format

Co-authored-by: xuhao <xuhao@cambricon.com>

.buildkite/check-wheel-size.py

@@ -0,0 +1,36 @@
+import os
+import zipfile
+
+MAX_SIZE_MB = 200
+
+
+def print_top_10_largest_files(zip_file):
+    with zipfile.ZipFile(zip_file, 'r') as z:
+        file_sizes = [(f, z.getinfo(f).file_size) for f in z.namelist()]
+        file_sizes.sort(key=lambda x: x[1], reverse=True)
+        for f, size in file_sizes[:10]:
+            print(f"{f}: {size/(1024*1024)} MBs uncompressed.")
+
+
+def check_wheel_size(directory):
+    for root, _, files in os.walk(directory):
+        for f in files:
+            if f.endswith(".whl"):
+                wheel_path = os.path.join(root, f)
+                wheel_size = os.path.getsize(wheel_path)
+                wheel_size_mb = wheel_size / (1024 * 1024)
+                if wheel_size_mb > MAX_SIZE_MB:
+                    print(
+                        f"Wheel {wheel_path} is too large ({wheel_size_mb} MB) "
+                        f"compare to the allowed size ({MAX_SIZE_MB} MB).")
+                    print_top_10_largest_files(wheel_path)
+                    return 1
+                else:
+                    print(f"Wheel {wheel_path} is within the allowed size "
+                          f"({wheel_size_mb} MB).")
+    return 0
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(check_wheel_size(sys.argv[1]))

.buildkite/download-images.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+set -ex
+set -o pipefail
+
+(which wget && which curl) || (apt-get update && apt-get install -y wget curl)
+
+# aws s3 sync s3://air-example-data-2/vllm_opensource_llava/ images/
+mkdir -p images
+cd images
+wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/stop_sign_pixel_values.pt
+wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/stop_sign_image_features.pt
+wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/cherry_blossom_pixel_values.pt
+wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/cherry_blossom_image_features.pt
+wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/stop_sign.jpg
+wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/cherry_blossom.jpg
+
+cd -

.buildkite/run-amd-test.sh

@@ -0,0 +1,73 @@
+# This script runs test inside the corresponding ROCm docker container.
+set -ex
+
+# Print ROCm version
+echo "--- ROCm info"
+rocminfo
+
+# cleanup older docker images
+cleanup_docker() {
+  # Get Docker's root directory
+  docker_root=$(docker info -f '{{.DockerRootDir}}')
+  if [ -z "$docker_root" ]; then
+    echo "Failed to determine Docker root directory."
+    exit 1
+  fi
+  echo "Docker root directory: $docker_root"
+  # Check disk usage of the filesystem where Docker's root directory is located
+  disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
+  # Define the threshold
+  threshold=70
+  if [ "$disk_usage" -gt "$threshold" ]; then
+    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
+    # Remove dangling images (those that are not tagged and not used by any container)
+    docker image prune -f
+    # Remove unused volumes
+    docker volume prune -f
+    echo "Docker images and volumes cleanup completed."
+  else
+    echo "Disk usage is below $threshold%. No cleanup needed."
+  fi
+}
+
+# Call the cleanup docker function
+cleanup_docker
+
+echo "--- Resetting GPUs"
+
+echo "reset" > /opt/amdgpu/etc/gpu_state
+
+while true; do
+        sleep 3
+        if grep -q clean /opt/amdgpu/etc/gpu_state; then
+                echo "GPUs state is \"clean\""
+                break
+        fi
+done
+
+echo "--- Building container"
+sha=$(git rev-parse --short HEAD)
+image_name=rocm_${sha}
+container_name=rocm_${sha}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)
+docker build \
+        -t ${image_name} \
+        -f Dockerfile.rocm \
+        --progress plain \
+        .
+
+remove_docker_container() {
+   docker rm -f ${container_name} || docker image rm -f ${image_name} || true
+}
+trap remove_docker_container EXIT
+
+echo "--- Running container"
+
+docker run \
+        --device /dev/kfd --device /dev/dri \
+        --network host \
+        --rm \
+        -e HF_TOKEN \
+        --name ${container_name} \
+        ${image_name} \
+        /bin/bash -c "${@}"
+

.buildkite/run-benchmarks.sh

@@ -9,10 +9,10 @@ cd "$(dirname "${BASH_SOURCE[0]}")/.."
 (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
 
 # run python-based benchmarks and upload the result to buildkite
-python3 benchmarks/benchmark_latency.py 2>&1 | tee benchmark_latency.txt
+python3 benchmarks/benchmark_latency.py --output-json latency_results.json 2>&1 | tee benchmark_latency.txt
 bench_latency_exit_code=$?
 
-python3 benchmarks/benchmark_throughput.py --input-len 256 --output-len 256 2>&1 | tee benchmark_throughput.txt
+python3 benchmarks/benchmark_throughput.py --input-len 256 --output-len 256 --output-json throughput_results.json 2>&1 | tee benchmark_throughput.txt
 bench_throughput_exit_code=$?
 
 # run server-based benchmarks and upload the result to buildkite
@@ -23,8 +23,9 @@ wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/r
 # wait for server to start, timeout after 600 seconds
 timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
 python3 benchmarks/benchmark_serving.py \
-    --backend openai \
-    --dataset ./ShareGPT_V3_unfiltered_cleaned_split.json \
+    --backend vllm \
+    --dataset-name sharegpt \
+    --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json \
     --model meta-llama/Llama-2-7b-chat-hf \
     --num-prompts 20 \
     --endpoint /v1/completions \
@@ -48,7 +49,14 @@ sed -n '$p' benchmark_throughput.txt >> benchmark_results.md # last line
 echo "### Serving Benchmarks" >> benchmark_results.md
 sed -n '1p' benchmark_serving.txt >> benchmark_results.md # first line
 echo "" >> benchmark_results.md
-tail -n 13 benchmark_serving.txt >> benchmark_results.md # last 13 lines
+echo '```' >> benchmark_results.md
+tail -n 20 benchmark_serving.txt >> benchmark_results.md # last 20 lines
+echo '```' >> benchmark_results.md
+
+# if the agent binary is not found, skip uploading the results, exit 0
+if [ ! -f /workspace/buildkite-agent ]; then
+    exit 0
+fi
 
 # upload the results to buildkite
 /workspace/buildkite-agent annotate --style "info" --context "benchmark-results" < benchmark_results.md
@@ -66,4 +74,5 @@ if [ $bench_serving_exit_code -ne 0 ]; then
     exit $bench_serving_exit_code
 fi
 
-/workspace/buildkite-agent artifact upload openai-*.json
+rm ShareGPT_V3_unfiltered_cleaned_split.json
+/workspace/buildkite-agent artifact upload "*.json"

.buildkite/run-cpu-test.sh

@@ -0,0 +1,14 @@
+# 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 cpu-test -f Dockerfile.cpu .
+
+# Setup cleanup
+remove_docker_container() { docker rm -f cpu-test || true; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Run the image and launch offline inference
+docker run --network host --env VLLM_CPU_KVCACHE_SPACE=1 --name cpu-test cpu-test python3 vllm/examples/offline_inference.py

.buildkite/run-neuron-test.sh

@@ -0,0 +1,51 @@
+# This script build the Neuron docker image and run the API server inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -e
+
+# Try building the docker image
+aws ecr get-login-password --region us-west-2 | docker login --username AWS --password-stdin 763104351884.dkr.ecr.us-west-2.amazonaws.com
+
+# prune old image and containers to save disk space, and only once a day
+# by using a timestamp file in tmp.
+if [ -f /tmp/neuron-docker-build-timestamp ]; then
+    last_build=$(cat /tmp/neuron-docker-build-timestamp)
+    current_time=$(date +%s)
+    if [ $((current_time - last_build)) -gt 86400 ]; then
+        docker system prune -f
+        echo $current_time > /tmp/neuron-docker-build-timestamp
+    fi
+else
+    echo $(date +%s) > /tmp/neuron-docker-build-timestamp
+fi
+
+docker build -t neuron -f Dockerfile.neuron .
+
+# Setup cleanup
+remove_docker_container() { docker rm -f neuron || true; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Run the image
+docker run --device=/dev/neuron0 --device=/dev/neuron1 --network host --name neuron neuron python3 -m vllm.entrypoints.api_server \
+       --model TinyLlama/TinyLlama-1.1B-Chat-v1.0 --max-num-seqs 8 --max-model-len 128 --block-size 128 --device neuron --tensor-parallel-size 2 &
+
+# Wait for the server to start
+wait_for_server_to_start() {
+    timeout=300
+    counter=0
+
+    while [ "$(curl -s -o /dev/null -w ''%{http_code}'' localhost:8000/health)" != "200" ]; do
+        sleep 1
+        counter=$((counter + 1))
+        if [ $counter -ge $timeout ]; then
+            echo "Timeout after $timeout seconds"
+            break
+        fi
+    done
+}
+wait_for_server_to_start
+
+# Test a simple prompt
+curl -X POST -H "Content-Type: application/json" \
+    localhost:8000/generate \
+    -d '{"prompt": "San Francisco is a"}'

.buildkite/test-pipeline.yaml

@@ -5,64 +5,161 @@
 
 steps:
 - label: Regression Test
+  mirror_hardwares: [amd]
   command: pytest -v -s test_regression.py
   working_dir: "/vllm-workspace/tests" # optional
 
 - label: AsyncEngine Test
+  #mirror_hardwares: [amd]
   command: pytest -v -s async_engine
 
 - label: Basic Correctness Test
-  command: pytest -v -s --forked basic_correctness
+  mirror_hardwares: [amd]
+  commands:
+  - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_basic_correctness.py
+  - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_basic_correctness.py
+  - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_chunked_prefill.py
+  - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_chunked_prefill.py
+  - VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
+
+- label: Core Test
+  mirror_hardwares: [amd]
+  command: pytest -v -s core
 
 - label: Distributed Comm Ops Test
-  command: pytest -v -s --forked test_comm_ops.py
-  working_dir: "/vllm-workspace/tests/distributed"
-  num_gpus: 2 # only support 1 or 2 for now.
+  #mirror_hardwares: [amd]
+  command: pytest -v -s distributed/test_comm_ops.py
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
 
-- label: Distributed Correctness Test
-  command: pytest -v -s --forked test_basic_distributed_correctness.py
-  working_dir: "/vllm-workspace/tests/distributed"
-  num_gpus: 2 # only support 1 or 2 for now.
+- label: Distributed Tests
+  mirror_hardwares: [amd]
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
+  commands:
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
+  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
+  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
+  - pytest -v -s spec_decode/e2e/test_integration_dist.py 
+
+- label: Distributed Tests (Multiple Groups)
+  #mirror_hardwares: [amd]
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  commands:
+  - pytest -v -s distributed/test_pynccl.py
 
 - label: Engine Test
-  command: pytest -v -s engine
+  mirror_hardwares: [amd]
+  command: pytest -v -s engine tokenization test_sequence.py test_config.py test_logger.py
 
 - label: Entrypoints Test
-  command: pytest -v -s entrypoints
+  mirror_hardwares: [amd]
 
-- label: Kernels Test
-  command: pytest -v -s kernels
-  soft_fail: true
+  commands:
+  - pytest -v -s test_inputs.py
+  - pytest -v -s entrypoints -m llm
+  - pytest -v -s entrypoints -m openai
+
+- label: Examples Test
+  working_dir: "/vllm-workspace/examples"
+  mirror_hardwares: [amd]
+  commands:
+    # install aws cli for llava_example.py
+    # install tensorizer for tensorize_vllm_model.py
+    - pip install awscli tensorizer
+    - python3 offline_inference.py
+    - python3 offline_inference_with_prefix.py
+    - python3 llm_engine_example.py
+    - python3 llava_example.py
+    - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
+
+- label: Kernels Test %N
+  #mirror_hardwares: [amd]
+  command: pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 4
 
 - label: Models Test
+  #mirror_hardwares: [amd]
   commands:
-    - pytest -v -s models --forked
-  soft_fail: true
+    - bash ../.buildkite/download-images.sh
+    - pytest -v -s models --ignore=models/test_llava.py
+
+- label: Llava Test
+  mirror_hardwares: [amd]
+  commands:
+    - bash ../.buildkite/download-images.sh
+    - pytest -v -s models/test_llava.py
 
 - label: Prefix Caching Test
+  mirror_hardwares: [amd]
   commands:
     - pytest -v -s prefix_caching
 
 - label: Samplers Test
-  command: pytest -v -s samplers --forked
+  #mirror_hardwares: [amd]
+  command: pytest -v -s samplers
+
+- label: LogitsProcessor Test
+  mirror_hardwares: [amd]
+  command: pytest -v -s test_logits_processor.py
+
+- label: Utils Test
+  command: pytest -v -s test_utils.py
 
 - label: Worker Test
+  mirror_hardwares: [amd]
   command: pytest -v -s worker
 
-- label: LoRA Test
-  command: pytest -v -s lora --forked
+- label: Speculative decoding tests
+  #mirror_hardwares: [amd]
+  command: pytest -v -s spec_decode
+
+- label: LoRA Test %N
+  #mirror_hardwares: [amd]
+  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py
+  parallelism: 4
+
+- label: LoRA Long Context (Distributed)
+  #mirror_hardwares: [amd]
+  num_gpus: 4
+  # This test runs llama 13B, so it is required to run on 4 GPUs.
+  commands:
+    # Temporarily run this way because we cannot clean up GPU mem usage
+    # for multi GPU tests.
+    # TODO(sang): Fix it.
+    - pytest -v -s lora/test_long_context.py::test_rotary_emb_replaced
+    - pytest -v -s lora/test_long_context.py::test_batched_rope_kernel
+    - pytest -v -s lora/test_long_context.py::test_self_consistency
+    - pytest -v -s lora/test_long_context.py::test_quality
+    - pytest -v -s lora/test_long_context.py::test_max_len
+
+- label: Tensorizer Test
+  #mirror_hardwares: [amd]
+  command: apt-get install curl libsodium23 && pytest -v -s tensorizer_loader
 
 - label: Metrics Test
+  mirror_hardwares: [amd]
   command: pytest -v -s metrics
 
+- label: Quantization Test
+  #mirror_hardwares: [amd]
+  command: pytest -v -s quantization
+
 - label: Benchmarks
   working_dir: "/vllm-workspace/.buildkite"
+  mirror_hardwares: [amd]
   commands:
   - pip install aiohttp
   - bash run-benchmarks.sh
 
 - label: Documentation Build
-  working_dir: "/vllm-workspace/docs"
+  working_dir: "/vllm-workspace/test_docs/docs"
   no_gpu: True
   commands:
   - pip install -r requirements-docs.txt

.buildkite/test-template.j2

@@ -4,7 +4,7 @@
 
 steps:
   - label: ":docker: build image"
-    commands:
+    commands: 
       - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."
       - "docker push {{ docker_image }}"
     env:
@@ -13,20 +13,55 @@ steps:
       automatic:
         - exit_status: -1  # Agent was lost
           limit: 5
+        - exit_status: -10  # Agent was lost
+          limit: 5
   - wait
 
+  - group: "AMD Tests"
+    depends_on: ~
+    steps:
+    {% for step in steps %}
+    {% if step.mirror_hardwares and "amd" in step.mirror_hardwares %}
+      - label: "AMD: {{ step.label }}"
+        agents:
+          queue: amd
+        command: bash .buildkite/run-amd-test.sh "cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" ; ")) | safe }}"
+        env:
+          DOCKER_BUILDKIT: "1"
+    {% endif %}
+    {% endfor %}
+
+  - label: "Neuron Test"
+    depends_on: ~
+    agents:
+      queue: neuron
+    command: bash .buildkite/run-neuron-test.sh
+    soft_fail: true
+
+  - label: "Intel Test"
+    depends_on: ~
+    command: bash .buildkite/run-cpu-test.sh
+
   {% for step in steps %}
   - label: "{{ step.label }}"
     agents:
       queue: kubernetes
     soft_fail: {{ step.soft_fail or false }}
+    {% if step.parallelism %}
+    parallelism: {{ step.parallelism }}
+    {% endif %}
     retry:
       automatic:
         - exit_status: -1  # Agent was lost
           limit: 5
+        - exit_status: -10  # Agent was lost
+          limit: 5
     plugins:
       - kubernetes:
           podSpec:
+            {% if step.num_gpus %}
+            priorityClassName: gpu-priority-cls-{{ step.num_gpus }}
+            {% endif %}
             volumes:
               - name: dshm
                 emptyDir:
@@ -45,6 +80,8 @@ steps:
                     nvidia.com/gpu: "{{ step.num_gpus or default_num_gpu }}"
                 {% endif %}
                 env:
+                  - name: VLLM_USAGE_SOURCE
+                    value: ci-test
                   - name: HF_TOKEN
                     valueFrom:
                       secretKeyRef:

.clang-format

@@ -0,0 +1,26 @@
+BasedOnStyle: Google
+UseTab: Never
+IndentWidth: 2
+ColumnLimit: 80
+
+# Force pointers to the type for C++.
+DerivePointerAlignment: false
+PointerAlignment: Left
+
+# Reordering #include statements can (and currently will) introduce errors
+SortIncludes: false
+
+# Style choices
+AlignConsecutiveAssignments: false
+AlignConsecutiveDeclarations: false
+IndentPPDirectives: BeforeHash
+
+IncludeCategories:
+  - Regex:           '^<'
+    Priority:        4
+  - Regex:           '^"(llvm|llvm-c|clang|clang-c|mlir|mlir-c)/'
+    Priority:        3
+  - Regex:           '^"(qoda|\.\.)/'
+    Priority:        2
+  - Regex:           '.*'
+    Priority:        1

.github/ISSUE_TEMPLATE/100-documentation.yml

@@ -0,0 +1,22 @@
+name: 📚 Documentation
+description: Report an issue related to https://docs.vllm.ai/
+title: "[Doc]: "
+labels: ["documentation"]
+
+body:
+- type: textarea
+  attributes:
+    label: 📚 The doc issue
+    description: >
+      A clear and concise description of what content in https://docs.vllm.ai/ is an issue.
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Suggest a potential alternative/fix
+    description: >
+      Tell us how we could improve the documentation in this regard.
+- type: markdown
+  attributes:
+    value: >
+      Thanks for contributing 🎉!

.github/ISSUE_TEMPLATE/200-installation.yml

@@ -0,0 +1,40 @@
+name: 🛠️ Installation
+description: Report an issue here when you hit errors during installation.
+title: "[Installation]: "
+labels: ["installation"]
+
+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: Your current environment
+    description: |
+      Please run the following and paste the output below.
+      ```sh
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      # For security purposes, please feel free to check the contents of collect_env.py before running it.
+      python collect_env.py
+      ```
+      It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
+    value: |
+      ```text
+      The output of `python collect_env.py`
+      ```
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: How you are installing vllm
+    description: |
+      Paste the full command you are trying to execute.
+    value: |
+      ```sh
+      pip install -vvv vllm
+      ```
+- type: markdown
+  attributes:
+    value: >
+      Thanks for contributing 🎉!

.github/ISSUE_TEMPLATE/300-usage.yml

@@ -0,0 +1,38 @@
+name: 💻 Usage
+description: Raise an issue here if you don't know how to use vllm.
+title: "[Usage]: "
+labels: ["usage"]
+
+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: Your current environment
+    description: |
+      Please run the following and paste the output below.
+      ```sh
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      # For security purposes, please feel free to check the contents of collect_env.py before running it.
+      python collect_env.py
+      ```
+      It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
+    value: |
+      ```text
+      The output of `python collect_env.py`
+      ```
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: How would you like to use vllm
+    description: |
+      A detailed description of how you want to use vllm.
+    value: |
+      I want to run inference of a [specific model](put link here). I don't know how to integrate it with vllm.
+- type: markdown
+  attributes:
+    value: >
+      Thanks for contributing 🎉!

.github/ISSUE_TEMPLATE/400-bug report.yml

@@ -0,0 +1,86 @@
+name: 🐛 Bug report
+description: Raise an issue here if you find a bug.
+title: "[Bug]: "
+labels: ["bug"]
+
+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: Your current environment
+    description: |
+      Please run the following and paste the output below.
+      ```sh
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      # For security purposes, please feel free to check the contents of collect_env.py before running it.
+      python collect_env.py
+      ```
+      It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
+    value: |
+      ```text
+      The output of `python collect_env.py`
+      ```
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: 🐛 Describe the bug
+    description: |
+      Please provide a clear and concise description of what the bug is.
+
+      If relevant, add a minimal example so that we can reproduce the error by running the code. It is very important for the snippet to be as succinct (minimal) as possible, so please take time to trim down any irrelevant code to help us debug efficiently. We are going to copy-paste your code and we expect to get the same result as you did: avoid any external data, and include the relevant imports, etc. For example:
+
+      ```python
+      from vllm import LLM, SamplingParams
+
+      prompts = [
+          "Hello, my name is",
+          "The president of the United States is",
+          "The capital of France is",
+          "The future of AI is",
+      ]
+      sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
+
+      llm = LLM(model="facebook/opt-125m")
+
+      outputs = llm.generate(prompts, sampling_params)
+
+      # Print the outputs.
+      for output in outputs:
+          prompt = output.prompt
+          generated_text = output.outputs[0].text
+          print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
+      ```
+
+      If the code is too long (hopefully, it isn't), feel free to put it in a public gist and link it in the issue: https://gist.github.com.
+
+      Please also paste or describe the results you observe instead of the expected results. If you observe an error, please paste the error message including the **full** traceback of the exception. It may be relevant to wrap error messages in ```` ```triple quotes blocks``` ````.
+
+      Please set the environment variable `export VLLM_LOGGING_LEVEL=DEBUG` to turn on more logging to help debugging potential issues.
+
+      If you experienced crashes or hangs, it would be helpful to run vllm with `export VLLM_TRACE_FUNCTION=1` . All the function calls in vllm will be recorded. Inspect these log files, and tell which function crashes or hangs.
+    placeholder: |
+      A clear and concise description of what the bug is.
+
+      ```python
+      # Sample code to reproduce the problem
+      ```
+
+      ```
+      The error message you got, with the full traceback.
+      ```
+  validations:
+    required: true
+- type: markdown
+  attributes:
+    value: >
+      ⚠️ Please separate bugs of `transformers` implementation or usage from bugs of `vllm`. If you think anything is wrong with the models' output:
+
+      - Try the counterpart of `transformers` first. If the error appears, please go to [their issues](https://github.com/huggingface/transformers/issues?q=is%3Aissue+is%3Aopen+sort%3Aupdated-desc).
+
+      - If the error only appears in vllm, please provide the detailed script of how you run `transformers` and `vllm`, also highlight the difference and what you expect.
+
+      Thanks for contributing 🎉!

.github/ISSUE_TEMPLATE/500-feature request.yml

@@ -0,0 +1,31 @@
+name: 🚀 Feature request
+description: Submit a proposal/request for a new vllm feature
+title: "[Feature]: "
+labels: ["feature request"]
+
+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: 🚀 The feature, motivation and pitch
+    description: >
+      A clear and concise description of the feature proposal. Please outline the motivation for the proposal. Is your feature request related to a specific problem? e.g., *"I'm working on X and would like Y to be possible"*. If this is related to another GitHub issue, please link here too.
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Alternatives
+    description: >
+      A description of any alternative solutions or features you've considered, if any.
+- type: textarea
+  attributes:
+    label: Additional context
+    description: >
+      Add any other context or screenshots about the feature request.
+- type: markdown
+  attributes:
+    value: >
+      Thanks for contributing 🎉!

.github/ISSUE_TEMPLATE/600-new model.yml

@@ -0,0 +1,33 @@
+name: 🤗 Support request for a new model from huggingface
+description: Submit a proposal/request for a new model from huggingface
+title: "[New Model]: "
+labels: ["new model"]
+
+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+).
+
+      #### We also highly recommend you read https://docs.vllm.ai/en/latest/models/adding_model.html first to understand how to add a new model.
+- type: textarea
+  attributes:
+    label: The model to consider.
+    description: >
+      A huggingface url, pointing to the model, e.g. https://huggingface.co/openai-community/gpt2 .
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: The closest model vllm already supports.
+    description: >
+      Here is the list of models already supported by vllm: https://github.com/vllm-project/vllm/tree/main/vllm/model_executor/models . Which model is the most similar to the model you want to add support for?
+- type: textarea
+  attributes:
+    label: What's your difficulty of supporting the model you want?
+    description: >
+      For example, any new operators or new architecture?
+- type: markdown
+  attributes:
+    value: >
+      Thanks for contributing 🎉!

.github/ISSUE_TEMPLATE/700-performance discussion.yml

@@ -0,0 +1,52 @@
+name: ⚡ Discussion on the performance of vllm
+description: Submit a proposal/discussion about the performance of vllm
+title: "[Performance]: "
+labels: ["performance"]
+
+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: Proposal to improve performance
+    description: >
+      How do you plan to improve vllm's performance?
+  validations:
+    required: false
+- type: textarea
+  attributes:
+    label: Report of performance regression
+    description: >
+      Please provide detailed description of performance comparison to confirm the regression. You may want to run the benchmark script at https://github.com/vllm-project/vllm/tree/main/benchmarks .
+  validations:
+    required: false
+- type: textarea
+  attributes:
+    label: Misc discussion on performance
+    description: >
+      Anything about the performance.
+  validations:
+    required: false
+- type: textarea
+  attributes:
+    label: Your current environment (if you think it is necessary)
+    description: |
+      Please run the following and paste the output below.
+      ```sh
+      wget https://raw.githubusercontent.com/vllm-project/vllm/main/collect_env.py
+      # For security purposes, please feel free to check the contents of collect_env.py before running it.
+      python collect_env.py
+      ```
+      It is suggested to download and execute the latest script, as vllm might frequently update the diagnosis information needed for accurately and quickly responding to issues.
+    value: |
+      ```text
+      The output of `python collect_env.py`
+      ```
+  validations:
+    required: false
+- type: markdown
+  attributes:
+    value: >
+      Thanks for contributing 🎉!

.github/ISSUE_TEMPLATE/750-RFC.yml

@@ -0,0 +1,49 @@
+name: 💬 Request for comments (RFC).
+description: Ask for feedback on major architectural changes or design choices.
+title: "[RFC]: "
+labels: ["RFC"]
+
+body:
+- type: markdown
+  attributes:
+    value: >
+      #### Please take a look at previous [RFCs](https://github.com/vllm-project/vllm/issues?q=label%3ARFC+sort%3Aupdated-desc) for reference.
+- type: textarea
+  attributes:
+    label: Motivation.
+    description: >
+      The motivation of the RFC.
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Proposed Change.
+    description: >
+      The proposed change of the RFC.
+  validations:
+    required: true
+- type: textarea
+  attributes:
+    label: Feedback Period.
+    description: >
+      The feedback period of the RFC. Usually at least one week.
+  validations:
+    required: false
+- type: textarea
+  attributes:
+    label: CC List.
+    description: >
+      The list of people you want to CC.
+  validations:
+    required: false
+- type: textarea
+  attributes:
+    label: Any Other Things.
+    description: >
+      Any other things you would like to mention.
+  validations:
+    required: false
+- type: markdown
+  attributes:
+    value: >
+      Thanks for contributing 🎉!

.github/ISSUE_TEMPLATE/800-misc discussion.yml

@@ -0,0 +1,21 @@
+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 🎉!

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1 @@
+blank_issues_enabled: false

.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,64 @@
+FILL IN THE PR DESCRIPTION HERE
+
+FIX #xxxx (*link existing issues this PR will resolve*)
+
+**BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE**
+
+---
+
+<details>
+<!-- inside this <details> section, markdown rendering does not work, so we use raw html here. -->
+<summary><b> PR Checklist (Click to Expand) </b></summary>
+
+<p>Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.</p>
+
+<h3>PR Title and Classification</h3>
+<p>Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:</p>
+<ul>
+    <li><code>[Bugfix]</code> for bug fixes.</li>
+    <li><code>[CI/Build]</code> for build or continuous integration improvements.</li>
+    <li><code>[Doc]</code> for documentation fixes and improvements.</li>
+    <li><code>[Model]</code> for adding a new model or improving an existing model. Model name should appear in the title.</li>
+    <li><code>[Frontend]</code> For changes on the vLLM frontend (e.g., OpenAI API server, <code>LLM</code> class, etc.) </li>
+    <li><code>[Kernel]</code> for changes affecting CUDA kernels or other compute kernels.</li>
+    <li><code>[Core]</code> for changes in the core vLLM logic (e.g., <code>LLMEngine</code>, <code>AsyncLLMEngine</code>, <code>Scheduler</code>, etc.)</li>
+    <li><code>[Hardware][Vendor]</code> for hardware-specific changes. Vendor name should appear in the prefix (e.g., <code>[Hardware][AMD]</code>).</li>
+    <li><code>[Misc]</code> for PRs that do not fit the above categories. Please use this sparingly.</li>
+</ul>
+<p><strong>Note:</strong> If the PR spans more than one category, please include all relevant prefixes.</p>
+
+<h3>Code Quality</h3>
+
+<p>The PR need to meet the following code quality standards:</p>
+
+<ul>
+    <li>We adhere to <a href="https://google.github.io/styleguide/pyguide.html">Google Python style guide</a> and <a href="https://google.github.io/styleguide/cppguide.html">Google C++ style guide</a>.</li>
+    <li>Pass all linter checks. Please use <a href="https://github.com/vllm-project/vllm/blob/main/format.sh"><code>format.sh</code></a> to format your code.</li>
+    <li>The code need to be well-documented to ensure future contributors can easily understand the code.</li>
+    <li>Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.</li>
+    <li>Please add documentation to <code>docs/source/</code> if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.</li>
+</ul>
+
+<h3>Notes for Large Changes</h3>
+<p>Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with <code>rfc-required</code> and might not go through the PR.</p>
+
+<h3>What to Expect for the Reviews</h3>
+
+<p>The goal of the vLLM team is to be a <i>transparent reviewing machine</i>. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process: </p>
+
+<ul>
+    <li> After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.</li>
+    <li> After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.</li>
+    <li> After the review, the reviewer will put an <code> action-required</code> label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.</li>
+    <li> Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.
+ </li>
+</ul>
+
+<h3>Thank You</h3>
+
+<p> Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone! </p>
+
+
+</details>
+
+

.github/workflows/clang-format.yml

@@ -0,0 +1,42 @@
+name: clang-format
+
+on:
+  # Trigger the workflow on push or pull request,
+  # but only for the main branch
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+
+jobs:
+  clang-format:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.11"]
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install clang-format==18.1.5
+    - name: Running clang-format
+      run: |
+        EXCLUDES=(
+            'csrc/moe/topk_softmax_kernels.cu'
+            'csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu'
+            'csrc/punica/bgmv/bgmv_config.h'
+            'csrc/punica/bgmv/bgmv_impl.cuh'
+            'csrc/punica/bgmv/vec_dtypes.cuh'
+            'csrc/punica/punica_ops.cu'
+            'csrc/punica/type_convert.h'
+        )
+        find csrc/ \( -name '*.h' -o -name '*.cpp' -o -name '*.cu' -o -name '*.cuh' \) -print \
+            | grep -vFf <(printf "%s\n" "${EXCLUDES[@]}") \
+            | xargs clang-format --dry-run --Werror

.github/workflows/mypy.yaml

@@ -0,0 +1,50 @@
+name: mypy
+
+on:
+  # Trigger the workflow on push or pull request,
+  # but only for the main branch
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+
+jobs:
+  ruff:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.8", "3.9", "3.10", "3.11"]
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install mypy==1.9.0
+        pip install types-setuptools
+        pip install types-PyYAML
+        pip install types-requests
+        pip install types-setuptools
+    - name: Mypy
+      run: |
+        mypy vllm/attention --config-file pyproject.toml
+        mypy vllm/core --config-file pyproject.toml
+        mypy vllm/distributed --config-file pyproject.toml
+        mypy vllm/entrypoints --config-file pyproject.toml
+        mypy vllm/executor --config-file pyproject.toml
+        mypy vllm/usage --config-file pyproject.toml
+        mypy vllm/*.py --config-file pyproject.toml
+        mypy vllm/transformers_utils --config-file pyproject.toml
+        mypy vllm/engine  --config-file pyproject.toml
+        mypy vllm/worker --config-file pyproject.toml
+        mypy vllm/spec_decode --config-file pyproject.toml
+        mypy vllm/model_executor  --config-file pyproject.toml
+        mypy vllm/lora --config-file pyproject.toml
+        mypy vllm/logging --config-file pyproject.toml
+        mypy vllm/model_executor --config-file pyproject.toml
+

.github/workflows/publish.yml

@@ -49,13 +49,19 @@ jobs:
       matrix:
           os: ['ubuntu-20.04']
           python-version: ['3.8', '3.9', '3.10', '3.11']
-          pytorch-version: ['2.1.2']  # Must be the most recent version that meets requirements.txt.
+          pytorch-version: ['2.3.0']  # Must be the most recent version that meets requirements-cuda.txt.
           cuda-version: ['11.8', '12.1']
 
     steps:
       - name: Checkout
         uses: actions/checkout@v3
 
+      - name: Setup ccache
+        uses: hendrikmuhs/ccache-action@v1.2
+        with:
+          create-symlink: true
+          key: ${{ github.job }}-${{ matrix.python-version }}-${{ matrix.cuda-version }}
+
       - name: Set up Linux Env
         if: ${{ runner.os == 'Linux' }}
         run: |
@@ -76,6 +82,8 @@ jobs:
 
       - name: Build wheel
         shell: bash
+        env:
+          CMAKE_BUILD_TYPE: Release # do not compile with debug symbol to reduce wheel size
         run: |
           bash -x .github/workflows/scripts/build.sh ${{ matrix.python-version }} ${{ matrix.cuda-version }}
           wheel_name=$(ls dist/*whl | xargs -n 1 basename)

.github/workflows/ruff.yml

@@ -15,7 +15,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.10"]
+        python-version: ["3.8", "3.9", "3.10", "3.11"]
     steps:
     - uses: actions/checkout@v2
     - name: Set up Python ${{ matrix.python-version }}
@@ -25,10 +25,13 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install ruff==0.1.5 codespell==2.2.6 tomli==2.0.1
+        pip install ruff==0.1.5 codespell==2.2.6 tomli==2.0.1 isort==5.13.2
     - name: Analysing the code with ruff
       run: |
-        ruff vllm tests
+        ruff .
     - name: Spelling check with codespell
       run: |
-         codespell --toml pyproject.toml
+        codespell --toml pyproject.toml
+    - name: Run isort
+      run: |
+        isort . --check-only

.github/workflows/scripts/build.sh

@@ -9,12 +9,13 @@ LD_LIBRARY_PATH=${cuda_home}/lib64:$LD_LIBRARY_PATH
 
 # Install requirements
 $python_executable -m pip install wheel packaging
-$python_executable -m pip install -r requirements.txt
+$python_executable -m pip install -r requirements-cuda.txt
 
 # Limit the number of parallel jobs to avoid OOM
 export MAX_JOBS=1
 # Make sure punica is built for the release (for LoRA)
 export VLLM_INSTALL_PUNICA_KERNELS=1
-
+# Make sure release wheels are built for the following architectures
+export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0+PTX"
 # Build
 $python_executable setup.py bdist_wheel --dist-dir=dist

.github/workflows/scripts/create_release.js

@@ -8,7 +8,7 @@ module.exports = async (github, context, core) => {
 			generate_release_notes: true,
 			name: process.env.RELEASE_TAG,
 			owner: context.repo.owner,
-			prerelease: false,
+			prerelease: true,
 			repo: context.repo.repo,
 			tag_name: process.env.RELEASE_TAG,
 		});

.github/workflows/yapf.yml

@@ -14,7 +14,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: ["3.10"]
+        python-version: ["3.8", "3.9", "3.10", "3.11"]
     steps:
     - uses: actions/checkout@v2
     - name: Set up Python ${{ matrix.python-version }}

.gitignore

@@ -70,6 +70,8 @@ instance/
 
 # Sphinx documentation
 docs/_build/
+docs/source/getting_started/examples/*.rst
+!**/*.template.rst
 
 # PyBuilder
 .pybuilder/
@@ -181,6 +183,7 @@ _build/
 # hip files generated by PyTorch
 *.hip
 *_hip*
+hip_compat.h
 
 # Benchmark dataset
 *.json

.yapfignore

@@ -0,0 +1 @@
+collect_env.py

CMakeLists.txt

@@ -0,0 +1,327 @@
+cmake_minimum_required(VERSION 3.21)
+
+project(vllm_extensions LANGUAGES CXX)
+
+option(VLLM_TARGET_DEVICE "Target device backend for vLLM" "cuda")
+
+message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
+message(STATUS "Target device: ${VLLM_TARGET_DEVICE}")
+
+include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)
+
+#
+# Supported python versions.  These versions will be searched in order, the
+# first match will be selected.  These should be kept in sync with setup.py.
+#
+set(PYTHON_SUPPORTED_VERSIONS "3.8" "3.9" "3.10" "3.11")
+
+# Supported NVIDIA architectures.
+set(CUDA_SUPPORTED_ARCHS "7.0;7.5;8.0;8.6;8.9;9.0")
+
+# Supported AMD GPU architectures.
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100")
+
+#
+# Supported/expected torch versions for CUDA/ROCm.
+#
+# Currently, having an incorrect pytorch version results in a warning
+# rather than an error.
+#
+# 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
+#
+set(TORCH_SUPPORTED_VERSION_CUDA "2.3.0")
+set(TORCH_SUPPORTED_VERSION_ROCM_5X "2.0.1")
+set(TORCH_SUPPORTED_VERSION_ROCM_6X "2.1.1")
+
+#
+# Try to find python package with an executable that exactly matches
+# `VLLM_PYTHON_EXECUTABLE` and is one of the supported versions.
+#
+if (VLLM_PYTHON_EXECUTABLE)
+  find_python_from_executable(${VLLM_PYTHON_EXECUTABLE} "${PYTHON_SUPPORTED_VERSIONS}")
+else()
+  message(FATAL_ERROR
+    "Please set VLLM_PYTHON_EXECUTABLE to the path of the desired python version"
+    " before running cmake configure.")
+endif()
+
+#
+# Update cmake's `CMAKE_PREFIX_PATH` with torch location.
+#
+append_cmake_prefix_path("torch" "torch.utils.cmake_prefix_path")
+
+# Ensure the 'nvcc' command is in the PATH
+find_program(NVCC_EXECUTABLE nvcc)
+if (CUDA_FOUND AND NOT NVCC_EXECUTABLE)
+    message(FATAL_ERROR "nvcc not found")
+endif()
+
+#
+# Import torch cmake configuration.
+# Torch also imports CUDA (and partially HIP) languages with some customizations,
+# so there is no need to do this explicitly with check_language/enable_language,
+# etc.
+#
+find_package(Torch REQUIRED)
+
+#
+# Normally `torch.utils.cpp_extension.CUDAExtension` would add
+# `libtorch_python.so` for linking against an extension. Torch's cmake
+# configuration does not include this library (presumably since the cmake
+# config is used for standalone C++ binaries that link against torch).
+# The `libtorch_python.so` library defines some of the glue code between
+# torch/python via pybind and is required by VLLM extensions for this
+# reason. So, add it by manually with `find_library` using torch's
+# installed library path.
+#
+find_library(torch_python_LIBRARY torch_python PATHS
+  "${TORCH_INSTALL_PREFIX}/lib")
+
+#
+# Forward the non-CUDA device extensions to external CMake scripts.
+#
+if (NOT VLLM_TARGET_DEVICE STREQUAL "cuda" AND
+    NOT VLLM_TARGET_DEVICE STREQUAL "rocm")
+    if (VLLM_TARGET_DEVICE STREQUAL "cpu")
+        include(${CMAKE_CURRENT_LIST_DIR}/cmake/cpu_extension.cmake)
+    else()
+        message(FATAL_ERROR "Unsupported vLLM target device: ${VLLM_TARGET_DEVICE}")
+    endif()
+    return()
+endif()
+
+#
+# Set up GPU language and check the torch version and warn if it isn't
+# what is expected.
+#
+if (NOT HIP_FOUND AND CUDA_FOUND)
+  set(VLLM_GPU_LANG "CUDA")
+
+  if (NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_CUDA})
+    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_CUDA} "
+      "expected for CUDA build, saw ${Torch_VERSION} instead.")
+  endif()
+elseif(HIP_FOUND)
+  set(VLLM_GPU_LANG "HIP")
+
+  # Importing torch recognizes and sets up some HIP/ROCm configuration but does
+  # not let cmake recognize .hip files. In order to get cmake to understand the
+  # .hip extension automatically, HIP must be enabled explicitly.
+  enable_language(HIP)
+
+  # ROCm 5.x
+  if (ROCM_VERSION_DEV_MAJOR EQUAL 5 AND
+      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM_5X})
+    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_ROCM_5X} "
+      "expected for ROCMm 5.x build, saw ${Torch_VERSION} instead.")
+  endif()
+
+  # ROCm 6.x
+  if (ROCM_VERSION_DEV_MAJOR EQUAL 6 AND
+      NOT Torch_VERSION VERSION_EQUAL ${TORCH_SUPPORTED_VERSION_ROCM_6X})
+    message(WARNING "Pytorch version ${TORCH_SUPPORTED_VERSION_ROCM_6X} "
+      "expected for ROCMm 6.x build, saw ${Torch_VERSION} instead.")
+  endif()
+else()
+  message(FATAL_ERROR "Can't find CUDA or HIP installation.")
+endif()
+
+#
+# Override the GPU architectures detected by cmake/torch and filter them by
+# the supported versions for the current language.
+# The final set of arches is stored in `VLLM_GPU_ARCHES`.
+#
+override_gpu_arches(VLLM_GPU_ARCHES
+  ${VLLM_GPU_LANG}
+  "${${VLLM_GPU_LANG}_SUPPORTED_ARCHS}")
+
+#
+# Query torch for additional GPU compilation flags for the given
+# `VLLM_GPU_LANG`.
+# The final set of arches is stored in `VLLM_GPU_FLAGS`.
+#
+get_torch_gpu_compiler_flags(VLLM_GPU_FLAGS ${VLLM_GPU_LANG})
+
+#
+# Set nvcc parallelism.
+#
+if(NVCC_THREADS AND VLLM_GPU_LANG STREQUAL "CUDA")
+  list(APPEND VLLM_GPU_FLAGS "--threads=${NVCC_THREADS}")
+endif()
+
+#
+# Define extension targets
+#
+
+#
+# _C extension
+#
+
+set(VLLM_EXT_SRC
+  "csrc/cache_kernels.cu"
+  "csrc/attention/attention_kernels.cu"
+  "csrc/pos_encoding_kernels.cu"
+  "csrc/activation_kernels.cu"
+  "csrc/layernorm_kernels.cu"
+  "csrc/quantization/squeezellm/quant_cuda_kernel.cu"
+  "csrc/quantization/gptq/q_gemm.cu"
+  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
+  "csrc/quantization/fp8/common.cu"
+  "csrc/cuda_utils_kernels.cu"
+  "csrc/moe_align_block_size_kernels.cu"
+  "csrc/pybind.cpp")
+
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  include(FetchContent)
+  SET(CUTLASS_ENABLE_HEADERS_ONLY=ON)
+  FetchContent_Declare(
+        cutlass
+        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
+        # CUTLASS 3.5.0
+        GIT_TAG 7d49e6c7e2f8896c47f586706e67e1fb215529dc
+  )
+  FetchContent_MakeAvailable(cutlass)
+
+  list(APPEND VLLM_EXT_SRC
+    "csrc/quantization/aqlm/gemm_kernels.cu"
+    "csrc/quantization/awq/gemm_kernels.cu"
+    "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
+    "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
+    "csrc/quantization/gptq_marlin/gptq_marlin.cu"
+    "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
+    "csrc/custom_all_reduce.cu"
+    "csrc/quantization/cutlass_w8a8/scaled_mm_dq_entry.cu"
+    "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c2x.cu"
+    "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu")
+
+  #
+  # The CUTLASS kernels for Hopper require sm90a to be enabled.
+  # This is done via the below gencode option, BUT that creates kernels for both sm90 and sm90a.
+  # That adds an extra 17MB to compiled binary, so instead we selectively enable it.
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0)
+    set_source_files_properties(
+          "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu"
+          PROPERTIES
+          COMPILE_FLAGS
+          "-gencode arch=compute_90a,code=sm_90a")
+  endif()
+
+endif()
+
+define_gpu_extension_target(
+  _C
+  DESTINATION vllm
+  LANGUAGE ${VLLM_GPU_LANG}
+  SOURCES ${VLLM_EXT_SRC}
+  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+  ARCHITECTURES ${VLLM_GPU_ARCHES}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
+  WITH_SOABI)
+
+#
+# _moe_C extension
+#
+
+set(VLLM_MOE_EXT_SRC
+  "csrc/moe/moe_ops.cpp"
+  "csrc/moe/topk_softmax_kernels.cu")
+
+define_gpu_extension_target(
+  _moe_C
+  DESTINATION vllm
+  LANGUAGE ${VLLM_GPU_LANG}
+  SOURCES ${VLLM_MOE_EXT_SRC}
+  COMPILE_FLAGS ${VLLM_GPU_FLAGS}
+  ARCHITECTURES ${VLLM_GPU_ARCHES}
+  WITH_SOABI)
+
+#
+# _punica_C extension
+#
+
+set(VLLM_PUNICA_EXT_SRC
+  "csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu"
+  "csrc/punica/bgmv/bgmv_bf16_fp32_bf16.cu"
+  "csrc/punica/bgmv/bgmv_fp16_fp16_fp16.cu"
+  "csrc/punica/bgmv/bgmv_fp16_fp32_fp16.cu"
+  "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
+  "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
+  "csrc/punica/punica_ops.cu"
+  "csrc/punica/punica_pybind.cpp")
+
+#
+# Copy GPU compilation flags+update for punica
+#
+set(VLLM_PUNICA_GPU_FLAGS ${VLLM_GPU_FLAGS})
+list(REMOVE_ITEM VLLM_PUNICA_GPU_FLAGS
+  "-D__CUDA_NO_HALF_OPERATORS__"
+  "-D__CUDA_NO_HALF_CONVERSIONS__"
+  "-D__CUDA_NO_BFLOAT16_CONVERSIONS__"
+  "-D__CUDA_NO_HALF2_OPERATORS__")
+
+#
+# Filter out CUDA architectures < 8.0 for punica.
+#
+if (${VLLM_GPU_LANG} STREQUAL "CUDA")
+  set(VLLM_PUNICA_GPU_ARCHES)
+  foreach(ARCH ${VLLM_GPU_ARCHES})
+    string_to_ver(CODE_VER ${ARCH})
+    if (CODE_VER GREATER_EQUAL 8.0)
+      list(APPEND VLLM_PUNICA_GPU_ARCHES ${ARCH})
+    endif()
+  endforeach()
+  message(STATUS "Punica target arches: ${VLLM_PUNICA_GPU_ARCHES}")
+elseif(${VLLM_GPU_LANG} STREQUAL "HIP")
+  set(VLLM_PUNICA_GPU_ARCHES ${VLLM_GPU_ARCHES})
+  message(STATUS "Punica target arches: ${VLLM_PUNICA_GPU_ARCHES}")
+endif()
+
+if (VLLM_PUNICA_GPU_ARCHES)
+  define_gpu_extension_target(
+    _punica_C
+    DESTINATION vllm
+    LANGUAGE ${VLLM_GPU_LANG}
+    SOURCES ${VLLM_PUNICA_EXT_SRC}
+    COMPILE_FLAGS ${VLLM_PUNICA_GPU_FLAGS}
+    ARCHITECTURES ${VLLM_PUNICA_GPU_ARCHES}
+    WITH_SOABI)
+else()
+  message(WARNING "Unable to create _punica_C target because none of the "
+    "requested architectures (${VLLM_GPU_ARCHES}) are supported, i.e. >= 8.0")
+endif()
+
+#
+# Add the `default` target which detects which extensions should be
+# built based on platform/architecture.  This is the same logic that
+# setup.py uses to select which extensions should be built and should
+# be kept in sync.
+#
+# The `default` target makes direct use of cmake easier since knowledge
+# of which extensions are supported has been factored in, e.g.
+#
+# mkdir build && cd build
+# cmake -G Ninja -DVLLM_PYTHON_EXECUTABLE=`which python3` -DCMAKE_LIBRARY_OUTPUT_DIRECTORY=../vllm ..
+# cmake --build . --target default
+#
+add_custom_target(default)
+
+if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
+  message(STATUS "Enabling C extension.")
+  add_dependencies(default _C)
+
+  # Enable punica if -DVLLM_INSTALL_PUNICA_KERNELS=ON or
+  # VLLM_INSTALL_PUNICA_KERNELS is set in the environment and
+  # there are supported target arches.
+  if (VLLM_PUNICA_GPU_ARCHES AND
+      (ENV{VLLM_INSTALL_PUNICA_KERNELS} OR VLLM_INSTALL_PUNICA_KERNELS))
+    message(STATUS "Enabling punica extension.")
+    add_dependencies(default _punica_C)
+  endif()
+endif()
+
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  message(STATUS "Enabling moe extension.")
+  add_dependencies(default _moe_C)
+endif()

CONTRIBUTING.md

@@ -21,7 +21,6 @@ Express your support on Twitter if vLLM aids you, or simply offer your appreciat
 ### Build from source
 
 ```bash
-pip install -r requirements.txt
 pip install -e .  # This may take several minutes.
 ```
 
@@ -30,6 +29,8 @@ pip install -e .  # This may take several minutes.
 ```bash
 pip install -r requirements-dev.txt
 
+# linting and formatting
+bash format.sh
 # Static type checking
 mypy
 # Unit tests
@@ -45,31 +46,9 @@ pytest tests/
 If you encounter a bug or have a feature request, please check our issues page first to see if someone else has already reported it.
 If not, please file a new issue, providing as much relevant information as possible.
 
-### Coding Style Guide
+### Pull Requests & Code Reviews
 
-In general, we adhere to [Google Python style guide](https://google.github.io/styleguide/pyguide.html) and [Google C++ style guide](https://google.github.io/styleguide/cppguide.html).
-
-We include a formatting script [`format.sh`](./format.sh) to format the code.
-
-### Pull Requests
-
-When submitting a pull request:
-
-1. Make sure your code has been rebased on top of the latest commit on the main branch.
-2. Ensure code is properly formatted by running [`format.sh`](./format.sh).
-3. Include a detailed description of the changes in the pull request.
-Explain why you made the changes you did.
-If your pull request fixes an open issue, please include a reference to it in the description.
-
-### Code Reviews
-
-All submissions, including submissions by project members, require a code review.
-To make the review process as smooth as possible, please:
-
-1. Keep your changes as concise as possible.
-If your pull request involves multiple unrelated changes, consider splitting it into separate pull requests.
-2. Respond to all comments within a reasonable time frame.
-If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.
+Please check the PR checklist in the [PR template](.github/PULL_REQUEST_TEMPLATE.md) for detailed guide for contribution.
 
 ### Thank You
 

Dockerfile

@@ -1,8 +1,13 @@
 # The vLLM Dockerfile is used to construct vLLM image that can be directly used
 # to run the OpenAI compatible server.
 
+# Please update any changes made here to
+# docs/source/dev/dockerfile/dockerfile.rst and
+# docs/source/assets/dev/dockerfile-stages-dependency.png
+
 #################### BASE BUILD IMAGE ####################
-FROM nvidia/cuda:12.1.0-devel-ubuntu22.04 AS dev
+# prepare basic build environment
+FROM nvidia/cuda:12.4.1-devel-ubuntu22.04 AS dev
 
 RUN apt-get update -y \
     && apt-get install -y python3-pip git
@@ -11,23 +16,31 @@ RUN apt-get update -y \
 # https://github.com/pytorch/pytorch/issues/107960 -- hopefully
 # this won't be needed for future versions of this docker image
 # or future versions of triton.
-RUN ldconfig /usr/local/cuda-12.1/compat/
+RUN ldconfig /usr/local/cuda-12.4/compat/
 
 WORKDIR /workspace
 
 # install build and runtime dependencies
-COPY requirements.txt requirements.txt
+COPY requirements-common.txt requirements-common.txt
+COPY requirements-cuda.txt requirements-cuda.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements.txt
+    pip install -r requirements-cuda.txt
 
 # install development dependencies
 COPY requirements-dev.txt requirements-dev.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install -r requirements-dev.txt
+
+# cuda arch list used by torch
+# can be useful for both `dev` and `test`
+# explicitly set the list to avoid issues with torch 2.2
+# see https://github.com/pytorch/pytorch/pull/123243
+ARG torch_cuda_arch_list='7.0 7.5 8.0 8.6 8.9 9.0+PTX'
+ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 #################### BASE BUILD IMAGE ####################
 
 
-#################### EXTENSION BUILD IMAGE ####################
+#################### WHEEL BUILD IMAGE ####################
 FROM dev AS build
 
 # install build dependencies
@@ -35,16 +48,19 @@ COPY requirements-build.txt requirements-build.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
     pip install -r requirements-build.txt
 
-# copy input files
+# install compiler cache to speed up compilation leveraging local or remote caching
+RUN apt-get update -y && apt-get install -y ccache
+
+# files and directories related to build wheels
 COPY csrc csrc
 COPY setup.py setup.py
-COPY requirements.txt requirements.txt
+COPY cmake cmake
+COPY CMakeLists.txt CMakeLists.txt
+COPY requirements-common.txt requirements-common.txt
+COPY requirements-cuda.txt requirements-cuda.txt
 COPY pyproject.toml pyproject.toml
-COPY vllm/__init__.py vllm/__init__.py
+COPY vllm vllm
 
-# cuda arch list used by torch
-ARG torch_cuda_arch_list='7.0 7.5 8.0 8.6 8.9 9.0+PTX'
-ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 # max jobs used by Ninja to build extensions
 ARG max_jobs=2
 ENV MAX_JOBS=${max_jobs}
@@ -54,52 +70,67 @@ ENV NVCC_THREADS=$nvcc_threads
 # make sure punica kernels are built (for LoRA)
 ENV VLLM_INSTALL_PUNICA_KERNELS=1
 
-RUN python3 setup.py build_ext --inplace
+ENV CCACHE_DIR=/root/.cache/ccache
+RUN --mount=type=cache,target=/root/.cache/ccache \
+    --mount=type=cache,target=/root/.cache/pip \
+    python3 setup.py bdist_wheel --dist-dir=dist
+
+# check the size of the wheel, we cannot upload wheels larger than 100MB
+COPY .buildkite/check-wheel-size.py check-wheel-size.py
+RUN python3 check-wheel-size.py dist
+
 #################### EXTENSION Build IMAGE ####################
 
+#################### vLLM installation IMAGE ####################
+# image with vLLM installed
+FROM nvidia/cuda:12.4.1-base-ubuntu22.04 AS vllm-base
+WORKDIR /vllm-workspace
+
+RUN apt-get update -y \
+    && apt-get install -y python3-pip git vim
+
+# Workaround for https://github.com/openai/triton/issues/2507 and
+# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
+# this won't be needed for future versions of this docker image
+# or future versions of triton.
+RUN ldconfig /usr/local/cuda-12.4/compat/
+
+# install vllm wheel first, so that torch etc will be installed
+RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
+    --mount=type=cache,target=/root/.cache/pip \
+    pip install dist/*.whl --verbose
+#################### vLLM installation IMAGE ####################
+
 
 #################### TEST IMAGE ####################
 # image to run unit testing suite
-FROM dev AS test
+# note that this uses vllm installed by `pip`
+FROM vllm-base AS test
 
-# copy pytorch extensions separately to avoid having to rebuild
-# when python code changes
-WORKDIR /vllm-workspace
-# ADD is used to preserve directory structure
 ADD . /vllm-workspace/
-COPY --from=build /workspace/vllm/*.so /vllm-workspace/vllm/
-# ignore build dependencies installation because we are using pre-complied extensions
-RUN rm pyproject.toml
-RUN --mount=type=cache,target=/root/.cache/pip VLLM_USE_PRECOMPILED=1 pip install . --verbose
-#################### TEST IMAGE ####################
 
-
-#################### RUNTIME BASE IMAGE ####################
-# We used base cuda image because pytorch installs its own cuda libraries.
-# However cupy depends on cuda libraries so we had to switch to the runtime image
-# In the future it would be nice to get a container with pytorch and cuda without duplicating cuda
-FROM nvidia/cuda:12.1.0-runtime-ubuntu22.04 AS vllm-base
-
-# libnccl required for ray
-RUN apt-get update -y \
-    && apt-get install -y python3-pip
-
-WORKDIR /workspace
-COPY requirements.txt requirements.txt
+# install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements.txt
-#################### RUNTIME BASE IMAGE ####################
+    pip install -r requirements-dev.txt
 
+# doc requires source code
+# we hide them inside `test_docs/` , so that this source code
+# will not be imported by other tests
+RUN mkdir test_docs
+RUN mv docs test_docs/
+RUN mv vllm test_docs/
+
+#################### TEST IMAGE ####################
 
 #################### OPENAI API SERVER ####################
 # openai api server alternative
 FROM vllm-base AS vllm-openai
+
 # install additional dependencies for openai api server
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install accelerate
+    pip install accelerate hf_transfer modelscope
 
-COPY --from=build /workspace/vllm/*.so /workspace/vllm/
-COPY vllm vllm
+ENV VLLM_USAGE_SOURCE production-docker-image
 
 ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
 #################### OPENAI API SERVER ####################

Dockerfile.cpu

@@ -0,0 +1,22 @@
+# This vLLM Dockerfile is used to construct image that can build and run vLLM on x86 CPU platform.
+
+FROM ubuntu:22.04
+
+RUN apt-get update  -y \
+    && apt-get install -y git wget vim numactl gcc-12 g++-12 python3 python3-pip \
+    && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
+
+RUN pip install --upgrade pip \
+    && pip install wheel packaging ninja setuptools>=49.4.0 numpy
+
+COPY ./ /workspace/vllm
+
+WORKDIR /workspace/vllm
+
+RUN pip install -v -r requirements-cpu.txt --extra-index-url https://download.pytorch.org/whl/cpu
+
+RUN VLLM_TARGET_DEVICE=cpu python3 setup.py install
+
+WORKDIR /workspace/
+
+CMD ["/bin/bash"]

Dockerfile.neuron

@@ -0,0 +1,36 @@
+# default base image
+ARG BASE_IMAGE="763104351884.dkr.ecr.us-west-2.amazonaws.com/pytorch-inference-neuronx:2.1.1-neuronx-py310-sdk2.17.0-ubuntu20.04"
+
+FROM $BASE_IMAGE
+
+RUN echo "Base image is $BASE_IMAGE"
+
+# Install some basic utilities
+RUN apt-get update && apt-get install python3 python3-pip -y
+
+### Mount Point ###
+# When launching the container, mount the code directory to /app
+ARG APP_MOUNT=/app
+VOLUME [ ${APP_MOUNT} ]
+WORKDIR ${APP_MOUNT}
+
+RUN python3 -m pip install --upgrade pip
+RUN python3 -m pip install --no-cache-dir fastapi ninja tokenizers pandas
+RUN python3 -m pip install sentencepiece transformers==4.36.2 -U
+RUN python3 -m pip install transformers-neuronx --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+RUN python3 -m pip install --pre neuronx-cc==2.12.* --extra-index-url=https://pip.repos.neuron.amazonaws.com -U
+
+COPY ./vllm /app/vllm/vllm
+COPY ./setup.py /app/vllm/setup.py
+COPY ./requirements-common.txt /app/vllm/requirements-common.txt
+COPY ./requirements-neuron.txt /app/vllm/requirements-neuron.txt
+
+RUN cd /app/vllm \
+    && python3 -m pip install -U -r requirements-neuron.txt
+
+ENV VLLM_BUILD_WITH_NEURON 1
+RUN cd /app/vllm \
+    && pip install -e . \
+    && cd ..
+
+CMD ["/bin/bash"]

Dockerfile.rocm

@@ -14,7 +14,7 @@ RUN echo "Base image is $BASE_IMAGE"
 ARG FA_GFX_ARCHS="gfx90a;gfx942"
 RUN echo "FA_GFX_ARCHS is $FA_GFX_ARCHS"
 
-ARG FA_BRANCH="3d2b6f5"
+ARG FA_BRANCH="ae7928c"
 RUN echo "FA_BRANCH is $FA_BRANCH"
 
 # whether to build flash-attention
@@ -23,6 +23,9 @@ RUN echo "FA_BRANCH is $FA_BRANCH"
 # In that case, we need to use the python reference attention implementation in vllm
 ARG BUILD_FA="1"
 
+# whether to build triton on rocm
+ARG BUILD_TRITON="1"
+
 # Install some basic utilities
 RUN apt-get update && apt-get install python3 python3-pip -y
 
@@ -43,7 +46,7 @@ RUN apt-get update && apt-get install -y \
 
 ### Mount Point ###
 # When launching the container, mount the code directory to /app
-ARG APP_MOUNT=/app
+ARG APP_MOUNT=/vllm-workspace
 VOLUME [ ${APP_MOUNT} ]
 WORKDIR ${APP_MOUNT}
 
@@ -70,26 +73,42 @@ RUN if [ "$BUILD_FA" = "1" ]; then \
     && cd ..; \
     fi
 
-COPY ./ /app/vllm
-
-RUN python3 -m pip install --upgrade pip
-RUN python3 -m pip install xformers==0.0.23 --no-deps
-
 # Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
 # Manually removed it so that later steps of numpy upgrade can continue
 RUN if [ "$BASE_IMAGE" = "rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1" ]; then \
     rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/; fi
 
-RUN cd /app \
-    && cd vllm \
-    && pip install -U -r requirements-rocm.txt \
-    && if [ "$BUILD_FA" = "1" ]; then \
-       bash patch_xformers.rocm.sh; fi \
-    && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h /app/vllm/rocm_patch/rocm_bf16.patch \
+# build triton
+RUN if [ "$BUILD_TRITON" = "1" ]; then \
+    mkdir -p libs \
+    && cd libs \
+    && pip uninstall -y triton \
+    && git clone https://github.com/ROCm/triton.git \
+    && cd triton/python \
+    && pip3 install . \
+    && cd ../..; \
+    fi
+
+WORKDIR /vllm-workspace
+COPY . .
+
+#RUN python3 -m pip install pynvml # to be removed eventually
+RUN python3 -m pip install --upgrade pip numba
+
+# make sure punica kernels are built (for LoRA)
+ENV VLLM_INSTALL_PUNICA_KERNELS=1
+# Workaround for ray >= 2.10.0
+ENV RAY_EXPERIMENTAL_NOSET_ROCR_VISIBLE_DEVICES=1
+
+ENV VLLM_NCCL_SO_PATH=/opt/rocm/lib/librccl.so
+
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip install -U -r requirements-rocm.txt \
+    && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h ./rocm_patch/rocm_bf16.patch \
     && python3 setup.py install \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_C.cpython-39-x86_64-linux-gnu.so vllm/ \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_punica_C.cpython-39-x86_64-linux-gnu.so vllm/ \
     && cd ..
 
-RUN python3 -m pip install --upgrade pip
-RUN python3 -m pip install --no-cache-dir ray[all]
 
 CMD ["/bin/bash"]

MANIFEST.in

@@ -1,4 +1,10 @@
 include LICENSE
-include requirements.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 *
 recursive-include csrc *

README.md

@@ -16,7 +16,17 @@ Easy, fast, and cheap LLM serving for everyone
 
 ---
 
+**The Fourth vLLM Bay Area Meetup (June 11th 5:30pm-8pm PT)**
+
+We are thrilled to announce our fourth vLLM Meetup!
+The vLLM team will share recent updates and roadmap.
+We will also have vLLM collaborators from BentoML and Cloudflare coming up to the stage to discuss their experience in deploying LLMs with vLLM.
+Please register [here](https://lu.ma/agivllm) and join us!
+
+---
+
 *Latest News* 🔥
+- [2024/04] We hosted [the third vLLM meetup](https://robloxandvllmmeetup2024.splashthat.com/) with Roblox! Please find the meetup slides [here](https://docs.google.com/presentation/d/1A--47JAK4BJ39t954HyTkvtfwn0fkqtsL8NGFuslReM/edit?usp=sharing).
 - [2024/01] We hosted [the second vLLM meetup](https://lu.ma/ygxbpzhl) in SF! Please find the meetup slides [here](https://docs.google.com/presentation/d/12mI2sKABnUw5RBWXDYY-HtHth4iMSNcEoQ10jDQbxgA/edit?usp=sharing).
 - [2024/01] Added ROCm 6.0 support to vLLM.
 - [2023/12] Added ROCm 5.7 support to vLLM.
@@ -52,34 +62,14 @@ vLLM is flexible and easy to use with:
 - (Experimental) Prefix caching support
 - (Experimental) Multi-lora support
 
-vLLM seamlessly supports many Hugging Face models, including the following architectures:
+vLLM seamlessly supports most popular open-source models on HuggingFace, including:
+- Transformer-like LLMs (e.g., Llama)
+- Mixture-of-Expert LLMs (e.g., Mixtral)
+- Multi-modal LLMs (e.g., LLaVA)
 
-- Aquila & Aquila2 (`BAAI/AquilaChat2-7B`, `BAAI/AquilaChat2-34B`, `BAAI/Aquila-7B`, `BAAI/AquilaChat-7B`, etc.)
-- Baichuan & Baichuan2 (`baichuan-inc/Baichuan2-13B-Chat`, `baichuan-inc/Baichuan-7B`, etc.)
-- BLOOM (`bigscience/bloom`, `bigscience/bloomz`, etc.)
-- ChatGLM (`THUDM/chatglm2-6b`, `THUDM/chatglm3-6b`, etc.)
-- DeciLM (`Deci/DeciLM-7B`, `Deci/DeciLM-7B-instruct`, etc.)
-- Falcon (`tiiuae/falcon-7b`, `tiiuae/falcon-40b`, `tiiuae/falcon-rw-7b`, etc.)
-- Gemma (`google/gemma-2b`, `google/gemma-7b`, etc.)
-- GPT-2 (`gpt2`, `gpt2-xl`, etc.)
-- GPT BigCode (`bigcode/starcoder`, `bigcode/gpt_bigcode-santacoder`, etc.)
-- GPT-J (`EleutherAI/gpt-j-6b`, `nomic-ai/gpt4all-j`, etc.)
-- GPT-NeoX (`EleutherAI/gpt-neox-20b`, `databricks/dolly-v2-12b`, `stabilityai/stablelm-tuned-alpha-7b`, etc.)
-- InternLM (`internlm/internlm-7b`, `internlm/internlm-chat-7b`, etc.)
-- InternLM2 (`internlm/internlm2-7b`, `internlm/internlm2-chat-7b`, etc.)
-- LLaMA & LLaMA-2 (`meta-llama/Llama-2-70b-hf`, `lmsys/vicuna-13b-v1.3`, `young-geng/koala`, `openlm-research/open_llama_13b`, etc.)
-- Mistral (`mistralai/Mistral-7B-v0.1`, `mistralai/Mistral-7B-Instruct-v0.1`, etc.)
-- Mixtral (`mistralai/Mixtral-8x7B-v0.1`, `mistralai/Mixtral-8x7B-Instruct-v0.1`, etc.)
-- MPT (`mosaicml/mpt-7b`, `mosaicml/mpt-30b`, etc.)
-- OLMo (`allenai/OLMo-1B`, `allenai/OLMo-7B`, etc.)
-- OPT (`facebook/opt-66b`, `facebook/opt-iml-max-30b`, etc.)
-- Orion (`OrionStarAI/Orion-14B-Base`, `OrionStarAI/Orion-14B-Chat`, etc.)
-- Phi (`microsoft/phi-1_5`, `microsoft/phi-2`, etc.)
-- Qwen (`Qwen/Qwen-7B`, `Qwen/Qwen-7B-Chat`, etc.)
-- Qwen2 (`Qwen/Qwen2-7B-beta`, `Qwen/Qwen-7B-Chat-beta`, etc.)
-- StableLM(`stabilityai/stablelm-3b-4e1t`, `stabilityai/stablelm-base-alpha-7b-v2`, etc.)
-- Starcoder2(`bigcode/starcoder2-3b`, `bigcode/starcoder2-7b`, `bigcode/starcoder2-15b`, etc.)
-- Yi (`01-ai/Yi-6B`, `01-ai/Yi-34B`, etc.)
+Find the full list of supported models [here](https://docs.vllm.ai/en/latest/models/supported_models.html).
+
+## Getting Started
 
 Install vLLM with pip or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):
 
@@ -87,9 +77,7 @@ Install vLLM with pip or [from source](https://vllm.readthedocs.io/en/latest/get
 pip install vllm
 ```
 
-## Getting Started
-
-Visit our [documentation](https://vllm.readthedocs.io/en/latest/) to get started.
+Visit our [documentation](https://vllm.readthedocs.io/en/latest/) to learn more.
 - [Installation](https://vllm.readthedocs.io/en/latest/getting_started/installation.html)
 - [Quickstart](https://vllm.readthedocs.io/en/latest/getting_started/quickstart.html)
 - [Supported Models](https://vllm.readthedocs.io/en/latest/models/supported_models.html)
@@ -99,6 +87,32 @@ Visit our [documentation](https://vllm.readthedocs.io/en/latest/) to get started
 We welcome and value any contributions and collaborations.
 Please check out [CONTRIBUTING.md](./CONTRIBUTING.md) for how to get involved.
 
+## Sponsors
+
+vLLM is a community project. Our compute resources for development and testing are supported by the following organizations. Thank you for your support!
+
+<!-- Note: Please sort them in alphabetical order. -->
+<!-- Note: Please keep these consistent with docs/source/community/sponsors.md -->
+
+- a16z
+- AMD
+- Anyscale
+- AWS
+- Crusoe Cloud
+- Databricks
+- DeepInfra
+- Dropbox
+- Lambda Lab
+- NVIDIA
+- Replicate
+- Roblox
+- RunPod
+- Trainy
+- UC Berkeley
+- UC San Diego
+
+We also have an official fundraising venue through [OpenCollective](https://opencollective.com/vllm). We plan to use the fund to support the development, maintenance, and adoption of vLLM.
+
 ## Citation
 
 If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs/2309.06180):

benchmarks/backend_request_func.py

@@ -1,8 +1,10 @@
 import json
 import os
+import sys
 import time
-from dataclasses import dataclass
-from typing import Optional
+import traceback
+from dataclasses import dataclass, field
+from typing import List, Optional
 
 import aiohttp
 from tqdm.asyncio import tqdm
@@ -25,9 +27,12 @@ class RequestFuncInput:
 class RequestFuncOutput:
     generated_text: str = ""
     success: bool = False
-    latency: float = 0
-    ttft: float = 0
+    latency: float = 0.0
+    ttft: float = 0.0  # Time to first token
+    itl: List[float] = field(
+        default_factory=list)  # List of inter-token latencies
     prompt_len: int = 0
+    error: str = ""
 
 
 async def async_request_tgi(
@@ -53,73 +58,44 @@ async def async_request_tgi(
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
 
-        ttft = 0
+        ttft = 0.0
         st = time.perf_counter()
+        most_recent_timestamp = st
         try:
             async with session.post(url=api_url, json=payload) as response:
                 if response.status == 200:
-                    async for data in response.content.iter_any():
-                        if ttft == 0:
+                    async for chunk_bytes in response.content:
+                        chunk_bytes = chunk_bytes.strip()
+                        if not chunk_bytes:
+                            continue
+
+                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                                              "data:")
+
+                        data = json.loads(chunk)
+                        timestamp = time.perf_counter()
+                        # First token
+                        if ttft == 0.0:
                             ttft = time.perf_counter() - st
                             output.ttft = ttft
-                    output.latency = time.perf_counter() - st
 
-                    body = data.decode("utf-8").lstrip("data:")
-                    output.generated_text = json.loads(body)["generated_text"]
+                        # Decoding phase
+                        else:
+                            output.itl.append(timestamp -
+                                              most_recent_timestamp)
+
+                        most_recent_timestamp = timestamp
+
+                    output.latency = most_recent_timestamp - st
                     output.success = True
+                    output.generated_text = data["generated_text"]
                 else:
+                    output.error = response.reason or ""
                     output.success = False
-        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
-            output.success = False
-
-        if pbar:
-            pbar.update(1)
-        return output
-
-
-async def async_request_vllm(
-    request_func_input: RequestFuncInput,
-    pbar: Optional[tqdm] = None,
-) -> RequestFuncOutput:
-    api_url = request_func_input.api_url
-    assert api_url.endswith("generate")
-
-    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
-        payload = {
-            "prompt": request_func_input.prompt,
-            "n": 1,
-            "best_of": request_func_input.best_of,
-            "use_beam_search": request_func_input.use_beam_search,
-            "temperature": 0.0 if request_func_input.use_beam_search else 1.0,
-            "top_p": 1.0,
-            "max_tokens": request_func_input.output_len,
-            "ignore_eos": True,
-            "stream": True,
-        }
-        output = RequestFuncOutput()
-        output.prompt_len = request_func_input.prompt_len
-
-        ttft = 0
-        st = time.perf_counter()
-        try:
-            async with session.post(url=api_url, json=payload) as response:
-                if response.status == 200:
-                    async for data in response.content.iter_any():
-                        if ttft == 0:
-                            ttft = time.perf_counter() - st
-                            output.ttft = ttft
-                    output.latency = time.perf_counter() - st
-
-                    # When streaming, '\0' is appended to the end of the response.
-                    body = data.decode("utf-8").strip("\0")
-                    output.generated_text = json.loads(
-                        body)["text"][0][len(request_func_input.prompt):]
-                    output.success = True
-
-                else:
-                    output.success = False
-        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
+        except Exception:
             output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
 
         if pbar:
             pbar.update(1)
@@ -146,26 +122,46 @@ async def async_request_trt_llm(
         }
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
-        ttft = 0
 
+        ttft = 0.0
         st = time.perf_counter()
+        most_recent_timestamp = st
         try:
-            async with session.post(url=api_url, json=payload) as resp:
-                if resp.status == 200:
-                    async for data in resp.content.iter_any():
-                        if ttft == 0:
+            async with session.post(url=api_url, json=payload) as response:
+                if response.status == 200:
+                    async for chunk_bytes in response.content:
+                        chunk_bytes = chunk_bytes.strip()
+                        if not chunk_bytes:
+                            continue
+
+                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                                              "data:")
+
+                        data = json.loads(chunk)
+                        output.generated_text += data["text_output"]
+                        timestamp = time.perf_counter()
+                        # First token
+                        if ttft == 0.0:
                             ttft = time.perf_counter() - st
                             output.ttft = ttft
-                    output.latency = time.perf_counter() - st
 
-                    body = data.decode("utf-8").lstrip("data:")
-                    output.generated_text = json.loads(body)["text_output"]
+                        # Decoding phase
+                        else:
+                            output.itl.append(timestamp -
+                                              most_recent_timestamp)
+
+                        most_recent_timestamp = timestamp
+
+                    output.latency = most_recent_timestamp - st
                     output.success = True
 
                 else:
+                    output.error = response.reason or ""
                     output.success = False
-        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
+        except Exception:
             output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
 
         if pbar:
             pbar.update(1)
@@ -181,34 +177,35 @@ async def async_request_deepspeed_mii(
         assert not request_func_input.use_beam_search
 
         payload = {
-            "prompts": request_func_input.prompt,
-            "max_new_tokens": request_func_input.output_len,
-            "ignore_eos": True,
-            "do_sample": True,
-            "temperature":
-            0.01,  # deepspeed-mii does not accept 0.0 temperature.
+            "prompt": request_func_input.prompt,
+            "max_tokens": request_func_input.output_len,
+            "temperature": 0.01,  # deepspeed-mii does not accept 0.0 temp.
             "top_p": 1.0,
         }
         output = RequestFuncOutput()
         output.prompt_len = request_func_input.prompt_len
 
-        # DeepSpeed-MII doesn't support streaming as of Jan 28 2024, will use 0 as placeholder.
-        # https://github.com/microsoft/DeepSpeed-MII/pull/311
+        # NOTE: DeepSpeed-MII doesn't support streaming as of Jan 28 2024,
+        # will use 0 as placeholder.
+        # See https://github.com/microsoft/DeepSpeed-MII/pull/311
         output.ttft = 0
 
         st = time.perf_counter()
         try:
             async with session.post(url=request_func_input.api_url,
-                                    json=payload) as resp:
-                if resp.status == 200:
-                    parsed_resp = await resp.json()
+                                    json=payload) as response:
+                if response.status == 200:
+                    parsed_resp = await response.json()
                     output.latency = time.perf_counter() - st
-                    output.generated_text = parsed_resp[0]["generated_text"]
+                    output.generated_text = parsed_resp["text"][0]
                     output.success = True
                 else:
+                    output.error = response.reason or ""
                     output.success = False
-        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
+        except Exception:
             output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
 
         if pbar:
             pbar.update(1)
@@ -220,7 +217,9 @@ async def async_request_openai_completions(
     pbar: Optional[tqdm] = None,
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
-    assert api_url.endswith("v1/completions")
+    assert api_url.endswith(
+        "v1/completions"
+    ), "OpenAI Completions API URL must end with 'v1/completions'."
 
     async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
         assert not request_func_input.use_beam_search
@@ -240,45 +239,157 @@ async def async_request_openai_completions(
         output.prompt_len = request_func_input.prompt_len
 
         generated_text = ""
-        ttft = 0
+        ttft = 0.0
         st = time.perf_counter()
+        most_recent_timestamp = st
         try:
             async with session.post(url=api_url, json=payload,
                                     headers=headers) as response:
                 if response.status == 200:
-                    async for chunk in response.content:
-                        if ttft == 0:
-                            ttft = time.perf_counter() - st
-                            output.ttft = ttft
-
-                        chunk = chunk.strip()
-                        if not chunk:
+                    async for chunk_bytes in response.content:
+                        chunk_bytes = chunk_bytes.strip()
+                        if not chunk_bytes:
                             continue
 
-                        chunk = chunk.decode("utf-8").lstrip("data: ")
+                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                                              "data: ")
                         if chunk == "[DONE]":
                             latency = time.perf_counter() - st
                         else:
-                            body = json.loads(chunk)
-                            generated_text += body["choices"][0]["text"]
+                            data = json.loads(chunk)
+
+                            if data["choices"][0]["text"]:
+                                timestamp = time.perf_counter()
+                                # First token
+                                if ttft == 0.0:
+                                    ttft = time.perf_counter() - st
+                                    output.ttft = ttft
+
+                                # Decoding phase
+                                # NOTE: Some completion API might have a last
+                                # usage summary response without a token so we
+                                # do not want to include as inter-token-latency
+                                elif data.get("usage", None) is None:
+                                    output.itl.append(timestamp -
+                                                      most_recent_timestamp)
+
+                                most_recent_timestamp = timestamp
+                                generated_text += data["choices"][0]["text"]
 
                     output.generated_text = generated_text
                     output.success = True
                     output.latency = latency
                 else:
+                    output.error = response.reason or ""
                     output.success = False
-        except (aiohttp.ClientOSError, aiohttp.ServerDisconnectedError):
+        except Exception:
             output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
 
     if pbar:
         pbar.update(1)
     return output
 
 
+async def async_request_openai_chat_completions(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith(
+        "v1/chat/completions"
+    ), "OpenAI Chat Completions API URL must end with 'v1/chat/completions'."
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert not request_func_input.use_beam_search
+        payload = {
+            "model": request_func_input.model,
+            "messages": [
+                {
+                    "role": "user",
+                    "content": request_func_input.prompt,
+                },
+            ],
+            "temperature": 0.0,
+            "max_tokens": request_func_input.output_len,
+            "stream": True,
+        }
+        headers = {
+            "Content-Type": "application/json",
+            "Authorization": f"Bearer {os.environ.get('OPENAI_API_KEY')}",
+        }
+
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        generated_text = ""
+        ttft = 0.0
+        st = time.perf_counter()
+        most_recent_timestamp = st
+        try:
+            async with session.post(url=api_url, json=payload,
+                                    headers=headers) as response:
+                if response.status == 200:
+                    async for chunk_bytes in response.content:
+                        chunk_bytes = chunk_bytes.strip()
+                        if not chunk_bytes:
+                            continue
+
+                        chunk = remove_prefix(chunk_bytes.decode("utf-8"),
+                                              "data: ")
+                        if chunk == "[DONE]":
+                            latency = time.perf_counter() - st
+                        else:
+                            timestamp = time.perf_counter()
+                            data = json.loads(chunk)
+
+                            delta = data["choices"][0]["delta"]
+                            if delta.get("content", None):
+                                # First token
+                                if ttft == 0.0:
+                                    ttft = time.perf_counter() - st
+                                    output.ttft = ttft
+
+                                # Decoding phase
+                                else:
+                                    output.itl.append(timestamp -
+                                                      most_recent_timestamp)
+
+                                generated_text += delta["content"]
+
+                            most_recent_timestamp = timestamp
+
+                    output.generated_text = generated_text
+                    output.success = True
+                    output.latency = latency
+                else:
+                    output.error = response.reason or ""
+                    output.success = False
+        except Exception:
+            output.success = False
+            exc_info = sys.exc_info()
+            output.error = "".join(traceback.format_exception(*exc_info))
+
+    if pbar:
+        pbar.update(1)
+    return output
+
+
+# Since vllm must support Python 3.8, we can't use str.removeprefix(prefix)
+# introduced in Python 3.9
+def remove_prefix(text: str, prefix: str) -> str:
+    if text.startswith(prefix):
+        return text[len(prefix):]
+    return text
+
+
 ASYNC_REQUEST_FUNCS = {
     "tgi": async_request_tgi,
-    "vllm": async_request_vllm,
+    "vllm": async_request_openai_completions,
+    "lmdeploy": async_request_openai_completions,
     "deepspeed-mii": async_request_deepspeed_mii,
     "openai": async_request_openai_completions,
+    "openai-chat": async_request_openai_chat_completions,
     "tensorrt-llm": async_request_trt_llm,
 }

benchmarks/benchmark_latency.py

@@ -1,14 +1,17 @@
 """Benchmark the latency of processing a single batch of requests."""
 import argparse
+import json
 import time
 from pathlib import Path
-from typing import Optional
+from typing import List, Optional
 
 import numpy as np
 import torch
 from tqdm import tqdm
 
 from vllm import LLM, SamplingParams
+from vllm.inputs import PromptStrictInputs
+from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
 
 
 def main(args: argparse.Namespace):
@@ -16,17 +19,24 @@ 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(
-        model=args.model,
-        tokenizer=args.tokenizer,
-        quantization=args.quantization,
-        tensor_parallel_size=args.tensor_parallel_size,
-        trust_remote_code=args.trust_remote_code,
-        dtype=args.dtype,
-        enforce_eager=args.enforce_eager,
-        kv_cache_dtype=args.kv_cache_dtype,
-        device=args.device,
-    )
+    llm = LLM(model=args.model,
+              speculative_model=args.speculative_model,
+              num_speculative_tokens=args.num_speculative_tokens,
+              tokenizer=args.tokenizer,
+              quantization=args.quantization,
+              tensor_parallel_size=args.tensor_parallel_size,
+              trust_remote_code=args.trust_remote_code,
+              dtype=args.dtype,
+              enforce_eager=args.enforce_eager,
+              kv_cache_dtype=args.kv_cache_dtype,
+              quantization_param_path=args.quantization_param_path,
+              device=args.device,
+              ray_workers_use_nsight=args.ray_workers_use_nsight,
+              use_v2_block_manager=args.use_v2_block_manager,
+              enable_chunked_prefill=args.enable_chunked_prefill,
+              download_dir=args.download_dir,
+              block_size=args.block_size,
+              gpu_memory_utilization=args.gpu_memory_utilization)
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -40,7 +50,9 @@ def main(args: argparse.Namespace):
     dummy_prompt_token_ids = np.random.randint(10000,
                                                size=(args.batch_size,
                                                      args.input_len))
-    dummy_prompt_token_ids = dummy_prompt_token_ids.tolist()
+    dummy_inputs: List[PromptStrictInputs] = [{
+        "prompt_token_ids": batch
+    } for batch in dummy_prompt_token_ids.tolist()]
 
     def run_to_completion(profile_dir: Optional[str] = None):
         if profile_dir:
@@ -51,13 +63,13 @@ def main(args: argparse.Namespace):
                     ],
                     on_trace_ready=torch.profiler.tensorboard_trace_handler(
                         str(profile_dir))) as p:
-                llm.generate(prompt_token_ids=dummy_prompt_token_ids,
+                llm.generate(dummy_inputs,
                              sampling_params=sampling_params,
                              use_tqdm=False)
             print(p.key_averages())
         else:
             start_time = time.perf_counter()
-            llm.generate(prompt_token_ids=dummy_prompt_token_ids,
+            llm.generate(dummy_inputs,
                          sampling_params=sampling_params,
                          use_tqdm=False)
             end_time = time.perf_counter()
@@ -65,7 +77,8 @@ def main(args: argparse.Namespace):
             return latency
 
     print("Warming up...")
-    run_to_completion(profile_dir=None)
+    for _ in tqdm(range(args.num_iters_warmup), desc="Warmup iterations"):
+        run_to_completion(profile_dir=None)
 
     if args.profile:
         profile_dir = args.profile_result_dir
@@ -81,7 +94,22 @@ def main(args: argparse.Namespace):
     latencies = []
     for _ in tqdm(range(args.num_iters), desc="Profiling iterations"):
         latencies.append(run_to_completion(profile_dir=None))
+    latencies = np.array(latencies)
+    percentages = [10, 25, 50, 75, 90]
+    percentiles = np.percentile(latencies, percentages)
     print(f'Avg latency: {np.mean(latencies)} seconds')
+    for percentage, percentile in zip(percentages, percentiles):
+        print(f'{percentage}% percentile latency: {percentile} seconds')
+
+    # Output JSON results if specified
+    if args.output_json:
+        results = {
+            "avg_latency": np.mean(latencies),
+            "latencies": latencies.tolist(),
+            "percentiles": dict(zip(percentages, percentiles.tolist())),
+        }
+        with open(args.output_json, "w") as f:
+            json.dump(results, f, indent=4)
 
 
 if __name__ == '__main__':
@@ -89,10 +117,12 @@ if __name__ == '__main__':
         description='Benchmark the latency of processing a single batch of '
         'requests till completion.')
     parser.add_argument('--model', type=str, default='facebook/opt-125m')
+    parser.add_argument('--speculative-model', type=str, default=None)
+    parser.add_argument('--num-speculative-tokens', type=int, default=None)
     parser.add_argument('--tokenizer', type=str, default=None)
     parser.add_argument('--quantization',
                         '-q',
-                        choices=['awq', 'gptq', 'squeezellm', None],
+                        choices=[*QUANTIZATION_METHODS, None],
                         default=None)
     parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
     parser.add_argument('--input-len', type=int, default=32)
@@ -103,9 +133,13 @@ if __name__ == '__main__':
                         default=1,
                         help='Number of generated sequences per prompt.')
     parser.add_argument('--use-beam-search', action='store_true')
+    parser.add_argument('--num-iters-warmup',
+                        type=int,
+                        default=10,
+                        help='Number of iterations to run for warmup.')
     parser.add_argument('--num-iters',
                         type=int,
-                        default=3,
+                        default=30,
                         help='Number of iterations to run.')
     parser.add_argument('--trust-remote-code',
                         action='store_true',
@@ -123,12 +157,23 @@ if __name__ == '__main__':
                         action='store_true',
                         help='enforce eager mode and disable CUDA graph')
     parser.add_argument(
-        "--kv-cache-dtype",
+        '--kv-cache-dtype',
         type=str,
-        choices=['auto', 'fp8_e5m2'],
-        default='auto',
-        help=
-        'Data type for kv cache storage. If "auto", will use model data type.')
+        choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
+        default="auto",
+        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(
+        '--quantization-param-path',
+        type=str,
+        default=None,
+        help='Path to the JSON file containing the KV cache scaling factors. '
+        'This should generally be supplied, when KV cache dtype is FP8. '
+        'Otherwise, KV cache scaling factors default to 1.0, which may cause '
+        'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
+        'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
+        'instead supported for common inference criteria.')
     parser.add_argument(
         '--profile',
         action='store_true',
@@ -143,7 +188,38 @@ if __name__ == '__main__':
         "--device",
         type=str,
         default="cuda",
-        choices=["cuda"],
-        help='device type for vLLM execution, supporting CUDA only currently.')
+        choices=["cuda", "cpu"],
+        help='device type for vLLM execution, supporting CUDA and CPU.')
+    parser.add_argument('--block-size',
+                        type=int,
+                        default=16,
+                        help='block size of key/value cache')
+    parser.add_argument(
+        '--enable-chunked-prefill',
+        action='store_true',
+        help='If True, the prefill requests can be chunked based on the '
+        'max_num_batched_tokens')
+    parser.add_argument('--use-v2-block-manager', action='store_true')
+    parser.add_argument(
+        "--ray-workers-use-nsight",
+        action='store_true',
+        help="If specified, use nsight to profile ray workers",
+    )
+    parser.add_argument('--download-dir',
+                        type=str,
+                        default=None,
+                        help='directory to download and load the weights, '
+                        'default to the default cache dir of huggingface')
+    parser.add_argument(
+        '--output-json',
+        type=str,
+        default=None,
+        help='Path to save the latency results in JSON format.')
+    parser.add_argument('--gpu-memory-utilization',
+                        type=float,
+                        default=0.9,
+                        help='the fraction of GPU memory to be used for '
+                        'the model executor, which can range from 0 to 1.'
+                        'If unspecified, will use the default value of 0.9.')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_prefix_caching.py

@@ -0,0 +1,62 @@
+import argparse
+import time
+
+from vllm import LLM, SamplingParams
+
+PROMPT = "You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as fellows. You need to answer my question about the table.\n# Table\n|Opening|Opening|Sl. No.|Film|Cast|Director|Music Director|Notes|\n|----|----|----|----|----|----|----|----|\n|J A N|9|1|Agni Pushpam|Jayabharathi, Kamalahasan|Jeassy|M. K. Arjunan||\n|J A N|16|2|Priyamvada|Mohan Sharma, Lakshmi, KPAC Lalitha|K. S. Sethumadhavan|V. Dakshinamoorthy||\n|J A N|23|3|Yakshagaanam|Madhu, Sheela|Sheela|M. S. Viswanathan||\n|J A N|30|4|Paalkkadal|Sheela, Sharada|T. K. Prasad|A. T. Ummer||\n|F E B|5|5|Amma|Madhu, Srividya|M. Krishnan Nair|M. K. Arjunan||\n|F E B|13|6|Appooppan|Thikkurissi Sukumaran Nair, Kamal Haasan|P. Bhaskaran|M. S. Baburaj||\n|F E B|20|7|Srishti|Chowalloor Krishnankutty, Ravi Alummoodu|K. T. Muhammad|M. S. Baburaj||\n|F E B|20|8|Vanadevatha|Prem Nazir, Madhubala|Yusufali Kechery|G. Devarajan||\n|F E B|27|9|Samasya|Madhu, Kamalahaasan|K. Thankappan|Shyam||\n|F E B|27|10|Yudhabhoomi|K. P. Ummer, Vidhubala|Crossbelt Mani|R. K. Shekhar||\n|M A R|5|11|Seemantha Puthran|Prem Nazir, Jayabharathi|A. B. Raj|M. K. Arjunan||\n|M A R|12|12|Swapnadanam|Rani Chandra, Dr. Mohandas|K. G. George|Bhaskar Chandavarkar||\n|M A R|19|13|Thulavarsham|Prem Nazir, sreedevi, Sudheer|N. Sankaran Nair|V. Dakshinamoorthy||\n|M A R|20|14|Aruthu|Kaviyoor Ponnamma, Kamalahasan|Ravi|G. Devarajan||\n|M A R|26|15|Swimming Pool|Kamal Haasan, M. G. Soman|J. Sasikumar|M. K. Arjunan||\n\n# Question\nWhat' s the content in the (1,1) cells\n"  # noqa: E501
+
+
+def test_prefix(llm=None, sampling_params=None, prompts=None):
+    start_time = time.time()
+
+    llm.generate(prompts, sampling_params=sampling_params)
+
+    end_time = time.time()
+    print(f"cost time {end_time - start_time}")
+
+
+def main(args):
+    llm = LLM(model=args.model,
+              tokenizer_mode='auto',
+              trust_remote_code=True,
+              enforce_eager=True,
+              use_v2_block_manager=args.use_v2_block_manager,
+              tensor_parallel_size=args.tensor_parallel_size,
+              enable_prefix_caching=args.enable_prefix_caching)
+
+    num_prompts = 100
+    prompts = [PROMPT] * num_prompts
+    sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
+
+    print("------warm up------")
+    test_prefix(
+        llm=llm,
+        prompts=prompts,
+        sampling_params=sampling_params,
+    )
+
+    print("------start generating------")
+    test_prefix(
+        llm=llm,
+        prompts=prompts,
+        sampling_params=sampling_params,
+    )
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description='Benchmark the performance with or without automatic '
+        'prefix caching.')
+    parser.add_argument('--model',
+                        type=str,
+                        default='baichuan-inc/Baichuan2-13B-Chat')
+    parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
+    parser.add_argument('--output-len', type=int, default=10)
+    parser.add_argument('--enable-prefix-caching',
+                        action='store_true',
+                        help='enable prefix caching')
+    parser.add_argument('--use-v2-block-manager',
+                        action='store_true',
+                        help='Use BlockSpaceMangerV2')
+    args = parser.parse_args()
+    main(args)

benchmarks/benchmark_serving.py

@@ -1,8 +1,8 @@
 """Benchmark online serving throughput.
 
 On the server side, run one of the following commands:
-    (vLLM backend)
-    python -m vllm.entrypoints.api_server \
+    vLLM OpenAI API server
+    python -m vllm.entrypoints.openai.api_server \
         --model <your_model> --swap-space 16 \
         --disable-log-requests
 
@@ -12,28 +12,34 @@ On the server side, run one of the following commands:
 On the client side, run:
     python benchmarks/benchmark_serving.py \
         --backend <backend> \
-        --tokenizer <your_model> --dataset <target_dataset> \
-        --request-rate <request_rate>
+        --model <your_model> \
+        --dataset-name sharegpt \
+        --dataset-path <path to dataset> \
+        --request-rate <request_rate> \ # By default <request_rate> is inf
+        --num-prompts <num_prompts> # By default <num_prompts> is 1000
+        
+    when using tgi backend, add
+        --endpoint /generate_stream
+    to the end of the command above.
 """
 import argparse
 import asyncio
 import json
+import os
 import random
 import time
+import warnings
 from dataclasses import dataclass
 from datetime import datetime
-from typing import AsyncGenerator, List, Tuple
+from typing import AsyncGenerator, List, Optional, Tuple
 
 import numpy as np
+from backend_request_func import (ASYNC_REQUEST_FUNCS, RequestFuncInput,
+                                  RequestFuncOutput)
 from tqdm.asyncio import tqdm
 from transformers import PreTrainedTokenizerBase
-from vllm.transformers_utils.tokenizer import get_tokenizer
 
-from backend_request_func import (
-    ASYNC_REQUEST_FUNCS,
-    RequestFuncInput,
-    RequestFuncOutput,
-)
+from vllm.transformers_utils.tokenizer import get_tokenizer
 
 
 @dataclass
@@ -52,11 +58,15 @@ class BenchmarkMetrics:
     p99_tpot_ms: float
 
 
-def sample_requests(
+def sample_sharegpt_requests(
     dataset_path: str,
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
+    fixed_output_len: Optional[int] = None,
 ) -> List[Tuple[str, int, int]]:
+    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)
@@ -66,37 +76,101 @@ def sample_requests(
     dataset = [(data["conversations"][0]["value"],
                 data["conversations"][1]["value"]) for data in dataset]
 
-    # some of these will be filtered out, so sample more than we need
-    sampled_indices = random.sample(range(len(dataset)),
-                                    int(num_requests * 1.2))
-    dataset = [dataset[i] for i in sampled_indices]
+    # Shuffle the dataset.
+    random.shuffle(dataset)
 
-    # Tokenize the prompts and completions.
-    prompts = [prompt for prompt, _ in dataset]
-    prompt_token_ids = tokenizer(prompts).input_ids
-    completions = [completion for _, completion in dataset]
-    completion_token_ids = tokenizer(completions).input_ids
-    tokenized_dataset = []
-    for i in range(len(dataset)):
-        output_len = len(completion_token_ids[i])
-        tokenized_dataset.append((prompts[i], prompt_token_ids[i], output_len))
-
-    # Filter out too long sequences.
+    # Filter out sequences that are too long or too short
     filtered_dataset: List[Tuple[str, int, int]] = []
-    for prompt, prompt_token_ids, output_len in tokenized_dataset:
+    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 output_len < 4:
             # Prune too short sequences.
-            # This is because TGI causes errors when the input or output length
-            # is too short.
             continue
         if prompt_len > 1024 or prompt_len + output_len > 2048:
             # Prune too long sequences.
             continue
         filtered_dataset.append((prompt, prompt_len, output_len))
 
-    # Sample the requests.
-    sampled_requests = random.sample(filtered_dataset, num_requests)
+    return filtered_dataset
+
+
+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]]:
+    assert (
+        input_len > prefix_len
+    ), "'args.sonnet-input-len' must be greater than 'args.prefix-input-len'."
+
+    # Load the dataset.
+    with open(dataset_path) 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):
+        sampled_lines = "".join(
+            prefix_lines +
+            random.sample(poem_lines, num_input_lines - num_prefix_lines))
+
+        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))
+
     return sampled_requests
 
 
@@ -122,37 +196,47 @@ def calculate_metrics(
     outputs: List[RequestFuncOutput],
     dur_s: float,
     tokenizer: PreTrainedTokenizerBase,
-) -> BenchmarkMetrics:
-    total_output = 0
+) -> Tuple[BenchmarkMetrics, List[int]]:
+    actual_output_lens = []
     total_input = 0
     completed = 0
-    per_token_latencies = []
+    tpots = []
     ttfts = []
     for i in range(len(outputs)):
         if outputs[i].success:
-            output_len = len(tokenizer.encode(outputs[i].generated_text))
-            total_output += output_len
+            output_len = len(tokenizer(outputs[i].generated_text).input_ids)
+            actual_output_lens.append(output_len)
             total_input += input_requests[i][1]
-            per_token_latencies.append(outputs[i].latency / output_len)
+            if output_len > 1:
+                tpots.append(
+                    (outputs[i].latency - outputs[i].ttft) / (output_len - 1))
             ttfts.append(outputs[i].ttft)
             completed += 1
+        else:
+            actual_output_lens.append(0)
 
+    if completed == 0:
+        warnings.warn(
+            "All requests failed. This is likely due to a misconfiguration "
+            "on the benchmark arguments.",
+            stacklevel=2)
     metrics = BenchmarkMetrics(
         completed=completed,
         total_input=total_input,
-        total_output=total_output,
+        total_output=sum(actual_output_lens),
         request_throughput=completed / dur_s,
         input_throughput=total_input / dur_s,
-        output_throughput=total_output / dur_s,
-        mean_ttft_ms=np.mean(ttfts) * 1000,
-        median_ttft_ms=np.median(ttfts) * 1000,
-        p99_ttft_ms=np.percentile(ttfts, 99) * 1000,
-        mean_tpot_ms=np.mean(per_token_latencies) * 1000,
-        median_tpot_ms=np.median(per_token_latencies) * 1000,
-        p99_tpot_ms=np.percentile(per_token_latencies, 99) * 1000,
+        output_throughput=sum(actual_output_lens) / dur_s,
+        mean_ttft_ms=np.mean(ttfts or 0) *
+        1000,  # ttfts is empty if streaming is not supported by backend
+        median_ttft_ms=np.median(ttfts or 0) * 1000,
+        p99_ttft_ms=np.percentile(ttfts or 0, 99) * 1000,
+        mean_tpot_ms=np.mean(tpots or 0) * 1000,
+        median_tpot_ms=np.median(tpots or 0) * 1000,
+        p99_tpot_ms=np.percentile(tpots or 0, 99) * 1000,
     )
 
-    return metrics
+    return metrics, actual_output_lens
 
 
 async def benchmark(
@@ -171,10 +255,28 @@ async def benchmark(
     else:
         raise ValueError(f"Unknown backend: {backend}")
 
-    pbar = None if disable_tqdm else tqdm(total=len(input_requests))
-
+    print("Starting initial single prompt test run...")
+    test_prompt, test_prompt_len, test_output_len = input_requests[0]
+    test_input = RequestFuncInput(
+        model=model_id,
+        prompt=test_prompt,
+        api_url=api_url,
+        prompt_len=test_prompt_len,
+        output_len=test_output_len,
+        best_of=best_of,
+        use_beam_search=use_beam_search,
+    )
+    test_output = await request_func(request_func_input=test_input)
+    if not test_output.success:
+        raise ValueError(
+            "Initial test run failed - Please make sure benchmark arguments "
+            f"are correctly specified. Error: {test_output.error}")
+    else:
+        print("Initial test run completed. Starting main benchmark run...")
     print(f"Traffic request rate: {request_rate}")
 
+    pbar = None if disable_tqdm else tqdm(total=len(input_requests))
+
     benchmark_start_time = time.perf_counter()
     tasks = []
     async for request in get_request(input_requests, request_rate):
@@ -192,40 +294,53 @@ async def benchmark(
             asyncio.create_task(
                 request_func(request_func_input=request_func_input,
                              pbar=pbar)))
-    outputs = await asyncio.gather(*tasks)
+    outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
 
     if not disable_tqdm:
         pbar.close()
 
     benchmark_duration = time.perf_counter() - benchmark_start_time
 
-    metrics = calculate_metrics(
+    metrics, actual_output_lens = calculate_metrics(
         input_requests=input_requests,
         outputs=outputs,
         dur_s=benchmark_duration,
         tokenizer=tokenizer,
     )
 
-    print(f"Successful requests: {metrics.completed}")
-    print(f"Benchmark duration: {benchmark_duration:2f} s")
-    print(f"Total input tokens: {metrics.total_input}")
-    print(f"Total generated tokens: {metrics.total_output}")
-    print(f"Request throughput: {metrics.request_throughput:.2f} requests/s")
-    print(f"Input token throughput: {metrics.input_throughput:.2f} tokens/s")
-    print(f"Output token throughput: {metrics.output_throughput:.2f} tokens/s")
-    print(f"Mean TTFT: {metrics.mean_ttft_ms:.2f} ms")
-    print(f"Median TTFT: {metrics.median_ttft_ms:.2f} ms")
-    print(f"P99 TTFT: {metrics.p99_ttft_ms:.2f} ms")
-    print(f"Mean TPOT: {metrics.mean_tpot_ms:.2f} ms")
-    print(f"Median TPOT: {metrics.median_tpot_ms:.2f} ms")
-    print(f"P99 TPOT: {metrics.p99_tpot_ms:.2f} ms")
+    print("{s:{c}^{n}}".format(s=' Serving Benchmark Result ', n=50, c='='))
+    print("{:<40} {:<10}".format("Successful requests:", metrics.completed))
+    print("{:<40} {:<10.2f}".format("Benchmark duration (s):",
+                                    benchmark_duration))
+    print("{:<40} {:<10}".format("Total input tokens:", metrics.total_input))
+    print("{:<40} {:<10}".format("Total generated tokens:",
+                                 metrics.total_output))
+    print("{:<40} {:<10.2f}".format("Request throughput (req/s):",
+                                    metrics.request_throughput))
+    print("{:<40} {:<10.2f}".format("Input token throughput (tok/s):",
+                                    metrics.input_throughput))
+    print("{:<40} {:<10.2f}".format("Output token throughput (tok/s):",
+                                    metrics.output_throughput))
+    print("{s:{c}^{n}}".format(s='Time to First Token', n=50, c='-'))
+    print("{:<40} {:<10.2f}".format("Mean TTFT (ms):", metrics.mean_ttft_ms))
+    print("{:<40} {:<10.2f}".format("Median TTFT (ms):",
+                                    metrics.median_ttft_ms))
+    print("{:<40} {:<10.2f}".format("P99 TTFT (ms):", metrics.p99_ttft_ms))
+    print("{s:{c}^{n}}".format(s='Time per Output Token (excl. 1st token)',
+                               n=50,
+                               c='-'))
+    print("{:<40} {:<10.2f}".format("Mean TPOT (ms):", metrics.mean_tpot_ms))
+    print("{:<40} {:<10.2f}".format("Median TPOT (ms):",
+                                    metrics.median_tpot_ms))
+    print("{:<40} {:<10.2f}".format("P99 TPOT (ms):", metrics.p99_tpot_ms))
+    print("=" * 50)
 
     result = {
         "duration": benchmark_duration,
         "completed": metrics.completed,
         "total_input_tokens": metrics.total_input,
         "total_output_tokens": metrics.total_output,
-        "request_inthroughput": metrics.request_throughput,
+        "request_throughput": metrics.request_throughput,
         "input_throughput": metrics.input_throughput,
         "output_throughput": metrics.output_throughput,
         "mean_ttft_ms": metrics.mean_ttft_ms,
@@ -233,7 +348,13 @@ async def benchmark(
         "p99_ttft_ms": metrics.p99_ttft_ms,
         "mean_tpot_ms": metrics.mean_tpot_ms,
         "median_tpot_ms": metrics.median_tpot_ms,
-        "p99_tpot_ms": metrics.p99_tpot_ms
+        "p99_tpot_ms": metrics.p99_tpot_ms,
+        "input_lens": [output.prompt_len for output in outputs],
+        "output_lens": actual_output_lens,
+        "ttfts": [output.ttft for output in outputs],
+        "itls": [output.itl for output in outputs],
+        "generated_texts": [output.generated_text for output in outputs],
+        "errors": [output.error for output in outputs],
     }
     return result
 
@@ -254,7 +375,60 @@ def main(args: argparse.Namespace):
 
     tokenizer = get_tokenizer(tokenizer_id,
                               trust_remote_code=args.trust_remote_code)
-    input_requests = sample_requests(args.dataset, args.num_prompts, tokenizer)
+
+    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,
+        )
+
+    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 == "sonnet":
+        # Do not format the prompt, pass to message directly
+        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)
+                              for prompt, prompt_formatted, prompt_len,
+                              output_len in input_requests]
+        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)
+                              for prompt, prompt_formatted, prompt_len,
+                              output_len in input_requests]
+
+    else:
+        raise ValueError(f"Unknown dataset: {args.dataset_name}")
 
     benchmark_result = asyncio.run(
         benchmark(
@@ -277,13 +451,23 @@ def main(args: argparse.Namespace):
         current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
         result_json["date"] = current_dt
         result_json["backend"] = backend
-        result_json["version"] = args.version
         result_json["model_id"] = model_id
         result_json["tokenizer_id"] = tokenizer_id
         result_json["best_of"] = args.best_of
         result_json["use_beam_search"] = args.use_beam_search
         result_json["num_prompts"] = args.num_prompts
 
+        # Metadata
+        if args.metadata:
+            for item in args.metadata:
+                if "=" in item:
+                    kvstring = item.split("=")
+                    result_json[kvstring[0].strip()] = kvstring[1].strip()
+                else:
+                    raise ValueError(
+                        "Invalid metadata format. Please use KEY=VALUE format."
+                    )
+
         # Traffic
         result_json["request_rate"] = (
             args.request_rate if args.request_rate < float("inf") else "inf")
@@ -293,7 +477,9 @@ def main(args: argparse.Namespace):
 
         # Save to file
         base_model_id = model_id.split("/")[-1]
-        file_name = f"{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"
+        file_name = f"{backend}-{args.request_rate}qps-{base_model_id}-{current_dt}.json"  #noqa
+        if args.result_dir:
+            file_name = os.path.join(args.result_dir, file_name)
         with open(file_name, "w") as outfile:
             json.dump(result_json, outfile)
 
@@ -307,12 +493,6 @@ if __name__ == "__main__":
         default="vllm",
         choices=list(ASYNC_REQUEST_FUNCS.keys()),
     )
-    parser.add_argument(
-        "--version",
-        type=str,
-        default="N/A",
-        help="Version of the serving backend/engine.",
-    )
     parser.add_argument(
         "--base-url",
         type=str,
@@ -324,12 +504,26 @@ if __name__ == "__main__":
     parser.add_argument(
         "--endpoint",
         type=str,
-        default="/generate",
+        default="/v1/completions",
         help="API endpoint.",
     )
-    parser.add_argument("--dataset",
+    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,
+        default="sharegpt",
+        choices=["sharegpt", "sonnet"],
+        help="Name of the dataset to benchmark on.",
+    )
+    parser.add_argument("--dataset-path",
                         type=str,
-                        required=True,
+                        default=None,
                         help="Path to the dataset.")
     parser.add_argument(
         "--model",
@@ -341,7 +535,7 @@ if __name__ == "__main__":
         "--tokenizer",
         type=str,
         help=
-        "Name or path of the tokenizer, if not using the default model tokenizer.",
+        "Name or path of the tokenizer, if not using the default tokenizer.",
     )
     parser.add_argument(
         "--best-of",
@@ -357,6 +551,33 @@ if __name__ == "__main__":
         default=1000,
         help="Number of prompts to process.",
     )
+    parser.add_argument(
+        "--sharegpt-output-len",
+        type=int,
+        default=None,
+        help="Output length for each request. Overrides the output length "
+        "from the ShareGPT dataset.")
+    parser.add_argument(
+        "--sonnet-input-len",
+        type=int,
+        default=550,
+        help=
+        "Number of input tokens per request, used only for sonnet dataset.",
+    )
+    parser.add_argument(
+        "--sonnet-output-len",
+        type=int,
+        default=150,
+        help=
+        "Number of output tokens per request, used only for sonnet dataset.",
+    )
+    parser.add_argument(
+        "--sonnet-prefix-len",
+        type=int,
+        default=200,
+        help=
+        "Number of prefix tokens per request, used only for sonnet dataset.",
+    )
     parser.add_argument(
         "--request-rate",
         type=float,
@@ -382,6 +603,21 @@ if __name__ == "__main__":
         action="store_true",
         help="Specify to save benchmark results to a json file",
     )
+    parser.add_argument(
+        "--metadata",
+        metavar="KEY=VALUE",
+        nargs="*",
+        help="Key-value pairs (e.g, --metadata version=0.3.3 tp=1) "
+        "for metadata of this run to be saved in the result JSON file "
+        "for record keeping purposes.",
+    )
+    parser.add_argument(
+        "--result-dir",
+        type=str,
+        default=None,
+        help="Specify directory to save benchmark json results."
+        "If not specified, results are saved in the current directory.",
+    )
 
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_throughput.py

@@ -6,9 +6,11 @@ import time
 from typing import List, Optional, Tuple
 
 import torch
+from tqdm import tqdm
 from transformers import (AutoModelForCausalLM, AutoTokenizer,
                           PreTrainedTokenizerBase)
-from tqdm import tqdm
+
+from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
 
 
 def sample_requests(
@@ -29,22 +31,23 @@ def sample_requests(
     dataset = [(data["conversations"][0]["value"],
                 data["conversations"][1]["value"]) for data in dataset]
 
-    # Tokenize the prompts and completions.
-    prompts = [prompt for prompt, _ in dataset]
-    prompt_token_ids = tokenizer(prompts).input_ids
-    completions = [completion for _, completion in dataset]
-    completion_token_ids = tokenizer(completions).input_ids
-    tokenized_dataset = []
-    for i in range(len(dataset)):
-        output_len = len(completion_token_ids[i])
-        if fixed_output_len is not None:
-            output_len = fixed_output_len
-        tokenized_dataset.append((prompts[i], prompt_token_ids[i], output_len))
+    # Shuffle the dataset.
+    random.shuffle(dataset)
 
-    # Filter out too long sequences.
+    # Filter out sequences that are too long or too short
     filtered_dataset: List[Tuple[str, int, int]] = []
-    for prompt, prompt_token_ids, output_len in tokenized_dataset:
+    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 output_len < 4:
             # Prune too short sequences.
             continue
@@ -53,9 +56,7 @@ def sample_requests(
             continue
         filtered_dataset.append((prompt, prompt_len, output_len))
 
-    # Sample the requests.
-    sampled_requests = random.sample(filtered_dataset, num_requests)
-    return sampled_requests
+    return filtered_dataset
 
 
 def run_vllm(
@@ -72,7 +73,13 @@ def run_vllm(
     max_model_len: Optional[int],
     enforce_eager: bool,
     kv_cache_dtype: str,
+    quantization_param_path: Optional[str],
     device: str,
+    enable_prefix_caching: bool,
+    enable_chunked_prefill: bool,
+    max_num_batched_tokens: int,
+    gpu_memory_utilization: float = 0.9,
+    download_dir: Optional[str] = None,
 ) -> float:
     from vllm import LLM, SamplingParams
     llm = LLM(
@@ -84,31 +91,34 @@ def run_vllm(
         trust_remote_code=trust_remote_code,
         dtype=dtype,
         max_model_len=max_model_len,
+        gpu_memory_utilization=gpu_memory_utilization,
         enforce_eager=enforce_eager,
         kv_cache_dtype=kv_cache_dtype,
+        quantization_param_path=quantization_param_path,
         device=device,
+        enable_prefix_caching=enable_prefix_caching,
+        download_dir=download_dir,
+        enable_chunked_prefill=enable_chunked_prefill,
+        max_num_batched_tokens=max_num_batched_tokens,
     )
 
     # Add the requests to the engine.
+    prompts = []
+    sampling_params = []
     for prompt, _, output_len in requests:
-        sampling_params = SamplingParams(
-            n=n,
-            temperature=0.0 if use_beam_search else 1.0,
-            top_p=1.0,
-            use_beam_search=use_beam_search,
-            ignore_eos=True,
-            max_tokens=output_len,
-        )
-        # FIXME(woosuk): Do not use internal method.
-        llm._add_request(
-            prompt=prompt,
-            prompt_token_ids=None,
-            sampling_params=sampling_params,
-        )
+        prompts.append(prompt)
+        sampling_params.append(
+            SamplingParams(
+                n=n,
+                temperature=0.0 if use_beam_search else 1.0,
+                top_p=1.0,
+                use_beam_search=use_beam_search,
+                ignore_eos=True,
+                max_tokens=output_len,
+            ))
 
     start = time.perf_counter()
-    # FIXME(woosuk): Do not use internal method.
-    llm._run_engine(use_tqdm=True)
+    llm.generate(prompts, sampling_params, use_tqdm=True)
     end = time.perf_counter()
     return end - start
 
@@ -179,13 +189,15 @@ def run_mii(
     tensor_parallel_size: int,
     output_len: int,
 ) -> float:
-    from mii import pipeline
-    llm = pipeline(model, tensor_parallel=tensor_parallel_size)
+    from mii import client, serve
+    llm = serve(model, tensor_parallel=tensor_parallel_size)
     prompts = [prompt for prompt, _, _ in requests]
 
     start = time.perf_counter()
-    llm(prompts, max_new_tokens=output_len)
+    llm.generate(prompts, max_new_tokens=output_len)
     end = time.perf_counter()
+    client = client(model)
+    client.terminate_server()
     return end - start
 
 
@@ -206,12 +218,15 @@ def main(args: argparse.Namespace):
                                    args.output_len)
 
     if args.backend == "vllm":
-        elapsed_time = run_vllm(requests, args.model, args.tokenizer,
-                                args.quantization, args.tensor_parallel_size,
-                                args.seed, args.n, args.use_beam_search,
-                                args.trust_remote_code, args.dtype,
-                                args.max_model_len, args.enforce_eager,
-                                args.kv_cache_dtype, args.device)
+        elapsed_time = run_vllm(
+            requests, args.model, args.tokenizer, args.quantization,
+            args.tensor_parallel_size, args.seed, args.n, args.use_beam_search,
+            args.trust_remote_code, args.dtype, args.max_model_len,
+            args.enforce_eager, args.kv_cache_dtype,
+            args.quantization_param_path, args.device,
+            args.enable_prefix_caching, args.enable_chunked_prefill,
+            args.max_num_batched_tokens, args.gpu_memory_utilization,
+            args.download_dir)
     elif args.backend == "hf":
         assert args.tensor_parallel_size == 1
         elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
@@ -227,6 +242,18 @@ def main(args: argparse.Namespace):
     print(f"Throughput: {len(requests) / elapsed_time:.2f} requests/s, "
           f"{total_num_tokens / elapsed_time:.2f} tokens/s")
 
+    # Output JSON results if specified
+    if args.output_json:
+        results = {
+            "elapsed_time": elapsed_time,
+            "num_requests": len(requests),
+            "total_num_tokens": total_num_tokens,
+            "requests_per_second": len(requests) / elapsed_time,
+            "tokens_per_second": total_num_tokens / elapsed_time,
+        }
+        with open(args.output_json, "w") as f:
+            json.dump(results, f, indent=4)
+
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="Benchmark the throughput.")
@@ -251,7 +278,7 @@ if __name__ == "__main__":
     parser.add_argument("--tokenizer", type=str, default=None)
     parser.add_argument('--quantization',
                         '-q',
-                        choices=['awq', 'gptq', 'squeezellm', None],
+                        choices=[*QUANTIZATION_METHODS, None],
                         default=None)
     parser.add_argument("--tensor-parallel-size", "-tp", type=int, default=1)
     parser.add_argument("--n",
@@ -286,22 +313,61 @@ if __name__ == "__main__":
         'The "auto" option will use FP16 precision '
         'for FP32 and FP16 models, and BF16 precision '
         'for BF16 models.')
+    parser.add_argument('--gpu-memory-utilization',
+                        type=float,
+                        default=0.9,
+                        help='the fraction of GPU memory to be used for '
+                        'the model executor, which can range from 0 to 1.'
+                        'If unspecified, will use the default value of 0.9.')
     parser.add_argument("--enforce-eager",
                         action="store_true",
                         help="enforce eager execution")
     parser.add_argument(
-        "--kv-cache-dtype",
+        '--kv-cache-dtype',
         type=str,
-        choices=["auto", "fp8_e5m2"],
+        choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
         default="auto",
-        help=
-        'Data type for kv cache storage. If "auto", will use model data type.')
+        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(
+        '--quantization-param-path',
+        type=str,
+        default=None,
+        help='Path to the JSON file containing the KV cache scaling factors. '
+        'This should generally be supplied, when KV cache dtype is FP8. '
+        'Otherwise, KV cache scaling factors default to 1.0, which may cause '
+        'accuracy issues. FP8_E5M2 (without scaling) is only supported on '
+        'cuda version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is '
+        'instead supported for common inference criteria.')
     parser.add_argument(
         "--device",
         type=str,
         default="cuda",
-        choices=["cuda"],
-        help='device type for vLLM execution, supporting CUDA only currently.')
+        choices=["cuda", "cpu"],
+        help='device type for vLLM execution, supporting CUDA and CPU.')
+    parser.add_argument(
+        "--enable-prefix-caching",
+        action='store_true',
+        help="enable automatic prefix caching for vLLM backend.")
+    parser.add_argument("--enable-chunked-prefill",
+                        action='store_true',
+                        help="enable chunked prefill for vLLM backend.")
+    parser.add_argument('--max-num-batched-tokens',
+                        type=int,
+                        default=None,
+                        help='maximum number of batched tokens per '
+                        'iteration')
+    parser.add_argument('--download-dir',
+                        type=str,
+                        default=None,
+                        help='directory to download and load the weights, '
+                        'default to the default cache dir of huggingface')
+    parser.add_argument(
+        '--output-json',
+        type=str,
+        default=None,
+        help='Path to save the throughput results in JSON format.')
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

benchmarks/kernels/benchmark_aqlm.py

@@ -0,0 +1,302 @@
+import argparse
+import os
+import sys
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+
+from vllm import _custom_ops as ops
+from vllm.model_executor.layers.quantization.aqlm import (
+    dequantize_weight, generic_dequantize_gemm, get_int_dtype,
+    optimized_dequantize_gemm)
+
+os.environ['CUDA_VISIBLE_DEVICES'] = '0'
+
+
+def torch_mult(
+        input: torch.Tensor,  #  [..., in_features]
+        weights: torch.Tensor,
+        scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+) -> torch.Tensor:
+    output = F.linear(input, weights)
+    return output
+
+
+def dequant_out_scale(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    if bias is None:
+        output = F.linear(input, weights, bias)
+        orig_shape = output.shape
+        flattened_output = output.view(-1, output.size(-1))
+        f_scales = scales.view(-1, scales.shape[0])
+        b_scales = f_scales.expand(flattened_output.shape[0], -1)
+        flattened_output *= b_scales
+        return flattened_output.view(orig_shape)
+    else:
+        b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
+            -1, weights.shape[1])
+        weights *= b_scales
+        return F.linear(input, weights, bias)
+
+
+def dequant_weight_scale(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    b_scales = scales.view(scales.shape[:-3] + (-1, )).expand(
+        -1, weights.shape[1])
+    weights *= b_scales
+    return F.linear(input, weights, bias)
+
+
+def dequant_no_scale(
+    input: torch.Tensor,  #  [..., in_features]
+    codes: torch.IntTensor,  #  [num_out_groups, num_in_groups, num_codebooks]
+    codebooks: torch.
+    Tensor,  #  [num_codebooks, codebook_size, out_group_size, in_group_size]
+    scales: torch.Tensor,  #  [num_out_groups, 1, 1, 1]
+    output_partition_sizes: torch.IntTensor,
+    bias: Optional[torch.Tensor],
+) -> torch.Tensor:
+
+    weights = ops.aqlm_dequant(codes, codebooks, output_partition_sizes)
+
+    return F.linear(input, weights, bias)
+
+
+# Compare the optimized 1x16 and 2x8 cuda decompression/dequant kernels against
+# the generic pytorch version.
+# Just visual comparison.
+def dequant_test(k: int, parts: torch.tensor, nbooks: int, bits: int) -> None:
+
+    n = parts.sum().item()
+
+    device = torch.device('cuda:0')
+
+    code_range = (1 << bits) // 2
+    ingroups = 8
+
+    codes = torch.randint(-code_range,
+                          code_range,
+                          size=(n, k // ingroups, nbooks),
+                          dtype=get_int_dtype(bits),
+                          device=device)
+
+    codebooks = torch.randn(size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
+                            dtype=torch.float16,
+                            device=device)
+
+    count = 0
+    for index in range(16):
+        for i in range(8):
+            for book in range(nbooks):
+                codebooks[book, index, 0, i] = count * (10**book)
+            count += 1
+
+    print("codes shape", codes.shape)
+
+    for i in range(16):
+        for book in range(nbooks):
+            codes[0, i, book] = i
+            codes[0, -i, book] = i
+
+    weights = dequantize_weight(codes, codebooks, None)
+    weights2 = ops.aqlm_dequant(codes, codebooks, parts)
+
+    print("weights shape:", weights.shape)
+    print("weights2 shape:", weights2.shape)
+
+    print("weights are:", weights)
+    print("weights2 are:", weights2)
+
+    print("first 128 weights are", weights[0, 0:128].to(torch.int32))
+    print("first 128 weights2 are:", weights2[0, 0:128].to(torch.int32))
+
+    print("last 128 weights are", weights[0, -128:])
+    print("last 128 weights2 are:", weights2[0, -128:])
+
+
+def main():
+
+    parser = argparse.ArgumentParser(description="Benchmark aqlm performance.")
+
+    # Add arguments
+    parser.add_argument("--nbooks",
+                        type=int,
+                        default=1,
+                        help="Number of codebooks (default: 1)")
+    parser.add_argument("--bits",
+                        type=int,
+                        default=16,
+                        help="Number of bits per code element (default: 16)")
+    parser.add_argument(
+        "--test",
+        type=bool,
+        default=False,
+        help="Run the decompression/dequant tester rather than benchmarking "
+        "(default: False)")
+
+    # Parse the arguments
+    args = parser.parse_args()
+
+    # Extract values
+    nbooks = args.nbooks
+    bits = args.bits
+
+    if args.test:
+        dequant_test(4096, torch.tensor((4096, )), nbooks, bits)
+        return
+
+    # Otherwise, benchmark.
+    methods = [
+        ops.aqlm_gemm,
+        dequant_out_scale,
+        generic_dequantize_gemm,
+        optimized_dequantize_gemm,
+        dequant_weight_scale,
+        torch_mult,
+        dequant_no_scale,
+    ]
+
+    filename = f"./aqlm_benchmark_{nbooks}x{bits}.csv"
+    print(f"writing benchmarks to file {filename}")
+    with open(filename, "w") as f:
+        sys.stdout = f
+
+        print('m | k | n | n parts', end='')
+        for method in methods:
+            print(f" | {method.__name__.replace('_', ' ')} (µs)", end='')
+        print('')
+
+        # These are reasonable prefill sizes.
+        ksandpartions = ((4096, (4096, 4096, 4096)), (4096, (4096, )),
+                         (4096, (11008, 11008)), (11008, (4096, )))
+
+        # reasonable ranges for m.
+        for m in [
+                1, 2, 4, 8, 10, 12, 14, 16, 24, 32, 48, 52, 56, 64, 96, 112,
+                128, 256, 512, 1024, 1536, 2048, 3072, 4096
+        ]:
+            print(f'{m}', file=sys.__stdout__)
+            for ksp in ksandpartions:
+                run_grid(m, ksp[0], torch.tensor(ksp[1]), nbooks, bits,
+                         methods)
+
+        sys.stdout = sys.__stdout__
+
+
+def run_grid(m: int, k: int, parts: torch.tensor, nbooks: int, bits: int,
+             methods):
+
+    # I didn't see visible improvements from increasing these, but feel free :)
+    num_warmup_trials = 1
+    num_trials = 1
+
+    num_calls = 100
+
+    # warmup.
+    for method in methods:
+        for _ in range(num_warmup_trials):
+            run_timing(
+                num_calls=num_calls,
+                m=m,
+                k=k,
+                parts=parts,
+                nbooks=nbooks,
+                bits=bits,
+                method=method,
+            )
+
+    n = parts.sum().item()
+    print(f'{m} | {k} | {n} | {parts.tolist()}', end='')
+
+    for method in methods:
+        best_time_us = 1e20
+        for _ in range(num_trials):
+            kernel_dur_ms = run_timing(
+                num_calls=num_calls,
+                m=m,
+                k=k,
+                parts=parts,
+                nbooks=nbooks,
+                bits=bits,
+                method=method,
+            )
+
+            kernel_dur_us = 1000 * kernel_dur_ms
+
+            if kernel_dur_us < best_time_us:
+                best_time_us = kernel_dur_us
+
+        print(f' | {kernel_dur_us:.0f}', end='')
+
+    print('')
+
+
+def run_timing(num_calls: int, m: int, k: int, parts: torch.tensor,
+               nbooks: int, bits: int, method) -> float:
+
+    n = parts.sum().item()
+
+    device = torch.device('cuda:0')
+
+    input = torch.randn((1, m, k), dtype=torch.float16, device=device)
+
+    code_range = (1 << bits) // 2
+    ingroups = 8
+
+    codes = torch.randint(-code_range,
+                          code_range,
+                          size=(n, k // ingroups, nbooks),
+                          dtype=get_int_dtype(bits),
+                          device=device)
+
+    codebooks = torch.randn(size=(parts.shape[0] * nbooks, 1 << bits, 1, 8),
+                            dtype=torch.float16,
+                            device=device)
+
+    scales = torch.randn(size=(n, 1, 1, 1), dtype=torch.float16, device=device)
+
+    # for comparison to just a pytorch mult.
+    weights = torch.randn((n, k), dtype=torch.float16, device=device)
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    start_event.record()
+
+    if method is torch_mult:
+        for i in range(num_calls):
+            torch_mult(input, weights, scales)
+    else:
+        for i in range(num_calls):
+            method(input, codes, codebooks, scales, parts, None)
+
+    end_event.record()
+    end_event.synchronize()
+
+    dur_ms = start_event.elapsed_time(end_event) / num_calls
+    return dur_ms
+
+
+if __name__ == "__main__":
+    sys.exit(main())

benchmarks/kernels/benchmark_marlin.py

@@ -0,0 +1,233 @@
+import argparse
+
+import torch
+import torch.utils.benchmark as benchmark
+from benchmark_shapes import WEIGHT_SHAPES
+
+from vllm import _custom_ops as ops
+from vllm.model_executor.layers.quantization.gptq_marlin import (
+    GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_N,
+    GPTQ_MARLIN_SUPPORTED_GROUP_SIZES, GPTQ_MARLIN_SUPPORTED_NUM_BITS)
+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_NUM_BITS)
+from vllm.model_executor.layers.quantization.utils.marlin_utils import (
+    MarlinWorkspace, marlin_24_quantize, marlin_quantize)
+from vllm.model_executor.layers.quantization.utils.quant_utils import (
+    gptq_pack, quantize_weights, sort_weights)
+
+DEFAULT_MODELS = ["meta-llama/Llama-2-7b-hf/TP1"]
+DEFAULT_BATCH_SIZES = [1, 16, 32, 64, 128, 256, 512]
+
+ACT_ORDER_OPTS = [False, True]
+K_FULL_OPTS = [False, True]
+
+
+def bench_run(results, model, act_order, is_k_full, num_bits, group_size,
+              size_m, size_k, size_n):
+    label = "Quant Matmul"
+
+    sub_label = ("{}, act={} k_full={}, b={}, g={}, "
+                 "MKN=({}x{}x{})".format(model, act_order, is_k_full, num_bits,
+                                         group_size, size_m, size_k, size_n))
+
+    print(f"Testing: {sub_label}")
+
+    a = torch.randn(size_m, size_k).to(torch.half).cuda()
+    b = torch.rand(size_k, size_n).to(torch.half).cuda()
+
+    a_tmp = (torch.zeros(size_m, size_k).to(torch.half).cuda())
+
+    # Marlin quant
+    (
+        marlin_w_ref,
+        marlin_q_w,
+        marlin_s,
+        marlin_g_idx,
+        marlin_sort_indices,
+        marlin_rand_perm,
+    ) = marlin_quantize(b, num_bits, group_size, act_order)
+
+    # Marlin_24 quant
+    (marlin_24_w_ref, marlin_24_q_w_comp, marlin_24_meta,
+     marlin_24_s) = marlin_24_quantize(b, num_bits, group_size)
+
+    # GPTQ quant
+    (w_ref, q_w, s, g_idx,
+     rand_perm) = quantize_weights(b, num_bits, group_size, act_order)
+    q_w_gptq = gptq_pack(q_w, num_bits, size_k, size_n)
+
+    # For act_order, sort the "weights" and "g_idx"
+    # so that group ids are increasing
+    repack_sort_indices = torch.empty(0, dtype=torch.int, device=b.device)
+    if act_order:
+        (q_w, g_idx, repack_sort_indices) = sort_weights(q_w, g_idx)
+
+    # Prepare
+    marlin_workspace = MarlinWorkspace(size_n, GPTQ_MARLIN_MIN_THREAD_N,
+                                       GPTQ_MARLIN_MAX_PARALLEL)
+
+    marlin_24_workspace = MarlinWorkspace(size_n, GPTQ_MARLIN_24_MIN_THREAD_N,
+                                          GPTQ_MARLIN_24_MAX_PARALLEL)
+
+    globals = {
+        # Gen params
+        "num_bits": num_bits,
+        "group_size": group_size,
+        "size_m": size_m,
+        "size_n": size_n,
+        "size_k": size_k,
+        "a": a,
+        "a_tmp": a_tmp,
+        # Marlin params
+        "marlin_w_ref": marlin_w_ref,
+        "marlin_q_w": marlin_q_w,
+        "marlin_s": marlin_s,
+        "marlin_g_idx": marlin_g_idx,
+        "marlin_sort_indices": marlin_sort_indices,
+        "marlin_rand_perm": marlin_rand_perm,
+        "marlin_workspace": marlin_workspace,
+        "is_k_full": is_k_full,
+        # Marlin_24 params
+        "marlin_24_w_ref": marlin_24_w_ref,
+        "marlin_24_q_w_comp": marlin_24_q_w_comp,
+        "marlin_24_meta": marlin_24_meta,
+        "marlin_24_s": marlin_24_s,
+        "marlin_24_workspace": marlin_24_workspace,
+        # GPTQ params
+        "q_w_gptq": q_w_gptq,
+        "repack_sort_indices": repack_sort_indices,
+        # Kernels
+        "gptq_marlin_gemm": ops.gptq_marlin_gemm,
+        "gptq_marlin_24_gemm": ops.gptq_marlin_24_gemm,
+        "gptq_marlin_repack": ops.gptq_marlin_repack,
+    }
+
+    min_run_time = 1
+
+    # Warmup pytorch
+    for i in range(5):
+        torch.matmul(a, marlin_w_ref)
+
+    results.append(
+        benchmark.Timer(
+            stmt="torch.matmul(a, marlin_w_ref)",
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="pytorch_gemm",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "output = gptq_marlin_gemm(a, marlin_q_w, marlin_s, marlin_g_idx, marlin_sort_indices, marlin_workspace.scratch, num_bits, size_m, size_n, size_k, is_k_full)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="gptq_marlin_gemm",
+        ).blocked_autorange(min_run_time=min_run_time))
+
+    if (num_bits in GPTQ_MARLIN_24_SUPPORTED_NUM_BITS
+            and group_size in GPTQ_MARLIN_24_SUPPORTED_GROUP_SIZES):
+        results.append(
+            benchmark.Timer(
+                stmt=
+                "output = gptq_marlin_24_gemm(a, marlin_24_q_w_comp, marlin_24_meta, marlin_24_s, marlin_24_workspace.scratch, num_bits, size_m, size_n, size_k)",  # noqa: E501
+                globals=globals,
+                label=label,
+                sub_label=sub_label,
+                description="gptq_marlin_24_gemm",
+            ).blocked_autorange(min_run_time=min_run_time))
+
+    results.append(
+        benchmark.Timer(
+            stmt=
+            "q_res = gptq_marlin_repack(q_w_gptq, repack_sort_indices, size_k, size_n, num_bits)",  # noqa: E501
+            globals=globals,
+            label=label,
+            sub_label=sub_label,
+            description="gptq_marlin_repack",
+        ).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 = []
+
+    for model in args.models:
+        for layer in WEIGHT_SHAPES[model]:
+            size_k = layer[0]
+            size_n = layer[1]
+
+            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 act_order in ACT_ORDER_OPTS:
+                if len(args.limit_act_order
+                       ) > 0 and act_order not in args.limit_act_order:
+                    continue
+
+                for is_k_full in K_FULL_OPTS:
+                    if len(args.limit_k_full
+                           ) > 0 and is_k_full not in args.limit_k_full:
+                        continue
+
+                    for num_bits in GPTQ_MARLIN_SUPPORTED_NUM_BITS:
+                        if len(args.limit_num_bits
+                               ) > 0 and num_bits not in args.limit_num_bits:
+                            continue
+
+                        for group_size in GPTQ_MARLIN_SUPPORTED_GROUP_SIZES:
+                            if len(
+                                    args.limit_group_size
+                            ) > 0 and group_size not in args.limit_group_size:
+                                continue
+
+                            # For act_order, the group_size must be less than
+                            # size_k
+                            if act_order and (group_size == size_k
+                                              or group_size == -1):
+                                continue
+
+                            for size_m in args.batch_sizes:
+                                bench_run(results, model, act_order, is_k_full,
+                                          num_bits, group_size, size_m, size_k,
+                                          size_n)
+
+    compare = benchmark.Compare(results)
+    compare.print()
+
+
+# For quick benchmarking use:
+#   python benchmark_marlin.py --batch-sizes 1 16 32 --limit-k 4096 --limit-n 4096 --limit-group-size 128 --limit-num-bits 4 --limit-act-order 0 --limit-k-full 1 # noqa E501
+#
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="Benchmark Marlin across specified models/shapes/batches")
+    parser.add_argument(
+        "--models",
+        nargs="+",
+        type=str,
+        default=DEFAULT_MODELS,
+        choices=WEIGHT_SHAPES.keys(),
+    )
+    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-group-size", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-num-bits", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-act-order", nargs="+", type=int, default=[])
+    parser.add_argument("--limit-k-full", nargs="+", type=int, default=[])
+
+    args = parser.parse_args()
+    main(args)

benchmarks/kernels/benchmark_mixtral_moe.py

@@ -1,70 +1,75 @@
+import argparse
 import json
 import os
 import sys
 
-os.environ['CUDA_VISIBLE_DEVICES'] = '0'
-
-from vllm.model_executor.layers.fused_moe import fused_moe
 import torch
 import torch.nn.functional as F
 import triton
+from tqdm import tqdm
+
+from vllm.model_executor.layers.fused_moe import (fused_moe,
+                                                  get_config_file_name)
 
 
-def main():
+def main(model, tp_size, gpu, dtype: str):
+    os.environ['CUDA_VISIBLE_DEVICES'] = str(gpu)
     method = fused_moe
     for bs in [
             1, 2, 4, 8, 16, 24, 32, 48, 64, 96, 128, 256, 512, 1024, 1536,
             2048, 3072, 4096
     ]:
-        run_grid(bs, method=method)
+        run_grid(bs,
+                 model=model,
+                 method=method,
+                 gpu=gpu,
+                 tp_size=tp_size,
+                 dtype=dtype)
 
 
-def run_grid(bs, method):
-    d_model = 4096
+def run_grid(bs, model, method, gpu, tp_size, dtype: str):
+    if model == '8x7B':
+        d_model = 4096
+        model_intermediate_size = 14336
+        num_layers = 32
+    elif model == '8x22B':
+        d_model = 6144
+        model_intermediate_size = 16384
+        num_layers = 56
+    else:
+        raise ValueError(f'Unsupported Mixtral model {model}')
     num_total_experts = 8
     top_k = 2
-    tp_size = 2
-    model_intermediate_size = 14336
-    num_layers = 32
+    # tp_size = 2
     num_calls = 100
 
     num_warmup_trials = 1
     num_trials = 1
 
     configs = []
-    if bs <= 16:
-        BLOCK_SIZES_M = [16]
-    elif bs <= 32:
-        BLOCK_SIZES_M = [16, 32]
-    elif bs <= 64:
-        BLOCK_SIZES_M = [16, 32, 64]
-    elif bs <= 128:
-        BLOCK_SIZES_M = [16, 32, 64, 128]
-    else:
-        BLOCK_SIZES_M = [16, 32, 64, 128, 256]
 
     for block_size_n in [32, 64, 128, 256]:
-        for block_size_m in BLOCK_SIZES_M:
+        for block_size_m in [16, 32, 64, 128, 256]:
             for block_size_k in [64, 128, 256]:
                 for group_size_m in [1, 16, 32, 64]:
                     for num_warps in [4, 8]:
-                        configs.append({
-                            "BLOCK_SIZE_M": block_size_m,
-                            "BLOCK_SIZE_N": block_size_n,
-                            "BLOCK_SIZE_K": block_size_k,
-                            "GROUP_SIZE_M": group_size_m,
-                            "num_warps": num_warps,
-                            "num_stages": 4,
-                        })
+                        for num_stages in [2, 3, 4, 5]:
+                            configs.append({
+                                "BLOCK_SIZE_M": block_size_m,
+                                "BLOCK_SIZE_N": block_size_n,
+                                "BLOCK_SIZE_K": block_size_k,
+                                "GROUP_SIZE_M": group_size_m,
+                                "num_warps": num_warps,
+                                "num_stages": num_stages,
+                            })
 
     best_config = None
     best_time_us = 1e20
 
-    for config in configs:
-        print(f'{tp_size=} {bs=}')
-        print(f'{config}')
+    print(f'{tp_size=} {bs=}')
+
+    for config in tqdm(configs):
         # warmup
-        print(f'warming up')
         try:
             for _ in range(num_warmup_trials):
                 run_timing(
@@ -77,12 +82,12 @@ def run_grid(bs, method):
                     model_intermediate_size=model_intermediate_size,
                     method=method,
                     config=config,
+                    dtype=dtype,
                 )
         except triton.runtime.autotuner.OutOfResources:
             continue
 
         # trial
-        print(f'benchmarking')
         for _ in range(num_trials):
             kernel_dur_ms = run_timing(
                 num_calls=num_calls,
@@ -94,6 +99,7 @@ def run_grid(bs, method):
                 model_intermediate_size=model_intermediate_size,
                 method=method,
                 config=config,
+                dtype=dtype,
             )
 
             kernel_dur_us = 1000 * kernel_dur_ms
@@ -103,42 +109,73 @@ def run_grid(bs, method):
                 best_config = config
                 best_time_us = kernel_dur_us
 
-            print(
-                f'{kernel_dur_us=:.1f} {model_dur_ms=:.1f} {bs=} {tp_size=} {top_k=} {num_total_experts=} {d_model=} {model_intermediate_size=} {num_layers=}'
-            )
+                tqdm.write(
+                    f'{kernel_dur_us=:.1f} {model_dur_ms=:.1f}'
+                    f' {bs=} {tp_size=} {top_k=} {num_total_experts=} '
+                    f'{d_model=} {model_intermediate_size=} {num_layers=}')
 
     print("best_time_us", best_time_us)
     print("best_config", best_config)
 
-    filename = "/tmp/config.jsonl"
+    # holds Dict[str, Dict[str, int]]
+    filename = get_config_file_name(num_total_experts,
+                                    model_intermediate_size // tp_size,
+                                    "float8" if dtype == "float8" else None)
     print(f"writing config to file {filename}")
-    with open(filename, "a") as f:
-        f.write(json.dumps({str(bs): best_config}) + "\n")
+    existing_content = {}
+    if os.path.exists(filename):
+        with open(filename, "r") as f:
+            existing_content = json.load(f)
+    existing_content[str(bs)] = best_config
+    with open(filename, "w") as f:
+        json.dump(existing_content, f, indent=4)
+        f.write("\n")
 
 
 def run_timing(num_calls: int, bs: int, d_model: int, num_total_experts: int,
                top_k: int, tp_size: int, model_intermediate_size: int, method,
-               config) -> float:
+               config, dtype: str) -> float:
     shard_intermediate_size = model_intermediate_size // tp_size
 
     hidden_states = torch.rand(
         (bs, d_model),
         device="cuda:0",
-        dtype=torch.bfloat16,
+        dtype=torch.float16,
     )
 
-    ws = torch.rand(
+    w1 = torch.rand(
         (num_total_experts, 2 * shard_intermediate_size, d_model),
         device=hidden_states.device,
         dtype=hidden_states.dtype,
     )
 
-    w2s = torch.rand(
+    w2 = torch.rand(
         (num_total_experts, d_model, shard_intermediate_size),
         device=hidden_states.device,
         dtype=hidden_states.dtype,
     )
 
+    w1_scale = None
+    w2_scale = None
+    a1_scale = None
+    a2_scale = None
+
+    if dtype == "float8":
+        w1 = w1.to(torch.float8_e4m3fn)
+        w2 = w2.to(torch.float8_e4m3fn)
+        w1_scale = torch.ones(num_total_experts,
+                              device=hidden_states.device,
+                              dtype=torch.float32)
+        w2_scale = torch.ones(num_total_experts,
+                              device=hidden_states.device,
+                              dtype=torch.float32)
+        a1_scale = torch.ones(1,
+                              device=hidden_states.device,
+                              dtype=torch.float32)
+        a2_scale = torch.ones(1,
+                              device=hidden_states.device,
+                              dtype=torch.float32)
+
     gating_output = F.softmax(torch.rand(
         (num_calls, bs, num_total_experts),
         device=hidden_states.device,
@@ -153,13 +190,18 @@ def run_timing(num_calls: int, bs: int, d_model: int, num_total_experts: int,
     for i in range(num_calls):
         hidden_states = method(
             hidden_states=hidden_states,
-            w1=ws,
-            w2=w2s,
+            w1=w1,
+            w2=w2,
+            w1_scale=w1_scale,
+            w2_scale=w2_scale,
+            a1_scale=a1_scale,
+            a2_scale=a2_scale,
             gating_output=gating_output[i],
             topk=2,
             renormalize=True,
             inplace=True,
             override_config=config,
+            use_fp8=dtype == "float8",
         )
     end_event.record()
     end_event.synchronize()
@@ -169,4 +211,29 @@ def run_timing(num_calls: int, bs: int, d_model: int, num_total_experts: int,
 
 
 if __name__ == "__main__":
-    sys.exit(main())
+    parser = argparse.ArgumentParser(
+        prog='benchmark_mixtral_moe',
+        description='Benchmark and tune the fused_moe kernel',
+    )
+    parser.add_argument(
+        '--dtype',
+        type=str,
+        default='auto',
+        choices=['float8', 'float16'],
+        help='Data type used for fused_moe kernel computations',
+    )
+    parser.add_argument('--model',
+                        type=str,
+                        default='8x7B',
+                        choices=['8x7B', '8x22B'],
+                        help='The Mixtral model to benchmark')
+    parser.add_argument('--tp-size',
+                        type=int,
+                        default=2,
+                        help='Tensor paralleli size')
+    parser.add_argument('--gpu',
+                        type=int,
+                        default=0,
+                        help="GPU ID for benchmarking")
+    args = parser.parse_args()
+    sys.exit(main(args.model, args.tp_size, args.gpu, args.dtype))

benchmarks/kernels/benchmark_paged_attention.py

@@ -1,12 +1,12 @@
-from typing import Optional
 import argparse
 import random
 import time
+from typing import Optional
 
 import torch
 
+from vllm import _custom_ops as ops
 from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, create_kv_caches_with_random
-from vllm._C import ops
 
 NUM_BLOCKS = 1024
 PARTITION_SIZE = 512
@@ -16,7 +16,7 @@ PARTITION_SIZE = 512
 def main(
     version: str,
     num_seqs: int,
-    context_len: int,
+    seq_len: int,
     num_query_heads: int,
     num_kv_heads: int,
     head_size: int,
@@ -48,12 +48,12 @@ def main(
                                    dtype=torch.float,
                                    device=device)
 
-    context_lens = [context_len for _ in range(num_seqs)]
-    max_context_len = max(context_lens)
-    context_lens = torch.tensor(context_lens, dtype=torch.int, device=device)
+    seq_lens = [seq_len for _ in range(num_seqs)]
+    max_seq_len = max(seq_lens)
+    seq_lens = torch.tensor(seq_lens, dtype=torch.int, device=device)
 
     # Create the block tables.
-    max_num_blocks_per_seq = (max_context_len + block_size - 1) // block_size
+    max_num_blocks_per_seq = (max_seq_len + block_size - 1) // block_size
     block_tables = []
     for _ in range(num_seqs):
         block_table = [
@@ -77,8 +77,7 @@ def main(
     # Prepare for the paged attention kernel.
     output = torch.empty_like(query)
     if version == "v2":
-        num_partitions = ((max_context_len + PARTITION_SIZE - 1) //
-                          PARTITION_SIZE)
+        num_partitions = ((max_seq_len + PARTITION_SIZE - 1) // PARTITION_SIZE)
         tmp_output = torch.empty(
             size=(num_seqs, num_query_heads, num_partitions, head_size),
             dtype=output.dtype,
@@ -97,6 +96,9 @@ def main(
             torch.cuda.cudart().cudaProfilerStart()
         start_time = time.perf_counter()
 
+        # Using default kv_scale
+        kv_scale = 1.0
+
         for _ in range(num_iters):
             if version == "v1":
                 ops.paged_attention_v1(
@@ -107,11 +109,12 @@ def main(
                     num_kv_heads,
                     scale,
                     block_tables,
-                    context_lens,
+                    seq_lens,
                     block_size,
-                    max_context_len,
+                    max_seq_len,
                     alibi_slopes,
                     kv_cache_dtype,
+                    kv_scale,
                 )
             elif version == "v2":
                 ops.paged_attention_v2(
@@ -125,11 +128,12 @@ def main(
                     num_kv_heads,
                     scale,
                     block_tables,
-                    context_lens,
+                    seq_lens,
                     block_size,
-                    max_context_len,
+                    max_seq_len,
                     alibi_slopes,
                     kv_cache_dtype,
+                    kv_scale,
                 )
             else:
                 raise ValueError(f"Invalid version: {version}")
@@ -161,12 +165,12 @@ if __name__ == '__main__':
                         choices=["v1", "v2"],
                         default="v2")
     parser.add_argument("--batch-size", type=int, default=8)
-    parser.add_argument("--context-len", type=int, default=4096)
+    parser.add_argument("--seq_len", type=int, default=4096)
     parser.add_argument("--num-query-heads", type=int, default=64)
     parser.add_argument("--num-kv-heads", type=int, default=8)
     parser.add_argument("--head-size",
                         type=int,
-                        choices=[64, 80, 96, 112, 128, 256],
+                        choices=[64, 80, 96, 112, 128, 192, 256],
                         default=128)
     parser.add_argument("--block-size", type=int, choices=[16, 32], default=16)
     parser.add_argument("--use-alibi", action="store_true")
@@ -179,11 +183,11 @@ if __name__ == '__main__':
     parser.add_argument(
         "--kv-cache-dtype",
         type=str,
-        choices=["auto", "fp8_e5m2"],
+        choices=["auto", "fp8", "fp8_e5m2", "fp8_e4m3"],
         default="auto",
-        help=
-        'Data type for kv cache storage. If "auto", will use model data type.')
-    parser.add_argument("--device", type=str, choices=["cuda"], default="cuda")
+        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)")
     args = parser.parse_args()
     print(args)
 
@@ -192,7 +196,7 @@ if __name__ == '__main__':
     main(
         version=args.version,
         num_seqs=args.batch_size,
-        context_len=args.context_len,
+        seq_len=args.seq_len,
         num_query_heads=args.num_query_heads,
         num_kv_heads=args.num_kv_heads,
         head_size=args.head_size,

benchmarks/kernels/benchmark_rope.py

@@ -0,0 +1,121 @@
+import argparse
+from itertools import accumulate
+from typing import Optional
+
+import nvtx
+import torch
+
+from vllm.model_executor.layers.rotary_embedding import get_rope
+
+
+def benchmark_rope_kernels_multi_lora(
+    is_neox_style: bool,
+    batch_size: int,
+    seq_len: int,
+    num_heads: int,
+    head_size: int,
+    rotary_dim: Optional[int],
+    dtype: torch.dtype,
+    seed: int,
+    device: str,
+    max_position: int = 8192,
+    base: int = 10000,
+) -> None:
+    torch.random.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed)
+    torch.set_default_device(device)
+    if rotary_dim is None:
+        rotary_dim = head_size
+    # silulating serving 4 LoRAs
+    scaling_factors = [1, 2, 4, 8]
+    # batched RoPE can take multiple scaling factors
+    batched_rope = get_rope(head_size, rotary_dim, max_position, base,
+                            is_neox_style, {
+                                "type": "linear",
+                                "factor": tuple(scaling_factors)
+                            })
+    # non-batched RoPE takes only one scaling factor, we create multiple
+    # instances to simulate the same behavior
+    non_batched_ropes = []
+    for scaling_factor in scaling_factors:
+        non_batched_ropes.append(
+            get_rope(head_size, rotary_dim, max_position, base, is_neox_style,
+                     {
+                         "type": "linear",
+                         "factor": (scaling_factor, )
+                     }))
+
+    positions = torch.randint(0, max_position, (batch_size, seq_len))
+    query = torch.randn(batch_size,
+                        seq_len,
+                        num_heads * head_size,
+                        dtype=dtype)
+    key = torch.randn_like(query)
+
+    # create query offsets for batched RoPE, we concat multiple kv cache
+    # together and each query needs to find the right kv cache of its type
+    offset_map = torch.tensor(
+        list(
+            accumulate([0] + [
+                max_position * scaling_factor * 2
+                for scaling_factor in scaling_factors[:-1]
+            ])))
+    query_types = torch.randint(0,
+                                len(scaling_factors), (batch_size, seq_len),
+                                device=device)
+    # map query types to offsets
+    query_offsets = offset_map[query_types]
+    # the kernel takes flattened offsets
+    flatten_offsets = query_offsets.flatten()
+
+    # batched queries of the same type together for non-batched RoPE
+    queries = [query[query_types == i] for i in range(len(scaling_factors))]
+    keys = [key[query_types == i] for i in range(len(scaling_factors))]
+    packed_qkr = zip(queries, keys, non_batched_ropes)
+    # synchronize before start timing
+    torch.cuda.synchronize()
+    with nvtx.annotate("non-batched", color="yellow"):
+        for q, k, r in packed_qkr:
+            r.forward(positions, q, k)
+    torch.cuda.synchronize()
+    with nvtx.annotate("batched", color="green"):
+        batched_rope.forward(positions, query, key, flatten_offsets)
+    torch.cuda.synchronize()
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(
+        description="Benchmark the rotary embedding kernels.")
+    parser.add_argument("--is-neox-style", type=bool, default=True)
+    parser.add_argument("--batch-size", type=int, default=16)
+    parser.add_argument("--seq-len", type=int, default=512)
+    parser.add_argument("--num-heads", type=int, default=8)
+    parser.add_argument("--head-size",
+                        type=int,
+                        choices=[64, 80, 96, 112, 128, 192, 256],
+                        default=128)
+    parser.add_argument("--rotary-dim", type=int, choices=[16, 32], default=32)
+    parser.add_argument("--dtype",
+                        type=str,
+                        choices=["bfloat16", "float"],
+                        default="float")
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--device",
+                        type=str,
+                        choices=["cuda:0", "cuda:1"],
+                        default="cuda:0")
+    args = parser.parse_args()
+    print(args)
+
+    benchmark_rope_kernels_multi_lora(
+        is_neox_style=args.is_neox_style,
+        batch_size=args.batch_size,
+        seq_len=args.seq_len,
+        num_heads=args.num_heads,
+        head_size=args.head_size,
+        rotary_dim=args.rotary_dim,
+        dtype=getattr(torch, args.dtype),
+        seed=args.seed,
+        device=args.device,
+    )

benchmarks/kernels/benchmark_shapes.py

@@ -0,0 +1,75 @@
+WEIGHT_SHAPES = {
+    "ideal": [[4 * 256 * 32, 256 * 32]],
+    "mistralai/Mistral-7B-v0.1/TP1": [
+        [4096, 6144],
+        [4096, 4096],
+        [4096, 28672],
+        [14336, 4096],
+    ],
+    "mistralai/Mistral-7B-v0.1/TP2": [
+        [4096, 3072],
+        [2048, 4096],
+        [4096, 14336],
+        [7168, 4096],
+    ],
+    "mistralai/Mistral-7B-v0.1/TP4": [
+        [4096, 1536],
+        [1024, 4096],
+        [4096, 7168],
+        [3584, 4096],
+    ],
+    "meta-llama/Llama-2-7b-hf/TP1": [
+        [4096, 12288],
+        [4096, 4096],
+        [4096, 22016],
+        [11008, 4096],
+    ],
+    "meta-llama/Llama-2-7b-hf/TP2": [
+        [4096, 6144],
+        [2048, 4096],
+        [4096, 11008],
+        [5504, 4096],
+    ],
+    "meta-llama/Llama-2-7b-hf/TP4": [
+        [4096, 3072],
+        [1024, 4096],
+        [4096, 5504],
+        [2752, 4096],
+    ],
+    "meta-llama/Llama-2-13b-hf/TP1": [
+        [5120, 15360],
+        [5120, 5120],
+        [5120, 27648],
+        [13824, 5120],
+    ],
+    "meta-llama/Llama-2-13b-hf/TP2": [
+        [5120, 7680],
+        [2560, 5120],
+        [5120, 13824],
+        [6912, 5120],
+    ],
+    "meta-llama/Llama-2-13b-hf/TP4": [
+        [5120, 3840],
+        [1280, 5120],
+        [5120, 6912],
+        [3456, 5120],
+    ],
+    "meta-llama/Llama-2-70b-hf/TP1": [
+        [8192, 10240],
+        [8192, 8192],
+        [8192, 57344],
+        [28672, 8192],
+    ],
+    "meta-llama/Llama-2-70b-hf/TP2": [
+        [8192, 5120],
+        [4096, 8192],
+        [8192, 28672],
+        [14336, 8192],
+    ],
+    "meta-llama/Llama-2-70b-hf/TP4": [
+        [8192, 2560],
+        [2048, 8192],
+        [8192, 14336],
+        [7168, 8192],
+    ],
+}

benchmarks/launch_tgi_server.sh

@@ -4,7 +4,7 @@ PORT=8000
 MODEL=$1
 TOKENS=$2
 
-docker run --gpus all --shm-size 1g -p $PORT:80 \
+docker run -e HF_TOKEN=$HF_TOKEN --gpus all --shm-size 1g -p $PORT:80 \
            -v $PWD/data:/data \
            ghcr.io/huggingface/text-generation-inference:1.4.0 \
            --model-id $MODEL \

benchmarks/overheads/benchmark_hashing.py

@@ -0,0 +1,63 @@
+import argparse
+import cProfile
+import pstats
+
+from vllm import LLM, SamplingParams
+
+# A very long prompt, total number of tokens is about 15k.
+LONG_PROMPT = ["You are an expert in large language models, aren't you?"
+               ] * 1000
+LONG_PROMPT = ' '.join(LONG_PROMPT)
+
+
+def main(args):
+    llm = LLM(
+        model=args.model,
+        enforce_eager=True,
+        enable_prefix_caching=True,
+        tensor_parallel_size=args.tensor_parallel_size,
+        use_v2_block_manager=args.use_v2_block_manager,
+    )
+
+    sampling_params = SamplingParams(temperature=0, max_tokens=args.output_len)
+    profiler = cProfile.Profile()
+
+    print("------warm up------")
+    for i in range(3):
+        output = llm.generate(LONG_PROMPT, sampling_params)
+        print(output[0].outputs[0].text)
+
+    print("------start generating------")
+    for i in range(3):
+        profiler.runctx('llm.generate(LONG_PROMPT, sampling_params)',
+                        globals(), locals())
+
+    # analyze the runtime of hashing function
+    stats = pstats.Stats(profiler)
+    stats.sort_stats('cumulative')
+    total_time = 0
+    total_calls = 0
+    for func in stats.stats:
+        if 'hash_of_block' in func[2]:
+            total_time = stats.stats[func][3]
+            total_calls = stats.stats[func][0]
+    percentage = (total_time / stats.total_tt) * 100
+    print(f"Hashing took {total_time:.2f} seconds,"
+          f"{percentage:.2f}% of the total runtime.")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description='Benchmark the performance of hashing function in'
+        'automatic prefix caching.')
+    parser.add_argument('--model', type=str, default='lmsys/longchat-7b-16k')
+    parser.add_argument('--tensor-parallel-size', '-tp', type=int, default=1)
+    parser.add_argument('--output-len', type=int, default=10)
+    parser.add_argument('--enable-prefix-caching',
+                        action='store_true',
+                        help='enable prefix caching')
+    parser.add_argument('--use-v2-block-manager',
+                        action='store_true',
+                        help='Use BlockSpaceMangerV2')
+    args = parser.parse_args()
+    main(args)

benchmarks/sonnet.txt

@@ -0,0 +1,518 @@
+FROM fairest creatures we desire increase,
+That thereby beauty's rose might never die,
+But as the riper should by time decease,
+His tender heir might bear his memory:
+But thou, contracted to thine own bright eyes,
+Feed'st thy light'st flame with self-substantial fuel,
+Making a famine where abundance lies,
+Thyself thy foe, to thy sweet self too cruel.
+Thou that art now the world's fresh ornament
+And only herald to the gaudy spring,
+Within thine own bud buriest thy content
+And, tender churl, makest waste in niggarding.
+Pity the world, or else this glutton be,
+To eat the world's due, by the grave and thee.
+When forty winters shall beseige thy brow,
+And dig deep trenches in thy beauty's field,
+Thy youth's proud livery, so gazed on now,
+Will be a tatter'd weed, of small worth held:
+Then being ask'd where all thy beauty lies,
+Where all the treasure of thy lusty days,
+To say, within thine own deep-sunken eyes,
+Were an all-eating shame and thriftless praise.
+How much more praise deserved thy beauty's use,
+If thou couldst answer 'This fair child of mine
+Shall sum my count and make my old excuse,'
+Proving his beauty by succession thine!
+This were to be new made when thou art old,
+And see thy blood warm when thou feel'st it cold.
+Look in thy glass, and tell the face thou viewest
+Now is the time that face should form another;
+Whose fresh repair if now thou not renewest,
+Thou dost beguile the world, unbless some mother.
+For where is she so fair whose unear'd womb
+Disdains the tillage of thy husbandry?
+Or who is he so fond will be the tomb
+Of his self-love, to stop posterity?
+Thou art thy mother's glass, and she in thee
+Calls back the lovely April of her prime:
+So thou through windows of thine age shall see
+Despite of wrinkles this thy golden time.
+But if thou live, remember'd not to be,
+Die single, and thine image dies with thee.
+Unthrifty loveliness, why dost thou spend
+Upon thyself thy beauty's legacy?
+Nature's bequest gives nothing but doth lend,
+And being frank she lends to those are free.
+Then, beauteous niggard, why dost thou abuse
+The bounteous largess given thee to give?
+Profitless usurer, why dost thou use
+So great a sum of sums, yet canst not live?
+For having traffic with thyself alone,
+Thou of thyself thy sweet self dost deceive.
+Then how, when nature calls thee to be gone,
+What acceptable audit canst thou leave?
+Thy unused beauty must be tomb'd with thee,
+Which, used, lives th' executor to be.
+Those hours, that with gentle work did frame
+The lovely gaze where every eye doth dwell,
+Will play the tyrants to the very same
+And that unfair which fairly doth excel:
+For never-resting time leads summer on
+To hideous winter and confounds him there;
+Sap cheque'd with frost and lusty leaves quite gone,
+Beauty o'ersnow'd and bareness every where:
+Then, were not summer's distillation left,
+A liquid prisoner pent in walls of glass,
+Beauty's effect with beauty were bereft,
+Nor it nor no remembrance what it was:
+But flowers distill'd though they with winter meet,
+Leese but their show; their substance still lives sweet.
+Then let not winter's ragged hand deface
+In thee thy summer, ere thou be distill'd:
+Make sweet some vial; treasure thou some place
+With beauty's treasure, ere it be self-kill'd.
+That use is not forbidden usury,
+Which happies those that pay the willing loan;
+That's for thyself to breed another thee,
+Or ten times happier, be it ten for one;
+Ten times thyself were happier than thou art,
+If ten of thine ten times refigured thee:
+Then what could death do, if thou shouldst depart,
+Leaving thee living in posterity?
+Be not self-will'd, for thou art much too fair
+To be death's conquest and make worms thine heir.
+Lo! in the orient when the gracious light
+Lifts up his burning head, each under eye
+Doth homage to his new-appearing sight,
+Serving with looks his sacred majesty;
+And having climb'd the steep-up heavenly hill,
+Resembling strong youth in his middle age,
+yet mortal looks adore his beauty still,
+Attending on his golden pilgrimage;
+But when from highmost pitch, with weary car,
+Like feeble age, he reeleth from the day,
+The eyes, 'fore duteous, now converted are
+From his low tract and look another way:
+So thou, thyself out-going in thy noon,
+Unlook'd on diest, unless thou get a son.
+Music to hear, why hear'st thou music sadly?
+Sweets with sweets war not, joy delights in joy.
+Why lovest thou that which thou receivest not gladly,
+Or else receivest with pleasure thine annoy?
+If the true concord of well-tuned sounds,
+By unions married, do offend thine ear,
+They do but sweetly chide thee, who confounds
+In singleness the parts that thou shouldst bear.
+Mark how one string, sweet husband to another,
+Strikes each in each by mutual ordering,
+Resembling sire and child and happy mother
+Who all in one, one pleasing note do sing:
+Whose speechless song, being many, seeming one,
+Sings this to thee: 'thou single wilt prove none.'
+Is it for fear to wet a widow's eye
+That thou consumest thyself in single life?
+Ah! if thou issueless shalt hap to die.
+The world will wail thee, like a makeless wife;
+The world will be thy widow and still weep
+That thou no form of thee hast left behind,
+When every private widow well may keep
+By children's eyes her husband's shape in mind.
+Look, what an unthrift in the world doth spend
+Shifts but his place, for still the world enjoys it;
+But beauty's waste hath in the world an end,
+And kept unused, the user so destroys it.
+No love toward others in that bosom sits
+That on himself such murderous shame commits.
+For shame! deny that thou bear'st love to any,
+Who for thyself art so unprovident.
+Grant, if thou wilt, thou art beloved of many,
+But that thou none lovest is most evident;
+For thou art so possess'd with murderous hate
+That 'gainst thyself thou stick'st not to conspire.
+Seeking that beauteous roof to ruinate
+Which to repair should be thy chief desire.
+O, change thy thought, that I may change my mind!
+Shall hate be fairer lodged than gentle love?
+Be, as thy presence is, gracious and kind,
+Or to thyself at least kind-hearted prove:
+Make thee another self, for love of me,
+That beauty still may live in thine or thee.
+As fast as thou shalt wane, so fast thou growest
+In one of thine, from that which thou departest;
+And that fresh blood which youngly thou bestowest
+Thou mayst call thine when thou from youth convertest.
+Herein lives wisdom, beauty and increase:
+Without this, folly, age and cold decay:
+If all were minded so, the times should cease
+And threescore year would make the world away.
+Let those whom Nature hath not made for store,
+Harsh featureless and rude, barrenly perish:
+Look, whom she best endow'd she gave the more;
+Which bounteous gift thou shouldst in bounty cherish:
+She carved thee for her seal, and meant thereby
+Thou shouldst print more, not let that copy die.
+When I do count the clock that tells the time,
+And see the brave day sunk in hideous night;
+When I behold the violet past prime,
+And sable curls all silver'd o'er with white;
+When lofty trees I see barren of leaves
+Which erst from heat did canopy the herd,
+And summer's green all girded up in sheaves
+Borne on the bier with white and bristly beard,
+Then of thy beauty do I question make,
+That thou among the wastes of time must go,
+Since sweets and beauties do themselves forsake
+And die as fast as they see others grow;
+And nothing 'gainst Time's scythe can make defence
+Save breed, to brave him when he takes thee hence.
+O, that you were yourself! but, love, you are
+No longer yours than you yourself here live:
+Against this coming end you should prepare,
+And your sweet semblance to some other give.
+So should that beauty which you hold in lease
+Find no determination: then you were
+Yourself again after yourself's decease,
+When your sweet issue your sweet form should bear.
+Who lets so fair a house fall to decay,
+Which husbandry in honour might uphold
+Against the stormy gusts of winter's day
+And barren rage of death's eternal cold?
+O, none but unthrifts! Dear my love, you know
+You had a father: let your son say so.
+Not from the stars do I my judgment pluck;
+And yet methinks I have astronomy,
+But not to tell of good or evil luck,
+Of plagues, of dearths, or seasons' quality;
+Nor can I fortune to brief minutes tell,
+Pointing to each his thunder, rain and wind,
+Or say with princes if it shall go well,
+By oft predict that I in heaven find:
+But from thine eyes my knowledge I derive,
+And, constant stars, in them I read such art
+As truth and beauty shall together thrive,
+If from thyself to store thou wouldst convert;
+Or else of thee this I prognosticate:
+Thy end is truth's and beauty's doom and date.
+When I consider every thing that grows
+Holds in perfection but a little moment,
+That this huge stage presenteth nought but shows
+Whereon the stars in secret influence comment;
+When I perceive that men as plants increase,
+Cheered and cheque'd even by the self-same sky,
+Vaunt in their youthful sap, at height decrease,
+And wear their brave state out of memory;
+Then the conceit of this inconstant stay
+Sets you most rich in youth before my sight,
+Where wasteful Time debateth with Decay,
+To change your day of youth to sullied night;
+And all in war with Time for love of you,
+As he takes from you, I engraft you new.
+But wherefore do not you a mightier way
+Make war upon this bloody tyrant, Time?
+And fortify yourself in your decay
+With means more blessed than my barren rhyme?
+Now stand you on the top of happy hours,
+And many maiden gardens yet unset
+With virtuous wish would bear your living flowers,
+Much liker than your painted counterfeit:
+So should the lines of life that life repair,
+Which this, Time's pencil, or my pupil pen,
+Neither in inward worth nor outward fair,
+Can make you live yourself in eyes of men.
+To give away yourself keeps yourself still,
+And you must live, drawn by your own sweet skill.
+Who will believe my verse in time to come,
+If it were fill'd with your most high deserts?
+Though yet, heaven knows, it is but as a tomb
+Which hides your life and shows not half your parts.
+If I could write the beauty of your eyes
+And in fresh numbers number all your graces,
+The age to come would say 'This poet lies:
+Such heavenly touches ne'er touch'd earthly faces.'
+So should my papers yellow'd with their age
+Be scorn'd like old men of less truth than tongue,
+And your true rights be term'd a poet's rage
+And stretched metre of an antique song:
+But were some child of yours alive that time,
+You should live twice; in it and in my rhyme.
+Shall I compare thee to a summer's day?
+Thou art more lovely and more temperate:
+Rough winds do shake the darling buds of May,
+And summer's lease hath all too short a date:
+Sometime too hot the eye of heaven shines,
+And often is his gold complexion dimm'd;
+And every fair from fair sometime declines,
+By chance or nature's changing course untrimm'd;
+But thy eternal summer shall not fade
+Nor lose possession of that fair thou owest;
+Nor shall Death brag thou wander'st in his shade,
+When in eternal lines to time thou growest:
+So long as men can breathe or eyes can see,
+So long lives this and this gives life to thee.
+Devouring Time, blunt thou the lion's paws,
+And make the earth devour her own sweet brood;
+Pluck the keen teeth from the fierce tiger's jaws,
+And burn the long-lived phoenix in her blood;
+Make glad and sorry seasons as thou fleets,
+And do whate'er thou wilt, swift-footed Time,
+To the wide world and all her fading sweets;
+But I forbid thee one most heinous crime:
+O, carve not with thy hours my love's fair brow,
+Nor draw no lines there with thine antique pen;
+Him in thy course untainted do allow
+For beauty's pattern to succeeding men.
+Yet, do thy worst, old Time: despite thy wrong,
+My love shall in my verse ever live young.
+A woman's face with Nature's own hand painted
+Hast thou, the master-mistress of my passion;
+A woman's gentle heart, but not acquainted
+With shifting change, as is false women's fashion;
+An eye more bright than theirs, less false in rolling,
+Gilding the object whereupon it gazeth;
+A man in hue, all 'hues' in his controlling,
+Much steals men's eyes and women's souls amazeth.
+And for a woman wert thou first created;
+Till Nature, as she wrought thee, fell a-doting,
+And by addition me of thee defeated,
+By adding one thing to my purpose nothing.
+But since she prick'd thee out for women's pleasure,
+Mine be thy love and thy love's use their treasure.
+So is it not with me as with that Muse
+Stirr'd by a painted beauty to his verse,
+Who heaven itself for ornament doth use
+And every fair with his fair doth rehearse
+Making a couplement of proud compare,
+With sun and moon, with earth and sea's rich gems,
+With April's first-born flowers, and all things rare
+That heaven's air in this huge rondure hems.
+O' let me, true in love, but truly write,
+And then believe me, my love is as fair
+As any mother's child, though not so bright
+As those gold candles fix'd in heaven's air:
+Let them say more than like of hearsay well;
+I will not praise that purpose not to sell.
+My glass shall not persuade me I am old,
+So long as youth and thou are of one date;
+But when in thee time's furrows I behold,
+Then look I death my days should expiate.
+For all that beauty that doth cover thee
+Is but the seemly raiment of my heart,
+Which in thy breast doth live, as thine in me:
+How can I then be elder than thou art?
+O, therefore, love, be of thyself so wary
+As I, not for myself, but for thee will;
+Bearing thy heart, which I will keep so chary
+As tender nurse her babe from faring ill.
+Presume not on thy heart when mine is slain;
+Thou gavest me thine, not to give back again.
+As an unperfect actor on the stage
+Who with his fear is put besides his part,
+Or some fierce thing replete with too much rage,
+Whose strength's abundance weakens his own heart.
+So I, for fear of trust, forget to say
+The perfect ceremony of love's rite,
+And in mine own love's strength seem to decay,
+O'ercharged with burden of mine own love's might.
+O, let my books be then the eloquence
+And dumb presagers of my speaking breast,
+Who plead for love and look for recompense
+More than that tongue that more hath more express'd.
+O, learn to read what silent love hath writ:
+To hear with eyes belongs to love's fine wit.
+Mine eye hath play'd the painter and hath stell'd
+Thy beauty's form in table of my heart;
+My body is the frame wherein 'tis held,
+And perspective it is the painter's art.
+For through the painter must you see his skill,
+To find where your true image pictured lies;
+Which in my bosom's shop is hanging still,
+That hath his windows glazed with thine eyes.
+Now see what good turns eyes for eyes have done:
+Mine eyes have drawn thy shape, and thine for me
+Are windows to my breast, where-through the sun
+Delights to peep, to gaze therein on thee;
+Yet eyes this cunning want to grace their art;
+They draw but what they see, know not the heart.
+Let those who are in favour with their stars
+Of public honour and proud titles boast,
+Whilst I, whom fortune of such triumph bars,
+Unlook'd for joy in that I honour most.
+Great princes' favourites their fair leaves spread
+But as the marigold at the sun's eye,
+And in themselves their pride lies buried,
+For at a frown they in their glory die.
+The painful warrior famoused for fight,
+After a thousand victories once foil'd,
+Is from the book of honour razed quite,
+And all the rest forgot for which he toil'd:
+Then happy I, that love and am beloved
+Where I may not remove nor be removed.
+Lord of my love, to whom in vassalage
+Thy merit hath my duty strongly knit,
+To thee I send this written embassage,
+To witness duty, not to show my wit:
+Duty so great, which wit so poor as mine
+May make seem bare, in wanting words to show it,
+But that I hope some good conceit of thine
+In thy soul's thought, all naked, will bestow it;
+Till whatsoever star that guides my moving
+Points on me graciously with fair aspect
+And puts apparel on my tatter'd loving,
+To show me worthy of thy sweet respect:
+Then may I dare to boast how I do love thee;
+Till then not show my head where thou mayst prove me.
+Weary with toil, I haste me to my bed,
+The dear repose for limbs with travel tired;
+But then begins a journey in my head,
+To work my mind, when body's work's expired:
+For then my thoughts, from far where I abide,
+Intend a zealous pilgrimage to thee,
+And keep my drooping eyelids open wide,
+Looking on darkness which the blind do see
+Save that my soul's imaginary sight
+Presents thy shadow to my sightless view,
+Which, like a jewel hung in ghastly night,
+Makes black night beauteous and her old face new.
+Lo! thus, by day my limbs, by night my mind,
+For thee and for myself no quiet find.
+How can I then return in happy plight,
+That am debarr'd the benefit of rest?
+When day's oppression is not eased by night,
+But day by night, and night by day, oppress'd?
+And each, though enemies to either's reign,
+Do in consent shake hands to torture me;
+The one by toil, the other to complain
+How far I toil, still farther off from thee.
+I tell the day, to please them thou art bright
+And dost him grace when clouds do blot the heaven:
+So flatter I the swart-complexion'd night,
+When sparkling stars twire not thou gild'st the even.
+But day doth daily draw my sorrows longer
+And night doth nightly make grief's strength seem stronger.
+When, in disgrace with fortune and men's eyes,
+I all alone beweep my outcast state
+And trouble deal heaven with my bootless cries
+And look upon myself and curse my fate,
+Wishing me like to one more rich in hope,
+Featured like him, like him with friends possess'd,
+Desiring this man's art and that man's scope,
+With what I most enjoy contented least;
+Yet in these thoughts myself almost despising,
+Haply I think on thee, and then my state,
+Like to the lark at break of day arising
+From sullen earth, sings hymns at heaven's gate;
+For thy sweet love remember'd such wealth brings
+That then I scorn to change my state with kings.
+When to the sessions of sweet silent thought
+I summon up remembrance of things past,
+I sigh the lack of many a thing I sought,
+And with old woes new wail my dear time's waste:
+Then can I drown an eye, unused to flow,
+For precious friends hid in death's dateless night,
+And weep afresh love's long since cancell'd woe,
+And moan the expense of many a vanish'd sight:
+Then can I grieve at grievances foregone,
+And heavily from woe to woe tell o'er
+The sad account of fore-bemoaned moan,
+Which I new pay as if not paid before.
+But if the while I think on thee, dear friend,
+All losses are restored and sorrows end.
+Thy bosom is endeared with all hearts,
+Which I by lacking have supposed dead,
+And there reigns love and all love's loving parts,
+And all those friends which I thought buried.
+How many a holy and obsequious tear
+Hath dear religious love stol'n from mine eye
+As interest of the dead, which now appear
+But things removed that hidden in thee lie!
+Thou art the grave where buried love doth live,
+Hung with the trophies of my lovers gone,
+Who all their parts of me to thee did give;
+That due of many now is thine alone:
+Their images I loved I view in thee,
+And thou, all they, hast all the all of me.
+If thou survive my well-contented day,
+When that churl Death my bones with dust shall cover,
+And shalt by fortune once more re-survey
+These poor rude lines of thy deceased lover,
+Compare them with the bettering of the time,
+And though they be outstripp'd by every pen,
+Reserve them for my love, not for their rhyme,
+Exceeded by the height of happier men.
+O, then vouchsafe me but this loving thought:
+'Had my friend's Muse grown with this growing age,
+A dearer birth than this his love had brought,
+To march in ranks of better equipage:
+But since he died and poets better prove,
+Theirs for their style I'll read, his for his love.'
+Full many a glorious morning have I seen
+Flatter the mountain-tops with sovereign eye,
+Kissing with golden face the meadows green,
+Gilding pale streams with heavenly alchemy;
+Anon permit the basest clouds to ride
+With ugly rack on his celestial face,
+And from the forlorn world his visage hide,
+Stealing unseen to west with this disgrace:
+Even so my sun one early morn did shine
+With all triumphant splendor on my brow;
+But out, alack! he was but one hour mine;
+The region cloud hath mask'd him from me now.
+Yet him for this my love no whit disdaineth;
+Suns of the world may stain when heaven's sun staineth.
+Why didst thou promise such a beauteous day,
+And make me travel forth without my cloak,
+To let base clouds o'ertake me in my way,
+Hiding thy bravery in their rotten smoke?
+'Tis not enough that through the cloud thou break,
+To dry the rain on my storm-beaten face,
+For no man well of such a salve can speak
+That heals the wound and cures not the disgrace:
+Nor can thy shame give physic to my grief;
+Though thou repent, yet I have still the loss:
+The offender's sorrow lends but weak relief
+To him that bears the strong offence's cross.
+Ah! but those tears are pearl which thy love sheds,
+And they are rich and ransom all ill deeds.
+No more be grieved at that which thou hast done:
+Roses have thorns, and silver fountains mud;
+Clouds and eclipses stain both moon and sun,
+And loathsome canker lives in sweetest bud.
+All men make faults, and even I in this,
+Authorizing thy trespass with compare,
+Myself corrupting, salving thy amiss,
+Excusing thy sins more than thy sins are;
+For to thy sensual fault I bring in sense--
+Thy adverse party is thy advocate--
+And 'gainst myself a lawful plea commence:
+Such civil war is in my love and hate
+That I an accessary needs must be
+To that sweet thief which sourly robs from me.
+Let me confess that we two must be twain,
+Although our undivided loves are one:
+So shall those blots that do with me remain
+Without thy help by me be borne alone.
+In our two loves there is but one respect,
+Though in our lives a separable spite,
+Which though it alter not love's sole effect,
+Yet doth it steal sweet hours from love's delight.
+I may not evermore acknowledge thee,
+Lest my bewailed guilt should do thee shame,
+Nor thou with public kindness honour me,
+Unless thou take that honour from thy name:
+But do not so; I love thee in such sort
+As, thou being mine, mine is thy good report.
+As a decrepit father takes delight
+To see his active child do deeds of youth,
+So I, made lame by fortune's dearest spite,
+Take all my comfort of thy worth and truth.
+For whether beauty, birth, or wealth, or wit,
+Or any of these all, or all, or more,
+Entitled in thy parts do crowned sit,
+I make my love engrafted to this store:
+So then I am not lame, poor, nor despised,
+Whilst that this shadow doth such substance give
+That I in thy abundance am sufficed
+And by a part of all thy glory live.
+Look, what is best, that best I wish in thee:
+This wish I have; then ten times happy me!

cmake/cpu_extension.cmake

@@ -0,0 +1,90 @@
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+#
+# Define environment variables for special configurations
+#
+if(DEFINED ENV{VLLM_CPU_AVX512BF16})
+    set(ENABLE_AVX512BF16 ON)
+endif()
+
+include_directories("${CMAKE_SOURCE_DIR}/csrc")
+
+#
+# Check the compile flags
+#
+list(APPEND CXX_COMPILE_FLAGS 
+    "-fopenmp"
+    "-DVLLM_CPU_EXTENSION")
+
+execute_process(COMMAND cat /proc/cpuinfo
+                RESULT_VARIABLE CPUINFO_RET
+                OUTPUT_VARIABLE CPUINFO)
+
+if (NOT CPUINFO_RET EQUAL 0)
+    message(FATAL_ERROR "Failed to check CPU features via /proc/cpuinfo")
+endif()
+
+function (find_isa CPUINFO TARGET OUT)
+    string(FIND ${CPUINFO} ${TARGET} ISA_FOUND)
+    if(NOT ISA_FOUND EQUAL -1)
+        set(${OUT} ON PARENT_SCOPE)
+    else()
+        set(${OUT} OFF PARENT_SCOPE)
+    endif()
+endfunction()
+
+find_isa(${CPUINFO} "avx512f" AVX512_FOUND)
+
+if (AVX512_FOUND)
+    list(APPEND CXX_COMPILE_FLAGS
+        "-mavx512f"
+        "-mavx512vl"
+        "-mavx512bw"
+        "-mavx512dq")
+
+    find_isa(${CPUINFO} "avx512_bf16" AVX512BF16_FOUND)
+    if (AVX512BF16_FOUND OR ENABLE_AVX512BF16)
+        if (CMAKE_CXX_COMPILER_ID STREQUAL "GNU" AND 
+            CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 12.3) 
+            list(APPEND CXX_COMPILE_FLAGS "-mavx512bf16")
+        else()
+            message(WARNING "Disable AVX512-BF16 ISA support, requires gcc/g++ >= 12.3")
+        endif()
+    else()
+        message(WARNING "Disable AVX512-BF16 ISA support, no avx512_bf16 found in local CPU flags." " If cross-compilation is required, please set env VLLM_CPU_AVX512BF16=1.")
+    endif()
+else()
+    message(FATAL_ERROR "vLLM CPU backend requires AVX512 ISA support.")
+endif()
+
+message(STATUS "CPU extension compile flags: ${CXX_COMPILE_FLAGS}")
+
+
+#
+# Define extension targets
+#
+
+#
+# _C extension
+#
+set(VLLM_EXT_SRC
+    "csrc/cpu/activation.cpp"
+    "csrc/cpu/attention.cpp"
+    "csrc/cpu/cache.cpp"
+    "csrc/cpu/layernorm.cpp"
+    "csrc/cpu/pos_encoding.cpp"
+    "csrc/cpu/pybind.cpp")
+
+define_gpu_extension_target(
+    _C
+    DESTINATION vllm
+    LANGUAGE CXX
+    SOURCES ${VLLM_EXT_SRC}
+    COMPILE_FLAGS ${CXX_COMPILE_FLAGS}
+    WITH_SOABI 
+)
+
+add_custom_target(default)
+message(STATUS "Enabling C extension.")
+add_dependencies(default _C)
+

cmake/hipify.py

@@ -0,0 +1,73 @@
+#!/usr/bin/env python3
+
+#
+# A command line tool for running pytorch's hipify preprocessor on CUDA
+# source files.
+#
+# See https://github.com/ROCm/hipify_torch
+# and <torch install dir>/utils/hipify/hipify_python.py
+#
+
+import argparse
+import os
+import shutil
+
+from torch.utils.hipify.hipify_python import hipify
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+
+    # Project directory where all the source + include files live.
+    parser.add_argument(
+        "-p",
+        "--project_dir",
+        help="The project directory.",
+    )
+
+    # Directory where hipified files are written.
+    parser.add_argument(
+        "-o",
+        "--output_dir",
+        help="The output directory.",
+    )
+
+    # Source files to convert.
+    parser.add_argument("sources",
+                        help="Source files to hipify.",
+                        nargs="*",
+                        default=[])
+
+    args = parser.parse_args()
+
+    # Limit include scope to project_dir only
+    includes = [os.path.join(args.project_dir, '*')]
+
+    # Get absolute path for all source files.
+    extra_files = [os.path.abspath(s) for s in args.sources]
+
+    # Copy sources from project directory to output directory.
+    # The directory might already exist to hold object files so we ignore that.
+    shutil.copytree(args.project_dir, args.output_dir, dirs_exist_ok=True)
+
+    hipify_result = hipify(project_directory=args.project_dir,
+                           output_directory=args.output_dir,
+                           header_include_dirs=[],
+                           includes=includes,
+                           extra_files=extra_files,
+                           show_detailed=True,
+                           is_pytorch_extension=True,
+                           hipify_extra_files_only=True)
+
+    hipified_sources = []
+    for source in args.sources:
+        s_abs = os.path.abspath(source)
+        hipified_s_abs = (hipify_result[s_abs].hipified_path if
+                          (s_abs in hipify_result
+                           and hipify_result[s_abs].hipified_path is not None)
+                          else s_abs)
+        hipified_sources.append(hipified_s_abs)
+
+    assert (len(hipified_sources) == len(args.sources))
+
+    # Print hipified source files.
+    print("\n".join(hipified_sources))

cmake/utils.cmake

@@ -0,0 +1,354 @@
+#
+# Attempt to find the python package that uses the same python executable as
+# `EXECUTABLE` and is one of the `SUPPORTED_VERSIONS`.
+#
+macro (find_python_from_executable EXECUTABLE SUPPORTED_VERSIONS)
+  file(REAL_PATH ${EXECUTABLE} EXECUTABLE)
+  set(Python_EXECUTABLE ${EXECUTABLE})
+  find_package(Python COMPONENTS Interpreter Development.Module)
+  if (NOT Python_FOUND)
+    message(FATAL_ERROR "Unable to find python matching: ${EXECUTABLE}.")
+  endif()
+  set(_VER "${Python_VERSION_MAJOR}.${Python_VERSION_MINOR}")
+  set(_SUPPORTED_VERSIONS_LIST ${SUPPORTED_VERSIONS} ${ARGN})
+  if (NOT _VER IN_LIST _SUPPORTED_VERSIONS_LIST)
+    message(FATAL_ERROR
+      "Python version (${_VER}) is not one of the supported versions: "
+      "${_SUPPORTED_VERSIONS_LIST}.")
+  endif()
+  message(STATUS "Found python matching: ${EXECUTABLE}.")
+endmacro()
+
+#
+# Run `EXPR` in python.  The standard output of python is stored in `OUT` and
+# has trailing whitespace stripped.  If an error is encountered when running
+# python, a fatal message `ERR_MSG` is issued.
+#
+function (run_python OUT EXPR ERR_MSG)
+  execute_process(
+    COMMAND
+    "${Python_EXECUTABLE}" "-c" "${EXPR}"
+    OUTPUT_VARIABLE PYTHON_OUT
+    RESULT_VARIABLE PYTHON_ERROR_CODE
+    ERROR_VARIABLE PYTHON_STDERR
+    OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+  if(NOT PYTHON_ERROR_CODE EQUAL 0)
+    message(FATAL_ERROR "${ERR_MSG}: ${PYTHON_STDERR}")
+  endif()
+  set(${OUT} ${PYTHON_OUT} PARENT_SCOPE)
+endfunction()
+
+# Run `EXPR` in python after importing `PKG`. Use the result of this to extend
+# `CMAKE_PREFIX_PATH` so the torch cmake configuration can be imported.
+macro (append_cmake_prefix_path PKG EXPR)
+  run_python(_PREFIX_PATH
+    "import ${PKG}; print(${EXPR})" "Failed to locate ${PKG} path")
+  list(APPEND CMAKE_PREFIX_PATH ${_PREFIX_PATH})
+endmacro()
+
+#
+# Add a target named `hipify${NAME}` that runs the hipify preprocessor on a set
+# of CUDA source files. The names of the corresponding "hipified" sources are
+# stored in `OUT_SRCS`.
+#
+function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
+  #
+  # Split into C++ and non-C++ (i.e. CUDA) sources.
+  #
+  set(SRCS ${ORIG_SRCS})
+  set(CXX_SRCS ${ORIG_SRCS})
+  list(FILTER SRCS EXCLUDE REGEX "\.(cc)|(cpp)$")
+  list(FILTER CXX_SRCS INCLUDE REGEX "\.(cc)|(cpp)$")
+
+  #
+  # Generate ROCm/HIP source file names from CUDA file names.
+  # Since HIP files are generated code, they will appear in the build area
+  # `CMAKE_CURRENT_BINARY_DIR` directory rather than the original csrc dir.
+  #
+  set(HIP_SRCS)
+  foreach (SRC ${SRCS})
+    string(REGEX REPLACE "\.cu$" "\.hip" SRC ${SRC})
+    string(REGEX REPLACE "cuda" "hip" SRC ${SRC})
+    list(APPEND HIP_SRCS "${CMAKE_CURRENT_BINARY_DIR}/${SRC}")
+  endforeach()
+
+  set(CSRC_BUILD_DIR ${CMAKE_CURRENT_BINARY_DIR}/csrc)
+  add_custom_target(
+    hipify${NAME}
+    COMMAND ${CMAKE_SOURCE_DIR}/cmake/hipify.py -p ${CMAKE_SOURCE_DIR}/csrc -o ${CSRC_BUILD_DIR} ${SRCS}
+    DEPENDS ${CMAKE_SOURCE_DIR}/cmake/hipify.py ${SRCS}
+    BYPRODUCTS ${HIP_SRCS}
+    COMMENT "Running hipify on ${NAME} extension source files.")
+
+  # Swap out original extension sources with hipified sources.
+  list(APPEND HIP_SRCS ${CXX_SRCS})
+  set(${OUT_SRCS} ${HIP_SRCS} PARENT_SCOPE)
+endfunction()
+
+#
+# Get additional GPU compiler flags from torch.
+#
+function (get_torch_gpu_compiler_flags OUT_GPU_FLAGS GPU_LANG)
+  if (${GPU_LANG} STREQUAL "CUDA")
+    #
+    # Get common NVCC flags from torch.
+    #
+    run_python(GPU_FLAGS
+      "from torch.utils.cpp_extension import COMMON_NVCC_FLAGS; print(';'.join(COMMON_NVCC_FLAGS))"
+      "Failed to determine torch nvcc compiler flags")
+
+    if (CUDA_VERSION VERSION_GREATER_EQUAL 11.8)
+      list(APPEND GPU_FLAGS "-DENABLE_FP8")
+    endif()
+    if (CUDA_VERSION VERSION_GREATER_EQUAL 12.0)
+      list(REMOVE_ITEM GPU_FLAGS
+        "-D__CUDA_NO_HALF_OPERATORS__"
+        "-D__CUDA_NO_HALF_CONVERSIONS__"
+        "-D__CUDA_NO_BFLOAT16_CONVERSIONS__"
+        "-D__CUDA_NO_HALF2_OPERATORS__")
+    endif()
+
+  elseif(${GPU_LANG} STREQUAL "HIP")
+    #
+    # Get common HIP/HIPCC flags from torch.
+    #
+    run_python(GPU_FLAGS
+      "import torch.utils.cpp_extension as t; print(';'.join(t.COMMON_HIP_FLAGS + t.COMMON_HIPCC_FLAGS))"
+      "Failed to determine torch nvcc compiler flags")
+
+    list(APPEND GPU_FLAGS
+      "-DUSE_ROCM"
+      "-DENABLE_FP8"
+      "-U__HIP_NO_HALF_CONVERSIONS__"
+      "-U__HIP_NO_HALF_OPERATORS__"
+      "-fno-gpu-rdc")
+
+  endif()
+  set(${OUT_GPU_FLAGS} ${GPU_FLAGS} PARENT_SCOPE)
+endfunction()
+
+# Macro for converting a `gencode` version number to a cmake version number.
+macro(string_to_ver OUT_VER IN_STR)
+  string(REGEX REPLACE "\([0-9]+\)\([0-9]\)" "\\1.\\2" ${OUT_VER} ${IN_STR})
+endmacro()
+
+#
+# Override the GPU architectures detected by cmake/torch and filter them by
+# `GPU_SUPPORTED_ARCHES`. Sets the final set of architectures in
+# `GPU_ARCHES`.
+#
+# Note: this is defined as a macro since it updates `CMAKE_CUDA_FLAGS`.
+#
+macro(override_gpu_arches GPU_ARCHES GPU_LANG GPU_SUPPORTED_ARCHES)
+  set(_GPU_SUPPORTED_ARCHES_LIST ${GPU_SUPPORTED_ARCHES} ${ARGN})
+  message(STATUS "${GPU_LANG} supported arches: ${_GPU_SUPPORTED_ARCHES_LIST}")
+
+  if (${GPU_LANG} STREQUAL "HIP")
+    #
+    # `GPU_ARCHES` controls the `--offload-arch` flags.
+    # `CMAKE_HIP_ARCHITECTURES` is set up by torch and can be controlled
+    # via the `PYTORCH_ROCM_ARCH` env variable.
+    #
+
+    #
+    # Find the intersection of the supported + detected architectures to
+    # set the module architecture flags.
+    #
+    set(${GPU_ARCHES})
+    foreach (_ARCH ${CMAKE_HIP_ARCHITECTURES})
+      if (_ARCH IN_LIST _GPU_SUPPORTED_ARCHES_LIST)
+        list(APPEND ${GPU_ARCHES} ${_ARCH})
+      endif()
+    endforeach()
+
+    if(NOT ${GPU_ARCHES})
+      message(FATAL_ERROR
+        "None of the detected ROCm architectures: ${CMAKE_HIP_ARCHITECTURES} is"
+        " supported. Supported ROCm architectures are: ${_GPU_SUPPORTED_ARCHES_LIST}.")
+    endif()
+
+  elseif(${GPU_LANG} STREQUAL "CUDA")
+    #
+    # Setup/process CUDA arch flags.
+    #
+    # The torch cmake setup hardcodes the detected architecture flags in
+    # `CMAKE_CUDA_FLAGS`.  Since `CMAKE_CUDA_FLAGS` is a "global" variable, it
+    # can't modified on a per-target basis, e.g. for the `punica` extension.
+    # So, all the `-gencode` flags need to be extracted and removed from
+    # `CMAKE_CUDA_FLAGS` for processing so they can be passed by another method.
+    # Since it's not possible to use `target_compiler_options` for adding target
+    # specific `-gencode` arguments, the target's `CUDA_ARCHITECTURES` property
+    # must be used instead.  This requires repackaging the architecture flags
+    # into a format that cmake expects for `CUDA_ARCHITECTURES`.
+    #
+    # This is a bit fragile in that it depends on torch using `-gencode` as opposed
+    # to one of the other nvcc options to specify architectures.
+    #
+    # Note: torch uses the `TORCH_CUDA_ARCH_LIST` environment variable to override
+    # detected architectures.
+    #
+    message(DEBUG "initial CMAKE_CUDA_FLAGS: ${CMAKE_CUDA_FLAGS}")
+
+    # Extract all `-gencode` flags from `CMAKE_CUDA_FLAGS`
+    string(REGEX MATCHALL "-gencode arch=[^ ]+" _CUDA_ARCH_FLAGS
+      ${CMAKE_CUDA_FLAGS})
+
+    # Remove all `-gencode` flags from `CMAKE_CUDA_FLAGS` since they will be modified
+    # and passed back via the `CUDA_ARCHITECTURES` property.
+    string(REGEX REPLACE "-gencode arch=[^ ]+ *" "" CMAKE_CUDA_FLAGS
+      ${CMAKE_CUDA_FLAGS})
+
+    # If this error is triggered, it might mean that torch has changed how it sets
+    # up nvcc architecture code generation flags.
+    if (NOT _CUDA_ARCH_FLAGS)
+      message(FATAL_ERROR
+        "Could not find any architecture related code generation flags in "
+        "CMAKE_CUDA_FLAGS. (${CMAKE_CUDA_FLAGS})")
+    endif()
+
+    message(DEBUG "final CMAKE_CUDA_FLAGS: ${CMAKE_CUDA_FLAGS}")
+    message(DEBUG "arch flags: ${_CUDA_ARCH_FLAGS}")
+
+    # Initialize the architecture lists to empty.
+    set(${GPU_ARCHES})
+
+    # Process each `gencode` flag.
+    foreach(_ARCH ${_CUDA_ARCH_FLAGS})
+      # For each flag, extract the version number and whether it refers to PTX
+      # or native code.
+      # Note: if a regex matches then `CMAKE_MATCH_1` holds the binding
+      # for that match.
+
+      string(REGEX MATCH "arch=compute_\([0-9]+a?\)" _COMPUTE ${_ARCH})
+      if (_COMPUTE)
+        set(_COMPUTE ${CMAKE_MATCH_1})
+      endif()
+
+      string(REGEX MATCH "code=sm_\([0-9]+a?\)" _SM ${_ARCH})
+      if (_SM)
+        set(_SM ${CMAKE_MATCH_1})
+      endif()
+
+      string(REGEX MATCH "code=compute_\([0-9]+a?\)" _CODE ${_ARCH})
+      if (_CODE)
+        set(_CODE ${CMAKE_MATCH_1})
+      endif()
+
+      # Make sure the virtual architecture can be matched.
+      if (NOT _COMPUTE)
+        message(FATAL_ERROR
+          "Could not determine virtual architecture from: ${_ARCH}.")
+      endif()
+
+      # One of sm_ or compute_ must exist.
+      if ((NOT _SM) AND (NOT _CODE))
+        message(FATAL_ERROR
+          "Could not determine a codegen architecture from: ${_ARCH}.")
+      endif()
+
+      if (_SM)
+        # -real suffix let CMake to only generate elf code for the kernels.
+        # we want this, otherwise the added ptx (default) will increase binary size.
+        set(_VIRT "-real")
+        set(_CODE_ARCH ${_SM})
+      else()
+        # -virtual suffix let CMake to generate ptx code for the kernels.
+        set(_VIRT "-virtual")
+        set(_CODE_ARCH ${_CODE})
+      endif()
+
+      # Check if the current version is in the supported arch list.
+      string_to_ver(_CODE_VER ${_CODE_ARCH})
+      if (NOT _CODE_VER IN_LIST _GPU_SUPPORTED_ARCHES_LIST)
+        message(STATUS "discarding unsupported CUDA arch ${_VER}.")
+        continue()
+      endif()
+
+      # Add it to the arch list.
+      list(APPEND ${GPU_ARCHES} "${_CODE_ARCH}${_VIRT}")
+    endforeach()
+  endif()
+  message(STATUS "${GPU_LANG} target arches: ${${GPU_ARCHES}}")
+endmacro()
+
+#
+# Define a target named `GPU_MOD_NAME` for a single extension. The
+# arguments are:
+#
+# DESTINATION <dest>         - Module destination directory.
+# LANGUAGE <lang>            - The GPU language for this module, e.g CUDA, HIP,
+#                              etc.
+# SOURCES <sources>          - List of source files relative to CMakeLists.txt
+#                              directory.
+#
+# Optional arguments:
+#
+# ARCHITECTURES <arches>     - A list of target GPU architectures in cmake
+#                              format.
+#                              Refer `CMAKE_CUDA_ARCHITECTURES` documentation
+#                              and `CMAKE_HIP_ARCHITECTURES` for more info.
+#                              ARCHITECTURES will use cmake's defaults if
+#                              not provided.
+# COMPILE_FLAGS <flags>      - Extra compiler flags passed to NVCC/hip.
+# INCLUDE_DIRECTORIES <dirs> - Extra include directories.
+# LIBRARIES <libraries>      - Extra link libraries.
+# WITH_SOABI                 - Generate library with python SOABI suffix name.
+#
+# Note: optimization level/debug info is set via cmake build type.
+#
+function (define_gpu_extension_target GPU_MOD_NAME)
+  cmake_parse_arguments(PARSE_ARGV 1
+    GPU
+    "WITH_SOABI"
+    "DESTINATION;LANGUAGE"
+    "SOURCES;ARCHITECTURES;COMPILE_FLAGS;INCLUDE_DIRECTORIES;LIBRARIES")
+
+  # Add hipify preprocessing step when building with HIP/ROCm.
+  if (GPU_LANGUAGE STREQUAL "HIP")
+    hipify_sources_target(GPU_SOURCES ${GPU_MOD_NAME} "${GPU_SOURCES}")
+  endif()
+
+  if (GPU_WITH_SOABI)
+    set(GPU_WITH_SOABI WITH_SOABI)
+  else()
+    set(GPU_WITH_SOABI)
+  endif()
+
+  Python_add_library(${GPU_MOD_NAME} MODULE "${GPU_SOURCES}" ${GPU_WITH_SOABI})
+
+  if (GPU_LANGUAGE STREQUAL "HIP")
+    # Make this target dependent on the hipify preprocessor step.
+    add_dependencies(${GPU_MOD_NAME} hipify${GPU_MOD_NAME})
+  endif()
+
+  if (GPU_ARCHITECTURES)
+    set_target_properties(${GPU_MOD_NAME} PROPERTIES
+      ${GPU_LANGUAGE}_ARCHITECTURES "${GPU_ARCHITECTURES}")
+  endif()
+
+  set_property(TARGET ${GPU_MOD_NAME} PROPERTY CXX_STANDARD 17)
+
+  target_compile_options(${GPU_MOD_NAME} PRIVATE
+    $<$<COMPILE_LANGUAGE:${GPU_LANGUAGE}>:${GPU_COMPILE_FLAGS}>)
+
+  target_compile_definitions(${GPU_MOD_NAME} PRIVATE
+    "-DTORCH_EXTENSION_NAME=${GPU_MOD_NAME}")
+
+  target_include_directories(${GPU_MOD_NAME} PRIVATE csrc
+    ${GPU_INCLUDE_DIRECTORIES})
+
+  target_link_libraries(${GPU_MOD_NAME} PRIVATE torch ${torch_python_LIBRARY}
+    ${GPU_LIBRARIES})
+
+  # Don't use `TORCH_LIBRARIES` for CUDA since it pulls in a bunch of
+  # dependencies that are not necessary and may not be installed.
+  if (GPU_LANGUAGE STREQUAL "CUDA")
+    target_link_libraries(${GPU_MOD_NAME} PRIVATE ${CUDA_CUDA_LIB}
+      ${CUDA_LIBRARIES})
+  else()
+    target_link_libraries(${GPU_MOD_NAME} PRIVATE ${TORCH_LIBRARIES})
+  endif()
+
+  install(TARGETS ${GPU_MOD_NAME} LIBRARY DESTINATION ${GPU_DESTINATION})
+endfunction()

collect_env.py

@@ -0,0 +1,721 @@
+# ruff: noqa
+# code borrowed from https://github.com/pytorch/pytorch/blob/main/torch/utils/collect_env.py
+
+# Unlike the rest of the PyTorch this file must be python2 compliant.
+# This script outputs relevant system environment info
+# Run it with `python collect_env.py` or `python -m torch.utils.collect_env`
+import datetime
+import locale
+import os
+import re
+import subprocess
+import sys
+from collections import namedtuple
+
+try:
+    import torch
+    TORCH_AVAILABLE = True
+except (ImportError, NameError, AttributeError, OSError):
+    TORCH_AVAILABLE = False
+
+# System Environment Information
+SystemEnv = namedtuple(
+    'SystemEnv',
+    [
+        'torch_version',
+        'is_debug_build',
+        'cuda_compiled_version',
+        'gcc_version',
+        'clang_version',
+        'cmake_version',
+        'os',
+        'libc_version',
+        'python_version',
+        'python_platform',
+        'is_cuda_available',
+        'cuda_runtime_version',
+        'cuda_module_loading',
+        'nvidia_driver_version',
+        'nvidia_gpu_models',
+        'cudnn_version',
+        'pip_version',  # 'pip' or 'pip3'
+        'pip_packages',
+        'conda_packages',
+        'hip_compiled_version',
+        'hip_runtime_version',
+        'miopen_runtime_version',
+        'caching_allocator_config',
+        'is_xnnpack_available',
+        'cpu_info',
+        'rocm_version',  # vllm specific field
+        'neuron_sdk_version',  # vllm specific field
+        'vllm_version',  # vllm specific field
+        'vllm_build_flags',  # vllm specific field
+        'gpu_topo',  # vllm specific field
+    ])
+
+DEFAULT_CONDA_PATTERNS = {
+    "torch",
+    "numpy",
+    "cudatoolkit",
+    "soumith",
+    "mkl",
+    "magma",
+    "triton",
+    "optree",
+    "nccl",
+}
+
+DEFAULT_PIP_PATTERNS = {
+    "torch",
+    "numpy",
+    "mypy",
+    "flake8",
+    "triton",
+    "optree",
+    "onnx",
+    "nccl",
+}
+
+
+def run(command):
+    """Return (return-code, stdout, stderr)."""
+    shell = True if type(command) is str else False
+    p = subprocess.Popen(command,
+                         stdout=subprocess.PIPE,
+                         stderr=subprocess.PIPE,
+                         shell=shell)
+    raw_output, raw_err = p.communicate()
+    rc = p.returncode
+    if get_platform() == 'win32':
+        enc = 'oem'
+    else:
+        enc = locale.getpreferredencoding()
+    output = raw_output.decode(enc)
+    err = raw_err.decode(enc)
+    return rc, output.strip(), err.strip()
+
+
+def run_and_read_all(run_lambda, command):
+    """Run command using run_lambda; reads and returns entire output if rc is 0."""
+    rc, out, _ = run_lambda(command)
+    if rc != 0:
+        return None
+    return out
+
+
+def run_and_parse_first_match(run_lambda, command, regex):
+    """Run command using run_lambda, returns the first regex match if it exists."""
+    rc, out, _ = run_lambda(command)
+    if rc != 0:
+        return None
+    match = re.search(regex, out)
+    if match is None:
+        return None
+    return match.group(1)
+
+
+def run_and_return_first_line(run_lambda, command):
+    """Run command using run_lambda and returns first line if output is not empty."""
+    rc, out, _ = run_lambda(command)
+    if rc != 0:
+        return None
+    return out.split('\n')[0]
+
+
+def get_conda_packages(run_lambda, patterns=None):
+    if patterns is None:
+        patterns = DEFAULT_CONDA_PATTERNS
+    conda = os.environ.get('CONDA_EXE', 'conda')
+    out = run_and_read_all(run_lambda, "{} list".format(conda))
+    if out is None:
+        return out
+
+    return "\n".join(line for line in out.splitlines()
+                     if not line.startswith("#") and any(name in line
+                                                         for name in patterns))
+
+
+def get_gcc_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, 'gcc --version', r'gcc (.*)')
+
+
+def get_clang_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, 'clang --version',
+                                     r'clang version (.*)')
+
+
+def get_cmake_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, 'cmake --version',
+                                     r'cmake (.*)')
+
+
+def get_nvidia_driver_version(run_lambda):
+    if get_platform() == 'darwin':
+        cmd = 'kextstat | grep -i cuda'
+        return run_and_parse_first_match(run_lambda, cmd,
+                                         r'com[.]nvidia[.]CUDA [(](.*?)[)]')
+    smi = get_nvidia_smi()
+    return run_and_parse_first_match(run_lambda, smi,
+                                     r'Driver Version: (.*?) ')
+
+
+def get_gpu_info(run_lambda):
+    if get_platform() == 'darwin' or (TORCH_AVAILABLE and hasattr(
+            torch.version, 'hip') and torch.version.hip is not None):
+        if TORCH_AVAILABLE and torch.cuda.is_available():
+            if torch.version.hip is not None:
+                prop = torch.cuda.get_device_properties(0)
+                if hasattr(prop, "gcnArchName"):
+                    gcnArch = " ({})".format(prop.gcnArchName)
+                else:
+                    gcnArch = "NoGCNArchNameOnOldPyTorch"
+            else:
+                gcnArch = ""
+            return torch.cuda.get_device_name(None) + gcnArch
+        return None
+    smi = get_nvidia_smi()
+    uuid_regex = re.compile(r' \(UUID: .+?\)')
+    rc, out, _ = run_lambda(smi + ' -L')
+    if rc != 0:
+        return None
+    # Anonymize GPUs by removing their UUID
+    return re.sub(uuid_regex, '', out)
+
+
+def get_running_cuda_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, 'nvcc --version',
+                                     r'release .+ V(.*)')
+
+
+def get_cudnn_version(run_lambda):
+    """Return a list of libcudnn.so; it's hard to tell which one is being used."""
+    if get_platform() == 'win32':
+        system_root = os.environ.get('SYSTEMROOT', 'C:\\Windows')
+        cuda_path = os.environ.get('CUDA_PATH', "%CUDA_PATH%")
+        where_cmd = os.path.join(system_root, 'System32', 'where')
+        cudnn_cmd = '{} /R "{}\\bin" cudnn*.dll'.format(where_cmd, cuda_path)
+    elif get_platform() == 'darwin':
+        # CUDA libraries and drivers can be found in /usr/local/cuda/. See
+        # https://docs.nvidia.com/cuda/cuda-installation-guide-mac-os-x/index.html#install
+        # https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html#installmac
+        # Use CUDNN_LIBRARY when cudnn library is installed elsewhere.
+        cudnn_cmd = 'ls /usr/local/cuda/lib/libcudnn*'
+    else:
+        cudnn_cmd = 'ldconfig -p | grep libcudnn | rev | cut -d" " -f1 | rev'
+    rc, out, _ = run_lambda(cudnn_cmd)
+    # find will return 1 if there are permission errors or if not found
+    if len(out) == 0 or (rc != 1 and rc != 0):
+        l = os.environ.get('CUDNN_LIBRARY')
+        if l is not None and os.path.isfile(l):
+            return os.path.realpath(l)
+        return None
+    files_set = set()
+    for fn in out.split('\n'):
+        fn = os.path.realpath(fn)  # eliminate symbolic links
+        if os.path.isfile(fn):
+            files_set.add(fn)
+    if not files_set:
+        return None
+    # Alphabetize the result because the order is non-deterministic otherwise
+    files = sorted(files_set)
+    if len(files) == 1:
+        return files[0]
+    result = '\n'.join(files)
+    return 'Probably one of the following:\n{}'.format(result)
+
+
+def get_nvidia_smi():
+    # Note: nvidia-smi is currently available only on Windows and Linux
+    smi = 'nvidia-smi'
+    if get_platform() == 'win32':
+        system_root = os.environ.get('SYSTEMROOT', 'C:\\Windows')
+        program_files_root = os.environ.get('PROGRAMFILES',
+                                            'C:\\Program Files')
+        legacy_path = os.path.join(program_files_root, 'NVIDIA Corporation',
+                                   'NVSMI', smi)
+        new_path = os.path.join(system_root, 'System32', smi)
+        smis = [new_path, legacy_path]
+        for candidate_smi in smis:
+            if os.path.exists(candidate_smi):
+                smi = '"{}"'.format(candidate_smi)
+                break
+    return smi
+
+
+def get_rocm_version(run_lambda):
+    """Returns the ROCm version if available, otherwise 'N/A'."""
+    return run_and_parse_first_match(run_lambda, 'hipcc --version',
+                                     r'HIP version: (\S+)')
+
+
+def get_neuron_sdk_version(run_lambda):
+    # Adapted from your install script
+    try:
+        result = run_lambda(["neuron-ls"])
+        return result if result[0] == 0 else 'N/A'
+    except Exception:
+        return 'N/A'
+
+
+def get_vllm_version():
+    try:
+        import vllm
+        return vllm.__version__
+    except ImportError:
+        return 'N/A'
+
+
+def summarize_vllm_build_flags():
+    # This could be a static method if the flags are constant, or dynamic if you need to check environment variables, etc.
+    return 'CUDA Archs: {}; ROCm: {}; Neuron: {}'.format(
+        os.environ.get('TORCH_CUDA_ARCH_LIST', 'Not Set'),
+        'Enabled' if os.environ.get('ROCM_HOME') else 'Disabled',
+        'Enabled' if os.environ.get('NEURON_CORES') else 'Disabled',
+    )
+
+
+def get_gpu_topo(run_lambda):
+    if get_platform() == 'linux':
+        return run_and_read_all(run_lambda, 'nvidia-smi topo -m')
+    return None
+
+
+# example outputs of CPU infos
+#  * linux
+#    Architecture:            x86_64
+#      CPU op-mode(s):        32-bit, 64-bit
+#      Address sizes:         46 bits physical, 48 bits virtual
+#      Byte Order:            Little Endian
+#    CPU(s):                  128
+#      On-line CPU(s) list:   0-127
+#    Vendor ID:               GenuineIntel
+#      Model name:            Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz
+#        CPU family:          6
+#        Model:               106
+#        Thread(s) per core:  2
+#        Core(s) per socket:  32
+#        Socket(s):           2
+#        Stepping:            6
+#        BogoMIPS:            5799.78
+#        Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr
+#                             sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon rep_good nopl
+#                             xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq monitor ssse3 fma cx16
+#                             pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand
+#                             hypervisor lahf_lm abm 3dnowprefetch invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced
+#                             fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap
+#                             avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1
+#                             xsaves wbnoinvd ida arat avx512vbmi pku ospke avx512_vbmi2 gfni vaes vpclmulqdq
+#                             avx512_vnni avx512_bitalg tme avx512_vpopcntdq rdpid md_clear flush_l1d arch_capabilities
+#    Virtualization features:
+#      Hypervisor vendor:     KVM
+#      Virtualization type:   full
+#    Caches (sum of all):
+#      L1d:                   3 MiB (64 instances)
+#      L1i:                   2 MiB (64 instances)
+#      L2:                    80 MiB (64 instances)
+#      L3:                    108 MiB (2 instances)
+#    NUMA:
+#      NUMA node(s):          2
+#      NUMA node0 CPU(s):     0-31,64-95
+#      NUMA node1 CPU(s):     32-63,96-127
+#    Vulnerabilities:
+#      Itlb multihit:         Not affected
+#      L1tf:                  Not affected
+#      Mds:                   Not affected
+#      Meltdown:              Not affected
+#      Mmio stale data:       Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
+#      Retbleed:              Not affected
+#      Spec store bypass:     Mitigation; Speculative Store Bypass disabled via prctl and seccomp
+#      Spectre v1:            Mitigation; usercopy/swapgs barriers and __user pointer sanitization
+#      Spectre v2:            Mitigation; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
+#      Srbds:                 Not affected
+#      Tsx async abort:       Not affected
+#  * win32
+#    Architecture=9
+#    CurrentClockSpeed=2900
+#    DeviceID=CPU0
+#    Family=179
+#    L2CacheSize=40960
+#    L2CacheSpeed=
+#    Manufacturer=GenuineIntel
+#    MaxClockSpeed=2900
+#    Name=Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz
+#    ProcessorType=3
+#    Revision=27142
+#
+#    Architecture=9
+#    CurrentClockSpeed=2900
+#    DeviceID=CPU1
+#    Family=179
+#    L2CacheSize=40960
+#    L2CacheSpeed=
+#    Manufacturer=GenuineIntel
+#    MaxClockSpeed=2900
+#    Name=Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz
+#    ProcessorType=3
+#    Revision=27142
+
+
+def get_cpu_info(run_lambda):
+    rc, out, err = 0, '', ''
+    if get_platform() == 'linux':
+        rc, out, err = run_lambda('lscpu')
+    elif get_platform() == 'win32':
+        rc, out, err = run_lambda(
+            'wmic cpu get Name,Manufacturer,Family,Architecture,ProcessorType,DeviceID, \
+        CurrentClockSpeed,MaxClockSpeed,L2CacheSize,L2CacheSpeed,Revision /VALUE'
+        )
+    elif get_platform() == 'darwin':
+        rc, out, err = run_lambda("sysctl -n machdep.cpu.brand_string")
+    cpu_info = 'None'
+    if rc == 0:
+        cpu_info = out
+    else:
+        cpu_info = err
+    return cpu_info
+
+
+def get_platform():
+    if sys.platform.startswith('linux'):
+        return 'linux'
+    elif sys.platform.startswith('win32'):
+        return 'win32'
+    elif sys.platform.startswith('cygwin'):
+        return 'cygwin'
+    elif sys.platform.startswith('darwin'):
+        return 'darwin'
+    else:
+        return sys.platform
+
+
+def get_mac_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, 'sw_vers -productVersion',
+                                     r'(.*)')
+
+
+def get_windows_version(run_lambda):
+    system_root = os.environ.get('SYSTEMROOT', 'C:\\Windows')
+    wmic_cmd = os.path.join(system_root, 'System32', 'Wbem', 'wmic')
+    findstr_cmd = os.path.join(system_root, 'System32', 'findstr')
+    return run_and_read_all(
+        run_lambda,
+        '{} os get Caption | {} /v Caption'.format(wmic_cmd, findstr_cmd))
+
+
+def get_lsb_version(run_lambda):
+    return run_and_parse_first_match(run_lambda, 'lsb_release -a',
+                                     r'Description:\t(.*)')
+
+
+def check_release_file(run_lambda):
+    return run_and_parse_first_match(run_lambda, 'cat /etc/*-release',
+                                     r'PRETTY_NAME="(.*)"')
+
+
+def get_os(run_lambda):
+    from platform import machine
+    platform = get_platform()
+
+    if platform == 'win32' or platform == 'cygwin':
+        return get_windows_version(run_lambda)
+
+    if platform == 'darwin':
+        version = get_mac_version(run_lambda)
+        if version is None:
+            return None
+        return 'macOS {} ({})'.format(version, machine())
+
+    if platform == 'linux':
+        # Ubuntu/Debian based
+        desc = get_lsb_version(run_lambda)
+        if desc is not None:
+            return '{} ({})'.format(desc, machine())
+
+        # Try reading /etc/*-release
+        desc = check_release_file(run_lambda)
+        if desc is not None:
+            return '{} ({})'.format(desc, machine())
+
+        return '{} ({})'.format(platform, machine())
+
+    # Unknown platform
+    return platform
+
+
+def get_python_platform():
+    import platform
+    return platform.platform()
+
+
+def get_libc_version():
+    import platform
+    if get_platform() != 'linux':
+        return 'N/A'
+    return '-'.join(platform.libc_ver())
+
+
+def get_pip_packages(run_lambda, patterns=None):
+    """Return `pip list` output. Note: will also find conda-installed pytorch and numpy packages."""
+    if patterns is None:
+        patterns = DEFAULT_PIP_PATTERNS
+
+    # People generally have `pip` as `pip` or `pip3`
+    # But here it is invoked as `python -mpip`
+    def run_with_pip(pip):
+        out = run_and_read_all(run_lambda, pip + ["list", "--format=freeze"])
+        return "\n".join(line for line in out.splitlines()
+                         if any(name in line for name in patterns))
+
+    pip_version = 'pip3' if sys.version[0] == '3' else 'pip'
+    out = run_with_pip([sys.executable, '-mpip'])
+
+    return pip_version, out
+
+
+def get_cachingallocator_config():
+    ca_config = os.environ.get('PYTORCH_CUDA_ALLOC_CONF', '')
+    return ca_config
+
+
+def get_cuda_module_loading_config():
+    if TORCH_AVAILABLE and torch.cuda.is_available():
+        torch.cuda.init()
+        config = os.environ.get('CUDA_MODULE_LOADING', '')
+        return config
+    else:
+        return "N/A"
+
+
+def is_xnnpack_available():
+    if TORCH_AVAILABLE:
+        import torch.backends.xnnpack
+        return str(
+            torch.backends.xnnpack.enabled)  # type: ignore[attr-defined]
+    else:
+        return "N/A"
+
+
+def get_env_info():
+    run_lambda = run
+    pip_version, pip_list_output = get_pip_packages(run_lambda)
+
+    if TORCH_AVAILABLE:
+        version_str = torch.__version__
+        debug_mode_str = str(torch.version.debug)
+        cuda_available_str = str(torch.cuda.is_available())
+        cuda_version_str = torch.version.cuda
+        if not hasattr(torch.version,
+                       'hip') or torch.version.hip is None:  # cuda version
+            hip_compiled_version = hip_runtime_version = miopen_runtime_version = 'N/A'
+        else:  # HIP version
+
+            def get_version_or_na(cfg, prefix):
+                _lst = [s.rsplit(None, 1)[-1] for s in cfg if prefix in s]
+                return _lst[0] if _lst else 'N/A'
+
+            cfg = torch._C._show_config().split('\n')
+            hip_runtime_version = get_version_or_na(cfg, 'HIP Runtime')
+            miopen_runtime_version = get_version_or_na(cfg, 'MIOpen')
+            cuda_version_str = 'N/A'
+            hip_compiled_version = torch.version.hip
+    else:
+        version_str = debug_mode_str = cuda_available_str = cuda_version_str = 'N/A'
+        hip_compiled_version = hip_runtime_version = miopen_runtime_version = 'N/A'
+
+    sys_version = sys.version.replace("\n", " ")
+
+    conda_packages = get_conda_packages(run_lambda)
+
+    rocm_version = get_rocm_version(run_lambda)
+    neuron_sdk_version = get_neuron_sdk_version(run_lambda)
+    vllm_version = get_vllm_version()
+    vllm_build_flags = summarize_vllm_build_flags()
+    gpu_topo = get_gpu_topo(run_lambda)
+
+    return SystemEnv(
+        torch_version=version_str,
+        is_debug_build=debug_mode_str,
+        python_version='{} ({}-bit runtime)'.format(
+            sys_version,
+            sys.maxsize.bit_length() + 1),
+        python_platform=get_python_platform(),
+        is_cuda_available=cuda_available_str,
+        cuda_compiled_version=cuda_version_str,
+        cuda_runtime_version=get_running_cuda_version(run_lambda),
+        cuda_module_loading=get_cuda_module_loading_config(),
+        nvidia_gpu_models=get_gpu_info(run_lambda),
+        nvidia_driver_version=get_nvidia_driver_version(run_lambda),
+        cudnn_version=get_cudnn_version(run_lambda),
+        hip_compiled_version=hip_compiled_version,
+        hip_runtime_version=hip_runtime_version,
+        miopen_runtime_version=miopen_runtime_version,
+        pip_version=pip_version,
+        pip_packages=pip_list_output,
+        conda_packages=conda_packages,
+        os=get_os(run_lambda),
+        libc_version=get_libc_version(),
+        gcc_version=get_gcc_version(run_lambda),
+        clang_version=get_clang_version(run_lambda),
+        cmake_version=get_cmake_version(run_lambda),
+        caching_allocator_config=get_cachingallocator_config(),
+        is_xnnpack_available=is_xnnpack_available(),
+        cpu_info=get_cpu_info(run_lambda),
+        rocm_version=rocm_version,
+        neuron_sdk_version=neuron_sdk_version,
+        vllm_version=vllm_version,
+        vllm_build_flags=vllm_build_flags,
+        gpu_topo=gpu_topo,
+    )
+
+
+env_info_fmt = """
+PyTorch version: {torch_version}
+Is debug build: {is_debug_build}
+CUDA used to build PyTorch: {cuda_compiled_version}
+ROCM used to build PyTorch: {hip_compiled_version}
+
+OS: {os}
+GCC version: {gcc_version}
+Clang version: {clang_version}
+CMake version: {cmake_version}
+Libc version: {libc_version}
+
+Python version: {python_version}
+Python platform: {python_platform}
+Is CUDA available: {is_cuda_available}
+CUDA runtime version: {cuda_runtime_version}
+CUDA_MODULE_LOADING set to: {cuda_module_loading}
+GPU models and configuration: {nvidia_gpu_models}
+Nvidia driver version: {nvidia_driver_version}
+cuDNN version: {cudnn_version}
+HIP runtime version: {hip_runtime_version}
+MIOpen runtime version: {miopen_runtime_version}
+Is XNNPACK available: {is_xnnpack_available}
+
+CPU:
+{cpu_info}
+
+Versions of relevant libraries:
+{pip_packages}
+{conda_packages}
+""".strip()
+
+env_info_fmt += """
+ROCM Version: {rocm_version}
+Neuron SDK Version: {neuron_sdk_version}
+vLLM Version: {vllm_version}
+vLLM Build Flags:
+{vllm_build_flags}
+GPU Topology:
+{gpu_topo}
+""".strip()
+
+
+def pretty_str(envinfo):
+
+    def replace_nones(dct, replacement='Could not collect'):
+        for key in dct.keys():
+            if dct[key] is not None:
+                continue
+            dct[key] = replacement
+        return dct
+
+    def replace_bools(dct, true='Yes', false='No'):
+        for key in dct.keys():
+            if dct[key] is True:
+                dct[key] = true
+            elif dct[key] is False:
+                dct[key] = false
+        return dct
+
+    def prepend(text, tag='[prepend]'):
+        lines = text.split('\n')
+        updated_lines = [tag + line for line in lines]
+        return '\n'.join(updated_lines)
+
+    def replace_if_empty(text, replacement='No relevant packages'):
+        if text is not None and len(text) == 0:
+            return replacement
+        return text
+
+    def maybe_start_on_next_line(string):
+        # If `string` is multiline, prepend a \n to it.
+        if string is not None and len(string.split('\n')) > 1:
+            return '\n{}\n'.format(string)
+        return string
+
+    mutable_dict = envinfo._asdict()
+
+    # If nvidia_gpu_models is multiline, start on the next line
+    mutable_dict['nvidia_gpu_models'] = \
+        maybe_start_on_next_line(envinfo.nvidia_gpu_models)
+
+    # If the machine doesn't have CUDA, report some fields as 'No CUDA'
+    dynamic_cuda_fields = [
+        'cuda_runtime_version',
+        'nvidia_gpu_models',
+        'nvidia_driver_version',
+    ]
+    all_cuda_fields = dynamic_cuda_fields + ['cudnn_version']
+    all_dynamic_cuda_fields_missing = all(mutable_dict[field] is None
+                                          for field in dynamic_cuda_fields)
+    if TORCH_AVAILABLE and not torch.cuda.is_available(
+    ) and all_dynamic_cuda_fields_missing:
+        for field in all_cuda_fields:
+            mutable_dict[field] = 'No CUDA'
+        if envinfo.cuda_compiled_version is None:
+            mutable_dict['cuda_compiled_version'] = 'None'
+
+    # Replace True with Yes, False with No
+    mutable_dict = replace_bools(mutable_dict)
+
+    # Replace all None objects with 'Could not collect'
+    mutable_dict = replace_nones(mutable_dict)
+
+    # If either of these are '', replace with 'No relevant packages'
+    mutable_dict['pip_packages'] = replace_if_empty(
+        mutable_dict['pip_packages'])
+    mutable_dict['conda_packages'] = replace_if_empty(
+        mutable_dict['conda_packages'])
+
+    # Tag conda and pip packages with a prefix
+    # If they were previously None, they'll show up as ie '[conda] Could not collect'
+    if mutable_dict['pip_packages']:
+        mutable_dict['pip_packages'] = prepend(
+            mutable_dict['pip_packages'], '[{}] '.format(envinfo.pip_version))
+    if mutable_dict['conda_packages']:
+        mutable_dict['conda_packages'] = prepend(
+            mutable_dict['conda_packages'], '[conda] ')
+    mutable_dict['cpu_info'] = envinfo.cpu_info
+    return env_info_fmt.format(**mutable_dict)
+
+
+def get_pretty_env_info():
+    return pretty_str(get_env_info())
+
+
+def main():
+    print("Collecting environment information...")
+    output = get_pretty_env_info()
+    print(output)
+
+    if TORCH_AVAILABLE and hasattr(torch, 'utils') and hasattr(
+            torch.utils, '_crash_handler'):
+        minidump_dir = torch.utils._crash_handler.DEFAULT_MINIDUMP_DIR
+        if sys.platform == "linux" and os.path.exists(minidump_dir):
+            dumps = [
+                os.path.join(minidump_dir, dump)
+                for dump in os.listdir(minidump_dir)
+            ]
+            latest = max(dumps, key=os.path.getctime)
+            ctime = os.path.getctime(latest)
+            creation_time = datetime.datetime.fromtimestamp(ctime).strftime(
+                '%Y-%m-%d %H:%M:%S')
+            msg = "\n*** Detected a minidump at {} created on {}, ".format(latest, creation_time) + \
+                  "if this is related to your bug please include it when you file a report ***"
+            print(msg, file=sys.stderr)
+
+
+if __name__ == '__main__':
+    main()

csrc/activation_kernels.cu

@@ -10,11 +10,11 @@
 namespace vllm {
 
 // Activation and gating kernel template.
-template<typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&)>
+template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&)>
 __global__ void act_and_mul_kernel(
-  scalar_t* __restrict__ out,               // [..., d]
-  const scalar_t* __restrict__ input,       // [..., 2, d]
-  const int d) {
+    scalar_t* __restrict__ out,          // [..., d]
+    const scalar_t* __restrict__ input,  // [..., 2, d]
+    const int d) {
   const int64_t token_idx = blockIdx.x;
   for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
     const scalar_t x = VLLM_LDG(&input[token_idx * 2 * d + idx]);
@@ -23,64 +23,78 @@ __global__ void act_and_mul_kernel(
   }
 }
 
-template<typename T>
+template <typename T>
 __device__ __forceinline__ T silu_kernel(const T& x) {
   // x * sigmoid(x)
-  return (T) (((float) x) / (1.0f + expf((float) -x)));
+  return (T)(((float)x) / (1.0f + expf((float)-x)));
 }
 
-template<typename T>
+template <typename T>
 __device__ __forceinline__ T gelu_kernel(const T& x) {
   // Equivalent to PyTorch GELU with 'none' approximation.
   // Refer to:
-  // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L38
-  const float f = (float) x;
+  // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L36-L38
+  const float f = (float)x;
   constexpr float ALPHA = M_SQRT1_2;
-  return (T) (f * 0.5f * (1.0f + ::erf(f * ALPHA)));
+  return (T)(f * 0.5f * (1.0f + ::erf(f * ALPHA)));
 }
 
-} // namespace vllm
+template <typename T>
+__device__ __forceinline__ T gelu_tanh_kernel(const T& x) {
+  // Equivalent to PyTorch GELU with 'tanh' approximation.
+  // Refer to:
+  // https://github.com/pytorch/pytorch/blob/8ac9b20d4b090c213799e81acf48a55ea8d437d6/aten/src/ATen/native/cuda/ActivationGeluKernel.cu#L25-L30
+  const float f = (float)x;
+  constexpr float BETA = M_SQRT2 * M_2_SQRTPI * 0.5f;
+  constexpr float KAPPA = 0.044715;
+  float x_cube = f * f * f;
+  float inner = BETA * (f + KAPPA * x_cube);
+  return (T)(0.5f * f * (1.0f + ::tanhf(inner)));
+}
+
+}  // namespace vllm
 
 // Launch activation and gating kernel.
-#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL)                                             \
-  int d = input.size(-1) / 2;                                                             \
-  int64_t num_tokens = input.numel() / input.size(-1);                                    \
-  dim3 grid(num_tokens);                                                                  \
-  dim3 block(std::min(d, 1024));                                                          \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));                       \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                           \
-  VLLM_DISPATCH_FLOATING_TYPES(                                                           \
-    input.scalar_type(),                                                                  \
-    "act_and_mul_kernel",                                                                 \
-    [&] {                                                                                 \
-      vllm::act_and_mul_kernel<scalar_t, KERNEL<scalar_t>><<<grid, block, 0, stream>>>(   \
-        out.data_ptr<scalar_t>(),                                                         \
-        input.data_ptr<scalar_t>(),                                                       \
-        d);                                                                               \
-    });
+#define LAUNCH_ACTIVATION_GATE_KERNEL(KERNEL)                            \
+  int d = input.size(-1) / 2;                                            \
+  int64_t num_tokens = input.numel() / input.size(-1);                   \
+  dim3 grid(num_tokens);                                                 \
+  dim3 block(std::min(d, 1024));                                         \
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));      \
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();          \
+  VLLM_DISPATCH_FLOATING_TYPES(                                          \
+      input.scalar_type(), "act_and_mul_kernel", [&] {                   \
+        vllm::act_and_mul_kernel<scalar_t, KERNEL<scalar_t>>             \
+            <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),       \
+                                         input.data_ptr<scalar_t>(), d); \
+      });
 
-void silu_and_mul(
-  torch::Tensor& out,      // [..., d]
-  torch::Tensor& input)    // [..., 2 * d]
+void silu_and_mul(torch::Tensor& out,    // [..., d]
+                  torch::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel);
 }
 
-void gelu_and_mul(
-  torch::Tensor& out,      // [..., d]
-  torch::Tensor& input)    // [..., 2 * d]
+void gelu_and_mul(torch::Tensor& out,    // [..., d]
+                  torch::Tensor& input)  // [..., 2 * d]
 {
   LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_kernel);
 }
 
+void gelu_tanh_and_mul(torch::Tensor& out,    // [..., d]
+                       torch::Tensor& input)  // [..., 2 * d]
+{
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::gelu_tanh_kernel);
+}
+
 namespace vllm {
 
 // Element-wise activation kernel template.
-template<typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&)>
+template <typename scalar_t, scalar_t (*ACT_FN)(const scalar_t&)>
 __global__ void activation_kernel(
-  scalar_t* __restrict__ out,               // [..., d]
-  const scalar_t* __restrict__ input,       // [..., d]
-  const int d) {
+    scalar_t* __restrict__ out,          // [..., d]
+    const scalar_t* __restrict__ input,  // [..., d]
+    const int d) {
   const int64_t token_idx = blockIdx.x;
   for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
     const scalar_t x = VLLM_LDG(&input[token_idx * d + idx]);
@@ -88,54 +102,49 @@ __global__ void activation_kernel(
   }
 }
 
-} // namespace vllm
+}  // namespace vllm
 
 // Launch element-wise activation kernel.
-#define LAUNCH_ACTIVATION_KERNEL(KERNEL)                                                  \
-  int d = input.size(-1);                                                                 \
-  int64_t num_tokens = input.numel() / d;                                                 \
-  dim3 grid(num_tokens);                                                                  \
-  dim3 block(std::min(d, 1024));                                                          \
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));                       \
-  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                           \
-  VLLM_DISPATCH_FLOATING_TYPES(                                                           \
-    input.scalar_type(),                                                                  \
-    "activation_kernel",                                                                  \
-    [&] {                                                                                 \
-      vllm::activation_kernel<scalar_t, KERNEL<scalar_t>><<<grid, block, 0, stream>>>(    \
-        out.data_ptr<scalar_t>(),                                                         \
-        input.data_ptr<scalar_t>(),                                                       \
-        d);                                                                               \
-    });
+#define LAUNCH_ACTIVATION_KERNEL(KERNEL)                                       \
+  int d = input.size(-1);                                                      \
+  int64_t num_tokens = input.numel() / d;                                      \
+  dim3 grid(num_tokens);                                                       \
+  dim3 block(std::min(d, 1024));                                               \
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));            \
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();                \
+  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "activation_kernel", [&] { \
+    vllm::activation_kernel<scalar_t, KERNEL<scalar_t>>                        \
+        <<<grid, block, 0, stream>>>(out.data_ptr<scalar_t>(),                 \
+                                     input.data_ptr<scalar_t>(), d);           \
+  });
 
 namespace vllm {
 
-template<typename T>
+template <typename T>
 __device__ __forceinline__ T gelu_new_kernel(const T& x) {
-  const float x3 = (float) (x * x * x);
-  const T t = (T) tanhf((T) (0.79788456f * (float) (x + (T) (0.044715f * x3))));
-  return ((T) 0.5) * x * (((T) 1.0) + t);
+  const float x3 = (float)(x * x * x);
+  const T t = (T)tanhf((T)(0.79788456f * (float)(x + (T)(0.044715f * x3))));
+  return ((T)0.5) * x * (((T)1.0) + t);
 }
 
-template<typename T>
+template <typename T>
 __device__ __forceinline__ T gelu_fast_kernel(const T& x) {
-  const float f = (float) x;
-  const T t = (T) tanhf(((T) (f * 0.79788456f)) * (((T) 1.0) + (T) (0.044715f * f) * x));
-  return ((T) 0.5) * x * (((T) 1.0) + t);
+  const float f = (float)x;
+  const T t =
+      (T)tanhf(((T)(f * 0.79788456f)) * (((T)1.0) + (T)(0.044715f * f) * x));
+  return ((T)0.5) * x * (((T)1.0) + t);
 }
 
-} // namespace vllm
+}  // namespace vllm
 
-void gelu_new(
-  torch::Tensor& out,     // [..., d]
-  torch::Tensor& input)   // [..., d]
+void gelu_new(torch::Tensor& out,    // [..., d]
+              torch::Tensor& input)  // [..., d]
 {
   LAUNCH_ACTIVATION_KERNEL(vllm::gelu_new_kernel);
 }
 
-void gelu_fast(
-  torch::Tensor& out,     // [..., d]
-  torch::Tensor& input)   // [..., d]
+void gelu_fast(torch::Tensor& out,    // [..., d]
+               torch::Tensor& input)  // [..., d]
 {
   LAUNCH_ACTIVATION_KERNEL(vllm::gelu_fast_kernel);
 }

csrc/attention/attention_dtypes.h

@@ -4,4 +4,4 @@
 #include "dtype_float16.cuh"
 #include "dtype_float32.cuh"
 #include "dtype_bfloat16.cuh"
-#include "dtype_fp8_e5m2.cuh"
+#include "dtype_fp8.cuh"

csrc/attention/attention_generic.cuh

@@ -1,5 +1,6 @@
 /*
- * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
+ * Adapted from
+ * https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
  * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
@@ -22,31 +23,31 @@
 namespace vllm {
 
 // A vector type to store Q, K, V elements.
-template<typename T, int VEC_SIZE>
+template <typename T, int VEC_SIZE>
 struct Vec {};
 
 // A vector type to store FP32 accumulators.
-template<typename T>
+template <typename T>
 struct FloatVec {};
 
 // Template vector operations.
-template<typename Acc, typename A, typename B>
+template <typename Acc, typename A, typename B>
 inline __device__ Acc mul(A a, B b);
 
-template<typename T>
+template <typename T>
 inline __device__ float sum(T v);
 
-template<typename T>
+template <typename T>
 inline __device__ float dot(T a, T b) {
   return sum(mul<T, T, T>(a, b));
 }
 
-template<typename A, typename T>
+template <typename A, typename T>
 inline __device__ float dot(T a, T b) {
   return sum(mul<A, T, T>(a, b));
 }
 
-template<typename T>
+template <typename T>
 inline __device__ void zero(T& dst) {
   constexpr int WORDS = sizeof(T) / 4;
   union {
@@ -61,4 +62,4 @@ inline __device__ void zero(T& dst) {
   dst = tmp.raw;
 }
 
-} // namespace vllm
+}  // namespace vllm

csrc/attention/attention_utils.cuh

@@ -1,5 +1,6 @@
 /*
- * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
+ * Adapted from
+ * https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
  * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
@@ -26,7 +27,7 @@
 namespace vllm {
 
 // Q*K^T operation.
-template<int THREAD_GROUP_SIZE, typename Vec, int N>
+template <int THREAD_GROUP_SIZE, typename Vec, int N>
 inline __device__ float qk_dot_(const Vec (&q)[N], const Vec (&k)[N]) {
   using A_vec = typename FloatVec<Vec>::Type;
   // Compute the parallel products for Q*K^T (treat vector lanes separately).
@@ -45,12 +46,12 @@ inline __device__ float qk_dot_(const Vec (&q)[N], const Vec (&k)[N]) {
   return qk;
 }
 
-template<typename T, int THREAD_GROUP_SIZE>
+template <typename T, int THREAD_GROUP_SIZE>
 struct Qk_dot {
-  template<typename Vec, int N>
+  template <typename Vec, int N>
   static inline __device__ float dot(const Vec (&q)[N], const Vec (&k)[N]) {
     return qk_dot_<THREAD_GROUP_SIZE>(q, k);
   }
 };
 
-} // namespace vllm
+}  // namespace vllm

csrc/attention/dtype_bfloat16.cuh

@@ -1,6 +1,8 @@
 /*
- * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
- * and https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
+ * Adapted from
+ * https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
+ * and
+ * https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
  * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
@@ -28,8 +30,8 @@
   #include <hip/hip_bf16.h>
   #include <hip/hip_fp16.h>
 
-  typedef __hip_bfloat162 __nv_bfloat162;
-  typedef __hip_bfloat16 __nv_bfloat16;
+typedef __hip_bfloat162 __nv_bfloat162;
+typedef __hip_bfloat16 __nv_bfloat16;
 #endif
 
 #include <stdint.h>
@@ -50,37 +52,37 @@ struct bf16_8_t {
 };
 
 // BF16 vector types for Q, K, V.
-template<>
+template <>
 struct Vec<__nv_bfloat16, 1> {
   using Type = __nv_bfloat16;
 };
-template<>
+template <>
 struct Vec<__nv_bfloat16, 2> {
   using Type = __nv_bfloat162;
 };
-template<>
+template <>
 struct Vec<__nv_bfloat16, 4> {
   using Type = bf16_4_t;
 };
-template<>
+template <>
 struct Vec<__nv_bfloat16, 8> {
   using Type = bf16_8_t;
 };
 
 // FP32 accumulator vector types corresponding to Vec.
-template<>
+template <>
 struct FloatVec<__nv_bfloat16> {
   using Type = float;
 };
-template<>
+template <>
 struct FloatVec<__nv_bfloat162> {
   using Type = float2;
 };
-template<>
+template <>
 struct FloatVec<bf16_4_t> {
   using Type = Float4_;
 };
-template<>
+template <>
 struct FloatVec<bf16_8_t> {
   using Type = Float8_;
 };
@@ -108,9 +110,9 @@ inline __device__ __nv_bfloat16 add(__nv_bfloat16 a, __nv_bfloat16 b) {
   assert(false);
 #else
   #ifndef USE_ROCM
-    return a + b;
+  return a + b;
   #else
-    return __hadd(a, b);
+  return __hadd(a, b);
   #endif
 #endif
 }
@@ -161,7 +163,7 @@ inline __device__ Float8_ add(bf16_8_t a, Float8_ fb) {
 }
 
 // Vector multiplication.
-template<>
+template <>
 inline __device__ __nv_bfloat16 mul(__nv_bfloat16 a, __nv_bfloat16 b) {
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
   assert(false);
@@ -170,7 +172,7 @@ inline __device__ __nv_bfloat16 mul(__nv_bfloat16 a, __nv_bfloat16 b) {
 #endif
 }
 
-template<>
+template <>
 inline __device__ __nv_bfloat162 mul(__nv_bfloat162 a, __nv_bfloat162 b) {
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
   assert(false);
@@ -179,12 +181,12 @@ inline __device__ __nv_bfloat162 mul(__nv_bfloat162 a, __nv_bfloat162 b) {
 #endif
 }
 
-template<>
+template <>
 inline __device__ __nv_bfloat162 mul(__nv_bfloat16 a, __nv_bfloat162 b) {
   return mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(bf162bf162(a), b);
 }
 
-template<>
+template <>
 inline __device__ bf16_4_t mul(bf16_4_t a, bf16_4_t b) {
   bf16_4_t c;
   c.x = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(a.x, b.x);
@@ -192,7 +194,7 @@ inline __device__ bf16_4_t mul(bf16_4_t a, bf16_4_t b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ bf16_4_t mul(__nv_bfloat16 a, bf16_4_t b) {
   __nv_bfloat162 s = bf162bf162(a);
   bf16_4_t c;
@@ -201,7 +203,7 @@ inline __device__ bf16_4_t mul(__nv_bfloat16 a, bf16_4_t b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ bf16_8_t mul(bf16_8_t a, bf16_8_t b) {
   bf16_8_t c;
   c.x = mul<__nv_bfloat162, __nv_bfloat162, __nv_bfloat162>(a.x, b.x);
@@ -211,7 +213,7 @@ inline __device__ bf16_8_t mul(bf16_8_t a, bf16_8_t b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ bf16_8_t mul(__nv_bfloat16 a, bf16_8_t b) {
   __nv_bfloat162 s = bf162bf162(a);
   bf16_8_t c;
@@ -222,26 +224,26 @@ inline __device__ bf16_8_t mul(__nv_bfloat16 a, bf16_8_t b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ float mul(__nv_bfloat16 a, __nv_bfloat16 b) {
   float fa = __bfloat162float(a);
   float fb = __bfloat162float(b);
   return fa * fb;
 }
 
-template<>
+template <>
 inline __device__ float2 mul(__nv_bfloat162 a, __nv_bfloat162 b) {
   float2 fa = bf1622float2(a);
   float2 fb = bf1622float2(b);
   return mul<float2, float2, float2>(fa, fb);
 }
 
-template<>
+template <>
 inline __device__ float2 mul(__nv_bfloat16 a, __nv_bfloat162 b) {
   return mul<float2, __nv_bfloat162, __nv_bfloat162>(bf162bf162(a), b);
 }
 
-template<>
+template <>
 inline __device__ Float4_ mul(bf16_4_t a, bf16_4_t b) {
   Float4_ fc;
   fc.x = mul<float2, __nv_bfloat162, __nv_bfloat162>(a.x, b.x);
@@ -249,7 +251,7 @@ inline __device__ Float4_ mul(bf16_4_t a, bf16_4_t b) {
   return fc;
 }
 
-template<>
+template <>
 inline __device__ Float4_ mul(__nv_bfloat16 a, bf16_4_t b) {
   __nv_bfloat162 s = bf162bf162(a);
   Float4_ fc;
@@ -258,7 +260,7 @@ inline __device__ Float4_ mul(__nv_bfloat16 a, bf16_4_t b) {
   return fc;
 }
 
-template<>
+template <>
 inline __device__ Float8_ mul(bf16_8_t a, bf16_8_t b) {
   Float8_ fc;
   fc.x = mul<float2, __nv_bfloat162, __nv_bfloat162>(a.x, b.x);
@@ -268,7 +270,7 @@ inline __device__ Float8_ mul(bf16_8_t a, bf16_8_t b) {
   return fc;
 }
 
-template<>
+template <>
 inline __device__ Float8_ mul(__nv_bfloat16 a, bf16_8_t b) {
   __nv_bfloat162 s = bf162bf162(a);
   Float8_ fc;
@@ -280,7 +282,8 @@ inline __device__ Float8_ mul(__nv_bfloat16 a, bf16_8_t b) {
 }
 
 // Vector fused multiply-add.
-inline __device__ __nv_bfloat162 fma(__nv_bfloat162 a, __nv_bfloat162 b, __nv_bfloat162 c) {
+inline __device__ __nv_bfloat162 fma(__nv_bfloat162 a, __nv_bfloat162 b,
+                                     __nv_bfloat162 c) {
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
   assert(false);
 #else
@@ -288,7 +291,8 @@ inline __device__ __nv_bfloat162 fma(__nv_bfloat162 a, __nv_bfloat162 b, __nv_bf
 #endif
 }
 
-inline __device__ __nv_bfloat162 fma(__nv_bfloat16 a, __nv_bfloat162 b, __nv_bfloat162 c) {
+inline __device__ __nv_bfloat162 fma(__nv_bfloat16 a, __nv_bfloat162 b,
+                                     __nv_bfloat162 c) {
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 800
   assert(false);
 #else
@@ -379,23 +383,23 @@ inline __device__ Float8_ fma(__nv_bfloat16 a, bf16_8_t b, Float8_ fc) {
 }
 
 // Vector sum.
-template<>
+template <>
 inline __device__ float sum(__nv_bfloat16 v) {
   return __bfloat162float(v);
 }
 
-template<>
+template <>
 inline __device__ float sum(__nv_bfloat162 v) {
   float2 vf = bf1622float2(v);
   return vf.x + vf.y;
 }
 
-template<>
+template <>
 inline __device__ float sum(bf16_4_t v) {
   return sum(v.x) + sum(v.y);
 }
 
-template<>
+template <>
 inline __device__ float sum(bf16_8_t v) {
   return sum(v.x) + sum(v.y) + sum(v.z) + sum(v.w);
 }
@@ -448,4 +452,4 @@ inline __device__ void zero(__nv_bfloat16& dst) {
 #endif
 }
 
-} // namespace vllm
+}  // namespace vllm

csrc/attention/dtype_float16.cuh

@@ -1,6 +1,8 @@
 /*
- * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
- * and https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
+ * Adapted from
+ * https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
+ * and
+ * https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
  * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
@@ -30,37 +32,37 @@
 namespace vllm {
 
 // FP16 vector types for Q, K, V.
-template<>
+template <>
 struct Vec<uint16_t, 1> {
   using Type = uint16_t;
 };
-template<>
+template <>
 struct Vec<uint16_t, 2> {
   using Type = uint32_t;
 };
-template<>
+template <>
 struct Vec<uint16_t, 4> {
   using Type = uint2;
 };
-template<>
+template <>
 struct Vec<uint16_t, 8> {
   using Type = uint4;
 };
 
 // FP32 accumulator vector types corresponding to Vec.
-template<>
+template <>
 struct FloatVec<uint16_t> {
   using Type = float;
 };
-template<>
+template <>
 struct FloatVec<uint32_t> {
   using Type = float2;
 };
-template<>
+template <>
 struct FloatVec<uint2> {
   using Type = Float4_;
 };
-template<>
+template <>
 struct FloatVec<uint4> {
   using Type = Float8_;
 };
@@ -73,8 +75,8 @@ inline __device__ uint32_t h0_h0(uint16_t a) {
   return b;
 #else
   union {
-   uint32_t u32;
-   uint16_t u16[2];
+    uint32_t u32;
+    uint16_t u16[2];
   } tmp;
   tmp.u16[0] = a;
   tmp.u16[1] = a;
@@ -130,10 +132,12 @@ inline __device__ uint32_t float2_to_half2(float2 f) {
   } tmp;
 #ifndef USE_ROCM
   #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
-    asm volatile("cvt.rn.f16x2.f32 %0, %1, %2;\n" : "=r"(tmp.u32) : "f"(f.y), "f"(f.x));
+  asm volatile("cvt.rn.f16x2.f32 %0, %1, %2;\n"
+               : "=r"(tmp.u32)
+               : "f"(f.y), "f"(f.x));
   #else
-    asm volatile("cvt.rn.f16.f32 %0, %1;\n" : "=h"(tmp.u16[0]) : "f"(f.x));
-    asm volatile("cvt.rn.f16.f32 %0, %1;\n" : "=h"(tmp.u16[1]) : "f"(f.y));
+  asm volatile("cvt.rn.f16.f32 %0, %1;\n" : "=h"(tmp.u16[0]) : "f"(f.x));
+  asm volatile("cvt.rn.f16.f32 %0, %1;\n" : "=h"(tmp.u16[1]) : "f"(f.y));
   #endif
 #else
   tmp.u16[0] = float_to_half(f.x);
@@ -201,7 +205,7 @@ inline __device__ Float8_ add(uint4 a, Float8_ fb) {
 }
 
 // Vector multiplication.
-template<>
+template <>
 inline __device__ uint16_t mul(uint16_t a, uint16_t b) {
   uint16_t c;
 #ifndef USE_ROCM
@@ -212,7 +216,7 @@ inline __device__ uint16_t mul(uint16_t a, uint16_t b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ uint32_t mul(uint32_t a, uint32_t b) {
   uint32_t c;
 #ifndef USE_ROCM
@@ -223,12 +227,12 @@ inline __device__ uint32_t mul(uint32_t a, uint32_t b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ uint32_t mul(uint16_t a, uint32_t b) {
   return mul<uint32_t, uint32_t, uint32_t>(h0_h0(a), b);
 }
 
-template<>
+template <>
 inline __device__ uint2 mul(uint2 a, uint2 b) {
   uint2 c;
   c.x = mul<uint32_t, uint32_t, uint32_t>(a.x, b.x);
@@ -236,7 +240,7 @@ inline __device__ uint2 mul(uint2 a, uint2 b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ uint2 mul(uint16_t a, uint2 b) {
   uint32_t s = h0_h0(a);
   uint2 c;
@@ -245,7 +249,7 @@ inline __device__ uint2 mul(uint16_t a, uint2 b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ uint4 mul(uint4 a, uint4 b) {
   uint4 c;
   c.x = mul<uint32_t, uint32_t, uint32_t>(a.x, b.x);
@@ -255,7 +259,7 @@ inline __device__ uint4 mul(uint4 a, uint4 b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ uint4 mul(uint16_t a, uint4 b) {
   uint32_t s = h0_h0(a);
   uint4 c;
@@ -266,26 +270,26 @@ inline __device__ uint4 mul(uint16_t a, uint4 b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ float mul(uint16_t a, uint16_t b) {
   float fa = half_to_float(a);
   float fb = half_to_float(b);
   return fa * fb;
 }
 
-template<>
+template <>
 inline __device__ float2 mul(uint32_t a, uint32_t b) {
   float2 fa = half2_to_float2(a);
   float2 fb = half2_to_float2(b);
   return mul<float2, float2, float2>(fa, fb);
 }
 
-template<>
+template <>
 inline __device__ float2 mul(uint16_t a, uint32_t b) {
   return mul<float2, uint32_t, uint32_t>(h0_h0(a), b);
 }
 
-template<>
+template <>
 inline __device__ Float4_ mul(uint2 a, uint2 b) {
   Float4_ fc;
   fc.x = mul<float2, uint32_t, uint32_t>(a.x, b.x);
@@ -293,7 +297,7 @@ inline __device__ Float4_ mul(uint2 a, uint2 b) {
   return fc;
 }
 
-template<>
+template <>
 inline __device__ Float4_ mul(uint16_t a, uint2 b) {
   uint32_t s = h0_h0(a);
   Float4_ fc;
@@ -302,7 +306,7 @@ inline __device__ Float4_ mul(uint16_t a, uint2 b) {
   return fc;
 }
 
-template<>
+template <>
 inline __device__ Float8_ mul(uint4 a, uint4 b) {
   Float8_ fc;
   fc.x = mul<float2, uint32_t, uint32_t>(a.x, b.x);
@@ -312,7 +316,7 @@ inline __device__ Float8_ mul(uint4 a, uint4 b) {
   return fc;
 }
 
-template<>
+template <>
 inline __device__ Float8_ mul(uint16_t a, uint4 b) {
   uint32_t s = h0_h0(a);
   Float8_ fc;
@@ -327,9 +331,13 @@ inline __device__ Float8_ mul(uint16_t a, uint4 b) {
 inline __device__ uint32_t fma(uint32_t a, uint32_t b, uint32_t c) {
   uint32_t d;
 #ifndef USE_ROCM
-  asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(d) : "r"(a), "r"(b), "r"(c));
+  asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n"
+               : "=r"(d)
+               : "r"(a), "r"(b), "r"(c));
 #else
-  asm volatile("v_pk_fma_f16 %0, %1, %2, %3;\n" : "=v"(d) : "v"(a), "v"(b), "v"(c));
+  asm volatile("v_pk_fma_f16 %0, %1, %2, %3;\n"
+               : "=v"(d)
+               : "v"(a), "v"(b), "v"(c));
 #endif
   return d;
 }
@@ -423,24 +431,24 @@ inline __device__ Float8_ fma(uint16_t a, uint4 b, Float8_ fc) {
 }
 
 // Vector sum.
-template<>
+template <>
 inline __device__ float sum(uint16_t v) {
   return half_to_float(v);
 }
 
-template<>
+template <>
 inline __device__ float sum(uint32_t v) {
   float2 tmp = half2_to_float2(v);
   return tmp.x + tmp.y;
 }
 
-template<>
+template <>
 inline __device__ float sum(uint2 v) {
   uint32_t c = add(v.x, v.y);
   return sum(c);
 }
 
-template<>
+template <>
 inline __device__ float sum(uint4 v) {
   uint32_t c = add(v.x, v.y);
   c = add(c, v.z);
@@ -470,13 +478,9 @@ inline __device__ void from_float(uint4& dst, Float8_ src) {
 }
 
 // From float16 to float32.
-inline __device__ float to_float(uint16_t u) {
-  return half_to_float(u);
-}
+inline __device__ float to_float(uint16_t u) { return half_to_float(u); }
 
-inline __device__ float2 to_float(uint32_t u) {
-  return half2_to_float2(u);
-}
+inline __device__ float2 to_float(uint32_t u) { return half2_to_float2(u); }
 
 inline __device__ Float4_ to_float(uint2 u) {
   Float4_ tmp;
@@ -495,8 +499,6 @@ inline __device__ Float8_ to_float(uint4 u) {
 }
 
 // Zero-out a variable.
-inline __device__ void zero(uint16_t& dst) {
-  dst = uint16_t(0);
-}
+inline __device__ void zero(uint16_t& dst) { dst = uint16_t(0); }
 
-} // namespace vllm
+}  // namespace vllm

csrc/attention/dtype_float32.cuh

@@ -1,6 +1,8 @@
 /*
- * Adapted from https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
- * and https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
+ * Adapted from
+ * https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention/decoder_masked_multihead_attention_template.hpp
+ * and
+ * https://github.com/NVIDIA/FasterTransformer/blob/release/v5.3_tag/src/fastertransformer/kernels/decoder_masked_multihead_attention_utils.h
  * Copyright (c) 2023, The vLLM team.
  * Copyright (c) 2020-2023, NVIDIA CORPORATION.  All rights reserved.
  *
@@ -38,37 +40,35 @@ struct Float8_ {
 };
 
 // FP32 vector types for Q, K, V.
-template<>
+template <>
 struct Vec<float, 1> {
   using Type = float;
 };
-template<>
+template <>
 struct Vec<float, 2> {
   using Type = float2;
 };
-template<>
+template <>
 struct Vec<float, 4> {
   using Type = float4;
 };
 
 // FP32 accumulator vector types corresponding to Vec.
-template<>
+template <>
 struct FloatVec<float> {
   using Type = float;
 };
-template<>
+template <>
 struct FloatVec<float2> {
   using Type = float2;
 };
-template<>
+template <>
 struct FloatVec<float4> {
   using Type = float4;
 };
 
 // Vector addition.
-inline __device__ float add(float a, float b) {
-  return a + b;
-}
+inline __device__ float add(float a, float b) { return a + b; }
 
 inline __device__ float2 add(float2 a, float2 b) {
   float2 c;
@@ -87,12 +87,12 @@ inline __device__ float4 add(float4 a, float4 b) {
 }
 
 // Vector multiplication.
-template<>
+template <>
 inline __device__ float mul<float, float>(float a, float b) {
   return a * b;
 }
 
-template<>
+template <>
 inline __device__ float2 mul(float2 a, float2 b) {
   float2 c;
   c.x = a.x * b.x;
@@ -100,7 +100,7 @@ inline __device__ float2 mul(float2 a, float2 b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ float2 mul(float a, float2 b) {
   float2 c;
   c.x = a * b.x;
@@ -108,7 +108,7 @@ inline __device__ float2 mul(float a, float2 b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ float4 mul(float4 a, float4 b) {
   float4 c;
   c.x = a.x * b.x;
@@ -118,7 +118,7 @@ inline __device__ float4 mul(float4 a, float4 b) {
   return c;
 }
 
-template<>
+template <>
 inline __device__ float4 mul(float a, float4 b) {
   float4 c;
   c.x = a * b.x;
@@ -129,9 +129,7 @@ inline __device__ float4 mul(float a, float4 b) {
 }
 
 // Vector fused multiply-add.
-inline __device__ float fma(float a, float b, float c) {
-  return a * b + c;
-}
+inline __device__ float fma(float a, float b, float c) { return a * b + c; }
 
 inline __device__ float2 fma(float2 a, float2 b, float2 c) {
   float2 d;
@@ -182,35 +180,33 @@ inline __device__ Float8_ fma(float a, Float8_ b, Float8_ c) {
 }
 
 // Vector sum.
-template<>
+template <>
 inline __device__ float sum(float v) {
   return v;
 }
 
-template<>
+template <>
 inline __device__ float sum(float2 v) {
   return v.x + v.y;
 }
 
-template<>
+template <>
 inline __device__ float sum(float4 v) {
   return v.x + v.y + v.z + v.w;
 }
 
-template<>
+template <>
 inline __device__ float sum(Float4_ v) {
   return v.x.x + v.x.y + v.y.x + v.y.y;
 }
 
-template<>
+template <>
 inline __device__ float sum(Float8_ v) {
   return v.x.x + v.x.y + v.y.x + v.y.y + v.z.x + v.z.y + v.w.x + v.w.y;
 }
 
 // Vector dot product.
-inline __device__ float dot(float a, float b) {
-  return a * b;
-}
+inline __device__ float dot(float a, float b) { return a * b; }
 
 inline __device__ float dot(float2 a, float2 b) {
   float2 c = mul<float2, float2, float2>(a, b);
@@ -232,42 +228,24 @@ inline __device__ float dot(Float8_ a, Float8_ b) {
 }
 
 // From float to float.
-inline __device__ void from_float(float& dst, float src) {
-  dst = src;
-}
+inline __device__ void from_float(float& dst, float src) { dst = src; }
 
-inline __device__ void from_float(float2& dst, float2 src) {
-  dst = src;
-}
+inline __device__ void from_float(float2& dst, float2 src) { dst = src; }
 
-inline __device__ void from_float(float4& dst, float4 src) {
-  dst = src;
-}
+inline __device__ void from_float(float4& dst, float4 src) { dst = src; }
 
 // From float to float.
-inline __device__ float to_float(float u) {
-  return u;
-}
+inline __device__ float to_float(float u) { return u; }
 
-inline __device__ float2 to_float(float2 u) {
-  return u;
-}
+inline __device__ float2 to_float(float2 u) { return u; }
 
-inline __device__ float4 to_float(float4 u) {
-  return u;
-}
+inline __device__ float4 to_float(float4 u) { return u; }
 
-inline __device__ Float4_ to_float(Float4_ u) {
-  return u;
-}
+inline __device__ Float4_ to_float(Float4_ u) { return u; }
 
-inline __device__ Float8_ to_float(Float8_ u) {
-  return u;
-}
+inline __device__ Float8_ to_float(Float8_ u) { return u; }
 
 // Zero-out a variable.
-inline __device__ void zero(float& dst) {
-  dst = 0.f;
-}
+inline __device__ void zero(float& dst) { dst = 0.f; }
 
-} // namespace vllm
+}  // namespace vllm

csrc/attention/dtype_fp8.cuh

@@ -0,0 +1,41 @@
+#pragma once
+
+#include "attention_generic.cuh"
+
+#include <stdint.h>
+#ifdef ENABLE_FP8
+  #ifndef USE_ROCM
+    #include <cuda_fp8.h>
+  #endif  // USE_ROCM
+#endif    // ENABLE_FP8
+
+namespace vllm {
+
+enum class Fp8KVCacheDataType {
+  kAuto = 0,
+  kFp8E4M3 = 1,
+  kFp8E5M2 = 2,
+};
+
+// fp8 vector types for quantization of kv cache
+template <>
+struct Vec<uint8_t, 1> {
+  using Type = uint8_t;
+};
+
+template <>
+struct Vec<uint8_t, 2> {
+  using Type = uint16_t;
+};
+
+template <>
+struct Vec<uint8_t, 4> {
+  using Type = uint32_t;
+};
+
+template <>
+struct Vec<uint8_t, 8> {
+  using Type = uint2;
+};
+
+}  // namespace vllm

csrc/attention/dtype_fp8_e5m2.cuh

@@ -1,35 +0,0 @@
-#pragma once
-
-#include "attention_generic.cuh"
-
-#include <stdint.h>
-#ifdef ENABLE_FP8_E5M2
-#include <cuda_fp8.h>
-#endif
-
-namespace vllm {
-#ifdef ENABLE_FP8_E5M2
-// fp8 vector types for quantization of kv cache
-
-template<>
-struct Vec<uint8_t, 1> {
-    using Type = uint8_t;
-};
-
-template<>
-struct Vec<uint8_t, 2> {
-    using Type = uint16_t;
-};
-
-template<>
-struct Vec<uint8_t, 4> {
-    using Type = uint32_t;
-};
-
-template<>
-struct Vec<uint8_t, 8> {
-    using Type = uint2;
-};
-#endif // ENABLE_FP8_E5M2
-
-} // namespace vllm

csrc/cache.h

@@ -5,25 +5,24 @@
 #include <map>
 #include <vector>
 
-void swap_blocks(
-  torch::Tensor& src,
-  torch::Tensor& dst,
-  const std::map<int64_t, int64_t>& block_mapping);
+void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
+                 const torch::Tensor& block_mapping);
 
-void copy_blocks(
-  std::vector<torch::Tensor>& key_caches,
-  std::vector<torch::Tensor>& value_caches,
-  const std::map<int64_t, std::vector<int64_t>>& block_mapping);
+void copy_blocks(std::vector<torch::Tensor>& key_caches,
+                 std::vector<torch::Tensor>& value_caches,
+                 const torch::Tensor& block_mapping);
 
-void reshape_and_cache(
-  torch::Tensor& key,
-  torch::Tensor& value,
-  torch::Tensor& key_cache,
-  torch::Tensor& value_cache,
-  torch::Tensor& slot_mapping,
-  const std::string& kv_cache_dtype);
+void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
+                       torch::Tensor& key_cache, torch::Tensor& value_cache,
+                       torch::Tensor& slot_mapping,
+                       const std::string& kv_cache_dtype, const float kv_scale);
+
+void reshape_and_cache_flash(torch::Tensor& key, torch::Tensor& value,
+                             torch::Tensor& key_cache,
+                             torch::Tensor& value_cache,
+                             torch::Tensor& slot_mapping,
+                             const std::string& kv_cache_dtype);
 
 // Just for unittest
-void convert_fp8_e5m2(
-  torch::Tensor& src_cache,
-  torch::Tensor& dst_cache);
+void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
+                 const float scale, const std::string& kv_cache_dtype);

csrc/cache_kernels.cu

@@ -4,8 +4,11 @@
 
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
-#ifdef ENABLE_FP8_E5M2
-#include "quantization/fp8_e5m2_kvcache/quant_utils.cuh"
+
+#ifdef USE_ROCM
+  #include "quantization/fp8/amd/quant_utils.cuh"
+#else
+  #include "quantization/fp8/nvidia/quant_utils.cuh"
 #endif
 
 #include <algorithm>
@@ -15,20 +18,17 @@
 
 #ifdef USE_ROCM
   #include <hip/hip_bf16.h>
-  typedef __hip_bfloat16 __nv_bfloat16;
+typedef __hip_bfloat16 __nv_bfloat16;
 #endif
 
-void swap_blocks(
-  torch::Tensor& src,
-  torch::Tensor& dst,
-  const std::map<int64_t, int64_t>& block_mapping) {
+void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
+                 const torch::Tensor& block_mapping) {
   torch::Device src_device = src.device();
   torch::Device dst_device = dst.device();
   cudaMemcpyKind memcpy_type;
   if (src_device.is_cuda() && dst_device.is_cuda()) {
-    TORCH_CHECK(
-      src_device.index() == dst_device.index(),
-      "src and dst must be on the same GPU");
+    TORCH_CHECK(src_device.index() == dst_device.index(),
+                "src and dst must be on the same GPU");
     memcpy_type = cudaMemcpyDeviceToDevice;
   } else if (src_device.is_cuda() && dst_device.is_cpu()) {
     memcpy_type = cudaMemcpyDeviceToHost;
@@ -38,41 +38,44 @@ void swap_blocks(
     TORCH_CHECK(false, "Invalid device combination");
   }
 
-  char *src_ptr = static_cast<char*>(src.data_ptr());
-  char *dst_ptr = static_cast<char*>(dst.data_ptr());
+  // NOTE(youkaichao): keep in mind that `block_mapping` should be
+  // a cpu tensor, otherwise every `item` call will require a gpu-cpu
+  // synchronization.
+  TORCH_CHECK(block_mapping.device().is_cpu(), "block_mapping must be on CPU");
+
+  char* src_ptr = static_cast<char*>(src.data_ptr());
+  char* dst_ptr = static_cast<char*>(dst.data_ptr());
 
   const int64_t block_size_in_bytes = src.element_size() * src[0].numel();
-  const at::cuda::OptionalCUDAGuard device_guard(src_device.is_cuda() ? src_device : dst_device);
+  const at::cuda::OptionalCUDAGuard device_guard(
+      src_device.is_cuda() ? src_device : dst_device);
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   // NOTE(woosuk): This can be slow if the number of blocks is large.
-  for (const auto& pair : block_mapping) {
-    int64_t src_block_number = pair.first;
-    int64_t dst_block_number = pair.second;
+  const int64_t num_blocks = block_mapping.size(0);
+  for (size_t i = 0; i < num_blocks; i++) {
+    int64_t src_block_number = block_mapping[i][0].item<int64_t>();
+    int64_t dst_block_number = block_mapping[i][1].item<int64_t>();
     int64_t src_offset = src_block_number * block_size_in_bytes;
     int64_t dst_offset = dst_block_number * block_size_in_bytes;
-    cudaMemcpyAsync(
-      dst_ptr + dst_offset,
-      src_ptr + src_offset,
-      block_size_in_bytes,
-      memcpy_type,
-      stream);
+    cudaMemcpyAsync(dst_ptr + dst_offset, src_ptr + src_offset,
+                    block_size_in_bytes, memcpy_type, stream);
   }
 }
 
 namespace vllm {
 
 // Grid: (num_layers, num_pairs)
-template<typename scalar_t>
-__global__ void copy_blocks_kernel(
-  int64_t* key_cache_ptrs,
-  int64_t* value_cache_ptrs,
-  const int64_t* __restrict__ block_mapping,
-  const int numel_per_block) {
+template <typename scalar_t>
+__global__ void copy_blocks_kernel(int64_t* key_cache_ptrs,
+                                   int64_t* value_cache_ptrs,
+                                   const int64_t* __restrict__ block_mapping,
+                                   const int numel_per_block) {
   const int layer_idx = blockIdx.x;
   const int pair_idx = blockIdx.y;
 
   scalar_t* key_cache = reinterpret_cast<scalar_t*>(key_cache_ptrs[layer_idx]);
-  scalar_t* value_cache = reinterpret_cast<scalar_t*>(value_cache_ptrs[layer_idx]);
+  scalar_t* value_cache =
+      reinterpret_cast<scalar_t*>(value_cache_ptrs[layer_idx]);
   int64_t src_block_number = block_mapping[2 * pair_idx];
   int64_t dst_block_number = block_mapping[2 * pair_idx + 1];
 
@@ -90,12 +93,11 @@ __global__ void copy_blocks_kernel(
   }
 }
 
-} // namespace vllm
+}  // namespace vllm
 
-void copy_blocks(
-  std::vector<torch::Tensor>& key_caches,
-  std::vector<torch::Tensor>& value_caches,
-  const std::map<int64_t, std::vector<int64_t>>& block_mapping) {
+void copy_blocks(std::vector<torch::Tensor>& key_caches,
+                 std::vector<torch::Tensor>& value_caches,
+                 const torch::Tensor& block_mapping) {
   int num_layers = key_caches.size();
   TORCH_CHECK(num_layers == value_caches.size());
   if (num_layers == 0) {
@@ -109,29 +111,23 @@ void copy_blocks(
   int64_t key_cache_ptrs[num_layers];
   int64_t value_cache_ptrs[num_layers];
   for (int layer_idx = 0; layer_idx < num_layers; ++layer_idx) {
-    key_cache_ptrs[layer_idx] = reinterpret_cast<int64_t>(key_caches[layer_idx].data_ptr());
-    value_cache_ptrs[layer_idx] = reinterpret_cast<int64_t>(value_caches[layer_idx].data_ptr());
+    key_cache_ptrs[layer_idx] =
+        reinterpret_cast<int64_t>(key_caches[layer_idx].data_ptr());
+    value_cache_ptrs[layer_idx] =
+        reinterpret_cast<int64_t>(value_caches[layer_idx].data_ptr());
   }
-  // Create block mapping array.
-  std::vector<int64_t> block_mapping_vec;
-  for (const auto& pair : block_mapping) {
-    int64_t src_block_number = pair.first;
-    for (int64_t dst_block_number : pair.second) {
-      block_mapping_vec.push_back(src_block_number);
-      block_mapping_vec.push_back(dst_block_number);
-    }
-  }
-  int64_t* block_mapping_array = block_mapping_vec.data();
-  int num_pairs = block_mapping_vec.size() / 2;
+
+  // block_mapping is a 2D tensor with shape (num_pairs, 2).
+  int num_pairs = block_mapping.size(0);
 
   // Move the data structures to the GPU.
   // NOTE: This synchronizes the CPU and GPU.
-  torch::Tensor key_cache_ptrs_tensor = torch::from_blob(
-    key_cache_ptrs, {num_layers}, torch::kInt64).to(cache_device);
-  torch::Tensor value_cache_ptrs_tensor = torch::from_blob(
-    value_cache_ptrs, {num_layers}, torch::kInt64).to(cache_device);
-  torch::Tensor block_mapping_tensor = torch::from_blob(
-    block_mapping_array, {2 * num_pairs}, torch::kInt64).to(cache_device);
+  torch::Tensor key_cache_ptrs_tensor =
+      torch::from_blob(key_cache_ptrs, {num_layers}, torch::kInt64)
+          .to(cache_device);
+  torch::Tensor value_cache_ptrs_tensor =
+      torch::from_blob(value_cache_ptrs, {num_layers}, torch::kInt64)
+          .to(cache_device);
 
   // Launch the kernel.
   const int numel_per_block = key_caches[0][0].numel();
@@ -140,30 +136,28 @@ void copy_blocks(
   const at::cuda::OptionalCUDAGuard device_guard(cache_device);
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   VLLM_DISPATCH_FLOATING_AND_BYTE_TYPES(
-    key_caches[0].scalar_type(), "copy_blocks_kernel", ([&] {
-      vllm::copy_blocks_kernel<scalar_t><<<grid, block, 0, stream>>>(
-        key_cache_ptrs_tensor.data_ptr<int64_t>(),
-        value_cache_ptrs_tensor.data_ptr<int64_t>(),
-        block_mapping_tensor.data_ptr<int64_t>(),
-        numel_per_block);
-    }));
+      key_caches[0].scalar_type(), "copy_blocks_kernel", ([&] {
+        vllm::copy_blocks_kernel<scalar_t><<<grid, block, 0, stream>>>(
+            key_cache_ptrs_tensor.data_ptr<int64_t>(),
+            value_cache_ptrs_tensor.data_ptr<int64_t>(),
+            block_mapping.data_ptr<int64_t>(), numel_per_block);
+      }));
 }
 
 namespace vllm {
 
-template<typename scalar_t, typename cache_t, bool is_fp8_e5m2_kv_cache>
+template <typename scalar_t, typename cache_t, Fp8KVCacheDataType kv_dt>
 __global__ void reshape_and_cache_kernel(
-  const scalar_t* __restrict__ key,           // [num_tokens, num_heads, head_size]
-  const scalar_t* __restrict__ value,         // [num_tokens, num_heads, head_size]
-  cache_t* __restrict__ key_cache,            // [num_blocks, num_heads, head_size/x, block_size, x]
-  cache_t* __restrict__ value_cache,          // [num_blocks, num_heads, head_size, block_size]
-  const int64_t* __restrict__ slot_mapping,   // [num_tokens]
-  const int key_stride,
-  const int value_stride,
-  const int num_heads,
-  const int head_size,
-  const int block_size,
-  const int x) {
+    const scalar_t* __restrict__ key,    // [num_tokens, num_heads, head_size]
+    const scalar_t* __restrict__ value,  // [num_tokens, num_heads, head_size]
+    cache_t* __restrict__ key_cache,     // [num_blocks, num_heads, head_size/x,
+                                         // block_size, x]
+    cache_t* __restrict__ value_cache,   // [num_blocks, num_heads, head_size,
+                                         // block_size]
+    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
+    const int key_stride, const int value_stride, const int num_heads,
+    const int head_size, const int block_size, const int x,
+    const float kv_scale) {
   const int64_t token_idx = blockIdx.x;
   const int64_t slot_idx = slot_mapping[token_idx];
   if (slot_idx < 0) {
@@ -184,55 +178,84 @@ __global__ void reshape_and_cache_kernel(
     const int x_idx = head_offset / x;
     const int x_offset = head_offset % x;
 
-    const int64_t tgt_key_idx = block_idx * num_heads * (head_size / x) * block_size * x
-                                + head_idx * (head_size / x) * block_size * x
-                                + x_idx * block_size * x
-                                + block_offset * x
-                                + x_offset;
-    const int64_t tgt_value_idx = block_idx * num_heads * head_size * block_size
-                                  + head_idx * head_size * block_size
-                                  + head_offset * block_size
-                                  + block_offset;
+    const int64_t tgt_key_idx =
+        block_idx * num_heads * (head_size / x) * block_size * x +
+        head_idx * (head_size / x) * block_size * x + x_idx * block_size * x +
+        block_offset * x + x_offset;
+    const int64_t tgt_value_idx =
+        block_idx * num_heads * head_size * block_size +
+        head_idx * head_size * block_size + head_offset * block_size +
+        block_offset;
     scalar_t tgt_key = key[src_key_idx];
     scalar_t tgt_value = value[src_value_idx];
-    if constexpr (is_fp8_e5m2_kv_cache) {
-#ifdef ENABLE_FP8_E5M2
-      key_cache[tgt_key_idx] = fp8_e5m2_unscaled::vec_conversion<uint8_t, scalar_t>(tgt_key);
-      value_cache[tgt_value_idx] = fp8_e5m2_unscaled::vec_conversion<uint8_t, scalar_t>(tgt_value);
-#else
-      assert(false);
-#endif
-    } else {
+    if constexpr (kv_dt == Fp8KVCacheDataType::kAuto) {
       key_cache[tgt_key_idx] = tgt_key;
       value_cache[tgt_value_idx] = tgt_value;
+    } else {
+      key_cache[tgt_key_idx] =
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_key, kv_scale);
+      value_cache[tgt_value_idx] =
+          fp8::scaled_convert<cache_t, scalar_t, kv_dt>(tgt_value, kv_scale);
     }
   }
 }
 
-} // namespace vllm
+template <typename scalar_t>
+__global__ void reshape_and_cache_flash_kernel(
+    const scalar_t* __restrict__ key,    // [num_tokens, num_heads, head_size]
+    const scalar_t* __restrict__ value,  // [num_tokens, num_heads, head_size]
+    scalar_t* __restrict__ k_cache,      // [num_blocks, block_size, num_heads,
+                                         // head_size]
+    scalar_t* __restrict__ v_cache,      // [num_blocks, block_size, num_heads,
+                                         // head_size]
+    const int64_t* __restrict__ slot_mapping,  // [num_tokens]
+    const int block_stride, const int key_stride, const int value_stride,
+    const int num_heads, const int head_size, const int block_size) {
+  const int64_t token_idx = blockIdx.x;
+  const int64_t slot_idx = slot_mapping[token_idx];
+  // NOTE: slot_idx can be -1 if the token is padded
+  if (slot_idx < 0) {
+    return;
+  }
+  const int64_t block_idx = slot_idx / block_size;
+  const int64_t block_offset = slot_idx % block_size;
+  const int n = num_heads * head_size;
+  for (int i = threadIdx.x; i < n; i += blockDim.x) {
+    const int64_t src_key_idx = token_idx * key_stride + i;
+    const int64_t src_value_idx = token_idx * value_stride + i;
+    const int head_idx = i / head_size;
+    const int head_offset = i % head_size;
+    const int64_t tgt_value_idx = block_idx * block_stride +
+                                  block_offset * num_heads * head_size +
+                                  head_idx * head_size + head_offset;
+    k_cache[tgt_value_idx] = key[src_key_idx];
+    v_cache[tgt_value_idx] = value[src_value_idx];
+  }
+}
+}  // namespace vllm
 
-#define CALL_RESHAPE_AND_CACHE(KV_T, CACHE_T, IS_FP8_E5M2_KV_CACHE)                                \
-  vllm::reshape_and_cache_kernel<KV_T, CACHE_T, IS_FP8_E5M2_KV_CACHE><<<grid, block, 0, stream>>>( \
-    reinterpret_cast<KV_T*>(key.data_ptr()),                                                       \
-    reinterpret_cast<KV_T*>(value.data_ptr()),                                                     \
-    reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),                                              \
-    reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),                                            \
-    slot_mapping.data_ptr<int64_t>(),                                                              \
-    key_stride,                                                                                    \
-    value_stride,                                                                                  \
-    num_heads,                                                                                     \
-    head_size,                                                                                     \
-    block_size,                                                                                    \
-    x);
+// KV_T is the stored data type of kv-cache.
+// CACHE_T is the data type of key and value tensors.
+// 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>             \
+      <<<grid, block, 0, stream>>>(                                   \
+          reinterpret_cast<KV_T*>(key.data_ptr()),                    \
+          reinterpret_cast<KV_T*>(value.data_ptr()),                  \
+          reinterpret_cast<CACHE_T*>(key_cache.data_ptr()),           \
+          reinterpret_cast<CACHE_T*>(value_cache.data_ptr()),         \
+          slot_mapping.data_ptr<int64_t>(), key_stride, value_stride, \
+          num_heads, head_size, block_size, x, kv_scale);
 
 void reshape_and_cache(
-  torch::Tensor& key,           // [num_tokens, num_heads, head_size]
-  torch::Tensor& value,         // [num_tokens, num_heads, head_size]
-  torch::Tensor& key_cache,     // [num_blocks, num_heads, head_size/x, block_size, x]
-  torch::Tensor& value_cache,   // [num_blocks, num_heads, head_size, block_size]
-  torch::Tensor& slot_mapping,  // [num_tokens]
-  const std::string& kv_cache_dtype)
-{
+    torch::Tensor& key,    // [num_tokens, num_heads, head_size]
+    torch::Tensor& value,  // [num_tokens, num_heads, head_size]
+    torch::Tensor&
+        key_cache,  // [num_blocks, num_heads, head_size/x, block_size, x]
+    torch::Tensor&
+        value_cache,  // [num_blocks, num_heads, head_size, block_size]
+    torch::Tensor& slot_mapping,  // [num_tokens]
+    const std::string& kv_cache_dtype, const float kv_scale) {
   int num_tokens = key.size(0);
   int num_heads = key.size(1);
   int head_size = key.size(2);
@@ -246,57 +269,80 @@ void reshape_and_cache(
   dim3 block(std::min(num_heads * head_size, 512));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(key));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  if (kv_cache_dtype == "auto") {
-    if (key.dtype() == at::ScalarType::Float) {
-      CALL_RESHAPE_AND_CACHE(float, float, false);
-    } else if (key.dtype() == at::ScalarType::Half) {
-      CALL_RESHAPE_AND_CACHE(uint16_t, uint16_t, false);
-    } else if (key.dtype() == at::ScalarType::BFloat16) {
-      CALL_RESHAPE_AND_CACHE(__nv_bfloat16, __nv_bfloat16, false);
-    }
-  } else if (kv_cache_dtype == "fp8_e5m2") {
-    if (key.dtype() == at::ScalarType::Float) {
-      CALL_RESHAPE_AND_CACHE(float, uint8_t, true);
-    } else if (key.dtype() == at::ScalarType::Half) {
-      CALL_RESHAPE_AND_CACHE(uint16_t, uint8_t, true);
-    } else if (key.dtype() == at::ScalarType::BFloat16) {
-      CALL_RESHAPE_AND_CACHE(__nv_bfloat16, uint8_t, true);
-    }
-  } else {
+
+  DISPATCH_BY_KV_CACHE_DTYPE(key.dtype(), kv_cache_dtype,
+                             CALL_RESHAPE_AND_CACHE)
+}
+
+void reshape_and_cache_flash(
+    torch::Tensor& key,      // [num_tokens, num_heads, head_size]
+    torch::Tensor& value,    // [num_tokens, num_heads, head_size]
+    torch::Tensor& k_cache,  // [num_blocks, block_size, num_heads, head_size]
+    torch::Tensor& v_cache,  // [num_blocks, block_size, num_heads, head_size]
+    torch::Tensor& slot_mapping,  // [num_tokens]
+    const std::string& kv_cache_dtype) {
+  // FIXME: only support auto datatype, does not support fp8
+  if (kv_cache_dtype != "auto") {
     TORCH_CHECK(false, "Unsupported data type of kv cache: ", kv_cache_dtype);
   }
+  int num_tokens = key.size(0);
+  int num_heads = key.size(1);
+  int head_size = key.size(2);
+  int block_size = k_cache.size(1);
+
+  int key_stride = key.stride(0);
+  int value_stride = value.stride(0);
+  int block_stride = k_cache.stride(0);
+  TORCH_CHECK(k_cache.stride(0) == v_cache.stride(0));
+
+  dim3 grid(num_tokens);
+  dim3 block(std::min(num_heads * head_size, 512));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(key));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  VLLM_DISPATCH_FLOATING_TYPES(
+      key.scalar_type(), "reshape_and_cache_flash", [&] {
+        vllm::reshape_and_cache_flash_kernel<scalar_t>
+            <<<grid, block, 0, stream>>>(
+                key.data_ptr<scalar_t>(), value.data_ptr<scalar_t>(),
+                k_cache.data_ptr<scalar_t>(), v_cache.data_ptr<scalar_t>(),
+                slot_mapping.data_ptr<int64_t>(), block_stride, key_stride,
+                value_stride, num_heads, head_size, block_size);
+      });
 }
 
 namespace vllm {
 
-template<typename Tout, typename Tin>
-__global__ void convert_fp8_e5m2_kernel(
-  const Tin* __restrict__ src_cache,
-  Tout* __restrict__ dst_cache,
-  const int64_t block_stride) {
+template <typename Tout, typename Tin, Fp8KVCacheDataType kv_dt>
+__global__ void convert_fp8_kernel(const Tin* __restrict__ src_cache,
+                                   Tout* __restrict__ dst_cache,
+                                   const float kv_scale,
+                                   const int64_t block_stride) {
   const int64_t block_idx = blockIdx.x;
   for (int i = threadIdx.x; i < block_stride; i += blockDim.x) {
     int64_t idx = block_idx * block_stride + i;
-#ifdef ENABLE_FP8_E5M2
-    dst_cache[idx] = fp8_e5m2_unscaled::vec_conversion<Tout, Tin>(src_cache[idx]);
-#else
-    assert(false);
-#endif
+    dst_cache[idx] =
+        fp8::scaled_convert<Tout, Tin, kv_dt>(src_cache[idx], kv_scale);
   }
 }
 
-} // namespace vllm
+}  // namespace vllm
 
-#define CALL_CONVERT_FP8_E5M2(Tout, Tin)                                 \
-  vllm::convert_fp8_e5m2_kernel<Tout, Tin><<<grid, block, 0, stream>>>(  \
-    reinterpret_cast<Tin*>(src_cache.data_ptr()),                        \
-    reinterpret_cast<Tout*>(dst_cache.data_ptr()),                       \
-    block_stride);
+#define CALL_CONVERT_FP8(Tout, Tin, KV_DTYPE)                                \
+  vllm::convert_fp8_kernel<Tout, Tin, KV_DTYPE><<<grid, block, 0, stream>>>( \
+      reinterpret_cast<Tin*>(src_cache.data_ptr()),                          \
+      reinterpret_cast<Tout*>(dst_cache.data_ptr()), kv_scale, block_stride);
+
+// Only for testing.
+void convert_fp8(torch::Tensor& dst_cache, torch::Tensor& src_cache,
+                 const float kv_scale, const std::string& kv_cache_dtype) {
+  torch::Device src_device = src_cache.device();
+  torch::Device dst_device = dst_cache.device();
+  TORCH_CHECK(src_device.is_cuda(), "src must be on a GPU")
+  TORCH_CHECK(dst_device.is_cuda(), "dst must be on a GPU")
+  TORCH_CHECK(src_device.index() == dst_device.index(),
+              "src and dst must be on the same GPU");
+  at::cuda::OptionalCUDAGuard device_guard(src_device);
 
-void convert_fp8_e5m2(
-  torch::Tensor& src_cache,
-  torch::Tensor& dst_cache)
-{
   int64_t num_blocks = src_cache.size(0);
   int64_t block_stride = src_cache.stride(0);
 
@@ -304,17 +350,37 @@ void convert_fp8_e5m2(
   dim3 block(std::min(block_stride, int64_t(512)));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
-  if (src_cache.dtype() == at::ScalarType::Float) {
-    CALL_CONVERT_FP8_E5M2(uint8_t, float);
-  } else if (src_cache.dtype() == at::ScalarType::Half) {
-    CALL_CONVERT_FP8_E5M2(uint8_t, uint16_t);
-  } else if (src_cache.dtype() == at::ScalarType::BFloat16) {
-    CALL_CONVERT_FP8_E5M2(uint8_t, __nv_bfloat16);
-  } else if (dst_cache.dtype() == at::ScalarType::Float) {
-    CALL_CONVERT_FP8_E5M2(float, uint8_t);
-  } else if (dst_cache.dtype() == at::ScalarType::Half) {
-    CALL_CONVERT_FP8_E5M2(uint16_t, uint8_t);
-  } else if (dst_cache.dtype() == at::ScalarType::BFloat16) {
-    CALL_CONVERT_FP8_E5M2(__nv_bfloat16, uint8_t);
+  if (kv_cache_dtype == "auto") {
+    if (src_cache.dtype() == at::ScalarType::Float) {
+      CALL_CONVERT_FP8(uint8_t, float, vllm::Fp8KVCacheDataType::kAuto);
+    } else if (src_cache.dtype() == at::ScalarType::Half) {
+      CALL_CONVERT_FP8(uint8_t, uint16_t, vllm::Fp8KVCacheDataType::kAuto);
+    } else if (src_cache.dtype() == at::ScalarType::BFloat16) {
+      CALL_CONVERT_FP8(uint8_t, __nv_bfloat16, vllm::Fp8KVCacheDataType::kAuto);
+    } else if (dst_cache.dtype() == at::ScalarType::Float) {
+      CALL_CONVERT_FP8(float, uint8_t, vllm::Fp8KVCacheDataType::kAuto);
+    } else if (dst_cache.dtype() == at::ScalarType::Half) {
+      CALL_CONVERT_FP8(uint16_t, uint8_t, vllm::Fp8KVCacheDataType::kAuto);
+    } else if (dst_cache.dtype() == at::ScalarType::BFloat16) {
+      CALL_CONVERT_FP8(__nv_bfloat16, uint8_t, vllm::Fp8KVCacheDataType::kAuto);
+    }
+  } else if (kv_cache_dtype == "fp8" || kv_cache_dtype == "fp8_e4m3") {
+    if (src_cache.dtype() == at::ScalarType::Float) {
+      CALL_CONVERT_FP8(uint8_t, float, vllm::Fp8KVCacheDataType::kFp8E4M3);
+    } else if (src_cache.dtype() == at::ScalarType::Half) {
+      CALL_CONVERT_FP8(uint8_t, uint16_t, vllm::Fp8KVCacheDataType::kFp8E4M3);
+    } else if (src_cache.dtype() == at::ScalarType::BFloat16) {
+      CALL_CONVERT_FP8(uint8_t, __nv_bfloat16,
+                       vllm::Fp8KVCacheDataType::kFp8E4M3);
+    } else if (dst_cache.dtype() == at::ScalarType::Float) {
+      CALL_CONVERT_FP8(float, uint8_t, vllm::Fp8KVCacheDataType::kFp8E4M3);
+    } else if (dst_cache.dtype() == at::ScalarType::Half) {
+      CALL_CONVERT_FP8(uint16_t, uint8_t, vllm::Fp8KVCacheDataType::kFp8E4M3);
+    } else if (dst_cache.dtype() == at::ScalarType::BFloat16) {
+      CALL_CONVERT_FP8(__nv_bfloat16, uint8_t,
+                       vllm::Fp8KVCacheDataType::kFp8E4M3);
+    }
+  } else {
+    TORCH_CHECK(false, "Unsupported data type: ", kv_cache_dtype);
   }
 }

csrc/cpu/activation.cpp

@@ -0,0 +1,144 @@
+#include "cpu_types.hpp"
+
+namespace {
+template <typename scalar_t, vec_op::FP32Vec8 (*func)(const vec_op::FP32Vec8&),
+          bool is_gated>
+void activation_kernel(int num_tokens, int d, scalar_t* __restrict__ input,
+                       scalar_t* __restrict__ output) {
+  using scalar_vec_t = vec_op::vec_t<scalar_t>;
+  constexpr int VEC_ELEM_NUM = scalar_vec_t::get_elem_num();
+
+  TORCH_CHECK(d % VEC_ELEM_NUM == 0);
+
+#pragma omp parallel for
+  for (int i = 0; i < num_tokens; ++i) {
+    for (int j = 0; j < d; j += VEC_ELEM_NUM) {
+      int start = i * d;
+      if constexpr (is_gated) {
+        start *= 2;
+      }
+
+      const scalar_vec_t x(input + start + j);
+      const vec_op::FP32Vec8 f32_x(x);
+      vec_op::FP32Vec8 f32_ans = func(f32_x);
+
+      if constexpr (is_gated) {
+        const scalar_vec_t y(input + start + d + j);
+        const vec_op::FP32Vec8 f32_y(y);
+        f32_ans = f32_y * f32_ans;
+      }
+
+      const scalar_vec_t result(f32_ans);
+      result.save(output + i * d + j);
+    }
+  }
+}
+
+FORCE_INLINE vec_op::FP32Vec8 silu_act(const vec_op::FP32Vec8& x) {
+  const vec_op::FP32Vec8 zeros(0.0);
+  const vec_op::FP32Vec8 ones(1.0);
+  return x / (ones + (zeros - x).exp());
+}
+
+FORCE_INLINE vec_op::FP32Vec8 gelu_new_act(const vec_op::FP32Vec8& x) {
+  const vec_op::FP32Vec8 ones(1.0);
+  const vec_op::FP32Vec8 w1(0.79788456f);
+  const vec_op::FP32Vec8 w2(0.044715f);
+  const vec_op::FP32Vec8 w3(0.5);
+  const vec_op::FP32Vec8 x3 = x * x * x;
+  const vec_op::FP32Vec8 t = (w1 * (x + w2 * x3)).tanh();
+  return w3 * x * (ones + t);
+}
+
+FORCE_INLINE vec_op::FP32Vec8 gelu_fast_act(const vec_op::FP32Vec8& x) {
+  const vec_op::FP32Vec8 ones(1.0);
+  const vec_op::FP32Vec8 w1(0.79788456f);
+  const vec_op::FP32Vec8 w2(0.044715f);
+  const vec_op::FP32Vec8 w3(0.5);
+  const vec_op::FP32Vec8 t = (x * w1 * (ones + x * w2 * x)).tanh();
+  return w3 * x * (ones + t);
+}
+
+FORCE_INLINE vec_op::FP32Vec8 gelu_act(const vec_op::FP32Vec8& x) {
+  const vec_op::FP32Vec8 ones(1.0);
+  const vec_op::FP32Vec8 w1(M_SQRT1_2);
+  const vec_op::FP32Vec8 w2(0.5);
+  return x * w2 * (ones + (x * w1).er());
+}
+
+FORCE_INLINE vec_op::FP32Vec8 gelu_tanh_act(const vec_op::FP32Vec8& x) {
+  const vec_op::FP32Vec8 ones(1.0);
+  const vec_op::FP32Vec8 w1(M_SQRT2 * M_2_SQRTPI * 0.5);
+  const vec_op::FP32Vec8 w2(0.5);
+  const vec_op::FP32Vec8 w3(0.044715);
+  const vec_op::FP32Vec8 x_3 = x * x * x;
+  const vec_op::FP32Vec8 inner = w1 * (x + x_3 * w3);
+  return x * w2 * (ones + inner.tanh());
+}
+};  // namespace
+
+void silu_and_mul(torch::Tensor& out, torch::Tensor& input) {
+  int num_tokens = input.numel() / input.size(-1);
+  int d = input.size(-1) / 2;
+
+  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "silu_and_mul_impl", [&] {
+    CPU_KERNEL_GUARD_IN(silu_and_mul_impl)
+    activation_kernel<scalar_t, silu_act, true>(
+        num_tokens, d, input.data_ptr<scalar_t>(), out.data_ptr<scalar_t>());
+    CPU_KERNEL_GUARD_OUT(silu_and_mul_impl)
+  });
+}
+
+void gelu_and_mul(torch::Tensor& out,    // [..., d]
+                  torch::Tensor& input)  // [..., 2 * d]
+{
+  int num_tokens = input.numel() / input.size(-1);
+  int d = input.size(-1) / 2;
+
+  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "gelu_and_mul_impl", [&] {
+    CPU_KERNEL_GUARD_IN(gelu_and_mul_impl)
+    activation_kernel<scalar_t, gelu_act, true>(
+        num_tokens, d, input.data_ptr<scalar_t>(), out.data_ptr<scalar_t>());
+    CPU_KERNEL_GUARD_OUT(gelu_and_mul_impl)
+  });
+}
+
+void gelu_tanh_and_mul(torch::Tensor& out,    // [..., d]
+                       torch::Tensor& input)  // [..., 2 * d]
+{
+  int num_tokens = input.numel() / input.size(-1);
+  int d = input.size(-1) / 2;
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "gelu_tanh_and_mul_impl", [&] {
+        CPU_KERNEL_GUARD_IN(gelu_tanh_and_mul_impl)
+        activation_kernel<scalar_t, gelu_tanh_act, true>(
+            num_tokens, d, input.data_ptr<scalar_t>(),
+            out.data_ptr<scalar_t>());
+        CPU_KERNEL_GUARD_OUT(gelu_tanh_and_mul_impl)
+      });
+}
+
+void gelu_new(torch::Tensor& out, torch::Tensor& input) {
+  int num_tokens = input.numel() / input.size(-1);
+  int d = input.size(-1);
+
+  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "gelu_new_impl", [&] {
+    CPU_KERNEL_GUARD_IN(gelu_new_impl)
+    activation_kernel<scalar_t, gelu_new_act, false>(
+        num_tokens, d, input.data_ptr<scalar_t>(), out.data_ptr<scalar_t>());
+    CPU_KERNEL_GUARD_OUT(gelu_new_impl)
+  });
+}
+
+void gelu_fast(torch::Tensor& out, torch::Tensor& input) {
+  int num_tokens = input.numel() / input.size(-1);
+  int d = input.size(-1);
+
+  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "gelu_fast_impl", [&] {
+    CPU_KERNEL_GUARD_IN(gelu_fast_impl)
+    activation_kernel<scalar_t, gelu_fast_act, false>(
+        num_tokens, d, input.data_ptr<scalar_t>(), out.data_ptr<scalar_t>());
+    CPU_KERNEL_GUARD_OUT(gelu_fast_impl)
+  });
+}

csrc/cpu/attention.cpp

@@ -0,0 +1,756 @@
+#include "cpu_types.hpp"
+
+namespace {
+
+template <typename scalar_t>
+struct KernelVecType {
+  using q_load_vec_type = void;
+  using q_vec_type = void;
+  using k_load_vec_type = void;
+  using k_vec_type = void;
+  using qk_acc_vec_type = void;
+  using v_load_vec_type = void;
+};
+
+template <>
+struct KernelVecType<float> {
+  using q_load_vec_type = vec_op::FP32Vec4;
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_load_vec_type = vec_op::FP32Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::FP32Vec16;
+};
+
+#ifdef __AVX512BF16__
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using q_load_vec_type = vec_op::BF16Vec8;
+  using q_vec_type = vec_op::BF16Vec32;
+  using k_load_vec_type = vec_op::BF16Vec32;
+  using k_vec_type = vec_op::BF16Vec32;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+#else
+template <>
+struct KernelVecType<c10::BFloat16> {
+  using q_load_vec_type = vec_op::BF16Vec8;
+  using q_vec_type = vec_op::FP32Vec16;
+  using k_load_vec_type = vec_op::BF16Vec16;
+  using k_vec_type = vec_op::FP32Vec16;
+  using qk_acc_vec_type = vec_op::FP32Vec16;
+  using v_load_vec_type = vec_op::BF16Vec16;
+};
+#endif
+
+template <typename T>
+FORCE_INLINE std::pair<T, T> reduceSoftmax(T* data, const int size,
+                                           const int capacity) {
+  T max = data[0];
+  for (int i = 1; i < size; ++i) {
+    max = max >= data[i] ? max : data[i];
+  }
+
+  T sum = 0;
+  for (int i = 0; i < size; ++i) {
+    data[i] = std::exp(data[i] - max);
+    sum += data[i];
+  }
+
+  int i = 0;
+  for (; i < size; ++i) {
+    data[i] /= sum;
+  }
+
+  for (; i < capacity; ++i) {
+    data[i] = 0;
+  }
+
+  return {max, sum};
+}
+
+template <typename T>
+FORCE_INLINE std::pair<T, T> reduceSoftmaxAlibi(T* data, const int size,
+                                                const int capacity,
+                                                const float alibi_slope,
+                                                const int start_index,
+                                                const int seq_len) {
+  data[0] += alibi_slope * (start_index - seq_len + 1);
+  T max = data[0];
+  for (int i = 1; i < size; ++i) {
+    T qk = data[i] + alibi_slope * (start_index + i - seq_len + 1);
+    data[i] = qk;
+    max = max >= qk ? max : qk;
+  }
+
+  T sum = 0;
+  for (int i = 0; i < size; ++i) {
+    data[i] = std::exp(data[i] - max);
+    sum += data[i];
+  }
+
+  int i = 0;
+  for (; i < size; ++i) {
+    data[i] /= sum;
+  }
+
+  for (; i < capacity; ++i) {
+    data[i] = 0;
+  }
+
+  return {max, sum};
+}
+
+template <typename T>
+FORCE_INLINE void reducePartitonSoftmax(const T* max_data, T* sum_data,
+                                        const int size) {
+  T max = max_data[0];
+  for (int i = 1; i < size; ++i) {
+    max = max >= max_data[i] ? max : max_data[i];
+  }
+
+  T rescaled_sum = 0;
+  for (int i = 0; i < size; ++i) {
+    T rescale_factor = std::exp(max_data[i] - max);
+    rescaled_sum += rescale_factor * sum_data[i];
+    sum_data[i] *= rescale_factor;
+  }
+  for (int i = 0; i < size; ++i) {
+    sum_data[i] /= rescaled_sum + 1e-8;
+  }
+}
+
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE, int x>
+struct reduceQKBlockKernel {
+  using q_load_vec_type = typename KernelVecType<scalar_t>::q_load_vec_type;
+  using q_vec_type = typename KernelVecType<scalar_t>::q_vec_type;
+  using k_load_vec_type = typename KernelVecType<scalar_t>::k_load_vec_type;
+  using k_vec_type = typename KernelVecType<scalar_t>::k_vec_type;
+  using qk_acc_vec_type = typename KernelVecType<scalar_t>::qk_acc_vec_type;
+
+  constexpr static int TOKEN_PER_GROUP = k_load_vec_type::get_elem_num() / x;
+  constexpr static int MAX_GROUP_NUM = 16 / TOKEN_PER_GROUP;
+  constexpr static int UNROLL_GROUP_NUM = MAX_GROUP_NUM / 4;
+
+  static_assert(MAX_GROUP_NUM == 8 || MAX_GROUP_NUM == 4);
+  static_assert(k_load_vec_type::get_elem_num() % x == 0);
+  static_assert(q_load_vec_type::get_elem_num() * sizeof(scalar_t) == 16);
+
+  FORCE_INLINE static void call(const scalar_t* __restrict__ q,
+                                const scalar_t* __restrict__ k_block,
+                                float* __restrict__ logits, float scale,
+                                const int token_num) {
+    const int group_num = (token_num + TOKEN_PER_GROUP - 1) / TOKEN_PER_GROUP;
+
+    qk_acc_vec_type group_accums[MAX_GROUP_NUM];
+    if (token_num == BLOCK_SIZE) {
+      for (int q_offset = 0; q_offset < HEAD_SIZE;
+           q_offset += x, k_block += x * BLOCK_SIZE) {
+        q_load_vec_type q_load_group_vec(q + q_offset);
+        q_vec_type q_group_vec(q_load_group_vec);
+
+        vec_op::unroll_loop<int, MAX_GROUP_NUM>(
+            [k_block, &q_group_vec, &group_accums](int token_group_idx) {
+              k_load_vec_type k_load_group_vec(k_block + token_group_idx * x *
+                                                             TOKEN_PER_GROUP);
+              k_vec_type k_group_vec(k_load_group_vec);
+              vec_op::fma(group_accums[token_group_idx], q_group_vec,
+                          k_group_vec);
+              vec_op::prefetch(k_block + x * BLOCK_SIZE +
+                               token_group_idx * x * TOKEN_PER_GROUP);
+            });
+      }
+    } else {
+      for (int q_offset = 0; q_offset < HEAD_SIZE;
+           q_offset += x, k_block += x * BLOCK_SIZE) {
+        q_load_vec_type q_load_group_vec(q + q_offset);
+        q_vec_type q_group_vec(q_load_group_vec);
+        for (int token_group_start = 0; token_group_start < group_num;
+             token_group_start += UNROLL_GROUP_NUM) {
+          vec_op::unroll_loop<int, UNROLL_GROUP_NUM>(
+              [token_group_start, k_block, &q_group_vec,
+               &group_accums](int token_group_idx) {
+                token_group_idx += token_group_start;
+                k_load_vec_type k_load_group_vec(k_block + token_group_idx * x *
+                                                               TOKEN_PER_GROUP);
+                k_vec_type k_group_vec(k_load_group_vec);
+                vec_op::fma(group_accums[token_group_idx], q_group_vec,
+                            k_group_vec);
+                vec_op::prefetch(k_block + x * BLOCK_SIZE +
+                                 token_group_idx * x * TOKEN_PER_GROUP);
+              });
+        }
+      }
+    }
+
+    for (int token_group_idx = 0; token_group_idx < group_num;
+         ++token_group_idx) {
+      vec_op::unroll_loop<int, TOKEN_PER_GROUP>(
+          [&group_accums, logits, scale, token_group_idx](int token_idx) {
+            float dot_v =
+                group_accums[token_group_idx]
+                    .template reduce_sub_sum<qk_acc_vec_type::get_elem_num() /
+                                             TOKEN_PER_GROUP>(token_idx);
+            logits[token_group_idx * TOKEN_PER_GROUP + token_idx] =
+                dot_v * scale;
+          });
+    }
+  }
+};
+
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE,
+          int HEAD_PARTITION_SIZE, typename acc_t>
+FORCE_INLINE void reduceValueBlock(const float* prob, const scalar_t* v_block,
+                                   acc_t&& acc) {
+  using v_load_vec_type = typename KernelVecType<scalar_t>::v_load_vec_type;
+  constexpr int ELEM_NUM = v_load_vec_type::get_elem_num();
+  static_assert(BLOCK_SIZE == ELEM_NUM);
+  vec_op::FP32Vec16 prob_vec(prob);
+
+  vec_op::unroll_loop<int, HEAD_PARTITION_SIZE>([&](int head_elem_idx) {
+    v_load_vec_type v_vec(v_block + BLOCK_SIZE * head_elem_idx);
+    vec_op::FP32Vec16 fp32_v_vec(v_vec);
+    acc[head_elem_idx] = acc[head_elem_idx] + prob_vec * fp32_v_vec;
+  });
+}
+};  // namespace
+
+// Paged attention v1
+namespace {
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE>
+struct paged_attention_v1_impl {
+  static void call(
+      scalar_t* __restrict__ out,            // [num_seqs, num_heads, head_size]
+      const scalar_t* __restrict__ q,        // [num_seqs, num_heads, head_size]
+      const scalar_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size/x, block_size, x]
+      const scalar_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size, block_size]
+      const int num_kv_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 float* __restrict__ alibi_slopes,  // [num_heads]
+      const int q_stride, const int kv_block_stride, const int kv_head_stride,
+      const int num_seqs, const int num_heads) {
+    constexpr int x = 16 / sizeof(scalar_t);
+    const int num_queries_per_kv = num_heads / num_kv_heads;
+
+    static_assert(BLOCK_SIZE == 16);
+
+    int max_seq_len = max_num_blocks_per_seq * BLOCK_SIZE;
+    int max_seq_len_padded = (max_seq_len + 15) & 0xFFFFFFF0;
+    TORCH_CHECK((max_seq_len_padded * sizeof(float)) % 64 == 0);
+
+    const int parallel_work_item_num = omp_get_max_threads();
+
+    size_t logits_bytes =
+        parallel_work_item_num * max_seq_len_padded * sizeof(float);
+    float* logits = (float*)std::aligned_alloc(
+        64, logits_bytes);  // Cacheline alignment for each context token.
+                            // [parallel_work_item_num, max_seq_len_padded]
+
+#pragma omp parallel for collapse(2) schedule(dynamic, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        int seq_len = seq_lens[seq_idx];
+        const int* seq_block_table =
+            block_tables + max_num_blocks_per_seq * seq_idx;
+        const int block_num = (seq_len + BLOCK_SIZE - 1) / BLOCK_SIZE;
+        const int64_t kv_head_idx = head_idx / num_queries_per_kv;
+        const scalar_t* __restrict__ q_vec_ptr =
+            q + seq_idx * q_stride + head_idx * HEAD_SIZE;
+        const int last_block_token_num = seq_len - (block_num - 1) * BLOCK_SIZE;
+        float* __restrict__ thread_block_logits =
+            logits + omp_get_thread_num() * max_seq_len_padded;
+
+        // Compute logits
+        for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+          const int64_t physical_block_idx = seq_block_table[block_idx];
+          const scalar_t* __restrict__ k_block_cache_ptr =
+              k_cache + physical_block_idx * kv_block_stride +
+              kv_head_idx * kv_head_stride;
+          float* __restrict__ head_block_logits =
+              thread_block_logits + block_idx * BLOCK_SIZE;
+
+          reduceQKBlockKernel<scalar_t, HEAD_SIZE, BLOCK_SIZE, x>::call(
+              q_vec_ptr, k_block_cache_ptr, head_block_logits, scale,
+              block_idx == block_num - 1 ? last_block_token_num : BLOCK_SIZE);
+        }
+
+        // Compute softmax
+        if (alibi_slopes) {
+          reduceSoftmaxAlibi(thread_block_logits, seq_len,
+                             block_num * BLOCK_SIZE, alibi_slopes[head_idx], 0,
+                             seq_len);
+        } else {
+          reduceSoftmax(thread_block_logits, seq_len, block_num * BLOCK_SIZE);
+        }
+
+        // Compute value
+        constexpr int head_elem_num_per_partition = 16;
+        constexpr int head_partition_num =
+            HEAD_SIZE / head_elem_num_per_partition;
+        for (int head_part_idx = 0; head_part_idx < head_partition_num;
+             ++head_part_idx) {
+          vec_op::FP32Vec16 accums[head_elem_num_per_partition];
+          scalar_t* __restrict__ out_ptr =
+              out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE +
+              head_part_idx * head_elem_num_per_partition;
+          for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+            const int64_t physical_block_idx = seq_block_table[block_idx];
+            const float* __restrict__ prob_vec_ptr =
+                thread_block_logits + block_idx * BLOCK_SIZE;
+            const scalar_t* __restrict__ v_block_cache_ptr =
+                v_cache + physical_block_idx * kv_block_stride +
+                kv_head_idx * kv_head_stride +
+                BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+            reduceValueBlock<scalar_t, HEAD_SIZE, BLOCK_SIZE,
+                             head_elem_num_per_partition>(
+                prob_vec_ptr, v_block_cache_ptr, accums);
+
+            if (block_idx != block_num - 1) {
+              const int64_t next_physical_block_idx =
+                  seq_block_table[block_idx + 1];
+              const scalar_t* __restrict__ next_v_block_cache_ptr =
+                  v_cache + next_physical_block_idx * kv_block_stride +
+                  kv_head_idx * kv_head_stride +
+                  BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+              vec_op::unroll_loop<int, head_elem_num_per_partition>(
+                  [&](int head_elem_idx) {
+                    if (head_elem_idx % 2 == 0) {
+                      vec_op::prefetch(next_v_block_cache_ptr +
+                                       BLOCK_SIZE * head_elem_idx);
+                    }
+                  });
+            }
+          }
+
+          vec_op::unroll_loop<int, head_elem_num_per_partition>(
+              [&](int head_elem_idx) {
+                float value = accums[head_elem_idx].reduce_sum();
+                vec_op::storeFP32(value, out_ptr + head_elem_idx);
+              });
+        }
+      }
+    }
+    std::free(logits);
+  }
+};
+
+#define LAUNCH_V1_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE)                   \
+  paged_attention_v1_impl<T, HEAD_SIZE, BLOCK_SIZE>::call(                     \
+      out_ptr, query_ptr, key_cache_ptr, value_cache_ptr, num_kv_heads, scale, \
+      block_tables_ptr, seq_lens_ptr, max_num_blocks_per_seq,                  \
+      alibi_slopes_ptr, q_stride, kv_block_stride, kv_head_stride, num_seqs,   \
+      num_heads);
+
+template <typename T, int BLOCK_SIZE>
+void paged_attention_v1_impl_launcher(
+    torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int max_seq_len,
+    const c10::optional<torch::Tensor>& alibi_slopes) {
+  int num_seqs = query.size(0);
+  int num_heads = query.size(1);
+  int head_size = query.size(2);
+  int max_num_blocks_per_seq = block_tables.size(1);
+  int q_stride = query.stride(0);
+  int kv_block_stride = key_cache.stride(0);
+  int kv_head_stride = key_cache.stride(1);
+
+  // NOTE: alibi_slopes is optional.
+  const float* alibi_slopes_ptr =
+      alibi_slopes
+          ? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
+          : nullptr;
+
+  T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
+  T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
+  T* key_cache_ptr = reinterpret_cast<T*>(key_cache.data_ptr());
+  T* value_cache_ptr = reinterpret_cast<T*>(value_cache.data_ptr());
+  int* block_tables_ptr = block_tables.data_ptr<int>();
+  int* seq_lens_ptr = seq_lens.data_ptr<int>();
+
+  switch (head_size) {
+    case 64:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 64, BLOCK_SIZE);
+      break;
+    case 80:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 80, BLOCK_SIZE);
+      break;
+    case 96:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 96, BLOCK_SIZE);
+      break;
+    case 112:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 112, BLOCK_SIZE);
+      break;
+    case 128:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 128, BLOCK_SIZE);
+      break;
+    case 192:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 192, BLOCK_SIZE);
+      break;
+    case 256:
+      LAUNCH_V1_ATTENTION_KERNEL(T, 256, BLOCK_SIZE);
+      break;
+    default:
+      TORCH_CHECK(false, "Unsupported head size: ", head_size);
+      break;
+  }
+}
+
+#define CALL_V1_KERNEL_LAUNCHER(T, BLOCK_SIZE)                               \
+  paged_attention_v1_impl_launcher<T, BLOCK_SIZE>(                           \
+      out, query, key_cache, value_cache, num_kv_heads, scale, block_tables, \
+      seq_lens, max_seq_len, alibi_slopes);
+
+#define CALL_V1_KERNEL_LAUNCHER_BLOCK_SIZE(T)                     \
+  switch (block_size) {                                           \
+    case 16:                                                      \
+      CALL_V1_KERNEL_LAUNCHER(T, 16);                             \
+      break;                                                      \
+    default:                                                      \
+      TORCH_CHECK(false, "Unsupported block size: ", block_size); \
+      break;                                                      \
+  }
+}  // namespace
+
+void paged_attention_v1(
+    torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int block_size,
+    int max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
+    const std::string& kv_cache_dtype, float kv_scale, const int tp_rank,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+  TORCH_CHECK(kv_scale == 1.0f);
+  TORCH_CHECK(blocksparse_vert_stride <= 1,
+              "CPU backend does not support blocksparse attention yet.");
+  VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v1_impl",
+                               [&] {
+                                 CPU_KERNEL_GUARD_IN(paged_attention_v1_impl)
+                                 CALL_V1_KERNEL_LAUNCHER_BLOCK_SIZE(scalar_t);
+                                 CPU_KERNEL_GUARD_OUT(paged_attention_v1_impl)
+                               });
+}
+
+// Paged attention v2
+namespace {
+template <typename scalar_t, int HEAD_SIZE, int BLOCK_SIZE, int PARTITION_SIZE>
+struct paged_attention_v2_impl {
+  static void call(
+      scalar_t* __restrict__ out,            // [num_seqs, num_heads, head_size]
+      float* __restrict__ exp_sums,          // [num_seqs, num_heads,
+                                             // max_num_partitions]
+      float* __restrict__ max_logits,        // [num_seqs, num_heads,
+                                             // max_num_partitions]
+      scalar_t* __restrict__ tmp_out,        // [num_seqs, num_heads,
+                                             // max_num_partitions, head_size]
+      const scalar_t* __restrict__ q,        // [num_seqs, num_heads, head_size]
+      const scalar_t* __restrict__ k_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size/x, block_size, x]
+      const scalar_t* __restrict__ v_cache,  // [num_blocks, num_kv_heads,
+                                             // head_size, block_size]
+      const int num_kv_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 float* __restrict__ alibi_slopes,  // [num_heads]
+      const int q_stride, const int kv_block_stride, const int kv_head_stride,
+      const int num_seqs, const int num_heads, const int max_num_partitions) {
+    constexpr int x = 16 / sizeof(scalar_t);
+    const int num_queries_per_kv = num_heads / num_kv_heads;
+
+    static_assert(BLOCK_SIZE == 16);
+    static_assert(PARTITION_SIZE * sizeof(float) % 64 == 0);
+    static_assert(PARTITION_SIZE % BLOCK_SIZE == 0);
+
+#pragma omp parallel for collapse(3) schedule(static, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int partition_idx = 0; partition_idx < max_num_partitions;
+           ++partition_idx) {
+        for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+          const int seq_len = seq_lens[seq_idx];
+          const int start_token_idx = partition_idx * PARTITION_SIZE;
+
+          if (start_token_idx >= seq_len) continue;
+
+          const int partition_num =
+              (seq_len + PARTITION_SIZE - 1) / PARTITION_SIZE;
+          const bool no_reduce = (partition_num == 1);
+          const int token_num =
+              (std::min(seq_len, start_token_idx + PARTITION_SIZE) -
+               start_token_idx);
+          const int block_num = (token_num + BLOCK_SIZE - 1) / BLOCK_SIZE;
+          const int last_block_token_num =
+              token_num - (block_num - 1) * BLOCK_SIZE;
+          const int* seq_block_table = block_tables +
+                                       max_num_blocks_per_seq * seq_idx +
+                                       start_token_idx / BLOCK_SIZE;
+          const int64_t kv_head_idx = head_idx / num_queries_per_kv;
+          const scalar_t* __restrict__ q_vec_ptr =
+              q + seq_idx * q_stride + head_idx * HEAD_SIZE;
+
+          float logits[PARTITION_SIZE] __attribute__((aligned(64))) = {0};
+
+          // Compute logits
+          for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+            const int64_t physical_block_idx = seq_block_table[block_idx];
+            const scalar_t* __restrict__ k_block_cache_ptr =
+                k_cache + physical_block_idx * kv_block_stride +
+                kv_head_idx * kv_head_stride;
+            float* __restrict__ head_block_logits =
+                logits + block_idx * BLOCK_SIZE;
+
+            reduceQKBlockKernel<scalar_t, HEAD_SIZE, BLOCK_SIZE, x>::call(
+                q_vec_ptr, k_block_cache_ptr, head_block_logits, scale,
+                block_idx == block_num - 1 ? last_block_token_num : BLOCK_SIZE);
+          }
+
+          std::pair<float, float> max_and_sum;
+          if (alibi_slopes) {
+            max_and_sum = reduceSoftmaxAlibi(
+                logits, token_num, block_num * BLOCK_SIZE,
+                alibi_slopes[head_idx], start_token_idx, seq_len);
+          } else {
+            max_and_sum =
+                reduceSoftmax(logits, token_num, block_num * BLOCK_SIZE);
+          }
+
+          auto&& [max_logit, exp_sum] = max_and_sum;
+
+          scalar_t* __restrict__ output_buffer = nullptr;
+          if (!no_reduce) {
+            auto idx = seq_idx * num_heads * max_num_partitions +
+                       head_idx * max_num_partitions + partition_idx;
+            max_logits[idx] = max_logit;
+            exp_sums[idx] = exp_sum;
+            output_buffer =
+                tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
+                head_idx * max_num_partitions * HEAD_SIZE +
+                partition_idx * HEAD_SIZE;
+          } else {
+            output_buffer =
+                out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE;
+          }
+
+          // Compute value
+          constexpr int head_elem_num_per_partition = 16;
+          constexpr int head_partition_num =
+              HEAD_SIZE / head_elem_num_per_partition;
+          for (int head_part_idx = 0; head_part_idx < head_partition_num;
+               ++head_part_idx) {
+            vec_op::FP32Vec16 accums[head_elem_num_per_partition];
+            scalar_t* __restrict__ out_ptr =
+                output_buffer + head_part_idx * head_elem_num_per_partition;
+            for (int block_idx = 0; block_idx < block_num; ++block_idx) {
+              const int64_t physical_block_idx = seq_block_table[block_idx];
+              const float* __restrict__ prob_vec_ptr =
+                  logits + block_idx * BLOCK_SIZE;
+              const scalar_t* __restrict__ v_block_cache_ptr =
+                  v_cache + physical_block_idx * kv_block_stride +
+                  kv_head_idx * kv_head_stride +
+                  BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+              reduceValueBlock<scalar_t, HEAD_SIZE, BLOCK_SIZE,
+                               head_elem_num_per_partition>(
+                  prob_vec_ptr, v_block_cache_ptr, accums);
+
+              if (block_idx != block_num - 1) {
+                const int64_t next_physical_block_idx =
+                    seq_block_table[block_idx + 1];
+                const scalar_t* __restrict__ next_v_block_cache_ptr =
+                    v_cache + next_physical_block_idx * kv_block_stride +
+                    kv_head_idx * kv_head_stride +
+                    BLOCK_SIZE * head_part_idx * head_elem_num_per_partition;
+                vec_op::unroll_loop<int, head_elem_num_per_partition>(
+                    [&](int head_elem_idx) {
+                      if (head_elem_idx % 2 == 0) {
+                        vec_op::prefetch(next_v_block_cache_ptr +
+                                         BLOCK_SIZE * head_elem_idx);
+                      }
+                    });
+              }
+            }
+
+            vec_op::unroll_loop<int, head_elem_num_per_partition>(
+                [&](int head_elem_idx) {
+                  float value = accums[head_elem_idx].reduce_sum();
+                  vec_op::storeFP32(value, out_ptr + head_elem_idx);
+                });
+          }
+        }
+      }
+    }
+
+    // Rescale partition softmax and store the factors to exp_sums
+#pragma omp parallel for collapse(2) schedule(static, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        const int seq_len = seq_lens[seq_idx];
+        const int partition_num =
+            (seq_len + PARTITION_SIZE - 1) / PARTITION_SIZE;
+
+        if (partition_num == 1) continue;
+
+        reducePartitonSoftmax(
+            max_logits + seq_idx * num_heads * max_num_partitions +
+                head_idx * max_num_partitions,
+            exp_sums + seq_idx * num_heads * max_num_partitions +
+                head_idx * max_num_partitions,
+            partition_num);
+      }
+    }
+
+    // Reduce values
+    using v_load_vec_type = typename KernelVecType<scalar_t>::v_load_vec_type;
+    static_assert(v_load_vec_type::get_elem_num() == BLOCK_SIZE);
+    constexpr int head_elem_num_per_group =
+        16;  // Note: didn't align with the cacheline size, due to some
+             // HEAD_SIZE didn't align with 64 bytes
+    static_assert(HEAD_SIZE % head_elem_num_per_group == 0);
+    constexpr int head_group_num = HEAD_SIZE / head_elem_num_per_group;
+    const float* __restrict__ rescale_factors = exp_sums;
+#pragma omp parallel for collapse(3) schedule(static, 1)
+    for (int seq_idx = 0; seq_idx < num_seqs; ++seq_idx) {
+      for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+        for (int group_idx = 0; group_idx < head_group_num; ++group_idx) {
+          const int seq_len = seq_lens[seq_idx];
+          const int partition_num =
+              (seq_len + PARTITION_SIZE - 1) / PARTITION_SIZE;
+
+          if (partition_num == 1) continue;
+
+          const float* __restrict__ seq_head_rescale_factors =
+              rescale_factors + seq_idx * num_heads * max_num_partitions +
+              head_idx * max_num_partitions;
+          const scalar_t* __restrict__ seq_head_tmp_out =
+              tmp_out + seq_idx * num_heads * max_num_partitions * HEAD_SIZE +
+              head_idx * max_num_partitions * HEAD_SIZE +
+              group_idx * head_elem_num_per_group;
+          scalar_t* __restrict__ seq_head_output =
+              out + seq_idx * num_heads * HEAD_SIZE + head_idx * HEAD_SIZE +
+              group_idx * head_elem_num_per_group;
+
+          vec_op::FP32Vec16 acc;
+          for (int i = 0; i < partition_num; ++i) {
+            vec_op::FP32Vec16 rescale_factor(seq_head_rescale_factors[i]);
+            v_load_vec_type value(seq_head_tmp_out + i * HEAD_SIZE);
+            vec_op::FP32Vec16 fp32_value(value);
+            acc = acc + fp32_value * rescale_factor;
+          }
+          v_load_vec_type cast_acc(acc);
+          cast_acc.save(seq_head_output);
+        }
+      }
+    }
+  }
+};
+
+#define LAUNCH_V2_ATTENTION_KERNEL(T, HEAD_SIZE, BLOCK_SIZE)                 \
+  paged_attention_v2_impl<T, HEAD_SIZE, BLOCK_SIZE, PARTITION_SIZE>::call(   \
+      out_ptr, exp_sums_ptr, max_logits_ptr, tmp_out_ptr, query_ptr,         \
+      key_cache_ptr, value_cache_ptr, num_kv_heads, scale, block_tables_ptr, \
+      seq_lens_ptr, max_num_blocks_per_seq, alibi_slopes_ptr, q_stride,      \
+      kv_block_stride, kv_head_stride, num_seqs, num_heads,                  \
+      max_num_partitions);
+
+template <typename T, int BLOCK_SIZE, int PARTITION_SIZE = 512>
+void paged_attention_v2_impl_launcher(
+    torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
+    torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int block_size,
+    int max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes) {
+  int num_seqs = query.size(0);
+  int num_heads = query.size(1);
+  int head_size = query.size(2);
+  int max_num_blocks_per_seq = block_tables.size(1);
+  int q_stride = query.stride(0);
+  int kv_block_stride = key_cache.stride(0);
+  int kv_head_stride = key_cache.stride(1);
+  int max_num_partitions = exp_sums.size(-1);
+
+  // NOTE: alibi_slopes is optional.
+  const float* alibi_slopes_ptr =
+      alibi_slopes
+          ? reinterpret_cast<const float*>(alibi_slopes.value().data_ptr())
+          : nullptr;
+
+  T* out_ptr = reinterpret_cast<T*>(out.data_ptr());
+  float* exp_sums_ptr = reinterpret_cast<float*>(exp_sums.data_ptr());
+  float* max_logits_ptr = reinterpret_cast<float*>(max_logits.data_ptr());
+  T* tmp_out_ptr = reinterpret_cast<T*>(tmp_out.data_ptr());
+  T* query_ptr = reinterpret_cast<T*>(query.data_ptr());
+  T* key_cache_ptr = reinterpret_cast<T*>(key_cache.data_ptr());
+  T* value_cache_ptr = reinterpret_cast<T*>(value_cache.data_ptr());
+  int* block_tables_ptr = block_tables.data_ptr<int>();
+  int* seq_lens_ptr = seq_lens.data_ptr<int>();
+
+  switch (head_size) {
+    case 64:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 64, BLOCK_SIZE);
+      break;
+    case 80:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 80, BLOCK_SIZE);
+      break;
+    case 96:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 96, BLOCK_SIZE);
+      break;
+    case 112:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 112, BLOCK_SIZE);
+      break;
+    case 128:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 128, BLOCK_SIZE);
+      break;
+    case 192:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 192, BLOCK_SIZE);
+      break;
+    case 256:
+      LAUNCH_V2_ATTENTION_KERNEL(T, 256, BLOCK_SIZE);
+      break;
+    default:
+      TORCH_CHECK(false, "Unsupported head size: ", head_size);
+      break;
+  }
+}
+
+#define CALL_V2_KERNEL_LAUNCHER(T, BLOCK_SIZE)                              \
+  paged_attention_v2_impl_launcher<T, BLOCK_SIZE>(                          \
+      out, exp_sums, max_logits, tmp_out, query, key_cache, value_cache,    \
+      num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len, \
+      alibi_slopes);
+
+#define CALL_V2_KERNEL_LAUNCHER_BLOCK_SIZE(T)                     \
+  switch (block_size) {                                           \
+    case 16:                                                      \
+      CALL_V2_KERNEL_LAUNCHER(T, 16);                             \
+      break;                                                      \
+    default:                                                      \
+      TORCH_CHECK(false, "Unsupported block size: ", block_size); \
+      break;                                                      \
+  }
+}  // namespace
+
+void paged_attention_v2(
+    torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
+    torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int block_size,
+    int max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
+    const std::string& kv_cache_dtype, float kv_scale, const int tp_rank,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step) {
+  TORCH_CHECK(kv_scale == 1.0f);
+  TORCH_CHECK(blocksparse_vert_stride <= 1,
+              "CPU backend does not support blocksparse attention yet.");
+  VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "paged_attention_v2_impl",
+                               [&] {
+                                 CPU_KERNEL_GUARD_IN(paged_attention_v2_impl)
+                                 CALL_V2_KERNEL_LAUNCHER_BLOCK_SIZE(scalar_t);
+                                 CPU_KERNEL_GUARD_OUT(paged_attention_v2_impl)
+                               });
+}

csrc/cpu/cache.cpp

@@ -0,0 +1,134 @@
+#include <map>
+#include <vector>
+
+#include "cpu_types.hpp"
+
+namespace {
+template <typename scalar_t>
+void copy_blocks_cpu_impl(std::vector<torch::Tensor>& key_caches,
+                          std::vector<torch::Tensor>& value_caches,
+                          const torch::Tensor& mapping_pairs,
+                          const int element_num_per_block,
+                          const int layer_num) {
+  const size_t pair_num = mapping_pairs.size(0);
+  const size_t block_bytes = sizeof(scalar_t) * element_num_per_block;
+#pragma omp parallel for collapse(2)
+  for (int layer = 0; layer < layer_num; ++layer) {
+    for (size_t pair = 0; pair < pair_num; ++pair) {
+      int64_t source_offset =
+          element_num_per_block * mapping_pairs[pair][0].item<int64_t>();
+      int64_t target_offset =
+          element_num_per_block * mapping_pairs[pair][1].item<int64_t>();
+      scalar_t* key_cache_ptr = key_caches[layer].data_ptr<scalar_t>();
+      scalar_t* source_ptr = key_cache_ptr + source_offset;
+      scalar_t* target_ptr = key_cache_ptr + target_offset;
+      std::memcpy(target_ptr, source_ptr, block_bytes);
+
+      scalar_t* value_cache_ptr = value_caches[layer].data_ptr<scalar_t>();
+      source_ptr = value_cache_ptr + source_offset;
+      target_ptr = value_cache_ptr + target_offset;
+      std::memcpy(target_ptr, source_ptr, block_bytes);
+    }
+  }
+}
+
+template <typename scalar_t>
+void reshape_and_cache_cpu_impl(
+    const scalar_t* __restrict__ key, const scalar_t* __restrict__ value,
+    scalar_t* __restrict__ key_cache, scalar_t* __restrict__ value_cache,
+    const int64_t* __restrict__ slot_mapping, const int num_tokens,
+    const int key_stride, const int value_stride, const int num_heads,
+    const int head_size, const int block_size, const int x) {
+  const int block_elem_num = num_heads * head_size * block_size;
+
+#pragma omp parallel for collapse(2)
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    for (int head_idx = 0; head_idx < num_heads; ++head_idx) {
+      const int64_t slot_idx = slot_mapping[token_idx];
+      if (slot_idx >= 0) {
+        int src_key_head_idx = token_idx * key_stride + head_idx * head_size;
+        int src_value_head_idx =
+            token_idx * value_stride + head_idx * head_size;
+        const scalar_t* src_key_head_ptr = key + src_key_head_idx;
+        const scalar_t* src_value_head_ptr = value + src_value_head_idx;
+        const int64_t block_index = slot_idx / block_size;
+        const int64_t block_offset = slot_idx % block_size;
+        scalar_t* target_key_head_ptr = key_cache +
+                                        block_elem_num * block_index +
+                                        head_idx * block_size * head_size;
+        scalar_t* target_value_head_ptr = value_cache +
+                                          block_elem_num * block_index +
+                                          head_idx * block_size * head_size;
+
+        for (int src_key_idx = 0; src_key_idx < head_size; src_key_idx += x) {
+          const int64_t target_offset =
+              src_key_idx * block_size + block_offset * x;
+          for (int i = 0; i < x; ++i) {
+            target_key_head_ptr[target_offset + i] =
+                src_key_head_ptr[src_key_idx + i];
+          }
+        }
+
+        for (int src_value_idx = 0; src_value_idx < head_size;
+             ++src_value_idx) {
+          const int64_t target_offset =
+              src_value_idx * block_size + block_offset;
+          target_value_head_ptr[target_offset] =
+              src_value_head_ptr[src_value_idx];
+        }
+      }
+    }
+  }
+}
+};  // namespace
+
+void copy_blocks(std::vector<torch::Tensor>& key_caches,
+                 std::vector<torch::Tensor>& value_caches,
+                 const torch::Tensor& block_mapping) {
+  unsigned num_layers = key_caches.size();
+  TORCH_CHECK(num_layers == value_caches.size());
+  if (num_layers == 0) {
+    return;
+  }
+
+  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)
+      });
+}
+
+void reshape_and_cache(torch::Tensor& key, torch::Tensor& value,
+                       torch::Tensor& key_cache, torch::Tensor& value_cache,
+                       torch::Tensor& slot_mapping,
+                       const std::string& kv_cache_dtype, float kv_scale) {
+  TORCH_CHECK(kv_scale == 1.0f);
+
+  int num_tokens = key.size(0);
+  int num_heads = key.size(1);
+  int head_size = key.size(2);
+  int block_size = key_cache.size(3);
+  int x = key_cache.size(4);
+
+  int key_stride = key.stride(0);
+  int value_stride = value.stride(0);
+
+  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)
+      });
+}
+
+void swap_blocks(torch::Tensor& src, torch::Tensor& dst,
+                 const torch::Tensor& block_mapping) {
+  TORCH_CHECK(false, "swap_blocks is unsupported on CPU.")
+}

csrc/cpu/cpu_types.hpp

@@ -0,0 +1,352 @@
+
+#ifndef CPU_TYPES_HPP
+#define CPU_TYPES_HPP
+
+#include <immintrin.h>
+#include <torch/extension.h>
+
+namespace vec_op {
+
+// 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; }
+};
+
+struct FP32Vec8;
+struct FP32Vec16;
+
+#ifdef __AVX512FP16__
+struct FP16Vec8 : public Vec<FP16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+
+  __m128h reg;
+
+  explicit FP16Vec8(_Float16 v) : reg(_mm_set1_ph(v)) {}
+
+  explicit FP16Vec8(const void *ptr) : reg(_mm_loadu_ph(ptr)) {}
+
+  explicit FP16Vec8(__m128h data) : reg(data) {}
+
+  FP16Vec8 operator*(const FP16Vec8 &b) const {
+    return FP16Vec8(_mm_mul_ph(reg, b.reg));
+  }
+
+  FP16Vec8 operator+(const FP16Vec8 &b) const {
+    return FP16Vec8(_mm_add_ph(reg, b.reg));
+  }
+
+  FP16Vec8 operator-(const FP16Vec8 &b) const {
+    return FP16Vec8(_mm_sub_ph(reg, b.reg));
+  }
+
+  FP16Vec8 operator/(const FP16Vec8 &b) const {
+    return FP16Vec8(_mm_div_ph(reg, b.reg));
+  }
+
+  void save(void *ptr) const { _mm_storeu_ph(ptr, reg); }
+};
+#endif
+
+struct BF16Vec8 : public Vec<BF16Vec8> {
+  constexpr static int VEC_ELEM_NUM = 8;
+
+  __m128i reg;
+
+  explicit BF16Vec8(const void *ptr)
+      : reg((__m128i)_mm_loadu_si128((__m128i *)ptr)) {}
+
+  explicit BF16Vec8(const FP32Vec8 &);
+
+  void save(void *ptr) const { *reinterpret_cast<__m128i *>(ptr) = reg; }
+};
+
+struct BF16Vec16 : public Vec<BF16Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+
+  __m256i reg;
+
+  explicit BF16Vec16(const void *ptr)
+      : reg((__m256i)_mm256_loadu_si256((__m256i *)ptr)) {}
+
+  explicit BF16Vec16(const FP32Vec16 &);
+
+  void save(void *ptr) const { *reinterpret_cast<__m256i *>(ptr) = reg; }
+};
+
+struct BF16Vec32 : public Vec<BF16Vec32> {
+  constexpr static int VEC_ELEM_NUM = 32;
+
+  __m512i reg;
+
+  explicit BF16Vec32(const void *ptr) : reg((__m512i)_mm512_loadu_si512(ptr)) {}
+
+  explicit BF16Vec32(__m512i data) : reg(data) {}
+
+  explicit BF16Vec32(BF16Vec8 &vec8_data)
+      : reg((__m512i)_mm512_inserti32x4(
+            _mm512_inserti32x4(_mm512_inserti32x4(_mm512_castsi128_si512(
+                                                      (__m128i)vec8_data.reg),
+                                                  (__m128i)vec8_data.reg, 1),
+                               (__m128i)vec8_data.reg, 2),
+            (__m128i)vec8_data.reg, 3)) {}
+
+  void save(void *ptr) const { *reinterpret_cast<__m512i *>(ptr) = reg; }
+};
+
+struct FP32Vec4 : public Vec<FP32Vec4> {
+  constexpr static int VEC_ELEM_NUM = 4;
+  union AliasReg {
+    __m128 reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  __m128 reg;
+
+  explicit FP32Vec4(float v) : reg(_mm_set1_ps(v)) {}
+
+  explicit FP32Vec4() : reg(_mm_set1_ps(0.0)) {}
+
+  explicit FP32Vec4(const float *ptr) : reg(_mm_loadu_ps(ptr)) {}
+
+  explicit FP32Vec4(__m128 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 {
+    __m256 reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  __m256 reg;
+
+  explicit FP32Vec8(float v) : reg(_mm256_set1_ps(v)) {}
+
+  explicit FP32Vec8() : reg(_mm256_set1_ps(0.0)) {}
+
+  explicit FP32Vec8(const float *ptr) : reg(_mm256_loadu_ps(ptr)) {}
+
+  explicit FP32Vec8(__m256 data) : reg(data) {}
+
+  explicit FP32Vec8(const FP32Vec8 &data) : reg(data.reg) {}
+
+#ifdef __AVX512FP16__
+  explicit FP32Vec8(__m128h v) : reg(_mm256_cvtph_ps(_mm_castph_si128(v))) {}
+#endif
+
+  explicit FP32Vec8(const BF16Vec8 &v)
+      : reg(_mm256_castsi256_ps(
+            _mm256_bslli_epi128(_mm256_cvtepu16_epi32(v.reg), 2))) {}
+
+  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 {
+    AliasReg ar;
+    ar.reg = reg;
+    return FP32Vec8(_mm256_set_ps(expf(ar.values[7]), expf(ar.values[6]),
+                                  expf(ar.values[5]), expf(ar.values[4]),
+                                  expf(ar.values[3]), expf(ar.values[2]),
+                                  expf(ar.values[1]), expf(ar.values[0])));
+  }
+
+  FP32Vec8 tanh() const {
+    AliasReg ar;
+    ar.reg = reg;
+    return FP32Vec8(_mm256_set_ps(tanhf(ar.values[7]), tanhf(ar.values[6]),
+                                  tanhf(ar.values[5]), tanhf(ar.values[4]),
+                                  tanhf(ar.values[3]), tanhf(ar.values[2]),
+                                  tanhf(ar.values[1]), tanhf(ar.values[0])));
+  }
+
+  FP32Vec8 er() const {
+    AliasReg ar;
+    ar.reg = reg;
+    return FP32Vec8(_mm256_set_ps(erf(ar.values[7]), erf(ar.values[6]),
+                                  erf(ar.values[5]), erf(ar.values[4]),
+                                  erf(ar.values[3]), erf(ar.values[2]),
+                                  erf(ar.values[1]), erf(ar.values[0])));
+  }
+
+  FP32Vec8 operator*(const FP32Vec8 &b) const {
+    return FP32Vec8(_mm256_mul_ps(reg, b.reg));
+  }
+
+  FP32Vec8 operator+(const FP32Vec8 &b) const {
+    return FP32Vec8(_mm256_add_ps(reg, b.reg));
+  }
+
+  FP32Vec8 operator-(const FP32Vec8 &b) const {
+    return FP32Vec8(_mm256_sub_ps(reg, b.reg));
+  }
+
+  FP32Vec8 operator/(const FP32Vec8 &b) const {
+    return FP32Vec8(_mm256_div_ps(reg, b.reg));
+  }
+
+  void save(float *ptr) const { _mm256_storeu_ps(ptr, reg); }
+};
+
+struct FP32Vec16 : public Vec<FP32Vec16> {
+  constexpr static int VEC_ELEM_NUM = 16;
+  union AliasReg {
+    __m512 reg;
+    float values[VEC_ELEM_NUM];
+  };
+
+  __m512 reg;
+
+  explicit FP32Vec16(float v) : reg(_mm512_set1_ps(v)) {}
+
+  explicit FP32Vec16() : reg(_mm512_set1_ps(0.0)) {}
+
+  explicit FP32Vec16(const float *ptr) : reg(_mm512_loadu_ps(ptr)) {}
+
+  explicit FP32Vec16(__m512 data) : reg(data) {}
+
+  explicit FP32Vec16(const FP32Vec16 &data) : reg(data.reg) {}
+
+  explicit FP32Vec16(const FP32Vec4 &data)
+      : reg((__m512)_mm512_inserti32x4(
+            _mm512_inserti32x4(
+                _mm512_inserti32x4(_mm512_castsi128_si512((__m128i)data.reg),
+                                   (__m128i)data.reg, 1),
+                (__m128i)data.reg, 2),
+            (__m128i)data.reg, 3)) {}
+
+  explicit FP32Vec16(const FP32Vec8 &data)
+      : reg((__m512)_mm512_inserti32x8(
+            _mm512_castsi256_si512((__m256i)data.reg), (__m256i)data.reg, 1)) {}
+
+  explicit FP32Vec16(const BF16Vec16 &v)
+      : reg(_mm512_castsi512_ps(
+            _mm512_bslli_epi128(_mm512_cvtepu16_epi32(v.reg), 2))) {}
+
+  explicit FP32Vec16(const BF16Vec8 &v) : FP32Vec16(FP32Vec8(v)) {}
+
+  FP32Vec16 operator*(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm512_mul_ps(reg, b.reg));
+  }
+
+  FP32Vec16 operator+(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm512_add_ps(reg, b.reg));
+  }
+
+  FP32Vec16 operator-(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm512_sub_ps(reg, b.reg));
+  }
+
+  FP32Vec16 operator/(const FP32Vec16 &b) const {
+    return FP32Vec16(_mm512_div_ps(reg, b.reg));
+  }
+
+  float reduce_sum() const { return _mm512_reduce_add_ps(reg); }
+
+  template <int group_size> float reduce_sub_sum(int idx) {
+    static_assert(VEC_ELEM_NUM % group_size == 0);
+    constexpr uint32_t base_mask = (0xFFFF >> (16 - group_size));
+    __mmask16 mask = _cvtu32_mask16(base_mask << (idx * group_size));
+    return _mm512_mask_reduce_add_ps(mask, reg);
+  }
+
+  void save(float *ptr) const { _mm512_storeu_ps(ptr, reg); }
+};
+
+template <typename T> struct VecType { using vec_type = void; };
+
+template <typename T> using vec_t = typename VecType<T>::vec_type;
+
+template <> struct VecType<float> { using vec_type = FP32Vec8; };
+
+#ifdef __AVX512FP16__
+template <> struct VecType<c10::Half> { using vec_type = FP16Vec16; };
+#endif
+
+template <> struct VecType<c10::BFloat16> { using vec_type = BF16Vec8; };
+
+template <typename T> void storeFP32(float v, T *ptr) { *ptr = v; }
+
+#ifdef __AVX512FP16__
+template <> inline void storeFP32<c10::Half>(float v, c10::Half *ptr) {
+  *reinterpret_cast<_Float16 *>(ptr) = v;
+}
+#endif
+
+inline void fma(FP32Vec16 &acc, FP32Vec16 &a, FP32Vec16 &b) {
+  acc = acc + a * b;
+}
+
+#ifdef __AVX512BF16__
+template <> inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) {
+  *reinterpret_cast<__bfloat16 *>(ptr) = _mm_cvtness_sbh(v);
+}
+
+inline BF16Vec8::BF16Vec8(const FP32Vec8 &v)
+    : reg((__m128i)_mm256_cvtneps_pbh(v.reg)) {}
+
+inline BF16Vec16::BF16Vec16(const FP32Vec16 &v)
+    : reg((__m256i)_mm512_cvtneps_pbh(v.reg)) {}
+
+inline void fma(FP32Vec16 &acc, BF16Vec32 &a, BF16Vec32 &b) {
+  acc.reg = _mm512_dpbf16_ps(acc.reg, (__m512bh)a.reg, (__m512bh)b.reg);
+}
+#else
+template <> inline void storeFP32<c10::BFloat16>(float v, c10::BFloat16 *ptr) {
+  c10::BFloat16 __attribute__((__may_alias__)) *v_ptr =
+      reinterpret_cast<c10::BFloat16 *>(&v);
+  *ptr = *(v_ptr + 1);
+}
+
+inline BF16Vec8::BF16Vec8(const FP32Vec8 &v)
+    : reg(_mm256_cvtepi32_epi16(
+          _mm256_bsrli_epi128(_mm256_castps_si256(v.reg), 2))) {}
+
+inline BF16Vec16::BF16Vec16(const FP32Vec16 &v)
+    : reg(_mm512_cvtepi32_epi16(
+          _mm512_bsrli_epi128(_mm512_castps_si512(v.reg), 2))) {}
+#endif
+
+inline void prefetch(const void *addr) { _mm_prefetch(addr, _MM_HINT_T1); }
+
+}; // namespace vec_op
+
+#endif

csrc/cpu/layernorm.cpp

@@ -0,0 +1,117 @@
+#include "cpu_types.hpp"
+
+namespace {
+template <typename scalar_t>
+void rms_norm_impl(scalar_t* __restrict__ out,
+                   const scalar_t* __restrict__ input,
+                   const scalar_t* __restrict__ weight, const float epsilon,
+                   const int num_tokens, const int hidden_size) {
+  using scalar_vec_t = vec_op::vec_t<scalar_t>;
+  constexpr int VEC_ELEM_NUM = scalar_vec_t::get_elem_num();
+  TORCH_CHECK(hidden_size % VEC_ELEM_NUM == 0);
+
+#pragma omp parallel for
+  for (int i = 0; i < num_tokens; ++i) {
+    vec_op::FP32Vec8 variance(0.0);
+    auto input_p = input + i * hidden_size;
+    auto output_p = out + i * hidden_size;
+    for (int j = 0; j < hidden_size; j += VEC_ELEM_NUM) {
+      scalar_vec_t x(input_p + j);
+      vec_op::FP32Vec8 fp32_x(x);
+      variance = variance + fp32_x * fp32_x;
+    }
+
+    float s_variance =
+        1.0f / sqrtf(variance.reduce_sum() / (float)hidden_size + epsilon);
+    vec_op::FP32Vec8 fp32_s_variance(s_variance);
+
+    for (int j = 0; j < hidden_size; j += VEC_ELEM_NUM) {
+      scalar_vec_t x(input_p + j);
+      scalar_vec_t w(weight + j);
+
+      vec_op::FP32Vec8 fp32_x(x);
+      vec_op::FP32Vec8 fp32_w(w);
+
+      vec_op::FP32Vec8 fp32_out = fp32_x * fp32_s_variance * fp32_w;
+
+      scalar_vec_t out(fp32_out);
+      out.save(output_p + j);
+    }
+  }
+}
+
+template <typename scalar_t>
+void fused_add_rms_norm_impl(scalar_t* __restrict__ input,
+                             scalar_t* __restrict__ residual,
+                             const scalar_t* __restrict__ weight,
+                             const float epsilon, const int num_tokens,
+                             const int hidden_size) {
+  using scalar_vec_t = vec_op::vec_t<scalar_t>;
+  constexpr int VEC_ELEM_NUM = scalar_vec_t::get_elem_num();
+  TORCH_CHECK(hidden_size % VEC_ELEM_NUM == 0);
+
+#pragma omp parallel for
+  for (int i = 0; i < num_tokens; ++i) {
+    vec_op::FP32Vec8 variance(0.0);
+    auto input_p = input + i * hidden_size;
+    auto residual_p = residual + i * hidden_size;
+    for (int j = 0; j < hidden_size; j += VEC_ELEM_NUM) {
+      scalar_vec_t x(input_p + j);
+      scalar_vec_t res(residual_p + j);
+      vec_op::FP32Vec8 fp32_x(x);
+      vec_op::FP32Vec8 fp32_res(res);
+
+      fp32_x = fp32_x + fp32_res;
+      variance = variance + fp32_x * fp32_x;
+      scalar_vec_t out(fp32_x);
+      out.save(residual_p + j);
+    }
+
+    float s_variance =
+        1.0f / sqrtf(variance.reduce_sum() / (float)hidden_size + epsilon);
+    vec_op::FP32Vec8 fp32_s_variance(s_variance);
+
+    for (int j = 0; j < hidden_size; j += VEC_ELEM_NUM) {
+      scalar_vec_t w(weight + j);
+      scalar_vec_t res(residual_p + j);
+
+      vec_op::FP32Vec8 fp32_w(w);
+      vec_op::FP32Vec8 fp32_res(res);
+
+      vec_op::FP32Vec8 fp32_out = fp32_res * fp32_s_variance * fp32_w;
+
+      scalar_vec_t out(fp32_out);
+      out.save(input_p + j);
+    }
+  }
+}
+}  // namespace
+
+void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
+              float epsilon) {
+  int hidden_size = input.size(-1);
+  int num_tokens = input.numel() / hidden_size;
+
+  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "rms_norm_impl", [&] {
+    CPU_KERNEL_GUARD_IN(rms_norm_impl)
+    rms_norm_impl(out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),
+                  weight.data_ptr<scalar_t>(), epsilon, num_tokens,
+                  hidden_size);
+    CPU_KERNEL_GUARD_OUT(rms_norm_impl)
+  });
+}
+
+void fused_add_rms_norm(torch::Tensor& input, torch::Tensor& residual,
+                        torch::Tensor& weight, float epsilon) {
+  int hidden_size = input.size(-1);
+  int num_tokens = input.numel() / hidden_size;
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      input.scalar_type(), "fused_add_rms_norm_impl", [&] {
+        CPU_KERNEL_GUARD_IN(fused_add_rms_norm_impl)
+        fused_add_rms_norm_impl(
+            input.data_ptr<scalar_t>(), residual.data_ptr<scalar_t>(),
+            weight.data_ptr<scalar_t>(), epsilon, num_tokens, hidden_size);
+        CPU_KERNEL_GUARD_OUT(fused_add_rms_norm_impl)
+      });
+}

csrc/cpu/pos_encoding.cpp

@@ -0,0 +1,198 @@
+
+#include "cpu_types.hpp"
+
+namespace {
+template <typename scalar_t>
+void rotary_embedding_impl(
+    const int64_t* __restrict__ positions,  // [batch_size, seq_len] or
+                                            // [num_tokens]
+    scalar_t* __restrict__ query,           /// [batch_size, seq_len, num_heads,
+                                   /// head_size] or [num_tokens, num_heads,
+                                   /// head_size]
+    scalar_t* __restrict__ key,  // [batch_size, seq_len, num_kv_heads,
+                                 // head_size] or [num_tokens, num_kv_heads,
+                                 // head_size]
+    const scalar_t* __restrict__ cos_sin_cache,  // [max_position, 2, rot_dim //
+                                                 // 2]
+    const int rot_dim, const int64_t query_stride, const int64_t key_stride,
+    const int num_heads, const int num_kv_heads, const int head_size,
+    const int num_tokens) {
+  using scalar_vec_t = vec_op::vec_t<scalar_t>;
+  constexpr int VEC_ELEM_NUM = scalar_vec_t::get_elem_num();
+
+  const int embed_dim = rot_dim / 2;
+  TORCH_CHECK(embed_dim % VEC_ELEM_NUM == 0);
+
+#pragma omp parallel for
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    int64_t pos = positions[token_idx];
+    const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
+
+    for (int i = 0; i < num_heads; ++i) {
+      const int head_idx = i;
+      const int64_t token_head =
+          token_idx * query_stride + head_idx * head_size;
+      for (int j = 0; j < embed_dim; j += VEC_ELEM_NUM) {
+        const int rot_offset = j;
+        const int x_index = rot_offset;
+        const int y_index = embed_dim + rot_offset;
+
+        const int64_t out_x = token_head + x_index;
+        const int64_t out_y = token_head + y_index;
+
+        const scalar_vec_t cos(cache_ptr + x_index);
+        const scalar_vec_t sin(cache_ptr + y_index);
+
+        const scalar_vec_t q_x(query + out_x);
+        const scalar_vec_t q_y(query + out_y);
+
+        vec_op::FP32Vec8 fp32_cos(cos);
+        vec_op::FP32Vec8 fp32_sin(sin);
+
+        vec_op::FP32Vec8 fp32_q_x(q_x);
+        vec_op::FP32Vec8 fp32_q_y(q_y);
+
+        auto out1 = fp32_q_x * fp32_cos - fp32_q_y * fp32_sin;
+        scalar_vec_t(out1).save(query + out_x);
+
+        auto out2 = fp32_q_y * fp32_cos + fp32_q_x * fp32_sin;
+        scalar_vec_t(out2).save(query + out_y);
+      }
+    }
+
+    for (int i = 0; i < num_kv_heads; ++i) {
+      const int head_idx = i;
+      const int64_t token_head = token_idx * key_stride + head_idx * head_size;
+      for (int j = 0; j < embed_dim; j += VEC_ELEM_NUM) {
+        const int rot_offset = j;
+        const int x_index = rot_offset;
+        const int y_index = embed_dim + rot_offset;
+
+        const int64_t out_x = token_head + x_index;
+        const int64_t out_y = token_head + y_index;
+
+        const scalar_vec_t cos(cache_ptr + x_index);
+        const scalar_vec_t sin(cache_ptr + y_index);
+
+        const scalar_vec_t k_x(key + out_x);
+        const scalar_vec_t k_y(key + out_y);
+
+        vec_op::FP32Vec8 fp32_cos(cos);
+        vec_op::FP32Vec8 fp32_sin(sin);
+
+        vec_op::FP32Vec8 fp32_k_x(k_x);
+        vec_op::FP32Vec8 fp32_k_y(k_y);
+
+        auto out1 = fp32_k_x * fp32_cos - fp32_k_y * fp32_sin;
+        scalar_vec_t(out1).save(key + out_x);
+        auto out2 = fp32_k_y * fp32_cos + fp32_k_x * fp32_sin;
+        scalar_vec_t(out2).save(key + out_y);
+      }
+    }
+  }
+}
+
+template <typename scalar_t>
+void rotary_embedding_gptj_impl(
+    const int64_t* __restrict__ positions,  // [batch_size, seq_len] or
+                                            // [num_tokens]
+    scalar_t* __restrict__ query,           /// [batch_size, seq_len, num_heads,
+                                   /// head_size] or [num_tokens, num_heads,
+                                   /// head_size]
+    scalar_t* __restrict__ key,  // [batch_size, seq_len, num_kv_heads,
+                                 // head_size] or [num_tokens, num_kv_heads,
+                                 // head_size]
+    const scalar_t* __restrict__ cos_sin_cache,  // [max_position, 2, rot_dim //
+                                                 // 2]
+    const int rot_dim, const int64_t query_stride, const int64_t key_stride,
+    const int num_heads, const int num_kv_heads, const int head_size,
+    const int num_tokens) {
+  const int embed_dim = rot_dim / 2;
+
+#pragma omp parallel for collapse(2)
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    for (int i = 0; i < num_heads; ++i) {
+      int64_t pos = positions[token_idx];
+      const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
+      const scalar_t* cos_cache_ptr = cache_ptr;
+      const scalar_t* sin_cache_ptr = cache_ptr + embed_dim;
+      const int head_idx = i;
+      const int64_t token_head =
+          token_idx * query_stride + head_idx * head_size;
+      scalar_t* head_query = token_head + query;
+      for (int j = 0; j < embed_dim; j += 1) {
+        const int rot_offset = j;
+        const int x_index = 2 * rot_offset;
+        const int y_index = 2 * rot_offset + 1;
+
+        const float cos = cos_cache_ptr[rot_offset];
+        const float sin = sin_cache_ptr[rot_offset];
+
+        const float x = head_query[x_index];
+        const float y = head_query[y_index];
+
+        head_query[x_index] = x * cos - y * sin;
+        head_query[y_index] = y * cos + x * sin;
+      }
+    }
+  }
+
+#pragma omp parallel for collapse(2)
+  for (int token_idx = 0; token_idx < num_tokens; ++token_idx) {
+    for (int i = 0; i < num_kv_heads; ++i) {
+      int64_t pos = positions[token_idx];
+      const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
+      const scalar_t* cos_cache_ptr = cache_ptr;
+      const scalar_t* sin_cache_ptr = cache_ptr + embed_dim;
+      const int head_idx = i;
+      const int64_t token_head = token_idx * key_stride + head_idx * head_size;
+      scalar_t* head_key = key + token_head;
+      for (int j = 0; j < embed_dim; j += 1) {
+        const int rot_offset = j;
+        const int x_index = 2 * rot_offset;
+        const int y_index = 2 * rot_offset + 1;
+
+        const float cos = cos_cache_ptr[rot_offset];
+        const float sin = sin_cache_ptr[rot_offset];
+
+        const float x = head_key[x_index];
+        const float y = head_key[y_index];
+
+        head_key[x_index] = x * cos - y * sin;
+        head_key[y_index] = y * cos + x * sin;
+      }
+    }
+  }
+}
+};  // namespace
+
+void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
+                      torch::Tensor& key, int head_size,
+                      torch::Tensor& cos_sin_cache, bool is_neox) {
+  int num_tokens = query.numel() / query.size(-1);
+  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;
+  int64_t key_stride = key.stride(-2);
+  int64_t query_stride = query.stride(-2);
+
+  VLLM_DISPATCH_FLOATING_TYPES(
+      query.scalar_type(), "rotary_embedding_impl", [&] {
+        CPU_KERNEL_GUARD_IN(rotary_embedding_impl)
+        if (is_neox) {
+          rotary_embedding_impl(
+              positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
+              key.data_ptr<scalar_t>(), cos_sin_cache.data_ptr<scalar_t>(),
+              rot_dim, query_stride, key_stride, num_heads, num_kv_heads,
+              head_size, num_tokens);
+        } else {
+          rotary_embedding_gptj_impl(
+              positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
+              key.data_ptr<scalar_t>(), cos_sin_cache.data_ptr<scalar_t>(),
+              rot_dim, query_stride, key_stride, num_heads, num_kv_heads,
+              head_size, num_tokens);
+        }
+
+        CPU_KERNEL_GUARD_OUT(rotary_embedding_impl)
+      });
+}

csrc/cpu/pybind.cpp

@@ -0,0 +1,44 @@
+#include "cache.h"
+#include "cuda_utils.h"
+#include "ops.h"
+#include <torch/extension.h>
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  // vLLM custom ops
+  pybind11::module ops = m.def_submodule("ops", "vLLM custom operators");
+
+  // Attention ops
+  ops.def("paged_attention_v1", &paged_attention_v1,
+          "Compute the attention between an input query and the cached "
+          "keys/values using PagedAttention.");
+  ops.def("paged_attention_v2", &paged_attention_v2, "PagedAttention V2.");
+
+  // Activation ops
+  ops.def("silu_and_mul", &silu_and_mul, "Activation function used in SwiGLU.");
+  ops.def("gelu_and_mul", &gelu_and_mul,
+          "Activation function used in GeGLU with `none` approximation.");
+  ops.def("gelu_tanh_and_mul", &gelu_tanh_and_mul,
+          "Activation function used in GeGLU with `tanh` approximation.");
+  ops.def("gelu_new", &gelu_new, "GELU implementation used in GPT-2.");
+  ops.def("gelu_fast", &gelu_fast, "Approximate GELU implementation.");
+
+  // Layernorm
+  ops.def("rms_norm", &rms_norm,
+          "Apply Root Mean Square (RMS) Normalization to the input tensor.");
+
+  ops.def("fused_add_rms_norm", &fused_add_rms_norm,
+          "In-place fused Add and RMS Normalization");
+
+  // Rotary embedding
+  ops.def("rotary_embedding", &rotary_embedding,
+          "Apply GPT-NeoX or GPT-J style rotary embedding to query and key");
+
+  // Cache ops
+  pybind11::module cache_ops = m.def_submodule("cache_ops", "vLLM cache ops");
+  cache_ops.def("swap_blocks", &swap_blocks,
+                "Swap in (out) the cache blocks from src to dst");
+  cache_ops.def("copy_blocks", &copy_blocks,
+                "Copy the cache blocks from src to dst");
+  cache_ops.def("reshape_and_cache", &reshape_and_cache,
+                "Reshape the key and value tensors and cache them");
+}

csrc/cuda_compat.h

@@ -1,5 +1,15 @@
 #pragma once
 
+#ifdef USE_ROCM
+  #include <hip/hip_runtime.h>
+#endif
+
+#ifndef USE_ROCM
+  #define WARP_SIZE 32
+#else
+  #define WARP_SIZE warpSize
+#endif
+
 #ifndef USE_ROCM
   #define VLLM_LDG(arg) __ldg(arg)
 #else
@@ -7,7 +17,8 @@
 #endif
 
 #ifndef USE_ROCM
-  #define VLLM_SHFL_XOR_SYNC(var, lane_mask) __shfl_xor_sync(uint32_t(-1), var, lane_mask)
+  #define VLLM_SHFL_XOR_SYNC(var, lane_mask) \
+    __shfl_xor_sync(uint32_t(-1), var, lane_mask)
 #else
   #define VLLM_SHFL_XOR_SYNC(var, lane_mask) __shfl_xor(var, lane_mask)
 #endif
@@ -18,6 +29,13 @@
   #define VLLM_SHFL_SYNC(var, src_lane) __shfl(var, src_lane)
 #endif
 
+#ifndef USE_ROCM
+  #define VLLM_SHFL_DOWN_SYNC(var, lane_delta) \
+    __shfl_down_sync(uint32_t(-1), var, lane_delta)
+#else
+  #define VLLM_SHFL_DOWN_SYNC(var, lane_delta) __shfl_down(var, lane_delta)
+#endif
+
 #ifndef USE_ROCM
   #define VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(FUNC, VAL) \
     cudaFuncSetAttribute(FUNC, cudaFuncAttributeMaxDynamicSharedMemorySize, VAL)
@@ -25,4 +43,3 @@
   #define VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(FUNC, VAL) \
     hipFuncSetAttribute(FUNC, hipFuncAttributeMaxDynamicSharedMemorySize, VAL)
 #endif
-

csrc/cuda_utils.h

@@ -2,9 +2,6 @@
 
 #include <torch/extension.h>
 
-int get_device_attribute(
-    int attribute,
-    int device_id);
+int get_device_attribute(int attribute, int device_id);
 
-int get_max_shared_memory_per_block_device_attribute(
-    int device_id);
+int get_max_shared_memory_per_block_device_attribute(int device_id);

csrc/cuda_utils_kernels.cu

@@ -2,34 +2,28 @@
   #include <hip/hip_runtime.h>
   #include <hip/hip_runtime_api.h>
 #endif
-int get_device_attribute(
-    int attribute,
-    int device_id)
-{
-    int device, value;
-    if (device_id < 0) {
-        cudaGetDevice(&device);
-    }
-    else {
-        device = device_id;
-    }
-    cudaDeviceGetAttribute(&value, static_cast<cudaDeviceAttr>(attribute), device);
-    return value;
+int get_device_attribute(int attribute, int device_id) {
+  int device, value;
+  if (device_id < 0) {
+    cudaGetDevice(&device);
+  } else {
+    device = device_id;
+  }
+  cudaDeviceGetAttribute(&value, static_cast<cudaDeviceAttr>(attribute),
+                         device);
+  return value;
 }
 
-
-int get_max_shared_memory_per_block_device_attribute(
-    int device_id)
-{
-int attribute;    
-// https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html
-// cudaDevAttrMaxSharedMemoryPerBlockOptin = 97 if not is_hip() else 74
+int get_max_shared_memory_per_block_device_attribute(int device_id) {
+  int attribute;
+  // https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html
+  // cudaDevAttrMaxSharedMemoryPerBlockOptin = 97 if not is_hip() else 74
 
 #ifdef USE_ROCM
-    attribute = hipDeviceAttributeMaxSharedMemoryPerBlock;
+  attribute = hipDeviceAttributeMaxSharedMemoryPerBlock;
 #else
-    attribute = cudaDevAttrMaxSharedMemoryPerBlockOptin;
+  attribute = cudaDevAttrMaxSharedMemoryPerBlockOptin;
 #endif
 
-    return get_device_attribute(attribute, device_id);
+  return get_device_attribute(attribute, device_id);
 }

csrc/custom_all_reduce.cu

@@ -7,11 +7,11 @@
 
 // fake pointer type
 using fptr_t = uint64_t;
-static_assert(sizeof(void *) == sizeof(fptr_t));
+static_assert(sizeof(void*) == sizeof(fptr_t));
 
-fptr_t init_custom_ar(torch::Tensor &meta, torch::Tensor &rank_data,
-                      const std::vector<std::string> &handles,
-                      const std::vector<int64_t> &offsets, int rank,
+fptr_t init_custom_ar(torch::Tensor& meta, torch::Tensor& rank_data,
+                      const std::vector<std::string>& handles,
+                      const std::vector<int64_t>& offsets, int rank,
                       bool full_nvlink) {
   int world_size = offsets.size();
   if (world_size > 8)
@@ -29,7 +29,7 @@ fptr_t init_custom_ar(torch::Tensor &meta, torch::Tensor &rank_data,
     std::memcpy(&ipc_handles[i], handles[i].data(), sizeof(cudaIpcMemHandle_t));
   }
   return (fptr_t) new vllm::CustomAllreduce(
-      reinterpret_cast<vllm::Metadata *>(meta.data_ptr()), rank_data.data_ptr(),
+      reinterpret_cast<vllm::Signal*>(meta.data_ptr()), rank_data.data_ptr(),
       rank_data.numel(), ipc_handles, offsets, rank, full_nvlink);
 }
 
@@ -49,46 +49,45 @@ fptr_t init_custom_ar(torch::Tensor &meta, torch::Tensor &rank_data,
  * 5. A[None].expand(2, -1, -1, -1): Not OK
  * 6. A[:, 1:, 1:]: Not OK
  */
-bool _is_weak_contiguous(torch::Tensor &t) {
+bool _is_weak_contiguous(torch::Tensor& t) {
   return t.is_contiguous() ||
          (t.storage().nbytes() - t.storage_offset() * t.element_size() ==
           t.numel() * t.element_size());
 }
 
-bool should_custom_ar(torch::Tensor &inp, int max_size, int world_size,
+bool should_custom_ar(torch::Tensor& inp, int max_size, int world_size,
                       bool full_nvlink) {
   auto inp_size = inp.numel() * inp.element_size();
   // custom allreduce requires input byte size to be multiples of 16
   if (inp_size % 16 != 0) return false;
   if (!_is_weak_contiguous(inp)) return false;
   if (world_size == 2 || full_nvlink) return inp_size <= max_size;
-  // 4 PCIE GPUs use 2 stage allreduce, and is only faster than NCCL when size
-  // <= 512k
-  return world_size <= 4 && inp_size <= 512 * 1024;
+  // for 4 or more non NVLink-capable GPUs, custom allreduce provides little
+  // performance improvement over NCCL.
+  return false;
 }
 
-void _all_reduce(fptr_t _fa, torch::Tensor &inp, torch::Tensor &out,
+void _all_reduce(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out,
                  cudaStream_t stream) {
-  auto fa = reinterpret_cast<vllm::CustomAllreduce *>(_fa);
+  auto fa = reinterpret_cast<vllm::CustomAllreduce*>(_fa);
   TORCH_CHECK(_is_weak_contiguous(out));
   switch (out.scalar_type()) {
     case at::ScalarType::Float: {
-      fa->allreduce<float>(stream, reinterpret_cast<float *>(inp.data_ptr()),
-                           reinterpret_cast<float *>(out.data_ptr()),
+      fa->allreduce<float>(stream, reinterpret_cast<float*>(inp.data_ptr()),
+                           reinterpret_cast<float*>(out.data_ptr()),
                            out.numel());
       break;
     }
     case at::ScalarType::Half: {
-      fa->allreduce<half>(stream, reinterpret_cast<half *>(inp.data_ptr()),
-                          reinterpret_cast<half *>(out.data_ptr()),
-                          out.numel());
+      fa->allreduce<half>(stream, reinterpret_cast<half*>(inp.data_ptr()),
+                          reinterpret_cast<half*>(out.data_ptr()), out.numel());
       break;
     }
 #if (__CUDA_ARCH__ >= 800 || !defined(__CUDA_ARCH__))
     case at::ScalarType::BFloat16: {
       fa->allreduce<nv_bfloat16>(
-          stream, reinterpret_cast<nv_bfloat16 *>(inp.data_ptr()),
-          reinterpret_cast<nv_bfloat16 *>(out.data_ptr()), out.numel());
+          stream, reinterpret_cast<nv_bfloat16*>(inp.data_ptr()),
+          reinterpret_cast<nv_bfloat16*>(out.data_ptr()), out.numel());
       break;
     }
 #endif
@@ -98,7 +97,7 @@ void _all_reduce(fptr_t _fa, torch::Tensor &inp, torch::Tensor &out,
   }
 }
 
-void all_reduce_reg(fptr_t _fa, torch::Tensor &inp, torch::Tensor &out) {
+void all_reduce_reg(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out) {
   const at::cuda::OptionalCUDAGuard device_guard(device_of(inp));
   auto stream = c10::cuda::getCurrentCUDAStream().stream();
   TORCH_CHECK_EQ(inp.scalar_type(), out.scalar_type());
@@ -106,8 +105,8 @@ void all_reduce_reg(fptr_t _fa, torch::Tensor &inp, torch::Tensor &out) {
   _all_reduce(_fa, inp, out, stream);
 }
 
-void all_reduce_unreg(fptr_t _fa, torch::Tensor &inp, torch::Tensor &reg_buffer,
-                      torch::Tensor &out) {
+void all_reduce_unreg(fptr_t _fa, torch::Tensor& inp, torch::Tensor& reg_buffer,
+                      torch::Tensor& out) {
   const at::cuda::OptionalCUDAGuard device_guard(device_of(inp));
   auto stream = c10::cuda::getCurrentCUDAStream().stream();
 
@@ -122,27 +121,27 @@ void all_reduce_unreg(fptr_t _fa, torch::Tensor &inp, torch::Tensor &reg_buffer,
 }
 
 void dispose(fptr_t _fa) {
-  auto fa = reinterpret_cast<vllm::CustomAllreduce *>(_fa);
+  auto fa = reinterpret_cast<vllm::CustomAllreduce*>(_fa);
   delete fa;
 }
 
-int meta_size() { return sizeof(vllm::Metadata); }
+int meta_size() { return sizeof(vllm::Signal); }
 
-void register_buffer(fptr_t _fa, torch::Tensor &t,
-                     const std::vector<std::string> &handles,
-                     const std::vector<int64_t> &offsets) {
-  auto fa = reinterpret_cast<vllm::CustomAllreduce *>(_fa);
+void register_buffer(fptr_t _fa, torch::Tensor& t,
+                     const std::vector<std::string>& handles,
+                     const std::vector<int64_t>& offsets) {
+  auto fa = reinterpret_cast<vllm::CustomAllreduce*>(_fa);
   fa->register_buffer(handles, offsets, t.data_ptr());
 }
 
 std::pair<std::vector<uint8_t>, std::vector<int64_t>> get_graph_buffer_ipc_meta(
     fptr_t _fa) {
-  auto fa = reinterpret_cast<vllm::CustomAllreduce *>(_fa);
+  auto fa = reinterpret_cast<vllm::CustomAllreduce*>(_fa);
   return fa->get_graph_buffer_ipc_meta();
 }
 
-void register_graph_buffers(fptr_t _fa, const std::vector<std::string> &handles,
-                            const std::vector<std::vector<int64_t>> &offsets) {
-  auto fa = reinterpret_cast<vllm::CustomAllreduce *>(_fa);
+void register_graph_buffers(fptr_t _fa, const std::vector<std::string>& handles,
+                            const std::vector<std::vector<int64_t>>& offsets) {
+  auto fa = reinterpret_cast<vllm::CustomAllreduce*>(_fa);
   fa->register_graph_buffers(handles, offsets);
 }

csrc/custom_all_reduce.cuh

@@ -23,29 +23,17 @@
 
 namespace vllm {
 
+constexpr int kMaxBlocks = 64;
+// note: we don't want to use atomics for signals because peer atomics are no
+// supported on PCIe links
 struct Signal {
-  alignas(64) union {
-    uint64_t flag;
-    unsigned char data[8];
-  } start;
-  alignas(64) union {
-    uint64_t flag;
-    unsigned char data[8];
-  } end;
+  alignas(128) uint32_t start[kMaxBlocks][8];
+  alignas(128) uint32_t end[kMaxBlocks][8];
 };
 
-struct Metadata {
-  alignas(128) Signal sg;
-  alignas(128) int counter;
-};
-static_assert(offsetof(Metadata, counter) == 128);
-static_assert(sizeof(Metadata) == 256);
+struct __align__(16) RankData { const void* __restrict__ ptrs[8]; };
 
-struct __align__(16) RankData { const void *__restrict__ ptrs[8]; };
-
-struct RankSignals {
-  volatile Signal *signals[8];
-};
+struct __align__(16) RankSignals { volatile Signal* signals[8]; };
 
 // like std::array, but aligned
 template <typename T, int sz>
@@ -80,11 +68,11 @@ DINLINE half downcast_s(float val) {
 // scalar add functions
 // for some reason when compiling with Pytorch, the + operator for half and
 // bfloat is disabled so we call the intrinsics directly
-DINLINE half &assign_add(half &a, half b) {
+DINLINE half& assign_add(half& a, half b) {
   a = __hadd(a, b);
   return a;
 }
-DINLINE float &assign_add(float &a, float b) { return a += b; }
+DINLINE float& assign_add(float& a, float b) { return a += b; }
 
 #if (__CUDA_ARCH__ >= 800 || !defined(__CUDA_ARCH__))
 DINLINE float upcast_s(nv_bfloat16 val) { return __bfloat162float(val); }
@@ -92,14 +80,14 @@ template <>
 DINLINE nv_bfloat16 downcast_s(float val) {
   return __float2bfloat16(val);
 }
-DINLINE nv_bfloat16 &assign_add(nv_bfloat16 &a, nv_bfloat16 b) {
+DINLINE nv_bfloat16& assign_add(nv_bfloat16& a, nv_bfloat16 b) {
   a = __hadd(a, b);
   return a;
 }
 #endif
 
 template <typename T, int N>
-DINLINE array_t<T, N> &packed_assign_add(array_t<T, N> &a, array_t<T, N> b) {
+DINLINE array_t<T, N>& packed_assign_add(array_t<T, N>& a, array_t<T, N> b) {
 #pragma unroll
   for (int i = 0; i < N; i++) {
     assign_add(a.data[i], b.data[i]);
@@ -135,74 +123,51 @@ DINLINE O downcast(array_t<float, O::size> val) {
   }
 }
 
-// compute flag at compile time
-__host__ __device__ constexpr uint64_t compute_flag(int ngpus) {
-  auto m = std::numeric_limits<uint64_t>::max();
-  return m >> ((8 - ngpus) * 8);
-}
-
+// This function is meant to be used as the first synchronization in the all
+// reduce kernel. Thus, it doesn't need to make any visibility guarantees for
+// prior memory accesses. Note: volatile writes will not be reordered against
+// other volatile writes.
 template <int ngpus>
-DINLINE void start_sync(const RankSignals &sg, volatile Metadata *meta,
+DINLINE void start_sync(const RankSignals& sg, volatile Signal* self_sg,
                         int rank) {
-  constexpr auto FLAG = compute_flag(ngpus);
-  if (blockIdx.x == 0) {
-    if (threadIdx.x < ngpus)
-      // simultaneously write to the corresponding byte to all other ranks.
-      // Latency = 1 p2p write
-      sg.signals[threadIdx.x]->start.data[rank] = 255;
-    else if (threadIdx.x == 32)
-      // reset
-      meta->sg.end.flag = 0;
-  }
-  if (threadIdx.x == 0) {
-    while (meta->sg.start.flag != FLAG)
-      ;
+  if (threadIdx.x < ngpus) {
+    // reset flag for next time
+    self_sg->end[blockIdx.x][threadIdx.x] = 0;
+    // simultaneously write to the corresponding flag of all ranks.
+    // Latency = 1 p2p write
+    sg.signals[threadIdx.x]->start[blockIdx.x][rank] = 1;
+    // wait until we got true from all ranks
+    while (!self_sg->start[blockIdx.x][threadIdx.x]);
   }
   __syncthreads();
 }
 
+// This function is meant to be used as the second or the final synchronization
+// barrier in the all reduce kernel. If it's the final synchronization barrier,
+// we don't need to make any visibility guarantees for prior memory accesses.
 template <int ngpus, bool final_sync = false>
-DINLINE void end_sync(const RankSignals &sg, volatile Metadata *meta,
+DINLINE void end_sync(const RankSignals& sg, volatile Signal* self_sg,
                       int rank) {
-  constexpr auto FLAG = compute_flag(ngpus);
   __syncthreads();
-  __shared__ int num;
-  if (threadIdx.x == 0) num = atomicAdd((int *)&meta->counter, 1);
-  __syncthreads();
-
-  // Only the last completing block can perform the end synchronization
-  // This can ensures when the final busy wait ends, all ranks must have
-  // finished reading each other's buffer.
-  if (num == gridDim.x - 1) {
-    if (threadIdx.x == 32) {
-      // reset in a different warp
-      meta->counter = 0;
-      meta->sg.start.flag = 0;
-    } else if (threadIdx.x < ngpus) {
-      // simultaneously write to the corresponding byte to all other ranks.
-      // Latency = 1 p2p write
-      sg.signals[threadIdx.x]->end.data[rank] = 255;
-    }
-    // if this is the final sync, only one block needs it
-    // because kernel exit can serve as sync
-    if constexpr (final_sync) {
-      if (threadIdx.x == 0) {
-        while (meta->sg.end.flag != FLAG)
-          ;
-      }
-    }
-  }
-  if constexpr (!final_sync) {
-    if (threadIdx.x == 0) {
-      while (meta->sg.end.flag != FLAG)
-        ;
-    }
-    __syncthreads();
+  // eliminate the case that prior writes are not visible after signals become
+  // visible. Note that I did not managed to make this happen through a lot of
+  // testing. Might be the case that hardware provides stronger guarantee than
+  // the memory model.
+  if constexpr (!final_sync) __threadfence_system();
+  if (threadIdx.x < ngpus) {
+    // reset flag for next time
+    self_sg->start[blockIdx.x][threadIdx.x] = 0;
+    // simultaneously write to the corresponding flag of all ranks.
+    // Latency = 1 p2p write
+    sg.signals[threadIdx.x]->end[blockIdx.x][rank] = 1;
+    // wait until we got true from all ranks
+    while (!self_sg->end[blockIdx.x][threadIdx.x]);
   }
+  if constexpr (!final_sync) __syncthreads();
 }
 
 template <typename P, int ngpus, typename A>
-DINLINE P packed_reduce(const P *ptrs[], int idx) {
+DINLINE P packed_reduce(const P* ptrs[], int idx) {
   A tmp = upcast(ptrs[0][idx]);
 #pragma unroll
   for (int i = 1; i < ngpus; i++) {
@@ -213,33 +178,32 @@ DINLINE P packed_reduce(const P *ptrs[], int idx) {
 
 template <typename T, int ngpus>
 __global__ void __launch_bounds__(512, 1)
-    cross_device_reduce_1stage(RankData *_dp, RankSignals sg,
-                               volatile Metadata *meta, T *__restrict__ result,
+    cross_device_reduce_1stage(RankData* _dp, RankSignals sg,
+                               volatile Signal* self_sg, T* __restrict__ result,
                                int rank, int size) {
   using P = typename packed_t<T>::P;
   using A = typename packed_t<T>::A;
   // note: we don't reorder the address so the accumulation order is the same
   // for all ranks, ensuring bitwise identical results
   auto dp = *_dp;
-  start_sync<ngpus>(sg, meta, rank);
+  start_sync<ngpus>(sg, self_sg, rank);
   // do the actual reduction
   for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
        idx += gridDim.x * blockDim.x) {
-    ((P *)result)[idx] =
-        packed_reduce<P, ngpus, A>((const P **)&dp.ptrs[0], idx);
+    ((P*)result)[idx] = packed_reduce<P, ngpus, A>((const P**)&dp.ptrs[0], idx);
   }
-  end_sync<ngpus, true>(sg, meta, rank);
+  end_sync<ngpus, true>(sg, self_sg, rank);
 }
 
 template <typename P>
-DINLINE P *get_tmp_buf(volatile Signal *sg) {
-  return (P *)(((Metadata *)sg) + 1);
+DINLINE P* get_tmp_buf(volatile Signal* sg) {
+  return (P*)(((Signal*)sg) + 1);
 }
 
 template <typename T, int ngpus>
 __global__ void __launch_bounds__(512, 1)
-    cross_device_reduce_2stage(RankData *_dp, RankSignals sg,
-                               volatile Metadata *meta, T *__restrict__ result,
+    cross_device_reduce_2stage(RankData* _dp, RankSignals sg,
+                               volatile Signal* self_sg, T* __restrict__ result,
                                int rank, int size) {
   int tid = blockIdx.x * blockDim.x + threadIdx.x;
   int stride = gridDim.x * blockDim.x;
@@ -248,84 +212,38 @@ __global__ void __launch_bounds__(512, 1)
   int part = size / ngpus;
   int start = rank * part;
   int end = rank == ngpus - 1 ? size : start + part;
-  const P *ptrs[ngpus];
-  P *tmps[ngpus];
+  int largest_part = part + size % ngpus;
+  const P* ptrs[ngpus];
+  P* tmps[ngpus];
 #pragma unroll
   for (int i = 0; i < ngpus; i++) {
     int target = (rank + i) % ngpus;
-    ptrs[i] = (const P *)_dp->ptrs[target];
+    ptrs[i] = (const P*)_dp->ptrs[target];
     tmps[i] = get_tmp_buf<P>(sg.signals[target]);
   }
   auto tmp_out = tmps[0];
-  start_sync<ngpus>(sg, meta, rank);
+  start_sync<ngpus>(sg, self_sg, rank);
   // stage 1: reduce scatter
   for (int idx = start + tid; idx < end; idx += stride) {
     tmp_out[idx - start] = packed_reduce<P, ngpus, A>(ptrs, idx);
   }
-  // Maybe TODO: replace this with per-block release-acquire
-  // can save about 1-2us (not a lot though)
-  end_sync<ngpus>(sg, meta, rank);
+  end_sync<ngpus>(sg, self_sg, rank);
 
-  // stage 2: allgather
-  for (int idx = tid; idx < part; idx += stride) {
+  // stage 2: allgather. Note: it's important to match the tid between
+  // the two stages, because visibility across devices is only guaranteed
+  // between threads that have the same tid. If thread i computes the sum of
+  // start + i in the first stage, then thread i also gathers start + i from all
+  // ranks.
+  for (int idx = tid; idx < largest_part; idx += stride) {
 #pragma unroll
     for (int i = 0; i < ngpus; i++) {
-      int dst_idx = ((rank + i) % ngpus) * part + idx;
-      ((P *)result)[dst_idx] = tmps[i][idx];
+      int gather_from_rank = ((rank + i) % ngpus);
+      if (gather_from_rank == ngpus - 1 || idx < part) {
+        int dst_idx = gather_from_rank * part + idx;
+        ((P*)result)[dst_idx] = tmps[i][idx];
+      }
     }
   }
-  // process the last larger partition
-  int remaining = size - part * ngpus;
-  if (tid < remaining) {
-    int dst_idx = tid + part * ngpus;
-    ((P *)result)[dst_idx] = get_tmp_buf<P>(sg.signals[ngpus - 1])[part + tid];
-  }
-
-  // faster than this
-  // for (int idx = tid; idx < size; idx += stride) {
-  //   int target_rank = idx / part;
-  //   if (target_rank == ngpus) target_rank -= 1;
-  //   ((P *)result)[idx] = tmps[target_rank][idx - target_rank * part];
-  // }
-}
-
-template <typename T, int ngpus>
-__global__ void __launch_bounds__(512, 1)
-    cross_device_reduce_half_butterfly(RankData *_dp, RankSignals sg,
-                                       volatile Metadata *meta,
-                                       T *__restrict__ result, int rank,
-                                       int size) {
-  int tid = blockIdx.x * blockDim.x + threadIdx.x;
-  int stride = gridDim.x * blockDim.x;
-  using P = typename packed_t<T>::P;
-  using A = typename packed_t<T>::A;
-  auto tmp_out = get_tmp_buf<P>(sg.signals[rank]);
-  constexpr int hg = ngpus / 2;
-  // Actually not quite half butterfly.
-  // This is an all-to-all within each group containing half of the ranks
-  // followed by cross-group add. Equivalent to half butterfly when there
-  // are 4 GPUs, a common case for PCIe cards like T4 and A10.
-  const P *ptrs[hg];
-  {
-    int start = rank - rank % hg;
-#pragma unroll
-    for (int i = 0; i < hg; i++) {
-      ptrs[i] = (const P *)_dp->ptrs[i + start];
-    }
-  }
-  start_sync<ngpus>(sg, meta, rank);
-  for (int idx = tid; idx < size; idx += stride) {
-    tmp_out[idx] = packed_reduce<P, hg, A>(ptrs, idx);
-  }
-  end_sync<ngpus>(sg, meta, rank);
-
-  auto src = get_tmp_buf<P>(sg.signals[(ngpus - 1) - rank % ngpus]);
-  // do the cross group reduction
-  for (int idx = tid; idx < size; idx += stride) {
-    auto tmp = tmp_out[idx];
-    packed_assign_add(tmp, src[idx]);
-    ((P *)result)[idx] = tmp;
-  }
 }
 
 using IPC_KEY = std::array<uint8_t, sizeof(cudaIpcMemHandle_t)>;
@@ -340,54 +258,54 @@ class CustomAllreduce {
 
   // below are device pointers
   RankSignals sg_;
-  std::unordered_map<void *, RankData *> buffers_;
-  Metadata *meta_;
+  std::unordered_map<void*, RankData*> buffers_;
+  Signal* self_sg_;
 
   // stores the registered device pointers from all ranks
   RankData *d_rank_data_base_, *d_rank_data_end_;
-  std::vector<void *> graph_unreg_buffers_;
+  std::vector<void*> graph_unreg_buffers_;
   // a map from IPC handles to opened IPC pointers
-  std::map<IPC_KEY, char *> ipc_handles_;
+  std::map<IPC_KEY, char*> ipc_handles_;
 
   /**
    * meta is a pointer to device metadata and temporary buffer for allreduce.
    *
-   * There's a total of sizeof(Metadata) of prefix before the actual data,
+   * There's a total of sizeof(Signal) of prefix before the actual data,
    * so meta + 1 points to actual temporary buffer.
    *
    * note: this class does not own any device memory. Any required buffers
    * are passed in from the constructor
    */
-  CustomAllreduce(Metadata *meta, void *rank_data, size_t rank_data_sz,
-                  const cudaIpcMemHandle_t *handles,
-                  const std::vector<int64_t> &offsets, int rank,
+  CustomAllreduce(Signal* meta, void* rank_data, size_t rank_data_sz,
+                  const cudaIpcMemHandle_t* handles,
+                  const std::vector<int64_t>& offsets, int rank,
                   bool full_nvlink = true)
       : rank_(rank),
         world_size_(offsets.size()),
         full_nvlink_(full_nvlink),
-        meta_(meta),
-        d_rank_data_base_(reinterpret_cast<RankData *>(rank_data)),
+        self_sg_(meta),
+        d_rank_data_base_(reinterpret_cast<RankData*>(rank_data)),
         d_rank_data_end_(d_rank_data_base_ + rank_data_sz / sizeof(RankData)) {
     for (int i = 0; i < world_size_; i++) {
-      Metadata *rank_meta;
+      Signal* rank_sg;
       if (i != rank_) {
-        char *handle = open_ipc_handle(&handles[i]);
+        char* handle = open_ipc_handle(&handles[i]);
         handle += offsets[i];
-        rank_meta = (Metadata *)handle;
+        rank_sg = (Signal*)handle;
       } else {
-        rank_meta = meta_;
+        rank_sg = self_sg_;
       }
-      sg_.signals[i] = &rank_meta->sg;
+      sg_.signals[i] = rank_sg;
     }
   }
 
-  char *open_ipc_handle(const void *ipc_handle) {
+  char* open_ipc_handle(const void* ipc_handle) {
     auto [it, new_handle] =
-        ipc_handles_.insert({*((IPC_KEY *)ipc_handle), nullptr});
+        ipc_handles_.insert({*((IPC_KEY*)ipc_handle), nullptr});
     if (new_handle) {
-      char *ipc_ptr;
-      CUDACHECK(cudaIpcOpenMemHandle((void **)&ipc_ptr,
-                                     *((const cudaIpcMemHandle_t *)ipc_handle),
+      char* ipc_ptr;
+      CUDACHECK(cudaIpcOpenMemHandle((void**)&ipc_ptr,
+                                     *((const cudaIpcMemHandle_t*)ipc_handle),
                                      cudaIpcMemLazyEnablePeerAccess));
       it->second = ipc_ptr;
     }
@@ -402,7 +320,7 @@ class CustomAllreduce {
     std::vector<int64_t> offsets(num_buffers);
     for (int i = 0; i < num_buffers; i++) {
       auto ptr = graph_unreg_buffers_[i];
-      void *base_ptr;
+      void* base_ptr;
       // note: must share the base address of each allocation, or we get wrong
       // address
       if (cuPointerGetAttribute(&base_ptr,
@@ -410,8 +328,8 @@ class CustomAllreduce {
                                 (CUdeviceptr)ptr) != CUDA_SUCCESS)
         throw std::runtime_error("failed to get pointer attr");
       CUDACHECK(cudaIpcGetMemHandle(
-          (cudaIpcMemHandle_t *)&handles[i * handle_sz], base_ptr));
-      offsets[i] = ((char *)ptr) - ((char *)base_ptr);
+          (cudaIpcMemHandle_t*)&handles[i * handle_sz], base_ptr));
+      offsets[i] = ((char*)ptr) - ((char*)base_ptr);
     }
     return std::make_pair(handles, offsets);
   }
@@ -423,13 +341,13 @@ class CustomAllreduce {
           std::to_string(d_rank_data_base_ + num - d_rank_data_end_));
   }
 
-  void register_buffer(const std::vector<std::string> &handles,
-                       const std::vector<int64_t> &offsets, void *self) {
+  void register_buffer(const std::vector<std::string>& handles,
+                       const std::vector<int64_t>& offsets, void* self) {
     check_rank_data_capacity();
     RankData data;
     for (int i = 0; i < world_size_; i++) {
       if (i != rank_) {
-        char *handle = open_ipc_handle(handles[i].data());
+        char* handle = open_ipc_handle(handles[i].data());
         handle += offsets[i];
         data.ptrs[i] = handle;
       } else {
@@ -450,17 +368,17 @@ class CustomAllreduce {
   // got a different address. IPC handles have internal reference counting
   // mechanism so overhead should be small.
   void register_graph_buffers(
-      const std::vector<std::string> &handles,
-      const std::vector<std::vector<int64_t>> &offsets) {
+      const std::vector<std::string>& handles,
+      const std::vector<std::vector<int64_t>>& offsets) {
     auto num_buffers = graph_unreg_buffers_.size();
     check_rank_data_capacity(num_buffers);
     std::vector<RankData> rank_data(num_buffers);
     for (int i = 0; i < num_buffers; i++) {
       auto self_ptr = graph_unreg_buffers_[i];
-      auto &rd = rank_data[i];
+      auto& rd = rank_data[i];
       for (int j = 0; j < world_size_; j++) {
         if (j != rank_) {
-          char *handle =
+          char* handle =
               open_ipc_handle(&handles[j][i * sizeof(cudaIpcMemHandle_t)]);
           handle += offsets[j][i];
           rd.ptrs[j] = handle;
@@ -484,7 +402,7 @@ class CustomAllreduce {
    * will cause contention on NVLink bus.
    */
   template <typename T>
-  void allreduce(cudaStream_t stream, T *input, T *output, int size,
+  void allreduce(cudaStream_t stream, T* input, T* output, int size,
                  int threads = 512, int block_limit = 36) {
     auto d = packed_t<T>::P::size;
     if (size % d != 0)
@@ -492,8 +410,12 @@ class CustomAllreduce {
           "custom allreduce currently requires input length to be multiple "
           "of " +
           std::to_string(d));
+    if (block_limit > kMaxBlocks)
+      throw std::runtime_error("max supported block limit is " +
+                               std::to_string(kMaxBlocks) + ". Got " +
+                               std::to_string(block_limit));
 
-    RankData *ptrs;
+    RankData* ptrs;
     cudaStreamCaptureStatus status;
     CUDACHECK(cudaStreamIsCapturing(stream, &status));
     if (status == cudaStreamCaptureStatusActive) {
@@ -512,9 +434,9 @@ class CustomAllreduce {
     size /= d;
     auto bytes = size * sizeof(typename packed_t<T>::P);
     int blocks = std::min(block_limit, (size + threads - 1) / threads);
-#define KL(ngpus, name) \
-  name<T, ngpus>        \
-      <<<blocks, threads, 0, stream>>>(ptrs, sg_, meta_, output, rank_, size);
+#define KL(ngpus, name)                                                       \
+  name<T, ngpus><<<blocks, threads, 0, stream>>>(ptrs, sg_, self_sg_, output, \
+                                                 rank_, size);
 #define REDUCE_CASE(ngpus)                            \
   case ngpus: {                                       \
     if (world_size_ == 2) {                           \
@@ -526,8 +448,6 @@ class CustomAllreduce {
       } else {                                        \
         KL(ngpus, cross_device_reduce_2stage);        \
       }                                               \
-    } else {                                          \
-      KL(ngpus, cross_device_reduce_half_butterfly);  \
     }                                                 \
     break;                                            \
   }
@@ -556,7 +476,7 @@ class CustomAllreduce {
 /**
  * To inspect PTX/SASS, copy paste this header file to compiler explorer and add
  a template instantiation:
- * template void CustomAllreduce::allreduce<half>(cudaStream_t, half *, half *,
- int, int, int);
+ * template void vllm::CustomAllreduce::allreduce<half>(cudaStream_t, half *,
+ half *, int, int, int);
 */
 }  // namespace vllm

csrc/custom_all_reduce_test.cu

@@ -48,7 +48,7 @@ __global__ void dummy_kernel() {
 }
 
 template <typename T>
-__global__ void set_data(T *data, int size, int myRank) {
+__global__ void set_data(T* data, int size, int myRank) {
   for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
        idx += gridDim.x * blockDim.x) {
     data[idx] = myRank * 0.11f;
@@ -56,8 +56,8 @@ __global__ void set_data(T *data, int size, int myRank) {
 }
 
 template <typename T>
-__global__ void convert_data(const T *data1, const T *data2, double *fdata1,
-                             double *fdata2, int size) {
+__global__ void convert_data(const T* data1, const T* data2, double* fdata1,
+                             double* fdata2, int size) {
   for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
        idx += gridDim.x * blockDim.x) {
     fdata1[idx] = data1[idx];
@@ -65,7 +65,7 @@ __global__ void convert_data(const T *data1, const T *data2, double *fdata1,
   }
 }
 
-__global__ void init_rand(curandState_t *state, int size, int nRanks) {
+__global__ void init_rand(curandState_t* state, int size, int nRanks) {
   for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
        idx += gridDim.x * blockDim.x) {
     for (int i = 0; i < nRanks; i++) {
@@ -75,7 +75,7 @@ __global__ void init_rand(curandState_t *state, int size, int nRanks) {
 }
 
 template <typename T>
-__global__ void gen_data(curandState_t *state, T *data, double *ground_truth,
+__global__ void gen_data(curandState_t* state, T* data, double* ground_truth,
                          int myRank, int nRanks, int size) {
   for (int idx = blockIdx.x * blockDim.x + threadIdx.x; idx < size;
        idx += gridDim.x * blockDim.x) {
@@ -91,9 +91,9 @@ __global__ void gen_data(curandState_t *state, T *data, double *ground_truth,
 }
 
 template <typename T>
-void run(int myRank, int nRanks, ncclComm_t &comm, int threads, int block_limit,
-         int data_size) {
-  T *result;
+void run(int myRank, int nRanks, ncclComm_t& comm, int threads, int block_limit,
+         int data_size, bool performance_test) {
+  T* result;
   cudaStream_t stream;
   CUDACHECK(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
   CUDACHECK(cudaMalloc(&result, data_size * sizeof(T)));
@@ -101,8 +101,8 @@ void run(int myRank, int nRanks, ncclComm_t &comm, int threads, int block_limit,
 
   cudaIpcMemHandle_t self_data_handle;
   cudaIpcMemHandle_t data_handles[8];
-  vllm::Metadata *buffer;
-  T *self_data_copy;
+  vllm::Signal* buffer;
+  T* self_data_copy;
   /**
    * Allocate IPC buffer
    *
@@ -115,9 +115,9 @@ void run(int myRank, int nRanks, ncclComm_t &comm, int threads, int block_limit,
    * convenience.
    */
   CUDACHECK(
-      cudaMalloc(&buffer, 2 * data_size * sizeof(T) + sizeof(vllm::Metadata)));
-  CUDACHECK(cudaMemset(buffer, 0,
-                       2 * data_size * sizeof(T) + sizeof(vllm::Metadata)));
+      cudaMalloc(&buffer, 2 * data_size * sizeof(T) + sizeof(vllm::Signal)));
+  CUDACHECK(
+      cudaMemset(buffer, 0, 2 * data_size * sizeof(T) + sizeof(vllm::Signal)));
   CUDACHECK(cudaMalloc(&self_data_copy, data_size * sizeof(T)));
   CUDACHECK(cudaIpcGetMemHandle(&self_data_handle, buffer));
 
@@ -125,32 +125,32 @@ void run(int myRank, int nRanks, ncclComm_t &comm, int threads, int block_limit,
                          MPI_BYTE, data_handles, sizeof(cudaIpcMemHandle_t),
                          MPI_BYTE, MPI_COMM_WORLD));
 
-  void *rank_data;
+  void* rank_data;
   size_t rank_data_sz = 16 * 1024 * 1024;
   CUDACHECK(cudaMalloc(&rank_data, rank_data_sz));
   std::vector<int64_t> offsets(nRanks, 0);
   vllm::CustomAllreduce fa(buffer, rank_data, rank_data_sz, data_handles,
                            offsets, myRank);
-  auto *self_data =
-      reinterpret_cast<T *>(reinterpret_cast<char *>(buffer) +
-                            sizeof(vllm::Metadata) + data_size * sizeof(T));
+  auto* self_data =
+      reinterpret_cast<T*>(reinterpret_cast<char*>(buffer) +
+                           sizeof(vllm::Signal) + data_size * sizeof(T));
   // hack buffer registration
   {
     std::vector<std::string> handles;
     handles.reserve(nRanks);
     for (int i = 0; i < nRanks; i++) {
-      char *begin = (char *)&data_handles[i];
-      char *end = (char *)&data_handles[i + 1];
+      char* begin = (char*)&data_handles[i];
+      char* end = (char*)&data_handles[i + 1];
       handles.emplace_back(begin, end);
     }
-    std::vector<int64_t> offsets(
-        nRanks, sizeof(vllm::Metadata) + data_size * sizeof(T));
+    std::vector<int64_t> offsets(nRanks,
+                                 sizeof(vllm::Signal) + data_size * sizeof(T));
     fa.register_buffer(handles, offsets, self_data);
   }
 
-  double *ground_truth;
+  double* ground_truth;
   CUDACHECK(cudaMallocHost(&ground_truth, data_size * sizeof(double)));
-  curandState_t *states;
+  curandState_t* states;
   CUDACHECK(cudaMalloc(&states, sizeof(curandState_t) * nRanks * data_size));
   init_rand<<<108, 1024, 0, stream>>>(states, data_size, nRanks);
   gen_data<T><<<108, 1024, 0, stream>>>(states, self_data, ground_truth, myRank,
@@ -169,81 +169,112 @@ void run(int myRank, int nRanks, ncclComm_t &comm, int threads, int block_limit,
   } else {
     ncclDtype = ncclFloat;
   }
-
-  dummy_kernel<<<1, 1, 0, stream>>>();
-  constexpr int warmup_iters = 5;
-  constexpr int num_iters = 25;
-  // warmup
-  for (int i = 0; i < warmup_iters; i++) {
-    NCCLCHECK(ncclAllReduce(result, result, data_size, ncclDtype, ncclSum, comm,
-                            stream));
-  }
-  CUDACHECK(cudaEventRecord(start, stream));
-  for (int i = 0; i < num_iters; i++) {
-    NCCLCHECK(ncclAllReduce(result, result, data_size, ncclDtype, ncclSum, comm,
-                            stream));
-  }
-  CUDACHECK(cudaEventRecord(stop, stream));
-  CUDACHECK(cudaStreamSynchronize(stream));
-  float allreduce_ms = 0;
-  cudaEventElapsedTime(&allreduce_ms, start, stop);
-
-  // if (myRank == 1) dummy_kernel<<<1, 1, 0, stream>>>();
-  // set_data<T><<<16, 1024, 0, stream>>>(self_data, data_size, myRank);
-
-  dummy_kernel<<<1, 1, 0, stream>>>();
-  // warm up
-  for (int i = 0; i < warmup_iters; i++) {
-    fa.allreduce<T>(stream, self_data, result, data_size, threads, block_limit);
-  }
-  CUDACHECK(cudaEventRecord(start, stream));
-  for (int i = 0; i < num_iters; i++) {
-    fa.allreduce<T>(stream, self_data, result, data_size, threads, block_limit);
-  }
-  CUDACHECK(cudaEventRecord(stop, stream));
-  CUDACHECK(cudaStreamSynchronize(stream));
-
-  float duration_ms = 0;
-  cudaEventElapsedTime(&duration_ms, start, stop);
-  if (myRank == 0)
-    printf(
-        "Rank %d done, nGPUs:%d, sz (kb): %d, %d, %d, my time:%.2fus, nccl "
-        "time:%.2fus\n",
-        myRank, nRanks, data_size * sizeof(T) / 1024, threads, block_limit,
-        duration_ms * 1e3 / num_iters, allreduce_ms * 1e3 / num_iters);
-
-  // And wait for all the queued up work to complete
-  CUDACHECK(cudaStreamSynchronize(stream));
-
-  NCCLCHECK(ncclAllReduce(self_data_copy, self_data, data_size, ncclDtype,
-                          ncclSum, comm, stream));
-
   double *nccl_result, *my_result;
   CUDACHECK(cudaMallocHost(&nccl_result, data_size * sizeof(double)));
   CUDACHECK(cudaMallocHost(&my_result, data_size * sizeof(double)));
-
-  convert_data<T><<<108, 1024, 0, stream>>>(self_data, result, nccl_result,
-                                            my_result, data_size);
-  CUDACHECK(cudaStreamSynchronize(stream));
-
-  for (unsigned long j = 0; j < data_size; j++) {
-    auto diff = abs(nccl_result[j] - my_result[j]);
-    if (diff >= 1e-2) {
-      printf("Rank %d: Verification mismatch at %lld: %f != (my) %f, gt=%f\n",
-             myRank, j, nccl_result[j], my_result[j], ground_truth[j]);
-      break;
+  if (performance_test) {
+    dummy_kernel<<<1, 1, 0, stream>>>();
+    constexpr int warmup_iters = 5;
+    constexpr int num_iters = 100;
+    // warmup
+    for (int i = 0; i < warmup_iters; i++) {
+      NCCLCHECK(ncclAllReduce(result, result, data_size, ncclDtype, ncclSum,
+                              comm, stream));
     }
-  }
+    CUDACHECK(cudaEventRecord(start, stream));
+    for (int i = 0; i < num_iters; i++) {
+      NCCLCHECK(ncclAllReduce(result, result, data_size, ncclDtype, ncclSum,
+                              comm, stream));
+    }
+    CUDACHECK(cudaEventRecord(stop, stream));
+    CUDACHECK(cudaStreamSynchronize(stream));
+    float allreduce_ms = 0;
+    cudaEventElapsedTime(&allreduce_ms, start, stop);
 
-  long double nccl_diffs = 0.0;
-  long double my_diffs = 0.0;
-  for (int j = 0; j < data_size; j++) {
-    nccl_diffs += abs(nccl_result[j] - ground_truth[j]);
-    my_diffs += abs(my_result[j] - ground_truth[j]);
+    dummy_kernel<<<1, 1, 0, stream>>>();
+    // warm up
+    for (int i = 0; i < warmup_iters; i++) {
+      fa.allreduce<T>(stream, self_data, result, data_size, threads,
+                      block_limit);
+    }
+    CUDACHECK(cudaEventRecord(start, stream));
+    for (int i = 0; i < num_iters; i++) {
+      fa.allreduce<T>(stream, self_data, result, data_size, threads,
+                      block_limit);
+    }
+    CUDACHECK(cudaEventRecord(stop, stream));
+    CUDACHECK(cudaStreamSynchronize(stream));
+
+    float duration_ms = 0;
+    cudaEventElapsedTime(&duration_ms, start, stop);
+    if (myRank == 0)
+      printf(
+          "Rank %d done, nGPUs:%d, sz (kb): %d, %d, %d, my time:%.2fus, nccl "
+          "time:%.2fus\n",
+          myRank, nRanks, data_size * sizeof(T) / 1024, threads, block_limit,
+          duration_ms * 1e3 / num_iters, allreduce_ms * 1e3 / num_iters);
+
+    // And wait for all the queued up work to complete
+    CUDACHECK(cudaStreamSynchronize(stream));
+
+    NCCLCHECK(ncclAllReduce(self_data_copy, self_data, data_size, ncclDtype,
+                            ncclSum, comm, stream));
+
+    convert_data<T><<<108, 1024, 0, stream>>>(self_data, result, nccl_result,
+                                              my_result, data_size);
+    CUDACHECK(cudaStreamSynchronize(stream));
+
+    for (unsigned long j = 0; j < data_size; j++) {
+      auto diff = abs(nccl_result[j] - my_result[j]);
+      if (diff >= 4e-2) {
+        printf("Rank %d: Verification mismatch at %lld: %f != (my) %f, gt=%f\n",
+               myRank, j, nccl_result[j], my_result[j], ground_truth[j]);
+        break;
+      }
+    }
+    long double nccl_diffs = 0.0;
+    long double my_diffs = 0.0;
+    for (int j = 0; j < data_size; j++) {
+      nccl_diffs += abs(nccl_result[j] - ground_truth[j]);
+      my_diffs += abs(my_result[j] - ground_truth[j]);
+    }
+    if (myRank == 0)
+      std::cout << "average abs diffs: nccl: " << nccl_diffs / data_size
+                << " me: " << my_diffs / data_size << std::endl;
+  } else {
+    for (int i = 0; i < 100; i++) {
+      fa.allreduce<T>(stream, self_data, result, data_size, threads,
+                      block_limit);
+      CUDACHECK(cudaStreamSynchronize(stream));
+      NCCLCHECK(ncclAllReduce(self_data, self_data_copy, data_size, ncclDtype,
+                              ncclSum, comm, stream));
+      convert_data<T><<<108, 1024, 0, stream>>>(
+          self_data_copy, result, nccl_result, my_result, data_size);
+      CUDACHECK(cudaStreamSynchronize(stream));
+
+      for (unsigned long j = 0; j < data_size; j++) {
+        auto diff = abs(nccl_result[j] - my_result[j]);
+        if (diff >= 4e-2) {
+          printf(
+              "Rank %d: Verification mismatch at %lld: %f != (my) %f, gt=%f\n",
+              myRank, j, nccl_result[j], my_result[j], ground_truth[j]);
+          break;
+        }
+      }
+    }
+    if (myRank == 0)
+      printf("Test passed: nGPUs:%d, sz (kb): %d, %d, %d\n", nRanks,
+             data_size * sizeof(T) / 1024, threads, block_limit);
+    // long double nccl_diffs = 0.0;
+    // long double my_diffs = 0.0;
+    // for (int j = 0; j < data_size; j++) {
+    //   nccl_diffs += abs(nccl_result[j] - ground_truth[j]);
+    //   my_diffs += abs(my_result[j] - ground_truth[j]);
+    // }
+    // if (myRank == 0)
+    //   std::cout << "average abs diffs: nccl: " << nccl_diffs / data_size
+    //             << " me: " << my_diffs / data_size << std::endl;
   }
-  if (myRank == 0)
-    std::cout << "average abs diffs: nccl: " << nccl_diffs / data_size
-              << " me: " << my_diffs / data_size << std::endl;
 
   CUDACHECK(cudaFree(result));
   CUDACHECK(cudaFree(self_data_copy));
@@ -256,7 +287,7 @@ void run(int myRank, int nRanks, ncclComm_t &comm, int threads, int block_limit,
   CUDACHECK(cudaStreamDestroy(stream));
 }
 
-int main(int argc, char **argv) {
+int main(int argc, char** argv) {
   int nRanks, myRank;
   MPICHECK(MPI_Init(&argc, &argv));
   MPICHECK(MPI_Comm_rank(MPI_COMM_WORLD, &myRank));
@@ -265,18 +296,19 @@ int main(int argc, char **argv) {
   ncclUniqueId id;
   ncclComm_t comm;
   if (myRank == 0) ncclGetUniqueId(&id);
-  MPICHECK(MPI_Bcast(static_cast<void *>(&id), sizeof(id), MPI_BYTE, 0,
+  MPICHECK(MPI_Bcast(static_cast<void*>(&id), sizeof(id), MPI_BYTE, 0,
                      MPI_COMM_WORLD));
   NCCLCHECK(ncclCommInitRank(&comm, nRanks, id, myRank));
 
+  bool performance_test = true;
   cudaProfilerStart();
   // for (int threads : {256, 512}) {
   //   for (int block_limit = 16; block_limit < 112; block_limit += 4) {
   //     run<half>(myRank, nRanks, comm, threads, block_limit, 4096 * 1024);
   //   }
   // }
-  for (int sz = 512; sz <= (32 << 20); sz *= 2) {
-    run<half>(myRank, nRanks, comm, 512, 36, sz + 8 * 50);
+  for (int sz = 512; sz <= (8 << 20); sz *= 2) {
+    run<half>(myRank, nRanks, comm, 512, 36, sz + 8 * 47, performance_test);
   }
 
   cudaProfilerStop();

csrc/dispatch_utils.h

@@ -6,32 +6,30 @@
 
 #include <torch/extension.h>
 
-#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)              \
-  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)      \
-  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)       \
+#define VLLM_DISPATCH_CASE_FLOATING_TYPES(...)         \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Half, __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__))
+#define VLLM_DISPATCH_FLOATING_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_TYPES(__VA_ARGS__))
 
-#define VLLM_DISPATCH_CASE_FLOATING_AND_BYTE_TYPES(...)     \
-  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)      \
-  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)       \
-  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)   \
+#define VLLM_DISPATCH_CASE_FLOATING_AND_BYTE_TYPES(...)   \
+  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__)    \
+  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)     \
+  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__) \
   AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__)
 
-#define VLLM_DISPATCH_FLOATING_AND_BYTE_TYPES(TYPE, NAME, ...)           \
-  AT_DISPATCH_SWITCH(                                                    \
-    TYPE, NAME, VLLM_DISPATCH_CASE_FLOATING_AND_BYTE_TYPES(__VA_ARGS__))
-    
-#define VLLM_DISPATCH_CASE_INTEGRAL_TYPES(...)             \
-  AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__)      \
-  AT_DISPATCH_CASE(at::ScalarType::Char, __VA_ARGS__)      \
-  AT_DISPATCH_CASE(at::ScalarType::Short, __VA_ARGS__)     \
-  AT_DISPATCH_CASE(at::ScalarType::Int, __VA_ARGS__)       \
+#define VLLM_DISPATCH_FLOATING_AND_BYTE_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME,                               \
+                     VLLM_DISPATCH_CASE_FLOATING_AND_BYTE_TYPES(__VA_ARGS__))
+
+#define VLLM_DISPATCH_CASE_INTEGRAL_TYPES(...)         \
+  AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__)  \
+  AT_DISPATCH_CASE(at::ScalarType::Char, __VA_ARGS__)  \
+  AT_DISPATCH_CASE(at::ScalarType::Short, __VA_ARGS__) \
+  AT_DISPATCH_CASE(at::ScalarType::Int, __VA_ARGS__)   \
   AT_DISPATCH_CASE(at::ScalarType::Long, __VA_ARGS__)
 
-#define VLLM_DISPATCH_INTEGRAL_TYPES(TYPE, NAME, ...)             \
-  AT_DISPATCH_SWITCH(                                             \
-    TYPE, NAME, VLLM_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__))
+#define VLLM_DISPATCH_INTEGRAL_TYPES(TYPE, NAME, ...) \
+  AT_DISPATCH_SWITCH(TYPE, NAME, VLLM_DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__))

csrc/layernorm_kernels.cu

@@ -4,23 +4,31 @@
 
 #include "dispatch_utils.h"
 #include "reduction_utils.cuh"
+#ifndef USE_ROCM
+  #include <cuda_bf16.h>
+  #include <cuda_fp16.h>
+#else
+  #include <hip/hip_bf16.h>
+  #include <hip/hip_fp16.h>
+
+using __nv_bfloat16 = __hip_bfloat16;
+using __nv_bfloat162 = __hip_bfloat162;
+#endif
 
 namespace vllm {
 
 // TODO(woosuk): Further optimize this kernel.
-template<typename scalar_t>
+template <typename scalar_t>
 __global__ void rms_norm_kernel(
-  scalar_t* __restrict__ out,             // [..., hidden_size]
-  const scalar_t* __restrict__ input,     // [..., hidden_size]
-  const scalar_t* __restrict__ weight,    // [hidden_size]
-  const float epsilon,
-  const int num_tokens,
-  const int hidden_size) {
+    scalar_t* __restrict__ out,           // [..., hidden_size]
+    const scalar_t* __restrict__ input,   // [..., hidden_size]
+    const scalar_t* __restrict__ weight,  // [hidden_size]
+    const float epsilon, const int num_tokens, const int hidden_size) {
   __shared__ float s_variance;
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    const float x = (float) input[blockIdx.x * hidden_size + idx];
+    const float x = (float)input[blockIdx.x * hidden_size + idx];
     variance += x * x;
   }
   variance = blockReduceSum<float>(variance);
@@ -30,48 +38,260 @@ __global__ void rms_norm_kernel(
   __syncthreads();
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float) input[blockIdx.x * hidden_size + idx];
-    out[blockIdx.x * hidden_size + idx] = ((scalar_t) (x * s_variance)) * weight[idx];
+    float x = (float)input[blockIdx.x * hidden_size + idx];
+    out[blockIdx.x * hidden_size + idx] =
+        ((scalar_t)(x * s_variance)) * weight[idx];
   }
 }
 
-// TODO: Further optimize this kernel.
-template<typename scalar_t>
-__global__ void fused_add_rms_norm_kernel(
-  scalar_t* __restrict__ input,           // [..., hidden_size]
-  scalar_t* __restrict__ residual,        // [..., hidden_size]
-  const scalar_t* __restrict__ weight,    // [hidden_size]
-  const float epsilon,
-  const int num_tokens,
-  const int hidden_size) {
+/* Converter structs for the conversion from torch types to HIP/CUDA types,
+   and the associated type conversions within HIP/CUDA. These helpers need
+   to be implemented for now because the relevant type conversion
+   operators/constructors are not consistently implemented by HIP/CUDA, so
+   a generic conversion via type casts cannot be implemented.
+
+   Each struct should have the member static constexpr bool `exists`:
+   If false, the optimized kernel is not used for the corresponding torch type.
+   If true, the struct should be fully defined as shown in the examples below.
+ */
+template <typename torch_type>
+struct _typeConvert {
+  static constexpr bool exists = false;
+};
+
+#if defined(USE_ROCM) || (defined(CUDA_VERSION) && (CUDA_VERSION >= 12000))
+// CUDA < 12.0 runs into issues with packed type conversion
+template <>
+struct _typeConvert<c10::Half> {
+  static constexpr bool exists = true;
+  using hip_type = __half;
+  using packed_hip_type = __half2;
+
+  __device__ static inline float convert(hip_type x) { return __half2float(x); }
+  __device__ static inline float2 convert(packed_hip_type x) {
+    return __half22float2(x);
+  }
+  __device__ static inline hip_type convert(float x) {
+    return __float2half_rn(x);
+  }
+  __device__ static inline packed_hip_type convert(float2 x) {
+    return __float22half2_rn(x);
+  }
+};
+
+  #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+// CUDA_ARCH < 800 does not have BF16 support
+// TODO: Add in ROCm support once public headers handle bf16 maturely
+template <>
+struct _typeConvert<c10::BFloat16> {
+  static constexpr bool exists = true;
+  using hip_type = __nv_bfloat16;
+  using packed_hip_type = __nv_bfloat162;
+
+  __device__ static inline float convert(hip_type x) {
+    return __bfloat162float(x);
+  }
+  __device__ static inline float2 convert(packed_hip_type x) {
+    return __bfloat1622float2(x);
+  }
+  __device__ static inline hip_type convert(float x) {
+    return __float2bfloat16(x);
+  }
+  __device__ static inline packed_hip_type convert(float2 x) {
+    return __float22bfloat162_rn(x);
+  }
+};
+  #endif  // defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+#endif    // defined(USE_ROCM) || (defined(CUDA_VERSION) && (CUDA_VERSION >=
+          // 12000))
+
+/* Vector POD struct to generate vectorized and packed FP16/BF16 ops
+   for appropriate specializations of fused_add_rms_norm_kernel.
+   Only functions that are necessary in that kernel are implemented.
+   Alignment to 16 bytes is required to use 128-bit global memory ops.
+ */
+template <typename scalar_t, int width>
+struct alignas(16) _f16Vec {
+  /* Not theoretically necessary that width is a power of 2 but should
+     almost always be the case for optimization purposes */
+  static_assert(width > 0 && (width & (width - 1)) == 0,
+                "Width is not a positive power of 2!");
+  using Converter = _typeConvert<scalar_t>;
+  using T1 = typename Converter::hip_type;
+  using T2 = typename Converter::packed_hip_type;
+  T1 data[width];
+
+  __device__ _f16Vec& operator+=(const _f16Vec<scalar_t, width>& other) {
+    if constexpr (width % 2 == 0) {
+#pragma unroll
+      for (int i = 0; i < width; i += 2) {
+        T2 temp{data[i], data[i + 1]};
+        temp += T2{other.data[i], other.data[i + 1]};
+        data[i] = temp.x;
+        data[i + 1] = temp.y;
+      }
+    } else {
+#pragma unroll
+      for (int i = 0; i < width; ++i) data[i] += other.data[i];
+    }
+    return *this;
+  }
+
+  __device__ _f16Vec& operator*=(const _f16Vec<scalar_t, width>& other) {
+    if constexpr (width % 2 == 0) {
+#pragma unroll
+      for (int i = 0; i < width; i += 2) {
+        T2 temp{data[i], data[i + 1]};
+        temp *= T2{other.data[i], other.data[i + 1]};
+        data[i] = temp.x;
+        data[i + 1] = temp.y;
+      }
+    } else {
+#pragma unroll
+      for (int i = 0; i < width; ++i) data[i] *= other.data[i];
+    }
+    return *this;
+  }
+
+  __device__ _f16Vec& operator*=(const float scale) {
+    if constexpr (width % 2 == 0) {
+#pragma unroll
+      for (int i = 0; i < width; i += 2) {
+        float2 temp_f = Converter::convert(T2{data[i], data[i + 1]});
+        temp_f.x *= scale;
+        temp_f.y *= scale;
+        T2 temp = Converter::convert(temp_f);
+        data[i] = temp.x;
+        data[i + 1] = temp.y;
+      }
+    } else {
+#pragma unroll
+      for (int i = 0; i < width; ++i) {
+        float temp = Converter::convert(data[i]) * scale;
+        data[i] = Converter::convert(temp);
+      }
+    }
+    return *this;
+  }
+
+  __device__ float sum_squares() const {
+    float result = 0.0f;
+    if constexpr (width % 2 == 0) {
+#pragma unroll
+      for (int i = 0; i < width; i += 2) {
+        float2 z = Converter::convert(T2{data[i], data[i + 1]});
+        result += z.x * z.x + z.y * z.y;
+      }
+    } else {
+#pragma unroll
+      for (int i = 0; i < width; ++i) {
+        float x = Converter::convert(data[i]);
+        result += x * x;
+      }
+    }
+    return result;
+  }
+};
+
+/* Function specialization in the case of FP16/BF16 tensors.
+   Additional optimizations we can make in this case are
+   packed and vectorized operations, which help with the
+   memory latency bottleneck. */
+template <typename scalar_t, int width>
+__global__ std::enable_if_t<(width > 0) && _typeConvert<scalar_t>::exists>
+fused_add_rms_norm_kernel(
+    scalar_t* __restrict__ input,         // [..., hidden_size]
+    scalar_t* __restrict__ residual,      // [..., hidden_size]
+    const scalar_t* __restrict__ weight,  // [hidden_size]
+    const float epsilon, const int num_tokens, const int hidden_size) {
+  // Sanity checks on our vector struct and type-punned pointer arithmetic
+  static_assert(std::is_pod_v<_f16Vec<scalar_t, width>>);
+  static_assert(sizeof(_f16Vec<scalar_t, width>) == sizeof(scalar_t) * width);
+
+  const int vec_hidden_size = hidden_size / width;
+  __shared__ float s_variance;
+  float variance = 0.0f;
+  /* These and the argument pointers are all declared `restrict` as they are
+     not aliased in practice. Argument pointers should not be dereferenced
+     in this kernel as that would be undefined behavior */
+  auto* __restrict__ input_v =
+      reinterpret_cast<_f16Vec<scalar_t, width>*>(input);
+  auto* __restrict__ residual_v =
+      reinterpret_cast<_f16Vec<scalar_t, width>*>(residual);
+  auto* __restrict__ weight_v =
+      reinterpret_cast<const _f16Vec<scalar_t, width>*>(weight);
+
+  for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
+    int id = blockIdx.x * vec_hidden_size + idx;
+    _f16Vec<scalar_t, width> temp = input_v[id];
+    temp += residual_v[id];
+    variance += temp.sum_squares();
+    residual_v[id] = temp;
+  }
+  /* Keep the following if-else block in sync with the
+     calculation of max_block_size in fused_add_rms_norm */
+  if (num_tokens < 256) {
+    variance = blockReduceSum<float, 1024>(variance);
+  } else
+    variance = blockReduceSum<float, 256>(variance);
+  if (threadIdx.x == 0) {
+    s_variance = rsqrtf(variance / hidden_size + epsilon);
+  }
+  __syncthreads();
+
+  for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
+    int id = blockIdx.x * vec_hidden_size + idx;
+    _f16Vec<scalar_t, width> temp = residual_v[id];
+    temp *= s_variance;
+    temp *= weight_v[idx];
+    input_v[id] = temp;
+  }
+}
+
+/* Generic fused_add_rms_norm_kernel
+   The width field is not used here but necessary for other specializations.
+ */
+template <typename scalar_t, int width>
+__global__ std::enable_if_t<(width == 0) || !_typeConvert<scalar_t>::exists>
+fused_add_rms_norm_kernel(
+    scalar_t* __restrict__ input,         // [..., hidden_size]
+    scalar_t* __restrict__ residual,      // [..., hidden_size]
+    const scalar_t* __restrict__ weight,  // [hidden_size]
+    const float epsilon, const int num_tokens, const int hidden_size) {
   __shared__ float s_variance;
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float) input[blockIdx.x * hidden_size + idx];
-    x += (float) residual[blockIdx.x * hidden_size + idx];
+    scalar_t z = input[blockIdx.x * hidden_size + idx];
+    z += residual[blockIdx.x * hidden_size + idx];
+    float x = (float)z;
     variance += x * x;
-    residual[blockIdx.x * hidden_size + idx] = (scalar_t) x;
+    residual[blockIdx.x * hidden_size + idx] = z;
   }
-  variance = blockReduceSum<float>(variance);
+  /* Keep the following if-else block in sync with the
+     calculation of max_block_size in fused_add_rms_norm */
+  if (num_tokens < 256) {
+    variance = blockReduceSum<float, 1024>(variance);
+  } else
+    variance = blockReduceSum<float, 256>(variance);
   if (threadIdx.x == 0) {
     s_variance = rsqrtf(variance / hidden_size + epsilon);
   }
   __syncthreads();
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float) residual[blockIdx.x * hidden_size + idx];
-    input[blockIdx.x * hidden_size + idx] = ((scalar_t) (x * s_variance)) * weight[idx];
+    float x = (float)residual[blockIdx.x * hidden_size + idx];
+    input[blockIdx.x * hidden_size + idx] =
+        ((scalar_t)(x * s_variance)) * weight[idx];
   }
 }
 
-} // namespace vllm
+}  // namespace vllm
 
-void rms_norm(
-  torch::Tensor& out,      // [..., hidden_size]
-  torch::Tensor& input,    // [..., hidden_size]
-  torch::Tensor& weight,   // [hidden_size]
-  float epsilon) {
+void rms_norm(torch::Tensor& out,     // [..., hidden_size]
+              torch::Tensor& input,   // [..., hidden_size]
+              torch::Tensor& weight,  // [hidden_size]
+              float epsilon) {
   int hidden_size = input.size(-1);
   int num_tokens = input.numel() / hidden_size;
 
@@ -79,42 +299,54 @@ void rms_norm(
   dim3 block(std::min(hidden_size, 1024));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  VLLM_DISPATCH_FLOATING_TYPES(
-    input.scalar_type(),
-    "rms_norm_kernel",
-    [&] {
-      vllm::rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
-        out.data_ptr<scalar_t>(),
-        input.data_ptr<scalar_t>(),
-        weight.data_ptr<scalar_t>(),
-        epsilon,
-        num_tokens,
-        hidden_size);
-    });
+  VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "rms_norm_kernel", [&] {
+    vllm::rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
+        out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),
+        weight.data_ptr<scalar_t>(), epsilon, num_tokens, hidden_size);
+  });
 }
 
-void fused_add_rms_norm(
-  torch::Tensor& input,    // [..., hidden_size]
-  torch::Tensor& residual, // [..., hidden_size]
-  torch::Tensor& weight,   // [hidden_size]
-  float epsilon) {
+#define LAUNCH_FUSED_ADD_RMS_NORM(width)                                       \
+  VLLM_DISPATCH_FLOATING_TYPES(                                                \
+      input.scalar_type(), "fused_add_rms_norm_kernel", [&] {                  \
+        vllm::fused_add_rms_norm_kernel<scalar_t, width>                       \
+            <<<grid, block, 0, stream>>>(input.data_ptr<scalar_t>(),           \
+                                         residual.data_ptr<scalar_t>(),        \
+                                         weight.data_ptr<scalar_t>(), epsilon, \
+                                         num_tokens, hidden_size);             \
+      });
+
+void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
+                        torch::Tensor& residual,  // [..., hidden_size]
+                        torch::Tensor& weight,    // [hidden_size]
+                        float epsilon) {
   int hidden_size = input.size(-1);
   int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
-  dim3 block(std::min(hidden_size, 1024));
+  /* This kernel is memory-latency bound in many scenarios.
+     When num_tokens is large, a smaller block size allows
+     for increased block occupancy on CUs and better latency
+     hiding on global mem ops. */
+  const int max_block_size = (num_tokens < 256) ? 1024 : 256;
+  dim3 block(std::min(hidden_size, max_block_size));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  VLLM_DISPATCH_FLOATING_TYPES(
-    input.scalar_type(),
-    "fused_add_rms_norm_kernel",
-    [&] {
-      vllm::fused_add_rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
-        input.data_ptr<scalar_t>(),
-        residual.data_ptr<scalar_t>(),
-        weight.data_ptr<scalar_t>(),
-        epsilon,
-        num_tokens,
-        hidden_size);
-    });
+  /*If the tensor types are FP16/BF16, try to use the optimized kernel
+    with packed + vectorized ops.
+    Max optimization is achieved with a width-8 vector of FP16/BF16s
+    since we can load at most 128 bits at once in a global memory op.
+    However, this requires each tensor's data to be aligned to 16
+    bytes.
+   */
+  auto inp_ptr = reinterpret_cast<std::uintptr_t>(input.data_ptr());
+  auto res_ptr = reinterpret_cast<std::uintptr_t>(residual.data_ptr());
+  auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight.data_ptr());
+  bool ptrs_are_aligned =
+      inp_ptr % 16 == 0 && res_ptr % 16 == 0 && wt_ptr % 16 == 0;
+  if (ptrs_are_aligned && hidden_size % 8 == 0) {
+    LAUNCH_FUSED_ADD_RMS_NORM(8);
+  } else {
+    LAUNCH_FUSED_ADD_RMS_NORM(0);
+  }
 }

csrc/moe/moe_ops.cpp

@@ -3,5 +3,6 @@
 #include <torch/extension.h>
 
 PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-  m.def("topk_softmax", &topk_softmax, "Apply topk softmax to the gating outputs.");
+  m.def("topk_softmax", &topk_softmax,
+        "Apply topk softmax to the gating outputs.");
 }

csrc/moe/moe_ops.h

@@ -2,8 +2,6 @@
 
 #include <torch/extension.h>
 
-void topk_softmax(
-  torch::Tensor& topk_weights,
-  torch::Tensor& topk_indices,
-  torch::Tensor& token_expert_indices,
-  torch::Tensor& gating_output);
+void topk_softmax(torch::Tensor& topk_weights, torch::Tensor& topk_indices,
+                  torch::Tensor& token_expert_indices,
+                  torch::Tensor& gating_output);

csrc/moe_align_block_size_kernels.cu

@@ -7,102 +7,128 @@
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
 
-const static size_t NUM_MAX_EXPERTS = 64;
-#define CEILDIV(x,y) (((x) + (y) - 1) / (y))
+#define CEILDIV(x, y) (((x) + (y) - 1) / (y))
 
 namespace vllm {
+
+namespace {
+__device__ __forceinline__ int32_t index(int32_t total_col, int32_t row,
+                                         int32_t col) {
+  // don't worry about overflow because num_experts is relatively small
+  return row * total_col + col;
+}
+}  // namespace
+
 template <typename scalar_t>
-__global__ void moe_align_block_size_kernel(scalar_t *__restrict__ topk_ids, 
-                                int32_t *sorted_token_ids, 
-                                int32_t *expert_ids, 
-                                int32_t *total_tokens_post_pad,
-                                int32_t num_experts, 
-                                int32_t block_size, 
-                                size_t numel) {
-    const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
-    const size_t start_idx = threadIdx.x * tokens_per_thread;
-    __shared__ int32_t tokens_cnts[NUM_MAX_EXPERTS + 1][NUM_MAX_EXPERTS];
-    __shared__ int32_t cumsum[NUM_MAX_EXPERTS + 1];
-    for (int i = 0; i < num_experts; ++i) {
-        tokens_cnts[threadIdx.x + 1][i] = 0;
-    }
+__global__ void moe_align_block_size_kernel(scalar_t* __restrict__ topk_ids,
+                                            int32_t* sorted_token_ids,
+                                            int32_t* expert_ids,
+                                            int32_t* total_tokens_post_pad,
+                                            int32_t num_experts,
+                                            int32_t block_size, size_t numel) {
+  const size_t tokens_per_thread = CEILDIV(numel, blockDim.x);
+  const size_t start_idx = threadIdx.x * tokens_per_thread;
 
-    /**
-    * In the first step we compute token_cnts[thread_index + 1][expert_index],
-    * which counts how many tokens in the token shard of thread_index are assigned
-    * to expert expert_index.
-    */
-    for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
-        ++tokens_cnts[threadIdx.x + 1][topk_ids[i]]; 
-    }
+  extern __shared__ int32_t shared_mem[];
 
-    __syncthreads();
+  int32_t* tokens_cnts =
+      shared_mem;  // 2d tensor with shape (num_experts + 1, num_experts)
+  int32_t* cumsum =
+      shared_mem + (num_experts + 1) *
+                       num_experts;  // 1d tensor with shape (num_experts + 1)
 
-    // For each expert we accumulate the token counts from the different threads.
-    tokens_cnts[0][threadIdx.x] = 0;
-    for (int i = 1; i <= blockDim.x; ++i) {
-        tokens_cnts[i][threadIdx.x] += tokens_cnts[i-1][threadIdx.x];
-    }
+  for (int i = 0; i < num_experts; ++i) {
+    tokens_cnts[index(num_experts, threadIdx.x + 1, i)] = 0;
+  }
 
-    __syncthreads();
-    
-    // We accumulate the token counts of all experts in thread 0.
-    if (threadIdx.x == 0) {
-        cumsum[0] = 0;
-        for (int i = 1; i <= num_experts; ++i) {
-            cumsum[i] = cumsum[i-1] + CEILDIV(tokens_cnts[blockDim.x][i - 1], block_size) * block_size;
-        }
-        *total_tokens_post_pad = cumsum[num_experts];
-    }
+  /**
+   * In the first step we compute token_cnts[thread_index + 1][expert_index],
+   * which counts how many tokens in the token shard of thread_index are
+   * assigned to expert expert_index.
+   */
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    ++tokens_cnts[index(num_experts, threadIdx.x + 1, topk_ids[i])];
+  }
 
-    __syncthreads();
+  __syncthreads();
 
-    /**
-    * For each expert, each thread processes the tokens of the corresponding blocks
-    * and stores the corresponding expert_id for each block.
-    */
-    for (int i = cumsum[threadIdx.x];i < cumsum[threadIdx.x + 1];i += block_size) {
-        expert_ids[i / block_size] = threadIdx.x;
-    }
-    
-    /**
-    * Each thread processes a token shard, calculating the index of each token after
-    * sorting by expert number. Given the example topk_ids = [0,1,2,1,2,3,0,3,4] and
-    * block_size = 4, then the output would be [0, 6, *, *, 1, 3, *, *, 2, 4, *, *, 5, 7, *, *, 8, *, *, *],
-    * where * represents a padding value(preset in python).
-    */
-    for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
-        int32_t expert_id = topk_ids[i];
-        /** The cumsum[expert_id] stores the starting index of the tokens that the
-        * expert with expert_id needs to process, and tokens_cnts[threadIdx.x][expert_id]
-        * stores the indices of the tokens processed by the expert with expert_id within
-        * the current thread's token shard.
-        */
-        int32_t rank_post_pad = tokens_cnts[threadIdx.x][expert_id] + cumsum[expert_id];
-        sorted_token_ids[rank_post_pad] = i;
-        ++tokens_cnts[threadIdx.x][expert_id];
+  // For each expert we accumulate the token counts from the different threads.
+  tokens_cnts[index(num_experts, 0, threadIdx.x)] = 0;
+  for (int i = 1; i <= blockDim.x; ++i) {
+    tokens_cnts[index(num_experts, i, threadIdx.x)] +=
+        tokens_cnts[index(num_experts, i - 1, threadIdx.x)];
+  }
+
+  __syncthreads();
+
+  // We accumulate the token counts of all experts in thread 0.
+  if (threadIdx.x == 0) {
+    cumsum[0] = 0;
+    for (int i = 1; i <= num_experts; ++i) {
+      cumsum[i] = cumsum[i - 1] +
+                  CEILDIV(tokens_cnts[index(num_experts, blockDim.x, i - 1)],
+                          block_size) *
+                      block_size;
     }
+    *total_tokens_post_pad = cumsum[num_experts];
+  }
+
+  __syncthreads();
+
+  /**
+   * For each expert, each thread processes the tokens of the corresponding
+   * blocks and stores the corresponding expert_id for each block.
+   */
+  for (int i = cumsum[threadIdx.x]; i < cumsum[threadIdx.x + 1];
+       i += block_size) {
+    expert_ids[i / block_size] = threadIdx.x;
+  }
+
+  /**
+   * Each thread processes a token shard, calculating the index of each token
+   * after sorting by expert number. Given the example topk_ids =
+   * [0,1,2,1,2,3,0,3,4] and block_size = 4, then the output would be [0, 6, *,
+   * *, 1, 3, *, *, 2, 4, *, *, 5, 7, *, *, 8, *, *, *], where * represents a
+   * padding value(preset in python).
+   */
+  for (int i = start_idx; i < numel && i < start_idx + tokens_per_thread; ++i) {
+    int32_t expert_id = topk_ids[i];
+    /** The cumsum[expert_id] stores the starting index of the tokens that the
+     * expert with expert_id needs to process, and
+     * tokens_cnts[threadIdx.x][expert_id] stores the indices of the tokens
+     * processed by the expert with expert_id within the current thread's token
+     * shard.
+     */
+    int32_t rank_post_pad =
+        tokens_cnts[index(num_experts, threadIdx.x, expert_id)] +
+        cumsum[expert_id];
+    sorted_token_ids[rank_post_pad] = i;
+    ++tokens_cnts[index(num_experts, threadIdx.x, expert_id)];
+  }
 }
-}
+}  // namespace vllm
 
-void moe_align_block_size(
-    torch::Tensor topk_ids,
-    int num_experts,
-    int block_size,
-    torch::Tensor sorted_token_ids,
-    torch::Tensor experts_ids,
-    torch::Tensor num_tokens_post_pad) {
-    const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-    assert(num_experts <= NUM_MAX_EXPERTS);
-    VLLM_DISPATCH_INTEGRAL_TYPES(
-        topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
-        vllm::moe_align_block_size_kernel<scalar_t><<<1, num_experts, 0, stream>>>(
-            topk_ids.data_ptr<scalar_t>(), 
-            sorted_token_ids.data_ptr<int32_t>(), 
-            experts_ids.data_ptr<int32_t>(), 
-            num_tokens_post_pad.data_ptr<int32_t>(), 
-            num_experts,
-            block_size,
+void moe_align_block_size(torch::Tensor topk_ids, int num_experts,
+                          int block_size, torch::Tensor sorted_token_ids,
+                          torch::Tensor experts_ids,
+                          torch::Tensor num_tokens_post_pad) {
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  VLLM_DISPATCH_INTEGRAL_TYPES(
+      topk_ids.scalar_type(), "moe_align_block_size_kernel", [&] {
+        // calc needed amount of shared mem for `tokens_cnts` and `cumsum`
+        // tensors
+        const int32_t shared_mem =
+            ((num_experts + 1) * num_experts + (num_experts + 1)) *
+            sizeof(int32_t);
+
+        // set dynamic shared mem
+        auto kernel = vllm::moe_align_block_size_kernel<scalar_t>;
+        AT_CUDA_CHECK(VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(
+            (void*)kernel, shared_mem));
+        kernel<<<1, num_experts, shared_mem, stream>>>(
+            topk_ids.data_ptr<scalar_t>(), sorted_token_ids.data_ptr<int32_t>(),
+            experts_ids.data_ptr<int32_t>(),
+            num_tokens_post_pad.data_ptr<int32_t>(), num_experts, block_size,
             topk_ids.numel());
-    });
+      });
 }

csrc/ops.h

@@ -3,143 +3,139 @@
 #include <torch/extension.h>
 
 void paged_attention_v1(
-  torch::Tensor& out,
-  torch::Tensor& query,
-  torch::Tensor& key_cache,
-  torch::Tensor& value_cache,
-  int num_kv_heads,
-  float scale,
-  torch::Tensor& block_tables,
-  torch::Tensor& context_lens,
-  int block_size,
-  int max_context_len,
-  const c10::optional<torch::Tensor>& alibi_slopes,
-  const std::string& kv_cache_dtype);
+    torch::Tensor& out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int block_size,
+    int max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
+    const std::string& kv_cache_dtype, float kv_scale, const int tp_rank,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step);
 
 void paged_attention_v2(
-  torch::Tensor& out,
-  torch::Tensor& exp_sums,
-  torch::Tensor& max_logits,
-  torch::Tensor& tmp_out,
-  torch::Tensor& query,
-  torch::Tensor& key_cache,
-  torch::Tensor& value_cache,
-  int num_kv_heads,
-  float scale,
-  torch::Tensor& block_tables,
-  torch::Tensor& context_lens,
-  int block_size,
-  int max_context_len,
-  const c10::optional<torch::Tensor>& alibi_slopes,
-  const std::string& kv_cache_dtype);
+    torch::Tensor& out, torch::Tensor& exp_sums, torch::Tensor& max_logits,
+    torch::Tensor& tmp_out, torch::Tensor& query, torch::Tensor& key_cache,
+    torch::Tensor& value_cache, int num_kv_heads, float scale,
+    torch::Tensor& block_tables, torch::Tensor& seq_lens, int block_size,
+    int max_seq_len, const c10::optional<torch::Tensor>& alibi_slopes,
+    const std::string& kv_cache_dtype, float kv_scale, const int tp_rank,
+    const int blocksparse_local_blocks, const int blocksparse_vert_stride,
+    const int blocksparse_block_size, const int blocksparse_head_sliding_step);
 
-void rms_norm(
-  torch::Tensor& out,
-  torch::Tensor& input,
-  torch::Tensor& weight,
-  float epsilon);
+void rms_norm(torch::Tensor& out, torch::Tensor& input, torch::Tensor& weight,
+              float epsilon);
 
-void fused_add_rms_norm(
-  torch::Tensor& input,
-  torch::Tensor& residual,
-  torch::Tensor& weight,
-  float epsilon);
+void fused_add_rms_norm(torch::Tensor& input, torch::Tensor& residual,
+                        torch::Tensor& weight, float epsilon);
 
-void rotary_embedding(
-  torch::Tensor& positions,
-  torch::Tensor& query,
-  torch::Tensor& key,
-  int head_size,
-  torch::Tensor& cos_sin_cache,
-  bool is_neox);
+void rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
+                      torch::Tensor& key, int head_size,
+                      torch::Tensor& cos_sin_cache, bool is_neox);
 
-void silu_and_mul(
-  torch::Tensor& out,
-  torch::Tensor& input);
+void batched_rotary_embedding(torch::Tensor& positions, torch::Tensor& query,
+                              torch::Tensor& key, int head_size,
+                              torch::Tensor& cos_sin_cache, bool is_neox,
+                              int rot_dim,
+                              torch::Tensor& cos_sin_cache_offsets);
 
-void gelu_and_mul(
-  torch::Tensor& out,
-  torch::Tensor& input);
+void silu_and_mul(torch::Tensor& out, torch::Tensor& input);
 
-void gelu_new(
-  torch::Tensor& out,
-  torch::Tensor& input);
+void gelu_and_mul(torch::Tensor& out, torch::Tensor& input);
 
-void gelu_fast(
-  torch::Tensor& out,
-  torch::Tensor& input);
+void gelu_tanh_and_mul(torch::Tensor& out, torch::Tensor& input);
+
+void gelu_new(torch::Tensor& out, torch::Tensor& input);
+
+void gelu_fast(torch::Tensor& out, torch::Tensor& input);
 
 #ifndef USE_ROCM
-torch::Tensor awq_gemm(
-  torch::Tensor _in_feats,
-  torch::Tensor _kernel,
-  torch::Tensor _scaling_factors,
-  torch::Tensor _zeros,
-  int split_k_iters);
+torch::Tensor aqlm_gemm(const torch::Tensor& input, const torch::Tensor& codes,
+                        const torch::Tensor& codebooks,
+                        const torch::Tensor& scales,
+                        const torch::Tensor& codebook_partition_sizes,
+                        const std::optional<torch::Tensor>& bias);
 
-torch::Tensor awq_dequantize(
-    torch::Tensor _kernel,
-    torch::Tensor _scaling_factors,
-    torch::Tensor _zeros,
-    int split_k_iters,
-    int thx,
-    int thy);
+torch::Tensor aqlm_dequant(const torch::Tensor& codes,
+                           const torch::Tensor& codebooks,
+                           const torch::Tensor& codebook_partition_sizes);
+
+torch::Tensor awq_gemm(torch::Tensor _in_feats, torch::Tensor _kernel,
+                       torch::Tensor _scaling_factors, torch::Tensor _zeros,
+                       int split_k_iters);
+
+torch::Tensor awq_dequantize(torch::Tensor _kernel,
+                             torch::Tensor _scaling_factors,
+                             torch::Tensor _zeros, int split_k_iters, int thx,
+                             int thy);
+
+torch::Tensor marlin_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
+                          torch::Tensor& b_scales, torch::Tensor& workspace,
+                          int64_t size_m, int64_t size_n, int64_t size_k);
+
+torch::Tensor gptq_marlin_24_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
+                                  torch::Tensor& b_meta,
+                                  torch::Tensor& b_scales,
+                                  torch::Tensor& workspace, int64_t num_bits,
+                                  int64_t size_m, int64_t size_n,
+                                  int64_t size_k);
+
+torch::Tensor gptq_marlin_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
+                               torch::Tensor& b_scales, torch::Tensor& g_idx,
+                               torch::Tensor& perm, torch::Tensor& workspace,
+                               int64_t num_bits, int64_t size_m, int64_t size_n,
+                               int64_t size_k, bool is_k_full);
+
+torch::Tensor gptq_marlin_repack(torch::Tensor& b_q_weight, torch::Tensor& perm,
+                                 int64_t size_k, int64_t size_n,
+                                 int64_t num_bits);
+
+int cutlass_scaled_mm_dq(torch::Tensor& out, torch::Tensor const& a,
+                         torch::Tensor const& b, torch::Tensor const& a_scales,
+                         torch::Tensor const& b_scales);
 
-torch::Tensor marlin_gemm(
-    torch::Tensor& a, 
-    torch::Tensor& b_q_weight,
-    torch::Tensor& b_scales, 
-    torch::Tensor& workspace,
-    int64_t size_m, 
-    int64_t size_n, 
-    int64_t size_k);
 #endif
 
-void squeezellm_gemm(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor lookup_table);
+void static_scaled_int8_quant(torch::Tensor& out, torch::Tensor& input,
+                              float scale);
 
-torch::Tensor gptq_gemm(
-  torch::Tensor a,
-  torch::Tensor b_q_weight,
-  torch::Tensor b_gptq_qzeros,
-  torch::Tensor b_gptq_scales,
-  torch::Tensor b_g_idx,
-  bool use_exllama,
-  int bit);
+void squeezellm_gemm(torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
+                     torch::Tensor lookup_table);
 
-void gptq_shuffle(
-  torch::Tensor q_weight,
-  torch::Tensor q_perm,
-  int bit);
+torch::Tensor gptq_gemm(torch::Tensor a, torch::Tensor b_q_weight,
+                        torch::Tensor b_gptq_qzeros,
+                        torch::Tensor b_gptq_scales, torch::Tensor b_g_idx,
+                        bool use_exllama, int bit);
 
-void moe_align_block_size(
-  torch::Tensor topk_ids,
-  int num_experts,
-  int block_size,
-  torch::Tensor sorted_token_ids,
-  torch::Tensor experts_ids,
-  torch::Tensor num_tokens_post_pad);
+void gptq_shuffle(torch::Tensor q_weight, torch::Tensor q_perm, int bit);
+
+void static_scaled_fp8_quant(torch::Tensor& out, torch::Tensor& input,
+                             torch::Tensor& scale);
+
+void dynamic_scaled_fp8_quant(torch::Tensor& out, torch::Tensor& input,
+                              torch::Tensor& scale);
+
+void moe_align_block_size(torch::Tensor topk_ids, int num_experts,
+                          int block_size, torch::Tensor sorted_token_ids,
+                          torch::Tensor experts_ids,
+                          torch::Tensor num_tokens_post_pad);
 
 #ifndef USE_ROCM
 using fptr_t = uint64_t;
-fptr_t init_custom_ar(torch::Tensor &meta, torch::Tensor &rank_data,
-                    const std::vector<std::string> &handles,
-                    const std::vector<int64_t> &offsets, int rank,
-                    bool full_nvlink);
-bool should_custom_ar(torch::Tensor &inp, int max_size, int world_size,
+fptr_t init_custom_ar(torch::Tensor& meta, torch::Tensor& rank_data,
+                      const std::vector<std::string>& handles,
+                      const std::vector<int64_t>& offsets, int rank,
                       bool full_nvlink);
-void all_reduce_reg(fptr_t _fa, torch::Tensor &inp, torch::Tensor &out);
-void all_reduce_unreg(fptr_t _fa, torch::Tensor &inp, torch::Tensor &reg_buffer,
-                      torch::Tensor &out);
+bool should_custom_ar(torch::Tensor& inp, int max_size, int world_size,
+                      bool full_nvlink);
+void all_reduce_reg(fptr_t _fa, torch::Tensor& inp, torch::Tensor& out);
+void all_reduce_unreg(fptr_t _fa, torch::Tensor& inp, torch::Tensor& reg_buffer,
+                      torch::Tensor& out);
 void dispose(fptr_t _fa);
 int meta_size();
-void register_buffer(fptr_t _fa, torch::Tensor &t,
-                     const std::vector<std::string> &handles,
-                     const std::vector<int64_t> &offsets);
-std::pair<std::vector<uint8_t>, std::vector<int64_t>> get_graph_buffer_ipc_meta(fptr_t _fa);
-void register_graph_buffers(fptr_t _fa, const std::vector<std::string> &handles,
-                            const std::vector<std::vector<int64_t>> &offsets);
+void register_buffer(fptr_t _fa, torch::Tensor& t,
+                     const std::vector<std::string>& handles,
+                     const std::vector<int64_t>& offsets);
+std::pair<std::vector<uint8_t>, std::vector<int64_t>> get_graph_buffer_ipc_meta(
+    fptr_t _fa);
+void register_graph_buffers(fptr_t _fa, const std::vector<std::string>& handles,
+                            const std::vector<std::vector<int64_t>>& offsets);
 #endif

csrc/pos_encoding_kernels.cu

@@ -7,14 +7,10 @@
 
 namespace vllm {
 
-template<typename scalar_t, bool IS_NEOX>
-inline __device__ void apply_rotary_embedding(
-  scalar_t* __restrict__ arr,
-  const scalar_t* __restrict__ cos_ptr,
-  const scalar_t* __restrict__ sin_ptr,
-  int rot_offset,
-  int embed_dim)
-{
+template <typename scalar_t, bool IS_NEOX>
+inline __device__ void apply_token_rotary_embedding(
+    scalar_t* __restrict__ arr, const scalar_t* __restrict__ cos_ptr,
+    const scalar_t* __restrict__ sin_ptr, int rot_offset, int embed_dim) {
   int x_index, y_index;
   scalar_t cos, sin;
   if (IS_NEOX) {
@@ -37,23 +33,17 @@ inline __device__ void apply_rotary_embedding(
   arr[y_index] = y * cos + x * sin;
 }
 
-template<typename scalar_t, bool IS_NEOX>
-__global__ void rotary_embedding_kernel(
-  const int64_t* __restrict__ positions,        // [batch_size, seq_len] or [num_tokens]
-  scalar_t* __restrict__ query,                 // [batch_size, seq_len, num_heads, head_size] or [num_tokens, num_heads, head_size]
-  scalar_t* __restrict__ key,                   // [batch_size, seq_len, num_kv_heads, head_size] or [num_tokens, num_kv_heads, head_size]
-  const scalar_t* __restrict__ cos_sin_cache,   // [max_position, 2, rot_dim // 2]
-  const int rot_dim,
-  const int64_t query_stride,
-  const int64_t key_stride,
-  const int num_heads,
-  const int num_kv_heads,
-  const int head_size) {
-  // Each thread block is responsible for one token.
-  const int token_idx = blockIdx.x;
-  int64_t pos = positions[token_idx];
-  const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
-
+template <typename scalar_t, bool IS_NEOX>
+inline __device__ void apply_rotary_embedding(
+    scalar_t* __restrict__ query,  // [batch_size, seq_len, num_heads,
+                                   // head_size] or [num_tokens, num_heads,
+                                   // head_size]
+    scalar_t* __restrict__ key,    // [batch_size, seq_len, num_kv_heads,
+                                   // head_size] or [num_tokens, num_kv_heads,
+                                   // head_size]
+    const scalar_t* cache_ptr, const int head_size, const int num_heads,
+    const int num_kv_heads, const int rot_dim, const int token_idx,
+    const int64_t query_stride, const int64_t key_stride) {
   const int embed_dim = rot_dim / 2;
   const scalar_t* cos_ptr = cache_ptr;
   const scalar_t* sin_ptr = cache_ptr + embed_dim;
@@ -63,8 +53,8 @@ __global__ void rotary_embedding_kernel(
     const int head_idx = i / embed_dim;
     const int64_t token_head = token_idx * query_stride + head_idx * head_size;
     const int rot_offset = i % embed_dim;
-    apply_rotary_embedding<scalar_t, IS_NEOX>(query + token_head, cos_ptr,
-                                              sin_ptr, rot_offset, embed_dim);
+    apply_token_rotary_embedding<scalar_t, IS_NEOX>(
+        query + token_head, cos_ptr, sin_ptr, rot_offset, embed_dim);
   }
 
   const int nk = num_kv_heads * embed_dim;
@@ -72,20 +62,74 @@ __global__ void rotary_embedding_kernel(
     const int head_idx = i / embed_dim;
     const int64_t token_head = token_idx * key_stride + head_idx * head_size;
     const int rot_offset = i % embed_dim;
-    apply_rotary_embedding<scalar_t, IS_NEOX>(key + token_head, cos_ptr,
-                                              sin_ptr, rot_offset, embed_dim);
+    apply_token_rotary_embedding<scalar_t, IS_NEOX>(
+        key + token_head, cos_ptr, sin_ptr, rot_offset, embed_dim);
   }
 }
 
-} // namespace vllm
+template <typename scalar_t, bool IS_NEOX>
+__global__ void rotary_embedding_kernel(
+    const int64_t* __restrict__ positions,  // [batch_size, seq_len] or
+                                            // [num_tokens]
+    scalar_t* __restrict__ query,           // [batch_size, seq_len, num_heads,
+                                   // head_size] or [num_tokens, num_heads,
+                                   // head_size]
+    scalar_t* __restrict__ key,  // [batch_size, seq_len, num_kv_heads,
+                                 // head_size] or [num_tokens, num_kv_heads,
+                                 // head_size]
+    const scalar_t* __restrict__ cos_sin_cache,  // [max_position, 2, rot_dim //
+                                                 // 2]
+    const int rot_dim, const int64_t query_stride, const int64_t key_stride,
+    const int num_heads, const int num_kv_heads, const int head_size) {
+  // Each thread block is responsible for one token.
+  const int token_idx = blockIdx.x;
+  int64_t pos = positions[token_idx];
+  const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
+
+  apply_rotary_embedding<scalar_t, IS_NEOX>(
+      query, key, cache_ptr, head_size, num_heads, num_kv_heads, rot_dim,
+      token_idx, query_stride, key_stride);
+}
+
+template <typename scalar_t, bool IS_NEOX>
+__global__ void batched_rotary_embedding_kernel(
+    const int64_t* __restrict__ positions,  // [batch_size, seq_len] or
+                                            // [num_tokens]
+    scalar_t* __restrict__ query,           // [batch_size, seq_len, num_heads,
+                                   // head_size] or [num_tokens, num_heads,
+                                   // head_size]
+    scalar_t* __restrict__ key,  // [batch_size, seq_len, num_kv_heads,
+                                 // head_size] or [num_tokens, num_kv_heads,
+                                 // head_size]
+    const scalar_t* __restrict__ cos_sin_cache,  // [max_position, 2, rot_dim //
+                                                 // 2]
+    const int64_t* __restrict__ cos_sin_cache_offsets,  // [batch_size, seq_len]
+                                                        // or [num_tokens]
+    const int rot_dim, const int64_t query_stride, const int64_t key_stride,
+    const int num_heads, const int num_kv_heads, const int head_size) {
+  // Each thread block is responsible for one token.
+  const int token_idx = blockIdx.x;
+  int64_t pos = positions[token_idx];
+  int64_t cos_sin_cache_offset = cos_sin_cache_offsets[token_idx];
+  const scalar_t* cache_ptr =
+      cos_sin_cache + (cos_sin_cache_offset + pos) * rot_dim;
+
+  apply_rotary_embedding<scalar_t, IS_NEOX>(
+      query, key, cache_ptr, head_size, num_heads, num_kv_heads, rot_dim,
+      token_idx, query_stride, key_stride);
+}
+
+}  // namespace vllm
 
 void rotary_embedding(
-  torch::Tensor& positions,         // [batch_size, seq_len] or [num_tokens]
-  torch::Tensor& query,             // [batch_size, seq_len, num_heads * head_size] or [num_tokens, num_heads * head_size]
-  torch::Tensor& key,               // [batch_size, seq_len, num_kv_heads * head_size] or [num_tokens, num_kv_heads * head_size]
-  int head_size,
-  torch::Tensor& cos_sin_cache,     // [max_position, rot_dim]
-  bool is_neox) {
+    torch::Tensor& positions,  // [batch_size, seq_len] or [num_tokens]
+    torch::Tensor& query,  // [batch_size, seq_len, num_heads * head_size] or
+                           // [num_tokens, num_heads * head_size]
+    torch::Tensor& key,    // [batch_size, seq_len, num_kv_heads * head_size] or
+                           // [num_tokens, num_kv_heads * head_size]
+    int head_size,
+    torch::Tensor& cos_sin_cache,  // [max_position, rot_dim]
+    bool is_neox) {
   int64_t num_tokens = query.numel() / query.size(-1);
   int rot_dim = cos_sin_cache.size(1);
   int num_heads = query.size(-1) / head_size;
@@ -97,34 +141,63 @@ void rotary_embedding(
   dim3 block(std::min(num_heads * rot_dim / 2, 512));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  VLLM_DISPATCH_FLOATING_TYPES(
-    query.scalar_type(),
-    "rotary_embedding",
-    [&] {
-      if (is_neox) {
-        vllm::rotary_embedding_kernel<scalar_t, true><<<grid, block, 0, stream>>>(
-          positions.data_ptr<int64_t>(),
-          query.data_ptr<scalar_t>(),
-          key.data_ptr<scalar_t>(),
-          cos_sin_cache.data_ptr<scalar_t>(),
-          rot_dim,
-          query_stride,
-          key_stride,
-          num_heads,
-          num_kv_heads,
-          head_size);
-      } else {
-        vllm::rotary_embedding_kernel<scalar_t, false><<<grid, block, 0, stream>>>(
-          positions.data_ptr<int64_t>(),
-          query.data_ptr<scalar_t>(),
-          key.data_ptr<scalar_t>(),
-          cos_sin_cache.data_ptr<scalar_t>(),
-          rot_dim,
-          query_stride,
-          key_stride,
-          num_heads,
-          num_kv_heads,
-          head_size);
-      }
-    });
+  VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "rotary_embedding", [&] {
+    if (is_neox) {
+      vllm::rotary_embedding_kernel<scalar_t, true><<<grid, block, 0, stream>>>(
+          positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
+          key.data_ptr<scalar_t>(), cos_sin_cache.data_ptr<scalar_t>(), rot_dim,
+          query_stride, key_stride, num_heads, num_kv_heads, head_size);
+    } else {
+      vllm::rotary_embedding_kernel<scalar_t, false>
+          <<<grid, block, 0, stream>>>(
+              positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
+              key.data_ptr<scalar_t>(), cos_sin_cache.data_ptr<scalar_t>(),
+              rot_dim, query_stride, key_stride, num_heads, num_kv_heads,
+              head_size);
+    }
+  });
+}
+
+/*
+Batched version of rotary embedding, pack multiple LoRAs together
+and process in batched manner.
+*/
+void batched_rotary_embedding(
+    torch::Tensor& positions,  // [batch_size, seq_len] or [num_tokens]
+    torch::Tensor& query,  // [batch_size, seq_len, num_heads * head_size] or
+                           // [num_tokens, num_heads * head_size]
+    torch::Tensor& key,    // [batch_size, seq_len, num_kv_heads * head_size] or
+                           // [num_tokens, num_kv_heads * head_size]
+    int head_size,
+    torch::Tensor& cos_sin_cache,  // [max_position, rot_dim]
+    bool is_neox, int rot_dim,
+    torch::Tensor& cos_sin_cache_offsets  // [num_tokens]
+) {
+  int64_t num_tokens = cos_sin_cache_offsets.size(0);
+  int num_heads = query.size(-1) / head_size;
+  int num_kv_heads = key.size(-1) / head_size;
+  int64_t query_stride = query.stride(-2);
+  int64_t key_stride = key.stride(-2);
+
+  dim3 grid(num_tokens);
+  dim3 block(std::min(num_heads * rot_dim / 2, 512));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(query));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  VLLM_DISPATCH_FLOATING_TYPES(query.scalar_type(), "rotary_embedding", [&] {
+    if (is_neox) {
+      vllm::batched_rotary_embedding_kernel<scalar_t, true>
+          <<<grid, block, 0, stream>>>(
+              positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
+              key.data_ptr<scalar_t>(), cos_sin_cache.data_ptr<scalar_t>(),
+              cos_sin_cache_offsets.data_ptr<int64_t>(), rot_dim, query_stride,
+              key_stride, num_heads, num_kv_heads, head_size);
+    } else {
+      vllm::batched_rotary_embedding_kernel<scalar_t, false>
+          <<<grid, block, 0, stream>>>(
+              positions.data_ptr<int64_t>(), query.data_ptr<scalar_t>(),
+              key.data_ptr<scalar_t>(), cos_sin_cache.data_ptr<scalar_t>(),
+              cos_sin_cache_offsets.data_ptr<int64_t>(), rot_dim, query_stride,
+              key_stride, num_heads, num_kv_heads, head_size);
+    }
+  });
 }

csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu

@@ -2,3 +2,4 @@
 #include "bgmv_impl.cuh"
 
 FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_bfloat16, nv_bfloat16)
+FOR_INST_BGMV_WIDE_NARROW(INST_BGMV_ONESIDE, nv_bfloat16, nv_bfloat16, nv_bfloat16)

csrc/punica/bgmv/bgmv_bf16_bf16_fp16.cu

@@ -1,4 +0,0 @@
-#include "bgmv_config.h"
-#include "bgmv_impl.cuh"
-
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_bfloat16, nv_half)

csrc/punica/bgmv/bgmv_bf16_fp16_bf16.cu

@@ -1,4 +0,0 @@
-#include "bgmv_config.h"
-#include "bgmv_impl.cuh"
-
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_half, nv_bfloat16)

csrc/punica/bgmv/bgmv_bf16_fp16_fp16.cu

@@ -1,4 +0,0 @@
-#include "bgmv_config.h"
-#include "bgmv_impl.cuh"
-
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_half, nv_half)

csrc/punica/bgmv/bgmv_bf16_fp32_bf16.cu

@@ -2,3 +2,4 @@
 #include "bgmv_impl.cuh"
 
 FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, float, nv_bfloat16)
+FOR_INST_BGMV_WIDE_NARROW(INST_BGMV_ONESIDE, nv_bfloat16, float, nv_bfloat16)

csrc/punica/bgmv/bgmv_bf16_fp32_fp16.cu

@@ -1,4 +0,0 @@
-#include "bgmv_config.h"
-#include "bgmv_impl.cuh"
-
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, float, nv_half)

csrc/punica/bgmv/bgmv_config.h

@@ -14,21 +14,31 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 128) \
     f(in_T, out_T, W_T, narrow, 256) \
     f(in_T, out_T, W_T, narrow, 512) \
+    f(in_T, out_T, W_T, narrow, 640) \
+    f(in_T, out_T, W_T, narrow, 768) \
     f(in_T, out_T, W_T, narrow, 1024) \
+    f(in_T, out_T, W_T, narrow, 1152) \
     f(in_T, out_T, W_T, narrow, 1280) \
+    f(in_T, out_T, W_T, narrow, 1536) \
     f(in_T, out_T, W_T, narrow, 1728) \
     f(in_T, out_T, W_T, narrow, 1792) \
     f(in_T, out_T, W_T, narrow, 2048) \
+    f(in_T, out_T, W_T, narrow, 2304) \
     f(in_T, out_T, W_T, narrow, 2560) \
     f(in_T, out_T, W_T, narrow, 2752) \
+    f(in_T, out_T, W_T, narrow, 2816) \
     f(in_T, out_T, W_T, narrow, 3072) \
+    f(in_T, out_T, W_T, narrow, 3328) \
     f(in_T, out_T, W_T, narrow, 3456) \
     f(in_T, out_T, W_T, narrow, 3584) \
     f(in_T, out_T, W_T, narrow, 4096) \
+    f(in_T, out_T, W_T, narrow, 4608) \
     f(in_T, out_T, W_T, narrow, 5120) \
     f(in_T, out_T, W_T, narrow, 5504) \
     f(in_T, out_T, W_T, narrow, 5632) \
     f(in_T, out_T, W_T, narrow, 6144) \
+    f(in_T, out_T, W_T, narrow, 6400) \
+    f(in_T, out_T, W_T, narrow, 6848) \
     f(in_T, out_T, W_T, narrow, 6912) \
     f(in_T, out_T, W_T, narrow, 7168) \
     f(in_T, out_T, W_T, narrow, 8192) \
@@ -36,11 +46,16 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 10240) \
     f(in_T, out_T, W_T, narrow, 11008) \
     f(in_T, out_T, W_T, narrow, 12288) \
+    f(in_T, out_T, W_T, narrow, 13696) \
     f(in_T, out_T, W_T, narrow, 13824) \
     f(in_T, out_T, W_T, narrow, 14336) \
+    f(in_T, out_T, W_T, narrow, 15360) \
     f(in_T, out_T, W_T, narrow, 16384) \
     f(in_T, out_T, W_T, narrow, 20480) \
+    f(in_T, out_T, W_T, narrow, 22016) \
     f(in_T, out_T, W_T, narrow, 24576) \
+    f(in_T, out_T, W_T, narrow, 27392) \
+    f(in_T, out_T, W_T, narrow, 27648) \
     f(in_T, out_T, W_T, narrow, 28672) \
     f(in_T, out_T, W_T, narrow, 32000) \
     f(in_T, out_T, W_T, narrow, 32256) \
@@ -48,9 +63,91 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 32768) \
     f(in_T, out_T, W_T, narrow, 33024) \
     f(in_T, out_T, W_T, narrow, 36864) \
+    f(in_T, out_T, W_T, narrow, 43264) \
     f(in_T, out_T, W_T, narrow, 49152) \
+    f(in_T, out_T, W_T, narrow, 64000) \
+    f(in_T, out_T, W_T, narrow, 64256) \
+    f(in_T, out_T, W_T, narrow, 64512) \
+    f(in_T, out_T, W_T, narrow, 102400) \
+    f(in_T, out_T, W_T, narrow, 102656) \
+    f(in_T, out_T, W_T, narrow, 102912) \
+    f(in_T, out_T, W_T, narrow, 128000) \
+    f(in_T, out_T, W_T, narrow, 128256) \
+    f(in_T, out_T, W_T, narrow, 128512) \
+// Keep above in sync with vllm/lora/layers::LogitsProcessorWithLoRA
+// and vllm/tests/lora/test_punica.py
+
+// Used for defining kernels going from the variety of 
+// dim in to the narrow dim out
+    // Using it for the fully sharded column 
+    // parallel LoRA A which splits the rank dim
+#define FOR_INST_BGMV_NARROW(f, in_T, out_T, W_T, narrow) \
+    f(in_T, out_T, W_T, 128, narrow) \
+    f(in_T, out_T, W_T, 256, narrow) \
+    f(in_T, out_T, W_T, 512, narrow) \
+    f(in_T, out_T, W_T, 640, narrow) \
+    f(in_T, out_T, W_T, 768, narrow) \
+    f(in_T, out_T, W_T, 1024, narrow) \
+    f(in_T, out_T, W_T, 1152, narrow) \
+    f(in_T, out_T, W_T, 1280, narrow) \
+    f(in_T, out_T, W_T, 1536, narrow) \
+    f(in_T, out_T, W_T, 1728, narrow) \
+    f(in_T, out_T, W_T, 1792, narrow) \
+    f(in_T, out_T, W_T, 2048, narrow) \
+    f(in_T, out_T, W_T, 2304, narrow) \
+    f(in_T, out_T, W_T, 2560, narrow) \
+    f(in_T, out_T, W_T, 2752, narrow) \
+    f(in_T, out_T, W_T, 2816, narrow) \
+    f(in_T, out_T, W_T, 3072, narrow) \
+    f(in_T, out_T, W_T, 3328, narrow) \
+    f(in_T, out_T, W_T, 3456, narrow) \
+    f(in_T, out_T, W_T, 3584, narrow) \
+    f(in_T, out_T, W_T, 4096, narrow) \
+    f(in_T, out_T, W_T, 4608, narrow) \
+    f(in_T, out_T, W_T, 5120, narrow) \
+    f(in_T, out_T, W_T, 5504, narrow) \
+    f(in_T, out_T, W_T, 5632, narrow) \
+    f(in_T, out_T, W_T, 6144, narrow) \
+    f(in_T, out_T, W_T, 6400, narrow) \
+    f(in_T, out_T, W_T, 6848, narrow) \
+    f(in_T, out_T, W_T, 6912, narrow) \
+    f(in_T, out_T, W_T, 7168, narrow) \
+    f(in_T, out_T, W_T, 8192, narrow) \
+    f(in_T, out_T, W_T, 9216, narrow) \
+    f(in_T, out_T, W_T, 10240, narrow) \
+    f(in_T, out_T, W_T, 11008, narrow) \
+    f(in_T, out_T, W_T, 12288, narrow) \
+    f(in_T, out_T, W_T, 13696, narrow) \
+    f(in_T, out_T, W_T, 13824, narrow) \
+    f(in_T, out_T, W_T, 14336, narrow) \
+    f(in_T, out_T, W_T, 15360, narrow) \
+    f(in_T, out_T, W_T, 16384, narrow) \
+    f(in_T, out_T, W_T, 20480, narrow) \
+    f(in_T, out_T, W_T, 22016, narrow) \
+    f(in_T, out_T, W_T, 24576, narrow) \
+    f(in_T, out_T, W_T, 27392, narrow) \
+    f(in_T, out_T, W_T, 27648, narrow) \
+    f(in_T, out_T, W_T, 28672, narrow) \
+    f(in_T, out_T, W_T, 32000, narrow) \
+    f(in_T, out_T, W_T, 32256, narrow) \
+    f(in_T, out_T, W_T, 32512, narrow) \
+    f(in_T, out_T, W_T, 32768, narrow) \
+    f(in_T, out_T, W_T, 33024, narrow) \
+    f(in_T, out_T, W_T, 36864, narrow) \
+    f(in_T, out_T, W_T, 43264, narrow) \
+    f(in_T, out_T, W_T, 49152, narrow) \
+    f(in_T, out_T, W_T, 64000, narrow) \
+    f(in_T, out_T, W_T, 64256, narrow) \
+    f(in_T, out_T, W_T, 64512, narrow) \
+    f(in_T, out_T, W_T, 102400, narrow) \
+    f(in_T, out_T, W_T, 102656, narrow) \
+    f(in_T, out_T, W_T, 102912, narrow) \
+    f(in_T, out_T, W_T, 128000, narrow) \
+    f(in_T, out_T, W_T, 128256, narrow) \
+    f(in_T, out_T, W_T, 128512, narrow) \
 // Keep above in sync with vllm/lora/layers::SamplerWithLoRA
 
+
 // Keep this in sync with vllm/config::LoRAConfig
 #define FOR_BGMV_WIDE_NARROW(f, in_T, out_T, W_T) \
     FOR_BGMV_WIDE(f, in_T, out_T, W_T, 8)  \
@@ -58,4 +155,14 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     FOR_BGMV_WIDE(f, in_T, out_T, W_T, 32) \
     FOR_BGMV_WIDE(f, in_T, out_T, W_T, 64)
 
+
+#define FOR_INST_BGMV_WIDE_NARROW(f, in_T, out_T, W_T) \
+    FOR_INST_BGMV_NARROW(f, in_T, out_T, W_T, 1) \
+    FOR_INST_BGMV_NARROW(f, in_T, out_T, W_T, 2) \
+    FOR_INST_BGMV_NARROW(f, in_T, out_T, W_T, 4) \
+    f(in_T, out_T, W_T, 8, 64) \
+    f(in_T, out_T, W_T, 16, 64) \
+    f(in_T, out_T, W_T, 32, 64) \
+    f(in_T, out_T, W_T, 64, 64)
+
 // clang-format on

csrc/punica/bgmv/bgmv_fp16_bf16_bf16.cu

@@ -1,4 +0,0 @@
-#include "bgmv_config.h"
-#include "bgmv_impl.cuh"
-
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_bfloat16, nv_bfloat16)

csrc/punica/bgmv/bgmv_fp16_bf16_fp16.cu

@@ -1,4 +0,0 @@
-#include "bgmv_config.h"
-#include "bgmv_impl.cuh"
-
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_bfloat16, nv_half)

csrc/punica/bgmv/bgmv_fp16_fp16_bf16.cu

@@ -1,4 +0,0 @@
-#include "bgmv_config.h"
-#include "bgmv_impl.cuh"
-
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_half, nv_bfloat16)

csrc/punica/bgmv/bgmv_fp16_fp16_fp16.cu

@@ -2,3 +2,4 @@
 #include "bgmv_impl.cuh"
 
 FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_half, nv_half)
+FOR_INST_BGMV_WIDE_NARROW(INST_BGMV_ONESIDE, nv_half, nv_half, nv_half)

csrc/punica/bgmv/bgmv_fp16_fp32_bf16.cu

@@ -1,4 +0,0 @@
-#include "bgmv_config.h"
-#include "bgmv_impl.cuh"
-
-FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, float, nv_bfloat16)

csrc/punica/bgmv/bgmv_fp16_fp32_fp16.cu

@@ -2,3 +2,4 @@
 #include "bgmv_impl.cuh"
 
 FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, float, nv_half)
+FOR_INST_BGMV_WIDE_NARROW(INST_BGMV_ONESIDE, nv_half, float, nv_half)