplingParams
     llm = LLM(**dataclasses.asdict(engine_args))
-
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (
+            request.prompt_len + request.expected_output_len)
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of"
+            " prompt_len and expected_output_len for all requests.")
     # Add the requests to the engine.
-    prompts: List[TextPrompt] = []
-    sampling_params: List[SamplingParams] = []
+    prompts: list[Union[TextPrompt, TokensPrompt]] = []
+    sampling_params: list[SamplingParams] = []
     for request in requests:
         prompts.append(
+            TokensPrompt(prompt_token_ids=request.prompt["prompt_token_ids"],
+                       multi_modal_data=request.multi_modal_data)
+            if "prompt_token_ids" in request.prompt else \
             TextPrompt(prompt=request.prompt,
                        multi_modal_data=request.multi_modal_data))
         sampling_params.append(
@@ -186,19 +61,21 @@ def run_vllm(
                 top_p=1.0,
                 ignore_eos=True,
                 max_tokens=request.expected_output_len,
+                detokenize=not disable_detokenize,
             ))
-    lora_requests: Optional[List[LoRARequest]] = None
+    lora_requests: Optional[list[LoRARequest]] = None
     if engine_args.enable_lora:
         lora_requests = [request.lora_request for request in requests]
 
     use_beam_search = False
 
+    outputs = None
     if not use_beam_search:
         start = time.perf_counter()
-        llm.generate(prompts,
-                     sampling_params,
-                     lora_request=lora_requests,
-                     use_tqdm=True)
+        outputs = llm.generate(prompts,
+                               sampling_params,
+                               lora_request=lora_requests,
+                               use_tqdm=True)
         end = time.perf_counter()
     else:
         assert lora_requests is None, "BeamSearch API does not support LoRA"
@@ -216,26 +93,75 @@ def run_vllm(
                 ignore_eos=True,
             ))
         end = time.perf_counter()
-    return end - start
+    return end - start, outputs
+
+
+def run_vllm_chat(
+        requests: list[SampleRequest],
+        n: int,
+        engine_args: EngineArgs,
+        disable_detokenize: bool = False) -> tuple[float, list[RequestOutput]]:
+    """
+    Run vLLM chat benchmark. This function is recommended ONLY for benchmarking
+    multimodal models as it properly handles multimodal inputs and chat
+    formatting. For non-multimodal models, use run_vllm() instead.
+    """
+    from vllm import LLM, SamplingParams
+    llm = LLM(**dataclasses.asdict(engine_args))
+
+    assert all(
+        llm.llm_engine.model_config.max_model_len >= (
+            request.prompt_len + request.expected_output_len)
+        for request in requests), (
+            "Please ensure that max_model_len is greater than the sum of "
+            "prompt_len and expected_output_len for all requests.")
+
+    prompts = []
+    sampling_params: list[SamplingParams] = []
+    for request in requests:
+        prompts.append(request.prompt)
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=1.0,
+                top_p=1.0,
+                ignore_eos=True,
+                max_tokens=request.expected_output_len,
+                detokenize=not disable_detokenize,
+            ))
+    start = time.perf_counter()
+    outputs = llm.chat(prompts, sampling_params, use_tqdm=True)
+    end = time.perf_counter()
+    return end - start, outputs
 
 
 async def run_vllm_async(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
     n: int,
     engine_args: AsyncEngineArgs,
     disable_frontend_multiprocessing: bool = False,
+    disable_detokenize: bool = False,
 ) -> float:
     from vllm import SamplingParams
 
     async with build_async_engine_client_from_engine_args(
             engine_args, disable_frontend_multiprocessing) as llm:
+        assert all(
+            llm.model_config.max_model_len >= (request.prompt_len +
+                                               request.expected_output_len)
+            for request in requests), (
+                "Please ensure that max_model_len is greater than the sum of"
+                " prompt_len and expected_output_len for all requests.")
 
         # Add the requests to the engine.
-        prompts: List[TextPrompt] = []
-        sampling_params: List[SamplingParams] = []
-        lora_requests: List[Optional[LoRARequest]] = []
+        prompts: list[Union[TextPrompt, TokensPrompt]] = []
+        sampling_params: list[SamplingParams] = []
+        lora_requests: list[Optional[LoRARequest]] = []
         for request in requests:
             prompts.append(
+                TokensPrompt(prompt_token_ids=request.prompt["prompt_token_ids"],
+                        multi_modal_data=request.multi_modal_data)
+                if "prompt_token_ids" in request.prompt else \
                 TextPrompt(prompt=request.prompt,
                            multi_modal_data=request.multi_modal_data))
             sampling_params.append(
@@ -245,6 +171,7 @@ async def run_vllm_async(
                     top_p=1.0,
                     ignore_eos=True,
                     max_tokens=request.expected_output_len,
+                    detokenize=not disable_detokenize,
                 ))
             lora_requests.append(request.lora_request)
 
@@ -265,12 +192,13 @@ async def run_vllm_async(
 
 
 def run_hf(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
     model: str,
     tokenizer: PreTrainedTokenizerBase,
     n: int,
     max_batch_size: int,
     trust_remote_code: bool,
+    disable_detokenize: bool = False,
 ) -> float:
     llm = AutoModelForCausalLM.from_pretrained(
         model, torch_dtype=torch.float16, trust_remote_code=trust_remote_code)
@@ -281,18 +209,21 @@ def run_hf(
 
     pbar = tqdm(total=len(requests))
     start = time.perf_counter()
-    batch: List[str] = []
+    batch: list[str] = []
     max_prompt_len = 0
     max_output_len = 0
     for i in range(len(requests)):
-        prompt, prompt_len, output_len = requests[i]
+        prompt = requests[i].prompt
+        prompt_len = requests[i].prompt_len
+        output_len = requests[i].expected_output_len
         # Add the prompt to the batch.
         batch.append(prompt)
         max_prompt_len = max(max_prompt_len, prompt_len)
         max_output_len = max(max_output_len, output_len)
         if len(batch) < max_batch_size and i != len(requests) - 1:
             # Check if we can add more requests to the batch.
-            _, next_prompt_len, next_output_len = requests[i + 1]
+            next_prompt_len = requests[i + 1].prompt_len
+            next_output_len = requests[i + 1].expected_output_len
             if (max(max_prompt_len, next_prompt_len) +
                     max(max_output_len, next_output_len)) <= 2048:
                 # We can add more requests to the batch.
@@ -310,8 +241,9 @@ def run_hf(
             use_cache=True,
             max_new_tokens=max_output_len,
         )
-        # Include the decoding time.
-        tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
+        if not disable_detokenize:
+            # Include the decoding time.
+            tokenizer.batch_decode(llm_outputs, skip_special_tokens=True)
         pbar.update(len(batch))
 
         # Clear the batch.
@@ -323,7 +255,7 @@ def run_hf(
 
 
 def run_mii(
-    requests: List[SampleRequest],
+    requests: list[SampleRequest],
     model: str,
     tensor_parallel_size: int,
     output_len: int,
@@ -341,7 +273,7 @@ def run_mii(
 
 
 def save_to_pytorch_benchmark_format(args: argparse.Namespace,
-                                     results: Dict[str, Any]) -> None:
+                                     results: dict[str, Any]) -> None:
     pt_records = convert_to_pytorch_benchmark_format(
         args=args,
         metrics={
@@ -355,62 +287,77 @@ def save_to_pytorch_benchmark_format(args: argparse.Namespace,
     if pt_records:
         # Don't use json suffix here as we don't want CI to pick it up
         pt_file = f"{os.path.splitext(args.output_json)[0]}.pytorch.json"
-        with open(pt_file, "w") as f:
-            json.dump(pt_records, f)
+        write_to_json(pt_file, pt_records)
+
+
+def get_requests(args, tokenizer):
+    # Common parameters for all dataset types.
+    common_kwargs = {
+        "dataset_path": args.dataset_path,
+        "random_seed": args.seed,
+    }
+    sample_kwargs = {
+        "tokenizer": tokenizer,
+        "lora_path": args.lora_path,
+        "max_loras": args.max_loras,
+        "num_requests": args.num_prompts,
+        "input_len": args.input_len,
+        "output_len": args.output_len,
+    }
+
+    if args.dataset_path is None or args.dataset_name == "random":
+        sample_kwargs["range_ratio"] = args.random_range_ratio
+        sample_kwargs["prefix_len"] = args.prefix_len
+        dataset_cls = RandomDataset
+    elif args.dataset_name == "sharegpt":
+        dataset_cls = ShareGPTDataset
+        if args.backend == "vllm-chat":
+            sample_kwargs["enable_multimodal_chat"] = True
+    elif args.dataset_name == "sonnet":
+        assert tokenizer.chat_template or tokenizer.default_chat_template, (
+            "Tokenizer/model must have chat template for sonnet dataset.")
+        dataset_cls = SonnetDataset
+        sample_kwargs["prefix_len"] = args.prefix_len
+        sample_kwargs["return_prompt_formatted"] = True
+    elif args.dataset_name == "burstgpt":
+        dataset_cls = BurstGPTDataset
+    elif args.dataset_name == "hf":
+        if args.dataset_path in VisionArenaDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = VisionArenaDataset
+            common_kwargs['dataset_subset'] = None
+            common_kwargs['dataset_split'] = "train"
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in InstructCoderDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = InstructCoderDataset
+            common_kwargs['dataset_split'] = "train"
+        elif args.dataset_path in ConversationDataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = ConversationDataset
+            common_kwargs['dataset_subset'] = args.hf_subset
+            common_kwargs['dataset_split'] = args.hf_split
+            sample_kwargs["enable_multimodal_chat"] = True
+        elif args.dataset_path in AIMODataset.SUPPORTED_DATASET_PATHS:
+            dataset_cls = AIMODataset
+            common_kwargs['dataset_subset'] = None
+            common_kwargs['dataset_split'] = "train"
+    else:
+        raise ValueError(f"Unknown dataset name: {args.dataset_name}")
+    # Remove None values
+    sample_kwargs = {k: v for k, v in sample_kwargs.items() if v is not None}
+    return dataset_cls(**common_kwargs).sample(**sample_kwargs)
 
 
 def main(args: argparse.Namespace):
+    if args.seed is None:
+        args.seed = 0
     print(args)
     random.seed(args.seed)
-
     # Sample the requests.
     tokenizer = AutoTokenizer.from_pretrained(
         args.tokenizer, trust_remote_code=args.trust_remote_code)
-    if args.dataset is None:
-        vocab_size = tokenizer.vocab_size
-        requests = []
-        for _ in range(args.num_prompts):
-
-            request_tokenizer = tokenizer
-            lora_request: Optional[LoRARequest] = None
-            if args.enable_lora:
-                lora_request, lora_tokenizer = get_random_lora_request(args)
-                if lora_tokenizer:
-                    request_tokenizer = lora_tokenizer
-
-            # Synthesize a prompt with the given input length.
-            candidate_ids = [
-                random.randint(0, vocab_size - 1)
-                for _ in range(args.input_len)
-            ]
-            # As tokenizer may add additional tokens like BOS, we need to try
-            # different lengths to get the desired input length.
-            for _ in range(5):  # Max attempts to correct
-                candidate_prompt = request_tokenizer.decode(candidate_ids)
-                tokenized_len = len(request_tokenizer.encode(candidate_prompt))
-
-                if tokenized_len == args.input_len:
-                    break
-
-                # Adjust length based on difference
-                diff = args.input_len - tokenized_len
-                if diff > 0:
-                    candidate_ids.extend([
-                        random.randint(100, vocab_size - 100)
-                        for _ in range(diff)
-                    ])
-                else:
-                    candidate_ids = candidate_ids[:diff]
-            requests.append(
-                SampleRequest(prompt=candidate_prompt,
-                              prompt_len=args.input_len,
-                              expected_output_len=args.output_len,
-                              lora_request=lora_request))
-    else:
-        requests = sample_requests(tokenizer, args)
-
+    requests = get_requests(args, tokenizer)
     is_multi_modal = any(request.multi_modal_data is not None
                          for request in requests)
+    request_outputs: Optional[list[RequestOutput]] = None
     if args.backend == "vllm":
         if args.async_engine:
             elapsed_time = uvloop.run(
@@ -419,31 +366,59 @@ def main(args: argparse.Namespace):
                     args.n,
                     AsyncEngineArgs.from_cli_args(args),
                     args.disable_frontend_multiprocessing,
+                    args.disable_detokenize,
                 ))
         else:
-            elapsed_time = run_vllm(requests, args.n,
-                                    EngineArgs.from_cli_args(args))
+            elapsed_time, request_outputs = run_vllm(
+                requests, args.n, EngineArgs.from_cli_args(args),
+                args.disable_detokenize)
     elif args.backend == "hf":
         assert args.tensor_parallel_size == 1
         elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
-                              args.hf_max_batch_size, args.trust_remote_code)
+                              args.hf_max_batch_size, args.trust_remote_code,
+                              args.disable_detokenize)
     elif args.backend == "mii":
         elapsed_time = run_mii(requests, args.model, args.tensor_parallel_size,
                                args.output_len)
+    elif args.backend == "vllm-chat":
+        elapsed_time, request_outputs = run_vllm_chat(
+            requests, args.n, EngineArgs.from_cli_args(args),
+            args.disable_detokenize)
     else:
         raise ValueError(f"Unknown backend: {args.backend}")
-    total_num_tokens = sum(request.prompt_len + request.expected_output_len
-                           for request in requests)
-    total_output_tokens = sum(request.expected_output_len
-                              for request in requests)
-    if is_multi_modal:
-        print("\033[91mWARNING\033[0m: Multi-modal request detected. The "
+
+    if request_outputs:
+        # Note: with the vllm and vllm-chat backends,
+        # we have request_outputs, which we use to count tokens.
+        total_prompt_tokens = 0
+        total_output_tokens = 0
+        for ro in request_outputs:
+            if not isinstance(ro, RequestOutput):
+                continue
+            total_prompt_tokens += len(
+                ro.prompt_token_ids) if ro.prompt_token_ids else 0
+            total_output_tokens += sum(
+                len(o.token_ids) for o in ro.outputs if o)
+        total_num_tokens = total_prompt_tokens + total_output_tokens
+    else:
+        total_num_tokens = sum(r.prompt_len + r.expected_output_len
+                               for r in requests)
+        total_output_tokens = sum(r.expected_output_len for r in requests)
+        total_prompt_tokens = total_num_tokens - total_output_tokens
+
+    if is_multi_modal and args.backend != "vllm-chat":
+        print("\033[91mWARNING\033[0m: Multi-modal request with "
+              f"{args.backend} backend detected. The "
               "following metrics are not accurate because image tokens are not"
               " counted. See vllm-project/vllm/issues/9778 for details.")
-        # TODO(vllm-project/vllm/issues/9778): Count molti-modal token length.
+        # TODO(vllm-project/vllm/issues/9778): Count multi-modal token length.
+        # vllm-chat backend counts the image tokens now
+
     print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
           f"{total_num_tokens / elapsed_time:.2f} total tokens/s, "
           f"{total_output_tokens / elapsed_time:.2f} output tokens/s")
+    print(f"Total num prompt tokens:  {total_prompt_tokens}")
+    print(f"Total num output tokens:  {total_output_tokens}")
 
     # Output JSON results if specified
     if args.output_json:
@@ -459,18 +434,120 @@ def main(args: argparse.Namespace):
         save_to_pytorch_benchmark_format(args, results)
 
 
+def validate_args(args):
+    """
+    Validate command-line arguments.
+    """
+
+    # === Deprecation and Defaulting ===
+    if args.dataset is not None:
+        warnings.warn(
+            "The '--dataset' argument will be deprecated in the next release. "
+            "Please use '--dataset-name' and '--dataset-path' instead.",
+            stacklevel=2)
+        args.dataset_path = args.dataset
+
+    if not getattr(args, "tokenizer", None):
+        args.tokenizer = args.model
+
+    # === Backend Validation ===
+    valid_backends = {"vllm", "hf", "mii", "vllm-chat"}
+    if args.backend not in valid_backends:
+        raise ValueError(f"Unsupported backend: {args.backend}")
+
+    # === Dataset Configuration ===
+    if not args.dataset and not args.dataset_path:
+        print(
+            "When dataset path is not set, it will default to random dataset")
+        args.dataset_name = 'random'
+        if args.input_len is None:
+            raise ValueError("input_len must be provided for a random dataset")
+
+    # === Dataset Name Specific Checks ===
+    # --hf-subset and --hf-split: only used
+    # when dataset_name is 'hf'
+    if args.dataset_name != "hf" and (
+            getattr(args, "hf_subset", None) is not None
+            or getattr(args, "hf_split", None) is not None):
+        warnings.warn("--hf-subset and --hf-split will be ignored \
+                since --dataset-name is not 'hf'.",
+                      stacklevel=2)
+    elif args.dataset_name == "hf":
+        if args.dataset_path in (
+                VisionArenaDataset.SUPPORTED_DATASET_PATHS.keys()
+                | ConversationDataset.SUPPORTED_DATASET_PATHS):
+            assert args.backend == "vllm-chat", f"{args.dataset_path} needs to use vllm-chat as the backend."  #noqa: E501
+        elif args.dataset_path in (InstructCoderDataset.SUPPORTED_DATASET_PATHS
+                                   | AIMODataset.SUPPORTED_DATASET_PATHS):
+            assert args.backend == "vllm", f"{args.dataset_path} needs to use vllm as the backend."  #noqa: E501
+        else:
+            raise ValueError(
+                f"{args.dataset_path} is not supported by hf dataset.")
+
+    # --random-range-ratio: only used when dataset_name is 'random'
+    if args.dataset_name != 'random' and args.random_range_ratio is not None:
+        warnings.warn("--random-range-ratio will be ignored since \
+                --dataset-name is not 'random'.",
+                      stacklevel=2)
+
+    # --prefix-len: only used when dataset_name is 'random', 'sonnet', or not
+    # set.
+    if args.dataset_name not in {"random", "sonnet", None
+                                 } and args.prefix_len is not None:
+        warnings.warn("--prefix-len will be ignored since --dataset-name\
+                 is not 'random', 'sonnet', or not set.",
+                      stacklevel=2)
+
+    # === LoRA Settings ===
+    if getattr(args, "enable_lora", False) and args.backend != "vllm":
+        raise ValueError(
+            "LoRA benchmarking is only supported for vLLM backend")
+    if getattr(args, "enable_lora", False) and args.lora_path is None:
+        raise ValueError("LoRA path must be provided when enable_lora is True")
+
+    # === Backend-specific Validations ===
+    if args.backend == "hf" and args.hf_max_batch_size is None:
+        raise ValueError("HF max batch size is required for HF backend")
+    if args.backend != "hf" and args.hf_max_batch_size is not None:
+        raise ValueError("HF max batch size is only for HF backend.")
+
+    if args.backend in {"hf", "mii"} and getattr(args, "quantization",
+                                                 None) is not None:
+        raise ValueError("Quantization is only for vLLM backend.")
+
+    if args.backend == "mii" and args.dtype != "auto":
+        raise ValueError("dtype must be auto for MII backend.")
+    if args.backend == "mii" and args.n != 1:
+        raise ValueError("n must be 1 for MII backend.")
+    if args.backend == "mii" and args.tokenizer != args.model:
+        raise ValueError(
+            "Tokenizer must be the same as the model for MII backend.")
+
+
 if __name__ == "__main__":
     parser = FlexibleArgumentParser(description="Benchmark the throughput.")
     parser.add_argument("--backend",
                         type=str,
-                        choices=["vllm", "hf", "mii"],
+                        choices=["vllm", "hf", "mii", "vllm-chat"],
                         default="vllm")
-    parser.add_argument("--dataset",
+    parser.add_argument(
+        "--dataset-name",
+        type=str,
+        choices=["sharegpt", "random", "sonnet", "burstgpt", "hf"],
+        help="Name of the dataset to benchmark on.",
+        default="sharegpt")
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default=None,
+        help="Path to the ShareGPT dataset, will be deprecated in\
+            the next release. The dataset is expected to "
+        "be a json in form of list[dict[..., conversations: "
+        "list[dict[..., value: <prompt_or_response>]]]]")
+    parser.add_argument("--dataset-path",
                         type=str,
                         default=None,
-                        help="Path to the dataset. The dataset is expected to "
-                        "be a json in form of List[Dict[..., conversations: "
-                        "List[Dict[..., value: <prompt_or_response>]]]]")
+                        help="Path to the dataset")
     parser.add_argument("--input-len",
                         type=int,
                         default=None,
@@ -505,6 +582,11 @@ if __name__ == "__main__":
                         action='store_true',
                         default=False,
                         help="Disable decoupled async engine frontend.")
+    parser.add_argument(
+        "--disable-detokenize",
+        action="store_true",
+        help=("Do not detokenize the response (i.e. do not include "
+              "detokenization time in the measurement)"))
     # LoRA
     parser.add_argument(
         "--lora-path",
@@ -512,43 +594,45 @@ if __name__ == "__main__":
         default=None,
         help="Path to the lora adapters to use. This can be an absolute path, "
         "a relative path, or a Hugging Face model identifier.")
+    parser.add_argument(
+        "--prefix-len",
+        type=int,
+        default=None,
+        help=f"Number of prefix tokens to be used in RandomDataset "
+        "and SonnetDataset. For RandomDataset, the total input "
+        "length is the sum of prefix-len (default: "
+        f"{RandomDataset.DEFAULT_PREFIX_LEN}) and a random context length "
+        "sampled from [input_len * (1 - range_ratio), "
+        "input_len * (1 + range_ratio)]. For SonnetDataset, "
+        f"prefix_len (default: {SonnetDataset.DEFAULT_PREFIX_LEN}) "
+        "controls how much of the input is fixed lines versus "
+        "random lines, but the total input length remains approximately "
+        "input_len tokens.")
+    # random dataset
+    parser.add_argument(
+        "--random-range-ratio",
+        type=float,
+        default=None,
+        help=f"Range ratio (default : {RandomDataset.DEFAULT_RANGE_RATIO}) "
+        "for sampling input/output length, "
+        "used only for RandomDataset. Must be in the range [0, 1) to "
+        "define a symmetric sampling range "
+        "[length * (1 - range_ratio), length * (1 + range_ratio)].",
+    )
+
+    # hf dtaset
+    parser.add_argument("--hf-subset",
+                        type=str,
+                        default=None,
+                        help="Subset of the HF dataset.")
+    parser.add_argument("--hf-split",
+                        type=str,
+                        default=None,
+                        help="Split of the HF dataset.")
 
     parser = AsyncEngineArgs.add_cli_args(parser)
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model
-    if args.dataset is None:
-        assert args.input_len is not None
-        assert args.output_len is not None
-    else:
-        assert args.input_len is None
-    if args.enable_lora:
-        assert args.lora_path is not None
-
-    if args.backend == "vllm":
-        if args.hf_max_batch_size is not None:
-            raise ValueError("HF max batch size is only for HF backend.")
-    elif args.backend == "hf":
-        if args.hf_max_batch_size is None:
-            raise ValueError("HF max batch size is required for HF backend.")
-        if args.quantization is not None:
-            raise ValueError("Quantization is only for vLLM backend.")
-        if args.enable_lora is not None:
-            raise ValueError("LoRA benchmarking is only supported for vLLM"
-                             " backend")
-    elif args.backend == "mii":
-        if args.dtype != "auto":
-            raise ValueError("dtype must be auto for MII backend.")
-        if args.n != 1:
-            raise ValueError("n must be 1 for MII backend.")
-        if args.quantization is not None:
-            raise ValueError("Quantization is only for vLLM backend.")
-        if args.hf_max_batch_size is not None:
-            raise ValueError("HF max batch size is only for HF backend.")
-        if args.tokenizer != args.model:
-            raise ValueError("Tokenizer must be the same as the model for MII "
-                             "backend.")
-        if args.enable_lora is not None:
-            raise ValueError("LoRA benchmarking is only supported for vLLM"
-                             " backend")
+    validate_args(args)
     main(args)

benchmarks/benchmark_utils.py

@@ -1,13 +1,15 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import argparse
+import json
+import math
 import os
-from typing import Any, Dict, List
+from typing import Any
 
 
 def convert_to_pytorch_benchmark_format(args: argparse.Namespace,
-                                        metrics: Dict[str, List],
-                                        extra_info: Dict[str, Any]) -> List:
+                                        metrics: dict[str, list],
+                                        extra_info: dict[str, Any]) -> list:
     """
     Save the benchmark results in the format used by PyTorch OSS benchmark with
     on metric per record
@@ -34,6 +36,34 @@ def convert_to_pytorch_benchmark_format(args: argparse.Namespace,
                 "extra_info": extra_info,
             },
         }
+
+        tp = record["benchmark"]["extra_info"]["args"].get(
+            "tensor_parallel_size")
+        # Save tensor_parallel_size parameter if it's part of the metadata
+        if not tp and "tensor_parallel_size" in extra_info:
+            record["benchmark"]["extra_info"]["args"][
+                "tensor_parallel_size"] = extra_info["tensor_parallel_size"]
+
         records.append(record)
 
     return records
+
+
+class InfEncoder(json.JSONEncoder):
+
+    def clear_inf(self, o: Any):
+        if isinstance(o, dict):
+            return {k: self.clear_inf(v) for k, v in o.items()}
+        elif isinstance(o, list):
+            return [self.clear_inf(v) for v in o]
+        elif isinstance(o, float) and math.isinf(o):
+            return "inf"
+        return o
+
+    def iterencode(self, o: Any, *args, **kwargs) -> Any:
+        return super().iterencode(self.clear_inf(o), *args, **kwargs)
+
+
+def write_to_json(filename: str, records: list) -> None:
+    with open(filename, "w") as f:
+        json.dump(records, f, cls=InfEncoder)

benchmarks/cutlass_benchmarks/sparse_benchmarks.py

@@ -5,7 +5,8 @@ import copy
 import itertools
 import pickle as pkl
 import time
-from typing import Callable, Iterable, List, Tuple
+from collections.abc import Iterable
+from typing import Callable
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -228,7 +229,7 @@ def print_timers(timers: Iterable[TMeasurement]):
 
 
 def run(dtype: torch.dtype,
-        MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
+        MKNs: Iterable[tuple[int, int, int]]) -> Iterable[TMeasurement]:
     results = []
     for m, k, n in MKNs:
         timers = bench(dtype, m, k, n, f"scaled-{dtype}-gemm",
@@ -241,7 +242,7 @@ def run(dtype: torch.dtype,
 
 # output makers
 def make_output(data: Iterable[TMeasurement],
-                MKNs: Iterable[Tuple[int, int, int]],
+                MKNs: Iterable[tuple[int, int, int]],
                 base_description: str,
                 timestamp=None):
     print(f"== All Results {base_description} ====")
@@ -282,7 +283,7 @@ def run_model_bench(args):
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")
 
-    def model_shapes(model_name: str, tp_size: int) -> List[Tuple[int, int]]:
+    def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]:
         KNs = []
         for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
             KN[tp_split_dim] = KN[tp_split_dim] // tp_size

benchmarks/cutlass_benchmarks/utils.py

@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 # Cutlass bench utils
-from typing import Iterable, Tuple
+from collections.abc import Iterable
 
 import torch
 
@@ -27,7 +27,7 @@ def to_fp16(tensor: torch.Tensor) -> torch.Tensor:
 
 
 def make_rand_tensors(dtype: torch.dtype, m: int, n: int,
-                      k: int) -> Tuple[torch.Tensor, torch.Tensor]:
+                      k: int) -> tuple[torch.Tensor, torch.Tensor]:
     a = torch.randn((m, k), device='cuda') * 5
     b = torch.randn((n, k), device='cuda').t() * 5
 
@@ -63,7 +63,7 @@ def prune_to_2_4(tensor):
 
 
 def make_rand_sparse_tensors(dtype: torch.dtype, m: int, n: int,
-                             k: int) -> Tuple[torch.Tensor, torch.Tensor]:
+                             k: int) -> tuple[torch.Tensor, torch.Tensor]:
     a = torch.randn((m, k), device='cuda') * 5
     b = torch.randn((n, k), device='cuda').t() * 5
 
@@ -88,7 +88,7 @@ def make_rand_sparse_tensors(dtype: torch.dtype, m: int, n: int,
 
 def make_n_rand_sparse_tensors(num_tensors: int, dtype: torch.dtype,
                         m: int, n: int, k: int) -> \
-                        Tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]:
+                        tuple[Iterable[torch.Tensor], Iterable[torch.Tensor]]:
     ABs = []
     for _ in range(num_tensors):
         b_comp, e, a, b = make_rand_sparse_tensors(dtype, m, n, k)

benchmarks/cutlass_benchmarks/w8a8_benchmarks.py

@@ -5,7 +5,8 @@ import copy
 import itertools
 import pickle as pkl
 import time
-from typing import Callable, Iterable, List, Optional, Tuple
+from collections.abc import Iterable
+from typing import Callable, Optional
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -49,7 +50,7 @@ def bench_int8(
         n: int,
         label: str,
         sub_label: str,
-        bench_kernels: Optional[List[str]] = None) -> Iterable[TMeasurement]:
+        bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
     """Benchmark INT8-based kernels."""
     assert dtype == torch.int8
     a, b = make_rand_tensors(torch.int8, m, n, k)
@@ -101,7 +102,7 @@ def bench_fp8(
         n: int,
         label: str,
         sub_label: str,
-        bench_kernels: Optional[List[str]] = None) -> Iterable[TMeasurement]:
+        bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
     """Benchmark FP8-based kernels."""
     assert dtype == torch.float8_e4m3fn
     a, b = make_rand_tensors(torch.float8_e4m3fn, m, n, k)
@@ -180,7 +181,7 @@ def bench(dtype: torch.dtype,
           n: int,
           label: str,
           sub_label: str,
-          bench_kernels: Optional[List[str]] = None) -> Iterable[TMeasurement]:
+          bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
     if dtype == torch.int8:
         return bench_int8(dtype, m, k, n, label, sub_label, bench_kernels)
     if dtype == torch.float8_e4m3fn:
@@ -195,8 +196,8 @@ def print_timers(timers: Iterable[TMeasurement]):
 
 
 def run(dtype: torch.dtype,
-        MKNs: Iterable[Tuple[int, int, int]],
-        bench_kernels: Optional[List[str]] = None) -> Iterable[TMeasurement]:
+        MKNs: Iterable[tuple[int, int, int]],
+        bench_kernels: Optional[list[str]] = None) -> Iterable[TMeasurement]:
     results = []
     for m, k, n in MKNs:
         timers = bench(dtype,
@@ -212,7 +213,7 @@ def run(dtype: torch.dtype,
 
 
 def make_output(data: Iterable[TMeasurement],
-                MKNs: Iterable[Tuple[int, int, int]],
+                MKNs: Iterable[tuple[int, int, int]],
                 base_description: str,
                 timestamp=None):
     print(f"== All Results {base_description} ====")
@@ -248,7 +249,7 @@ def run_model_bench(args):
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")
 
-    def model_shapes(model_name: str, tp_size: int) -> List[Tuple[int, int]]:
+    def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]:
         KNs = []
         for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
             KN[tp_split_dim] = KN[tp_split_dim] // tp_size

benchmarks/fused_kernels/layernorm_rms_benchmarks.py

@@ -2,9 +2,10 @@
 
 import pickle as pkl
 import time
+from collections.abc import Iterable
 from dataclasses import dataclass
 from itertools import product
-from typing import Callable, Iterable, List, Optional
+from typing import Callable, Optional
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -29,7 +30,7 @@ class bench_params_t:
                 f'x DT {self.dtype}')
 
 
-def get_bench_params() -> List[bench_params_t]:
+def get_bench_params() -> list[bench_params_t]:
     ## Test Fixtures
     NUM_TOKENS = [2**x for x in range(11)]
     HIDDEN_SIZES = list(range(1024, 8129, 1024))

benchmarks/kernels/benchmark_grouped_gemm_cutlass.py

@@ -0,0 +1,340 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import torch
+import torch.utils.benchmark as benchmark
+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.fused_moe import (cutlass_moe_fp8,
+                                                            fused_experts,
+                                                            fused_topk)
+from vllm.utils import FlexibleArgumentParser
+
+DEFAULT_MODELS = [
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1", "nm-testing/deepseekv2-lite",
+    "ibm-granite/granite-3.0-1b-a400m", "ibm-granite/granite-3.0-3b-a800m"
+]
+DEFAULT_BATCH_SIZES = [1, 4, 8, 16, 32, 64, 128, 256, 512]
+DEFAULT_TP_SIZES = [1]
+
+PER_ACT_TOKEN_OPTS = [False]
+PER_OUT_CH_OPTS = [False]
+
+
+def to_fp8(tensor: torch.Tensor):
+    finfo = torch.finfo(torch.float8_e4m3fn)
+    return torch.round(tensor.clamp(
+        min=finfo.min, max=finfo.max)).to(dtype=torch.float8_e4m3fn)
+
+
+def bench_run(results: list[benchmark.Measurement], model: str,
+              num_experts: int, topk: int, per_act_token: bool,
+              per_out_ch: bool, mkn: tuple[int, int, int]):
+    label = "Quant Matmul"
+
+    sub_label = (
+        "{}, num_experts={}, topk={}, per_act_token={} per_out_ch={}, "
+        "MKN=({})".format(model, num_experts, topk, per_act_token, per_out_ch,
+                          mkn))
+
+    print(f"Testing: {sub_label}")
+
+    (m, k, n) = mkn
+
+    dtype = torch.half
+
+    a = torch.randn((m, k), device="cuda", dtype=dtype) / 10
+    w1 = torch.randn((num_experts, 2 * n, k), device="cuda", dtype=dtype) / 10
+    w2 = torch.randn((num_experts, k, n), device="cuda", dtype=dtype) / 10
+
+    _, a_scale = ops.scaled_fp8_quant(a)
+
+    w1_q = torch.empty((num_experts, 2 * n, k),
+                       device="cuda",
+                       dtype=torch.float8_e4m3fn)
+    w2_q = torch.empty((num_experts, k, n),
+                       device="cuda",
+                       dtype=torch.float8_e4m3fn)
+    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)
+
+    topk_weights, topk_ids = fused_topk(a, score, topk, renormalize=False)
+
+    def run_triton_moe(a: torch.Tensor, w1: torch.Tensor, w2: torch.Tensor,
+                       topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+                       w1_scale: torch.Tensor, w2_scale: torch.Tensor,
+                       a_scale: torch.Tensor, num_repeats: int):
+        for _ in range(num_repeats):
+            fused_experts(a,
+                          w1,
+                          w2,
+                          topk_weights,
+                          topk_ids,
+                          use_fp8_w8a8=True,
+                          w1_scale=w1_scale,
+                          w2_scale=w2_scale,
+                          a1_scale=a_scale)
+
+    def run_cutlass_moe(a: torch.Tensor, a_scale: torch.Tensor,
+                        w1: torch.Tensor, w2: torch.Tensor,
+                        w1_scale: torch.Tensor, 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,
+                        num_repeats: int):
+        for _ in range(num_repeats):
+            cutlass_moe_fp8(a,
+                            w1,
+                            w2,
+                            w1_scale,
+                            w2_scale,
+                            topk_weights,
+                            topk_ids,
+                            ab_strides1,
+                            c_strides1,
+                            ab_strides2,
+                            c_strides2,
+                            a1_scale=a_scale)
+
+    def run_cutlass_from_graph(
+            a: torch.Tensor, a_scale: torch.Tensor, w1_q: torch.Tensor,
+            w2_q: torch.Tensor, w1_scale: torch.Tensor, 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))):
+            return cutlass_moe_fp8(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)
+
+    def run_triton_from_graph(a: torch.Tensor, w1: torch.Tensor,
+                              w2: torch.Tensor, topk_weights: torch.Tensor,
+                              topk_ids: torch.Tensor, w1_scale: torch.Tensor,
+                              w2_scale: torch.Tensor, a_scale: torch.Tensor):
+        with set_current_vllm_config(
+                VllmConfig(parallel_config=ParallelConfig(
+                    pipeline_parallel_size=1))):
+            return fused_experts(a,
+                                 w1,
+                                 w2,
+                                 topk_weights,
+                                 topk_ids,
+                                 use_fp8_w8a8=True,
+                                 w1_scale=w1_scale,
+                                 w2_scale=w2_scale,
+                                 a1_scale=a_scale)
+
+    def replay_graph(graph, num_repeats):
+        for _ in range(num_repeats):
+            graph.replay()
+        torch.cuda.synchronize()
+
+    cutlass_stream = torch.cuda.Stream()
+    cutlass_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(cutlass_graph, stream=cutlass_stream):
+        run_cutlass_from_graph(a, a_scale, w1_q, w2_q, w1_scale, w2_scale,
+                               topk_weights, topk_ids, ab_strides1, c_strides1,
+                               ab_strides2, c_strides2)
+    torch.cuda.synchronize()
+
+    triton_stream = torch.cuda.Stream()
+    triton_graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(triton_graph, stream=triton_stream):
+        run_triton_from_graph(a, w1_q_notransp, w2_q_notransp, topk_weights,
+                              topk_ids, w1_scale, w2_scale, a_scale)
+    torch.cuda.synchronize()
+
+    min_run_time = 5
+    num_warmup = 5
+    num_runs = 25
+
+    globals = {
+        # Baseline params
+        "w1": w1,
+        "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,
+        # cuda graph params
+        "cutlass_graph": cutlass_graph,
+        "triton_graph": triton_graph,
+        # Gen params
+        "a": a,
+        "topk_weights": topk_weights,
+        "topk_ids": topk_ids,
+        "num_runs": num_runs,
+        # Kernels
+        "run_triton_moe": run_triton_moe,
+        "run_cutlass_moe": run_cutlass_moe,
+        "replay_graph": replay_graph,
+    }
+
+    # Warmup
+    run_triton_moe(a, w1_q_notransp, w2_q_notransp, topk_weights, topk_ids,
+                   w1_scale, w2_scale, a_scale, num_warmup)
+
+    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
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="triton_moe",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    replay_graph(triton_graph, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt="replay_graph(triton_graph, num_runs)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="triton_moe_cuda_graphs",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    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_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
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="grouped_gemm_moe",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    # Warmup
+    replay_graph(cutlass_graph, num_warmup)
+
+    results.append(
+        benchmark.Timer(
+            stmt="replay_graph(cutlass_graph, num_runs)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="grouped_gemm_moe_cuda_graphs",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+
+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:
+        for tp in args.tp_sizes:
+            for layer in WEIGHT_SHAPES_MOE[model]:
+                num_experts = layer[0]
+                topk = layer[1]
+                size_k = layer[2]
+                size_n = layer[3] // tp
+
+                if len(args.limit_k) > 0 and size_k not in args.limit_k:
+                    continue
+
+                if len(args.limit_n) > 0 and size_n not in args.limit_n:
+                    continue
+
+                for per_act_token in PER_ACT_TOKEN_OPTS:
+                    for per_out_ch in PER_OUT_CH_OPTS:
+                        for size_m in DEFAULT_BATCH_SIZES:
+                            mkn = (size_m, size_k, size_n)
+                            bench_run(results, model, num_experts, topk,
+                                      per_act_token, per_out_ch, mkn)
+
+    compare = benchmark.Compare(results)
+    compare.print()
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="Benchmark Marlin across specified models/shapes/batches")
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=DEFAULT_MODELS,
+        choices=WEIGHT_SHAPES_MOE.keys(),
+    )
+    parser.add_argument("--tp-sizes",
+                        nargs="+",
+                        type=int,
+                        default=DEFAULT_TP_SIZES)
+    parser.add_argument("--batch-sizes",
+                        nargs="+",
+                        type=int,
+                        default=DEFAULT_BATCH_SIZES)
+    parser.add_argument("--limit-k", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-n", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-num-groups", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-per-act-token",
+                        nargs="+",
+                        type=int,
+                        default=[])
+    parser.add_argument("--limit-per-out-ch", nargs="+", type=int, default=[])
+
+    args = parser.parse_args()
+    main(args)

benchmarks/kernels/benchmark_layernorm.py

@@ -40,7 +40,7 @@ def main(num_tokens: int,
 
         end_time = time.perf_counter()
         if profile:
-            torch.cuda.cudart().cudaProfilerStart()
+            torch.cuda.cudart().cudaProfilerStop()
         return (end_time - start_time) / num_iters
 
     # Warmup.

benchmarks/kernels/benchmark_lora.py

@@ -9,7 +9,7 @@ from dataclasses import dataclass
 from enum import Enum, auto
 from itertools import product
 from pathlib import Path
-from typing import Any, Callable, Dict, List, Optional, Tuple
+from typing import Any, Callable, Optional
 
 import torch
 import torch.utils.benchmark as TBenchmark
@@ -17,11 +17,7 @@ from torch.utils.benchmark import Measurement as TMeasurement
 from utils import ArgPool, Bench, CudaGraphBenchParams
 from weight_shapes import WEIGHT_SHAPES
 
-from vllm.lora.ops.triton_ops.bgmv_expand import bgmv_expand
-from vllm.lora.ops.triton_ops.bgmv_expand_slice import bgmv_expand_slice
-from vllm.lora.ops.triton_ops.bgmv_shrink import bgmv_shrink
-from vllm.lora.ops.triton_ops.sgmv_expand import sgmv_expand
-from vllm.lora.ops.triton_ops.sgmv_shrink import sgmv_shrink
+from vllm.lora.ops.triton_ops import LoRAKernelMeta, lora_expand, lora_shrink
 from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT
 from vllm.utils import FlexibleArgumentParser
 
@@ -61,15 +57,15 @@ def make_rand_lora_weight_tensor(k: int,
 
 
 def make_rand_tensors(
-    a_shape: Tuple[int],
-    b_shape: Tuple[int],
-    c_shape: Tuple[int],
+    a_shape: tuple[int],
+    b_shape: tuple[int],
+    c_shape: tuple[int],
     a_dtype: torch.dtype,
     b_dtype: torch.dtype,
     c_dtype: torch.dtype,
     num_slices: int,
     device: str = "cuda",
-) -> Tuple[torch.Tensor, List[torch.Tensor], torch.Tensor]:
+) -> tuple[torch.Tensor, list[torch.Tensor], torch.Tensor]:
     """
     Make LoRA input/output matrices.
     """
@@ -89,7 +85,7 @@ def make_prompt_lora_mapping(num_prompts: int, num_active_loras: int,
                              sort_by_lora_id: bool,
                              device: str) -> torch.Tensor:
     """
-    All prompts are mapped to a Lora ID in range [0, num_active_loras).
+    All prompts are mapped to a LoRA ID in range [0, num_active_loras).
     where 0 refers to first lora, 1 refers to second lora and so on.
     """
     assert num_active_loras > 0
@@ -135,7 +131,7 @@ def make_token_lora_mapping(num_tokens: int, num_prompts: int,
 
 
 def ref_group_gemm(ref_out: torch.Tensor, input: torch.Tensor,
-                   lora_weights: List[torch.Tensor],
+                   lora_weights: list[torch.Tensor],
                    seq_lens_cpu: torch.Tensor,
                    prompt_lora_mapping_cpu: torch.Tensor, scaling: float,
                    add_inputs: Optional[bool]):
@@ -153,7 +149,6 @@ def ref_group_gemm(ref_out: torch.Tensor, input: torch.Tensor,
         result = torch.nn.functional.linear(x, w)
         result *= scaling
         out_list.append(result)
-    torch.cat(out_list, dim=0)
 
     cat_result = torch.cat(out_list, dim=0)
 
@@ -167,62 +162,35 @@ class OpType(Enum):
     """
     LoRA Ops to benchmark and its properties.
     """
-    SGMV_SHRINK = auto()
-    BGMV_SHRINK = auto()
-    SGMV_EXPAND = auto()
-    BGMV_EXPAND = auto()
-    BGMV_EXPAND_SLICE = auto()
+    LORA_SHRINK = auto()
+    LORA_EXPAND = auto()
 
     @staticmethod
     def from_str(s: str) -> "OpType":
-        if s.lower() == 'sgmv_shrink':
-            return OpType.SGMV_SHRINK
-        if s.lower() == 'sgmv_expand':
-            return OpType.SGMV_EXPAND
-        if s.lower() == 'bgmv_shrink':
-            return OpType.BGMV_SHRINK
-        if s.lower() == 'bgmv_expand':
-            return OpType.BGMV_EXPAND
-        if s.lower() == "bgmv_expand_slice":
-            return OpType.BGMV_EXPAND_SLICE
+        if s.lower() == "lora_shrink":
+            return OpType.LORA_SHRINK
+        if s.lower() == "lora_expand":
+            return OpType.LORA_EXPAND
         raise ValueError(f"Unrecognized str {s} to convert to OpType")
 
     def is_shrink_fn(self) -> bool:
-        return self in [OpType.SGMV_SHRINK, OpType.BGMV_SHRINK]
+        return self in [OpType.LORA_SHRINK]
 
     def is_expand_fn(self) -> bool:
-        return self in [OpType.SGMV_EXPAND, OpType.BGMV_EXPAND]
+        return self in [OpType.LORA_EXPAND]
 
-    def is_prefill_op(self) -> bool:
-        return self in [OpType.SGMV_SHRINK, OpType.SGMV_EXPAND]
-
-    def is_decode_op(self) -> bool:
-        return self in [
-            OpType.BGMV_SHRINK, OpType.BGMV_EXPAND, OpType.BGMV_EXPAND_SLICE
-        ]
-
-    def is_expand_slice_fn(self) -> bool:
-        return self in [OpType.BGMV_EXPAND_SLICE]
-
-    def num_slices(self) -> List[int]:
-        if self in [OpType.SGMV_EXPAND, OpType.SGMV_SHRINK]:
-            # SGMV kernels supports slices
-            return [1, 2, 3]
-        if self in [OpType.BGMV_SHRINK, OpType.BGMV_EXPAND]:
-            return [1]
-        if self in [OpType.BGMV_EXPAND_SLICE]:
-            return [2, 3]
-        raise ValueError(f"Unrecognized OpType {self}")
+    def num_slices(self) -> list[int]:
+        return [1, 2, 3]
 
     def mkn(self, batch_size: int, seq_length: int, hidden_size: int,
-            lora_rank: int) -> Tuple[int, int, int]:
+            lora_rank: int) -> tuple[int, int, int]:
         num_tokens = batch_size * seq_length
         if self.is_shrink_fn():
             m = num_tokens
             k = hidden_size
             n = lora_rank
         else:
-            assert self.is_expand_fn() or self.is_expand_slice_fn()
+            assert self.is_expand_fn()
             m = num_tokens
             k = lora_rank
             n = hidden_size
@@ -230,20 +198,20 @@ class OpType(Enum):
 
     def matmul_dtypes(
             self, op_dtype: torch.dtype
-    ) -> Tuple[torch.dtype, torch.dtype, torch.dtype]:
+    ) -> tuple[torch.dtype, torch.dtype, torch.dtype]:
         """
         return a type, b type and c type for A x B = C
         """
         if self.is_shrink_fn():
             return op_dtype, op_dtype, torch.float32
         else:
-            assert self.is_expand_fn() or self.is_expand_slice_fn()
+            assert self.is_expand_fn()
             return torch.float32, op_dtype, op_dtype
 
     def matmul_shapes(
             self, batch_size: int, seq_length: int, hidden_size: int,
             lora_rank: int, num_loras: int,
-            num_slices: int) -> Tuple[Tuple[int], Tuple[int], Tuple[int]]:
+            num_slices: int) -> tuple[tuple[int], tuple[int], tuple[int]]:
         """
         Given num_slices, return the shapes of the A, B, and C matrices
         in A x B = C, for the op_type
@@ -251,56 +219,39 @@ class OpType(Enum):
         m, k, n = self.mkn(batch_size, seq_length, hidden_size, lora_rank)
 
         b_shape = (num_loras, n, k)  # col-major
-        if self == OpType.SGMV_SHRINK:
-            # SGMV shrink supports num_slices inherently in the kernel
+        if self in [OpType.LORA_SHRINK]:
+            # LoRA shrink kernels support num_slices inherently in the kernel.
             return ((m, k), b_shape, (num_slices, m, n))
-        if self == OpType.SGMV_EXPAND:
-            # SGMV expand supports num_slices inherently in the kernel
+        if self in [OpType.LORA_EXPAND]:
+            # LoRA expand kernels support num_slices inherently in the kernel
             return ((num_slices, m, k), b_shape, (m, n * num_slices))
-        if self == OpType.BGMV_SHRINK:
-            return ((m, k), b_shape, (m, n))
-        if self == OpType.BGMV_EXPAND:
-            return ((m, k), b_shape, (m, n))
-        if self == OpType.BGMV_EXPAND_SLICE:
-            return ((num_slices, m, k), b_shape, (m, n * num_slices))
-
         raise ValueError(f"Unrecognized op_type {self}")
 
     def bench_fn(self) -> Callable:
+        if self == OpType.LORA_SHRINK:
+            return lora_shrink
+        if self == OpType.LORA_EXPAND:
+            return lora_expand
 
-        def emulate_bgmv_expand_slice(kwargs_list: List[Dict[str, Any]]):
-            for x in kwargs_list:
-                bgmv_expand_slice(**x)
-
-        if self == OpType.SGMV_SHRINK:
-            return sgmv_shrink
-        if self == OpType.SGMV_EXPAND:
-            return sgmv_expand
-        if self == OpType.BGMV_SHRINK:
-            return bgmv_shrink
-        if self == OpType.BGMV_EXPAND:
-            return bgmv_expand
-        if self == OpType.BGMV_EXPAND_SLICE:
-            return emulate_bgmv_expand_slice
         raise ValueError(f"Unrecognized optype {self}")
 
     def run_ref_group_gemm(self, output: torch.Tensor, input: torch.Tensor,
-                           lora_weights: List[torch.Tensor],
+                           lora_weights: list[torch.Tensor],
                            **kwargs) -> Callable:
-        """Each benchmark operation expected the input, lora_weights and outputs
+        """Each benchmark operation expects the input, lora_weights and outputs
            in a slightly different format. Refer to self.matmul_shapes().
            run_ref_group_gemm accounts for those differences in executing a
            reference group gemm for correctness testing.
         """
         w_dtype = lora_weights[0].dtype
         num_slices = len(lora_weights)
-        if self == OpType.SGMV_SHRINK:
+        if self in [OpType.LORA_SHRINK]:
             for slice_idx in range(num_slices):
                 ref_group_gemm(ref_out=output[slice_idx, :],
                                input=input,
                                lora_weights=lora_weights[slice_idx],
                                **kwargs)
-        if self == OpType.SGMV_EXPAND:
+        elif self in [OpType.LORA_EXPAND]:
             hidden_size = lora_weights[0].shape[1]
             for slice_idx in range(num_slices):
                 slice_offset = slice_idx * hidden_size
@@ -309,28 +260,8 @@ class OpType(Enum):
                     input=input[slice_idx].clone().to(dtype=w_dtype),
                     lora_weights=lora_weights[slice_idx],
                     **kwargs)
-        if self == OpType.BGMV_SHRINK:
-            assert num_slices == 1
-            ref_group_gemm(ref_out=output,
-                           input=input,
-                           lora_weights=lora_weights[0],
-                           **kwargs)
-        if self == OpType.BGMV_EXPAND:
-            assert num_slices == 1
-            ref_group_gemm(ref_out=output,
-                           input=input.clone().to(dtype=w_dtype),
-                           lora_weights=lora_weights[0],
-                           **kwargs)
-        if self == OpType.BGMV_EXPAND_SLICE:
-            hidden_size = lora_weights[0].shape[1]
-            for slice_idx in range(num_slices):
-                slice_offset = slice_idx * hidden_size
-                ref_group_gemm(
-                    ref_out=output[:, slice_offset:slice_offset + hidden_size],
-                    input=input[slice_idx].clone().to(dtype=w_dtype),
-                    lora_weights=lora_weights[slice_idx],
-                    **kwargs)
-        raise ValueError(f"Unrecognized optype {self}")
+        else:
+            raise ValueError(f"Unrecognized optype {self}")
 
 
 @dataclass
@@ -384,13 +315,13 @@ class BenchmarkTensors:
     """
     # matmul tensors
     input: torch.Tensor
-    lora_weights_lst: List[torch.Tensor]
+    lora_weights_lst: list[torch.Tensor]
     output: torch.Tensor
-    # metadata tensors
+    # LoRA kernel metadata
+    lora_kernel_meta: LoRAKernelMeta
+    # Metadata tensors used in testing correctness
     seq_lens: torch.Tensor
-    seq_start_loc: torch.Tensor
     prompt_lora_mapping: torch.Tensor
-    token_lora_mapping: torch.Tensor
 
     def io_types(self) -> str:
         return (f"{dtype_to_str(self.input.dtype)}x"
@@ -417,26 +348,29 @@ class BenchmarkTensors:
         assert ctx.num_active_loras <= ctx.num_loras
         total_tokens = ctx.batch_size * ctx.seq_length
 
+        # Make metadata tensors involved in correctness testing.
         # Prepare seq lens tensor
         seq_len_tensor = torch.randint(ctx.seq_length, ctx.seq_length + 1,
                                        (ctx.batch_size, ))
-        # Prepare seq_start_loc tensor
-        seq_start_loc_tensor = torch.cumsum(torch.tensor(
-            [0] + seq_len_tensor[:-1].tolist(), dtype=torch.long),
-                                            dim=0)
         assert total_tokens == seq_len_tensor.sum()
         # Prepare prompt lora indices tensor
         prompt_lora_indices_tensor = make_prompt_lora_mapping(
             ctx.batch_size, ctx.num_active_loras, ctx.sort_by_lora_id, "cpu")
-        # Prepare token lora indices tensor
+
+        # Make LoRAKernelMeta
         token_lora_indices_tensor = make_token_lora_mapping(
             total_tokens, ctx.batch_size, prompt_lora_indices_tensor,
             seq_len_tensor, "cpu")
+        lora_kernel_meta = LoRAKernelMeta.make(
+            max_loras=ctx.num_loras,
+            max_num_tokens=token_lora_indices_tensor.size(0),
+            device="cpu")
+        lora_kernel_meta.prepare_tensors(
+            token_lora_mapping=token_lora_indices_tensor)
 
         return BenchmarkTensors(input_tensor, lora_weights, output_tensor,
-                                seq_len_tensor, seq_start_loc_tensor,
-                                prompt_lora_indices_tensor,
-                                token_lora_indices_tensor)
+                                lora_kernel_meta, seq_len_tensor,
+                                prompt_lora_indices_tensor)
 
     def sanity_check(self) -> None:
         """
@@ -446,9 +380,9 @@ class BenchmarkTensors:
         # check metadata tensors
         assert torch.sum(self.seq_lens) == num_tokens
         num_seqs = self.seq_lens.shape[0]
-        assert self.seq_start_loc.shape[0] == num_seqs
+        #assert self.seq_start_loc.shape[0] == num_seqs
         assert self.prompt_lora_mapping.shape[0] == num_seqs
-        assert self.token_lora_mapping.shape[0] == num_tokens
+        assert self.lora_kernel_meta.token_lora_mapping.shape[0] == num_tokens
 
     def to_device(self, device: str):
         """
@@ -463,54 +397,31 @@ class BenchmarkTensors:
         self.input = to_device(self.input)
         self.output = to_device(self.output)
         self.seq_lens = to_device(self.seq_lens)
-        self.seq_start_loc = to_device(self.seq_start_loc)
         self.prompt_lora_mapping = to_device(self.prompt_lora_mapping)
-        self.token_lora_mapping = to_device(self.token_lora_mapping)
         for i in range(len(self.lora_weights_lst)):
             self.lora_weights_lst[i] = to_device(self.lora_weights_lst[i])
 
-    def metadata(self) -> Tuple[int, int, int]:
+        # LoRA meta
+        for field_name in LoRAKernelMeta.__dataclass_fields__:
+            field = getattr(self.lora_kernel_meta, field_name)
+            assert isinstance(field, torch.Tensor)
+            setattr(self.lora_kernel_meta, field_name, to_device(field))
+
+    def metadata(self) -> tuple[int, int, int]:
         """
         Return num_seqs, num_tokens and max_seq_len
         """
         num_seqs = self.seq_lens.shape[0]
-        num_tokens = self.token_lora_mapping.shape[0]
+        num_tokens = self.lora_kernel_meta.token_lora_mapping.shape[0]
         max_seq_len = torch.max(self.seq_lens).item()
         num_slices = len(self.lora_weights_lst)
         return num_seqs, num_tokens, max_seq_len, num_slices
 
-    def convert_to_sgmv_benchmark_tensors(self):
-        """
-        For sgmv punica kernels, when consecutive sequences have the
-        same LoRA ID, we just merge them together.
-        This happens in punica.py::compute_metadata
-        """
-
-        # Collapse seq_lens and seq_start_loc
-        _, seq_lens = torch.unique_consecutive(self.token_lora_mapping,
-                                               return_counts=True)
-        cum_result = torch.cumsum(seq_lens, dim=0)
-        seq_start_loc = torch.zeros_like(seq_lens)
-        seq_start_loc[1:].copy_(cum_result[:-1])
-
-        # Collapse prompt mapping
-        prompt_lora_mapping = torch.unique_consecutive(
-            self.prompt_lora_mapping)
-
-        assert torch.sum(seq_lens) == torch.sum(self.seq_lens), \
-         f"dont match - new {torch.sum(seq_lens)} vs {torch.sum(self.seq_lens)}"
-
-        self.prompt_lora_mapping = prompt_lora_mapping.to(
-            dtype=self.prompt_lora_mapping.dtype)
-        self.seq_lens = seq_lens.to(dtype=self.seq_lens.dtype)
-        self.seq_start_loc = seq_start_loc.to(dtype=self.seq_start_loc.dtype)
-
-    def as_sgmv_shrink_kwargs(self) -> Dict[str, Any]:
-        self.convert_to_sgmv_benchmark_tensors()
+    def as_lora_shrink_kwargs(self) -> dict[str, Any]:
         self.sanity_check()
         self.to_device(self.input.device)
 
-        num_seqs, num_tokens, max_seq_len, num_slices = self.metadata()
+        _, num_tokens, _, num_slices = self.metadata()
 
         # Sanity check matrix shapes.
         i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
@@ -531,22 +442,20 @@ class BenchmarkTensors:
             'inputs': self.input,
             'lora_a_weights': self.lora_weights_lst,
             'output_tensor': self.output,
-            'b_seq_start_loc': self.seq_start_loc,
-            'seq_len_tensor': self.seq_lens,
-            'lora_indices_tensor': self.prompt_lora_mapping,
-            'batches': num_seqs,
-            'max_seq_length': max_seq_len,
-            'token_nums': num_tokens,
+            'token_lora_mapping': self.lora_kernel_meta.token_lora_mapping,
+            'token_indices_sorted_by_lora_ids':
+            self.lora_kernel_meta.token_indices_sorted_by_lora_ids,
+            'num_tokens_per_lora': self.lora_kernel_meta.num_tokens_per_lora,
+            'lora_token_start_loc': self.lora_kernel_meta.lora_token_start_loc,
+            'lora_ids': self.lora_kernel_meta.active_lora_ids,
             'scaling': 1.0,
         }
 
-    def as_sgmv_expand_kwargs(self, add_inputs: bool) -> Dict[str, Any]:
-
-        self.convert_to_sgmv_benchmark_tensors()
+    def as_lora_expand_kwargs(self, add_inputs: bool) -> dict[str, Any]:
         self.sanity_check()
         self.to_device(self.input.device)
 
-        num_seqs, num_tokens, max_seq_len, num_slices = self.metadata()
+        _, num_tokens, _, num_slices = self.metadata()
 
         # Sanity check matrix shapes.
         i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
@@ -568,124 +477,28 @@ class BenchmarkTensors:
             'inputs': self.input,
             'lora_b_weights': self.lora_weights_lst,
             'output_tensor': self.output,
-            'b_seq_start_loc': self.seq_start_loc,
-            'seq_len_tensor': self.seq_lens,
-            'lora_indices_tensor': self.prompt_lora_mapping,
-            'batches': num_seqs,
-            'max_seq_length': max_seq_len,
-            'token_nums': num_tokens,
+            'token_lora_mapping': self.lora_kernel_meta.token_lora_mapping,
+            'token_indices_sorted_by_lora_ids':
+            self.lora_kernel_meta.token_indices_sorted_by_lora_ids,
+            'num_tokens_per_lora': self.lora_kernel_meta.num_tokens_per_lora,
+            'lora_token_start_loc': self.lora_kernel_meta.lora_token_start_loc,
+            'lora_ids': self.lora_kernel_meta.active_lora_ids,
             'offset_start': 0,
             'add_inputs': add_inputs,
         }
 
-    def as_bgmv_shrink_kwargs(self) -> Dict[str, Any]:
-        assert len(self.lora_weights_lst) == 1
-        self.to_device(self.input.device)
-
-        _, num_tokens, _, _ = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_tokens, hidden_size]
-        assert len(i_shape) == 2
-        assert i_shape[0] == num_tokens
-        hidden_size = i_shape[1]
-        # Expected lora weight shape [num_loras, lora_rank, hidden_size]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == hidden_size
-        lora_rank = lw_shape[1]
-        # Expected output shape [num_tokens, lora_rank]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, lora_rank)
-
-        return {
-            'inputs': self.input,
-            'lora_a_weights': self.lora_weights_lst[0],
-            'output_tensor': self.output,
-            'lora_indices_tensor': self.token_lora_mapping,
-            'scaling': 1.0
-        }
-
-    def as_bgmv_expand_kwargs(self, add_inputs: bool):
-        assert len(self.lora_weights_lst) == 1
-        self.to_device(self.input.device)
-
-        _, num_tokens, _, _ = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_tokens, lora_rank]
-        assert len(i_shape) == 2
-        assert i_shape[0] == num_tokens
-        lora_rank = i_shape[1]
-        # Expected lora weight shape [num_loras, hidden_size, lora_rank]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == lora_rank
-        hidden_size = lw_shape[1]
-        # Expected output shape [num_tokens, hidden_size]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, hidden_size)
-
-        return {
-            'inputs': self.input,
-            'lora_b_weights': self.lora_weights_lst[0],
-            'output_tensor': self.output,
-            'lora_indices_tensor': self.token_lora_mapping,
-            'add_inputs': add_inputs
-        }
-
-    def as_bgmv_expand_slice_kwargs(self, add_inputs: bool) -> Dict[str, Any]:
-
-        _, num_tokens, _, num_slices = self.metadata()
-        # Sanity check shapes
-        i_shape, lw_shape, o_shape = self.input.shape, self.lora_weights_lst[
-            0].shape, self.output.shape
-        # Expected input shape [num_slices, num_tokens, lora_rank]
-        assert len(i_shape) == 3
-        assert i_shape[0] == num_slices
-        assert i_shape[1] == num_tokens
-        lora_rank = i_shape[2]
-        # Expected lora weight shape [num_loras, hidden_size, lora_rank]
-        assert len(lw_shape) == 3
-        assert lw_shape[2] == lora_rank
-        hidden_size = lw_shape[1]
-        # Expected output shape [num_tokens, hidden_size * num_slices]
-        assert len(o_shape) == 2
-        assert o_shape == (num_tokens, hidden_size * num_slices)
-
-        self.to_device(self.input.device)
-
-        kwargs_list = []
-        for i in range(num_slices):
-            kwargs_list.append({
-                'inputs': self.input[i],
-                'lora_b_weights': self.lora_weights_lst[i],
-                'output_tensor': self.output,
-                'lora_indices_tensor': self.token_lora_mapping,
-                'slice_offset': i * hidden_size,
-                'slice_size': hidden_size,
-                'add_inputs': add_inputs,
-            })
-        return {'kwargs_list': kwargs_list}
-
     def bench_fn_kwargs(self,
                         op_type: OpType,
-                        add_inputs: Optional[bool] = None) -> Dict[str, Any]:
+                        add_inputs: Optional[bool] = None) -> dict[str, Any]:
         if op_type.is_shrink_fn():
             assert add_inputs is None
         else:
             assert add_inputs is not None
 
-        if op_type == OpType.SGMV_SHRINK:
-            return self.as_sgmv_shrink_kwargs()
-        if op_type == OpType.SGMV_EXPAND:
-            return self.as_sgmv_expand_kwargs(add_inputs)
-        if op_type == OpType.BGMV_SHRINK:
-            return self.as_bgmv_shrink_kwargs()
-        if op_type == OpType.BGMV_EXPAND:
-            return self.as_bgmv_expand_kwargs(add_inputs)
-        if op_type == OpType.BGMV_EXPAND_SLICE:
-            return self.as_bgmv_expand_slice_kwargs(add_inputs)
+        if op_type == OpType.LORA_SHRINK:
+            return self.as_lora_shrink_kwargs()
+        if op_type == OpType.LORA_EXPAND:
+            return self.as_lora_expand_kwargs(add_inputs)
         raise ValueError(f"Unrecognized optype {self}")
 
     def test_correctness(self, op_type: OpType,
@@ -734,7 +547,7 @@ def bench_optype(ctx: BenchmarkContext,
         assert expand_fn_add_inputs is not None
 
     # BenchmarkContext -> BenchmarkTensors
-    bench_tensors : List[BenchmarkTensors] = \
+    bench_tensors : list[BenchmarkTensors] = \
         [BenchmarkTensors.make(ctx, op_type) for _ in range(arg_pool_size)]
     for bt in bench_tensors:
         bt.sanity_check()
@@ -746,7 +559,7 @@ def bench_optype(ctx: BenchmarkContext,
             for bt in bench_tensors
         ])
 
-    # BenchmarkTensors -> Dict (kwargs)
+    # BenchmarkTensors -> dict (kwargs)
     kwargs_list = [
         bt.bench_fn_kwargs(op_type, add_inputs=expand_fn_add_inputs)
         for bt in bench_tensors
@@ -841,7 +654,7 @@ def use_cuda_graph_recommendation() -> str:
             """
 
 
-def print_timers(timers: List[TMeasurement],
+def print_timers(timers: list[TMeasurement],
                  args: Optional[argparse.Namespace] = None):
     compare = TBenchmark.Compare(timers)
     compare.print()
@@ -861,7 +674,7 @@ def print_timers(timers: List[TMeasurement],
           "small num_loras the goal should be to match the torch.mm numbers.")
 
 
-def run(args: argparse.Namespace, bench_ctxs: List[BenchmarkContext]):
+def run(args: argparse.Namespace, bench_ctxs: list[BenchmarkContext]):
 
     if args.cuda_graph_nops is not None:
         assert args.cuda_graph_nops > 0
@@ -873,14 +686,7 @@ def run(args: argparse.Namespace, bench_ctxs: List[BenchmarkContext]):
     timers = []
     for bench_ctx in bench_ctxs:
         for seq_len in args.seq_lengths:
-            bench_ops: List[OpType] = []
-            if seq_len == 1:
-                # bench all decode ops
-                bench_ops = [op for op in args.op_types if op.is_decode_op()]
-            else:
-                # bench all prefill ops
-                bench_ops = [op for op in args.op_types if op.is_prefill_op()]
-
+            bench_ops: list[OpType] = args.op_types
             seq_len_timers = []
             for bench_op in bench_ops:
                 for num_slices in bench_op.num_slices():
@@ -921,10 +727,10 @@ def run(args: argparse.Namespace, bench_ctxs: List[BenchmarkContext]):
             pickle.dump(timers, f)
 
 
-def as_benchmark_contexts(hidden_sizes: List[int], lora_ranks: List[int],
-                          args: argparse.Namespace) -> List[BenchmarkContext]:
+def as_benchmark_contexts(hidden_sizes: list[int], lora_ranks: list[int],
+                          args: argparse.Namespace) -> list[BenchmarkContext]:
 
-    ctxs: List[BenchmarkContext] = []
+    ctxs: list[BenchmarkContext] = []
     for batch_size, hidden_size, lora_rank, num_loras, sort_by_lora_id in product(  # noqa
             args.batch_sizes, list(hidden_sizes), lora_ranks, args.num_loras,
             args.sort_by_lora_id):
@@ -954,7 +760,7 @@ def run_list_bench(args: argparse.Namespace):
           f"  LoRA Ranks {args.lora_ranks}")
 
     # Get all benchmarking contexts
-    bench_contexts: List[BenchmarkContext] = as_benchmark_contexts(
+    bench_contexts: list[BenchmarkContext] = as_benchmark_contexts(
         hidden_sizes=args.hidden_sizes, lora_ranks=args.lora_ranks, args=args)
 
     run(args, bench_contexts)
@@ -975,7 +781,7 @@ def run_range_bench(args: argparse.Namespace):
           f" LoRA Ranks {lora_ranks}")
 
     # Get all benchmarking contexts
-    bench_contexts: List[BenchmarkContext] = as_benchmark_contexts(
+    bench_contexts: list[BenchmarkContext] = as_benchmark_contexts(
         hidden_sizes=hidden_sizes, lora_ranks=lora_ranks, args=args)
 
     run(args, bench_contexts)
@@ -1002,7 +808,7 @@ def run_model_bench(args: argparse.Namespace):
           f" LoRA Ranks {args.lora_ranks}")
 
     # Get all benchmarking contexts
-    bench_contexts: List[BenchmarkContext] = as_benchmark_contexts(
+    bench_contexts: list[BenchmarkContext] = as_benchmark_contexts(
         hidden_sizes=hidden_sizes, lora_ranks=args.lora_ranks, args=args)
 
     run(args, bench_contexts)
@@ -1090,13 +896,13 @@ Benchmark LoRA kernels:
     {use_cuda_graph_recommendation()}
 
     list_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32
+        python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32
 
     model_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16  --lora-ranks 16 --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 
+        python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16  --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 
 
     range_bench example:
-        python3 benchmarks/kernels/benchmark_lora.py range_bench  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16   --num-loras 1 4 --op-types bgmv_shrink bgmv_expand sgmv_shrink sgmv_expand bgmv_expand_slice --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 
+        python3 benchmarks/kernels/benchmark_lora.py range_bench  --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16   --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 
             """,  # noqa: E501
         formatter_class=argparse.RawTextHelpFormatter)
 

benchmarks/kernels/benchmark_machete.py

@@ -7,9 +7,10 @@ import math
 import os
 import pickle as pkl
 import time
+from collections.abc import Iterable
 from dataclasses import dataclass
 from itertools import product
-from typing import Callable, Iterable, List, Optional, Tuple
+from typing import Callable, Optional
 
 import pandas as pd
 import torch
@@ -44,7 +45,6 @@ def terse_type_name(dt):
         torch.float16: "fp16",
         torch.int8: "int8",
         torch.float8_e4m3fn: "fp8",
-        torch.bfloat16: "bf16",
         torch.float: "float",
         torch.int: "int",
     }[dt]
@@ -102,8 +102,8 @@ def quantize_and_pack(atype: torch.dtype,
     return w_ref, w_q, w_s, w_zp
 
 
-def create_bench_tensors(shape: Tuple[int, int, int], types: TypeConfig,
-                         group_size: Optional[int]) -> List[BenchmarkTensors]:
+def create_bench_tensors(shape: tuple[int, int, int], types: TypeConfig,
+                         group_size: Optional[int]) -> list[BenchmarkTensors]:
     m, n, k = shape
 
     # we want to make sure that weights don't fit into L2 cache between runs so
@@ -114,7 +114,7 @@ def create_bench_tensors(shape: Tuple[int, int, int], types: TypeConfig,
 
     a = rand_data((m, k), types.act_type, scale=5)
 
-    benchmark_tensors: List[BenchmarkTensors] = []
+    benchmark_tensors: list[BenchmarkTensors] = []
     for _ in range(num_weights):
         w = rand_data((k, n), types.act_type, scale=5)
 
@@ -258,7 +258,7 @@ def machete_create_bench_fn(bt: BenchmarkTensors,
 
     return lambda: ops.machete_mm(
         a=bt.a,
-        b_q=bt.w_q,
+        b_q=w_q,
         b_type=bt.wtype,
         b_group_scales=bt.w_g_s,
         b_group_zeros=w_g_zp,
@@ -276,7 +276,7 @@ def machete_create_bench_fn(bt: BenchmarkTensors,
 
 
 def bench_fns(label: str, sub_label: str, description: str,
-              fns: List[Callable]):
+              fns: list[Callable]):
 
     min_run_time = 1 if not NVTX_PROFILE else 0.1
     res = TBenchmark.Timer(
@@ -311,7 +311,7 @@ def bench(types: TypeConfig,
           n: int,
           label: str,
           sub_label: str,
-          sweep_schedules: bool = True) -> List[TMeasurement]:
+          sweep_schedules: bool = True) -> list[TMeasurement]:
     benchmark_tensors = create_bench_tensors((m, n, k), types, group_size)
     sub_label += f", L={len(benchmark_tensors)}"
 
@@ -414,12 +414,12 @@ def bench(types: TypeConfig,
 
 
 # runner
-def print_timers(timers: List[TMeasurement]):
+def print_timers(timers: list[TMeasurement]):
     compare = TBenchmark.Compare(timers)
     compare.print()
 
 
-def run(args, MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
+def run(args, MKNs: Iterable[tuple[int, int, int]]) -> Iterable[TMeasurement]:
     types = TypeConfig(
         act_type=args.act_type,
         weight_type=scalar_types.uint4b8 if args.group_zero_type is None \
@@ -431,7 +431,7 @@ def run(args, MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
         token_scale_type=args.token_scale_type,
     )
 
-    results: List[TMeasurement] = []
+    results: list[TMeasurement] = []
     for m, k, n in MKNs:
         timers = bench(types,
                        args.group_size,
@@ -449,8 +449,8 @@ def run(args, MKNs: Iterable[Tuple[int, int, int]]) -> Iterable[TMeasurement]:
 
 # output makers
 def make_output(
-    data: List[TMeasurement],
-    MKNs: Iterable[Tuple[int, int, int]],
+    data: list[TMeasurement],
+    MKNs: Iterable[tuple[int, int, int]],
     base_description: str,
     timestamp=None,
 ):
@@ -497,7 +497,7 @@ def run_model_bench(args):
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")
 
-    def model_shapes(model_name: str, tp_size: int) -> List[Tuple[int, int]]:
+    def model_shapes(model_name: str, tp_size: int) -> list[tuple[int, int]]:
         KNs = []
         for KN, tp_split_dim in copy.deepcopy(WEIGHT_SHAPES[model_name]):
             KN[tp_split_dim] = KN[tp_split_dim] // tp_size

benchmarks/kernels/benchmark_marlin.py

@@ -1,7 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
 
-from typing import List
-
 import torch
 import torch.utils.benchmark as benchmark
 from benchmark_shapes import WEIGHT_SHAPES
@@ -10,6 +8,8 @@ from vllm import _custom_ops as ops
 from vllm.model_executor.layers.quantization.gptq_marlin_24 import (
     GPTQ_MARLIN_24_MAX_PARALLEL, GPTQ_MARLIN_24_MIN_THREAD_N,
     GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_24_SUPPORTED_QUANT_TYPES)
+from vllm.model_executor.layers.quantization.utils.allspark_utils import (
+    ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD, ALLSPARK_SUPPORTED_QUANT_TYPES)
 from vllm.model_executor.layers.quantization.utils.marlin_utils import (
     GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
     MARLIN_SUPPORTED_GROUP_SIZES, query_marlin_supported_quant_types)
@@ -18,18 +18,18 @@ from vllm.model_executor.layers.quantization.utils.marlin_utils_test import (
 from vllm.model_executor.layers.quantization.utils.marlin_utils_test_24 import (
     marlin_24_quantize)
 from vllm.model_executor.layers.quantization.utils.quant_utils import (
-    gptq_pack, gptq_quantize_weights, sort_weights)
+    gptq_pack, gptq_quantize_weights, quantize_weights, sort_weights)
 from vllm.scalar_type import ScalarType
 from vllm.utils import FlexibleArgumentParser
 
 DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
-DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
+DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8192]
 
 ACT_ORDER_OPTS = [False, True]
 K_FULL_OPTS = [False, True]
 
 
-def bench_run(results: List[benchmark.Measurement], model: str,
+def bench_run(results: list[benchmark.Measurement], model: str,
               act_order: bool, is_k_full: bool, quant_type: ScalarType,
               group_size: int, size_m: int, size_k: int, size_n: int):
     label = "Quant Matmul"
@@ -81,6 +81,27 @@ def bench_run(results: List[benchmark.Measurement], model: str,
                                           GPTQ_MARLIN_24_MAX_PARALLEL)
     marlin_zp = torch.zeros_like(marlin_s, dtype=torch.int)
 
+    # AllSpark W8A16 quant
+    as_supported_case = (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)
+            CUBLAS_M_THRESHOLD = ALLSPARK_AMPERE_M_CUBLAS_THRESHOLD
+
     globals = {
         # Gen params
         "quant_type": quant_type,
@@ -109,10 +130,19 @@ def bench_run(results: List[benchmark.Measurement], model: str,
         # GPTQ params
         "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,
         # Kernels
         "gptq_marlin_gemm": ops.gptq_marlin_gemm,
         "gptq_marlin_24_gemm": ops.gptq_marlin_24_gemm,
         "gptq_marlin_repack": ops.gptq_marlin_repack,
+        "allspark_w8a16_gemm": ops.allspark_w8a16_gemm,
     }
 
     min_run_time = 1
@@ -172,13 +202,24 @@ def bench_run(results: List[benchmark.Measurement], model: str,
             description="gptq_marlin_repack",
         ).blocked_autorange(min_run_time=min_run_time))
 
+    if as_supported_case:
+        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
+                globals=globals,
+                label=label,
+                sub_label=sub_label,
+                description="allspark_w8a16_gemm_fp32",
+            ).blocked_autorange(min_run_time=min_run_time))
+
 
 def main(args):
     print("Benchmarking models:")
     for i, model in enumerate(args.models):
         print(f"[{i}]  {model}")
 
-    results: List[benchmark.Measurement] = []
+    results: list[benchmark.Measurement] = []
 
     for model in args.models:
         for layer in WEIGHT_SHAPES[model]:

benchmarks/kernels/benchmark_moe.py

@@ -1,10 +1,12 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import argparse
+import json
 import time
+from contextlib import nullcontext
 from datetime import datetime
 from itertools import product
-from typing import Any, Dict, List, Tuple, TypedDict
+from typing import Any, TypedDict
 
 import ray
 import torch
@@ -16,8 +18,7 @@ from vllm.model_executor.layers.fused_moe.fused_moe import *
 from vllm.platforms import current_platform
 from vllm.utils import FlexibleArgumentParser
 
-FP8_DTYPE = torch.float8_e4m3fnuz if current_platform.is_rocm(
-) else torch.float8_e4m3fn
+FP8_DTYPE = current_platform.fp8_dtype()
 
 
 class BenchmarkConfig(TypedDict):
@@ -29,18 +30,18 @@ class BenchmarkConfig(TypedDict):
     num_stages: int
 
 
-def benchmark_config(
-    config: BenchmarkConfig,
-    num_tokens: int,
-    num_experts: int,
-    shard_intermediate_size: int,
-    hidden_size: int,
-    topk: int,
-    dtype: torch.dtype,
-    use_fp8_w8a8: bool,
-    use_int8_w8a16: bool,
-    num_iters: int = 100,
-) -> float:
+def benchmark_config(config: BenchmarkConfig,
+                     num_tokens: int,
+                     num_experts: int,
+                     shard_intermediate_size: int,
+                     hidden_size: int,
+                     topk: int,
+                     dtype: torch.dtype,
+                     use_fp8_w8a8: bool,
+                     use_int8_w8a16: bool,
+                     num_iters: int = 100,
+                     block_quant_shape: List[int] = None,
+                     use_deep_gemm: bool = False) -> float:
     init_dtype = torch.float16 if use_fp8_w8a8 else dtype
     x = torch.randn(num_tokens, hidden_size, dtype=dtype)
     if use_int8_w8a16:
@@ -81,8 +82,24 @@ def benchmark_config(
                                dtype=torch.float32)
         w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
     if use_fp8_w8a8:
-        w1_scale = torch.randn(num_experts, dtype=torch.float32)
-        w2_scale = torch.randn(num_experts, dtype=torch.float32)
+        if block_quant_shape:
+            block_n, block_k = block_quant_shape[0], block_quant_shape[1]
+            E = num_experts
+            N = shard_intermediate_size // 2
+            K = hidden_size
+            factor_for_scale = 1e-2
+            n_tiles_w1 = (2 * N + block_n - 1) // block_n
+            n_tiles_w2 = (K + block_n - 1) // block_n
+            k_tiles_w1 = (K + block_k - 1) // block_k
+            k_tiles_w2 = (N + block_k - 1) // block_k
+            w1_scale = torch.rand((E, n_tiles_w1, k_tiles_w1),
+                                  dtype=torch.float32) * factor_for_scale
+            w2_scale = torch.rand((E, n_tiles_w2, k_tiles_w2),
+                                  dtype=torch.float32) * factor_for_scale
+        else:
+            w1_scale = torch.randn(num_experts, dtype=torch.float32)
+            w2_scale = torch.randn(num_experts, dtype=torch.float32)
+
         a1_scale = torch.randn(1, dtype=torch.float32)
         a2_scale = torch.randn(1, dtype=torch.float32)
 
@@ -97,21 +114,41 @@ def benchmark_config(
     def run():
         from vllm.model_executor.layers.fused_moe import override_config
         with override_config(config):
-            fused_moe(
-                x,
-                w1,
-                w2,
-                input_gating,
-                topk,
-                renormalize=True,
-                inplace=True,
-                use_fp8_w8a8=use_fp8_w8a8,
-                use_int8_w8a16=use_int8_w8a16,
-                w1_scale=w1_scale,
-                w2_scale=w2_scale,
-                a1_scale=a1_scale,
-                a2_scale=a2_scale,
-            )
+            if use_deep_gemm:
+                topk_weights, topk_ids = fused_topk(x, input_gating, topk,
+                                                    False)
+                return fused_experts(
+                    x,
+                    w1,
+                    w2,
+                    topk_weights,
+                    topk_ids,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                    allow_deep_gemm=True,
+                )
+            else:
+                fused_moe(
+                    x,
+                    w1,
+                    w2,
+                    input_gating,
+                    topk,
+                    renormalize=True,
+                    inplace=True,
+                    use_fp8_w8a8=use_fp8_w8a8,
+                    use_int8_w8a16=use_int8_w8a16,
+                    w1_scale=w1_scale,
+                    w2_scale=w2_scale,
+                    a1_scale=a1_scale,
+                    a2_scale=a2_scale,
+                    block_shape=block_quant_shape,
+                )
 
     # JIT compilation & warmup
     run()
@@ -132,7 +169,7 @@ def benchmark_config(
     start_event = torch.cuda.Event(enable_timing=True)
     end_event = torch.cuda.Event(enable_timing=True)
 
-    latencies: List[float] = []
+    latencies: list[float] = []
     for i in range(num_iters):
         prepare(i)
         torch.cuda.synchronize()
@@ -175,8 +212,9 @@ def get_rocm_tuning_space(use_fp16):
     return param_ranges
 
 
-def get_configs_compute_bound(use_fp16) -> List[Dict[str, int]]:
-    configs: List[BenchmarkConfig] = []
+def get_configs_compute_bound(use_fp16,
+                              block_quant_shape) -> list[dict[str, int]]:
+    configs: list[BenchmarkConfig] = []
 
     if current_platform.is_rocm():
         param_ranges = get_rocm_tuning_space(use_fp16)
@@ -204,17 +242,27 @@ def get_configs_compute_bound(use_fp16) -> List[Dict[str, int]]:
     for config_values in product(*values):
         config = dict(zip(keys, config_values))
         configs.append(config)
+
+    # Remove configs that are not compatible with fp8 block quantization
+    # BLOCK_SIZE_K must be a multiple of block_k
+    # BLOCK_SIZE_N must be a multiple of block_n
+    if block_quant_shape is not None and not use_fp16:
+        block_n, block_k = block_quant_shape[0], block_quant_shape[1]
+        for config in configs[:]:
+            if config["BLOCK_SIZE_K"] % block_k != 0 or config[
+                    "BLOCK_SIZE_N"] % block_n != 0:
+                configs.remove(config)
     return configs
 
 
 def prune_rocm_search_space(num_tokens, shard_intermediate_size, hidden_size,
-                            search_space, is_fp16):
+                            search_space, is_fp16, topk):
     N1, K1 = shard_intermediate_size, hidden_size
     N2, K2 = hidden_size, shard_intermediate_size // 2
-    pruned_space_1 = prune_rocm_configs(num_tokens * 2, N1, K1, search_space,
-                                        is_fp16)
-    pruned_space_2 = prune_rocm_configs(num_tokens * 2, N2, K2, search_space,
-                                        is_fp16)
+    pruned_space_1 = prune_rocm_configs(num_tokens * topk, N1, K1,
+                                        search_space, is_fp16)
+    pruned_space_2 = prune_rocm_configs(num_tokens * topk, N2, K2,
+                                        search_space, is_fp16)
     search_space = merge_unique_dicts(pruned_space_1, pruned_space_2)
     return search_space
 
@@ -335,7 +383,9 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
-    ) -> Tuple[Dict[str, int], float]:
+        block_quant_shape: List[int] = None,
+        use_deep_gemm: bool = False,
+    ) -> tuple[dict[str, int], float]:
         current_platform.seed_everything(self.seed)
         dtype_str = get_config_dtype_str(dtype,
                                          use_int8_w8a16=use_int8_w8a16,
@@ -355,10 +405,18 @@ class BenchmarkWorker:
         else:
             config = op_config[min(op_config.keys(),
                                    key=lambda x: abs(x - num_tokens))]
-        kernel_time = benchmark_config(config, num_tokens, num_experts,
-                                       shard_intermediate_size, hidden_size,
-                                       topk, dtype, use_fp8_w8a8,
-                                       use_int8_w8a16)
+        kernel_time = benchmark_config(config,
+                                       num_tokens,
+                                       num_experts,
+                                       shard_intermediate_size,
+                                       hidden_size,
+                                       topk,
+                                       dtype,
+                                       use_fp8_w8a8,
+                                       use_int8_w8a16,
+                                       num_iters=100,
+                                       block_quant_shape=block_quant_shape,
+                                       use_deep_gemm=use_deep_gemm)
         return config, kernel_time
 
     def tune(
@@ -371,8 +429,10 @@ class BenchmarkWorker:
         dtype: torch.dtype,
         use_fp8_w8a8: bool,
         use_int8_w8a16: bool,
-        search_space: List[Dict[str, int]],
-    ) -> Dict[str, int]:
+        search_space: list[dict[str, int]],
+        block_quant_shape: list[int],
+        use_deep_gemm: bool,
+    ) -> dict[str, int]:
         best_config = None
         best_time = float("inf")
         if current_platform.is_rocm():
@@ -380,21 +440,25 @@ class BenchmarkWorker:
             search_space = prune_rocm_search_space(num_tokens,
                                                    shard_intermediate_size,
                                                    hidden_size, search_space,
-                                                   is_fp16)
+                                                   is_fp16, topk)
 
-        with torch.cuda.device(self.device_id):
+        with torch.cuda.device(self.device_id) if current_platform.is_rocm(
+        ) else nullcontext():
             for config in tqdm(search_space):
                 try:
-                    kernel_time = benchmark_config(config,
-                                                   num_tokens,
-                                                   num_experts,
-                                                   shard_intermediate_size,
-                                                   hidden_size,
-                                                   topk,
-                                                   dtype,
-                                                   use_fp8_w8a8,
-                                                   use_int8_w8a16,
-                                                   num_iters=20)
+                    kernel_time = benchmark_config(
+                        config,
+                        num_tokens,
+                        num_experts,
+                        shard_intermediate_size,
+                        hidden_size,
+                        topk,
+                        dtype,
+                        use_fp8_w8a8,
+                        use_int8_w8a16,
+                        num_iters=20,
+                        block_quant_shape=block_quant_shape,
+                        use_deep_gemm=use_deep_gemm)
                 except triton.runtime.autotuner.OutOfResources:
                     # Some configurations may be invalid and fail to compile.
                     continue
@@ -434,10 +498,10 @@ def sort_config(config: BenchmarkConfig) -> BenchmarkConfig:
     }
 
 
-def save_configs(configs: Dict[int, BenchmarkConfig], num_experts: int,
+def save_configs(configs: dict[int, BenchmarkConfig], num_experts: int,
                  shard_intermediate_size: int, hidden_size: int, topk: int,
-                 dtype: torch.dtype, use_fp8_w8a8: bool,
-                 use_int8_w8a16: bool) -> None:
+                 dtype: torch.dtype, use_fp8_w8a8: bool, use_int8_w8a16: bool,
+                 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)
@@ -445,7 +509,7 @@ def save_configs(configs: Dict[int, BenchmarkConfig], num_experts: int,
     # NOTE(woosuk): The current naming convention uses w2.shape[2], which
     # is the intermediate size after silu_and_mul.
     filename = get_config_file_name(num_experts, shard_intermediate_size // 2,
-                                    dtype_str)
+                                    dtype_str, block_quant_shape)
 
     print(f"Writing best config to {filename}...")
     with open(filename, "w") as f:
@@ -453,9 +517,17 @@ def save_configs(configs: Dict[int, BenchmarkConfig], num_experts: int,
         f.write("\n")
 
 
+def get_weight_block_size_safety(config, default_value=None):
+
+    quantization_config = getattr(config, 'quantization_config', {})
+    if isinstance(quantization_config, dict):
+        return quantization_config.get('weight_block_size', default_value)
+    return default_value
+
+
 def main(args: argparse.Namespace):
     print(args)
-
+    block_quant_shape = None
     config = AutoConfig.from_pretrained(
         args.model, trust_remote_code=args.trust_remote_code)
     if config.architectures[0] == "DbrxForCausalLM":
@@ -468,12 +540,22 @@ def main(args: argparse.Namespace):
         topk = config.num_experts_per_tok
         intermediate_size = config.intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
-    elif config.architectures[0] == "DeepseekV3ForCausalLM":
+    elif (config.architectures[0] == "DeepseekV3ForCausalLM"
+          or config.architectures[0] == "DeepseekV2ForCausalLM"):
         E = config.n_routed_experts
         topk = config.num_experts_per_tok
         intermediate_size = config.moe_intermediate_size
         shard_intermediate_size = 2 * intermediate_size // args.tp_size
+        block_quant_shape = get_weight_block_size_safety(config)
+    elif config.architectures[0] == "Qwen2MoeForCausalLM":
+        E = config.num_experts
+        topk = config.num_experts_per_tok
+        intermediate_size = config.moe_intermediate_size
+        shard_intermediate_size = 2 * intermediate_size // args.tp_size
     else:
+        if not hasattr(config, "hidden_size"):
+            # Support for llama4
+            config = config.text_config
         # Default: Mixtral.
         E = config.num_local_experts
         topk = config.num_experts_per_tok
@@ -493,11 +575,13 @@ def main(args: argparse.Namespace):
     else:
         batch_sizes = [args.batch_size]
 
+    use_deep_gemm = bool(args.use_deep_gemm)
+
     ray.init()
     num_gpus = int(ray.available_resources()["GPU"])
     workers = [BenchmarkWorker.remote(args.seed) for _ in range(num_gpus)]
 
-    def _distribute(method: str, inputs: List[Any]) -> List[Any]:
+    def _distribute(method: str, inputs: list[Any]) -> list[Any]:
         outputs = []
         worker_idx = 0
         for input_args in inputs:
@@ -510,27 +594,30 @@ def main(args: argparse.Namespace):
 
     if args.tune:
         is_fp16 = not (use_fp8_w8a8 or use_int8_w8a16)
-        search_space = get_configs_compute_bound(is_fp16)
+        search_space = get_configs_compute_bound(is_fp16, block_quant_shape)
         print(f"Start tuning over {len(search_space)} configurations...")
 
         start = time.time()
         configs = _distribute(
             "tune", [(batch_size, E, shard_intermediate_size, hidden_size,
-                      topk, dtype, use_fp8_w8a8, use_int8_w8a16, search_space)
+                      topk, dtype, use_fp8_w8a8, use_int8_w8a16, search_space,
+                      block_quant_shape, use_deep_gemm)
                      for batch_size in batch_sizes])
         best_configs = {
             M: sort_config(config)
             for M, config in zip(batch_sizes, configs)
         }
         save_configs(best_configs, E, shard_intermediate_size, hidden_size,
-                     topk, dtype, use_fp8_w8a8, use_int8_w8a16)
+                     topk, dtype, use_fp8_w8a8, use_int8_w8a16,
+                     block_quant_shape)
         end = time.time()
         print(f"Tuning took {end - start:.2f} seconds")
     else:
         outputs = _distribute(
-            "benchmark", [(batch_size, E, shard_intermediate_size, hidden_size,
-                           topk, dtype, use_fp8_w8a8, use_int8_w8a16)
-                          for batch_size in batch_sizes])
+            "benchmark",
+            [(batch_size, E, shard_intermediate_size, hidden_size, topk, dtype,
+              use_fp8_w8a8, use_int8_w8a16, block_quant_shape, use_deep_gemm)
+             for batch_size in batch_sizes])
 
         for batch_size, (config, kernel_time) in zip(batch_sizes, outputs):
             print(f"Batch size: {batch_size}, config: {config}")
@@ -551,6 +638,7 @@ if __name__ == "__main__":
                         type=str,
                         choices=["auto", "fp8_w8a8", "int8_w8a16"],
                         default="auto")
+    parser.add_argument("--use-deep-gemm", action="store_true")
     parser.add_argument("--seed", type=int, default=0)
     parser.add_argument("--batch-size", type=int, required=False)
     parser.add_argument("--tune", action="store_true")

benchmarks/kernels/benchmark_paged_attention.py

@@ -2,17 +2,21 @@
 
 import random
 import time
-from typing import List, Optional
+from typing import Optional
 
 import torch
 
 from vllm import _custom_ops as ops
+from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
                         create_kv_caches_with_random)
 
-NUM_BLOCKS = 1024
+logger = init_logger(__name__)
+
+NUM_BLOCKS = 128 * 1024
 PARTITION_SIZE = 512
+PARTITION_SIZE_ROCM = 256
 
 
 @torch.inference_mode()
@@ -54,7 +58,7 @@ def main(
 
     # Create the block tables.
     max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size
-    block_tables_lst: List[List[int]] = []
+    block_tables_lst: list[list[int]] = []
     for _ in range(num_seqs):
         block_table = [
             random.randint(0, NUM_BLOCKS - 1)
@@ -80,6 +84,12 @@ def main(
     # Prepare for the paged attention kernel.
     output = torch.empty_like(query)
     if version == "v2":
+        if current_platform.is_rocm():
+            global PARTITION_SIZE
+            if not args.custom_paged_attn:
+                PARTITION_SIZE = 1024
+            else:
+                PARTITION_SIZE = PARTITION_SIZE_ROCM
         num_partitions = ((max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE)
         tmp_output = torch.empty(
             size=(num_seqs, num_query_heads, num_partitions, head_size),
@@ -123,32 +133,53 @@ def main(
                     v_scale,
                 )
             elif version == "v2":
-                ops.paged_attention_v2(
-                    output,
-                    exp_sums,
-                    max_logits,
-                    tmp_output,
-                    query,
-                    key_cache,
-                    value_cache,
-                    num_kv_heads,
-                    scale,
-                    block_tables,
-                    seq_lens,
-                    block_size,
-                    max_seq_len,
-                    alibi_slopes,
-                    kv_cache_dtype,
-                    k_scale,
-                    v_scale,
-                )
+                if not args.custom_paged_attn:
+                    ops.paged_attention_v2(
+                        output,
+                        exp_sums,
+                        max_logits,
+                        tmp_output,
+                        query,
+                        key_cache,
+                        value_cache,
+                        num_kv_heads,
+                        scale,
+                        block_tables,
+                        seq_lens,
+                        block_size,
+                        max_seq_len,
+                        alibi_slopes,
+                        kv_cache_dtype,
+                        k_scale,
+                        v_scale,
+                    )
+                else:
+                    ops.paged_attention_rocm(
+                        output,
+                        exp_sums,
+                        max_logits,
+                        tmp_output,
+                        query,
+                        key_cache,
+                        value_cache,
+                        num_kv_heads,
+                        scale,
+                        block_tables,
+                        seq_lens,
+                        block_size,
+                        max_seq_len,
+                        alibi_slopes,
+                        kv_cache_dtype,
+                        k_scale,
+                        v_scale,
+                    )
             else:
                 raise ValueError(f"Invalid version: {version}")
         torch.cuda.synchronize()
 
         end_time = time.perf_counter()
         if profile:
-            torch.cuda.cudart().cudaProfilerStart()
+            torch.cuda.cudart().cudaProfilerStop()
         return (end_time - start_time) / num_iters
 
     # Warmup.
@@ -165,6 +196,9 @@ def main(
 
 
 if __name__ == '__main__':
+    logger.warning("This script benchmarks the paged attention kernel. "
+                   "By default this is no longer used in vLLM inference.")
+
     parser = FlexibleArgumentParser(
         description="Benchmark the paged attention kernel.")
     parser.add_argument("--version",
@@ -195,6 +229,9 @@ if __name__ == '__main__':
         help="Data type for kv cache storage. If 'auto', will use model "
         "data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. "
         "ROCm (AMD GPU) supports fp8 (=fp8_e4m3)")
+    parser.add_argument("--custom-paged-attn",
+                        action="store_true",
+                        help="Use custom paged attention")
     args = parser.parse_args()
     print(args)
 

benchmarks/kernels/benchmark_quant.py

@@ -40,7 +40,7 @@ def main(num_tokens: int,
 
         end_time = time.perf_counter()
         if profile:
-            torch.cuda.cudart().cudaProfilerStart()
+            torch.cuda.cudart().cudaProfilerStop()
         return (end_time - start_time) / num_iters
 
     # Warmup.

benchmarks/kernels/benchmark_rmsnorm.py

@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import itertools
-from typing import Optional, Tuple, Union
+from typing import Optional, Union
 
 import torch
 import triton
@@ -22,7 +22,7 @@ class HuggingFaceRMSNorm(nn.Module):
         self,
         x: torch.Tensor,
         residual: Optional[torch.Tensor] = None,
-    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
         orig_dtype = x.dtype
         x = x.to(torch.float32)
         if residual is not None:
@@ -139,7 +139,7 @@ def calculate_diff(batch_size, seq_len, hidden_size, use_residual=True):
 
     print(f"Naive output={output_naive}")
     print(f"FlashInfer output={output_flashinfer}")
-    print(f"VLLM output={output_vllm}")
+    print(f"vLLM output={output_vllm}")
 
     if torch.allclose(output_naive, output_flashinfer, atol=1e-2,
                       rtol=1e-2) and torch.allclose(

benchmarks/kernels/benchmark_rope.py

@@ -1,7 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 from itertools import accumulate
-from typing import List, Optional
+from typing import Optional
 
 import nvtx
 import torch
@@ -39,7 +39,7 @@ def benchmark_rope_kernels_multi_lora(
                             })
     # non-batched RoPE takes only one scaling factor, we create multiple
     # instances to simulate the same behavior
-    non_batched_ropes: List[RotaryEmbedding] = []
+    non_batched_ropes: list[RotaryEmbedding] = []
     for scaling_factor in scaling_factors:
         non_batched_ropes.append(
             get_rope(head_size, rotary_dim, max_position, base, is_neox_style,

benchmarks/kernels/benchmark_shapes.py

@@ -75,3 +75,19 @@ WEIGHT_SHAPES = {
         [7168, 8192],
     ],
 }
+
+WEIGHT_SHAPES_MOE = {
+    "nm-testing/Mixtral-8x7B-Instruct-v0.1": [
+        [8, 2, 4096, 28672],
+        [8, 2, 14336, 4096],
+    ],
+    "nm-testing/deepseekv2-lite": [
+        [64, 6, 2048, 1408],
+    ],
+    "ibm-granite/granite-3.0-1b-a400m": [
+        [32, 8, 1024, 1024],
+    ],
+    "ibm-granite/granite-3.0-3b-a800m": [
+        [40, 8, 1024, 1536],
+    ],
+}

benchmarks/kernels/benchmark_w8a8_block_fp8.py

@@ -0,0 +1,420 @@
+# SPDX-License-Identifier: Apache-2.0
+# Adapted from sglang quantization/tuning_block_wise_kernel.py
+
+import argparse
+import json
+import multiprocessing as mp
+import os
+import time
+from datetime import datetime
+from typing import Any
+
+import torch
+import tqdm
+import triton
+
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    _w8a8_block_fp8_matmul)
+from vllm.platforms import current_platform
+from vllm.utils import FlexibleArgumentParser
+
+mp.set_start_method("spawn", force=True)
+
+assert current_platform.is_cuda(
+), "Only support tune w8a8 block fp8 kernel on CUDA device."
+
+DTYPE_MAP = {
+    "float32": torch.float32,
+    "float16": torch.float16,
+    "half": torch.half,
+    "bfloat16": torch.bfloat16,
+}
+
+
+def w8a8_block_matmul(
+    A: torch.Tensor,
+    B: torch.Tensor,
+    As: torch.Tensor,
+    Bs: torch.Tensor,
+    block_size: list[int],
+    config: dict[str, Any],
+    output_dtype: torch.dtype = torch.float16,
+) -> torch.Tensor:
+    """This function performs matrix multiplication with 
+    block-wise quantization.
+
+    It takes two input tensors `A` and `B` with scales `As` and `Bs`.
+    The output is returned in the specified `output_dtype`.
+
+    Args:
+        A: The input tensor, e.g., activation.
+        B: The input tensor, e.g., weight.
+        As: The per-token-group quantization scale for `A`.
+        Bs: The per-block quantization scale for `B`.
+        block_size: The block size for per-block quantization. 
+                    It should be 2-dim, e.g., [128, 128].
+        output_dytpe: The dtype of the returned tensor.
+
+    Returns:
+        torch.Tensor: The result of matmul.
+    """
+    assert len(block_size) == 2
+    block_n, block_k = block_size[0], block_size[1]
+
+    assert A.shape[-1] == B.shape[-1]
+    assert A.shape[:-1] == As.shape[:-1] and A.is_contiguous()
+    assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
+    M = A.numel() // A.shape[-1]
+
+    assert B.ndim == 2 and B.is_contiguous() and Bs.ndim == 2
+    N, K = B.shape
+    assert triton.cdiv(N, block_n) == Bs.shape[0]
+    assert triton.cdiv(K, block_k) == Bs.shape[1]
+
+    C_shape = A.shape[:-1] + (N, )
+    C = A.new_empty(C_shape, dtype=output_dtype)
+
+    def grid(META):
+        return (triton.cdiv(M, META["BLOCK_SIZE_M"]) *
+                triton.cdiv(N, META["BLOCK_SIZE_N"]), )
+
+    if A.dtype == torch.float8_e4m3fn:
+        kernel = _w8a8_block_fp8_matmul
+    else:
+        raise RuntimeError(
+            "Currently, only support tune w8a8 block fp8 kernel.")
+
+    kernel[grid](
+        A,
+        B,
+        C,
+        As,
+        Bs,
+        M,
+        N,
+        K,
+        block_n,
+        block_k,
+        A.stride(-2),
+        A.stride(-1),
+        B.stride(1),
+        B.stride(0),
+        C.stride(-2),
+        C.stride(-1),
+        As.stride(-2),
+        As.stride(-1),
+        Bs.stride(1),
+        Bs.stride(0),
+        **config,
+    )
+
+    return C
+
+
+def get_configs_compute_bound():
+    configs = []
+    for num_stages in [2, 3, 4, 5]:
+        for block_m in [16, 32, 64, 128, 256]:
+            for block_k in [64, 128]:
+                for block_n in [32, 64, 128, 256]:
+                    for num_warps in [4, 8]:
+                        for group_size in [1, 16, 32, 64]:
+                            configs.append({
+                                "BLOCK_SIZE_M": block_m,
+                                "BLOCK_SIZE_N": block_n,
+                                "BLOCK_SIZE_K": block_k,
+                                "GROUP_SIZE_M": group_size,
+                                "num_warps": num_warps,
+                                "num_stages": num_stages,
+                            })
+    return configs
+
+
+def get_weight_shapes(tp_size):
+    # NOTE(HandH1998): The weight shapes only works for DeepSeek-V3.
+    # Modify them, if you tune for another different model.
+    # cannot TP
+    total = [
+        (512 + 64, 7168),
+        ((128 + 64) * 128, 7168),
+        (128 * (128 + 128), 512),
+        (7168, 16384),
+        (7168, 18432),
+    ]
+    # N can TP
+    n_tp = [
+        (18432 * 2, 7168),
+        ((128 + 64) * 128, 7168),
+        (128 * (128 + 128), 512),
+        (24576, 1536),
+        (12288, 7168),
+        (4096, 7168),
+    ]
+    # K can TP
+    k_tp = [(7168, 18432), (7168, 16384), (7168, 2048)]
+
+    weight_shapes = []
+    for t in total:
+        weight_shapes.append(t)
+    for n_t in n_tp:
+        new_t = (n_t[0] // tp_size, n_t[1])
+        weight_shapes.append(new_t)
+    for k_t in k_tp:
+        new_t = (k_t[0], k_t[1] // tp_size)
+        weight_shapes.append(new_t)
+    return weight_shapes
+
+
+def benchmark_config(A,
+                     B,
+                     As,
+                     Bs,
+                     block_size,
+                     config,
+                     out_dtype=torch.float16,
+                     num_iters=10):
+
+    def run():
+        w8a8_block_matmul(A, B, As, Bs, block_size, config, out_dtype)
+
+    torch.cuda.synchronize()
+    # JIT complication & warmup
+    for _ in range(5):
+        run()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: list[float] = []
+    for i in range(num_iters):
+        torch.cuda.synchronize()
+        start_event.record()
+        run()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    avg = sum(latencies) / (num_iters * 10) * 1000  # us
+    return avg
+
+
+def tune(M, N, K, block_size, out_dtype, search_space, input_type):
+    factor_for_scale = 1e-2
+
+    if input_type == "fp8":
+        fp8_info = torch.finfo(torch.float8_e4m3fn)
+        fp8_max, fp8_min = fp8_info.max, fp8_info.min
+
+        A_fp32 = (
+            (torch.rand(M, K, dtype=torch.float32, device="cuda") - 0.5) * 2 *
+            fp8_max)
+        A = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+
+        B_fp32 = (
+            (torch.rand(N, K, dtype=torch.float32, device="cuda") - 0.5) * 2 *
+            fp8_max)
+        B = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn)
+    else:
+        raise RuntimeError(
+            "Currently, only support tune w8a8 block fp8 kernel.")
+
+    block_n, block_k = block_size[0], block_size[1]
+    n_tiles = (N + block_n - 1) // block_n
+    k_tiles = (K + block_k - 1) // block_k
+
+    As = torch.rand(M, k_tiles, dtype=torch.float32,
+                    device="cuda") * factor_for_scale
+    Bs = (torch.rand(n_tiles, k_tiles, dtype=torch.float32, device="cuda") *
+          factor_for_scale)
+
+    best_config = None
+    best_time = float("inf")
+    for config in tqdm(search_space):
+        try:
+            kernel_time = benchmark_config(
+                A,
+                B,
+                As,
+                Bs,
+                block_size,
+                config,
+                out_dtype,
+                num_iters=10,
+            )
+        except triton.runtime.autotuner.OutOfResources:
+            # Some configurations may be invalid and fail to compile.
+            continue
+
+        if kernel_time < best_time:
+            best_time = kernel_time
+            best_config = config
+    now = datetime.now()
+    print(f"{now.ctime()}] Completed tuning for batch_size={M}")
+    assert best_config is not None
+    return best_config
+
+
+def save_configs(
+    N,
+    K,
+    block_n,
+    block_k,
+    configs,
+    save_path,
+    input_type="fp8",
+) -> None:
+    os.makedirs(save_path, exist_ok=True)
+    device_name = current_platform.get_device_name().replace(" ", "_")
+    json_file_name = (
+        f"N={N},K={K},device_name={device_name},dtype={input_type}_w8a8,"
+        f"block_shape=[{block_n},{block_k}].json")
+
+    config_file_path = os.path.join(save_path, json_file_name)
+    print(f"Writing best config to {config_file_path}...")
+
+    with open(config_file_path, "w") as f:
+        json.dump(configs, f, indent=4)
+        f.write("\n")
+
+
+def tune_on_gpu(args_dict):
+    """Run tuning on a specific GPU."""
+    gpu_id = args_dict["gpu_id"]
+    batch_sizes = args_dict["batch_sizes"]
+    weight_shapes = args_dict["weight_shapes"]
+    args = args_dict["args"]
+
+    torch.cuda.set_device(gpu_id)
+    print(f"Starting tuning on GPU {gpu_id} with batch sizes {batch_sizes}")
+
+    block_n = args.block_n
+    block_k = args.block_k
+    out_dtype = DTYPE_MAP[args.out_dtype]
+    save_path = args.save_path
+    input_type = args.input_type
+
+    search_space = get_configs_compute_bound()
+    search_space = [
+        config for config in search_space
+        if block_k % config["BLOCK_SIZE_K"] == 0
+    ]
+
+    start = time.time()
+    for shape in tqdm(weight_shapes, desc=f"GPU {gpu_id} - Shapes"):
+        N, K = shape[0], shape[1]
+        print(f"[GPU {gpu_id}] Tune for weight shape of `N: {N}, K: {K}`")
+        benchmark_results = [
+            tune(
+                batch_size,
+                N,
+                K,
+                [block_n, block_k],
+                out_dtype,
+                search_space,
+                input_type,
+            ) for batch_size in tqdm(batch_sizes,
+                                     desc=f"GPU {gpu_id} - Batch sizes")
+        ]
+        best_configs = {
+            M: config
+            for M, config in zip(batch_sizes, benchmark_results)
+        }
+        save_configs(N, K, block_n, block_k, best_configs, save_path,
+                     input_type)
+
+    end = time.time()
+    print(f"Tuning on GPU {gpu_id} took {end - start:.2f} seconds")
+
+
+def distribute_batch_sizes(batch_sizes, num_gpus):
+    """Distribute batch sizes across available GPUs."""
+    batches_per_gpu = []
+    for i in range(num_gpus):
+        start_idx = i * len(batch_sizes) // num_gpus
+        end_idx = (i + 1) * len(batch_sizes) // num_gpus
+        batches_per_gpu.append(batch_sizes[start_idx:end_idx])
+    return batches_per_gpu
+
+
+def main(args):
+    print(args)
+    num_gpus = torch.cuda.device_count()
+    if num_gpus == 0:
+        raise RuntimeError("No GPU available for tuning")
+    print(f"Found {num_gpus} GPUs for parallel tuning")
+
+    torch.cuda.init()
+
+    if args.batch_size is None:
+        batch_sizes = [
+            1,
+            2,
+            4,
+            8,
+            16,
+            24,
+            32,
+            48,
+            64,
+            96,
+            128,
+            256,
+            512,
+            1024,
+            1536,
+            2048,
+            3072,
+            4096,
+        ]
+    else:
+        batch_sizes = [args.batch_size]
+        num_gpus = 1  # If only one batch size, use only one GPU
+
+    weight_shapes = get_weight_shapes(args.tp_size)
+
+    batches_per_gpu = distribute_batch_sizes(batch_sizes, num_gpus)
+
+    process_args = []
+    for gpu_id in range(num_gpus):
+        process_args.append({
+            "gpu_id": gpu_id,
+            "batch_sizes": batches_per_gpu[gpu_id],
+            "weight_shapes":
+            weight_shapes,  # Each GPU processes all weight shapes
+            "args": args,
+        })
+
+    ctx = mp.get_context("spawn")
+    with ctx.Pool(num_gpus) as pool:
+        pool.map(tune_on_gpu, process_args)
+
+    print("Multi-GPU tuning completed")
+
+
+if __name__ == "__main__":
+    parser = FlexibleArgumentParser(
+        description="""
+Tune triton w8a8 block fp8 for DeepSeek-V3/DeepSeek-R1:
+    python3 benchmark_w8a8_block_fp8.py --tp-size 8 --input-type fp8
+Then copy to model_executor/layers/quantization/utils/configs
+        """,
+        formatter_class=argparse.RawTextHelpFormatter)
+
+    parser.add_argument("--tp-size", "-tp", type=int, default=8)
+    parser.add_argument("--input-type",
+                        type=str,
+                        choices=["fp8"],
+                        default="fp8")
+    parser.add_argument(
+        "--out-dtype",
+        type=str,
+        choices=["float32", "float16", "bfloat16", "half"],
+        default="float16",
+    )
+    parser.add_argument("--block-n", type=int, default=128)
+    parser.add_argument("--block-k", type=int, default=128)
+    parser.add_argument("--batch-size", type=int, required=False)
+    parser.add_argument("--save-path", type=str, default="./")
+    args = parser.parse_args()
+
+    main(args)

benchmarks/kernels/deepgemm/README.md

@@ -0,0 +1,129 @@
+# DeepSeek DeepGEMM Kernels Benchmark
+
+This directory includes benchmarks between DeepSeek's DeepGEMM block fp8 kernels against vLLM's existing triton and CUTLASS-based kernels.
+
+Currently this just includes dense GEMMs and only works on Hopper GPUs.
+
+## Setup
+
+You need to install vLLM in your usual fashion, then install DeepGEMM from source in its own directory:
+
+```
+git clone --recursive https://github.com/deepseek-ai/DeepGEMM
+cd DeepGEMM
+python setup.py install
+uv pip install -e .
+```
+
+## Usage
+
+```
+python benchmark_fp8_block_dense_gemm.py
+INFO 02-26 21:55:13 [__init__.py:207] Automatically detected platform cuda.
+===== STARTING FP8 GEMM BENCHMARK =====
+PyTorch version: 2.5.1+cu124
+CUDA version: 12.4
+Triton version: 3.1.0
+Using device: NVIDIA H100 80GB HBM3
+WARNING 02-26 21:55:15 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+INFO 02-26 21:55:15 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+WARNING 02-26 21:55:16 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=18432,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+WARNING 02-26 21:55:17 [fp8_utils.py:458] Using default W8A8 Block FP8 kernel config. Performance might be sub-optimal! Config file not found at /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json
+INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+INFO 02-26 21:55:17 [fp8_utils.py:449] Using configuration from /home/mgoin/code/vllm/vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json for W8A8 Block FP8 kernel.
+
+===== PERFORMANCE COMPARISON =====
+
+DeepGEMM Implementation:
++------+-------+-------+-----------+--------+--------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   |
++------+-------+-------+-----------+--------+--------+
+|    8 |  4096 |  7168 | 102.9     | 4.6    | 286.4  |
+|    8 |  7168 | 18432 | 70.8      | 29.8   | 1868.8 |
+|    8 | 18432 |  7168 | 69.3      | 30.5   | 1911.8 |
+|   64 |  4096 |  7168 | 69.1      | 54.4   | 439.0  |
+|   64 |  7168 | 18432 | 69.4      | 243.6  | 1933.6 |
+|   64 | 18432 |  7168 | 70.4      | 240.3  | 1917.2 |
+|   64 | 24576 |  1536 | 70.1      | 68.9   | 584.6  |
+|   64 | 32768 |   512 | 68.4      | 31.4   | 307.1  |
+|   64 |  7168 | 16384 | 69.5      | 216.3  | 1718.5 |
+|  128 |  4096 |  7168 | 141.1     | 53.3   | 222.1  |
+|  128 |  7168 | 18432 | 71.9      | 470.5  | 1896.1 |
+|  128 | 18432 |  7168 | 69.3      | 488.2  | 1988.2 |
+| 1024 |  4096 |  7168 | 89.7      | 670.1  | 502.5  |
+| 1024 | 18432 |  7168 | 279.0     | 969.8  | 635.2  |
+| 2048 |  4096 |  7168 | 175.1     | 687.0  | 347.4  |
+| 4096 |  4096 |  7168 | 335.4     | 717.0  | 275.1  |
++------+-------+-------+-----------+--------+--------+
+
+vLLM Triton Implementation:
++------+-------+-------+-----------+--------+--------+--------------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   | vs DeepGEMM  |
++------+-------+-------+-----------+--------+--------+--------------+
+|    8 |  4096 |  7168 | 74.0      | 6.3    | 398.2  | 1.39x faster |
+|    8 |  7168 | 18432 | 89.6      | 23.6   | 1478.1 | 0.79x slower |
+|    8 | 18432 |  7168 | 113.2     | 18.7   | 1170.4 | 0.61x slower |
+|   64 |  4096 |  7168 | 79.4      | 47.3   | 382.2  | 0.87x slower |
+|   64 |  7168 | 18432 | 98.5      | 171.7  | 1363.0 | 0.70x slower |
+|   64 | 18432 |  7168 | 119.5     | 141.5  | 1129.4 | 0.59x slower |
+|   64 | 24576 |  1536 | 37.6      | 128.4  | 1089.7 | 1.86x faster |
+|   64 | 32768 |   512 | 38.7      | 55.5   | 542.6  | 1.77x faster |
+|   64 |  7168 | 16384 | 86.1      | 174.5  | 1386.4 | 0.81x slower |
+|  128 |  4096 |  7168 | 90.7      | 82.9   | 345.4  | 1.56x faster |
+|  128 |  7168 | 18432 | 144.0     | 234.9  | 946.9  | 0.50x slower |
+|  128 | 18432 |  7168 | 229.5     | 147.4  | 600.1  | 0.30x slower |
+| 1024 |  4096 |  7168 | 242.3     | 248.2  | 186.1  | 0.37x slower |
+| 1024 | 18432 |  7168 | 897.8     | 301.4  | 197.4  | 0.31x slower |
+| 2048 |  4096 |  7168 | 463.0     | 259.7  | 131.4  | 0.38x slower |
+| 4096 |  4096 |  7168 | 901.8     | 266.7  | 102.3  | 0.37x slower |
++------+-------+-------+-----------+--------+--------+--------------+
+
+vLLM CUTLASS Implementation:
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+| m    | n     | k     | Time (μs) | TFLOPS | GB/s   | vs DeepGEMM  | vs Triton    |
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+|    8 |  4096 |  7168 | 34.6      | 13.6   | 852.3  | 2.98x faster | 2.14x faster |
+|    8 |  7168 | 18432 | 78.9      | 26.8   | 1677.3 | 0.90x slower | 1.13x faster |
+|    8 | 18432 |  7168 | 81.2      | 26.0   | 1631.1 | 0.85x slower | 1.39x faster |
+|   64 |  4096 |  7168 | 36.9      | 101.9  | 822.9  | 1.87x faster | 2.15x faster |
+|   64 |  7168 | 18432 | 87.4      | 193.4  | 1535.2 | 0.79x slower | 1.13x faster |
+|   64 | 18432 |  7168 | 85.0      | 199.0  | 1587.6 | 0.83x slower | 1.41x faster |
+|   64 | 24576 |  1536 | 28.0      | 172.8  | 1465.8 | 2.51x faster | 1.35x faster |
+|   64 | 32768 |   512 | 28.8      | 74.5   | 728.5  | 2.37x faster | 1.34x faster |
+|   64 |  7168 | 16384 | 77.9      | 193.0  | 1532.8 | 0.89x slower | 1.11x faster |
+|  128 |  4096 |  7168 | 39.1      | 192.4  | 802.0  | 3.61x faster | 2.32x faster |
+|  128 |  7168 | 18432 | 93.7      | 360.8  | 1454.2 | 0.77x slower | 1.54x faster |
+|  128 | 18432 |  7168 | 85.7      | 394.8  | 1608.0 | 0.81x slower | 2.68x faster |
+| 1024 |  4096 |  7168 | 99.7      | 603.1  | 452.2  | 0.90x slower | 2.43x faster |
+| 1024 | 18432 |  7168 | 331.3     | 816.7  | 534.9  | 0.84x slower | 2.71x faster |
+| 2048 |  4096 |  7168 | 198.3     | 606.6  | 306.7  | 0.88x slower | 2.34x faster |
+| 4096 |  4096 |  7168 | 392.2     | 613.2  | 235.3  | 0.86x slower | 2.30x faster |
++------+-------+-------+-----------+--------+--------+--------------+--------------+
+
+===== AVERAGE PERFORMANCE =====
++----------------+------------+----------+---------------+
+| Implementation | Avg TFLOPS | Avg GB/s | Avg Time (ms) |
++----------------+------------+----------+---------------+
+| DeepGEMM       | 310.98     | 1052.10  | 0.11          |
+| vLLM Triton    | 144.30     | 715.60   | 0.23          |
+| vLLM CUTLASS   | 286.78     | 1076.67  | 0.11          |
++----------------+------------+----------+---------------+
+
+===== AVERAGE SPEEDUPS =====
++-----------------------------+--------------+
+| Comparison                  | Speedup      |
++-----------------------------+--------------+
+| DeepGEMM vs vLLM Triton     | 1.71x faster |
+| DeepGEMM vs vLLM CUTLASS    | 0.94x slower |
+| vLLM CUTLASS vs vLLM Triton | 1.84x faster |
++-----------------------------+--------------+
+
+===== ACCURACY COMPARISON =====
++----------------+-----------------------+
+| Implementation | Avg Diff vs Reference |
++----------------+-----------------------+
+| DeepGEMM       | 0.000684              |
+| vLLM Triton    | 0.000684              |
+| vLLM CUTLASS   | 0.000684              |
++----------------+-----------------------+
+```

benchmarks/kernels/deepgemm/benchmark_fp8_block_dense_gemm.py

@@ -0,0 +1,464 @@
+# SPDX-License-Identifier: Apache-2.0
+# fmt: off
+# ruff: noqa: E501
+import time
+
+# Import DeepGEMM functions
+import deep_gemm
+import torch
+import triton
+from deep_gemm import calc_diff, ceil_div, get_col_major_tma_aligned_tensor
+
+# Import vLLM functions
+from vllm import _custom_ops as ops
+from vllm.model_executor.layers.quantization.utils.fp8_utils import (
+    per_token_group_quant_fp8, w8a8_block_fp8_matmul)
+
+
+# Copied from
+# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L9
+def per_token_cast_to_fp8(
+        x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    """Convert tensor to FP8 format with per-token scaling."""
+    assert x.dim() == 2 and x.size(1) % 128 == 0
+    m, n = x.shape
+    x_view = x.view(m, -1, 128)
+    x_amax = x_view.abs().float().amax(dim=2).view(m, -1).clamp(1e-4)
+    return (x_view * (448.0 / x_amax.unsqueeze(2))).to(
+        torch.float8_e4m3fn).view(m, n), (x_amax / 448.0).view(m, -1)
+
+
+# Copied from
+# https://github.com/deepseek-ai/DeepGEMM/blob/78cacf70d41d15d688bd493ebc85845f7f2a3d5d/tests/test_core.py#L17
+def per_block_cast_to_fp8(
+        x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+    """Convert tensor to FP8 format with per-block scaling."""
+    assert x.dim() == 2
+    m, n = x.shape
+    x_padded = torch.zeros((ceil_div(m, 128) * 128, ceil_div(n, 128) * 128),
+                           dtype=x.dtype,
+                           device=x.device)
+    x_padded[:m, :n] = x
+    x_view = x_padded.view(-1, 128, x_padded.size(1) // 128, 128)
+    x_amax = x_view.abs().float().amax(dim=(1, 3), keepdim=True).clamp(1e-4)
+    x_scaled = (x_view * (448.0 / x_amax)).to(torch.float8_e4m3fn)
+    return x_scaled.view_as(x_padded)[:m, :n].contiguous(), (
+        x_amax / 448.0).view(x_view.size(0), x_view.size(2))
+
+
+def benchmark_shape(m: int,
+                    n: int,
+                    k: int,
+                    warmup: int = 100,
+                    repeat: int = 10000,
+                    verbose: bool = False) -> dict:
+    """Benchmark all implementations for a specific (m, n, k) shape."""
+    if verbose:
+        print(f"\n=== Benchmarking shape: m={m}, n={n}, k={k} ===")
+
+    # Create test tensors
+    A = torch.randn((m, k), device='cuda', dtype=torch.bfloat16)
+    B = torch.randn((n, k), device='cuda', dtype=torch.bfloat16)
+
+    # Reference result in BF16
+    torch.cuda.synchronize()
+    C_ref = A @ B.t()
+
+    # Pre-quantize B for all implementations
+    # (weights can be pre-quantized offline)
+    B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B)
+    B_vllm, B_scale_vllm = per_block_cast_to_fp8(B)
+
+    # Block size configuration
+    block_size = [128, 128]
+
+    # Pre-quantize A for all implementations
+    A_deepgemm, A_scale_deepgemm = per_token_cast_to_fp8(A)
+    A_scale_deepgemm = get_col_major_tma_aligned_tensor(A_scale_deepgemm)
+    C_deepgemm = torch.empty((m, n), device='cuda', dtype=torch.bfloat16)
+    A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1])
+    A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8(
+        A, block_size[1], column_major_scales=True)
+
+    # === 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)
+        return C_deepgemm
+
+    # === 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,
+                                     B_scale_vllm,
+                                     block_size,
+                                     output_dtype=torch.bfloat16)
+
+    # === 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,
+                                     scale_b=B_scale_vllm.T,
+                                     out_dtype=torch.bfloat16)
+
+    # Run correctness check first
+    if verbose:
+        print("Running correctness check...")
+    C_deepgemm = deepgemm_gemm()
+    C_vllm_triton = vllm_triton_gemm()
+    C_vllm_cutlass = vllm_cutlass_gemm()
+
+    deepgemm_diff = calc_diff(C_deepgemm, C_ref)
+    vllm_triton_diff = calc_diff(C_vllm_triton, C_ref)
+    vllm_cutlass_diff = calc_diff(C_vllm_cutlass, C_ref)
+
+    if verbose:
+        print(f"DeepGEMM vs Reference difference: {deepgemm_diff:.6f}")
+        print(f"vLLM Triton vs Reference difference: {vllm_triton_diff:.6f}")
+        print(f"vLLM CUTLASS vs Reference difference: {vllm_cutlass_diff:.6f}")
+        print("vLLM Triton vs DeepGEMM difference: "
+              f"{calc_diff(C_vllm_triton, C_deepgemm):.6f}")
+        print("vLLM CUTLASS vs DeepGEMM difference: "
+              f"{calc_diff(C_vllm_cutlass, C_deepgemm):.6f}")
+
+    # Benchmark implementations
+    implementations = {
+        "DeepGEMM": deepgemm_gemm,
+        "vLLM Triton": vllm_triton_gemm,
+        "vLLM CUTLASS": vllm_cutlass_gemm
+    }
+
+    benchmark_results = {
+        "shape": {
+            "m": m,
+            "n": n,
+            "k": k
+        },
+        "implementations": {}
+    }
+
+    for name, func in implementations.items():
+        # Warmup
+        for _ in range(warmup):
+            func()
+            torch.cuda.synchronize()
+
+        # Timing loop
+        torch.cuda.synchronize()
+        start = time.time()
+        for _ in range(repeat):
+            func()
+        torch.cuda.synchronize()
+        end = time.time()
+
+        # Calculate timing and TFLOPS
+        avg_time_ms = (end - start) / repeat * 1000
+        avg_time_us = avg_time_ms * 1000
+        tflops = 2 * m * n * k / (avg_time_ms * 1e-3) / 1e12
+        gb_s = (m * k + k * n + m * n * 2) / 1e9 / (avg_time_ms * 1e-3)
+
+        benchmark_results["implementations"][name] = {
+            "time_ms": avg_time_ms,
+            "time_us": avg_time_us,
+            "tflops": tflops,
+            "gb_s": gb_s,
+            "diff": {
+                "DeepGEMM":
+                0.0 if name == "DeepGEMM" else calc_diff(func(), C_deepgemm),
+                "Reference":
+                deepgemm_diff if name == "DeepGEMM" else
+                (vllm_triton_diff
+                 if name == "vLLM Triton" else vllm_cutlass_diff)
+            }
+        }
+
+        if verbose:
+            print(
+                f"{name}: {avg_time_ms:.3f} ms, {tflops:.2f} TFLOPS, {gb_s:.2f} GB/s"
+            )
+
+    # Calculate speedups
+    baseline = benchmark_results["implementations"]["DeepGEMM"]["time_ms"]
+    for name, data in benchmark_results["implementations"].items():
+        if name != "DeepGEMM":
+            speedup = baseline / data["time_ms"]
+            benchmark_results["implementations"][name][
+                "speedup_vs_deepgemm"] = speedup
+            if verbose:
+                print(f"DeepGEMM is {1/speedup:.2f}x "
+                      f"{'faster' if 1/speedup > 1 else 'slower'} than {name}")
+
+    vllm_triton_time = benchmark_results["implementations"]["vLLM Triton"][
+        "time_ms"]
+    vllm_cutlass_time = benchmark_results["implementations"]["vLLM CUTLASS"][
+        "time_ms"]
+    cutlass_vs_triton = vllm_triton_time / vllm_cutlass_time
+    benchmark_results["implementations"]["vLLM CUTLASS"][
+        "speedup_vs_triton"] = cutlass_vs_triton
+    if verbose:
+        print(
+            f"vLLM CUTLASS is {cutlass_vs_triton:.2f}x "
+            f"{'faster' if cutlass_vs_triton > 1 else 'slower'} than vLLM Triton"
+        )
+
+    return benchmark_results
+
+
+def format_table_row(values, widths):
+    """Format a row with specified column widths."""
+    return "| " + " | ".join(f"{val:{w}}"
+                             for val, w in zip(values, widths)) + " |"
+
+
+def print_table(headers, rows, title=None):
+    """Print a table with headers and rows."""
+    if title:
+        print(f"\n{title}")
+
+    # Calculate column widths based on headers and data
+    widths = [
+        max(len(str(h)), max(len(str(row[i])) for row in rows))
+        for i, h in enumerate(headers)
+    ]
+
+    # Create separator line
+    separator = "+-" + "-+-".join("-" * w for w in widths) + "-+"
+
+    # Print table
+    print(separator)
+    print(format_table_row(headers, widths))
+    print(separator)
+    for row in rows:
+        print(format_table_row(row, widths))
+    print(separator)
+
+
+def format_speedup(value):
+    """Format speedup value with indicator if it's faster or slower."""
+    return f"{value:.2f}x {'faster' if value > 1.0 else 'slower'}"
+
+
+def run_benchmarks(verbose: bool = False):
+    """Run benchmarks for a set of common shapes."""
+    print("===== STARTING FP8 GEMM BENCHMARK =====")
+
+    # Make sure we're using the GPU
+    if not torch.cuda.is_available():
+        print("CUDA not available! Tests require GPU.")
+        return
+
+    # Print system information
+    print(f"PyTorch version: {torch.__version__}")
+    print(f"CUDA version: {torch.version.cuda}")
+    print(f"Triton version: {triton.__version__}")
+    print(f"Using device: {torch.cuda.get_device_name()}")
+
+    # Enable TF32 for better performance
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+
+    # Set seeds for reproducibility
+    torch.manual_seed(42)
+    torch.cuda.manual_seed(42)
+
+    # Define benchmark shapes (m, n, k)
+    shapes = [
+        (8, 4096, 7168),
+        (8, 7168, 18432),
+        (8, 18432, 7168),
+        (64, 4096, 7168),
+        (64, 7168, 18432),
+        (64, 18432, 7168),
+        (64, 24576, 1536),
+        (64, 32768, 512),
+        (64, 7168, 16384),
+        (128, 4096, 7168),
+        (128, 7168, 18432),
+        (128, 18432, 7168),
+        (1024, 4096, 7168),
+        (1024, 18432, 7168),
+        (2048, 4096, 7168),
+        (4096, 4096, 7168),
+    ]
+    shapes = [
+        # (64, 2112, 7168),
+        (64, 24576, 1536),
+        (64, 32768, 512),
+        (64, 7168, 16384),
+        (64, 4096, 7168),
+        (64, 7168, 2048),
+        # (128, 2112, 7168),
+        (128, 24576, 1536),
+        (128, 32768, 512),
+        (128, 7168, 16384),
+        (128, 4096, 7168),
+        (128, 7168, 2048),
+        # (4096, 2112, 7168),
+        (4096, 24576, 1536),
+        (4096, 32768, 512),
+        (4096, 7168, 16384),
+        (4096, 4096, 7168),
+        (4096, 7168, 2048),
+    ]
+
+    all_results = []
+    for m, n, k in shapes:
+        result = benchmark_shape(m, n, k, verbose=verbose)
+        all_results.append(result)
+
+    # Print results in a nicely formatted table
+    print("\n===== PERFORMANCE COMPARISON =====")
+
+    # Print DeepGEMM table
+    deepgemm_headers = ["m", "n", "k", "Time (μs)", "TFLOPS", "GB/s"]
+    deepgemm_rows = []
+    for result in all_results:
+        shape = result["shape"]
+        impl_data = result["implementations"]["DeepGEMM"]
+        deepgemm_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}"
+        ])
+
+    print_table(deepgemm_headers,
+                deepgemm_rows,
+                title="DeepGEMM Implementation:")
+
+    # Print vLLM Triton table
+    triton_headers = [
+        "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM"
+    ]
+    triton_rows = []
+    for result in all_results:
+        shape = result["shape"]
+        impl_data = result["implementations"]["vLLM Triton"]
+        speedup = impl_data.get("speedup_vs_deepgemm", 1.0)
+        triton_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}",
+            format_speedup(speedup)
+        ])
+
+    print_table(triton_headers,
+                triton_rows,
+                title="vLLM Triton Implementation:")
+
+    # Print vLLM CUTLASS table
+    cutlass_headers = [
+        "m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM",
+        "vs Triton"
+    ]
+    cutlass_rows = []
+    for result in all_results:
+        shape = result["shape"]
+        impl_data = result["implementations"]["vLLM CUTLASS"]
+        vs_deepgemm = impl_data.get("speedup_vs_deepgemm", 1.0)
+        vs_triton = impl_data.get("speedup_vs_triton", 1.0)
+        cutlass_rows.append([
+            shape["m"], shape["n"], shape["k"], f"{impl_data['time_us']:.1f}",
+            f"{impl_data['tflops']:.1f}", f"{impl_data['gb_s']:.1f}",
+            format_speedup(vs_deepgemm),
+            format_speedup(vs_triton)
+        ])
+
+    print_table(cutlass_headers,
+                cutlass_rows,
+                title="vLLM CUTLASS Implementation:")
+
+    # Calculate and print averages
+    print("\n===== AVERAGE PERFORMANCE =====")
+
+    implementations = ["DeepGEMM", "vLLM Triton", "vLLM CUTLASS"]
+    avg_metrics = {
+        impl: {
+            "tflops": 0,
+            "gb_s": 0,
+            "time_ms": 0
+        }
+        for impl in implementations
+    }
+
+    for result in all_results:
+        for impl in implementations:
+            impl_data = result["implementations"][impl]
+            avg_metrics[impl]["tflops"] += impl_data["tflops"]
+            avg_metrics[impl]["gb_s"] += impl_data["gb_s"]
+            avg_metrics[impl]["time_ms"] += impl_data["time_ms"]
+
+    num_shapes = len(all_results)
+    avg_headers = ["Implementation", "Avg TFLOPS", "Avg GB/s", "Avg Time (ms)"]
+    avg_rows = []
+
+    for impl in implementations:
+        avg_tflops = avg_metrics[impl]["tflops"] / num_shapes
+        avg_mem_bw = avg_metrics[impl]["gb_s"] / num_shapes
+        avg_time = avg_metrics[impl]["time_ms"] / num_shapes
+        avg_rows.append([
+            impl, f"{avg_tflops:.2f}", f"{avg_mem_bw:.2f}", f"{avg_time:.2f}"
+        ])
+
+    print_table(avg_headers, avg_rows)
+
+    # Calculate average speedups
+    avg_speedups = {
+        "DeepGEMM vs vLLM Triton": 0,
+        "DeepGEMM vs vLLM CUTLASS": 0,
+        "vLLM CUTLASS vs vLLM Triton": 0
+    }
+
+    for result in all_results:
+        deepgemm_time = result["implementations"]["DeepGEMM"]["time_ms"]
+        vllm_triton_time = result["implementations"]["vLLM Triton"]["time_ms"]
+        vllm_cutlass_time = result["implementations"]["vLLM CUTLASS"][
+            "time_ms"]
+
+        avg_speedups[
+            "DeepGEMM vs vLLM Triton"] += vllm_triton_time / deepgemm_time
+        avg_speedups[
+            "DeepGEMM vs vLLM CUTLASS"] += vllm_cutlass_time / deepgemm_time
+        avg_speedups[
+            "vLLM CUTLASS vs vLLM Triton"] += vllm_triton_time / vllm_cutlass_time
+
+    print("\n===== AVERAGE SPEEDUPS =====")
+    speedup_headers = ["Comparison", "Speedup"]
+    speedup_rows = []
+    for comparison, total in avg_speedups.items():
+        avg_speedup = total / num_shapes
+        status = "faster" if avg_speedup > 1 else "slower"
+        speedup_rows.append([comparison, f"{avg_speedup:.2f}x {status}"])
+
+    print_table(speedup_headers, speedup_rows)
+
+    # Average accuracy comparison
+    print("\n===== ACCURACY COMPARISON =====")
+    avg_diff = {impl: 0 for impl in implementations}
+
+    for result in all_results:
+        for impl in implementations:
+            avg_diff[impl] += result["implementations"][impl]["diff"][
+                "Reference"]
+
+    diff_headers = ["Implementation", "Avg Diff vs Reference"]
+    diff_rows = []
+    for impl in implementations:
+        diff_rows.append([impl, f"{avg_diff[impl] / num_shapes:.6f}"])
+
+    print_table(diff_headers, diff_rows)
+
+
+if __name__ == "__main__":
+    run_benchmarks(verbose=False)

benchmarks/kernels/graph_machete_bench.py

@@ -4,7 +4,6 @@ import math
 import pickle
 import re
 from collections import defaultdict
-from typing import List
 
 import matplotlib.pyplot as plt
 import pandas as pd
@@ -23,7 +22,7 @@ if __name__ == "__main__":
 
     with open(args.filename, 'rb') as f:
         data = pickle.load(f)
-        raw_results: List[TMeasurement] = data["results"]
+        raw_results: list[TMeasurement] = data["results"]
 
     results = defaultdict(lambda: list())
     for v in raw_results:

benchmarks/kernels/utils.py

@@ -1,7 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
 
 import dataclasses
-from typing import Any, Callable, Iterable, Optional
+from collections.abc import Iterable
+from typing import Any, Callable, Optional
 
 import torch
 import torch.utils.benchmark as TBenchmark

benchmarks/launch_tgi_server.sh

@@ -1,16 +0,0 @@
-#!/bin/bash
-
-PORT=8000
-MODEL=$1
-TOKENS=$2
-
-docker run -e "HF_TOKEN=$HF_TOKEN" --gpus all --shm-size 1g -p $PORT:80 \
-           -v "$PWD/data:/data" \
-           ghcr.io/huggingface/text-generation-inference:2.2.0 \
-           --model-id "$MODEL" \
-           --sharded false  \
-           --max-input-length 1024 \
-           --max-total-tokens 2048 \
-           --max-best-of 5 \
-           --max-concurrent-requests 5000 \
-           --max-batch-total-tokens "$TOKENS"

benchmarks/run_structured_output_benchmark.sh

@@ -0,0 +1,65 @@
+#!/bin/bash
+
+# Define the model to use
+MODEL=${1:-"Qwen/Qwen2.5-7B-Instruct"}
+
+# Define the backend to use
+BACKEND=${2:-"vllm"}
+
+# Define the dataset to use
+DATASET=${3:-"xgrammar_bench"}
+
+# Define the guided decoding backend
+GUIDED_BACKEND=${4:-"xgrammar"}
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+OUTPUT_DIR=${5:-"$SCRIPT_DIR/structured_output_benchmark_results"}
+
+GUIDED_RATIO=${6:-0.5}
+
+# Create output directory if it doesn't exist
+mkdir -p "$OUTPUT_DIR"
+
+# Define QPS values to test
+QPS_VALUES=(70 60 50 25 20 15 10)
+
+# Common parameters
+COMMON_PARAMS="--backend $BACKEND \
+               --model $MODEL \
+               --dataset $DATASET \
+               --structured-output-backend $GUIDED_BACKEND \
+               --structured-output-ratio $GUIDED_RATIO \
+               --save-results \
+               --result-dir $OUTPUT_DIR"
+
+echo "Starting structured output benchmark with model: $MODEL"
+echo "Backend: $BACKEND"
+echo "Dataset: $DATASET"
+echo "Structured output backend: $GUIDED_BACKEND"
+echo "Results will be saved to: $OUTPUT_DIR"
+echo "----------------------------------------"
+
+# Run benchmarks with different QPS values
+for qps in "${QPS_VALUES[@]}"; do
+  echo "Running benchmark with QPS: $qps"
+
+  # Get git hash and branch for the filename
+  GIT_HASH=$(git rev-parse --short HEAD 2>/dev/null || echo "unknown")
+  GIT_BRANCH=$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "unknown")
+
+  # Construct filename for this run
+  FILENAME="${GUIDED_BACKEND}_${BACKEND}_${qps}qps_$(basename $MODEL)_${DATASET}_${GIT_HASH}.json"
+
+  # Run the benchmark
+  python "$SCRIPT_DIR/benchmark_serving_structured_output.py" $COMMON_PARAMS \
+    --request-rate $qps \
+    --result-filename "$FILENAME" \
+    --tokenizer-mode ${TOKENIZER_MODE:-"auto"} \
+    --port ${PORT:-8000}
+
+  echo "Completed benchmark with QPS: $qps"
+  echo "----------------------------------------"
+done
+
+echo "All benchmarks completed!"
+echo "Results saved to: $OUTPUT_DIR"

benchmarks/structured_schemas/structured_schema_1.json

@@ -1,113 +1,19 @@
 {
-    "$schema":
-    "https://json-schema.org/draft/2020-12/schema",
-    "title":
-    "User Profile",
-    "type":
-    "object",
+    "type": "object",
     "properties": {
-        "userId": {
-            "type": "string",
-            "description": "Unique identifier for the user."
-        },
-        "personalInfo": {
-            "type": "object",
-            "properties": {
-                "firstName": {
-                    "type": "string",
-                    "description": "The user's first name."
-                },
-                "lastName": {
-                    "type": "string",
-                    "description": "The user's last name."
-                },
-                "age": {
-                    "type": "integer",
-                    "minimum": 0,
-                    "description": "The user's age."
-                },
-                "phoneNumbers": {
-                    "type":
-                    "array",
-                    "items": {
-                        "type": "object",
-                        "properties": {
-                            "type": {
-                                "type": "string",
-                                "enum": ["home", "work", "mobile"],
-                                "description": "Type of phone number."
-                            },
-                            "number": {
-                                "type": "string",
-                                "pattern": "^\\+?[1-9]\\d{1,14}$",
-                                "description": "Phone number in E.164 format."
-                            }
-                        },
-                        "required": ["type", "number"]
-                    },
-                    "description":
-                    "List of phone numbers associated with the user."
-                }
-            },
-            "required": ["firstName", "lastName"]
-        },
-        "address": {
-            "type": "object",
-            "properties": {
-                "street": {
-                    "type": "string",
-                    "description": "Street address."
-                },
-                "city": {
-                    "type": "string",
-                    "description": "City name."
-                },
-                "state": {
-                    "type": "string",
-                    "description": "State or province."
-                },
-                "postalCode": {
-                    "type": "string",
-                    "pattern": "^\\d{5}(-\\d{4})?$",
-                    "description": "Postal code."
-                },
-                "country": {
-                    "type": "string",
-                    "description": "Country name."
-                }
-            },
-            "required": ["street", "city", "state", "postalCode", "country"]
-        },
-        "preferences": {
-            "type": "object",
-            "properties": {
-                "newsletterSubscribed": {
-                    "type":
-                    "boolean",
-                    "description":
-                    "Indicates if the user is subscribed to the newsletter."
-                },
-                "favoriteCategories": {
-                    "type": "array",
-                    "items": {
-                        "type": "string"
-                    },
-                    "description": "List of user's favorite categories."
-                }
-            },
-            "required": ["newsletterSubscribed"]
-        },
-        "accountStatus": {
-            "type": "string",
-            "enum": ["active", "inactive", "suspended"],
-            "description": "Current status of the user's account."
-        },
-        "registrationDate": {
-            "type": "string",
-            "format": "date-time",
-            "description": "ISO 8601 formatted date-time of user registration."
-        }
+      "name": { "type": "string" },
+      "email": { "type": "string" },
+      "street": { "type": "string" },
+      "city": { "type": "string" },
+      "state": { "type": "string" },
+      "zip": { "type": "string" },
+      "phone": { "type": "string" },
+      "website": { "type": "string" },
+      "company": { "type": "string" },
+      "age": { "type": "integer" }
     },
-    "required":
-    ["userId", "personalInfo", "address", "accountStatus", "registrationDate"]
+    "required": [
+      "name",
+      "email"
+    ]
 }

cmake/cpu_extension.cmake

@@ -33,8 +33,6 @@ endif()
 
 if(MACOSX_FOUND)
     list(APPEND CXX_COMPILE_FLAGS
-        "-Xpreprocessor"
-        "-fopenmp"
         "-DVLLM_CPU_EXTENSION")
 else()
     list(APPEND CXX_COMPILE_FLAGS
@@ -81,6 +79,7 @@ else()
     find_isa(${CPUINFO} "POWER9" POWER9_FOUND)
     find_isa(${CPUINFO} "asimd" ASIMD_FOUND) # Check for ARM NEON support
     find_isa(${CPUINFO} "bf16" ARM_BF16_FOUND) # Check for ARM BF16 support
+    find_isa(${CPUINFO} "S390" S390_FOUND)
 endif()
 
 
@@ -129,8 +128,16 @@ elseif (ASIMD_FOUND)
 elseif(APPLE_SILICON_FOUND)
     message(STATUS "Apple Silicon Detected")
     set(ENABLE_NUMA OFF)
+elseif (S390_FOUND)
+    message(STATUS "S390 detected")
+    # Check for S390 VXE support
+    list(APPEND CXX_COMPILE_FLAGS
+        "-mvx"
+        "-mzvector"
+        "-march=native"
+        "-mtune=native")
 else()
-    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA or ARMv8 support.")
+    message(FATAL_ERROR "vLLM CPU backend requires AVX512, AVX2, Power9+ ISA, S390X ISA or ARMv8 support.")
 endif()
 
 #
@@ -140,7 +147,7 @@ if (AVX512_FOUND AND NOT AVX512_DISABLED)
     FetchContent_Declare(
         oneDNN
         GIT_REPOSITORY https://github.com/oneapi-src/oneDNN.git
-        GIT_TAG  v3.6
+        GIT_TAG  v3.7.1
         GIT_PROGRESS TRUE
         GIT_SHALLOW TRUE
     )
@@ -181,12 +188,14 @@ set(VLLM_EXT_SRC
     "csrc/cpu/cache.cpp"
     "csrc/cpu/utils.cpp"
     "csrc/cpu/layernorm.cpp"
+    "csrc/cpu/mla_decode.cpp"
     "csrc/cpu/pos_encoding.cpp"
     "csrc/cpu/torch_bindings.cpp")
 
 if (AVX512_FOUND AND NOT AVX512_DISABLED)
     set(VLLM_EXT_SRC
         "csrc/cpu/quant.cpp"
+        "csrc/cpu/shm.cpp"
         ${VLLM_EXT_SRC})
 endif()
 

cmake/external_projects/flashmla.cmake

@@ -0,0 +1,66 @@
+include(FetchContent)
+
+# If FLASH_MLA_SRC_DIR is set, flash-mla is installed from that directory 
+# instead of downloading.
+# It can be set as an environment variable or passed as a cmake argument.
+# The environment variable takes precedence.
+if (DEFINED ENV{FLASH_MLA_SRC_DIR})
+  set(FLASH_MLA_SRC_DIR $ENV{FLASH_MLA_SRC_DIR})
+endif()
+
+if(FLASH_MLA_SRC_DIR)
+  FetchContent_Declare(
+        flashmla 
+        SOURCE_DIR ${FLASH_MLA_SRC_DIR}
+        CONFIGURE_COMMAND ""
+        BUILD_COMMAND ""
+  )
+else()
+  FetchContent_Declare(
+        flashmla
+        GIT_REPOSITORY https://github.com/vllm-project/FlashMLA.git
+        GIT_TAG 575f7724b9762f265bbee5889df9c7d630801845
+        GIT_PROGRESS TRUE
+        CONFIGURE_COMMAND ""
+        BUILD_COMMAND ""
+  )
+endif()
+
+
+FetchContent_MakeAvailable(flashmla)
+message(STATUS "FlashMLA is available at ${flashmla_SOURCE_DIR}")
+
+# The FlashMLA kernels only work on hopper and require CUDA 12.3 or later.
+# Only build FlashMLA kernels if we are building for something compatible with 
+# sm90a
+cuda_archs_loose_intersection(FLASH_MLA_ARCHS "9.0a" "${CUDA_ARCHS}")
+if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.3 AND FLASH_MLA_ARCHS)
+    set(FlashMLA_SOURCES
+        ${flashmla_SOURCE_DIR}/csrc/flash_api.cpp
+        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_bf16_sm90.cu
+        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_fp16_sm90.cu
+        ${flashmla_SOURCE_DIR}/csrc/flash_fwd_mla_metadata.cu)
+
+    set(FlashMLA_INCLUDES
+        ${flashmla_SOURCE_DIR}/csrc/cutlass/include
+        ${flashmla_SOURCE_DIR}/csrc/include)
+
+    set_gencode_flags_for_srcs(
+        SRCS "${FlashMLA_SOURCES}"
+        CUDA_ARCHS "${FLASH_MLA_ARCHS}")
+
+    define_gpu_extension_target(
+        _flashmla_C
+        DESTINATION vllm
+        LANGUAGE ${VLLM_GPU_LANG}
+        SOURCES ${FlashMLA_SOURCES}
+        COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+        ARCHITECTURES ${VLLM_GPU_ARCHES}
+        INCLUDE_DIRECTORIES ${FlashMLA_INCLUDES}
+        USE_SABI 3
+        WITH_SOABI)
+else()
+    # Create an empty target for setup.py when not targeting sm90a systems
+    add_custom_target(_flashmla_C)
+endif()
+

cmake/external_projects/vllm_flash_attn.cmake

@@ -0,0 +1,67 @@
+# vLLM flash attention requires VLLM_GPU_ARCHES to contain the set of target
+# arches in the CMake syntax (75-real, 89-virtual, etc), since we clear the
+# arches in the CUDA case (and instead set the gencodes on a per file basis)
+# we need to manually set VLLM_GPU_ARCHES here.
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  foreach(_ARCH ${CUDA_ARCHS})
+    string(REPLACE "." "" _ARCH "${_ARCH}")
+    list(APPEND VLLM_GPU_ARCHES "${_ARCH}-real")
+  endforeach()
+endif()
+
+#
+# Build vLLM flash attention from source
+#
+# IMPORTANT: This has to be the last thing we do, because vllm-flash-attn uses the same macros/functions as vLLM.
+# Because functions all belong to the global scope, vllm-flash-attn's functions overwrite vLLMs.
+# They should be identical but if they aren't, this is a massive footgun.
+#
+# The vllm-flash-attn install rules are nested under vllm to make sure the library gets installed in the correct place.
+# To only install vllm-flash-attn, use --component _vllm_fa2_C (for FA2) or --component _vllm_fa3_C (for FA3).
+# If no component is specified, vllm-flash-attn is still installed.
+
+# If VLLM_FLASH_ATTN_SRC_DIR is set, vllm-flash-attn is installed from that directory instead of downloading.
+# This is to enable local development of vllm-flash-attn within vLLM.
+# It can be set as an environment variable or passed as a cmake argument.
+# The environment variable takes precedence.
+if (DEFINED ENV{VLLM_FLASH_ATTN_SRC_DIR})
+  set(VLLM_FLASH_ATTN_SRC_DIR $ENV{VLLM_FLASH_ATTN_SRC_DIR})
+endif()
+
+if(VLLM_FLASH_ATTN_SRC_DIR)
+  FetchContent_Declare(
+          vllm-flash-attn SOURCE_DIR 
+          ${VLLM_FLASH_ATTN_SRC_DIR}
+          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
+  )
+else()
+  FetchContent_Declare(
+          vllm-flash-attn
+          GIT_REPOSITORY https://github.com/vllm-project/flash-attention.git
+          GIT_TAG 0a721daebe4fa7149f06ecf3d3eabeb6dcd0f1fa
+          GIT_PROGRESS TRUE
+          # Don't share the vllm-flash-attn build between build types
+          BINARY_DIR ${CMAKE_BINARY_DIR}/vllm-flash-attn
+  )
+endif()
+
+
+# Fetch the vllm-flash-attn library
+FetchContent_MakeAvailable(vllm-flash-attn)
+message(STATUS "vllm-flash-attn is available at ${vllm-flash-attn_SOURCE_DIR}")
+
+# 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
+  COMPONENT _vllm_fa2_C
+  FILES_MATCHING PATTERN "*.py"
+)
+
+install(
+  DIRECTORY ${vllm-flash-attn_SOURCE_DIR}/vllm_flash_attn/
+  DESTINATION vllm_flash_attn
+  COMPONENT _vllm_fa3_C
+  FILES_MATCHING PATTERN "*.py"
+)

csrc/attention/merge_attn_states.cu

@@ -0,0 +1,178 @@
+#include <optional>
+#include <torch/all.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <algorithm>
+
+#include "attention_dtypes.h"
+#include "attention_utils.cuh"
+
+namespace vllm {
+
+// Implements section 2.2 of https://www.arxiv.org/pdf/2501.01005
+// can be used to combine partial attention results (in the split-KV case)
+template <typename scalar_t, const uint NUM_THREADS>
+__global__ void merge_attn_states_kernel(
+    scalar_t* output, float* output_lse, const scalar_t* prefix_output,
+    const float* prefix_lse, const scalar_t* suffix_output,
+    const float* suffix_lse, const uint num_tokens, const uint num_heads,
+    const uint head_size) {
+  using pack_128b_t = uint4;
+  const uint pack_size = 16 / sizeof(scalar_t);
+  const uint threads_per_head = head_size / pack_size;
+
+  const uint global_idx = blockIdx.x * NUM_THREADS + threadIdx.x;
+  const uint token_head_threads = num_tokens * num_heads * threads_per_head;
+
+  if (global_idx >= token_head_threads) return;
+
+  // global_idx -> token_idx + head_idx + pack_idx
+  const uint token_head_idx = global_idx / threads_per_head;
+  const uint pack_idx = global_idx % threads_per_head;
+
+  const uint token_idx = token_head_idx / num_heads;
+  const uint head_idx = token_head_idx % num_heads;
+
+  const uint pack_offset = pack_idx * pack_size;  // (0~15)*8, etc.
+  const uint head_offset =
+      token_idx * num_heads * head_size + head_idx * head_size;
+  const scalar_t* prefix_head_ptr = prefix_output + head_offset;
+  const scalar_t* suffix_head_ptr = suffix_output + head_offset;
+  scalar_t* output_head_ptr = output + head_offset;
+
+  float p_lse = prefix_lse[head_idx * num_tokens + token_idx];
+  float s_lse = suffix_lse[head_idx * num_tokens + token_idx];
+  p_lse = std::isinf(p_lse) ? -std::numeric_limits<float>::infinity() : p_lse;
+  s_lse = std::isinf(s_lse) ? -std::numeric_limits<float>::infinity() : s_lse;
+
+  const float max_lse = fmaxf(p_lse, s_lse);
+  p_lse = p_lse - max_lse;
+  s_lse = s_lse - max_lse;
+  const float p_se = expf(p_lse);
+  const float s_se = expf(s_lse);
+  const float out_se = p_se + s_se;
+  const float p_scale = p_se / out_se;
+  const float s_scale = s_se / out_se;
+
+  if (pack_offset < head_size) {
+    // Pack 128b load
+    pack_128b_t p_out_pack = reinterpret_cast<const pack_128b_t*>(
+        prefix_head_ptr)[pack_offset / pack_size];
+    pack_128b_t s_out_pack = reinterpret_cast<const pack_128b_t*>(
+        suffix_head_ptr)[pack_offset / pack_size];
+    pack_128b_t o_out_pack;
+
+#pragma unroll
+    for (uint i = 0; i < pack_size; ++i) {
+      // Always use float for FMA to keep high precision.
+      // half(uint16_t), bfloat16, float -> float.
+      const float p_out_f =
+          vllm::to_float(reinterpret_cast<const scalar_t*>(&p_out_pack)[i]);
+      const float s_out_f =
+          vllm::to_float(reinterpret_cast<const scalar_t*>(&s_out_pack)[i]);
+      // fma: a * b + c = p_out_f * p_scale + (s_out_f * s_scale)
+      const float o_out_f = p_out_f * p_scale + (s_out_f * s_scale);
+      // float -> half(uint16_t), bfloat16, float.
+      vllm::from_float(reinterpret_cast<scalar_t*>(&o_out_pack)[i], o_out_f);
+    }
+
+    // Pack 128b storage
+    reinterpret_cast<pack_128b_t*>(output_head_ptr)[pack_offset / pack_size] =
+        o_out_pack;
+  }
+  // We only need to write to output_lse once per head.
+  if (output_lse != nullptr && pack_idx == 0) {
+    float out_lse = logf(out_se) + max_lse;
+    output_lse[head_idx * num_tokens + token_idx] = out_lse;
+  }
+}
+
+}  // namespace vllm
+
+// The following macro is used to dispatch the conversion function based on
+// the output data type. The FN is a macro that calls a function with
+// template<typename scalar_t>.
+#define DISPATCH_BY_SCALAR_DTYPE(scalar_dtype, fn)                      \
+  {                                                                     \
+    if (scalar_dtype == at::ScalarType::Float) {                        \
+      fn(float);                                                        \
+    } else if (scalar_dtype == at::ScalarType::Half) {                  \
+      fn(uint16_t);                                                     \
+    } else if (scalar_dtype == at::ScalarType::BFloat16) {              \
+      fn(__nv_bfloat16);                                                \
+    } else {                                                            \
+      TORCH_CHECK(false, "Unsupported data type of O: ", scalar_dtype); \
+    }                                                                   \
+  }
+
+#define LAUNCH_MERGE_ATTN_STATES(scalar_t, NUM_THREADS)                     \
+  {                                                                         \
+    vllm::merge_attn_states_kernel<scalar_t, NUM_THREADS>                   \
+        <<<grid, block, 0, stream>>>(                                       \
+            reinterpret_cast<scalar_t*>(output.data_ptr()), output_lse_ptr, \
+            reinterpret_cast<scalar_t*>(prefix_output.data_ptr()),          \
+            reinterpret_cast<float*>(prefix_lse.data_ptr()),                \
+            reinterpret_cast<scalar_t*>(suffix_output.data_ptr()),          \
+            reinterpret_cast<float*>(suffix_lse.data_ptr()), num_tokens,    \
+            num_heads, head_size);                                          \
+  }
+
+/*@brief Merges the attention states from prefix and suffix
+ * into the output tensor. NUM_TOKENS: n, NUM_HEADS: h, HEAD_SIZE: d
+ *
+ * @param output [n,h,d] The output tensor to store the merged attention states.
+ * @param output_lse [h,d] Optional tensor to store the log-sum-exp values.
+ * @param prefix_output [n,h,d] The prefix attention states.
+ * @param prefix_lse [h,n] The log-sum-exp values for the prefix attention
+ * states.
+ * @param suffix_output [n,h,d] The suffix attention states.
+ * @param suffix_lse [h,n] The log-sum-exp values for the suffix attention
+ * states.
+ */
+template <typename scalar_t>
+void merge_attn_states_launcher(torch::Tensor& output,
+                                std::optional<torch::Tensor> output_lse,
+                                const torch::Tensor& prefix_output,
+                                const torch::Tensor& prefix_lse,
+                                const torch::Tensor& suffix_output,
+                                const torch::Tensor& suffix_lse) {
+  constexpr uint NUM_THREADS = 128;
+  const uint num_tokens = output.size(0);
+  const uint num_heads = output.size(1);
+  const uint head_size = output.size(2);
+  const uint pack_size = 16 / sizeof(scalar_t);
+  TORCH_CHECK(head_size % pack_size == 0,
+              "headsize must be multiple of pack_size:", pack_size);
+  float* output_lse_ptr = nullptr;
+  if (output_lse.has_value()) {
+    output_lse_ptr = output_lse.value().data_ptr<float>();
+  }
+  // Process one pack elements per thread. for float, the
+  // pack_size is 4 for half/bf16, the pack_size is 8.
+  const uint threads_per_head = head_size / pack_size;
+  const uint total_threads = num_tokens * num_heads * threads_per_head;
+
+  dim3 block(NUM_THREADS);
+  dim3 grid((total_threads + NUM_THREADS - 1) / NUM_THREADS);
+
+  const c10::cuda::OptionalCUDAGuard device_guard(prefix_output.device());
+  auto stream = at::cuda::getCurrentCUDAStream();
+
+  LAUNCH_MERGE_ATTN_STATES(scalar_t, NUM_THREADS);
+}
+
+#define CALL_MERGE_ATTN_STATES_LAUNCHER(scalar_t)                           \
+  {                                                                         \
+    merge_attn_states_launcher<scalar_t>(output, output_lse, prefix_output, \
+                                         prefix_lse, suffix_output,         \
+                                         suffix_lse);                       \
+  }
+
+void merge_attn_states(torch::Tensor& output,
+                       std::optional<torch::Tensor> output_lse,
+                       const torch::Tensor& prefix_output,
+                       const torch::Tensor& prefix_lse,
+                       const torch::Tensor& suffix_output,
+                       const torch::Tensor& suffix_lse) {
+  DISPATCH_BY_SCALAR_DTYPE(output.dtype(), CALL_MERGE_ATTN_STATES_LAUNCHER);
+}

csrc/cache.h

@@ -39,3 +39,10 @@ void concat_and_cache_mla(torch::Tensor& kv_c, torch::Tensor& k_pe,
 // Just for unittest
 void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
                  const double scale, const std::string& kv_cache_dtype);
+
+void gather_cache(
+    torch::Tensor const& src_cache,    // [NUM_BLOCKS, BLOCK_SIZE, ENTRIES...]
+    torch::Tensor const& dst,          // [TOT_TOKENS, ENTRIES...]
+    torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
+    torch::Tensor const& cu_seq_lens,  // [BATCH+1]
+    int64_t batch_size, std::optional<torch::Tensor> seq_starts = std::nullopt);

csrc/cache_kernels.cu

@@ -2,6 +2,7 @@
 #include <ATen/cuda/CUDAContext.h>
 #include <c10/cuda/CUDAGuard.h>
 
+#include "cuda_utils.h"
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
 
@@ -349,8 +350,8 @@ __global__ void concat_and_cache_mla_kernel(
 
 }  // namespace vllm
 
-// KV_T is the stored data type of kv-cache.
-// CACHE_T is the data type of key and value tensors.
+// KV_T is the data type of key and value tensors.
+// CACHE_T is the stored data type of kv-cache.
 // KV_DTYPE is the real data type of kv-cache.
 #define CALL_RESHAPE_AND_CACHE(KV_T, CACHE_T, KV_DTYPE)               \
   vllm::reshape_and_cache_kernel<KV_T, CACHE_T, KV_DTYPE>             \
@@ -374,7 +375,7 @@ void reshape_and_cache(
     torch::Tensor& slot_mapping,  // [num_tokens]
     const std::string& kv_cache_dtype, torch::Tensor& k_scale,
     torch::Tensor& v_scale) {
-  int num_tokens = key.size(0);
+  int num_tokens = slot_mapping.size(0);
   int num_heads = key.size(1);
   int head_size = key.size(2);
   int block_size = key_cache.size(3);
@@ -392,8 +393,8 @@ void reshape_and_cache(
                              CALL_RESHAPE_AND_CACHE)
 }
 
-// KV_T is the stored data type of kv-cache.
-// CACHE_T is the data type of key and value tensors.
+// KV_T is the data type of key and value tensors.
+// CACHE_T is the stored data type of kv-cache.
 // KV_DTYPE is the real data type of kv-cache.
 #define CALL_RESHAPE_AND_CACHE_FLASH(KV_T, CACHE_T, KV_DTYPE)         \
   vllm::reshape_and_cache_flash_kernel<KV_T, CACHE_T, KV_DTYPE>       \
@@ -445,8 +446,8 @@ void reshape_and_cache_flash(
                              CALL_RESHAPE_AND_CACHE_FLASH);
 }
 
-// KV_T is the stored data type of kv-cache.
-// CACHE_T is the data type of key and value tensors.
+// KV_T is the data type of key and value tensors.
+// CACHE_T is the stored data type of kv-cache.
 // KV_DTYPE is the real data type of kv-cache.
 #define CALL_CONCAT_AND_CACHE_MLA(KV_T, CACHE_T, KV_DTYPE)              \
   vllm::concat_and_cache_mla_kernel<KV_T, CACHE_T, KV_DTYPE>            \
@@ -570,3 +571,161 @@ void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
     TORCH_CHECK(false, "Unsupported data type: ", kv_cache_dtype);
   }
 }
+
+namespace vllm {
+
+// grid is launched with dimensions (batch, num_splits)
+template <typename scalar_t>
+__global__ void gather_cache(
+    const scalar_t* __restrict__ src_cache,   // [NUM_BLOCKS, BLOCK_SIZE,
+                                              // ENTRIES...]
+    scalar_t* __restrict__ dst,               // [TOT_TOKENS, ENTRIES...]
+    const int32_t* __restrict__ block_table,  // [BATCH, BLOCK_INDICES]
+    const int32_t* __restrict__ cu_seq_lens,  // [BATCH+1]
+    const int32_t block_size, const int32_t entry_size,
+    const int64_t block_table_stride, const int64_t cache_block_stride,
+    const int64_t cache_entry_stride, const int64_t dst_entry_stride,
+    const int32_t* __restrict__ seq_starts) {  // Optional: starting offsets per
+                                               // batch
+
+  const int64_t bid = blockIdx.x;  // Batch ID
+  const int32_t num_splits = gridDim.y;
+  const int32_t split = blockIdx.y;
+  const int32_t seq_start = cu_seq_lens[bid];
+  const int32_t seq_end = cu_seq_lens[bid + 1];
+  const int32_t seq_len = seq_end - seq_start;
+  const int32_t tot_blocks = cuda_utils::ceil_div(seq_len, block_size);
+  const int32_t split_blocks = cuda_utils::ceil_div(tot_blocks, num_splits);
+
+  const int32_t split_start = split * split_blocks;
+  const int32_t split_end = min((split + 1) * split_blocks, tot_blocks);
+
+  const bool is_active_split = (split_start < tot_blocks);
+  const bool is_last_split = (split_end == tot_blocks);
+
+  if (!is_active_split) return;
+
+  int32_t full_blocks_end = split_end;
+  int32_t partial_block_size = 0;
+
+  // Adjust the pointer for the block_table for this batch.
+  // If seq_starts is provided, compute an offset based on (seq_starts[bid] /
+  // page_size)
+  const int32_t batch_offset = bid * block_table_stride;
+  int32_t offset = 0;
+  if (seq_starts != nullptr) {
+    offset = seq_starts[bid] / block_size;
+  }
+  const int32_t* batch_block_table = block_table + batch_offset + offset;
+
+  // Adjust dst pointer based on the cumulative sequence lengths.
+  dst += seq_start * dst_entry_stride;
+
+  if (is_last_split) {
+    partial_block_size = seq_len % block_size;
+    if (partial_block_size) full_blocks_end -= 1;
+  }
+
+  auto copy_entry = [&](const scalar_t* __restrict__ _src,
+                        scalar_t* __restrict__ _dst) {
+    for (int i = threadIdx.x; i < entry_size; i += blockDim.x)
+      _dst[i] = _src[i];
+  };
+
+  for (int pid = split_start; pid < full_blocks_end; ++pid) {
+    auto block_id = batch_block_table[pid];
+    auto block_start_ptr = src_cache + block_id * cache_block_stride;
+    auto block_dst_ptr = dst + pid * block_size * dst_entry_stride;
+    for (int eid = 0; eid < block_size; ++eid) {
+      copy_entry(block_start_ptr + eid * cache_entry_stride,
+                 block_dst_ptr + eid * dst_entry_stride);
+    }
+  }
+
+  if (partial_block_size) {
+    auto block_id = batch_block_table[full_blocks_end];
+    auto block_start_ptr = src_cache + block_id * cache_block_stride;
+    auto block_dst_ptr = dst + full_blocks_end * block_size * dst_entry_stride;
+    for (int eid = 0; eid < partial_block_size; ++eid) {
+      copy_entry(block_start_ptr + eid * cache_entry_stride,
+                 block_dst_ptr + eid * dst_entry_stride);
+    }
+  }
+}
+
+}  // namespace vllm
+
+// Macro to dispatch the kernel based on the data type.
+#define CALL_GATHER_CACHE(CPY_DTYPE)                                    \
+  vllm::gather_cache<CPY_DTYPE><<<grid, block, 0, stream>>>(            \
+      reinterpret_cast<CPY_DTYPE*>(src_cache.data_ptr()),               \
+      reinterpret_cast<CPY_DTYPE*>(dst.data_ptr()),                     \
+      block_table.data_ptr<int32_t>(), cu_seq_lens.data_ptr<int32_t>(), \
+      block_size, entry_size, block_table_stride, cache_block_stride,   \
+      cache_entry_stride, dst_entry_stride, seq_starts_ptr);
+
+// Gather sequences from the cache into the destination tensor.
+//  - cu_seq_lens contains the cumulative sequence lengths for each batch
+//  - block_table contains the cache block indices for each sequence
+//  - Optionally, seq_starts (if provided) offsets the starting block index by
+//  (seq_starts[bid] / page_size)
+void gather_cache(
+    torch::Tensor const& src_cache,    // [NUM_BLOCKS, BLOCK_SIZE, ENTRIES...]
+    torch::Tensor const& dst,          // [TOT_TOKENS, ENTRIES...]
+    torch::Tensor const& block_table,  // [BATCH, BLOCK_INDICES]
+    torch::Tensor const& cu_seq_lens,  // [BATCH+1]
+    int64_t batch_size,
+    std::optional<torch::Tensor> seq_starts = std::nullopt) {
+  at::cuda::OptionalCUDAGuard device_guard(src_cache.device());
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  int32_t block_size = src_cache.size(1);
+  int32_t entry_size = src_cache.flatten(2, -1).size(2);
+
+  TORCH_CHECK(block_table.dtype() == torch::kInt32,
+              "block_table must be int32");
+  TORCH_CHECK(cu_seq_lens.dtype() == torch::kInt32,
+              "cu_seq_lens must be int32");
+  if (seq_starts.has_value()) {
+    TORCH_CHECK(seq_starts.value().dtype() == torch::kInt32,
+                "seq_starts must be int32");
+  }
+
+  TORCH_CHECK(src_cache.device() == dst.device(),
+              "src_cache and dst must be on the same device");
+  TORCH_CHECK(src_cache.device() == block_table.device(),
+              "src_cache and block_table must be on the same device");
+  TORCH_CHECK(src_cache.device() == cu_seq_lens.device(),
+              "src_cache and cu_seq_lens must be on the same device");
+  if (seq_starts.has_value()) {
+    TORCH_CHECK(src_cache.device() == seq_starts.value().device(),
+                "src_cache and seq_starts must be on the same device");
+  }
+
+  int64_t block_table_stride = block_table.stride(0);
+  int64_t cache_block_stride = src_cache.stride(0);
+  int64_t cache_entry_stride = src_cache.stride(1);
+  int64_t dst_entry_stride = dst.stride(0);
+
+  // Decide on the number of splits based on the batch size.
+  int num_splits = batch_size > 128 ? 2 : batch_size > 64 ? 4 : 16;
+  dim3 grid(batch_size, num_splits);
+  dim3 block(1024);
+
+  TORCH_CHECK(src_cache.dtype() == dst.dtype(),
+              "src_cache and dst must have the same dtype");
+
+  const int dtype_bits = src_cache.element_size() * 8;
+  const int32_t* seq_starts_ptr =
+      seq_starts.has_value() ? seq_starts.value().data_ptr<int32_t>() : nullptr;
+
+  if (dtype_bits == 32) {
+    CALL_GATHER_CACHE(uint32_t);
+  } else if (dtype_bits == 16) {
+    CALL_GATHER_CACHE(uint16_t);
+  } else if (dtype_bits == 8) {
+    CALL_GATHER_CACHE(uint8_t);
+  } else {
+    TORCH_CHECK(false, "Unsupported data type width: ", dtype_bits);
+  }
+}

csrc/core/math.hpp

@@ -7,8 +7,3 @@ inline constexpr uint32_t next_pow_2(uint32_t const num) {
   if (num <= 1) return num;
   return 1 << (CHAR_BIT * sizeof(num) - __builtin_clz(num - 1));
 }
-
-template <typename T>
-inline constexpr std::enable_if_t<std::is_integral_v<T>, T> ceil_div(T a, T b) {
-  return (a + b - 1) / b;
-}

csrc/cpu/attention.cpp

@@ -24,8 +24,8 @@ struct KernelVecType<float> {
 
 template <>
 struct KernelVecType<c10::Half> {
-#ifdef __powerpc64__
-  // Power architecture-specific vector types
+#if defined(__powerpc64__) || defined(__s390x__)
+  // Power and s390x architecture-specific vector types
   using q_load_vec_type = vec_op::FP32Vec8;
   using k_load_vec_type = vec_op::FP32Vec16;
   using v_load_vec_type = vec_op::FP32Vec16;

csrc/cpu/cache.cpp

@@ -3,6 +3,12 @@
 
 #include "cpu_types.hpp"
 
+#if defined(__x86_64__)
+  #define DISPATCH_MACRO VLLM_DISPATCH_FLOATING_TYPES_WITH_E5M2
+#else
+  #define DISPATCH_MACRO VLLM_DISPATCH_FLOATING_TYPES
+#endif
+
 namespace {
 template <typename scalar_t>
 void copy_blocks_cpu_impl(std::vector<torch::Tensor> const& key_caches,
@@ -82,6 +88,48 @@ void reshape_and_cache_cpu_impl(
 }
 };  // namespace
 
+template <typename scalar_t>
+void concat_and_cache_mla_cpu_impl(
+    const scalar_t* __restrict__ kv_c,  // [num_tokens, kv_lora_rank]
+    const scalar_t* __restrict__ k_pe,  // [num_tokens, pe_dim]
+    scalar_t* __restrict__ kv_cache,  // [num_blocks, block_size, (kv_lora_rank
+                                      // + pe_dim)]
+    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
+    const int num_tokens,                      //
+    const int block_stride,                    //
+    const int entry_stride,                    //
+    const int kv_c_stride,                     //
+    const int k_pe_stride,                     //
+    const int kv_lora_rank,                    //
+    const int pe_dim,                          //
+    const int block_size                       //
+) {
+#pragma omp parallel for
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    const int64_t slot_idx = slot_mapping[token_idx];
+    // NOTE: slot_idx can be -1 if the token is padded
+    if (slot_idx < 0) {
+      continue;
+    }
+    const int64_t block_idx = slot_idx / block_size;
+    const int64_t block_offset = slot_idx % block_size;
+
+    auto copy = [&](const scalar_t* __restrict__ src,
+                    scalar_t* __restrict__ dst, int src_stride, int dst_stride,
+                    int size, int offset) {
+      for (int i = 0; i < size; i++) {
+        const int64_t src_idx = token_idx * src_stride + i;
+        const int64_t dst_idx =
+            block_idx * block_stride + block_offset * entry_stride + i + offset;
+        dst[dst_idx] = src[src_idx];
+      }
+    };
+
+    copy(kv_c, kv_cache, kv_c_stride, block_stride, kv_lora_rank, 0);
+    copy(k_pe, kv_cache, k_pe_stride, block_stride, pe_dim, kv_lora_rank);
+  }
+}
+
 // Note: the key_caches and value_caches vectors are constant but
 // not the Tensors they contain. The vectors need to be const refs
 // in order to satisfy pytorch's C++ operator registration code.
@@ -95,13 +143,12 @@ void copy_blocks(std::vector<torch::Tensor> const& key_caches,
   }
 
   const int element_num_per_block = key_caches[0][0].numel();
-  VLLM_DISPATCH_FLOATING_TYPES(
-      key_caches[0].scalar_type(), "copy_blocks_cpu_impl", [&] {
-        CPU_KERNEL_GUARD_IN(copy_blocks_cpu_impl)
-        copy_blocks_cpu_impl<scalar_t>(key_caches, value_caches, block_mapping,
-                                       element_num_per_block, num_layers);
-        CPU_KERNEL_GUARD_OUT(copy_blocks_cpu_impl)
-      });
+  DISPATCH_MACRO(key_caches[0].scalar_type(), "copy_blocks_cpu_impl", [&] {
+    CPU_KERNEL_GUARD_IN(copy_blocks_cpu_impl)
+    copy_blocks_cpu_impl<scalar_t>(key_caches, value_caches, block_mapping,
+                                   element_num_per_block, num_layers);
+    CPU_KERNEL_GUARD_OUT(copy_blocks_cpu_impl)
+  });
 }
 
 void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
@@ -118,15 +165,46 @@ void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
   int key_stride = key.stride(0);
   int value_stride = value.stride(0);
 
+  DISPATCH_MACRO(key.scalar_type(), "reshape_and_cache_cpu_impl", [&] {
+    CPU_KERNEL_GUARD_IN(reshape_and_cache_cpu_impl)
+    reshape_and_cache_cpu_impl<scalar_t>(
+        key.data_ptr<scalar_t>(), value.data_ptr<scalar_t>(),
+        key_cache.data_ptr<scalar_t>(), value_cache.data_ptr<scalar_t>(),
+        slot_mapping.data_ptr<int64_t>(), num_tokens, key_stride, value_stride,
+        num_heads, head_size, block_size, x);
+    CPU_KERNEL_GUARD_OUT(reshape_and_cache_cpu_impl)
+  });
+}
+
+void concat_and_cache_mla(
+    torch::Tensor& kv_c,          // [num_tokens, kv_lora_rank]
+    torch::Tensor& k_pe,          // [num_tokens, pe_dim]
+    torch::Tensor& kv_cache,      // [num_blocks, block_size, (kv_lora_rank +
+                                  // pe_dim)]
+    torch::Tensor& slot_mapping,  // [num_tokens] or [num_actual_tokens]
+    const std::string& kv_cache_dtype, torch::Tensor& scale) {
+  int num_tokens = slot_mapping.size(0);
+  int kv_lora_rank = kv_c.size(1);
+  int pe_dim = k_pe.size(1);
+  int block_size = kv_cache.size(1);
+
+  TORCH_CHECK(kv_cache.size(2) == kv_lora_rank + pe_dim);
+  TORCH_CHECK(kv_cache_dtype != "fp8");
+
+  int kv_c_stride = kv_c.stride(0);
+  int k_pe_stride = k_pe.stride(0);
+  int block_stride = kv_cache.stride(0);
+  int entry_stride = kv_cache.stride(1);
+
   VLLM_DISPATCH_FLOATING_TYPES(
-      key.scalar_type(), "reshape_and_cache_cpu_impl", [&] {
-        CPU_KERNEL_GUARD_IN(reshape_and_cache_cpu_impl)
-        reshape_and_cache_cpu_impl<scalar_t>(
-            key.data_ptr<scalar_t>(), value.data_ptr<scalar_t>(),
-            key_cache.data_ptr<scalar_t>(), value_cache.data_ptr<scalar_t>(),
-            slot_mapping.data_ptr<int64_t>(), num_tokens, key_stride,
-            value_stride, num_heads, head_size, block_size, x);
-        CPU_KERNEL_GUARD_OUT(reshape_and_cache_cpu_impl)
+      kv_c.scalar_type(), "concat_and_cache_mla_cpu_impl", [&] {
+        CPU_KERNEL_GUARD_IN(concat_and_cache_mla_cpu_impl)
+        concat_and_cache_mla_cpu_impl<scalar_t>(
+            kv_c.data_ptr<scalar_t>(), k_pe.data_ptr<scalar_t>(),
+            kv_cache.data_ptr<scalar_t>(), slot_mapping.data_ptr<int64_t>(),
+            num_tokens, block_stride, entry_stride, kv_c_stride, k_pe_stride,
+            kv_lora_rank, pe_dim, block_size);
+        CPU_KERNEL_GUARD_OUT(concat_and_cache_mla_cpu_impl)
       });
 }
 

csrc/cpu/cpu_types.hpp

@@ -7,6 +7,9 @@
 #elif defined(__POWER9_VECTOR__)
   // ppc implementation
   #include "cpu_types_vsx.hpp"
+#elif defined(__s390x__)
+  // s390 implementation
+  #include "cpu_types_vxe.hpp"
 #elif defined(__aarch64__)
   // arm implementation
   #include "cpu_types_arm.hpp"

csrc/cpu/cpu_types_arm.hpp

@@ -2,6 +2,10 @@
 #include <torch/all.h>
 #include <cmath>
 
+#if defined(__APPLE__)
+  #include "omp.h"
+#endif
+
 namespace vec_op {
 
 #ifdef ARM_BF16_SUPPORT

csrc/cpu/cpu_types_vxe.hpp

@@ -0,0 +1,480 @@
+
+#ifndef CPU_TYPES_VXE_HPP
+#define CPU_TYPES_VXE_HPP
+
+#include <vecintrin.h>
+#include <cmath>
+#include <torch/all.h>
+namespace vec_op {
+
+#define vec_neg(a) (-(a))
+#define vec_add(a, b) ((a) + (b))
+#define vec_sub(a, b) ((a) - (b))
+#define vec_mul(a, b) ((a) * (b))
+#define vec_div(a, b) ((a) / (b))
+#define vec_sr(a, b) ((a) >> (b))  // Vector Shift Right Algebaic
+#define vec_sl(a, b) ((a) << (b))  // Vector Shift Left
+
+// FIXME: FP16 is not fully supported in Torch-CPU
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)
+
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
+
+#ifndef CPU_OP_GUARD
+  #define CPU_KERNEL_GUARD_IN(NAME)
+  #define CPU_KERNEL_GUARD_OUT(NAME)
+#else
+  #define CPU_KERNEL_GUARD_IN(NAME) \
+    std::cout << #NAME << " invoked." << std::endl;
+  #define CPU_KERNEL_GUARD_OUT(NAME) \
+    std::cout << #NAME << " exit." << std::endl;
+#endif
+
+#define FORCE_INLINE __attribute__((always_inline)) inline
+
+namespace {
+template <typename T, T... indexes, typename F>
+constexpr void unroll_loop_item(std::integer_sequence<T, indexes...>, F&& f) {
+  (f(std::integral_constant<T, indexes>{}), ...);
+}
+};  // namespace
+
+template <typename T, T count, typename F,
+          typename = std::enable_if_t<std::is_invocable_v<F, T>>>
+constexpr void unroll_loop(F&& f) {
+  unroll_loop_item(std::make_integer_sequence<T, count>{}, std::forward<F>(f));
+}
+
+template <typename T>
+struct Vec {
+  constexpr static int get_elem_num() { return T::VEC_ELEM_NUM; }
+};
+
+typedef struct ss16x8x2_t {
+  __vector signed short val[2];
+} ss16x8x2_t;
+
+typedef struct ss16x8x4_t {
+  __vector signed short val[4];
+} ss16x8x4_t;
+
+typedef struct f32x4x2_t {
+  __vector float val[2];
+} f32x4x2_t;
+
+typedef struct f32x4x4_t {
+  __vector float val[4];
+} f32x4x4_t;
+
+struct FP32Vec8;
+struct FP32Vec16;
+
+struct BF16Vec8 : public Vec<BF16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+
+  __vector signed short reg;
+
+  explicit BF16Vec8(const void* ptr) : reg(*(__vector signed short*)ptr) {}
+  explicit BF16Vec8(const FP32Vec8&);
+
+  void save(void* ptr) const {
+    *reinterpret_cast<__vector signed short*>(ptr) = reg;
+  }
+};
+
+struct BF16Vec16 : public Vec<BF16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+
+  ss16x8x2_t reg;
+
+  explicit BF16Vec16(const void* ptr) {
+    // Load 256 bits in two parts
+    reg.val[0] = (__vector signed short)vec_xl(0, (signed short*)ptr);
+    reg.val[1] = (__vector signed short)vec_xl(16, (signed short*)ptr);
+  }
+
+  explicit BF16Vec16(const FP32Vec16&);
+
+  void save(void* ptr) const {
+    // Save 256 bits in two parts
+    vec_xst(reg.val[0], 0, (signed short*)ptr);
+    vec_xst(reg.val[1], 16, (signed short*)ptr);
+  }
+};
+
+const static __vector signed short zero = vec_splats((signed short)0);
+
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+
+  ss16x8x4_t reg;
+  explicit BF16Vec32(const void* ptr)
+      : reg(*reinterpret_cast<const ss16x8x4_t*>(ptr)) {}
+
+  explicit BF16Vec32(ss16x8x4_t data) : reg(data) {}
+
+  explicit BF16Vec32(const BF16Vec8& vec8_data)
+      : reg({vec8_data.reg, vec8_data.reg, vec8_data.reg, vec8_data.reg}) {}
+
+  void save(void* ptr) const { *reinterpret_cast<ss16x8x4_t*>(ptr) = reg; }
+};
+
+struct FP32Vec4 : public Vec<FP32Vec4> {
+  constexpr static int VEC_ELEM_NUM = 4;
+  union AliasReg {
+    __vector float reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  __vector float reg;
+
+  explicit FP32Vec4(float v) : reg(vec_splats(v)) {}
+
+  explicit FP32Vec4() : reg(vec_splats(0.0f)) {}
+
+  explicit FP32Vec4(const float* ptr) : reg(vec_xl(0, ptr)) {}
+
+  explicit FP32Vec4(__vector float data) : reg(data) {}
+
+  explicit FP32Vec4(const FP32Vec4& data) : reg(data.reg) {}
+};
+
+struct FP32Vec8 : public Vec<FP32Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+  union AliasReg {
+    f32x4x2_t reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  f32x4x2_t reg;
+
+  explicit FP32Vec8(float v) {
+    reg.val[0] = vec_splats(v);
+    reg.val[1] = vec_splats(v);
+  }
+
+  explicit FP32Vec8() {
+    reg.val[0] = vec_splats(0.0f);
+    reg.val[1] = vec_splats(0.0f);
+  }
+
+  explicit FP32Vec8(const float* ptr) {
+    reg.val[0] = vec_xl(0, ptr);
+    reg.val[1] = vec_xl(16, ptr);
+  }
+
+  explicit FP32Vec8(f32x4x2_t data) : reg(data) {}
+
+  explicit FP32Vec8(const FP32Vec8& data) {
+    reg.val[0] = data.reg.val[0];
+    reg.val[1] = data.reg.val[1];
+  }
+
+  explicit FP32Vec8(const BF16Vec8& v) {
+    reg.val[0] = (__vector float)vec_mergeh(zero, v.reg);
+    reg.val[1] = (__vector float)vec_mergel(zero, v.reg);
+  }
+
+  float reduce_sum() const {
+    AliasReg ar;
+    ar.reg = reg;
+    float result = 0;
+    unroll_loop<int, VEC_ELEM_NUM>(
+        [&result, &ar](int i) { result += ar.values[i]; });
+
+    return result;
+  }
+
+  FP32Vec8 exp() const {
+    // TODO: Vectorize this
+    AliasReg ar;
+    ar.reg = reg;
+    f32x4x4_t ret;
+    ret.val[0][0] = std::exp(ar.values[0]);
+    ret.val[0][1] = std::exp(ar.values[1]);
+    ret.val[0][2] = std::exp(ar.values[2]);
+    ret.val[0][3] = std::exp(ar.values[3]);
+    ret.val[1][0] = std::exp(ar.values[4]);
+    ret.val[1][1] = std::exp(ar.values[5]);
+    ret.val[1][2] = std::exp(ar.values[6]);
+    ret.val[1][3] = std::exp(ar.values[7]);
+    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
+  }
+
+  FP32Vec8 tanh() const {
+    // TODO: Vectorize this
+    AliasReg ar;
+    ar.reg = reg;
+    f32x4x4_t ret;
+    ret.val[0][0] = std::tanh(ar.values[0]);
+    ret.val[0][1] = std::tanh(ar.values[1]);
+    ret.val[0][2] = std::tanh(ar.values[2]);
+    ret.val[0][3] = std::tanh(ar.values[3]);
+    ret.val[1][0] = std::tanh(ar.values[4]);
+    ret.val[1][1] = std::tanh(ar.values[5]);
+    ret.val[1][2] = std::tanh(ar.values[6]);
+    ret.val[1][3] = std::tanh(ar.values[7]);
+    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
+  }
+
+  FP32Vec8 er() const {
+    // TODO: Vectorize this
+    AliasReg ar;
+    ar.reg = reg;
+    f32x4x4_t ret;
+    ret.val[0][0] = std::erf(ar.values[0]);
+    ret.val[0][1] = std::erf(ar.values[1]);
+    ret.val[0][2] = std::erf(ar.values[2]);
+    ret.val[0][3] = std::erf(ar.values[3]);
+    ret.val[1][0] = std::erf(ar.values[4]);
+    ret.val[1][1] = std::erf(ar.values[5]);
+    ret.val[1][2] = std::erf(ar.values[6]);
+    ret.val[1][3] = std::erf(ar.values[7]);
+    return FP32Vec8(f32x4x2_t({ret.val[0], ret.val[1]}));
+  }
+
+  FP32Vec8 operator*(const FP32Vec8& b) const {
+    return FP32Vec8(
+        {vec_mul(reg.val[0], b.reg.val[0]), vec_mul(reg.val[1], b.reg.val[1])});
+  }
+
+  FP32Vec8 operator+(const FP32Vec8& b) const {
+    return FP32Vec8(
+        {vec_add(reg.val[0], b.reg.val[0]), vec_add(reg.val[1], b.reg.val[1])});
+  }
+
+  FP32Vec8 operator-(const FP32Vec8& b) const {
+    return FP32Vec8(
+        {vec_sub(reg.val[0], b.reg.val[0]), vec_sub(reg.val[1], b.reg.val[1])});
+  }
+
+  FP32Vec8 operator/(const FP32Vec8& b) const {
+    return FP32Vec8(
+        {vec_div(reg.val[0], b.reg.val[0]), vec_div(reg.val[1], b.reg.val[1])});
+  }
+
+  void save(float* ptr) const {
+    vec_xst(reg.val[0], 0, ptr);
+    vec_xst(reg.val[1], 16, ptr);
+  }
+};
+
+struct FP32Vec16 : public Vec<FP32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  union AliasReg {
+    f32x4x4_t reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  f32x4x4_t reg;
+
+  explicit FP32Vec16(float v) {
+    reg.val[0] = vec_splats(v);
+    reg.val[1] = vec_splats(v);
+    reg.val[2] = vec_splats(v);
+    reg.val[3] = vec_splats(v);
+  }
+
+  explicit FP32Vec16() {
+    reg.val[0] = vec_splats(0.0f);
+    reg.val[1] = vec_splats(0.0f);
+    reg.val[2] = vec_splats(0.0f);
+    reg.val[3] = vec_splats(0.0f);
+  }
+
+  explicit FP32Vec16(const float* ptr) {
+    reg.val[0] = vec_xl(0, ptr);
+    reg.val[1] = vec_xl(16, ptr);
+    reg.val[2] = vec_xl(32, ptr);
+    reg.val[3] = vec_xl(48, ptr);
+  }
+
+  explicit FP32Vec16(f32x4x4_t data) : reg(data) {}
+
+  explicit FP32Vec16(const FP32Vec16& data) {
+    reg.val[0] = data.reg.val[0];
+    reg.val[1] = data.reg.val[1];
+    reg.val[2] = data.reg.val[2];
+    reg.val[3] = data.reg.val[3];
+  }
+
+  explicit FP32Vec16(const FP32Vec4& data) {
+    reg.val[0] = data.reg;
+    reg.val[1] = data.reg;
+    reg.val[2] = data.reg;
+    reg.val[3] = data.reg;
+  }
+
+  explicit FP32Vec16(const FP32Vec8& data) {
+    reg.val[0] = data.reg.val[0];
+    reg.val[1] = data.reg.val[1];
+    reg.val[2] = data.reg.val[0];
+    reg.val[3] = data.reg.val[1];
+  }
+
+  explicit FP32Vec16(const BF16Vec16& v) {
+    reg.val[0] = (__vector float)vec_mergeh(zero, v.reg.val[0]);
+    reg.val[1] = (__vector float)vec_mergel(zero, v.reg.val[0]);
+    reg.val[2] = (__vector float)vec_mergeh(zero, v.reg.val[1]);
+    reg.val[3] = (__vector float)vec_mergel(zero, v.reg.val[1]);
+  }
+
+  explicit FP32Vec16(const BF16Vec8& v) : FP32Vec16(FP32Vec8(v)) {}
+
+  FP32Vec16 operator*(const FP32Vec16& b) const {
+    return FP32Vec16(f32x4x4_t({vec_mul(reg.val[0], b.reg.val[0]),
+                                vec_mul(reg.val[1], b.reg.val[1]),
+                                vec_mul(reg.val[2], b.reg.val[2]),
+                                vec_mul(reg.val[3], b.reg.val[3])}));
+  }
+
+  FP32Vec16 operator+(const FP32Vec16& b) const {
+    return FP32Vec16(f32x4x4_t({vec_add(reg.val[0], b.reg.val[0]),
+                                vec_add(reg.val[1], b.reg.val[1]),
+                                vec_add(reg.val[2], b.reg.val[2]),
+                                vec_add(reg.val[3], b.reg.val[3])}));
+  }
+
+  FP32Vec16 operator-(const FP32Vec16& b) const {
+    return FP32Vec16(f32x4x4_t({vec_sub(reg.val[0], b.reg.val[0]),
+                                vec_sub(reg.val[1], b.reg.val[1]),
+                                vec_sub(reg.val[2], b.reg.val[2]),
+                                vec_sub(reg.val[3], b.reg.val[3])}));
+  }
+
+  FP32Vec16 operator/(const FP32Vec16& b) const {
+    return FP32Vec16(f32x4x4_t({vec_div(reg.val[0], b.reg.val[0]),
+                                vec_div(reg.val[1], b.reg.val[1]),
+                                vec_div(reg.val[2], b.reg.val[2]),
+                                vec_div(reg.val[3], b.reg.val[3])}));
+  }
+
+  float reduce_sum() const {
+    AliasReg ar;
+    ar.reg = reg;
+    float result = 0;
+    unroll_loop<int, VEC_ELEM_NUM>(
+        [&result, &ar](int i) { result += ar.values[i]; });
+
+    return result;
+  }
+
+  template <int group_size>
+  float reduce_sub_sum(int idx) {
+    static_assert(VEC_ELEM_NUM % group_size == 0);
+
+    AliasReg ar;
+    ar.reg = reg;
+    float result = 0;
+    const int start = idx * group_size;
+    unroll_loop<int, group_size>(
+        [&result, &start, ar](int i) { result += ar.values[start + i]; });
+
+    return result;
+  }
+
+  void save(float* ptr) const {
+    vec_xst(reg.val[0], 0, ptr);
+    vec_xst(reg.val[1], 16, ptr);
+    vec_xst(reg.val[2], 32, ptr);
+    vec_xst(reg.val[3], 48, ptr);
+  }
+};
+
+template <typename T>
+struct VecType {
+  using vec_type = void;
+};
+
+template <typename T>
+using vec_t = typename VecType<T>::vec_type;
+
+template <>
+struct VecType<float> {
+  using vec_type = FP32Vec8;
+};
+
+template <>
+struct VecType<c10::BFloat16> {
+  using vec_type = BF16Vec8;
+};
+
+template <typename T>
+void storeFP32(float v, T* ptr) {
+  *ptr = v;
+}
+
+inline void fma(FP32Vec16& acc, FP32Vec16& a, FP32Vec16& b) {
+  acc = acc + a * b;
+}
+
+namespace c10 {
+struct BFloat16 {
+  uint16_t value;  // Assume BFloat16 is defined as a struct containing a 16-bit
+                   // value.
+};
+}  // namespace c10
+
+template <>
+inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16* ptr) {
+  c10::BFloat16 __attribute__((__may_alias__))* v_ptr =
+      reinterpret_cast<c10::BFloat16*>(&v);
+  *ptr = *(v_ptr + 1);
+}
+
+#ifndef __VEC_CLASS_FP_NAN
+  #define __VEC_CLASS_FP_NAN (1 << 6)
+#endif
+
+const static __vector unsigned char omask = {2,  3,  6,  7,  10, 11, 14, 15,
+                                             18, 19, 22, 23, 26, 27, 30, 31};
+const static __vector unsigned int bias = {0x00007fff, 0x00007fff, 0x00007fff,
+                                           0x00007fff};
+const static __vector unsigned int nan = {0x7fc00000, 0x7fc00000, 0x7fc00000,
+                                          0x7fc00000};
+const static __vector unsigned int sh16 = {16, 16, 16, 16};
+const static __vector unsigned int one = {1, 1, 1, 1};
+
+inline BF16Vec8::BF16Vec8(const FP32Vec8& v) {
+  __vector unsigned int inp0 = (__vector unsigned int)(v.reg.val[0]);
+  __vector unsigned int inp1 = (__vector unsigned int)(v.reg.val[1]);
+  int cc;
+  __vector __bool int sel0 =
+      vec_fp_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN, &cc);
+  __vector __bool int sel1 =
+      vec_fp_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN, &cc);
+  inp0 = vec_sel(inp0, nan, sel0) >> sh16;
+  inp1 = vec_sel(inp1, nan, sel1) >> sh16;
+  reg = (__vector signed short)vec_perm(inp0, inp1, omask);
+}
+
+inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
+  __vector unsigned int inp0 = (__vector unsigned int)(v.reg.val[0]);
+  __vector unsigned int inp1 = (__vector unsigned int)(v.reg.val[1]);
+  __vector unsigned int inp2 = (__vector unsigned int)(v.reg.val[2]);
+  __vector unsigned int inp3 = (__vector unsigned int)(v.reg.val[3]);
+  int cc;
+  __vector __bool int sel0 =
+      vec_fp_test_data_class(v.reg.val[0], __VEC_CLASS_FP_NAN, &cc);
+  __vector __bool int sel1 =
+      vec_fp_test_data_class(v.reg.val[1], __VEC_CLASS_FP_NAN, &cc);
+  __vector __bool int sel2 =
+      vec_fp_test_data_class(v.reg.val[2], __VEC_CLASS_FP_NAN, &cc);
+  __vector __bool int sel3 =
+      vec_fp_test_data_class(v.reg.val[3], __VEC_CLASS_FP_NAN, &cc);
+  inp0 = vec_sel(inp0, nan, sel0) >> sh16;
+  inp1 = vec_sel(inp1, nan, sel1) >> sh16;
+  inp2 = vec_sel(inp2, nan, sel2) >> sh16;
+  inp3 = vec_sel(inp3, nan, sel3) >> sh16;
+  reg.val[0] = (__vector signed short)vec_perm(inp0, inp1, omask);
+  reg.val[1] = (__vector signed short)vec_perm(inp2, inp3, omask);
+}
+
+inline void prefetch(const void* addr) { void __dcbt(const void* addr); }
+
+};  // namespace vec_op
+
+#endif

csrc/cpu/cpu_types_x86.hpp

@@ -16,9 +16,18 @@ namespace vec_op {
   AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
   AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)
 
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES_FP8(...)        \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)    \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Float8_e5m2, __VA_ARGS__)
+
 #define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
   AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
 
+#define VLLM_DISPATCH_FLOATING_TYPES_WITH_E5M2(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME,                                \
+                     VLLM_DISPATCH_CASE_FLOATING_TYPES_FP8(__VA_ARGS__))
+
 #ifndef CPU_OP_GUARD
   #define CPU_KERNEL_GUARD_IN(NAME)
   #define CPU_KERNEL_GUARD_OUT(NAME)
@@ -69,9 +78,14 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
 
   __m256i reg;
 
+  // normal load
   explicit FP16Vec16(const void* ptr)
       : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
+  // non-temproal load
+  explicit FP16Vec16(bool, void* ptr)
+      : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
+
   explicit FP16Vec16(const FP32Vec16&);
 
   void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
@@ -101,9 +115,14 @@ struct BF16Vec16 : public Vec<BF16Vec16> {
 
   __m256i reg;
 
+  // normal load
   explicit BF16Vec16(const void* ptr)
       : reg((__m256i)_mm256_loadu_si256((__m256i*)ptr)) {}
 
+  // non-temproal load
+  explicit BF16Vec16(bool, void* ptr)
+      : reg(_mm256_stream_load_si256((__m256i*)ptr)) {}
+
   explicit BF16Vec16(const FP32Vec16&);
 
   void save(void* ptr) const { *reinterpret_cast<__m256i*>(ptr) = reg; }
@@ -121,6 +140,8 @@ struct BF16Vec32 : public Vec<BF16Vec32> {
 
   __m512i reg;
 
+  explicit BF16Vec32() : reg(_mm512_setzero_si512()) {}
+
   explicit BF16Vec32(const void* ptr) : reg((__m512i)_mm512_loadu_si512(ptr)) {}
 
   explicit BF16Vec32(__m512i data) : reg(data) {}
@@ -302,8 +323,13 @@ struct FP32Vec16 : public Vec<FP32Vec16> {
 
   explicit FP32Vec16() : reg(_mm512_set1_ps(0.0)) {}
 
+  // normal load
   explicit FP32Vec16(const float* ptr) : reg(_mm512_loadu_ps(ptr)) {}
 
+  // non-temproal load
+  explicit FP32Vec16(bool, void* ptr)
+      : reg((__m512)_mm512_stream_load_si512(ptr)) {}
+
   explicit FP32Vec16(__m512 data) : reg(data) {}
 
   explicit FP32Vec16(const FP32Vec4& data)
@@ -536,6 +562,33 @@ struct INT8Vec16 : public Vec<INT8Vec16> {
     _mm_mask_storeu_epi8(ptr, mask, reg);
   }
 };
+
+struct INT8Vec64 : public Vec<INT8Vec64> {
+  constexpr static int VEC_ELEM_NUM = 64;
+  union AliasReg {
+    __m512i reg;
+    int8_t values[VEC_ELEM_NUM];
+  };
+
+  __m512i reg;
+
+  // normal load
+  explicit INT8Vec64(void* ptr) : reg(_mm512_loadu_epi8(ptr)) {}
+
+  // non-temproal load
+  explicit INT8Vec64(bool, void* ptr) : reg(_mm512_stream_load_si512(ptr)) {}
+
+  void save(void* ptr) const { _mm512_storeu_epi8(ptr, reg); }
+
+  void save(int8_t* ptr, const int elem_num) const {
+    constexpr uint64_t M = 0xFFFFFFFFFFFFFFFF;
+    __mmask64 mask = _cvtu64_mask64(M >> (64 - elem_num));
+    _mm512_mask_storeu_epi8(ptr, mask, reg);
+  }
+
+  // non-temproal save
+  void nt_save(int8_t* ptr) { _mm512_stream_si512((__m512i*)ptr, reg); }
+};
 #endif
 
 template <typename T>
@@ -646,6 +699,22 @@ inline BF16Vec16::BF16Vec16(const FP32Vec16& v) {
 
 inline void prefetch(const void* addr) { _mm_prefetch(addr, _MM_HINT_T1); }
 
+#ifdef __AVX512F__
+inline void non_temporal_save(FP16Vec16& vec, void* ptr) {
+  _mm256_stream_si256((__m256i*)ptr, vec.reg);
+}
+inline void non_temporal_save(BF16Vec32& vec, void* ptr) {
+  _mm512_stream_si512((__m512i*)ptr, vec.reg);
+}
+inline void non_temporal_save(BF16Vec16& vec, void* ptr) {
+  _mm256_stream_si256((__m256i*)ptr, vec.reg);
+}
+inline void non_temporal_save(FP32Vec16& vec, void* ptr) {
+  _mm512_stream_ps((float*)ptr, vec.reg);
+}
+#endif
+
+inline void mem_barrier() { _mm_mfence(); }
 };  // namespace vec_op
 
 #endif

csrc/cpu/mla_decode.cpp

@@ -0,0 +1,393 @@
+#include "cpu_types.hpp"
+#include <float.h>
+
+namespace {
+template <typename scalar_t>
+struct KernelVecType {
+  using qk_load_vec_type = void;
+  using qk_vec_type = void;
+  using v_load_vec_type = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using qk_load_vec_type = vec_op::FP32Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+};
+
+template <>
+struct KernelVecType<c10::Half> {
+#if defined(__powerpc64__) || defined(__s390x__)
+  // Power and s390x architecture-specific vector types
+  using qk_load_vec_type = vec_op::FP32Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+#else
+  // Fallback for other architectures, including x86
+  using qk_load_vec_type = vec_op::FP16Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP16Vec16;
+#endif
+};
+
+#ifdef __AVX512BF16__
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using qk_load_vec_type = vec_op::BF16Vec32;
+  using qk_vec_type = vec_op::BF16Vec32;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+#elif defined(__aarch64__) && !defined(ARM_BF16_SUPPORT)
+// pass
+#else
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using qk_load_vec_type = vec_op::BF16Vec16;
+  using qk_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+#endif
+
+template <int HEAD_DIM, int V_HEAD_DIM, int BLOCK_SIZE, int HEAD_UNROLL,
+          typename qk_vec_type>
+void mla_decode_block_head(
+    const qk_vec_type* __restrict__ q_vecs,          // [HEAD_UNROLL, head_dim]
+    const qk_vec_type* __restrict__ k_vecs,          // [block_size, head_dim]
+    const vec_op::FP32Vec16* __restrict v_vecs_f32,  // [block_size, v_head_dim]
+    float* __restrict__ acc_out,  // [HEAD_UNROLL, v_head_dim]
+    float* __restrict__ acc_lse,  // [HEAD_UNROLL]
+    const float scale, const int num_tokens) {
+  using f32_vec_type = vec_op::FP32Vec16;
+  constexpr int QK_NUM_ELEM = qk_vec_type::VEC_ELEM_NUM;
+  constexpr int V_NUM_ELEM = f32_vec_type::VEC_ELEM_NUM;
+
+  float logits[BLOCK_SIZE][HEAD_UNROLL] = {};  // initialize to zeros
+  float max_val[HEAD_UNROLL];
+  std::fill(max_val, max_val + HEAD_UNROLL, -FLT_MAX);
+
+  f32_vec_type acc_vec[BLOCK_SIZE][HEAD_UNROLL];
+  for (int i = 0; i < HEAD_DIM; i += QK_NUM_ELEM) {
+    // load to registers
+    qk_vec_type q_vec[HEAD_UNROLL];
+
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+      q_vec[unroll] =
+          qk_vec_type{q_vecs[(i + unroll * HEAD_DIM) / QK_NUM_ELEM]};
+
+    for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+      qk_vec_type k_vec(k_vecs[(block_offset * HEAD_DIM + i) / QK_NUM_ELEM]);
+
+#pragma unroll
+      for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+        vec_op::fma(acc_vec[block_offset][unroll], q_vec[unroll], k_vec);
+    }
+  }
+
+  for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+      const float acc = acc_vec[block_offset][unroll].reduce_sum() * scale;
+      logits[block_offset][unroll] = acc;
+      max_val[unroll] = std::max(max_val[unroll], acc);
+    }
+  }
+
+  float sum_exp[HEAD_UNROLL] = {};
+  for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+      const float val =
+          std::exp(logits[block_offset][unroll] - max_val[unroll]);
+      logits[block_offset][unroll] = val;
+      sum_exp[unroll] += val;
+    }
+  }
+
+  f32_vec_type this_out[V_HEAD_DIM / V_NUM_ELEM][HEAD_UNROLL];
+
+  for (int block_offset = 0; block_offset < num_tokens; ++block_offset) {
+    // load to registers
+    f32_vec_type scale_[HEAD_UNROLL];
+
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+      scale_[unroll] =
+          f32_vec_type{logits[block_offset][unroll] / sum_exp[unroll]};
+
+    for (int i = 0; i < V_HEAD_DIM; i += V_NUM_ELEM) {
+      f32_vec_type v_vec(
+          v_vecs_f32[(block_offset * HEAD_DIM + i) / V_NUM_ELEM]);
+
+#pragma unroll
+      for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll)
+        vec_op::fma(this_out[i / V_NUM_ELEM][unroll], v_vec, scale_[unroll]);
+    }
+  }
+
+  // merge attention state
+  // section 2.2 in https://arxiv.org/pdf/2501.01005
+  f32_vec_type prev_scale[HEAD_UNROLL];
+  f32_vec_type curr_scale[HEAD_UNROLL];
+
+#pragma unroll
+  for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+    const float prev_lse = acc_lse[unroll];
+    const float curr_lse = std::log(sum_exp[unroll]) +
+                           max_val[unroll];  // add back max_val to get true lse
+    // softmax trick
+    const float max_lse = std::max(prev_lse, curr_lse);
+    const float prev_sum_exp = std::exp(prev_lse - max_lse);
+    const float curr_sum_exp = std::exp(curr_lse - max_lse);
+
+    const float new_sum_exp = prev_sum_exp + curr_sum_exp;
+    acc_lse[unroll] = std::log(new_sum_exp) + max_lse;
+
+    prev_scale[unroll] = f32_vec_type{prev_sum_exp / new_sum_exp};
+    curr_scale[unroll] = f32_vec_type{curr_sum_exp / new_sum_exp};
+  }
+
+  for (int i = 0; i < V_HEAD_DIM; i += V_NUM_ELEM) {
+#pragma unroll
+    for (int unroll = 0; unroll < HEAD_UNROLL; ++unroll) {
+      f32_vec_type o_vec(acc_out + i + V_HEAD_DIM * unroll);
+      o_vec = o_vec * prev_scale[unroll] +
+              this_out[i / V_NUM_ELEM][unroll] * curr_scale[unroll];
+      o_vec.save(acc_out + i + V_HEAD_DIM * unroll);
+    }
+  }
+
+  q_vecs += HEAD_DIM / QK_NUM_ELEM * HEAD_UNROLL;
+  acc_out += V_HEAD_DIM * HEAD_UNROLL;
+}
+
+template <typename scalar_t, int HEAD_DIM, int V_HEAD_DIM, int BLOCK_SIZE,
+          typename qk_vec_type>
+void mla_decode_block(
+    const qk_vec_type* __restrict__ q_vecs,  // [num_heads, head_dim]
+    const scalar_t* __restrict__ kv_cache,   // [block_size, head_dim]
+    float* __restrict__ acc_out,             // [num_heads, v_head_dim]
+    float* __restrict__ acc_lse,             // [num_heads]
+    const int num_heads, const float scale, const int num_tokens) {
+  using qk_load_vec_type = typename KernelVecType<scalar_t>::qk_load_vec_type;
+  static_assert(
+      std::is_same<qk_vec_type,
+                   typename KernelVecType<scalar_t>::qk_vec_type>::value);
+  using v_load_vec_type = typename KernelVecType<scalar_t>::v_load_vec_type;
+  using f32_vec_type = vec_op::FP32Vec16;
+  static_assert(qk_load_vec_type::VEC_ELEM_NUM == qk_vec_type::VEC_ELEM_NUM);
+  static_assert(v_load_vec_type::VEC_ELEM_NUM == f32_vec_type::VEC_ELEM_NUM);
+  constexpr int QK_NUM_ELEM = qk_vec_type::VEC_ELEM_NUM;
+  constexpr int V_NUM_ELEM = v_load_vec_type::VEC_ELEM_NUM;
+
+  const qk_vec_type* k_vecs;
+  const f32_vec_type* v_vecs_f32;
+  float* kv_cache_f32 = nullptr;
+
+  if constexpr (!std::is_same<scalar_t, float>::value) {
+    // convert KV cache block to FP32 to reuse it across query heads and
+    // attn @ V computation, since FP16/BF16->FP32 is expensive.
+    // TODO: move malloc outside of this fn to reuse across iterations.
+    const int nbytes = BLOCK_SIZE * HEAD_DIM * sizeof(float);
+    kv_cache_f32 = static_cast<float*>(std::aligned_alloc(64, nbytes));
+
+    for (int block_offset = 0; block_offset < num_tokens; ++block_offset)
+      for (int i = 0; i < HEAD_DIM; i += V_NUM_ELEM) {
+        v_load_vec_type kv_load_vec(kv_cache + block_offset * HEAD_DIM + i);
+        f32_vec_type kv_vec_f32(kv_load_vec);
+        kv_vec_f32.save(kv_cache_f32 + block_offset * HEAD_DIM + i);
+      }
+
+    if constexpr (std::is_same<qk_load_vec_type, qk_vec_type>::value) {
+      // for AVX512_BF16, Q @ K.T uses BF16 for K (no conversion)
+      // NOTE: in this case, we only need to convert the V section to FP32.
+      // But for simplicity, we will convert the whole KV block to FP32.
+      k_vecs = reinterpret_cast<const qk_vec_type*>(kv_cache);
+    } else {
+      k_vecs = reinterpret_cast<const qk_vec_type*>(kv_cache_f32);
+    }
+
+    // attn @ V always use FP32 for V, since attn is FP32.
+    v_vecs_f32 = reinterpret_cast<const f32_vec_type*>(kv_cache_f32);
+
+  } else {
+    // KV cache is FP32. don't need to do anything.
+    k_vecs = reinterpret_cast<const qk_vec_type*>(kv_cache);
+    v_vecs_f32 = reinterpret_cast<const f32_vec_type*>(kv_cache);
+  }
+
+  // compute 2 heads at the same time to improve ILP and
+  // take advantage of register cache for K and V.
+  constexpr int HEAD_UNROLL = 2;
+  for (int iter = 0; iter < num_heads / HEAD_UNROLL; ++iter) {
+    mla_decode_block_head<HEAD_DIM, V_HEAD_DIM, BLOCK_SIZE, HEAD_UNROLL>(
+        q_vecs, k_vecs, v_vecs_f32, acc_out, acc_lse, scale, num_tokens);
+
+    q_vecs += HEAD_UNROLL * HEAD_DIM / QK_NUM_ELEM;
+    acc_out += HEAD_UNROLL * V_HEAD_DIM;
+    acc_lse += HEAD_UNROLL;
+  }
+
+  // take care of the remaining heads
+  for (int iter = 0; iter < num_heads % HEAD_UNROLL; ++iter) {
+    mla_decode_block_head<HEAD_DIM, V_HEAD_DIM, BLOCK_SIZE, 1>(
+        q_vecs, k_vecs, v_vecs_f32, acc_out, acc_lse, scale, num_tokens);
+
+    q_vecs += HEAD_DIM / QK_NUM_ELEM;
+    acc_out += V_HEAD_DIM;
+    acc_lse += 1;
+  }
+
+  if (kv_cache_f32 != nullptr) {
+    std::free(kv_cache_f32);
+  }
+}
+}  // namespace
+
+template <typename scalar_t, int HEAD_DIM, int V_HEAD_DIM, int BLOCK_SIZE>
+void mla_decode_kvcache_cpu_impl(
+    scalar_t* __restrict__ out,             // [num_seqs, num_heads, v_head_dim]
+    const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_dim]
+    const scalar_t* __restrict__ kv_cache,  // [num_blocks, block_size,
+                                            // head_dim]
+    const int num_heads, const float scale,
+    const int* __restrict__ block_tables,  // [num_seqs, max_num_blocks_per_seq]
+    const int* __restrict__ seq_lens,      // [num_seqs]
+    const int max_num_blocks_per_seq, const int o_stride, const int q_stride,
+    const int kv_stride, const int num_seqs) {
+  using qk_load_vec_type = typename KernelVecType<scalar_t>::qk_load_vec_type;
+  using qk_vec_type = typename KernelVecType<scalar_t>::qk_vec_type;
+  constexpr int QK_NUM_ELEM = qk_vec_type::VEC_ELEM_NUM;
+
+  // shared across threads
+  const int max_threads = omp_get_max_threads();
+  const int acc_out_nbytes =
+      max_threads * num_heads * V_HEAD_DIM * sizeof(float);
+  float* acc_out = static_cast<float*>(std::aligned_alloc(64, acc_out_nbytes));
+  std::vector<float> acc_lse(max_threads * num_heads);
+
+  // allocate memory to pre-convert query to FP32 later
+  float* q_f32;
+  constexpr bool PRE_CONVERT_QUERY =
+      !std::is_same<scalar_t, float>::value &&
+      std::is_same<qk_vec_type, vec_op::FP32Vec16>::value;
+  if constexpr (PRE_CONVERT_QUERY) {
+    const int q_f32_nbytes = num_heads * HEAD_DIM * sizeof(float);
+    q_f32 = static_cast<float*>(std::aligned_alloc(64, q_f32_nbytes));
+  }
+
+#pragma omp parallel
+  {
+    const int num_threads = omp_get_num_threads();
+    const int thread_id = omp_get_thread_num();
+    float* __restrict__ acc_out_thread =
+        acc_out + thread_id * num_heads * V_HEAD_DIM;
+    float* __restrict__ acc_lse_thread = acc_lse.data() + thread_id * num_heads;
+
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      // reset accumulator
+      std::fill(acc_out_thread, acc_out_thread + num_heads * V_HEAD_DIM, 0.0f);
+      std::fill(acc_lse_thread, acc_lse_thread + num_heads, -FLT_MAX);
+
+      const int seq_len = seq_lens[seq_idx];
+      const int block_num = (seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+      const int last_block_size = seq_len - (block_num - 1) * BLOCK_SIZE;
+
+      const qk_vec_type* q_vecs;
+      if constexpr (PRE_CONVERT_QUERY) {
+// pre-convert query to FP32 since FP16/BF16->FP32 is slow.
+#pragma omp for
+        for (int i = 0; i < num_heads * HEAD_DIM; i += QK_NUM_ELEM) {
+          qk_load_vec_type q_load_vec(q + seq_idx * q_stride + i);
+          qk_vec_type q_vec(q_load_vec);
+          q_vec.save(q_f32 + i);
+        }
+        q_vecs = reinterpret_cast<const qk_vec_type*>(q_f32);
+      } else {
+        q_vecs = reinterpret_cast<const qk_vec_type*>(q + seq_idx * q_stride);
+      }
+
+#pragma omp for
+      for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+        const int physical_block_idx =
+            block_tables[seq_idx * max_num_blocks_per_seq + block_idx];
+        const int num_tokens =
+            block_idx < block_num - 1 ? BLOCK_SIZE : last_block_size;
+
+        mla_decode_block<scalar_t, HEAD_DIM, V_HEAD_DIM, BLOCK_SIZE>(
+            q_vecs, kv_cache + physical_block_idx * kv_stride, acc_out_thread,
+            acc_lse_thread, num_heads, scale, num_tokens);
+      }
+
+// merge attention states across threads
+// section 2.2 in https://arxiv.org/pdf/2501.01005
+// each thread is responsible for 1 head
+#pragma omp for
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        float* acc_lse_head = acc_lse.data() + head_idx;
+        float* acc_out_head = acc_out + head_idx * V_HEAD_DIM;
+
+        float max_val = -FLT_MAX;
+        for (int thread_id_ = 0; thread_id_ < num_threads; ++thread_id_) {
+          max_val = std::max(max_val, acc_lse_head[thread_id_ * num_heads]);
+        }
+
+        float sum_exp = 0.0f;
+        for (int thread_id_ = 0; thread_id_ < num_threads; ++thread_id_) {
+          float val = std::exp(acc_lse_head[thread_id_ * num_heads] - max_val);
+          acc_lse_head[thread_id_ * num_heads] = val;
+          sum_exp += val;
+        }
+
+        float inv_sum = 1.0f / sum_exp;
+        float out_head[V_HEAD_DIM] = {};
+        for (int thread_id_ = 0; thread_id_ < num_threads; ++thread_id_) {
+          float scale_ = acc_lse_head[thread_id_ * num_heads] * inv_sum;
+          for (int i = 0; i < V_HEAD_DIM; ++i) {
+            out_head[i] +=
+                acc_out_head[thread_id_ * num_heads * V_HEAD_DIM + i] * scale_;
+          }
+        }
+
+        for (int i = 0; i < V_HEAD_DIM; ++i) {
+          vec_op::storeFP32(out_head[i], out + seq_idx * o_stride +
+                                             head_idx * V_HEAD_DIM + i);
+        }
+      }
+    }
+  }
+  if (PRE_CONVERT_QUERY) {
+    std::free(q_f32);
+  }
+  std::free(acc_out);
+}
+
+void mla_decode_kvcache(torch::Tensor& out, torch::Tensor& query,
+                        torch::Tensor& kv_cache, double scale,
+                        torch::Tensor& block_tables, torch::Tensor& seq_lens) {
+  const int num_seqs = query.size(0);
+  const int num_heads = query.size(1);
+  const int head_dim = query.size(2);
+  const int block_size = kv_cache.size(1);
+  const int v_head_dim = out.size(2);
+
+  const int max_num_blocks_per_seq = block_tables.size(1);
+  const int o_stride = out.stride(0);
+  const int q_stride = query.stride(0);
+  const int kv_stride = kv_cache.stride(0);
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      query.scalar_type(), "mla_decode_kvcache_cpu_impl", [&] {
+        CPU_KERNEL_GUARD_IN(mla_decode_kvcache_cpu_impl)
+        if (head_dim == 576 && v_head_dim == 512 && block_size == 16)
+          mla_decode_kvcache_cpu_impl<scalar_t, 576, 512, 16>(
+              out.data_ptr<scalar_t>(), query.data_ptr<scalar_t>(),
+              kv_cache.data_ptr<scalar_t>(), num_heads, scale,
+              block_tables.data_ptr<int>(), seq_lens.data_ptr<int>(),
+              max_num_blocks_per_seq, o_stride, q_stride, kv_stride, num_seqs);
+        else
+          TORCH_CHECK(false, "Unsupported block size: ", block_size);
+        CPU_KERNEL_GUARD_OUT(mla_decode_kvcache_cpu_impl)
+      });
+}

csrc/cpu/pos_encoding.cpp

@@ -170,7 +170,7 @@ void rotary_embedding_gptj_impl(
 void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
                       torch::Tensor& key, int64_t head_size,
                       torch::Tensor& cos_sin_cache, bool is_neox) {
-  int num_tokens = query.numel() / query.size(-1);
+  int num_tokens = positions.numel();
   int rot_dim = cos_sin_cache.size(1);
   int num_heads = query.size(-1) / head_size;
   int num_kv_heads = key.size(-1) / head_size;