[Misc] Reduce supported Punica dtypes (#4304)

Fixes fp8 iterface which broke in AQLM merge.

.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 +1,38 @@
+# This script build the ROCm docker image and run the API server inside the container.
+# It serves a sanity check for compilation and basic model usage.
+set -ex
+
+# Print ROCm version
+rocminfo
+
+# Try building the docker image
+docker build -t rocm -f Dockerfile.rocm .
+
+# Setup cleanup
+remove_docker_container() { docker rm -f rocm || true; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Run the image
+docker run --device /dev/kfd --device /dev/dri --network host --name rocm rocm python3 -m vllm.entrypoints.api_server &
+
+# 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/run-benchmarks.sh

@@ -6,15 +6,16 @@ set -o pipefail
 # cd into parent directory of this file
 cd "$(dirname "${BASH_SOURCE[0]}")/.."
 
-(wget && curl) || (apt-get update && apt-get install -y wget curl)
+(which wget && which curl) || (apt-get update && apt-get install -y wget curl)
 
-# run benchmarks and upload the result to buildkite
+# run python-based benchmarks and upload the result to buildkite
 python3 benchmarks/benchmark_latency.py 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
 bench_throughput_exit_code=$?
 
+# run server-based benchmarks and upload the result to buildkite
 python3 -m vllm.entrypoints.openai.api_server --model meta-llama/Llama-2-7b-chat-hf &
 server_pid=$!
 wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
@@ -22,11 +23,15 @@ wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/r
 # wait for server to start, timeout after 600 seconds
 timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1
 python3 benchmarks/benchmark_serving.py \
-    --dataset ./ShareGPT_V3_unfiltered_cleaned_split.json \
+    --backend vllm \
+    --dataset-name sharegpt \
+    --dataset-path ./ShareGPT_V3_unfiltered_cleaned_split.json \
     --model meta-llama/Llama-2-7b-chat-hf \
     --num-prompts 20 \
     --endpoint /v1/completions \
-    --tokenizer meta-llama/Llama-2-7b-chat-hf 2>&1 | tee benchmark_serving.txt
+    --tokenizer meta-llama/Llama-2-7b-chat-hf \
+    --save-result \
+    2>&1 | tee benchmark_serving.txt
 bench_serving_exit_code=$?
 kill $server_pid
 
@@ -44,7 +49,9 @@ 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 5 benchmark_serving.txt >> benchmark_results.md # last 5 lines
+echo '```' >> benchmark_results.md
+tail -n 20 benchmark_serving.txt >> benchmark_results.md # last 20 lines
+echo '```' >> benchmark_results.md
 
 # upload the results to buildkite
 /workspace/buildkite-agent annotate --style "info" --context "benchmark-results" < benchmark_results.md
@@ -61,3 +68,5 @@ fi
 if [ $bench_serving_exit_code -ne 0 ]; then
     exit $bench_serving_exit_code
 fi
+
+/workspace/buildkite-agent artifact upload openai-*.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 examples/offline_inference.py

.buildkite/run-neuron-test.sh

@@ -0,0 +1,37 @@
+# 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
+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

@@ -11,41 +11,103 @@ steps:
 - label: AsyncEngine Test
   command: pytest -v -s async_engine
 
-- label: Distributed Test
+- label: Basic Correctness Test
+  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
+
+- label: Core Test
+  command: pytest -v -s core
+
+- label: Distributed Comm Ops Test
   command: pytest -v -s test_comm_ops.py
   working_dir: "/vllm-workspace/tests/distributed"
   num_gpus: 2 # only support 1 or 2 for now.
 
+- label: Distributed Tests
+  working_dir: "/vllm-workspace/tests/distributed"
+  num_gpus: 2 # only support 1 or 2 for now.
+  commands:
+  - pytest -v -s test_pynccl.py
+  - pytest -v -s test_pynccl_library.py
+  - TEST_DIST_MODEL=facebook/opt-125m pytest -v -s test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf pytest -v -s test_basic_distributed_correctness.py
+  - TEST_DIST_MODEL=facebook/opt-125m pytest -v -s test_chunked_prefill_distributed.py
+  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf pytest -v -s test_chunked_prefill_distributed.py
+
 - label: Engine Test
-  command: pytest -v -s engine
+  command: pytest -v -s engine tokenization test_sequence.py test_config.py test_logger.py
 
 - label: Entrypoints Test
-  command: pytest -v -s entrypoints
+  commands:
+  # these tests have to be separated, because each one will allocate all posible GPU memory
+  - pytest -v -s entrypoints --ignore=entrypoints/test_server_oot_registration.py
+  - pytest -v -s entrypoints/test_server_oot_registration.py
 
-- label: Kernels Test
-  command: pytest -v -s kernels
-  soft_fail: true
+- label: Examples Test
+  working_dir: "/vllm-workspace/examples"
+  commands:
+    # install aws cli for llava_example.py
+    - pip install awscli
+    - python3 offline_inference.py
+    - python3 offline_inference_with_prefix.py
+    - python3 llm_engine_example.py
+    - python3 llava_example.py
+
+- label: Kernels Test %N
+  command: pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 4
 
 - label: Models Test
   commands:
-    - pytest -v -s models --forked
-  soft_fail: true
+    - bash ../.buildkite/download-images.sh
+    - pytest -v -s models --ignore=models/test_llava.py --ignore=models/test_mistral.py
+
+- label: Llava Test
+  commands:
+    - bash ../.buildkite/download-images.sh
+    - pytest -v -s models/test_llava.py
 
 - label: Prefix Caching Test
   commands:
     - pytest -v -s prefix_caching
 
 - label: Samplers Test
-  command: pytest -v -s samplers --forked
+  command: pytest -v -s samplers
+
+- label: LogitsProcessor Test
+  command: pytest -v -s test_logits_processor.py
 
 - label: Worker Test
   command: pytest -v -s worker
 
-- label: LoRA Test
-  command: pytest -v -s lora
+- label: Speculative decoding tests
+  command: pytest -v -s spec_decode
+
+- label: LoRA Test %N
+  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  parallelism: 4
+
+- label: Tensorizer Test
+  command: apt-get install curl libsodium23 && pytest -v -s tensorizer_loader
+
+- label: Metrics Test
+  command: pytest -v -s metrics
+
+- label: Quantization Test
+  command: pytest -v -s quantization
 
 - label: Benchmarks
   working_dir: "/vllm-workspace/.buildkite"
   commands:
   - pip install aiohttp
   - bash run-benchmarks.sh
+
+- label: Documentation Build
+  working_dir: "/vllm-workspace/test_docs/docs"
+  no_gpu: True
+  commands:
+  - pip install -r requirements-docs.txt
+  - SPHINXOPTS=\"-W\" make html

.buildkite/test-template.j2

@@ -3,6 +3,7 @@
 {% set default_working_dir = "/vllm-workspace/tests" %}
 
 steps:
+
   - label: ":docker: build image"
     commands:
       - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."
@@ -15,11 +16,27 @@ steps:
           limit: 5
   - wait
 
+  - label: "AMD Test"
+    agents:
+      queue: amd
+    command: bash .buildkite/run-amd-test.sh
+
+  - label: "Neuron Test"
+    agents:
+      queue: neuron
+    command: bash .buildkite/run-neuron-test.sh
+
+  - label: "CPU Test"
+    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
@@ -35,14 +52,18 @@ steps:
               - image: "{{ docker_image }}"
                 command: ["bash"]
                 args:
-                - "-c"
+                - '-c'
                 - "'cd {{ (step.working_dir or default_working_dir) | safe  }} && {{ step.command  or (step.commands | join(' && ')) | safe }}'"
+                {% if not step.no_gpu %}
                 resources:
                   requests:
                     nvidia.com/gpu: "{{ step.num_gpus or default_num_gpu }}"
                   limits:
                     nvidia.com/gpu: "{{ step.num_gpus or default_num_gpu }}"
+                {% endif %}
                 env:
+                  - name: VLLM_USAGE_SOURCE
+                    value: ci-test
                   - name: HF_TOKEN
                     valueFrom:
                       secretKeyRef:

.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 +1,84 @@
+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``` ````.
+
+      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/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/mypy.yaml

@@ -0,0 +1,51 @@
+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
+        # TODO(sang): Fix nested dir
+        mypy vllm/core/*.py --follow-imports=skip --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
+        # TODO(sang): Fix nested dir
+        mypy vllm/model_executor/*.py  --config-file pyproject.toml
+        # TODO(sang): Fix nested dir
+        # mypy vllm/lora/*.py --config-file pyproject.toml
+

.github/workflows/publish.yml

@@ -49,13 +49,16 @@ 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: ['2.2.1']  # 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
+
       - name: Set up Linux Env
         if: ${{ runner.os == 'Linux' }}
         run: |

.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 }}
@@ -25,7 +25,13 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install ruff==0.1.5
+        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
+    - 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/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 +1,292 @@
+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 "2.2.1")
+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/fp8/fp8_cuda_kernels.cu"
+  "csrc/cuda_utils_kernels.cu"
+  "csrc/moe_align_block_size_kernels.cu"
+  "csrc/pybind.cpp")
+
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  list(APPEND VLLM_EXT_SRC
+    "csrc/quantization/aqlm/gemm_kernels.cu"
+    "csrc/quantization/awq/gemm_kernels.cu"
+    "csrc/quantization/marlin/marlin_cuda_kernel.cu"
+    "csrc/custom_all_reduce.cu")
+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}
+  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.cc")
+
+#
+# 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}")
+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)
+endif()
+
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  message(STATUS "Enabling moe extension.")
+  add_dependencies(default _moe_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()

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

@@ -2,26 +2,41 @@
 # to run the OpenAI compatible server.
 
 #################### BASE BUILD IMAGE ####################
+# prepare basic build environment
 FROM nvidia/cuda:12.1.0-devel-ubuntu22.04 AS dev
 
 RUN apt-get update -y \
     && apt-get install -y python3-pip git
 
+# 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-12.1/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
@@ -29,16 +44,19 @@ 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}
@@ -48,50 +66,90 @@ 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
+
+# the `vllm_nccl` package must be installed from source distribution
+# pip is too smart to store a wheel in the cache, and other CI jobs
+# will directly use the wheel from the cache, which is not what we want.
+# we need to remove it manually
+RUN --mount=type=cache,target=/root/.cache/pip \
+    pip cache remove vllm_nccl*
 #################### EXTENSION Build IMAGE ####################
 
+#################### FLASH_ATTENTION Build IMAGE ####################
+FROM dev as flash-attn-builder
+# max jobs used for build
+ARG max_jobs=2
+ENV MAX_JOBS=${max_jobs}
+# flash attention version
+ARG flash_attn_version=v2.5.6
+ENV FLASH_ATTN_VERSION=${flash_attn_version}
+
+WORKDIR /usr/src/flash-attention-v2
+
+# Download the wheel or build it if a pre-compiled release doesn't exist
+RUN pip --verbose wheel flash-attn==${FLASH_ATTN_VERSION} \
+    --no-build-isolation --no-deps --no-cache-dir
+
+#################### FLASH_ATTENTION Build IMAGE ####################
+
+#################### vLLM installation IMAGE ####################
+# image with vLLM installed
+FROM nvidia/cuda:12.1.0-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-12.1/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
+
+RUN --mount=type=bind,from=flash-attn-builder,src=/usr/src/flash-attention-v2,target=/usr/src/flash-attention-v2 \
+    --mount=type=cache,target=/root/.cache/pip \
+    pip install /usr/src/flash-attention-v2/*.whl --no-cache-dir
+#################### 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 ####################
-# use CUDA base as CUDA runtime dependencies are already installed via pip
-FROM nvidia/cuda:12.1.0-base-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 +1,20 @@
+# 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
+
+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

@@ -10,16 +10,22 @@ RUN echo "Base image is $BASE_IMAGE"
 # BASE_IMAGE for ROCm_5.7: "rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1"
 # BASE_IMAGE for ROCm_6.0: "rocm/pytorch:rocm6.0_ubuntu20.04_py3.9_pytorch_2.1.1"
 
-# this does not always work for all rocm versions
-RUN LLVM_GFX_ARCH=$(/opt/rocm/llvm/bin/amdgpu-offload-arch) && \
-    echo "LLVM_GFX_ARCH is $LLVM_GFX_ARCH"
 
 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
+# if 0, will not build flash attention
+# this is useful for gfx target where flash-attention is not supported
+# 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
 
@@ -53,9 +59,10 @@ ENV LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/opt/rocm/lib/:/libtorch/lib:
 ENV CPLUS_INCLUDE_PATH=$CPLUS_INCLUDE_PATH:/libtorch/include:/libtorch/include/torch/csrc/api/include/:/opt/rocm/include/:
 
 # Install ROCm flash-attention
-RUN mkdir libs \
+RUN if [ "$BUILD_FA" = "1" ]; then \
+    mkdir libs \
     && cd libs \
-    && git clone https://github.com/ROCmSoftwarePlatform/flash-attention.git \
+    && git clone https://github.com/ROCm/flash-attention.git \
     && cd flash-attention \
     && git checkout ${FA_BRANCH} \
     && git submodule update --init \
@@ -63,26 +70,37 @@ RUN mkdir libs \
     && if [ "$BASE_IMAGE" = "rocm/pytorch:rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1" ]; then \
         patch /opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/utils/hipify/hipify_python.py hipify_patch.patch; fi \
     && python3 setup.py install \
-    && cd ..
-
-COPY ./ /app/vllm
-
-RUN python3 -m pip install --upgrade pip
-RUN python3 -m pip install xformers==0.0.23 --no-deps
+    && cd ..; \
+    fi
 
 # 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
 
+# 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
+
+COPY ./ /app/vllm
+
+RUN python3 -m pip install --upgrade pip numba
+
 RUN cd /app \
     && cd vllm \
     && pip install -U -r requirements-rocm.txt \
-    && bash patch_xformers.rocm.sh \
+    && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h /app/vllm/rocm_patch/rocm_bf16.patch \
     && python3 setup.py install \
     && cd ..
 
 RUN python3 -m pip install --upgrade pip
-RUN python3 -m pip install --no-cache-dir ray[all]
+RUN python3 -m pip install --no-cache-dir ray[all]==2.9.3
 
 CMD ["/bin/bash"]

MANIFEST.in

@@ -1,4 +1,7 @@
 include LICENSE
-include requirements.txt
+include requirements-common.txt
+include requirements-cuda.txt
+include CMakeLists.txt
 
+recursive-include cmake *
 recursive-include csrc *

README.md

@@ -14,18 +14,9 @@ Easy, fast, and cheap LLM serving for everyone
 
 </p>
 
----
-
-**The Second vLLM Bay Area Meetup (Jan 31st 5pm-7:30pm PT)**
-
-We are thrilled to announce our second vLLM Meetup!
-The vLLM team will share recent updates and roadmap.
-We will also have vLLM collaborators from IBM coming up to the stage to discuss their insights on LLM optimizations.
-Please register [here](https://lu.ma/ygxbpzhl) 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.
 - [2023/10] We hosted [the first vLLM meetup](https://lu.ma/first-vllm-meetup) in SF! Please find the meetup slides [here](https://docs.google.com/presentation/d/1QL-XPFXiFpDBh86DbEegFXBXFXjix4v032GhShbKf3s/edit?usp=sharing).
@@ -66,22 +57,33 @@ vLLM seamlessly supports many Hugging Face models, including the following archi
 - 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.)
+- Command-R (`CohereForAI/c4ai-command-r-v01`, etc.)
+- DBRX (`databricks/dbrx-base`, `databricks/dbrx-instruct` 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.)
-- LLaMA & LLaMA-2 (`meta-llama/Llama-2-70b-hf`, `lmsys/vicuna-13b-v1.3`, `young-geng/koala`, `openlm-research/open_llama_13b`, etc.)
+- InternLM2 (`internlm/internlm2-7b`, `internlm/internlm2-chat-7b`, etc.)
+- Jais (`core42/jais-13b`, `core42/jais-13b-chat`, `core42/jais-30b-v3`, `core42/jais-30b-chat-v3`, etc.)
+- LLaMA, Llama 2, and Meta Llama 3 (`meta-llama/Meta-Llama-3-8B-Instruct`, `meta-llama/Meta-Llama-3-70B-Instruct`, `meta-llama/Llama-2-70b-hf`, `lmsys/vicuna-13b-v1.3`, `young-geng/koala`, `openlm-research/open_llama_13b`, etc.)
+- MiniCPM (`openbmb/MiniCPM-2B-sft-bf16`, `openbmb/MiniCPM-2B-dpo-bf16`, 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.)
+- Mixtral (`mistralai/Mixtral-8x7B-v0.1`, `mistralai/Mixtral-8x7B-Instruct-v0.1`, `mistral-community/Mixtral-8x22B-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.)
+- Qwen2 (`Qwen/Qwen1.5-7B`, `Qwen/Qwen1.5-7B-Chat`, etc.)
+- Qwen2MoE (`Qwen/Qwen1.5-MoE-A2.7B`, `Qwen/Qwen1.5-MoE-A2.7B-Chat`, etc.)
 - StableLM(`stabilityai/stablelm-3b-4e1t`, `stabilityai/stablelm-base-alpha-7b-v2`, etc.)
+- Starcoder2(`bigcode/starcoder2-3b`, `bigcode/starcoder2-7b`, `bigcode/starcoder2-15b`, etc.)
+- Xverse (`xverse/XVERSE-7B-Chat`, `xverse/XVERSE-13B-Chat`, `xverse/XVERSE-65B-Chat`, etc.)
 - Yi (`01-ai/Yi-6B`, `01-ai/Yi-34B`, etc.)
 
 Install vLLM with pip or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source):

benchmarks/backend_request_func.py

@@ -0,0 +1,389 @@
+import json
+import os
+import sys
+import time
+import traceback
+from dataclasses import dataclass, field
+from typing import List, Optional
+
+import aiohttp
+from tqdm.asyncio import tqdm
+
+AIOHTTP_TIMEOUT = aiohttp.ClientTimeout(total=6 * 60 * 60)
+
+
+@dataclass
+class RequestFuncInput:
+    prompt: str
+    api_url: str
+    prompt_len: int
+    output_len: int
+    model: str
+    best_of: int = 1
+    use_beam_search: bool = False
+
+
+@dataclass
+class RequestFuncOutput:
+    generated_text: str = ""
+    success: bool = False
+    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(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith("generate_stream")
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert not request_func_input.use_beam_search
+        params = {
+            "best_of": request_func_input.best_of,
+            "max_new_tokens": request_func_input.output_len,
+            "do_sample": True,
+            "temperature": 0.01,  # TGI does not accept 0.0 temperature.
+            "top_p": 0.99,  # TGI does not accept 1.0 top_p.
+        }
+        payload = {
+            "inputs": request_func_input.prompt,
+            "parameters": params,
+        }
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        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 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
+
+                        # 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"]
+        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_trt_llm(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    assert api_url.endswith("generate_stream")
+
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert not request_func_input.use_beam_search
+        assert request_func_input.best_of == 1
+        payload = {
+            "accumulate_tokens": True,
+            "text_input": request_func_input.prompt,
+            "temperature": 0.0,
+            "top_p": 1.0,
+            "max_tokens": request_func_input.output_len,
+            "stream": True,
+        }
+        output = RequestFuncOutput()
+        output.prompt_len = request_func_input.prompt_len
+
+        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 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
+
+                        # 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 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_deepspeed_mii(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
+        assert request_func_input.best_of == 1
+        assert not request_func_input.use_beam_search
+
+        payload = {
+            "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
+
+        # 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 response:
+                if response.status == 200:
+                    parsed_resp = await response.json()
+                    output.latency = time.perf_counter() - st
+                    output.generated_text = parsed_resp["text"][0]
+                    output.success = True
+                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
+
+
+async def async_request_openai_completions(
+    request_func_input: RequestFuncInput,
+    pbar: Optional[tqdm] = None,
+) -> RequestFuncOutput:
+    api_url = request_func_input.api_url
+    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
+        payload = {
+            "model": request_func_input.model,
+            "prompt": request_func_input.prompt,
+            "temperature": 0.0,
+            "best_of": request_func_input.best_of,
+            "max_tokens": request_func_input.output_len,
+            "stream": True,
+        }
+        headers = {
+            "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:
+                            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
+        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_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

@@ -9,6 +9,7 @@ import torch
 from tqdm import tqdm
 
 from vllm import LLM, SamplingParams
+from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
 
 
 def main(args: argparse.Namespace):
@@ -16,16 +17,20 @@ 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,
-    )
+    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,
+              quantization_param_path=args.quantization_param_path,
+              device=args.device,
+              ray_workers_use_nsight=args.ray_workers_use_nsight,
+              enable_chunked_prefill=args.enable_chunked_prefill,
+              download_dir=args.download_dir,
+              block_size=args.block_size)
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -36,7 +41,10 @@ def main(args: argparse.Namespace):
         max_tokens=args.output_len,
     )
     print(sampling_params)
-    dummy_prompt_token_ids = [[0] * args.input_len] * args.batch_size
+    dummy_prompt_token_ids = np.random.randint(10000,
+                                               size=(args.batch_size,
+                                                     args.input_len))
+    dummy_prompt_token_ids = dummy_prompt_token_ids.tolist()
 
     def run_to_completion(profile_dir: Optional[str] = None):
         if profile_dir:
@@ -61,7 +69,8 @@ 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
@@ -70,14 +79,19 @@ def main(args: argparse.Namespace):
                 "."
             ) / "vllm_benchmark_result" / f"latency_result_{time.time()}"
         print(f"Profiling (results will be saved to '{profile_dir}')...")
-        run_to_completion(profile_dir=args.profile_result_dir)
+        run_to_completion(profile_dir=profile_dir)
         return
 
     # Benchmark.
     latencies = []
     for _ in tqdm(range(args.num_iters), desc="Profiling iterations"):
         latencies.append(run_to_completion(profile_dir=None))
+    latencies = np.array(latencies)
+    percentages = [10, 25, 50, 75, 90]
+    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')
 
 
 if __name__ == '__main__':
@@ -88,7 +102,7 @@ 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('--input-len', type=int, default=32)
@@ -99,9 +113,13 @@ 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',
@@ -121,10 +139,23 @@ if __name__ == '__main__':
     parser.add_argument(
         "--kv-cache-dtype",
         type=str,
-        choices=['auto', 'fp8_e5m2'],
+        choices=['auto', 'fp8'],
         default='auto',
         help=
-        'Data type for kv cache storage. If "auto", will use model data type.')
+        'Data type for kv cache storage. If "auto", will use model data type. '
+        '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(
+        '--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',
@@ -135,5 +166,30 @@ if __name__ == '__main__':
         default=None,
         help=('path to save the pytorch profiler output. Can be visualized '
               'with ui.perfetto.dev or Tensorboard.'))
+    parser.add_argument(
+        "--device",
+        type=str,
+        default="cuda",
+        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(
+        "--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')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_prefix_caching.py

@@ -0,0 +1,52 @@
+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="baichuan-inc/Baichuan2-13B-Chat",
+              tokenizer_mode='auto',
+              trust_remote_code=True,
+              enforce_eager=True,
+              enable_prefix_caching=args.enable_prefix_caching)
+
+    num_prompts = 100
+    prompts = [PROMPT] * num_prompts
+    sampling_params = SamplingParams(temperature=0, max_tokens=100)
+
+    print("------warm up------")
+    test_prefix(
+        llm=llm,
+        prompts=prompts[:1],
+        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('--enable-prefix-caching',
+                        action='store_true',
+                        help='enable prefix caching')
+    args = parser.parse_args()
+    main(args)

benchmarks/benchmark_serving.py

@@ -1,38 +1,60 @@
 """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
 
     (TGI backend)
-    ./launch_hf_server.sh <your_model>
+    ./launch_tgi_server.sh <your_model> <max_batch_total_tokens>
 
 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
 """
 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
 
-import aiohttp
 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
 
-# (prompt len, output len, latency)
-REQUEST_LATENCY: List[Tuple[int, int, float]] = []
+
+@dataclass
+class BenchmarkMetrics:
+    completed: int
+    total_input: int
+    total_output: int
+    request_throughput: float
+    input_throughput: float
+    output_throughput: float
+    mean_ttft_ms: float
+    median_ttft_ms: float
+    p99_ttft_ms: float
+    mean_tpot_ms: float
+    median_tpot_ms: float
+    p99_tpot_ms: float
 
 
-def sample_requests(
+def sample_sharegpt_requests(
     dataset_path: str,
     num_requests: int,
     tokenizer: PreTrainedTokenizerBase,
@@ -46,6 +68,11 @@ 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]
+
     # Tokenize the prompts and completions.
     prompts = [prompt for prompt, _ in dataset]
     prompt_token_ids = tokenizer(prompts).input_ids
@@ -75,6 +102,76 @@ def sample_requests(
     return sampled_requests
 
 
+def sample_sonnet_requests(
+    dataset_path: str,
+    num_requests: int,
+    input_len: int,
+    output_len: int,
+    prefix_len: int,
+    tokenizer: PreTrainedTokenizerBase,
+) -> List[Tuple[str, str, int, int]]:
+    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
+
+
 async def get_request(
     input_requests: List[Tuple[str, int, int]],
     request_rate: float,
@@ -92,80 +189,149 @@ async def get_request(
         await asyncio.sleep(interval)
 
 
-async def send_request(backend: str, model: str, api_url: str, prompt: str,
-                       prompt_len: int, output_len: int, best_of: int,
-                       use_beam_search: bool, pbar: tqdm) -> None:
-    request_start_time = time.perf_counter()
+def calculate_metrics(
+    input_requests: List[Tuple[str, int, int]],
+    outputs: List[RequestFuncOutput],
+    dur_s: float,
+    tokenizer: PreTrainedTokenizerBase,
+) -> Tuple[BenchmarkMetrics, List[int]]:
+    actual_output_lens = []
+    total_input = 0
+    completed = 0
+    tpots = []
+    ttfts = []
+    for i in range(len(outputs)):
+        if outputs[i].success:
+            output_len = len(tokenizer(outputs[i].generated_text).input_ids)
+            actual_output_lens.append(output_len)
+            total_input += input_requests[i][1]
+            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)
 
-    headers = {"User-Agent": "Benchmark Client"}
-    if backend == "vllm":
-        pload = {
-            "prompt": prompt,
-            "n": 1,
-            "best_of": best_of,
-            "use_beam_search": use_beam_search,
-            "temperature": 0.0 if use_beam_search else 1.0,
-            "top_p": 1.0,
-            "max_tokens": output_len,
-            "ignore_eos": True,
-            "stream": False,
-        }
-        if model is not None:
-            pload["model"] = model
-    elif backend == "tgi":
-        assert not use_beam_search
-        params = {
-            "best_of": best_of,
-            "max_new_tokens": output_len,
-            "do_sample": True,
-        }
-        pload = {
-            "inputs": prompt,
-            "parameters": params,
-        }
-    else:
-        raise ValueError(f"Unknown backend: {backend}")
+    metrics = BenchmarkMetrics(
+        completed=completed,
+        total_input=total_input,
+        total_output=sum(actual_output_lens),
+        request_throughput=completed / dur_s,
+        input_throughput=total_input / dur_s,
+        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) * 1000,
+        median_tpot_ms=np.median(tpots) * 1000,
+        p99_tpot_ms=np.percentile(tpots, 99) * 1000,
+    )
 
-    timeout = aiohttp.ClientTimeout(total=3 * 3600)
-    async with aiohttp.ClientSession(timeout=timeout) as session:
-        while True:
-            async with session.post(api_url, headers=headers,
-                                    json=pload) as response:
-                chunks = []
-                async for chunk, _ in response.content.iter_chunks():
-                    chunks.append(chunk)
-            output = b"".join(chunks).decode("utf-8")
-            output = json.loads(output)
-
-            # Re-send the request if it failed.
-            if "error" not in output:
-                break
-
-    request_end_time = time.perf_counter()
-    request_latency = request_end_time - request_start_time
-    REQUEST_LATENCY.append((prompt_len, output_len, request_latency))
-    pbar.update(1)
+    return metrics, actual_output_lens
 
 
 async def benchmark(
     backend: str,
-    model: str,
     api_url: str,
+    model_id: str,
+    tokenizer: PreTrainedTokenizerBase,
     input_requests: List[Tuple[str, int, int]],
     best_of: int,
     use_beam_search: bool,
     request_rate: float,
-) -> None:
-    tasks: List[asyncio.Task] = []
-    pbar = tqdm(total=len(input_requests))
+    disable_tqdm: bool,
+):
+    if backend in ASYNC_REQUEST_FUNCS:
+        request_func = ASYNC_REQUEST_FUNCS.get(backend)
+    else:
+        raise ValueError(f"Unknown backend: {backend}")
+
+    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):
         prompt, prompt_len, output_len = request
-        task = asyncio.create_task(
-            send_request(backend, model, api_url, prompt, prompt_len,
-                         output_len, best_of, use_beam_search, pbar))
-        tasks.append(task)
-    await asyncio.gather(*tasks)
-    pbar.close()
+        request_func_input = RequestFuncInput(
+            model=model_id,
+            prompt=prompt,
+            api_url=api_url,
+            prompt_len=prompt_len,
+            output_len=output_len,
+            best_of=best_of,
+            use_beam_search=use_beam_search,
+        )
+        tasks.append(
+            asyncio.create_task(
+                request_func(request_func_input=request_func_input,
+                             pbar=pbar)))
+    outputs: List[RequestFuncOutput] = await asyncio.gather(*tasks)
+
+    if not disable_tqdm:
+        pbar.close()
+
+    benchmark_duration = time.perf_counter() - benchmark_start_time
+
+    metrics, actual_output_lens = calculate_metrics(
+        input_requests=input_requests,
+        outputs=outputs,
+        dur_s=benchmark_duration,
+        tokenizer=tokenizer,
+    )
+
+    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_throughput": metrics.request_throughput,
+        "input_throughput": metrics.input_throughput,
+        "output_throughput": metrics.output_throughput,
+        "mean_ttft_ms": metrics.mean_ttft_ms,
+        "median_ttft_ms": metrics.median_ttft_ms,
+        "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,
+        "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
 
 
 def main(args: argparse.Namespace):
@@ -173,77 +339,252 @@ def main(args: argparse.Namespace):
     random.seed(args.seed)
     np.random.seed(args.seed)
 
-    api_url = f"{args.protocol}://{args.host}:{args.port}{args.endpoint}"
-    tokenizer = get_tokenizer(args.tokenizer,
+    backend = args.backend
+    model_id = args.model
+    tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
+
+    if args.base_url is not None:
+        api_url = f"{args.base_url}{args.endpoint}"
+    else:
+        api_url = f"http://{args.host}:{args.port}{args.endpoint}"
+
+    tokenizer = get_tokenizer(tokenizer_id,
                               trust_remote_code=args.trust_remote_code)
-    input_requests = sample_requests(args.dataset, args.num_prompts, tokenizer)
 
-    benchmark_start_time = time.perf_counter()
-    asyncio.run(
-        benchmark(args.backend, args.model, api_url, input_requests,
-                  args.best_of, args.use_beam_search, args.request_rate))
-    benchmark_end_time = time.perf_counter()
-    benchmark_time = benchmark_end_time - benchmark_start_time
-    print(f"Total time: {benchmark_time:.2f} s")
-    print(f"Throughput: {args.num_prompts / benchmark_time:.2f} requests/s")
+    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,
+        )
 
-    # Compute the latency statistics.
-    avg_latency = np.mean([latency for _, _, latency in REQUEST_LATENCY])
-    print(f"Average latency: {avg_latency:.2f} s")
-    avg_per_token_latency = np.mean([
-        latency / (prompt_len + output_len)
-        for prompt_len, output_len, latency in REQUEST_LATENCY
-    ])
-    print(f"Average latency per token: {avg_per_token_latency:.2f} s")
-    avg_per_output_token_latency = np.mean(
-        [latency / output_len for _, output_len, latency in REQUEST_LATENCY])
-    print("Average latency per output token: "
-          f"{avg_per_output_token_latency:.2f} s")
+    elif args.dataset_name == "sharegpt":
+        input_requests = sample_sharegpt_requests(
+            dataset_path=args.dataset_path,
+            num_requests=args.num_prompts,
+            tokenizer=tokenizer,
+        )
+
+    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(
+            backend=backend,
+            api_url=api_url,
+            model_id=model_id,
+            tokenizer=tokenizer,
+            input_requests=input_requests,
+            best_of=args.best_of,
+            use_beam_search=args.use_beam_search,
+            request_rate=args.request_rate,
+            disable_tqdm=args.disable_tqdm,
+        ))
+
+    # Save config and results to json
+    if args.save_result:
+        result_json = {}
+
+        # Setup
+        current_dt = datetime.now().strftime("%Y%m%d-%H%M%S")
+        result_json["date"] = current_dt
+        result_json["backend"] = backend
+        result_json["model_id"] = model_id
+        result_json["tokenizer_id"] = tokenizer_id
+        result_json["best_of"] = args.best_of
+        result_json["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")
+
+        # Merge with benchmark result
+        result_json = {**result_json, **benchmark_result}
+
+        # Save to file
+        base_model_id = model_id.split("/")[-1]
+        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)
 
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
         description="Benchmark the online serving throughput.")
-    parser.add_argument("--backend",
-                        type=str,
-                        default="vllm",
-                        choices=["vllm", "tgi"])
-    parser.add_argument("--protocol",
-                        type=str,
-                        default="http",
-                        choices=["http", "https"])
+    parser.add_argument(
+        "--backend",
+        type=str,
+        default="vllm",
+        choices=list(ASYNC_REQUEST_FUNCS.keys()),
+    )
+    parser.add_argument(
+        "--base-url",
+        type=str,
+        default=None,
+        help="Server or API base url if not using http host and port.",
+    )
     parser.add_argument("--host", type=str, default="localhost")
     parser.add_argument("--port", type=int, default=8000)
-    parser.add_argument("--endpoint", type=str, default="/generate")
-    parser.add_argument("--model", type=str, default=None)
-    parser.add_argument("--dataset",
+    parser.add_argument(
+        "--endpoint",
+        type=str,
+        default="/v1/completions",
+        help="API endpoint.",
+    )
+    parser.add_argument(
+        "--dataset",
+        type=str,
+        default=None,
+        help="Path to the ShareGPT dataset, will be deprecated in the "
+        "next release.",
+    )
+    parser.add_argument(
+        "--dataset-name",
+        type=str,
+        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("--tokenizer",
-                        type=str,
-                        required=True,
-                        help="Name or path of the tokenizer.")
-    parser.add_argument("--best-of",
-                        type=int,
-                        default=1,
-                        help="Generates `best_of` sequences per prompt and "
-                        "returns the best one.")
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="Name of the model.",
+    )
+    parser.add_argument(
+        "--tokenizer",
+        type=str,
+        help=
+        "Name or path of the tokenizer, if not using the default tokenizer.",
+    )
+    parser.add_argument(
+        "--best-of",
+        type=int,
+        default=1,
+        help="Generates `best_of` sequences per prompt and "
+        "returns the best one.",
+    )
     parser.add_argument("--use-beam-search", action="store_true")
-    parser.add_argument("--num-prompts",
-                        type=int,
-                        default=1000,
-                        help="Number of prompts to process.")
-    parser.add_argument("--request-rate",
-                        type=float,
-                        default=float("inf"),
-                        help="Number of requests per second. If this is inf, "
-                        "then all the requests are sent at time 0. "
-                        "Otherwise, we use Poisson process to synthesize "
-                        "the request arrival times.")
+    parser.add_argument(
+        "--num-prompts",
+        type=int,
+        default=1000,
+        help="Number of prompts to process.",
+    )
+    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,
+        default=float("inf"),
+        help="Number of requests per second. If this is inf, "
+        "then all the requests are sent at time 0. "
+        "Otherwise, we use Poisson process to synthesize "
+        "the request arrival times.",
+    )
     parser.add_argument("--seed", type=int, default=0)
-    parser.add_argument('--trust-remote-code',
-                        action='store_true',
-                        help='trust remote code from huggingface')
+    parser.add_argument(
+        "--trust-remote-code",
+        action="store_true",
+        help="Trust remote code from huggingface",
+    )
+    parser.add_argument(
+        "--disable-tqdm",
+        action="store_true",
+        help="Specify to disable tqdm progress bar.",
+    )
+    parser.add_argument(
+        "--save-result",
+        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(
@@ -72,6 +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(
@@ -83,8 +91,15 @@ 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.
@@ -177,13 +192,15 @@ 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
 
 
@@ -204,12 +221,15 @@ 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)
+        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,
@@ -249,7 +269,7 @@ 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",
@@ -284,16 +304,58 @@ 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",
         type=str,
-        choices=["auto", "fp8_e5m2"],
+        choices=["auto", "fp8"],
         default="auto",
         help=
-        'Data type for kv cache storage. If "auto", will use model data type.')
+        'Data type for kv cache storage. If "auto", will use model data type. '
+        '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(
+        '--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", "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')
     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._C import 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_mixtral_moe.py

@@ -0,0 +1,182 @@
+import json
+import os
+import sys
+
+import torch
+import torch.nn.functional as F
+import triton
+
+from vllm.model_executor.layers.fused_moe import (fused_moe,
+                                                  get_config_file_name)
+
+os.environ['CUDA_VISIBLE_DEVICES'] = '0'
+
+
+def main():
+    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)
+
+
+def run_grid(bs, method):
+    d_model = 4096
+    num_total_experts = 8
+    top_k = 2
+    tp_size = 2
+    model_intermediate_size = 14336
+    num_layers = 32
+    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_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,
+                        })
+
+    best_config = None
+    best_time_us = 1e20
+
+    for config in configs:
+        print(f'{tp_size=} {bs=}')
+        print(f'{config}')
+        # warmup
+        print('warming up')
+        try:
+            for _ in range(num_warmup_trials):
+                run_timing(
+                    num_calls=num_calls,
+                    bs=bs,
+                    d_model=d_model,
+                    num_total_experts=num_total_experts,
+                    top_k=top_k,
+                    tp_size=tp_size,
+                    model_intermediate_size=model_intermediate_size,
+                    method=method,
+                    config=config,
+                )
+        except triton.runtime.autotuner.OutOfResources:
+            continue
+
+        # trial
+        print('benchmarking')
+        for _ in range(num_trials):
+            kernel_dur_ms = run_timing(
+                num_calls=num_calls,
+                bs=bs,
+                d_model=d_model,
+                num_total_experts=num_total_experts,
+                top_k=top_k,
+                tp_size=tp_size,
+                model_intermediate_size=model_intermediate_size,
+                method=method,
+                config=config,
+            )
+
+            kernel_dur_us = 1000 * kernel_dur_ms
+            model_dur_ms = kernel_dur_ms * num_layers
+
+            if kernel_dur_us < best_time_us:
+                best_config = config
+                best_time_us = kernel_dur_us
+
+            print(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)
+
+    # holds Dict[str, Dict[str, int]]
+    filename = get_config_file_name(num_total_experts,
+                                    model_intermediate_size // tp_size)
+    print(f"writing config to file {filename}")
+    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:
+    shard_intermediate_size = model_intermediate_size // tp_size
+
+    hidden_states = torch.rand(
+        (bs, d_model),
+        device="cuda:0",
+        dtype=torch.bfloat16,
+    )
+
+    ws = torch.rand(
+        (num_total_experts, 2 * shard_intermediate_size, d_model),
+        device=hidden_states.device,
+        dtype=hidden_states.dtype,
+    )
+
+    w2s = torch.rand(
+        (num_total_experts, d_model, shard_intermediate_size),
+        device=hidden_states.device,
+        dtype=hidden_states.dtype,
+    )
+
+    gating_output = F.softmax(torch.rand(
+        (num_calls, bs, num_total_experts),
+        device=hidden_states.device,
+        dtype=torch.float32,
+    ),
+                              dim=-1)
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    start_event.record()
+    for i in range(num_calls):
+        hidden_states = method(
+            hidden_states=hidden_states,
+            w1=ws,
+            w2=w2s,
+            gating_output=gating_output[i],
+            topk=2,
+            renormalize=True,
+            inplace=True,
+            override_config=config,
+        )
+    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_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
@@ -25,18 +25,20 @@ def main(
     dtype: torch.dtype,
     seed: int,
     do_profile: bool,
+    device: str = "cuda",
     kv_cache_dtype: Optional[str] = None,
 ) -> None:
     random.seed(seed)
     torch.random.manual_seed(seed)
-    torch.cuda.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed)
 
     scale = float(1.0 / (head_size**0.5))
     query = torch.empty(num_seqs,
                         num_query_heads,
                         head_size,
                         dtype=dtype,
-                        device="cuda")
+                        device=device)
     query.uniform_(-scale, scale)
 
     assert num_query_heads % num_kv_heads == 0
@@ -44,11 +46,11 @@ def main(
     if use_alibi:
         alibi_slopes = torch.randn(num_query_heads,
                                    dtype=torch.float,
-                                   device="cuda")
+                                   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="cuda")
+    context_lens = torch.tensor(context_lens, dtype=torch.int, device=device)
 
     # Create the block tables.
     max_num_blocks_per_seq = (max_context_len + block_size - 1) // block_size
@@ -59,12 +61,17 @@ def main(
             for _ in range(max_num_blocks_per_seq)
         ]
         block_tables.append(block_table)
-    block_tables = torch.tensor(block_tables, dtype=torch.int, device="cuda")
+    block_tables = torch.tensor(block_tables, dtype=torch.int, device=device)
 
     # Create the KV cache.
-    key_caches, value_caches = create_kv_caches_with_random(
-        NUM_BLOCKS, block_size, 1, num_kv_heads, head_size, kv_cache_dtype,
-        dtype)
+    key_caches, value_caches = create_kv_caches_with_random(NUM_BLOCKS,
+                                                            block_size,
+                                                            1,
+                                                            num_kv_heads,
+                                                            head_size,
+                                                            kv_cache_dtype,
+                                                            dtype,
+                                                            device=device)
     key_cache, value_cache = key_caches[0], value_caches[0]
 
     # Prepare for the paged attention kernel.
@@ -84,12 +91,15 @@ def main(
         )
         max_logits = torch.empty_like(exp_sums)
 
-    def run_benchmark(num_iters: int, profile: bool = False) -> float:
+    def run_cuda_benchmark(num_iters: int, profile: bool = False) -> float:
         torch.cuda.synchronize()
         if profile:
             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(
@@ -105,6 +115,7 @@ def main(
                     max_context_len,
                     alibi_slopes,
                     kv_cache_dtype,
+                    kv_scale,
                 )
             elif version == "v2":
                 ops.paged_attention_v2(
@@ -123,6 +134,7 @@ def main(
                     max_context_len,
                     alibi_slopes,
                     kv_cache_dtype,
+                    kv_scale,
                 )
             else:
                 raise ValueError(f"Invalid version: {version}")
@@ -135,6 +147,7 @@ def main(
 
     # Warmup.
     print("Warming up...")
+    run_benchmark = run_cuda_benchmark
     run_benchmark(num_iters=3, profile=False)
 
     # Benchmark.
@@ -171,10 +184,13 @@ if __name__ == '__main__':
     parser.add_argument(
         "--kv-cache-dtype",
         type=str,
-        choices=["auto", "fp8_e5m2"],
+        choices=["auto", "fp8"],
         default="auto",
         help=
-        'Data type for kv cache storage. If "auto", will use model data type.')
+        'Data type for kv cache storage. If "auto", will use model data type. '
+        '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.')
     args = parser.parse_args()
     print(args)
 

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, 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/launch_tgi_server.sh

@@ -6,7 +6,7 @@ TOKENS=$2
 
 docker run --gpus all --shm-size 1g -p $PORT:80 \
            -v $PWD/data:/data \
-           ghcr.io/huggingface/text-generation-inference:0.8 \
+           ghcr.io/huggingface/text-generation-inference:1.4.0 \
            --model-id $MODEL \
            --sharded false  \
            --max-input-length 1024 \

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_E5M2")
+    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(${GPU_LANG} STREQUAL "HIP")
+    #
+    # Get common HIP/HIPCC flags from torch.
+    #
+    run_python(GPU_FLAGS
+      "import torch.utils.cpp_extension as t; print(';'.join(t.COMMON_HIP_FLAGS + t.COMMON_HIPCC_FLAGS))"
+      "Failed to determine torch nvcc compiler flags")
+
+    list(APPEND GPU_FLAGS
+      "-DUSE_ROCM"
+      "-DENABLE_FP8_E4M3"
+      "-U__HIP_NO_HALF_CONVERSIONS__"
+      "-U__HIP_NO_HALF_OPERATORS__"
+      "-fno-gpu-rdc")
+
+  endif()
+  set(${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

@@ -2,19 +2,16 @@
 #include <torch/extension.h>
 #include <c10/cuda/CUDAGuard.h>
 
+#include <cmath>
+
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
 
 namespace vllm {
 
-template<typename T>
-__device__ __forceinline__ T silu(const T& x) {
-  // x * sigmoid(x)
-  return (T) (((float) x) / (1.0f + expf((float) -x)));
-}
-
-template<typename scalar_t>
-__global__ void silu_and_mul_kernel(
+// Activation and gating kernel template.
+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) {
@@ -22,32 +19,78 @@ __global__ void silu_and_mul_kernel(
   for (int64_t idx = threadIdx.x; idx < d; idx += blockDim.x) {
     const scalar_t x = VLLM_LDG(&input[token_idx * 2 * d + idx]);
     const scalar_t y = VLLM_LDG(&input[token_idx * 2 * d + d + idx]);
-    out[token_idx * d + idx] = silu(x) * y;
+    out[token_idx * d + idx] = ACT_FN(x) * y;
   }
 }
 
+template<typename T>
+__device__ __forceinline__ T silu_kernel(const T& x) {
+  // x * sigmoid(x)
+  return (T) (((float) x) / (1.0f + expf((float) -x)));
+}
+
+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#L36-L38
+  const float f = (float) x;
+  constexpr float ALPHA = M_SQRT1_2;
+  return (T) (f * 0.5f * (1.0f + ::erf(f * ALPHA)));
+}
+
+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);                                                                               \
+    });
+
 void silu_and_mul(
   torch::Tensor& out,      // [..., d]
   torch::Tensor& input)    // [..., 2 * d]
 {
-  int64_t num_tokens = input.numel() / input.size(-1);
-  int d = input.size(-1) / 2;
+  LAUNCH_ACTIVATION_GATE_KERNEL(vllm::silu_kernel);
+}
 
-  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(),
-    "silu_and_mul_kernel",
-    [&] {
-      vllm::silu_and_mul_kernel<scalar_t><<<grid, block, 0, stream>>>(
-        out.data_ptr<scalar_t>(),
-        input.data_ptr<scalar_t>(),
-        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 {

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_kernels.cu

@@ -15,9 +15,6 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#ifdef USE_ROCM
-#include <hip/hip_runtime.h>
-#endif
 
 #include <torch/extension.h>
 #include <ATen/cuda/CUDAContext.h>
@@ -25,15 +22,26 @@
 
 #include "attention_dtypes.h"
 #include "attention_utils.cuh"
+
+#if defined(ENABLE_FP8_E5M2)
 #include "../quantization/fp8_e5m2_kvcache/quant_utils.cuh"
+#elif defined(ENABLE_FP8_E4M3)
+#include "../quantization/fp8/amd_detail/quant_utils.cuh"
+#endif
 
 #include <algorithm>
 
+#ifdef USE_ROCM
+  #include <hip/hip_bf16.h>
+  typedef __hip_bfloat16 __nv_bfloat16;
+#endif
+
 #ifndef USE_ROCM
 #define WARP_SIZE 32
 #else
 #define WARP_SIZE warpSize
 #endif
+
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define DIVIDE_ROUND_UP(a, b) (((a) + (b) - 1) / (b))
@@ -84,7 +92,7 @@ template<
   int HEAD_SIZE,
   int BLOCK_SIZE,
   int NUM_THREADS,
-  bool IS_FP8_E5M2_KV_CACHE,
+  bool IS_FP8_KV_CACHE,
   int PARTITION_SIZE = 0> // Zero means no partitioning.
 __device__ void paged_attention_kernel(
   float* __restrict__ exp_sums,           // [num_seqs, num_heads, max_num_partitions]
@@ -101,7 +109,8 @@ __device__ void paged_attention_kernel(
   const float* __restrict__ alibi_slopes, // [num_heads]
   const int q_stride,
   const int kv_block_stride,
-  const int kv_head_stride) {
+  const int kv_head_stride,
+  const float kv_scale) {
   const int seq_idx = blockIdx.y;
   const int partition_idx = blockIdx.z;
   const int max_num_partitions = gridDim.z;
@@ -148,7 +157,7 @@ __device__ void paged_attention_kernel(
   constexpr int VEC_SIZE = MAX(16 / (THREAD_GROUP_SIZE * sizeof(scalar_t)), 1);
   using K_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
   using Q_vec = typename Vec<scalar_t, VEC_SIZE>::Type;
-#ifdef ENABLE_FP8_E5M2
+#if defined(ENABLE_FP8_E5M2) || defined(ENABLE_FP8_E4M3)
   using Quant_vec = typename Vec<cache_t, VEC_SIZE>::Type;
 #endif
 
@@ -214,11 +223,16 @@ __device__ void paged_attention_kernel(
         const int vec_idx = thread_group_offset + j * THREAD_GROUP_SIZE;
         const int offset1 = (vec_idx * VEC_SIZE) / x;
         const int offset2 = (vec_idx * VEC_SIZE) % x;
-        if constexpr (IS_FP8_E5M2_KV_CACHE) {
-#ifdef ENABLE_FP8_E5M2
+        if constexpr (IS_FP8_KV_CACHE) {
+#if defined(ENABLE_FP8_E5M2)
           Quant_vec k_vec_quant = *reinterpret_cast<const Quant_vec*>(k_ptr + offset1 * BLOCK_SIZE * x + offset2);
           // Vector conversion from Quant_vec to K_vec.
           k_vecs[j] = fp8_e5m2_unscaled::vec_conversion<K_vec, Quant_vec>(k_vec_quant);
+#elif defined(ENABLE_FP8_E4M3)
+          Quant_vec k_vec_quant = *reinterpret_cast<const Quant_vec*>(k_ptr + offset1 * BLOCK_SIZE * x + offset2);
+          // Vector conversion from Quant_vec to K_vec. Use scaled_vec_conversion to convert FP8_E4M3 quantized k
+          // cache vec to k vec in higher precision (FP16, BFloat16, etc.)
+          k_vecs[j] = fp8_e4m3::scaled_vec_conversion<K_vec, Quant_vec>(k_vec_quant, kv_scale);
 #else
           assert(false);
 #endif
@@ -298,7 +312,7 @@ __device__ void paged_attention_kernel(
   constexpr int V_VEC_SIZE = MIN(16 / sizeof(scalar_t), BLOCK_SIZE);
   using V_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
   using L_vec = typename Vec<scalar_t, V_VEC_SIZE>::Type;
-#ifdef ENABLE_FP8_E5M2
+#if defined(ENABLE_FP8_E5M2) || defined(ENABLE_FP8_E4M3)
   using V_quant_vec = typename Vec<cache_t, V_VEC_SIZE>::Type;
 #endif
   using Float_L_vec = typename FloatVec<L_vec>::Type;
@@ -334,11 +348,16 @@ __device__ void paged_attention_kernel(
       if (row_idx < HEAD_SIZE) {
         const int offset = row_idx * BLOCK_SIZE + physical_block_offset;
         V_vec v_vec;
-        if constexpr (IS_FP8_E5M2_KV_CACHE) {
-#ifdef ENABLE_FP8_E5M2
+        if constexpr (IS_FP8_KV_CACHE) {
+#if defined(ENABLE_FP8_E5M2)
           V_quant_vec v_quant_vec = *reinterpret_cast<const V_quant_vec*>(v_ptr + offset);
           // Vector conversion from V_quant_vec to V_vec.
           v_vec = fp8_e5m2_unscaled::vec_conversion<V_vec, V_quant_vec>(v_quant_vec);
+#elif defined(ENABLE_FP8_E4M3)
+          V_quant_vec v_quant_vec = *reinterpret_cast<const V_quant_vec*>(v_ptr + offset);
+          // Vector conversion from V_quant_vec to V_vec. Use scaled_vec_conversion to convert
+          // FP8_E4M3 quantized v cache vec to v vec in higher precision (FP16, BFloat16, etc.)
+          v_vec = fp8_e4m3::scaled_vec_conversion<V_vec, V_quant_vec>(v_quant_vec, kv_scale);
 #else
           assert(false);
 #endif
@@ -429,7 +448,7 @@ template<
   int HEAD_SIZE,
   int BLOCK_SIZE,
   int NUM_THREADS,
-  bool IS_FP8_E5M2_KV_CACHE>
+  bool IS_FP8_KV_CACHE>
 __global__ void paged_attention_v1_kernel(
   scalar_t* __restrict__ out,             // [num_seqs, num_heads, head_size]
   const scalar_t* __restrict__ q,         // [num_seqs, num_heads, head_size]
@@ -443,11 +462,12 @@ __global__ void paged_attention_v1_kernel(
   const float* __restrict__ alibi_slopes, // [num_heads]
   const int q_stride,
   const int kv_block_stride,
-  const int kv_head_stride) {
-  paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS, IS_FP8_E5M2_KV_CACHE>(
+  const int kv_head_stride,
+  const float kv_scale) {
+  paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS, IS_FP8_KV_CACHE>(
     /* exp_sums */ nullptr, /* max_logits */ nullptr,
     out, q, k_cache, v_cache, num_kv_heads, scale, block_tables, context_lens,
-    max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride, kv_head_stride);
+    max_num_blocks_per_seq, alibi_slopes, q_stride, kv_block_stride, kv_head_stride, kv_scale);
 }
 
 // Grid: (num_heads, num_seqs, max_num_partitions).
@@ -457,7 +477,7 @@ template<
   int HEAD_SIZE,
   int BLOCK_SIZE,
   int NUM_THREADS,
-  bool IS_FP8_E5M2_KV_CACHE,
+  bool IS_FP8_KV_CACHE,
   int PARTITION_SIZE>
 __global__ void paged_attention_v2_kernel(
   float* __restrict__ exp_sums,           // [num_seqs, num_heads, max_num_partitions]
@@ -474,11 +494,12 @@ __global__ void paged_attention_v2_kernel(
   const float* __restrict__ alibi_slopes, // [num_heads]
   const int q_stride,
   const int kv_block_stride,
-  const int kv_head_stride) {
-  paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS, IS_FP8_E5M2_KV_CACHE, PARTITION_SIZE>(
+  const int kv_head_stride,
+  const float kv_scale) {
+  paged_attention_kernel<scalar_t, cache_t, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS, IS_FP8_KV_CACHE, PARTITION_SIZE>(
     exp_sums, max_logits, tmp_out, q, k_cache, v_cache, num_kv_heads, scale,
     block_tables, context_lens, max_num_blocks_per_seq, alibi_slopes,
-    q_stride, kv_block_stride, kv_head_stride);
+    q_stride, kv_block_stride, kv_head_stride, kv_scale);
 }
 
 // Grid: (num_heads, num_seqs).
@@ -585,9 +606,9 @@ __global__ void paged_attention_v2_reduce_kernel(
 #define LAUNCH_PAGED_ATTENTION_V1(HEAD_SIZE)                                                  \
   VLLM_DevFuncAttribute_SET_MaxDynamicSharedMemorySize(                                       \
     ((void*)vllm::paged_attention_v1_kernel<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,   \
-      IS_FP8_E5M2_KV_CACHE>), shared_mem_size);                                               \
+      IS_FP8_KV_CACHE>), shared_mem_size);                                                    \
   vllm::paged_attention_v1_kernel<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,             \
-  IS_FP8_E5M2_KV_CACHE><<<grid, block, shared_mem_size, stream>>>(                            \
+  IS_FP8_KV_CACHE><<<grid, block, shared_mem_size, stream>>>(                                 \
     out_ptr,                                                                                  \
     query_ptr,                                                                                \
     key_cache_ptr,                                                                            \
@@ -600,14 +621,15 @@ __global__ void paged_attention_v2_reduce_kernel(
     alibi_slopes_ptr,                                                                         \
     q_stride,                                                                                 \
     kv_block_stride,                                                                          \
-    kv_head_stride);
+    kv_head_stride,                                                                           \
+    kv_scale);
 
 // TODO(woosuk): Tune NUM_THREADS.
 template<
   typename T,
   typename CACHE_T,
   int BLOCK_SIZE,
-  bool IS_FP8_E5M2_KV_CACHE,
+  bool IS_FP8_KV_CACHE,
   int NUM_THREADS = 128>
 void paged_attention_v1_launcher(
   torch::Tensor& out,
@@ -619,7 +641,8 @@ void paged_attention_v1_launcher(
   torch::Tensor& block_tables,
   torch::Tensor& context_lens,
   int max_context_len,
-  const c10::optional<torch::Tensor>& alibi_slopes) {
+  const c10::optional<torch::Tensor>& alibi_slopes,
+  float kv_scale) {
   int num_seqs = query.size(0);
   int num_heads = query.size(1);
   int head_size = query.size(2);
@@ -683,8 +706,8 @@ void paged_attention_v1_launcher(
   }
 }
 
-#define CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_E5M2_KV_CACHE)       \
-  paged_attention_v1_launcher<T, CACHE_T, BLOCK_SIZE, IS_FP8_E5M2_KV_CACHE>( \
+#define CALL_V1_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE)            \
+  paged_attention_v1_launcher<T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE>(      \
     out,                                                                     \
     query,                                                                   \
     key_cache,                                                               \
@@ -694,20 +717,21 @@ void paged_attention_v1_launcher(
     block_tables,                                                            \
     context_lens,                                                            \
     max_context_len,                                                         \
-    alibi_slopes);
+    alibi_slopes,                                                            \
+    kv_scale);
 
 // NOTE(woosuk): To reduce the compilation time, we omitted block sizes
 // 1, 2, 4, 64, 128, 256.
-#define CALL_V1_LAUNCHER_BLOCK_SIZE(T, CACHE_T, IS_FP8_E5M2_KV_CACHE) \
+#define CALL_V1_LAUNCHER_BLOCK_SIZE(T, CACHE_T, IS_FP8_KV_CACHE)      \
   switch (block_size) {                                               \
     case 8:                                                           \
-      CALL_V1_LAUNCHER(T, CACHE_T, 8, IS_FP8_E5M2_KV_CACHE);          \
+      CALL_V1_LAUNCHER(T, CACHE_T, 8, IS_FP8_KV_CACHE);               \
       break;                                                          \
     case 16:                                                          \
-      CALL_V1_LAUNCHER(T, CACHE_T, 16, IS_FP8_E5M2_KV_CACHE);         \
+      CALL_V1_LAUNCHER(T, CACHE_T, 16, IS_FP8_KV_CACHE);              \
       break;                                                          \
     case 32:                                                          \
-      CALL_V1_LAUNCHER(T, CACHE_T, 32, IS_FP8_E5M2_KV_CACHE);         \
+      CALL_V1_LAUNCHER(T, CACHE_T, 32, IS_FP8_KV_CACHE);              \
       break;                                                          \
     default:                                                          \
       TORCH_CHECK(false, "Unsupported block size: ", block_size);     \
@@ -726,7 +750,8 @@ void paged_attention_v1(
   int block_size,
   int max_context_len,
   const c10::optional<torch::Tensor>& alibi_slopes,
-  const std::string& kv_cache_dtype) {
+  const std::string& kv_cache_dtype,
+  float kv_scale) {
   if (kv_cache_dtype == "auto") {
     if (query.dtype() == at::ScalarType::Float) {
       CALL_V1_LAUNCHER_BLOCK_SIZE(float, float, false);
@@ -737,7 +762,7 @@ void paged_attention_v1(
     } else {
       TORCH_CHECK(false, "Unsupported data type: ", query.dtype());
     }
-  } else if (kv_cache_dtype == "fp8_e5m2") {
+  } else if (kv_cache_dtype == "fp8") {
     if (query.dtype() == at::ScalarType::Float) {
       CALL_V1_LAUNCHER_BLOCK_SIZE(float, uint8_t, true);
     } else if (query.dtype() == at::ScalarType::Half) {
@@ -754,7 +779,7 @@ void paged_attention_v1(
 
 #define LAUNCH_PAGED_ATTENTION_V2(HEAD_SIZE)                                                  \
   vllm::paged_attention_v2_kernel<T, CACHE_T, HEAD_SIZE, BLOCK_SIZE, NUM_THREADS,             \
-  IS_FP8_E5M2_KV_CACHE, PARTITION_SIZE>                                                       \
+  IS_FP8_KV_CACHE, PARTITION_SIZE>                                                            \
   <<<grid, block, shared_mem_size, stream>>>(                                                 \
     exp_sums_ptr,                                                                             \
     max_logits_ptr,                                                                           \
@@ -770,7 +795,8 @@ void paged_attention_v1(
     alibi_slopes_ptr,                                                                         \
     q_stride,                                                                                 \
     kv_block_stride,                                                                          \
-    kv_head_stride);                                                                          \
+    kv_head_stride,                                                                           \
+    kv_scale);                                                                                \
   vllm::paged_attention_v2_reduce_kernel<T, HEAD_SIZE, NUM_THREADS, PARTITION_SIZE>           \
   <<<reduce_grid, block, reduce_shared_mem_size, stream>>>(                                   \
     out_ptr,                                                                                  \
@@ -784,7 +810,7 @@ template<
   typename T,
   typename CACHE_T,
   int BLOCK_SIZE,
-  bool IS_FP8_E5M2_KV_CACHE,
+  bool IS_FP8_KV_CACHE,
   int NUM_THREADS = 128,
   int PARTITION_SIZE = 512>
 void paged_attention_v2_launcher(
@@ -800,7 +826,8 @@ void paged_attention_v2_launcher(
   torch::Tensor& block_tables,
   torch::Tensor& context_lens,
   int max_context_len,
-  const c10::optional<torch::Tensor>& alibi_slopes) {
+  const c10::optional<torch::Tensor>& alibi_slopes,
+  float kv_scale) {
   int num_seqs = query.size(0);
   int num_heads = query.size(1);
   int head_size = query.size(2);
@@ -870,8 +897,8 @@ void paged_attention_v2_launcher(
   }
 }
 
-#define CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_E5M2_KV_CACHE)           \
-  paged_attention_v2_launcher<T, CACHE_T, BLOCK_SIZE, IS_FP8_E5M2_KV_CACHE>(     \
+#define CALL_V2_LAUNCHER(T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE)                \
+  paged_attention_v2_launcher<T, CACHE_T, BLOCK_SIZE, IS_FP8_KV_CACHE>(          \
     out,                                                                         \
     exp_sums,                                                                    \
     max_logits,                                                                  \
@@ -884,20 +911,21 @@ void paged_attention_v2_launcher(
     block_tables,                                                                \
     context_lens,                                                                \
     max_context_len,                                                             \
-    alibi_slopes);
+    alibi_slopes,                                                                \
+    kv_scale);
 
 // NOTE(woosuk): To reduce the compilation time, we omitted block sizes
 // 1, 2, 4, 64, 128, 256.
-#define CALL_V2_LAUNCHER_BLOCK_SIZE(T, CACHE_T, IS_FP8_E5M2_KV_CACHE)       \
+#define CALL_V2_LAUNCHER_BLOCK_SIZE(T, CACHE_T, IS_FP8_KV_CACHE)            \
   switch (block_size) {                                                     \
     case 8:                                                                 \
-      CALL_V2_LAUNCHER(T, CACHE_T, 8, IS_FP8_E5M2_KV_CACHE);                \
+      CALL_V2_LAUNCHER(T, CACHE_T, 8, IS_FP8_KV_CACHE);                     \
       break;                                                                \
     case 16:                                                                \
-      CALL_V2_LAUNCHER(T, CACHE_T, 16, IS_FP8_E5M2_KV_CACHE);               \
+      CALL_V2_LAUNCHER(T, CACHE_T, 16, IS_FP8_KV_CACHE);                    \
       break;                                                                \
     case 32:                                                                \
-      CALL_V2_LAUNCHER(T, CACHE_T, 32, IS_FP8_E5M2_KV_CACHE);               \
+      CALL_V2_LAUNCHER(T, CACHE_T, 32, IS_FP8_KV_CACHE);                    \
       break;                                                                \
     default:                                                                \
       TORCH_CHECK(false, "Unsupported block size: ", block_size);           \
@@ -919,7 +947,8 @@ void paged_attention_v2(
   int block_size,
   int max_context_len,
   const c10::optional<torch::Tensor>& alibi_slopes,
-  const std::string& kv_cache_dtype) {
+  const std::string& kv_cache_dtype,
+  float kv_scale) {
   if (kv_cache_dtype == "auto") {
     if (query.dtype() == at::ScalarType::Float) {
       CALL_V2_LAUNCHER_BLOCK_SIZE(float, float, false);
@@ -930,7 +959,7 @@ void paged_attention_v2(
     } else {
       TORCH_CHECK(false, "Unsupported data type: ", query.dtype());
     }
-  } else if (kv_cache_dtype == "fp8_e5m2") {
+  } else if (kv_cache_dtype == "fp8") {
     if (query.dtype() == at::ScalarType::Float) {
       CALL_V2_LAUNCHER_BLOCK_SIZE(float, uint8_t, true);
     } else if (query.dtype() == at::ScalarType::Half) {

csrc/attention/dtype_fp8.cuh

@@ -8,7 +8,7 @@
 #endif
 
 namespace vllm {
-#ifdef ENABLE_FP8_E5M2
+#if defined(ENABLE_FP8_E5M2) || defined(ENABLE_FP8_E4M3)
 // fp8 vector types for quantization of kv cache
 
 template<>

csrc/cache.h

@@ -21,16 +21,10 @@ void reshape_and_cache(
   torch::Tensor& key_cache,
   torch::Tensor& value_cache,
   torch::Tensor& slot_mapping,
-  const std::string& kv_cache_dtype);
-
-void gather_cached_kv(
-  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);
 
 // Just for unittest
-void convert_fp8_e5m2(
+void convert_fp8(
   torch::Tensor& src_cache,
   torch::Tensor& dst_cache);

csrc/cache_kernels.cu

@@ -4,13 +4,22 @@
 
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
+#if defined(ENABLE_FP8_E5M2)
 #include "quantization/fp8_e5m2_kvcache/quant_utils.cuh"
+#elif defined(ENABLE_FP8_E4M3)
+#include "quantization/fp8/amd_detail/quant_utils.cuh"
+#endif
 
 #include <algorithm>
 #include <cassert>
 #include <map>
 #include <vector>
 
+#ifdef USE_ROCM
+  #include <hip/hip_bf16.h>
+  typedef __hip_bfloat16 __nv_bfloat16;
+#endif
+
 void swap_blocks(
   torch::Tensor& src,
   torch::Tensor& dst,
@@ -144,7 +153,7 @@ void copy_blocks(
 
 namespace vllm {
 
-template<typename scalar_t, typename cache_t, bool is_fp8_e5m2_kv_cache>
+template<typename scalar_t, typename cache_t, bool is_fp8_kv_cache>
 __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]
@@ -156,7 +165,8 @@ __global__ void reshape_and_cache_kernel(
   const int num_heads,
   const int head_size,
   const int block_size,
-  const int x) {
+  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) {
@@ -188,10 +198,13 @@ __global__ void reshape_and_cache_kernel(
                                   + 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
+    if constexpr (is_fp8_kv_cache) {
+#if defined(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);
+#elif defined(ENABLE_FP8_E4M3)
+      key_cache[tgt_key_idx] = fp8_e4m3::scaled_vec_conversion<uint8_t, scalar_t>(tgt_key, kv_scale);
+      value_cache[tgt_value_idx] = fp8_e4m3::scaled_vec_conversion<uint8_t, scalar_t>(tgt_value, kv_scale);
 #else
       assert(false);
 #endif
@@ -204,8 +217,8 @@ __global__ void reshape_and_cache_kernel(
 
 } // 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>>>( \
+#define CALL_RESHAPE_AND_CACHE(KV_T, CACHE_T, IS_FP8_KV_CACHE)                                     \
+  vllm::reshape_and_cache_kernel<KV_T, CACHE_T, IS_FP8_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()),                                              \
@@ -216,7 +229,8 @@ __global__ void reshape_and_cache_kernel(
     num_heads,                                                                                     \
     head_size,                                                                                     \
     block_size,                                                                                    \
-    x);
+    x,                                                                                             \
+    kv_scale);
 
 void reshape_and_cache(
   torch::Tensor& key,           // [num_tokens, num_heads, head_size]
@@ -224,7 +238,8 @@ void reshape_and_cache(
   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 std::string& kv_cache_dtype,
+  const float kv_scale)
 {
   int num_tokens = key.size(0);
   int num_heads = key.size(1);
@@ -247,7 +262,7 @@ void reshape_and_cache(
     } else if (key.dtype() == at::ScalarType::BFloat16) {
       CALL_RESHAPE_AND_CACHE(__nv_bfloat16, __nv_bfloat16, false);
     }
-  } else if (kv_cache_dtype == "fp8_e5m2") {
+  } else if (kv_cache_dtype == "fp8") {
     if (key.dtype() == at::ScalarType::Float) {
       CALL_RESHAPE_AND_CACHE(float, uint8_t, true);
     } else if (key.dtype() == at::ScalarType::Half) {
@@ -262,177 +277,18 @@ void reshape_and_cache(
 
 namespace vllm {
 
-// Grid: (num_blocks, block_size).
-template<typename scalar_t>
-__global__ void gather_cached_kv_kernel(
-  scalar_t* __restrict__ key,             // [num_tokens, [stride], num_heads, head_size]
-  scalar_t* __restrict__ value,           // [num_tokens, [stride], num_heads, head_size]
-  const scalar_t* __restrict__ key_cache,   // [num_blocks, num_heads, head_size/x, block_size, x]
-  const scalar_t* __restrict__ value_cache,   // [num_blocks, num_heads, head_size, block_size]
-  const int* __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 int token_idx = blockIdx.x;
-    const int slot_idx = slot_mapping[token_idx];
-    const int block_idx = slot_idx / block_size;
-    const int block_offset = slot_idx % block_size;
-
-    const int num_tokens = num_heads * head_size;
-    for (int i = threadIdx.x; i < num_tokens; i += blockDim.x) {
-      const int tgt_key_idx = token_idx * key_stride + i;
-      const int tgt_value_idx = token_idx * value_stride + i;
-
-      const int head_idx = i / head_size;
-      const int head_offset = i % head_size;
-      const int x_idx = head_offset / x;  // the offset of the [head_size/x] dimension
-      const int x_offset = head_offset % x;
-
-      const int src_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 int src_value_idx = block_idx * num_heads * head_size * block_size
-                                + head_idx * head_size * block_size
-                                + head_offset * block_size
-                                + block_offset;
-
-      key[tgt_key_idx] = VLLM_LDG(&key_cache[src_key_idx]);
-      value[tgt_value_idx] = VLLM_LDG(&value_cache[src_value_idx]);
-    }
-}
-
-template <typename scalar_t>
-__global__ void gather_cached_kv_kernel_optimized(
-    scalar_t *__restrict__ key,             // [num_tokens, [stride], num_heads, head_size]
-    scalar_t *__restrict__ value,           // [num_tokens, [stride], num_heads, head_size]
-    const scalar_t *__restrict__ key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
-    const scalar_t *__restrict__ value_cache, // [num_blocks, num_heads, head_size, block_size]
-    const int *__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 int token_idx = blockIdx.x;
-    const int slot_idx = slot_mapping[token_idx];
-    const int block_idx = slot_idx / block_size;
-    const int block_offset = slot_idx % block_size;
-
-    const int dim = num_heads * head_size;
-    assert(dim % 4 == 0);  // this is true for known use cases
-    const int unroll_factor = 4;
-    const int unrolled_dim = dim / unroll_factor;
-
-    for (int i = threadIdx.x; i < unrolled_dim; i += blockDim.x)
-    {
-        int tgt_key_indices[unroll_factor];
-        int tgt_value_indices[unroll_factor];
-        int src_key_indices[unroll_factor];
-        int src_value_indices[unroll_factor];
-        scalar_t keys_to_store[unroll_factor];
-        scalar_t values_to_store[unroll_factor];
-
-        #pragma unroll
-        for (int j = 0; j < unroll_factor; ++j)
-        {
-            int index = i + j * unrolled_dim;
-
-            const int tgt_key_idx = token_idx * key_stride + index;
-            const int tgt_value_idx = token_idx * value_stride + index;
-
-            const int head_idx = index / head_size;
-            const int head_offset = index % head_size;
-            const int x_idx = head_offset / x;
-            const int x_offset = head_offset % x;
-
-            const int src_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 int src_value_idx = block_idx * num_heads * head_size * block_size
-                                      + head_idx * head_size * block_size
-                                      + head_offset * block_size
-                                      + block_offset;
-
-            tgt_key_indices[j] = tgt_key_idx;
-            tgt_value_indices[j] = tgt_value_idx;
-            src_key_indices[j] = src_key_idx;
-            src_value_indices[j] = src_value_idx;
-
-            keys_to_store[j] = VLLM_LDG(&key_cache[src_key_idx]);
-            values_to_store[j] = VLLM_LDG(&value_cache[src_value_idx]);
-        }
-
-        #pragma unroll
-        for (int j = 0; j < unroll_factor; ++j)
-        {
-            key[tgt_key_indices[j]] = keys_to_store[j];
-            value[tgt_value_indices[j]] = values_to_store[j];
-        }
-    }
-}
-
-} // namespace vllm
-
-void gather_cached_kv(
-  torch::Tensor& key,           // [out] [num_tokens, num_heads, head_size]
-  torch::Tensor& value,         // [out] [num_tokens, num_heads, head_size]
-  torch::Tensor& key_cache,     // [in]  [num_blocks, num_heads, head_size/x, block_size, x]
-  torch::Tensor& value_cache,   // [in]  [num_blocks, num_heads, head_size, block_size]
-  torch::Tensor& slot_mapping)  // [in]  [num_tokens]
-{
-  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);
-
-  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_AND_BYTE_TYPES(
-    key.scalar_type(),
-    "gather_cached_kv_kernel_optimized",
-    [&] {
-      vllm::gather_cached_kv_kernel_optimized<scalar_t><<<grid, block, 0, stream>>>(
-        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<int>(),
-        key_stride,
-        value_stride,
-        num_heads,
-        head_size,
-        block_size,
-        x);
-    });
-}
-
-namespace vllm {
-
 template<typename Tout, typename Tin>
-__global__ void convert_fp8_e5m2_kernel(
+__global__ void convert_fp8_kernel(
   const Tin* __restrict__ src_cache,
   Tout* __restrict__ dst_cache,
   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
+#if defined(ENABLE_FP8_E5M2)
     dst_cache[idx] = fp8_e5m2_unscaled::vec_conversion<Tout, Tin>(src_cache[idx]);
+#elif defined(ENABLE_FP8_E4M3)
+    dst_cache[idx] = fp8_e4m3::vec_conversion<Tout, Tin>(src_cache[idx]);
 #else
     assert(false);
 #endif
@@ -441,16 +297,25 @@ __global__ void convert_fp8_e5m2_kernel(
 
 } // 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()),                       \
+#define CALL_CONVERT_FP8(Tout, Tin)                                 \
+  vllm::convert_fp8_kernel<Tout, Tin><<<grid, block, 0, stream>>>(  \
+    reinterpret_cast<Tin*>(src_cache.data_ptr()),                   \
+    reinterpret_cast<Tout*>(dst_cache.data_ptr()),                  \
     block_stride);
 
-void convert_fp8_e5m2(
+void convert_fp8(
   torch::Tensor& src_cache,
   torch::Tensor& dst_cache)
 {
+  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);
+
   int64_t num_blocks = src_cache.size(0);
   int64_t block_stride = src_cache.stride(0);
 
@@ -459,16 +324,16 @@ void convert_fp8_e5m2(
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
   if (src_cache.dtype() == at::ScalarType::Float) {
-    CALL_CONVERT_FP8_E5M2(uint8_t, float);
+    CALL_CONVERT_FP8(uint8_t, float);
   } else if (src_cache.dtype() == at::ScalarType::Half) {
-    CALL_CONVERT_FP8_E5M2(uint8_t, uint16_t);
+    CALL_CONVERT_FP8(uint8_t, uint16_t);
   } else if (src_cache.dtype() == at::ScalarType::BFloat16) {
-    CALL_CONVERT_FP8_E5M2(uint8_t, __nv_bfloat16);
+    CALL_CONVERT_FP8(uint8_t, __nv_bfloat16);
   } else if (dst_cache.dtype() == at::ScalarType::Float) {
-    CALL_CONVERT_FP8_E5M2(float, uint8_t);
+    CALL_CONVERT_FP8(float, uint8_t);
   } else if (dst_cache.dtype() == at::ScalarType::Half) {
-    CALL_CONVERT_FP8_E5M2(uint16_t, uint8_t);
+    CALL_CONVERT_FP8(uint16_t, uint8_t);
   } else if (dst_cache.dtype() == at::ScalarType::BFloat16) {
-    CALL_CONVERT_FP8_E5M2(__nv_bfloat16, uint8_t);
+    CALL_CONVERT_FP8(__nv_bfloat16, uint8_t);
   }
 }

csrc/cpu/activation.cpp

@@ -0,0 +1,148 @@
+#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 +1,746 @@
+#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 context_len) {
+  data[0] += alibi_slope * (start_index - context_len + 1);
+  T max = data[0];
+  for (int i = 1; i < size; ++i) {
+    T qk = data[i] + alibi_slope * (start_index + i - context_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__ context_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_context_len = max_num_blocks_per_seq * BLOCK_SIZE;
+    int max_context_len_padded = (max_context_len + 15) & 0xFFFFFFF0;
+    TORCH_CHECK((max_context_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_context_len_padded * sizeof(float);
+    float *logits = (float *)std::aligned_alloc(
+        64, logits_bytes); // Cacheline alignment for each context token.
+                           // [parallel_work_item_num, max_context_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 context_len = context_lens[seq_idx];
+        const int *seq_block_table =
+            block_tables + max_num_blocks_per_seq * seq_idx;
+        const int block_num = (context_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 =
+            context_len - (block_num - 1) * BLOCK_SIZE;
+        float *__restrict__ thread_block_logits =
+            logits + omp_get_thread_num() * max_context_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, context_len,
+                             block_num * BLOCK_SIZE, alibi_slopes[head_idx], 0,
+                             context_len);
+        } else {
+          reduceSoftmax(thread_block_logits, context_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, context_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 &context_lens,
+    int max_context_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 *context_lens_ptr = context_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 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,   \
+      context_lens, max_context_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 &context_lens, int block_size,
+                        int max_context_len,
+                        const c10::optional<torch::Tensor> &alibi_slopes,
+                        const std::string &kv_cache_dtype, float kv_scale) {
+  TORCH_CHECK(kv_scale == 1.0f);
+  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__ context_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 context_len = context_lens[seq_idx];
+          const int start_token_idx = partition_idx * PARTITION_SIZE;
+
+          if (start_token_idx >= context_len)
+            continue;
+
+          const int partition_num =
+              (context_len + PARTITION_SIZE - 1) / PARTITION_SIZE;
+          const bool no_reduce = (partition_num == 1);
+          const int context_token_num =
+              (std::min(context_len, start_token_idx + PARTITION_SIZE) -
+               start_token_idx);
+          const int block_num =
+              (context_token_num + BLOCK_SIZE - 1) / BLOCK_SIZE;
+          const int last_block_token_num =
+              context_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, context_token_num, block_num * BLOCK_SIZE,
+                alibi_slopes[head_idx], start_token_idx, context_len);
+          } else {
+            max_and_sum = reduceSoftmax(logits, context_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 context_len = context_lens[seq_idx];
+        const int partition_num =
+            (context_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 context_len = context_lens[seq_idx];
+          const int partition_num =
+              (context_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,   \
+      context_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 &context_lens, int block_size,
+    int max_context_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 *context_lens_ptr = context_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 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, context_lens, block_size,             \
+      max_context_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 &context_lens, int block_size,
+                        int max_context_len,
+                        const c10::optional<torch::Tensor> &alibi_slopes,
+                        const std::string &kv_cache_dtype, float kv_scale) {
+  TORCH_CHECK(kv_scale == 1.0f);
+  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 +1,141 @@
+#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 std::vector<std::pair<int64_t, int64_t>> mapping_pairs,
+    const int element_num_per_block, const int layer_num) {
+  const size_t pair_num = mapping_pairs.size();
+  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].first;
+      int64_t target_offset =
+          element_num_per_block * mapping_pairs[pair].second;
+      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 std::map<int64_t, std::vector<int64_t>> &block_mapping) {
+  int num_layers = key_caches.size();
+  TORCH_CHECK(num_layers == value_caches.size());
+  if (num_layers == 0) {
+    return;
+  }
+
+  std::vector<std::pair<int64_t, int64_t>> mapping_pairs;
+  mapping_pairs.reserve(block_mapping.size());
+  for (const auto &pair : block_mapping) {
+    for (const auto &dst : pair.second) {
+      mapping_pairs.emplace_back(pair.first, dst);
+    }
+  }
+
+  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, mapping_pairs,
+                                       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 std::map<int64_t, int64_t> &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 +1,199 @@
+
+#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();
+  constexpr int ELEM_SIZE = sizeof(scalar_t);
+
+  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 +1,73 @@
+#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

csrc/custom_all_reduce.cu

@@ -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);
 }
 
@@ -62,9 +62,9 @@ bool should_custom_ar(torch::Tensor &inp, int max_size, int world_size,
   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,
@@ -126,7 +126,7 @@ void dispose(fptr_t _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,

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 RankSignals {
-  volatile Signal *signals[8];
-};
+struct __align__(16) RankSignals { volatile Signal *signals[8]; };
 
 // like std::array, but aligned
 template <typename T, int sz>
@@ -135,70 +123,49 @@ 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>
@@ -214,32 +181,32 @@ 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,
+                               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);
   }
-  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);
+  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,
+                               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;
@@ -248,6 +215,7 @@ __global__ void __launch_bounds__(512, 1)
   int part = size / ngpus;
   int start = rank * part;
   int end = rank == ngpus - 1 ? size : start + part;
+  int largest_part = part + size % ngpus;
   const P *ptrs[ngpus];
   P *tmps[ngpus];
 #pragma unroll
@@ -257,75 +225,28 @@ __global__ void __launch_bounds__(512, 1)
     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)>;
@@ -341,7 +262,7 @@ class CustomAllreduce {
   // below are device pointers
   RankSignals sg_;
   std::unordered_map<void *, RankData *> buffers_;
-  Metadata *meta_;
+  Signal *self_sg_;
 
   // stores the registered device pointers from all ranks
   RankData *d_rank_data_base_, *d_rank_data_end_;
@@ -352,32 +273,32 @@ class CustomAllreduce {
   /**
    * 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,
+  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),
+        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]);
         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;
     }
   }
 
@@ -492,6 +413,10 @@ 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;
     cudaStreamCaptureStatus status;
@@ -512,9 +437,9 @@ 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 +451,6 @@ class CustomAllreduce {
       } else {                                        \
         KL(ngpus, cross_device_reduce_2stage);        \
       }                                               \
-    } else {                                          \
-      KL(ngpus, cross_device_reduce_half_butterfly);  \
     }                                                 \
     break;                                            \
   }
@@ -556,7 +479,7 @@ 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

@@ -92,7 +92,7 @@ __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) {
+         int data_size, bool performance_test) {
   T *result;
   cudaStream_t stream;
   CUDACHECK(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
@@ -101,7 +101,7 @@ 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;
+  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));
 
@@ -133,7 +133,7 @@ void run(int myRank, int nRanks, ncclComm_t &comm, int threads, int block_limit,
                            offsets, myRank);
   auto *self_data =
       reinterpret_cast<T *>(reinterpret_cast<char *>(buffer) +
-                            sizeof(vllm::Metadata) + data_size * sizeof(T));
+                            sizeof(vllm::Signal) + data_size * sizeof(T));
   // hack buffer registration
   {
     std::vector<std::string> handles;
@@ -143,8 +143,8 @@ void run(int myRank, int nRanks, ncclComm_t &comm, int threads, int block_limit,
       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);
   }
 
@@ -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));
@@ -269,14 +300,15 @@ int main(int argc, char **argv) {
                      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/layernorm_kernels.cu

@@ -4,6 +4,16 @@
 
 #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 {
 
@@ -35,9 +45,201 @@ __global__ void rms_norm_kernel(
   }
 }
 
-// TODO: Further optimize this kernel.
-template<typename scalar_t>
-__global__ void fused_add_rms_norm_kernel(
+
+/* 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]
@@ -48,12 +250,17 @@ __global__ void fused_add_rms_norm_kernel(
   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);
   }
@@ -93,6 +300,21 @@ void rms_norm(
     });
 }
 
+#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]
@@ -102,19 +324,29 @@ void fused_add_rms_norm(
   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

@@ -0,0 +1,7 @@
+#include "moe_ops.h"
+
+#include <torch/extension.h>
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def("topk_softmax", &topk_softmax, "Apply topk softmax to the gating outputs.");
+}

csrc/moe/moe_ops.h

@@ -0,0 +1,9 @@
+#pragma once
+
+#include <torch/extension.h>
+
+void topk_softmax(
+  torch::Tensor& topk_weights,
+  torch::Tensor& topk_indices,
+  torch::Tensor& token_expert_indices,
+  torch::Tensor& gating_output);

csrc/moe/topk_softmax_kernels.cu

@@ -0,0 +1,499 @@
+/*
+ * Adapted from https://github.com/NVIDIA/TensorRT-LLM/blob/v0.7.1/cpp/tensorrt_llm/kernels/mixtureOfExperts/moe_kernels.cu
+ * Copyright (c) 2024, The vLLM team.
+ * SPDX-FileCopyrightText: Copyright (c) 1993-2023 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include <torch/extension.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include <cub/cub.cuh>
+#include <cub/util_type.cuh>
+
+namespace vllm {
+namespace moe {
+
+static constexpr int WARP_SIZE = 32;
+
+/// Aligned array type
+template <
+    typename T,
+    /// Number of elements in the array
+    int N,
+    /// Alignment requirement in bytes
+    int Alignment = sizeof(T) * N
+>
+class alignas(Alignment) AlignedArray {
+    float data[N];
+};
+
+// ====================== Softmax things ===============================
+// We have our own implementation of softmax here so we can support transposing the output
+// in the softmax kernel when we extend this module to support expert-choice routing.
+template <int TPB>
+__launch_bounds__(TPB) __global__
+    void moeSoftmax(const float* input, const bool* finished, float* output, const int num_cols)
+{
+    using BlockReduce = cub::BlockReduce<float, TPB>;
+    __shared__ typename BlockReduce::TempStorage tmpStorage;
+
+    __shared__ float normalizing_factor;
+    __shared__ float float_max;
+
+    const int thread_row_offset = blockIdx.x * num_cols;
+
+    cub::Sum sum;
+    float threadData(-FLT_MAX);
+
+    // Don't touch finished rows.
+    if ((finished != nullptr) && finished[blockIdx.x])
+    {
+        return;
+    }
+
+    for (int ii = threadIdx.x; ii < num_cols; ii += TPB)
+    {
+        const int idx = thread_row_offset + ii;
+        threadData = max(static_cast<float>(input[idx]), threadData);
+    }
+
+    const float maxElem = BlockReduce(tmpStorage).Reduce(threadData, cub::Max());
+    if (threadIdx.x == 0)
+    {
+        float_max = maxElem;
+    }
+    __syncthreads();
+
+    threadData = 0;
+
+    for (int ii = threadIdx.x; ii < num_cols; ii += TPB)
+    {
+        const int idx = thread_row_offset + ii;
+        threadData += exp((static_cast<float>(input[idx]) - float_max));
+    }
+
+    const auto Z = BlockReduce(tmpStorage).Reduce(threadData, sum);
+
+    if (threadIdx.x == 0)
+    {
+        normalizing_factor = 1.f / Z;
+    }
+    __syncthreads();
+
+    for (int ii = threadIdx.x; ii < num_cols; ii += TPB)
+    {
+        const int idx = thread_row_offset + ii;
+        const float val = exp((static_cast<float>(input[idx]) - float_max)) * normalizing_factor;
+        output[idx] = val;
+    }
+}
+
+template <int TPB>
+__launch_bounds__(TPB) __global__ void moeTopK(const float* inputs_after_softmax, const bool* finished, float* output,
+    int* indices, int* source_rows, const int num_experts, const int k, const int start_expert, const int end_expert)
+{
+
+    using cub_kvp = cub::KeyValuePair<int, float>;
+    using BlockReduce = cub::BlockReduce<cub_kvp, TPB>;
+    __shared__ typename BlockReduce::TempStorage tmpStorage;
+
+    cub_kvp thread_kvp;
+    cub::ArgMax arg_max;
+
+    const int num_rows = gridDim.x;
+    const int block_row = blockIdx.x;
+
+    const bool row_is_active = finished ? !finished[block_row] : true;
+    const int thread_read_offset = blockIdx.x * num_experts;
+    for (int k_idx = 0; k_idx < k; ++k_idx)
+    {
+        thread_kvp.key = 0;
+        thread_kvp.value = -1.f; // This is OK because inputs are probabilities
+
+        cub_kvp inp_kvp;
+        for (int expert = threadIdx.x; expert < num_experts; expert += TPB)
+        {
+            const int idx = thread_read_offset + expert;
+            inp_kvp.key = expert;
+            inp_kvp.value = inputs_after_softmax[idx];
+
+            for (int prior_k = 0; prior_k < k_idx; ++prior_k)
+            {
+                const int prior_winning_expert = indices[k * block_row + prior_k];
+
+                if (prior_winning_expert == expert)
+                {
+                    inp_kvp = thread_kvp;
+                }
+            }
+
+            thread_kvp = arg_max(inp_kvp, thread_kvp);
+        }
+
+        const cub_kvp result_kvp = BlockReduce(tmpStorage).Reduce(thread_kvp, arg_max);
+        if (threadIdx.x == 0)
+        {
+            // Ignore experts the node isn't responsible for with expert parallelism
+            const int expert = result_kvp.key;
+            const bool node_uses_expert = expert >= start_expert && expert < end_expert;
+            const bool should_process_row = row_is_active && node_uses_expert;
+
+            const int idx = k * block_row + k_idx;
+            output[idx] = result_kvp.value;
+            indices[idx] = should_process_row ? (expert - start_expert) : num_experts;
+            assert(indices[idx] >= 0);
+            source_rows[idx] = k_idx * num_rows + block_row;
+        }
+        __syncthreads();
+    }
+}
+
+// ====================== TopK softmax things ===============================
+
+/*
+  A Top-K gating softmax written to exploit when the number of experts in the MoE layers
+  are a small power of 2. This allows us to cleanly share the rows among the threads in
+  a single warp and eliminate communication between warps (so no need to use shared mem).
+
+  It fuses the softmax, max and argmax into a single kernel.
+
+  Limitations:
+  1) This implementation is intended for when the number of experts is a small power of 2.
+  2) This implementation assumes k is small, but will work for any k.
+*/
+
+template <int VPT, int NUM_EXPERTS, int WARPS_PER_CTA, int BYTES_PER_LDG>
+__launch_bounds__(WARPS_PER_CTA* WARP_SIZE) __global__
+    void topkGatingSoftmax(const float* input, const bool* finished, float* output, const int num_rows, int* indices,
+        int* source_rows, const int k, const int start_expert, const int end_expert)
+{
+    // We begin by enforcing compile time assertions and setting up compile time constants.
+    static_assert(VPT == (VPT & -VPT), "VPT must be power of 2");
+    static_assert(NUM_EXPERTS == (NUM_EXPERTS & -NUM_EXPERTS), "NUM_EXPERTS must be power of 2");
+    static_assert(BYTES_PER_LDG == (BYTES_PER_LDG & -BYTES_PER_LDG), "BYTES_PER_LDG must be power of 2");
+    static_assert(BYTES_PER_LDG <= 16, "BYTES_PER_LDG must be leq 16");
+
+    // Number of bytes each thread pulls in per load
+    static constexpr int ELTS_PER_LDG = BYTES_PER_LDG / sizeof(float);
+    static constexpr int ELTS_PER_ROW = NUM_EXPERTS;
+    static constexpr int THREADS_PER_ROW = ELTS_PER_ROW / VPT;
+    static constexpr int LDG_PER_THREAD = VPT / ELTS_PER_LDG;
+
+    // Restrictions based on previous section.
+    static_assert(VPT % ELTS_PER_LDG == 0, "The elements per thread must be a multiple of the elements per ldg");
+    static_assert(WARP_SIZE % THREADS_PER_ROW == 0, "The threads per row must cleanly divide the threads per warp");
+    static_assert(THREADS_PER_ROW == (THREADS_PER_ROW & -THREADS_PER_ROW), "THREADS_PER_ROW must be power of 2");
+    static_assert(THREADS_PER_ROW <= WARP_SIZE, "THREADS_PER_ROW can be at most warp size");
+
+    // We have NUM_EXPERTS elements per row. We specialize for small #experts
+    static constexpr int ELTS_PER_WARP = WARP_SIZE * VPT;
+    static constexpr int ROWS_PER_WARP = ELTS_PER_WARP / ELTS_PER_ROW;
+    static constexpr int ROWS_PER_CTA = WARPS_PER_CTA * ROWS_PER_WARP;
+
+    // Restrictions for previous section.
+    static_assert(ELTS_PER_WARP % ELTS_PER_ROW == 0, "The elts per row must cleanly divide the total elt per warp");
+
+    // ===================== From this point, we finally start computing run-time variables. ========================
+
+    // Compute CTA and warp rows. We pack multiple rows into a single warp, and a block contains WARPS_PER_CTA warps.
+    // This, each block processes a chunk of rows. We start by computing the start row for each block.
+    const int cta_base_row = blockIdx.x * ROWS_PER_CTA;
+
+    // Now, using the base row per thread block, we compute the base row per warp.
+    const int warp_base_row = cta_base_row + threadIdx.y * ROWS_PER_WARP;
+
+    // The threads in a warp are split into sub-groups that will work on a row.
+    // We compute row offset for each thread sub-group
+    const int thread_row_in_warp = threadIdx.x / THREADS_PER_ROW;
+    const int thread_row = warp_base_row + thread_row_in_warp;
+
+    // Threads with indices out of bounds should early exit here.
+    if (thread_row >= num_rows)
+    {
+        return;
+    }
+    const bool row_is_active = finished ? !finished[thread_row] : true;
+
+    // We finally start setting up the read pointers for each thread. First, each thread jumps to the start of the
+    // row it will read.
+    const float* thread_row_ptr = input + thread_row * ELTS_PER_ROW;
+
+    // Now, we compute the group each thread belong to in order to determine the first column to start loads.
+    const int thread_group_idx = threadIdx.x % THREADS_PER_ROW;
+    const int first_elt_read_by_thread = thread_group_idx * ELTS_PER_LDG;
+    const float* thread_read_ptr = thread_row_ptr + first_elt_read_by_thread;
+
+    // Determine the pointer type to use to read in the data depending on the BYTES_PER_LDG template param. In theory,
+    // this can support all powers of 2 up to 16.
+    // NOTE(woosuk): The original implementation uses CUTLASS aligned array here.
+    // We defined our own aligned array and use it here to avoid the dependency on CUTLASS.
+    using AccessType = AlignedArray<float, ELTS_PER_LDG>;
+
+    // Finally, we pull in the data from global mem
+    float row_chunk[VPT];
+    AccessType* row_chunk_vec_ptr = reinterpret_cast<AccessType*>(&row_chunk);
+    const AccessType* vec_thread_read_ptr = reinterpret_cast<const AccessType*>(thread_read_ptr);
+#pragma unroll
+    for (int ii = 0; ii < LDG_PER_THREAD; ++ii)
+    {
+        row_chunk_vec_ptr[ii] = vec_thread_read_ptr[ii * THREADS_PER_ROW];
+    }
+
+    // First, we perform a max reduce within the thread. We can do the max in fp16 safely (I think) and just
+    // convert to float afterwards for the exp + sum reduction.
+    float thread_max = row_chunk[0];
+#pragma unroll
+    for (int ii = 1; ii < VPT; ++ii)
+    {
+        thread_max = max(thread_max, row_chunk[ii]);
+    }
+
+// Now, we find the max within the thread group and distribute among the threads. We use a butterfly reduce.
+#pragma unroll
+    for (int mask = THREADS_PER_ROW / 2; mask > 0; mask /= 2)
+    {
+        thread_max = max(thread_max, __shfl_xor_sync(0xFFFFFFFF, thread_max, mask, THREADS_PER_ROW));
+    }
+
+    // From this point, thread max in all the threads have the max within the row.
+    // Now, we subtract the max from each element in the thread and take the exp. We also compute the thread local sum.
+    float row_sum = 0;
+#pragma unroll
+    for (int ii = 0; ii < VPT; ++ii)
+    {
+        row_chunk[ii] = expf(row_chunk[ii] - thread_max);
+        row_sum += row_chunk[ii];
+    }
+
+// Now, we perform the sum reduce within each thread group. Similar to the max reduce, we use a bufferfly pattern.
+#pragma unroll
+    for (int mask = THREADS_PER_ROW / 2; mask > 0; mask /= 2)
+    {
+        row_sum += __shfl_xor_sync(0xFFFFFFFF, row_sum, mask, THREADS_PER_ROW);
+    }
+
+    // From this point, all threads have the max and the sum for their rows in the thread_max and thread_sum variables
+    // respectively. Finally, we can scale the rows for the softmax. Technically, for top-k gating we don't need to
+    // compute the entire softmax row. We can likely look at the maxes and only compute for the top-k values in the row.
+    // However, this kernel will likely not be a bottle neck and it seems better to closer match torch and find the
+    // argmax after computing the softmax.
+    const float reciprocal_row_sum = 1.f / row_sum;
+
+#pragma unroll
+    for (int ii = 0; ii < VPT; ++ii)
+    {
+        row_chunk[ii] = row_chunk[ii] * reciprocal_row_sum;
+    }
+
+    // Now, softmax_res contains the softmax of the row chunk. Now, I want to find the topk elements in each row, along
+    // with the max index.
+    int start_col = first_elt_read_by_thread;
+    static constexpr int COLS_PER_GROUP_LDG = ELTS_PER_LDG * THREADS_PER_ROW;
+
+    for (int k_idx = 0; k_idx < k; ++k_idx)
+    {
+        // First, each thread does the local argmax
+        float max_val = row_chunk[0];
+        int expert = start_col;
+#pragma unroll
+        for (int ldg = 0, col = start_col; ldg < LDG_PER_THREAD; ++ldg, col += COLS_PER_GROUP_LDG)
+        {
+#pragma unroll
+            for (int ii = 0; ii < ELTS_PER_LDG; ++ii)
+            {
+                float val = row_chunk[ldg * ELTS_PER_LDG + ii];
+
+                // No check on the experts here since columns with the smallest index are processed first and only
+                // updated if > (not >=)
+                if (val > max_val)
+                {
+                    max_val = val;
+                    expert = col + ii;
+                }
+            }
+        }
+
+// Now, we perform the argmax reduce. We use the butterfly pattern so threads reach consensus about the max.
+// This will be useful for K > 1 so that the threads can agree on "who" had the max value. That thread can
+// then blank out their max with -inf and the warp can run more iterations...
+#pragma unroll
+        for (int mask = THREADS_PER_ROW / 2; mask > 0; mask /= 2)
+        {
+            float other_max = __shfl_xor_sync(0xFFFFFFFF, max_val, mask, THREADS_PER_ROW);
+            int other_expert = __shfl_xor_sync(0xFFFFFFFF, expert, mask, THREADS_PER_ROW);
+
+            // We want lower indices to "win" in every thread so we break ties this way
+            if (other_max > max_val || (other_max == max_val && other_expert < expert))
+            {
+                max_val = other_max;
+                expert = other_expert;
+            }
+        }
+
+        // Write the max for this k iteration to global memory.
+        if (thread_group_idx == 0)
+        {
+            // Add a guard to ignore experts not included by this node
+            const bool node_uses_expert = expert >= start_expert && expert < end_expert;
+            const bool should_process_row = row_is_active && node_uses_expert;
+
+            // The lead thread from each sub-group will write out the final results to global memory. (This will be a
+            // single) thread per row of the input/output matrices.
+            const int idx = k * thread_row + k_idx;
+            output[idx] = max_val;
+            indices[idx] = should_process_row ? (expert - start_expert) : NUM_EXPERTS;
+            source_rows[idx] = k_idx * num_rows + thread_row;
+        }
+
+        // Finally, we clear the value in the thread with the current max if there is another iteration to run.
+        if (k_idx + 1 < k)
+        {
+            const int ldg_group_for_expert = expert / COLS_PER_GROUP_LDG;
+            const int thread_to_clear_in_group = (expert / ELTS_PER_LDG) % THREADS_PER_ROW;
+
+            // Only the thread in the group which produced the max will reset the "winning" value to -inf.
+            if (thread_group_idx == thread_to_clear_in_group)
+            {
+                const int offset_for_expert = expert % ELTS_PER_LDG;
+                // Safe to set to any negative value since row_chunk values must be between 0 and 1.
+                row_chunk[ldg_group_for_expert * ELTS_PER_LDG + offset_for_expert] = -10000.f;
+            }
+        }
+    }
+}
+
+namespace detail
+{
+// Constructs some constants needed to partition the work across threads at compile time.
+template <int EXPERTS, int BYTES_PER_LDG>
+struct TopkConstants
+{
+    static constexpr int ELTS_PER_LDG = BYTES_PER_LDG / sizeof(float);
+    static_assert(EXPERTS / (ELTS_PER_LDG * WARP_SIZE) == 0 || EXPERTS % (ELTS_PER_LDG * WARP_SIZE) == 0, "");
+    static constexpr int VECs_PER_THREAD = std::max(1, EXPERTS / (ELTS_PER_LDG * WARP_SIZE));
+    static constexpr int VPT = VECs_PER_THREAD * ELTS_PER_LDG;
+    static constexpr int THREADS_PER_ROW = EXPERTS / VPT;
+    static constexpr int ROWS_PER_WARP = WARP_SIZE / THREADS_PER_ROW;
+};
+} // namespace detail
+
+template <int EXPERTS, int WARPS_PER_TB>
+void topkGatingSoftmaxLauncherHelper(const float* input, const bool* finished, float* output, int* indices,
+    int* source_row, const int num_rows, const int k, const int start_expert, const int end_expert, cudaStream_t stream)
+{
+    static constexpr std::size_t MAX_BYTES_PER_LDG = 16;
+
+    static constexpr int BYTES_PER_LDG = std::min(MAX_BYTES_PER_LDG, sizeof(float) * EXPERTS);
+    using Constants = detail::TopkConstants<EXPERTS, BYTES_PER_LDG>;
+    static constexpr int VPT = Constants::VPT;
+    static constexpr int ROWS_PER_WARP = Constants::ROWS_PER_WARP;
+    const int num_warps = (num_rows + ROWS_PER_WARP - 1) / ROWS_PER_WARP;
+    const int num_blocks = (num_warps + WARPS_PER_TB - 1) / WARPS_PER_TB;
+
+    dim3 block_dim(WARP_SIZE, WARPS_PER_TB);
+    topkGatingSoftmax<VPT, EXPERTS, WARPS_PER_TB, BYTES_PER_LDG><<<num_blocks, block_dim, 0, stream>>>(
+        input, finished, output, num_rows, indices, source_row, k, start_expert, end_expert);
+}
+
+#define LAUNCH_SOFTMAX(NUM_EXPERTS, WARPS_PER_TB)                       \
+    topkGatingSoftmaxLauncherHelper<NUM_EXPERTS, WARPS_PER_TB>(         \
+        gating_output, nullptr, topk_weights, topk_indicies,            \
+        token_expert_indices, num_tokens, topk, 0, num_experts,         \
+        stream);
+
+void topkGatingSoftmaxKernelLauncher(
+    const float* gating_output,
+    float* topk_weights,
+    int* topk_indicies,
+    int* token_expert_indices,
+    float* softmax_workspace,
+    const int num_tokens,
+    const int num_experts,
+    const int topk,
+    cudaStream_t stream) {
+    static constexpr int WARPS_PER_TB = 4;
+    switch (num_experts) {
+        case 1:
+            LAUNCH_SOFTMAX(1, WARPS_PER_TB);
+            break;
+        case 2:
+            LAUNCH_SOFTMAX(2, WARPS_PER_TB);
+            break;
+        case 4:
+            LAUNCH_SOFTMAX(4, WARPS_PER_TB);
+            break;
+        case 8:
+            LAUNCH_SOFTMAX(8, WARPS_PER_TB);
+            break;
+        case 16:
+            LAUNCH_SOFTMAX(16, WARPS_PER_TB);
+            break;
+        case 32:
+            LAUNCH_SOFTMAX(32, WARPS_PER_TB);
+            break;
+        case 64:
+            LAUNCH_SOFTMAX(64, WARPS_PER_TB);
+            break;
+        case 128:
+            LAUNCH_SOFTMAX(128, WARPS_PER_TB);
+            break;
+        case 256:
+            LAUNCH_SOFTMAX(256, WARPS_PER_TB);
+            break;
+        default: {
+            TORCH_CHECK(softmax_workspace != nullptr,
+                "softmax_workspace must be provided for num_experts that are not a power of 2.");
+            static constexpr int TPB = 256;
+            moeSoftmax<TPB><<<num_tokens, TPB, 0, stream>>>(
+                gating_output, nullptr, softmax_workspace, num_experts);
+            moeTopK<TPB><<<num_tokens, TPB, 0, stream>>>(
+                softmax_workspace, nullptr, topk_weights, topk_indicies, token_expert_indices,
+                num_experts, topk, 0, num_experts);
+        }
+    }
+}
+
+} // namespace moe
+} // namespace vllm
+
+void topk_softmax(
+    torch::Tensor& topk_weights,                // [num_tokens, topk]
+    torch::Tensor& topk_indices,                // [num_tokens, topk]
+    torch::Tensor& token_expert_indices,        // [num_tokens, topk]
+    torch::Tensor& gating_output)               // [num_tokens, num_experts]
+{
+    const int num_experts = gating_output.size(-1);
+    const int num_tokens = gating_output.numel() / num_experts;
+    const int topk = topk_weights.size(-1);
+
+    const bool is_pow_2 = (num_experts != 0) && ((num_experts & (num_experts - 1)) == 0);
+    const bool needs_workspace = !is_pow_2 || num_experts > 256;
+    const int64_t workspace_size = needs_workspace ? num_tokens * num_experts : 0;
+
+    const at::cuda::OptionalCUDAGuard device_guard(device_of(gating_output));
+    const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+    torch::Tensor softmax_workspace = torch::empty({workspace_size}, gating_output.options());
+    vllm::moe::topkGatingSoftmaxKernelLauncher(
+        gating_output.data_ptr<float>(),
+        topk_weights.data_ptr<float>(),
+        topk_indices.data_ptr<int>(),
+        token_expert_indices.data_ptr<int>(),
+        softmax_workspace.data_ptr<float>(),
+        num_tokens,
+        num_experts,
+        topk,
+        stream);
+}

csrc/moe_align_block_size_kernels.cu

@@ -7,10 +7,17 @@
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
 
-const static size_t NUM_MAX_EXPERTS = 64;
 #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;
+}
+}
+
 template <typename scalar_t>
 __global__ void moe_align_block_size_kernel(scalar_t *__restrict__ topk_ids, 
                                 int32_t *sorted_token_ids, 
@@ -21,10 +28,14 @@ __global__ void moe_align_block_size_kernel(scalar_t *__restrict__ topk_ids,
                                 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];
+
+    extern __shared__ int32_t shared_mem[];
+
+    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 (int i = 0; i < num_experts; ++i) {
-        tokens_cnts[threadIdx.x + 1][i] = 0;
+        tokens_cnts[index(num_experts, threadIdx.x + 1, i)] = 0;
     }
 
     /**
@@ -33,15 +44,15 @@ __global__ void moe_align_block_size_kernel(scalar_t *__restrict__ topk_ids,
     * 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]]; 
+        ++tokens_cnts[index(num_experts, threadIdx.x + 1, topk_ids[i])]; 
     }
 
     __syncthreads();
 
     // For each expert we accumulate the token counts from the different threads.
-    tokens_cnts[0][threadIdx.x] = 0;
+    tokens_cnts[index(num_experts, 0, threadIdx.x)] = 0;
     for (int i = 1; i <= blockDim.x; ++i) {
-        tokens_cnts[i][threadIdx.x] += tokens_cnts[i-1][threadIdx.x];
+        tokens_cnts[index(num_experts, i, threadIdx.x)] += tokens_cnts[index(num_experts, i-1, threadIdx.x)];
     }
 
     __syncthreads();
@@ -50,7 +61,7 @@ __global__ void moe_align_block_size_kernel(scalar_t *__restrict__ topk_ids,
     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;
+            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];
     }
@@ -78,9 +89,9 @@ __global__ void moe_align_block_size_kernel(scalar_t *__restrict__ topk_ids,
         * 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];
+        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[threadIdx.x][expert_id];
+        ++tokens_cnts[index(num_experts, threadIdx.x, expert_id)];
     }
 }
 }
@@ -93,11 +104,17 @@ void moe_align_block_size(
     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>(), 
+        // 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>(), 

csrc/ops.h

@@ -14,7 +14,8 @@ void paged_attention_v1(
   int block_size,
   int max_context_len,
   const c10::optional<torch::Tensor>& alibi_slopes,
-  const std::string& kv_cache_dtype);
+  const std::string& kv_cache_dtype,
+  float kv_scale);
 
 void paged_attention_v2(
   torch::Tensor& out,
@@ -31,7 +32,8 @@ void paged_attention_v2(
   int block_size,
   int max_context_len,
   const c10::optional<torch::Tensor>& alibi_slopes,
-  const std::string& kv_cache_dtype);
+  const std::string& kv_cache_dtype,
+  float kv_scale);
 
 void rms_norm(
   torch::Tensor& out,
@@ -53,10 +55,28 @@ void rotary_embedding(
   torch::Tensor& cos_sin_cache,
   bool is_neox);
 
+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 silu_and_mul(
   torch::Tensor& out,
   torch::Tensor& input);
 
+void gelu_and_mul(
+  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);
@@ -66,6 +86,21 @@ void gelu_fast(
   torch::Tensor& input);
 
 #ifndef USE_ROCM
+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 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,
@@ -80,6 +115,15 @@ torch::Tensor awq_dequantize(
     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);
 #endif
 
 void squeezellm_gemm(
@@ -94,11 +138,18 @@ torch::Tensor gptq_gemm(
   torch::Tensor b_gptq_qzeros,
   torch::Tensor b_gptq_scales,
   torch::Tensor b_g_idx,
-  bool use_exllama);
+  bool use_exllama,
+  int bit);
 
 void gptq_shuffle(
   torch::Tensor q_weight,
-  torch::Tensor q_perm);
+  torch::Tensor q_perm,
+  int bit);
+
+void scaled_fp8_quant(
+  torch::Tensor& out,
+  torch::Tensor& input,
+  torch::Tensor& scale);
 
 void moe_align_block_size(
   torch::Tensor topk_ids,

csrc/pos_encoding_kernels.cu

@@ -8,7 +8,7 @@
 namespace vllm {
 
 template<typename scalar_t, bool IS_NEOX>
-inline __device__ void apply_rotary_embedding(
+inline __device__ void apply_token_rotary_embedding(
   scalar_t* __restrict__ arr,
   const scalar_t* __restrict__ cos_ptr,
   const scalar_t* __restrict__ sin_ptr,
@@ -37,6 +37,42 @@ inline __device__ void apply_rotary_embedding(
   arr[y_index] = y * cos + x * sin;
 }
 
+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;
+
+  const int nq = num_heads * embed_dim;
+  for (int i = threadIdx.x; i < nq; i += blockDim.x) {
+    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_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;
+  for (int i = threadIdx.x; i < nk; i += blockDim.x) {
+    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_token_rotary_embedding<scalar_t, IS_NEOX>(key + token_head, cos_ptr,
+                                              sin_ptr, rot_offset, embed_dim);
+  }
+}
+
 template<typename scalar_t, bool IS_NEOX>
 __global__ void rotary_embedding_kernel(
   const int64_t* __restrict__ positions,        // [batch_size, seq_len] or [num_tokens]
@@ -54,27 +90,29 @@ __global__ void rotary_embedding_kernel(
   int64_t pos = positions[token_idx];
   const scalar_t* cache_ptr = cos_sin_cache + pos * rot_dim;
 
-  const int embed_dim = rot_dim / 2;
-  const scalar_t* cos_ptr = cache_ptr;
-  const scalar_t* sin_ptr = cache_ptr + embed_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);
+}
 
-  const int nq = num_heads * embed_dim;
-  for (int i = threadIdx.x; i < nq; i += blockDim.x) {
-    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);
-  }
+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;
 
-  const int nk = num_kv_heads * embed_dim;
-  for (int i = threadIdx.x; i < nk; i += blockDim.x) {
-    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_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
@@ -128,3 +166,61 @@ void rotary_embedding(
       }
     });
 }
+
+/*
+Batched version of rotary embedding, pack multiple LoRAs together
+and process in batched manner.
+*/
+void batched_rotary_embedding(
+  torch::Tensor& positions,         // [batch_size, seq_len] or [num_tokens]
+  torch::Tensor& query,             // [batch_size, seq_len, num_heads * head_size] or [num_tokens, num_heads * head_size]
+  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_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_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

@@ -14,20 +14,29 @@ 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, 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, 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) \
@@ -35,10 +44,15 @@ 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, 28672) \
     f(in_T, out_T, W_T, narrow, 32000) \
     f(in_T, out_T, W_T, narrow, 32256) \
@@ -46,8 +60,19 @@ 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) \
-// Keep above in sync with vllm/lora/layers::SamplerWithLoRA
+    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
 
 // Keep this in sync with vllm/config::LoRAConfig
 #define FOR_BGMV_WIDE_NARROW(f, in_T, out_T, W_T) \

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_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_fp32_bf16_fp16.cu

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

csrc/punica/bgmv/bgmv_fp32_fp16_bf16.cu

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

csrc/punica/bgmv/bgmv_fp32_fp32_bf16.cu

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

csrc/punica/bgmv/bgmv_fp32_fp32_fp16.cu

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

csrc/punica/bgmv/generator.py

@@ -10,7 +10,7 @@ TEMPLATE = """
 #include "bgmv_impl.cuh"
 
 FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, {input_dtype}, {output_dtype}, {weight_dtype})
-""".lstrip()
+""".lstrip()  # noqa: E501
 
 for input_dtype in DTYPES:
     for output_dtype in DTYPES:
@@ -18,6 +18,26 @@ for input_dtype in DTYPES:
             if weight_dtype == "fp32":
                 # FP32 weights are not supported.
                 continue
+            if output_dtype == "fp32":
+                # LoRA A matrix.
+                if input_dtype != weight_dtype:
+                    # NOTE(woosuk): While Punica supports the case where the
+                    # input and weight dtypes are different, we only generate
+                    # the kernels the same dtypes to reduce the binary size.
+                    continue
+            elif input_dtype == "fp32":
+                # LoRA B matrix.
+                if output_dtype != weight_dtype:
+                    # NOTE(woosuk): While Punica supports the case where the
+                    # output and weight dtypes are different, we only generate
+                    # the kernels the same dtypes to reduce the binary size.
+                    continue
+            elif not (input_dtype == output_dtype == weight_dtype):
+                # NOTE(woosuk): While Punica supports mixed data types for
+                # input, output, and weight, we only generate the kernels with
+                # the same data types to reduce the binary size.
+                continue
+
             kernel_definition = TEMPLATE.format(
                 input_dtype=DTYPE_MAP[input_dtype],
                 output_dtype=DTYPE_MAP[output_dtype],

csrc/punica/punica_ops.cc

@@ -1,7 +1,7 @@
 #include <cuda_bf16.h>
 #include <cuda_fp16.h>
 #include <torch/extension.h>
-
+#include <c10/cuda/CUDAGuard.h>
 #include <cstdint>
 
 #include "bgmv/bgmv_config.h"
@@ -20,8 +20,8 @@ inline void check_shape(const torch::Tensor &a, const torch::Tensor &b,
   }
 }
 
-inline constexpr uint32_t pack_u16(uint16_t a, uint16_t b) {
-  return (uint32_t(a) << 16) | uint32_t(b);
+inline constexpr uint64_t pack_u32(uint32_t a, uint32_t b) {
+  return (uint64_t(a) << 32) | uint64_t(b);
 }
 
 #define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")
@@ -46,13 +46,30 @@ inline constexpr uint32_t pack_u16(uint16_t a, uint16_t b) {
 template <typename in_T, typename out_T, typename W_T>
 inline bool launch_bgmv_kernel(out_T *Y, const in_T *X, const W_T *W,
                                const int64_t *lora_indices,
-                               uint16_t in_features, uint16_t out_features,
+                               uint32_t in_features, uint32_t out_features,
                                int64_t y_offset, int64_t full_y_size,
                                int64_t batch_size, int64_t num_layers,
                                int64_t layer_idx, float scale) {
-  switch (pack_u16(in_features, out_features)) {
+  // NOTE(woosuk): While Punica supports various combinations of input/output
+  // data types, we limit the supported data types to reduce the binary size.
+  constexpr bool is_input_float = std::is_same<in_T, float>::value;
+  constexpr bool is_output_float = std::is_same<out_T, float>::value;
+  if (is_input_float) {
+    if (!std::is_same<out_T, W_T>::value) {
+      return false;
+    }
+  } else if (is_output_float) {
+    if (!std::is_same<in_T, W_T>::value) {
+      return false;
+    }
+  } else if (!(std::is_same<in_T, W_T>::value &&
+               std::is_same<out_T, W_T>::value)) {
+    return false;
+  }
+
+  switch (pack_u32(in_features, out_features)) {
 #define CASE_ONESIDE(_in_T, _out_T, _W_T, feat_in, feat_out)                   \
-  case pack_u16(feat_in, feat_out):                                            \
+  case pack_u32(feat_in, feat_out):                                            \
     bgmv_kernel<feat_in, feat_out>(Y, X, W, lora_indices, y_offset,            \
                                    full_y_size, batch_size, num_layers,        \
                                    layer_idx, scale);                          \
@@ -91,8 +108,9 @@ void dispatch_bgmv(torch::Tensor y, torch::Tensor x, torch::Tensor w,
   CHECK_EQ(w.size(2), h_out);
   CHECK_EQ(indicies.size(0), x.size(0));
   CHECK_EQ(y.size(0), x.size(0));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
   bool ok = false;
-  if (h_in < 65536 && h_out < 65536) {
+  if (h_in <= 128512 && h_out <= 128512) {
     // TODO: See if we can get rid of this massive nested switch
     switch (x.scalar_type()) {
     case at::ScalarType::Half:
@@ -322,8 +340,9 @@ void dispatch_bgmv_low_level(torch::Tensor y, torch::Tensor x, torch::Tensor w,
   CHECK_EQ(w.size(2), h_out);
   CHECK_EQ(indicies.size(0), x.size(0));
   CHECK_EQ(y.size(0), x.size(0));
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(x));
   bool ok = false;
-  if (h_in < 65536 && h_out < 65536) {
+  if (h_in <= 128512 && h_out <= 128512) {
     // TODO: See if we can get rid of this massive nested switch
     switch (x.scalar_type()) {
     case at::ScalarType::Half:

csrc/pybind.cpp

@@ -22,6 +22,14 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     "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,
@@ -48,14 +56,24 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     &rotary_embedding,
     "Apply GPT-NeoX or GPT-J style rotary embedding to query and key");
 
+  ops.def(
+    "batched_rotary_embedding",
+    &batched_rotary_embedding,
+    "Apply GPT-NeoX or GPT-J style rotary embedding to query and key (supports multiple loras)");
+
+// Quantization ops
 #ifndef USE_ROCM
-  // Quantization ops
+  ops.def("aqlm_gemm", &aqlm_gemm, "Quantized GEMM for AQLM");
+  ops.def("aqlm_dequant", &aqlm_dequant, "Decompression method for AQLM");
   ops.def("awq_gemm", &awq_gemm, "Quantized GEMM for AWQ");
+  ops.def("marlin_gemm", &marlin_gemm, "Marlin Optimized Quantized GEMM for GPTQ");
   ops.def("awq_dequantize", &awq_dequantize, "Dequantization for AWQ");
 #endif
+ 
   ops.def("gptq_gemm", &gptq_gemm, "Quantized GEMM for GPTQ");
   ops.def("gptq_shuffle", &gptq_shuffle, "Post processing for GPTQ");
   ops.def("squeezellm_gemm", &squeezellm_gemm, "Quantized GEMM for SqueezeLLM");
+  ops.def("scaled_fp8_quant", &scaled_fp8_quant, "Compute FP8 quantized tensor and scaling factor");
   ops.def(
     "moe_align_block_size",
     &moe_align_block_size,
@@ -76,13 +94,9 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     &reshape_and_cache,
     "Reshape the key and value tensors and cache them");
   cache_ops.def(
-    "gather_cached_kv",
-    &gather_cached_kv,
-    "Gather key and value from the cache into contiguous QKV tensors");
-  cache_ops.def(
-    "convert_fp8_e5m2",
-    &convert_fp8_e5m2,
-    "Convert the key and value cache to fp8_e5m2 data type");
+    "convert_fp8",
+    &convert_fp8,
+    "Convert the key and value cache to fp8 data type");
 
   // Cuda utils
   pybind11::module cuda_utils = m.def_submodule("cuda_utils", "vLLM cuda utils");

csrc/quantization/aqlm/gemm_kernels.cu

@@ -0,0 +1,712 @@
+/*
+ * Modified by Neural Magic
+ * Adapted from https://github.com/Vahe1994/AQLM
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *         http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <cuda.h>
+#include <cuda_fp16.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+#include <c10/cuda/CUDAStream.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include <iostream>
+#include <cstdlib>
+
+
+namespace vllm {
+namespace aqlm {
+
+__global__ void Code1x16MatVec(
+  const int4* __restrict__ A,
+  const int4* __restrict__ B,
+        int4* __restrict__ C,
+  const int4* __restrict__ codebook,
+  const int prob_m,
+  const int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long.
+  const int codebook_stride // as int4.
+) {
+  int a_gl_stride = prob_k / 8 / 8;
+  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  bool pred = a_gl_rd < prob_m;
+
+  if (pred)
+  {
+    // advance to the correct codebook, this easy because we only multiply one column of the codebook.
+    auto codebook_size = &codebook_a_sizes.x;
+    while (a_gl_rd >= *codebook_size)
+    {
+        codebook += codebook_stride;
+        ++codebook_size;
+    }
+  }
+
+  int b_gl_rd = 0;
+  int c_gl_wr = a_gl_rd;
+  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
+  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
+
+  __shared__ int4 sh_b[32 * 9];
+  float res = 0;
+
+  int iters = (prob_k / 8 + 8 * 32 - 1) / (8 * 32);
+  while (iters--) {
+    // We pad shared memory to avoid bank conflicts during reads
+    __syncthreads();
+    for (int i = threadIdx.x; i < 32 * 8; i += blockDim.x) {
+      if (b_gl_rd + i < prob_k / 8)
+        sh_b[9 * (i / 8) + i % 8] = B[b_gl_rd + i];
+    }
+    __syncthreads();
+    b_gl_rd += 32 * 8;
+
+    int b_sh_rd = 9 * (threadIdx.x % 32);
+    if (pred && a_gl_rd < a_gl_end) {
+      const uint16_t* enc = reinterpret_cast<const uint16_t*>(&A[a_gl_rd]);
+      #pragma unroll
+      for (int i = 0; i < 8; i++) {
+        uint32_t dec[4];
+        // We bypass the L1 cache to avoid massive amounts of memory streaming that doesn't
+        // actually help us; this brings > 2x speedup.
+        asm volatile (
+          "ld.cg.global.v4.u32 {%0, %1, %2, %3}, [%4];"
+          : "=r"(dec[0]), "=r"(dec[1]), "=r"(dec[2]), "=r"(dec[3])
+          : "l"((void*) &codebook[enc[i]])
+        );
+        half2* a = reinterpret_cast<half2*>(&dec);
+        half2* b = reinterpret_cast<half2*>(&sh_b[b_sh_rd]);
+        half2 res2 = {};
+        #pragma unroll
+        for (int j = 0; j < 4; j++)
+          res2 = __hfma2(a[j], b[j], res2);
+        res += __half2float(res2.x) + __half2float(res2.y);
+        b_sh_rd++;
+      }
+      a_gl_rd += 32;
+    }
+  }
+
+  if (pred) {
+    #pragma unroll
+    for (int i = 16; i > 0; i /= 2)
+      res += __shfl_down_sync(0xffffffff, res, i);
+    if (threadIdx.x % 32 == 0)
+      reinterpret_cast<__half*>(C)[c_gl_wr] = __float2half(res);
+  }
+}
+
+__global__ void Code2x8MatVec(
+  const int4* __restrict__ A,
+  const int4* __restrict__ B,
+        int4* __restrict__ C,
+  const int4* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long.
+  const int codebook_stride // as int4.
+
+) {
+  int a_gl_stride = prob_k / 8 / 8;
+  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  bool pred = a_gl_rd < prob_m;
+
+  if (pred)
+  {
+    // advance to the correct codebook, this easy because we only multiply one column of the codebook.
+    auto codebook_size = &codebook_a_sizes.x;
+    while (a_gl_rd >= *codebook_size)
+    {
+        codebook += codebook_stride;
+        ++codebook_size;
+    }
+  }
+
+  int b_gl_rd = 0;
+  int c_gl_wr = a_gl_rd;
+  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
+  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
+  int lane = threadIdx.x % 8;
+
+  extern __shared__ int4 sh[];
+  int4* sh_b = sh;
+  int4* sh_code = sh_b + 32 * 9;
+  int4* sh_code0 = sh_code;
+  int4* sh_code1 = sh_code + 256 * 8;
+
+  for (int i = threadIdx.x; i < 2 * 256; i += blockDim.x) {
+    int4 dec = codebook[i];
+    #pragma unroll
+    for (int j = 0; j < 8; j++)
+      sh_code[8 * i + (j + lane) % 8] = dec;
+  }
+  __syncthreads();
+
+  float res = 0;
+
+  int iters = (prob_k / 8 + 8 * 32 - 1) / (8 * 32);
+  while (iters--) {
+    // We pad shared memory to avoid bank conflicts during reads
+    __syncthreads();
+    for (int i = threadIdx.x; i < 32 * 8; i += blockDim.x) {
+      if (b_gl_rd + i < prob_k / 8)
+        sh_b[9 * (i / 8) + i % 8] = B[b_gl_rd + i];
+    }
+    __syncthreads();
+    b_gl_rd += 32 * 8;
+
+    int b_sh_rd = 9 * (threadIdx.x % 32);
+    if (pred && a_gl_rd < a_gl_end) {
+      const uint8_t* enc = reinterpret_cast<const uint8_t*>(&A[a_gl_rd]);
+      #pragma unroll
+      for (int i = 0; i < 8; i++) {
+        half2* a0 = reinterpret_cast<half2*>(&sh_code0[8 * enc[2 * i + 0] + lane]);
+        half2* a1 = reinterpret_cast<half2*>(&sh_code1[8 * enc[2 * i + 1] + lane]);
+        half2*  b = reinterpret_cast<half2*>(&sh_b[b_sh_rd]);
+        half2 res2 = {};
+        #pragma unroll
+        for (int j = 0; j < 4; j++)
+          res2 = __hfma2(__hadd2(a0[j], a1[j]), b[j], res2);
+        res += __half2float(res2.x) + __half2float(res2.y);
+        b_sh_rd++;
+      }
+      a_gl_rd += 32;
+    }
+  }
+
+  if (pred) {
+    #pragma unroll
+    for (int i = 16; i > 0; i /= 2)
+      res += __shfl_down_sync(0xffffffff, res, i);
+    if (threadIdx.x % 32 == 0)
+      reinterpret_cast<__half*>(C)[c_gl_wr] = __float2half(res);
+  }
+}
+
+
+__global__ void Code1x16Dequant(
+  const int4* __restrict__ A,
+        int4* __restrict__ C,
+  const int4* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long, sums to m.
+  const int codebook_stride // as int4
+) {
+  int a_gl_stride = prob_k / 8 / 8;
+  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  bool pred = a_gl_rd < prob_m;
+
+  if (pred)
+  {
+    // advance to the correct codebook, this easy because we only multiply one column of the codebook.
+    auto codebook_size = &codebook_a_sizes.x;
+    while (a_gl_rd >= *codebook_size)
+    {
+        codebook += codebook_stride;
+        ++codebook_size;
+    }
+  }
+
+  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
+  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
+
+  int c_gl_stride = prob_k / 8;
+  int c_gl_wr = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  c_gl_wr = c_gl_stride * c_gl_wr + (threadIdx.x % 32) * 8;
+
+  int iters = (prob_k / 8 - 1) / (8 * 32) + 1;
+  while (iters--) {
+    if (pred && a_gl_rd < a_gl_end) {
+      const uint16_t* enc = reinterpret_cast<const uint16_t*>(&A[a_gl_rd]);
+      #pragma unroll
+      for (int i = 0; i < 8; i++) {
+        int4 chunk;
+        auto dec = reinterpret_cast<uint32_t*>(&chunk);
+        // We bypass the L1 cache to avoid massive amounts of memory streaming that doesn't
+        // actually help us; this brings > 2x speedup.
+        asm volatile (
+          "ld.cg.global.v4.u32 {%0, %1, %2, %3}, [%4];"
+          : "=r"(dec[0]), "=r"(dec[1]), "=r"(dec[2]), "=r"(dec[3])
+          : "l"((void*) &codebook[enc[i]])
+        );
+
+        C[a_gl_rd * 8 + i] = chunk;
+      }
+    }
+    a_gl_rd += 32;
+  }
+}
+
+
+__global__ void Code2x8Dequant(
+  const int4* __restrict__ A,
+        int4* __restrict__ C,
+  const int4* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long, corresponds to cols.
+  const int codebook_stride // as int4
+) {
+  int a_gl_stride = prob_k / 8 / 8;
+  int a_gl_rd = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  bool pred = a_gl_rd < prob_m;
+
+  if (pred)
+  {
+    // advance to the correct codebook, this easy because we only multiply one column of the codebook.
+    auto codebook_size = &codebook_a_sizes.x;
+    while (a_gl_rd >= *codebook_size)
+    {
+        codebook += codebook_stride;
+        ++codebook_size;
+    }
+  }
+
+  a_gl_rd = a_gl_stride * a_gl_rd + threadIdx.x % 32;
+  int a_gl_end = a_gl_rd + a_gl_stride - threadIdx.x % 32;
+  int lane = threadIdx.x % 8;
+
+  int c_gl_stride = prob_k / 8;
+  int c_gl_wr = (blockDim.x / 32) * blockIdx.x + (threadIdx.x / 32);
+  c_gl_wr = c_gl_stride * c_gl_wr + (threadIdx.x % 32) * 8;
+
+  extern __shared__ int4 sh[];
+  int4* sh_code = sh;
+  int4* sh_code0 = sh_code;
+  int4* sh_code1 = sh_code + 256 * 8;
+
+  for (int i = threadIdx.x; i < 2 * 256; i += blockDim.x) {
+    int4 dec = codebook[i];
+    #pragma unroll
+    for (int j = 0; j < 8; j++)
+      sh_code[8 * i + (j + lane) % 8] = dec;
+  }
+  __syncthreads();
+
+  float res = 0;
+
+  int iters = (prob_k / 8 - 1) / (8 * 32) + 1;
+  while (iters--) {
+    if (pred && a_gl_rd < a_gl_end) {
+      const uint8_t* enc = reinterpret_cast<const uint8_t*>(&A[a_gl_rd]);
+      #pragma unroll
+      for (int i = 0; i < 8; i++) {
+        int4 chunk;
+        half2* a0 = reinterpret_cast<half2*>(&sh_code0[8 * enc[2 * i + 0] + lane]);
+        half2* a1 = reinterpret_cast<half2*>(&sh_code1[8 * enc[2 * i + 1] + lane]);
+        #pragma unroll
+        for (int j = 0; j < 4; j++)
+          reinterpret_cast<half2*>(&chunk)[j] = __hadd2(a0[j], a1[j]);
+        C[a_gl_rd * 8 + i] = chunk;
+      }
+    }
+    a_gl_rd += 32;
+  }
+}
+
+inline int ceildiv(int a, int b) {
+  return (a + b - 1) / b;
+}
+
+const int THREAD_M = 16;
+
+void  code1x16_matvec_cuda(
+  const void* __restrict__ A,
+  const void* __restrict__ B,
+        void* __restrict__ C,
+  const void* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,
+  const int codebook_stride
+) {
+  int sms;
+  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
+  int waves = 0;
+  int thread_m;
+  do {
+    waves++;
+    thread_m = ceildiv(prob_m, waves * sms);
+  } while (thread_m > THREAD_M);
+
+  int blocks = ceildiv(prob_m, thread_m);
+  int threads = 32 * thread_m;
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
+  Code1x16MatVec<<<blocks, threads, 16*32*9, stream>>>(
+    (const int4*) A,
+    (const int4*) B,
+    (int4*) C,
+    (const int4*) codebook,
+    prob_m,
+    prob_k,
+    codebook_a_sizes,
+    codebook_stride
+  );
+}
+
+void  code2x8_matvec_cuda(
+  const void* __restrict__ A,
+  const void* __restrict__ B,
+        void* __restrict__ C,
+  const void* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,
+  const int codebook_stride
+) {
+  int sms;
+  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
+  int waves = 0;
+  int thread_m;
+  do {
+    waves++;
+    thread_m = ceildiv(prob_m, waves * sms);
+  } while (thread_m > THREAD_M);
+
+  int blocks = ceildiv(prob_m, thread_m);
+  int threads = 32 * thread_m;
+  int shared = 16 * (2 * 256 * 8 + 32 * 9);
+  cudaFuncSetAttribute(
+    Code2x8MatVec, cudaFuncAttributeMaxDynamicSharedMemorySize, shared
+  );
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
+  Code2x8MatVec<<<blocks, threads, shared, stream>>>(
+    (const int4*) A,
+    (const int4*) B,
+    (int4*) C,
+    (const int4*) codebook,
+    prob_m,
+    prob_k,
+    codebook_a_sizes,
+    codebook_stride
+  );
+}
+
+void code1x16_dequant_cuda(
+  const void* __restrict__ A,
+        void* __restrict__ C,
+  const void* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long.
+  const int codebook_stride // as int4.
+) {
+  int sms;
+  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
+  int waves = 0;
+  int thread_m;
+  do {
+    waves++;
+    thread_m = ceildiv(prob_m, waves * sms);
+  } while (thread_m > THREAD_M);
+
+  int blocks = ceildiv(prob_m, thread_m);
+  int threads = 32 * thread_m;
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
+  Code1x16Dequant<<<blocks, threads, 0, stream>>>(
+    (const int4*) A,
+    (int4*) C,
+    (const int4*) codebook,
+    prob_m,
+    prob_k,
+    codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long.
+    codebook_stride // as int4.
+  );
+}
+
+// Dequantizes the code and codebook into weights.
+void  code2x8_dequant_cuda(
+  const void* __restrict__ A,
+        void* __restrict__ C,
+  const void* __restrict__ codebook,
+  int prob_m,
+  int prob_k,
+  const int4 codebook_a_sizes,  // cumulative sizes of A spanning each codebook, at most 3 long, corresponds to cols.
+  const int codebook_stride // as int4
+) {
+  int sms;
+  cudaDeviceGetAttribute(&sms, cudaDevAttrMultiProcessorCount, 0);
+  int waves = 0;
+  int thread_m;
+  do {
+    waves++;
+    thread_m = ceildiv(prob_m, waves * sms);
+  } while (thread_m > THREAD_M);
+
+  int blocks = ceildiv(prob_m, thread_m);
+  int threads = 32 * thread_m;
+  int shared = 16 * (2 * 256 * 8 + 32 * 9);
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream().stream();
+
+  cudaFuncSetAttribute(
+    Code2x8Dequant, cudaFuncAttributeMaxDynamicSharedMemorySize, shared
+  );
+  Code2x8Dequant<<<blocks, threads, shared, stream>>>(
+    (const int4*) A,
+    (int4*) C,
+    (const int4*) codebook,
+    prob_m,
+    prob_k,
+    codebook_a_sizes,
+    codebook_stride
+  );
+}
+
+int codebook_stride(const torch::Tensor& codebooks)
+{
+  return codebooks.stride(0) * codebooks.element_size() / sizeof(int4);
+}
+
+void code1x16_matvec(
+  const torch::Tensor& A,
+  const torch::Tensor& B,
+        torch::Tensor& C,
+  const torch::Tensor& codebook,
+  const int4 codebook_a_sizes  // cumulative sizes of A spanning each codebook, at most 3 long.
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(A));
+  int prob_m = C.size(0);
+  int prob_k = B.size(0);
+
+  code1x16_matvec_cuda(
+    A.data_ptr(),
+    B.data_ptr(),
+    C.data_ptr(),
+    codebook.data_ptr(),
+    prob_m,
+    prob_k,
+    codebook_a_sizes,
+    codebook_stride(codebook)
+  );
+}
+
+torch::Tensor code1x16_matmat(
+  const torch::Tensor& input,
+  const torch::Tensor& codes,
+  const torch::Tensor& codebooks,
+  const torch::Tensor& scales,
+  const int4 codebook_a_sizes,
+  const std::optional<torch::Tensor>& bias) {
+  auto input_sizes = input.sizes();
+  auto out_features = codes.size(0) * codebooks.size(2);
+  auto flat_input = input.reshape({-1, input.size(-1)});
+  auto flat_output = torch::empty({flat_input.size(0), out_features},
+    torch::TensorOptions()
+      .dtype(input.dtype())
+      .device(input.device())
+  );
+
+  for (int i = 0; i < flat_input.size(0); ++i) {
+    auto input_vec = flat_input.index({i});
+    auto output_vec = flat_output.index({i});
+    code1x16_matvec(
+      codes.squeeze(2),
+      input_vec,
+      output_vec,
+      codebooks,
+      codebook_a_sizes
+    );
+  }
+  flat_output *= scales.flatten().unsqueeze(0);
+
+  if (bias.has_value()) {
+    flat_output += bias->unsqueeze(0);
+  }
+
+  auto output_sizes = input_sizes.vec();
+  output_sizes.pop_back();
+  output_sizes.push_back(-1);
+  auto output = flat_output.reshape(output_sizes);
+  return output;
+}
+
+void code2x8_matvec(
+  const torch::Tensor& A,
+  const torch::Tensor& B,
+        torch::Tensor& C,
+  const torch::Tensor& codebook,
+  const int4 codebook_a_sizes
+) {
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(A));
+  int prob_m = C.size(0);
+  int prob_k = B.size(0);
+  code2x8_matvec_cuda(
+    A.data_ptr(),
+    B.data_ptr(),
+    C.data_ptr(),
+    codebook.data_ptr(),
+    prob_m,
+    prob_k,
+    codebook_a_sizes,
+    2 * codebook_stride(codebook)
+  );
+}
+
+torch::Tensor code2x8_matmat(
+  const torch::Tensor& input,
+  const torch::Tensor& codes,
+  const torch::Tensor& codebooks,
+  const torch::Tensor& scales,
+  const int4 codebook_a_sizes,
+  const std::optional<torch::Tensor>& bias
+) {
+  auto input_sizes = input.sizes();
+  auto out_features = codes.size(0) * codebooks.size(2);
+  auto flat_input = input.reshape({-1, input.size(-1)});
+  auto flat_output = torch::empty({flat_input.size(0), out_features},
+    torch::TensorOptions()
+      .dtype(input.dtype())
+      .device(input.device())
+  );
+
+  for (int i = 0; i < flat_input.size(0); ++i) {
+    auto input_vec = flat_input.index({i});
+    auto output_vec = flat_output.index({i});
+    code2x8_matvec(
+      codes.squeeze(2),
+      input_vec,
+      output_vec,
+      codebooks,
+      codebook_a_sizes
+    );
+  }
+  flat_output *= scales.flatten().unsqueeze(0);
+  if (bias.has_value()) {
+    flat_output += bias->unsqueeze(0);
+  }
+
+  auto output_sizes = input_sizes.vec();
+  output_sizes.pop_back();
+  output_sizes.push_back(-1);
+  auto output = flat_output.reshape(output_sizes);
+  return output;
+}
+
+// Accumulate the partition sizes.
+int4 accumulate_sizes(const torch::Tensor& codebook_partition_sizes)
+{
+  int4 cumulative_sizes;
+  auto cumulative_size = &cumulative_sizes.x;
+  int i = 0;
+  int last = 0;
+  assert(codebook_partition_sizes.size(0) <= 4);
+  for (; i <  codebook_partition_sizes.size(0); ++i, ++cumulative_size)
+  {
+    *cumulative_size = codebook_partition_sizes[i].item<int>() + last;
+    last = *cumulative_size;
+  }
+  // fill in the rest with unreachable.
+  for (; i < 4; ++i, ++cumulative_size)
+  {
+    *cumulative_size = last*10;
+  }
+  return cumulative_sizes;
+}
+
+} // namespace aqlm
+} // namespace vllm
+
+
+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
+)
+{
+  int4 cumulative_sizes = vllm::aqlm::accumulate_sizes(codebook_partition_sizes);
+
+  int const nbooks = codebooks.size(0) / codebook_partition_sizes.size(0);
+  int const entries = codebooks.size(1);
+
+  if (nbooks == 1 && entries == (1 << 16))
+  { 
+    return vllm::aqlm::code1x16_matmat(input, codes, codebooks, scales, cumulative_sizes, bias);
+  }
+  if (nbooks == 2 && entries == (1 << 8))
+  {
+    return vllm::aqlm::code2x8_matmat(input, codes, codebooks, scales, cumulative_sizes, bias);
+  }
+
+  TORCH_CHECK(false, "AQLM with ", nbooks, " codebooks and ", entries, " entries is not currently supported.")
+  return {};
+}
+
+torch::Tensor aqlm_dequant(
+  const torch::Tensor& codes,
+  const torch::Tensor& codebooks,
+  const torch::Tensor& codebook_partition_sizes
+)
+{
+  int4 cumulative_sizes = vllm::aqlm::accumulate_sizes(codebook_partition_sizes);
+
+  int const nbooks = codebooks.size(0) / codebook_partition_sizes.size(0);
+  int const entries = codebooks.size(1);
+
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(codes));
+  int rows = codes.size(1);
+  int cols = codes.size(0);
+
+  auto in_features = codes.size(1) * 8;
+  auto out_features = codes.size(0);
+
+  assert(out_features = codebook_partition_sizes.sum().item<int>());
+
+  auto weights = torch::empty({out_features, in_features},
+    torch::TensorOptions()
+      .dtype(codebooks.dtype())
+      .device(codebooks.device())
+  );
+
+  if (nbooks == 1 && entries == (1 << 16))
+  {
+    vllm::aqlm::code1x16_dequant_cuda(
+      codes.data_ptr(),
+      weights.data_ptr(),
+      codebooks.data_ptr(),
+      out_features,
+      in_features,
+      cumulative_sizes,
+      vllm::aqlm::codebook_stride(codebooks));
+
+    // if you wanted to flip to scaling the weights, (though it's 30%-ish slower and not consistent with gemv implementation.)
+    // weights *= scales.index({"...", 0, 0});
+
+     return weights;
+  }
+
+  if (nbooks == 2 && entries == (1 << 8))
+  {
+     vllm::aqlm::code2x8_dequant_cuda(
+        codes.data_ptr(), 
+        weights.data_ptr(), 
+        codebooks.data_ptr(), 
+        out_features,
+        in_features, 
+        cumulative_sizes, 
+        vllm::aqlm::codebook_stride(codebooks));
+
+    // if you wanted to flip to scaling the weights, (though it's 30%-ish slower and not consistent with gemv implementation)
+    // weights *= scales.index({"...", 0, 0});
+
+     return weights;
+  }
+
+  TORCH_CHECK(false, "AQLM with ", nbooks, " codebooks and ", entries, " entries is not currently supported.")
+  return {};
+}

csrc/quantization/awq/gemm_kernels.cu

@@ -27,72 +27,85 @@ __pack_half2(const half x, const half y) {
   return (v1 << 16) | v0;
 }
 
-__global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, int split_k_iters, half* __restrict__ A, int* __restrict__ B, half* __restrict__ scaling_factors, int* __restrict__ zeros, int M, int IC, int OC, half* __restrict__ C) 
+template<int N>
+__global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16nXk32(
+  int G,
+  int split_k_iters,
+  half* __restrict__ A,
+  int* __restrict__ B,
+  half* __restrict__ scaling_factors,
+  int* __restrict__ zeros,
+  int M,
+  int IC,
+  int OC,
+  half* __restrict__ C)
 {
+  // Only support matrix n = 64 or 128
+  assert(N == 64 || N == 128);
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 750
   assert(false);
 #else
   static constexpr uint32_t ZERO = 0x0;
   float C_warp[32];
   __shared__ half A_shared[16 * (32 + 8)];
-  __shared__ half B_shared[32 * (128 + 8)];
-  
-  __shared__ half scaling_factors_shared[128];
-  __shared__ half zeros_shared[128];
+  __shared__ half B_shared[32 * (N + 8)];
 
-  int j_factors1 = ((OC + 128 - 1) / 128);
+  __shared__ half scaling_factors_shared[N];
+  __shared__ half zeros_shared[N];
+
+  int j_factors1 = ((OC + N - 1) / N);
   int blockIdx_x = 0;
   int blockIdx_y = blockIdx.x % ((M + 16 - 1) / 16 * j_factors1);
   int blockIdx_z = blockIdx.x / ((M + 16 - 1) / 16 * j_factors1);
 
   half A_shared_warp[8];
-  half B_shared_warp[32];
-  for (int j_0_4_init = 0; j_0_4_init < 4; ++j_0_4_init) {
+  half B_shared_warp[N / 4];
+  for (int j_0_4_init = 0; j_0_4_init < N / 32; ++j_0_4_init) {
     for (int i = 0; i < 8; ++i) {
       C_warp[(j_0_4_init * 8) + i] = 0.0;
     }
   }
 
   static constexpr int row_stride_warp = 32 * 8 / 32;
-  static constexpr int row_stride = 2 * 32 * 8 / 128;
-  bool ld_zero_flag = (threadIdx.y * 32 + threadIdx.x) * 8 < 128;
+  static constexpr int row_stride = 2 * 32 * 8 / N;
+  bool ld_zero_flag = (threadIdx.y * 32 + threadIdx.x) * 8 < N;
   // TODO: Haotian: blockIdx_y / j_factors1 in A loading to support bsz > 16
   bool ld_A_flag = (blockIdx_y / j_factors1 * 16 + threadIdx.y * row_stride_warp + threadIdx.x * 8 / 32) < M;     // threadIdx.y is warp_id
   // bool wb_C_flag = (threadIdx.x / 4) < M;
 
-  half* A_ptr = A 
+  half* A_ptr = A
                 + (((int)blockIdx_y) / j_factors1 * 16 + (((int)threadIdx.y) * row_stride_warp) + ((int)threadIdx.x) / (32 / 8)) * IC
                 + (((int)threadIdx.x) % (32 / 8)) * 8;
-  
+
   int* B_ptr = B
-            + ((int)threadIdx.y) * (OC / 8) * 2
-            + (((int)threadIdx.x) / (128 / 8)) * (OC / 8)
-            + (((int)blockIdx_y) % j_factors1) * (128 / 8)
-            + (((int)threadIdx.x) % (128 / 8)) * 1;
+            + ((int)threadIdx.y) * (OC / 8) * (256 / N)
+            + (((int)threadIdx.x) / (N / 8)) * (OC / 8)
+            + (((int)blockIdx_y) % j_factors1) * (N / 8)
+            + (((int)threadIdx.x) % (N / 8)) * 1;
 // Why * 1 in the above line?
-                        
-  half* A_shared_ptr = A_shared 
-                    + ((int)threadIdx.y) * row_stride_warp * (32 + 8) 
+
+  half* A_shared_ptr = A_shared
+                    + ((int)threadIdx.y) * row_stride_warp * (32 + 8)
                     + (((int)threadIdx.x) / (32 / 8)) * (32 + 8)
                     + (((int)threadIdx.x) % (32 / 8) ) * 8;
 
   half* B_shared_ptr = B_shared
-                    + ((int)threadIdx.y) * (row_stride / 2) * (128 + 8)
-                    + (((int)threadIdx.x) / (128 / 8)) * (128 + 8)
-                    + (((int)threadIdx.x) % (128 / 8)) * 8;
-  
-  int* zeros_ptr = zeros
-                + (((int)blockIdx_y) % j_factors1) * (128 / 8)
-                + ((int)threadIdx.x) % (128 / 8);
-  
-  half* scaling_factors_ptr = scaling_factors
-                            + (((int)blockIdx_y) % j_factors1) * (128) 
-                            + (((int)threadIdx.x) % (128 / 8)) * 8;
+                    + ((int)threadIdx.y) * (row_stride / 2) * (N + 8)
+                    + (((int)threadIdx.x) / (N / 8)) * (N + 8)
+                    + (((int)threadIdx.x) % (N / 8)) * 8;
 
-  half* C_ptr = C 
+  int* zeros_ptr = zeros
+                + (((int)blockIdx_y) % j_factors1) * (N / 8)
+                + ((int)threadIdx.x) % (N / 8);
+
+  half* scaling_factors_ptr = scaling_factors
+                            + (((int)blockIdx_y) % j_factors1) * N
+                            + (((int)threadIdx.x) % (N / 8)) * 8;
+
+  half* C_ptr = C
               + static_cast<long long>(blockIdx_z) * M * OC        // blockIdz.x -> split_k dim
-              + (((int)blockIdx_y) % j_factors1) * 128
-              + ((int)threadIdx.y) * 64
+              + (((int)blockIdx_y) % j_factors1) * N
+              + ((int)threadIdx.y) * (N / 2)
               + (((int)threadIdx.x) % 4) * 2;
 
   // preload s.f. and zeros
@@ -123,13 +136,13 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
     // uint4 B_loaded_scale = make_uint4(0, 0, 0, 0);
     int* B_ptr_local = B_ptr + k_0_0 * 32 * (OC / 8);
 
-    for (int ax0_ax1_fused_0 = 0; ax0_ax1_fused_0 < 8; ++ax0_ax1_fused_0) {
+    for (int ax0_ax1_fused_0 = 0; ax0_ax1_fused_0 < N / 16; ++ax0_ax1_fused_0) {
 
       // B: 32 x 136 (128+8) float16
       // each warp: 32 x 4
       // each thr: read 32 bit -> convert to 8xFP16 (a UINT4) -> scale and minus zero -> WB UINT4
       // *(uint4*)(B_shared + ((((ax0_ax1_fused_0 * 544) + (((int)threadIdx.y) * 272)) + ((((int)threadIdx.x) >> 4) * 136)) + ((((int)threadIdx.x) & 15) * 8))) = *(uint4*)(B + ((((((k_0_0 * 163840) + (ax0_ax1_fused_0 * 20480)) + (((int)threadIdx.y) * 10240)) + ((((int)threadIdx.x) >> 4) * 5120)) + (((int)blockIdx_y) * 128)) + ((((int)threadIdx.x) & 15) * 8)));
-      // row stride in shared memory: (NWARPS * 32 * 8 / cta_N) 
+      // row stride in shared memory: (NWARPS * 32 * 8 / cta_N)
       uint32_t B_loaded = *(uint32_t*)(B_ptr_local + ax0_ax1_fused_0 * row_stride * (OC / 8));
       uint4 B_loaded_fp16 = dequantize_s4_to_fp16x2(B_loaded);
       //uint4 B_loaded_zero = *(uint4*)(zeros_shared + (threadIdx.x % (cta_N / 8)) * 8);
@@ -152,7 +165,7 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
       */
 
       // write back
-      *(uint4*)(B_shared_ptr + ax0_ax1_fused_0 * row_stride * (128 + 8)) = B_loaded_fp16;
+      *(uint4*)(B_shared_ptr + ax0_ax1_fused_0 * row_stride * (N + 8)) = B_loaded_fp16;
     }
     __syncthreads();
 
@@ -174,13 +187,13 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
         );
       }
 
-      for (int ax1_0 = 0; ax1_0 < 4; ++ax1_0) {
+      for (int ax1_0 = 0; ax1_0 < N / 32; ++ax1_0) {
         {
           unsigned int addr;
           __asm__ __volatile__(
             "{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
             : "=r"(addr)
-            : "l"((void *)((&(B_shared[(((k_0_1 * 2176) + (((int)threadIdx.y) * 64)) + (ax1_0 * 16))])) + (((((int)threadIdx.x) & 15) * 136) + ((((int)threadIdx.x) >> 4) * 8))))
+            : "l"((void *)((&(B_shared[(((k_0_1 * (N * 16 + 128)) + (((int)threadIdx.y) * (N / 2))) + (ax1_0 * 16))])) + (((((int)threadIdx.x) & 15) * (N + 8)) + ((((int)threadIdx.x) >> 4) * 8))))
           );
           __asm__ __volatile__(
             "ldmatrix.sync.aligned.m8n8.x4.trans.shared.b16"
@@ -190,7 +203,7 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
           );
         }
       }
-      for (int j_0_4 = 0; j_0_4 < 4; ++j_0_4) {
+      for (int j_0_4 = 0; j_0_4 < N / 32; ++j_0_4) {
 #if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == 750
         {
           __asm__ __volatile__(
@@ -258,241 +271,6 @@ __global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n128k32(int G, i
 #endif
 }
 
-
-__global__ void __launch_bounds__(64) gemm_forward_4bit_cuda_m16n64k32(int G, int split_k_iters, half* __restrict__ A, int* __restrict__ B, half* __restrict__ scaling_factors, int* __restrict__ zeros, int M, int IC, int OC, half* __restrict__ C) 
-{
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 750
-  assert(false);
-#else
-  static constexpr uint32_t ZERO = 0x0;
-  float C_warp[32];
-  __shared__ half A_shared[16 * (32 + 8)];
-  __shared__ half B_shared[32 * (64 + 8)];
-  
-  __shared__ half scaling_factors_shared[64];
-  __shared__ half zeros_shared[64];
-
-  int j_factors1 = ((OC + 64 - 1) / 64);
-
-  int blockIdx_x = 0;
-  int blockIdx_y = blockIdx.x % ((M + 16 - 1) / 16 * j_factors1);
-  int blockIdx_z = blockIdx.x / ((M + 16 - 1) / 16 * j_factors1);
-
-  half A_shared_warp[8];
-  half B_shared_warp[16];
-  for (int j_0_4_init = 0; j_0_4_init < 2; ++j_0_4_init) {
-    for (int i = 0; i < 8; ++i) {
-      C_warp[(j_0_4_init * 8) + i] = 0.0;
-    }
-  }
-
-  static constexpr int row_stride_warp = 32 * 8 / 32;
-  static constexpr int row_stride = 2 * 32 * 8 / 64;
-  bool ld_zero_flag = (threadIdx.y * 32 + threadIdx.x) * 8 < 64;
-  // TODO: Haotian: blockIdx_y / j_factors1 in A loading to support bsz > 16
-  bool ld_A_flag = (blockIdx_y / j_factors1 * 16 + threadIdx.y * row_stride_warp + threadIdx.x * 8 / 32) < M;     // threadIdx.y is warp_id
-  // bool wb_C_flag = (threadIdx.x / 4) < M;
-
-  half* A_ptr = A 
-                + (((int)blockIdx_y) / j_factors1 * 16 + (((int)threadIdx.y) * row_stride_warp) + ((int)threadIdx.x) / (32 / 8)) * IC
-                + (((int)threadIdx.x) % (32 / 8)) * 8;
-  
-  int* B_ptr = B
-            + ((int)threadIdx.y) * (OC / 8) * 4
-            + (((int)threadIdx.x) / (64 / 8)) * (OC / 8)
-            + (((int)blockIdx_y) % j_factors1) * (64 / 8)
-            + (((int)threadIdx.x) % (64 / 8)) * 1;
-// Why * 1 in the above line?
-                        
-  half* A_shared_ptr = A_shared 
-                    + ((int)threadIdx.y) * row_stride_warp * (32 + 8) 
-                    + (((int)threadIdx.x) / (32 / 8)) * (32 + 8)
-                    + (((int)threadIdx.x) % (32 / 8) ) * 8;
-
-  half* B_shared_ptr = B_shared
-                    + ((int)threadIdx.y) * (row_stride / 2) * (64 + 8)
-                    + (((int)threadIdx.x) / (64 / 8)) * (64 + 8)
-                    + (((int)threadIdx.x) % (64 / 8)) * 8;
-  
-  int* zeros_ptr = zeros
-                + (((int)blockIdx_y) % j_factors1) * (64 / 8)
-                + ((int)threadIdx.x) % (64 / 8);
-  
-  half* scaling_factors_ptr = scaling_factors
-                            + (((int)blockIdx_y) % j_factors1) * (64) 
-                            + (((int)threadIdx.x) % (64 / 8)) * 8;
-
-  half* C_ptr = C 
-              + static_cast<long long>(blockIdx_z) * M * OC        // blockIdz.x -> split_k dim
-              + (((int)blockIdx_y) % j_factors1) * 64
-              + ((int)threadIdx.y) * 32
-              + (((int)threadIdx.x) % 4) * 2;
-
-  // preload s.f. and zeros
-  int k_bound = (IC / 32 + split_k_iters - 1) / split_k_iters;
-  if ((k_bound - 1) * split_k_iters * 32 + blockIdx_z * 32 >= IC) k_bound -= 1;
-  for (int _k_0_0 = 0; _k_0_0 < k_bound; ++_k_0_0) {
-    int k_0_0 = _k_0_0 * split_k_iters + blockIdx_z;
-    __syncthreads();
-    // TODO: Haotian: blockIdx_y / j_factors1 in A loading to support bsz > 16
-    if (ld_A_flag)
-    {
-      *(uint4*)(A_shared_ptr) = *(uint4*)(A_ptr + (k_0_0 * 32));
-    }
-    else
-    {
-      *(uint4*)(A_shared_ptr) = make_uint4(0, 0, 0, 0);
-    }
-
-    // for (int ax0_ax1_fused_0 = 0; ax0_ax1_fused_0 < 2; ++ax0_ax1_fused_0) {
-    uint32_t zeros_loaded = *(uint32_t*)(zeros_ptr + k_0_0 * 32 / G * (OC / 8));
-    uint4 B_loaded_zero = dequantize_s4_to_fp16x2(zeros_loaded);
-    uint4 B_loaded_scale = *(uint4*)(scaling_factors_ptr + k_0_0 * 32 / G * (OC));
-    /*
-    if (blockIdx_z == 0 && blockIdx_y == 0 && k_0_0 == 0 && threadIdx.x == 0 && threadIdx.y == 0){
-      printf("%x %x %x %x %x %x %x %x\n", B_loaded_scale.x, B_loaded_scale.y, B_loaded_scale.z, B_loaded_scale.w, B_loaded_zero.x, B_loaded_zero.y, B_loaded_zero.z, B_loaded_zero.w);
-    }
-    */
-    // uint4 B_loaded_scale = make_uint4(0, 0, 0, 0);
-    int* B_ptr_local = B_ptr + k_0_0 * 32 * (OC / 8);
-
-    for (int ax0_ax1_fused_0 = 0; ax0_ax1_fused_0 < 4; ++ax0_ax1_fused_0) {
-
-      // B: 32 x 136 (128+8) float16
-      // each warp: 32 x 4
-      // each thr: read 32 bit -> convert to 8xFP16 (a UINT4) -> scale and minus zero -> WB UINT4
-      // *(uint4*)(B_shared + ((((ax0_ax1_fused_0 * 544) + (((int)threadIdx.y) * 272)) + ((((int)threadIdx.x) >> 4) * 136)) + ((((int)threadIdx.x) & 15) * 8))) = *(uint4*)(B + ((((((k_0_0 * 163840) + (ax0_ax1_fused_0 * 20480)) + (((int)threadIdx.y) * 10240)) + ((((int)threadIdx.x) >> 4) * 5120)) + (((int)blockIdx_y) * 128)) + ((((int)threadIdx.x) & 15) * 8)));
-      // row stride in shared memory: (NWARPS * 32 * 8 / cta_N) 
-      uint32_t B_loaded = *(uint32_t*)(B_ptr_local + ax0_ax1_fused_0 * row_stride * (OC / 8));
-      uint4 B_loaded_fp16 = dequantize_s4_to_fp16x2(B_loaded);
-      //uint4 B_loaded_zero = *(uint4*)(zeros_shared + (threadIdx.x % (cta_N / 8)) * 8);
-
-      // uint4 B_loaded_scale = *(uint4*)(scaling_factors_shared + (threadIdx.x % (cta_N / 8)) * 8);
-      // - zero and * scale
-      // TODO (Haotian): can save 4 assembly instructions if sormulate as deq = q * scale - zero * scale.
-      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_zero.x));
-      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_scale.x), "r"(ZERO));
-      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_zero.y));
-      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_scale.y), "r"(ZERO));
-      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_zero.z));
-      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_scale.z), "r"(ZERO));
-      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_zero.w));
-      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_scale.w), "r"(ZERO));
-      /*
-      if (ax0_ax1_fused_0 == 0 && blockIdx_z == 0 && blockIdx_y == 0 && k_0_0 == 0 && threadIdx.x == 17 && threadIdx.y == 0){
-        printf("[x] %X %X %X %X\n", B_loaded_fp16.x, B_loaded_fp16.y, B_loaded_fp16.z, B_loaded_fp16.w);
-      }
-      */
-
-      // write back
-      *(uint4*)(B_shared_ptr + ax0_ax1_fused_0 * row_stride * (64 + 8)) = B_loaded_fp16;
-    }
-    __syncthreads();
-
-    for (int k_0_1 = 0; k_0_1 < 2; ++k_0_1) 
-    {
-      {
-        unsigned int addr;
-        __asm__ __volatile__(
-          "{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
-          : "=r"(addr)
-          : "l"((void *)((&(A_shared[(k_0_1 * 16)])) + (((((int)threadIdx.x) & 15) * 40) + ((((int)threadIdx.x) >> 4) * 8))))
-        );
-        __asm__ __volatile__(
-          "ldmatrix.sync.aligned.m8n8.x4.shared.b16"
-          "{%0, %1, %2, %3}, [%4];\n"
-          : "=r"(((unsigned *)(A_shared_warp + 0))[0]), "=r"(((unsigned *)(A_shared_warp + 0))[1]), "=r"(((unsigned *)(A_shared_warp + 0))[2]), "=r"(((unsigned *)(A_shared_warp + 0))[3])
-          : "r"(addr)
-        );
-      }
-        
-
-      for (int ax1_0 = 0; ax1_0 < 2; ++ax1_0) 
-      {
-        {
-          unsigned int addr;
-          __asm__ __volatile__(
-            "{ .reg .u64 addr; cvta.to.shared.u64 addr, %1; cvt.u32.u64 %0, addr; }\n"
-            : "=r"(addr)
-            : "l"((void *)((&(B_shared[(((k_0_1 * 1152) + (((int)threadIdx.y) * 32)) + (ax1_0 * 16))])) + (((((int)threadIdx.x) & 15) * 72) + ((((int)threadIdx.x) >> 4) * 8))))
-          );
-          __asm__ __volatile__(
-            "ldmatrix.sync.aligned.m8n8.x4.trans.shared.b16"
-            "{%0, %1, %2, %3}, [%4];\n"
-            : "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[0]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[1]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[2]), "=r"(((unsigned *)(B_shared_warp + (ax1_0 * 8)))[3])
-            : "r"(addr)
-          );
-        }
-      }
-      
-      for (int j_0_4 = 0; j_0_4 < 2; ++j_0_4) 
-      {
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ == 750
-        {
-          __asm__ __volatile__(
-            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
-            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
-            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
-            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
-        }
-
-        {
-          __asm__ __volatile__(
-            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
-            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
-            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
-            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
-        }
-
-        {
-          __asm__ __volatile__(
-            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
-            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
-            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
-            : "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
-        }
-
-        {
-          __asm__ __volatile__(
-            "mma.sync.aligned.m16n8k8.row.col.f32.f16.f16.f32"
-            "{%0, %1, %2, %3}, {%4, %5}, {%6}, {%7, %8, %9, %10};\n"
-            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
-            : "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
-        }
-#else
-        {
-          __asm__ __volatile__(
-            "mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
-            "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%10, %11, %12, %13};\n"
-            :  "=f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "=f"(((float *)(C_warp + (j_0_4 * 8)))[3])
-            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[0]), "r"(((unsigned *)(B_shared_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[0]), "f"(((float *)(C_warp + (j_0_4 * 8)))[1]), "f"(((float *)(C_warp + (j_0_4 * 8)))[2]), "f"(((float *)(C_warp + (j_0_4 * 8)))[3]));
-        }
-
-        {
-          __asm__ __volatile__(
-            "mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32"
-            "{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%10, %11, %12, %13};\n"
-            :  "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "=f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3])
-            : "r"(((unsigned *)(A_shared_warp + 0))[0]), "r"(((unsigned *)(A_shared_warp + 0))[1]), "r"(((unsigned *)(A_shared_warp + 0))[2]), "r"(((unsigned *)(A_shared_warp + 0))[3]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[0]), "r"(((unsigned *)(B_shared_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[0]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[1]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[2]), "f"(((float *)(C_warp + ((j_0_4 * 8) + 4)))[3]));
-        }
-#endif
-      }
-    }
-  }
-
-// TODO: Shang: Hoist loop invariance.
-  for (int ax1_0_1 = 0; ax1_0_1 < 2; ++ax1_0_1) {
-    for (int local_id = 0; local_id < 8; ++local_id) {
-      int row_offset = (((int)blockIdx_y) / j_factors1) * 16 + ((int)threadIdx.x) / 4 + (local_id % 4) / 2 * 8;
-      if (row_offset < M)
-      {
-        *(C_ptr + ax1_0_1 * 16 + row_offset * OC + (local_id / 4) * 8 + local_id % 2) = __float2half(C_warp[(ax1_0_1 * 8) + local_id]);
-      }
-    }
-  }
-#endif
-}
-
 __global__ void __launch_bounds__(64) dequantize_weights(
     int* __restrict__ B,
     half* __restrict__ scaling_factors,
@@ -526,26 +304,24 @@ __global__ void __launch_bounds__(64) dequantize_weights(
   int index4 = 8 * col + (int)(row / G) * N * 8;
   half* scaling_factors_ptr2 = scaling_factors + index4;
 
+  uint32_t zeros_loaded = *(uint32_t*)(zeros_ptr2);
+  uint4 B_loaded_zero = dequantize_s4_to_fp16x2(zeros_loaded);
+  uint4 B_loaded_scale = *(uint4*)(scaling_factors_ptr2);
 
-    uint32_t zeros_loaded = *(uint32_t*)(zeros_ptr2);
-    uint4 B_loaded_zero = dequantize_s4_to_fp16x2(zeros_loaded);
-    uint4 B_loaded_scale = *(uint4*)(scaling_factors_ptr2);
-int j=0;
+  uint32_t B_loaded = *(uint32_t*)B_ptr2;
+  uint4 B_loaded_fp16 = dequantize_s4_to_fp16x2(B_loaded);
+  asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_zero.x));
+  asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_scale.x), "r"(ZERO));
+  asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_zero.y));
+  asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_scale.y), "r"(ZERO));
+  asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_zero.z));
+  asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_scale.z), "r"(ZERO));
+  asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_zero.w));
+  asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_scale.w), "r"(ZERO));
 
-      uint32_t B_loaded = *(uint32_t*)(B_ptr2 + j);
-      uint4 B_loaded_fp16 = dequantize_s4_to_fp16x2(B_loaded);
-      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_zero.x));
-      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.x) : "r"(B_loaded_fp16.x), "r"(B_loaded_scale.x), "r"(ZERO));
-      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_zero.y));
-      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.y) : "r"(B_loaded_fp16.y), "r"(B_loaded_scale.y), "r"(ZERO));
-      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_zero.z));
-      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.z) : "r"(B_loaded_fp16.z), "r"(B_loaded_scale.z), "r"(ZERO));
-      asm volatile("sub.f16x2 %0, %1, %2;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_zero.w));
-      asm volatile("fma.rn.f16x2 %0, %1, %2, %3;\n" : "=r"(B_loaded_fp16.w) : "r"(B_loaded_fp16.w), "r"(B_loaded_scale.w), "r"(ZERO));
+  *(uint4*)B_shared_ptr2 = B_loaded_fp16;
 
-      *(uint4*)(B_shared_ptr2 + j) = B_loaded_fp16;
-
-  for (int i=0; i<8; ++i) {
+  for (int i = 0; i < 8; ++i) {
     *(C_ptr2 + i) = B_shared[i];
   }
 }
@@ -650,19 +426,21 @@ torch::Tensor awq_gemm(
         // threadIdx.x: 32
         // threadIdx.y: i_factors[2] * j_factors[2]
         dim3 threads_per_block(32, 2);
-        vllm::awq::gemm_forward_4bit_cuda_m16n128k32<<<num_blocks, threads_per_block, 0, stream>>>(
-            group_size, split_k_iters, in_feats, kernel, scaling_factors, zeros, num_in_feats, num_in_channels, num_out_channels, out_feats);
+        vllm::awq::gemm_forward_4bit_cuda_m16nXk32<128><<<num_blocks, threads_per_block, 0, stream>>>(
+            group_size, split_k_iters, in_feats, kernel, scaling_factors, zeros, num_in_feats, num_in_channels,
+            num_out_channels, out_feats);
     }
     else if (num_out_channels % 64 == 0)
     {
         int j_factors1 = num_out_channels / 64 / 1;
         dim3 num_blocks(1 * (num_out_feats + 16 - 1) / 16 * j_factors1 * split_k_iters);
-    
+
         // threadIdx.x: 32
         // threadIdx.y: i_factors[2] * j_factors[2]
         dim3 threads_per_block(32, 2);
-        vllm::awq::gemm_forward_4bit_cuda_m16n64k32<<<num_blocks, threads_per_block, 0, stream>>>(
-            group_size, split_k_iters, in_feats, kernel, scaling_factors, zeros, num_in_feats, num_in_channels, num_out_channels, out_feats);
+        vllm::awq::gemm_forward_4bit_cuda_m16nXk32<64><<<num_blocks, threads_per_block, 0, stream>>>(
+            group_size, split_k_iters, in_feats, kernel, scaling_factors, zeros, num_in_feats, num_in_channels,
+            num_out_channels, out_feats);
     }
     return _out_feats.sum(0);
 }

csrc/quantization/fp8/amd_detail/hip_float8.h

@@ -0,0 +1,167 @@
+#pragma once
+
+#ifdef __HIPCC__
+#include <hip/hip_runtime.h>
+#else
+#include <type_traits>
+#include <stdint.h>
+#include <math.h>
+#include <iostream>
+#endif
+
+#include "hip_float8_impl.h"
+
+struct alignas(1) hip_fp8
+{
+    struct from_bits_t
+    {
+    };
+    HIP_FP8_HOST_DEVICE static constexpr from_bits_t from_bits() { return from_bits_t(); }
+    uint8_t data;
+
+    hip_fp8() = default;
+    HIP_FP8_HOST_DEVICE constexpr hip_fp8(const hip_fp8&) = default;
+    HIP_FP8_HOST_DEVICE constexpr hip_fp8(uint8_t v) = delete;
+    explicit HIP_FP8_HOST_DEVICE constexpr hip_fp8(uint8_t v, from_bits_t)
+        : data(v)
+    {
+    }
+
+#ifdef __HIP__MI300__
+    // NOTE: ON-DEVICE... always optimal bias
+    explicit HIP_FP8_DEVICE hip_fp8(float v)
+        : data(hip_fp8_impl::to_fp8_from_fp32(v))
+    {
+    }
+
+    explicit HIP_FP8_DEVICE hip_fp8(_Float16 v)
+        : hip_fp8(static_cast<float>(v))
+    {
+    }
+
+    // Host only implementation using s/w simulation
+    explicit HIP_FP8_HOST
+#else  // __HIP__MI300__
+    // both Host and DEVICE for non-MI300 using s/w simulation
+    explicit HIP_FP8_HOST_DEVICE
+#endif // __HIP__MI300__
+    hip_fp8(float v)
+    {
+        data = hip_fp8_impl::to_float8<4, 3, float, true /*negative_zero_nan*/, true /*clip*/>(v);
+    }
+
+    explicit HIP_FP8_HOST_DEVICE hip_fp8(double v)
+        : hip_fp8(static_cast<float>(v))
+    {
+    }
+
+#ifdef __HIP__MI300__
+    // upcast using device specific intrinsic
+    explicit inline HIP_FP8_DEVICE operator float() const
+    {
+        float fval;
+        uint32_t i32val = static_cast<uint32_t>(data);
+
+        // upcast
+        asm volatile("v_cvt_f32_fp8 %0, %1 src0_sel:BYTE_0" : "=v"(fval) : "v"(i32val));
+
+        return fval;
+    }
+
+    explicit inline HIP_FP8_HOST operator float() const
+#else  // __HIP__MI300__
+    explicit inline HIP_FP8_HOST_DEVICE operator float() const
+#endif // __HIP__MI300__
+    {
+        return hip_fp8_impl::from_float8<4, 3, float, true /*negative_zero_nan*/>(data);
+    }
+};
+
+namespace std
+{
+inline hip_fp8 sin(hip_fp8 a)
+{
+    return hip_fp8(sinf(float(a)));
+}
+inline hip_fp8 cos(hip_fp8 a)
+{
+    return hip_fp8(cosf(float(a)));
+}
+HIP_FP8_HOST_DEVICE constexpr hip_fp8 real(const hip_fp8& a)
+{
+    return a;
+}
+} // namespace std
+
+// Special operator overloading
+inline std::ostream& operator<<(std::ostream& os, const hip_fp8& f8)
+{
+    return os << float(f8);
+}
+
+// all + operator overloading with mixed types
+// mixed types, always converts to f32, does computation in f32, and returns float
+inline HIP_FP8_HOST_DEVICE float operator+(const float fa, hip_fp8 b)
+{
+    return (fa + float(b));
+}
+
+inline HIP_FP8_HOST_DEVICE float operator+(hip_fp8 a, const float fb)
+{
+    return (float(a) + fb);
+}
+
+inline HIP_FP8_HOST_DEVICE hip_fp8 operator+(hip_fp8 a, hip_fp8 b)
+{
+    return hip_fp8(float(a) + float(b));
+}
+
+inline HIP_FP8_HOST_DEVICE hip_fp8& operator+=(hip_fp8& a, hip_fp8 b)
+{
+    return a = hip_fp8(float(a) + float(b));
+}
+
+// overloading multiplication, always returns float,
+inline HIP_FP8_HOST_DEVICE float operator*(hip_fp8 a, hip_fp8 b)
+{
+    return float(a) * float(b);
+}
+
+inline HIP_FP8_HOST_DEVICE float operator*(float a, hip_fp8 b)
+{
+    return (a * float(b));
+}
+
+inline HIP_FP8_HOST_DEVICE float operator*(hip_fp8 a, float b)
+{
+    return (float(a) * b);
+}
+
+inline HIP_FP8_HOST_DEVICE float operator*(int32_t a, hip_fp8 b)
+{
+    return ((float)a * float(b));
+}
+
+inline HIP_FP8_HOST_DEVICE float operator*(double a, hip_fp8 b)
+{
+    return ((float)a * float(b));
+}
+
+// overloading for compare
+inline HIP_FP8_HOST_DEVICE bool operator==(hip_fp8 a, hip_fp8 b)
+{
+    return (a.data == b.data);
+}
+inline HIP_FP8_HOST_DEVICE bool operator!=(hip_fp8 a, hip_fp8 b)
+{
+    return (a.data != b.data);
+}
+
+inline HIP_FP8_HOST_DEVICE bool operator>=(hip_fp8 a, hip_fp8 b)
+{
+    return static_cast<float>(a) >= static_cast<float>(b);
+}
+inline HIP_FP8_HOST_DEVICE bool operator>(hip_fp8 a, hip_fp8 b)
+{
+    return static_cast<float>(a) > static_cast<float>(b);
+}

csrc/quantization/fp8/amd_detail/hip_float8_impl.h

@@ -0,0 +1,316 @@
+#pragma once
+
+#if defined(__HIPCC__) && (defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+#define __HIP__MI300__
+#endif
+
+#ifdef __HIPCC__
+#define HIP_FP8_HOST_DEVICE __host__ __device__
+#define HIP_FP8_HOST __host__
+#define HIP_FP8_DEVICE __device__
+#else
+#define HIP_FP8_HOST_DEVICE
+#define HIP_FP8_HOST
+#define HIP_FP8_DEVICE
+#endif
+
+namespace hip_fp8_impl
+{
+
+#ifdef __HIP__MI300__
+HIP_FP8_DEVICE uint8_t to_fp8_from_fp32(float v)
+{
+    uint8_t i8data;
+    union {
+        float fval;
+        uint32_t i32val;
+        uint8_t i8val[4]; // NOTE: not endian independent
+    } val;
+
+    uint32_t ival = 0;
+    val.fval = v;
+
+    if ((val.i32val & 0x7F800000) != 0x7F800000) { /// propagate NAN/INF, no clipping
+        val.fval = __builtin_amdgcn_fmed3f(val.fval, 240.0, -240.0);
+    }
+
+    ival = __builtin_amdgcn_cvt_pk_fp8_f32(val.fval, val.fval, ival,
+        false); // false -> WORD0
+    val.i32val = ival;
+    i8data = val.i8val[0];
+
+    return i8data;
+}
+#endif // __HIP__MI300__
+
+HIP_FP8_HOST inline int clz(uint32_t x)
+{
+    return __builtin_clz(x);
+}
+#if defined(__HIPCC__) || defined(__CUDA_ARCH__)
+HIP_FP8_DEVICE inline int clz(uint32_t x)
+{
+    return __clz(x);
+}
+#endif
+
+template <int we, int wm, typename T, bool negative_zero_nan, bool clip>
+HIP_FP8_HOST_DEVICE uint8_t to_float8(T _x, bool stoch = false, uint32_t rng = 0)
+{
+#ifdef __HIPCC__
+    constexpr bool is_half = std::is_same<T, _Float16>::value;
+#else
+    constexpr bool is_half = false;
+#endif
+    constexpr bool is_float = std::is_same<T, float>::value;
+    static_assert(wm + we == 7, "wm+we==7");
+    static_assert(is_half || is_float, "Only half and float can be cast to f8");
+
+    const int mfmt = (sizeof(T) == 4) ? 23 : 10;
+    uint32_t x;
+    if (sizeof(T) == 4) {
+        x = reinterpret_cast<uint32_t&>(_x);
+    } else {
+        x = reinterpret_cast<uint16_t&>(_x);
+    }
+
+    uint32_t head, mantissa;
+    int exponent, bias;
+    uint32_t sign;
+
+    if (sizeof(T) == 4) {
+        head = x & 0xFF800000;
+        mantissa = x & 0x7FFFFF;
+        exponent = (head >> 23) & 0xFF;
+        sign = head >> 31;
+        bias = 127;
+    } else {
+        head = x & 0xFC00;
+        mantissa = x & 0x3FF;
+        exponent = (head >> 10) & 0x1F;
+        sign = head >> 15;
+        bias = 15;
+    }
+
+    uint32_t signed_inf = (sign << 7) + (((1 << we) - 1) << wm);
+
+    // Deal with inf and NaNs
+    if (negative_zero_nan) {
+        if (sizeof(T) == 4) {
+            if ((x & 0x7F800000) == 0x7F800000) {
+                return 0x80;
+            }
+        } else {
+            // if(__hisinf(x) || __hisnan(x))
+            if ((x & 0x7C00) == 0x7C00) {
+                return 0x80;
+            }
+        }
+    } else {
+        if (sizeof(T) == 4) {
+            if ((x & 0x7F800000) == 0x7F800000) {
+                return signed_inf + (mantissa != 0 ? 1 : 0);
+            }
+        } else {
+            if ((x & 0x7C00) == 0x7C00) {
+                return signed_inf + (mantissa != 0 ? 1 : 0);
+            }
+        }
+    }
+    if (x == 0) {
+        return 0;
+    }
+
+    // First need to check if it is normal or denorm as there is a difference of
+    // implicit 1 Then need to adjust the exponent to align with the F8 exponent,
+    // in the meanwhile, shift The mantissa. Then for stochastic rounding, add rng
+    // to mantissa and truncate. And for RNE, no need to add rng. Then probably
+    // need to check whether there is carry and adjust exponent and mantissa again
+
+    // For IEEE bias mode, the bias is 2^(k-1) -1 where k is the width of exponent
+    // bits
+    const int f8_bias = (1 << (we - 1)) - 1 + (negative_zero_nan ? 1 : 0);
+    const int f8_denormal_act_exponent = 1 - f8_bias; // actual exponent of f8 denormal
+    // act_exponent is the actual exponent of fp32/fp16 (after subtracting bias)
+    // f8_exponent is the converted f8 exponent with bias encoding
+    // exponent_diff is the diff between fp32/fp16 exponent and f8 exponent,
+    // the difference needs to be adjusted and mantissa shifted
+    int act_exponent, f8_exponent, exponent_diff;
+
+    if (exponent == 0) { // fp32/fp16 is in denormal.
+        /* fp32 denormal is below 2^-127 so it is usually not a concern here, we
+mostly concern fp16 here. In this case, f8 is usually in denormal. But there
+could be exceptions. fp16 denormal has exponent bias 15 while bf8 with NANOO has
+exponent bias 16. It means that there are some numbers in fp16 denormal but they
+are bf8 (NANOO) normals - smallest bf8 (NANOO) normal is 2^-15. fp16 numbers
+where exponent==0 (actual exponent -14) and highest bit of mantissa is 1 are bf8
+(NANOO) normal. In this case, the fp16 mantissa should be shift left by 1  */
+        act_exponent = exponent - bias + 1;
+        exponent_diff = f8_denormal_act_exponent - act_exponent; // actual exponent is exponent-bias+1 as it is denormal
+    } else {                                                     // fp32/fp16 is normal with implicit 1
+        act_exponent = exponent - bias;
+        if (act_exponent <= f8_denormal_act_exponent) {
+            /* This is the case where fp32/fp16 is normal but it is in f8 denormal
+ range. For example fp8 nanoo mode, denormal exponent is -7, but if the
+ fp32/fp16 actual exponent is -7, it is actually larger due to the implicit 1,
+ Therefore it needs to be adjust to -6 and mantissa shift right by 1.
+ So for fp32/fp16, exponent -8 is the cut point to convert to fp8 nanoo */
+            exponent_diff = f8_denormal_act_exponent - act_exponent;
+        } else {               // both fp32/fp16 and f8 are in normal range
+            exponent_diff = 0; // exponent_diff=0 does not mean there is no difference
+                               // for this case,
+                               // act_exponent could be larger. Just that it does not need shift mantissa
+        }
+        mantissa += (1 << mfmt); // Add the implicit 1 into mantissa
+    }
+
+    bool midpoint = (mantissa & ((1 << (mfmt - wm + exponent_diff)) - 1)) ==
+                    static_cast<uint32_t>(1 << (mfmt - wm + exponent_diff - 1));
+    /* This part is a bit tricky. The judgment of whether it is a tie needs to be
+   done before we shift right as shift right could rip off some residual part
+   and make something not midpoint look like midpoint. For example, the fp16
+   number 0x1002 (0 00100 0000000010), it is larger than midpoint, but after
+   shift right by 4 bits, it would look like midpoint.
+*/
+
+    if (exponent_diff > 0) {
+        mantissa >>= exponent_diff;
+    } else if (exponent_diff == -1) {
+        mantissa <<= -exponent_diff;
+    }
+    bool implicit_one = mantissa & (1 << mfmt);
+    // if there is no implicit 1, it  means the f8 is denormal and need to adjust
+    // to denorm exponent
+    f8_exponent = (act_exponent + exponent_diff) /*actual f8 exponent*/ + f8_bias - (implicit_one ? 0 : 1);
+
+    // Now we have the exponent and mantissa adjusted
+    uint32_t drop_mask = (1 << (mfmt - wm)) - 1;
+    bool odd = mantissa & (1 << (mfmt - wm)); // if the least significant bit that
+                                              // is not truncated is 1
+    mantissa += (stoch ? rng : (midpoint ? (odd ? mantissa : mantissa - 1) : mantissa)) & drop_mask;
+
+    // Now we deal with overflow
+    if (f8_exponent == 0) {
+        if ((1 << mfmt) & mantissa) {
+            f8_exponent = 1; // denormal overflow to become normal, promote exponent
+        }
+    } else {
+        if ((1 << (mfmt + 1)) & mantissa) {
+            mantissa >>= 1;
+            f8_exponent++;
+        }
+    }
+
+    mantissa >>= (mfmt - wm);
+
+    // above range: quantize to maximum possible float of the same sign
+    const int max_exp = (1 << we) - (negative_zero_nan ? 1 : 2);
+    if (f8_exponent > max_exp) {
+        if (clip) {
+            mantissa = (1 << wm) - 1;
+            f8_exponent = max_exp;
+        } else {
+            return signed_inf;
+        }
+    }
+
+    if (f8_exponent == 0 && mantissa == 0) {
+        return negative_zero_nan ? 0 : (sign << 7);
+    }
+    mantissa &= (1 << wm) - 1;
+    return (sign << 7) | (f8_exponent << wm) | mantissa;
+}
+
+template <int we, int wm, typename T = float, bool negative_zero_nan = true>
+inline HIP_FP8_HOST_DEVICE T from_float8(uint8_t x)
+{
+#ifdef __HIPCC__
+    constexpr bool is_half = std::is_same<T, _Float16>::value;
+#else
+    constexpr bool is_half = false;
+#endif
+    constexpr bool is_float = std::is_same<T, float>::value;
+    static_assert(is_half || is_float, "only half and float are supported");
+
+    constexpr int weo = is_half ? 5 : 8;
+    constexpr int wmo = is_half ? 10 : (is_float ? 23 : 7);
+
+    T fInf, fNegInf, fNaN, fNeg0;
+
+#ifdef __HIPCC__
+    if (is_half) {
+        const uint16_t ihInf = 0x7C00;
+        const uint16_t ihNegInf = 0xFC00;
+        const uint16_t ihNaN = 0x7C01;
+        const uint16_t ihNeg0 = 0x8000;
+        fInf = reinterpret_cast<const _Float16&>(ihInf);
+        fNegInf = reinterpret_cast<const _Float16&>(ihNegInf);
+        fNaN = reinterpret_cast<const _Float16&>(ihNaN);
+        fNeg0 = reinterpret_cast<const _Float16&>(ihNeg0);
+    } else
+#endif
+        if (is_float) {
+        const uint32_t ifInf = 0x7F800000;
+        const uint32_t ifNegInf = 0xFF800000;
+        const uint32_t ifNaN = 0x7F800001;
+        const uint32_t ifNeg0 = 0x80000000;
+        fInf = reinterpret_cast<const float&>(ifInf);
+        fNegInf = reinterpret_cast<const float&>(ifNegInf);
+        fNaN = reinterpret_cast<const float&>(ifNaN);
+        fNeg0 = reinterpret_cast<const float&>(ifNeg0);
+    }
+
+    if (x == 0) {
+        return 0;
+    }
+
+    uint32_t sign = x >> 7;
+    uint32_t mantissa = x & ((1 << wm) - 1);
+    int exponent = (x & 0x7F) >> wm;
+    if (negative_zero_nan) {
+        if (x == 0x80) {
+            return fNaN;
+        }
+    } else {
+        if (x == 0x80) {
+            return fNeg0;
+        }
+        if (exponent == ((1 << we) - 1)) {
+            return (mantissa == 0) ? (sign ? fNegInf : fInf) : fNaN;
+        }
+    }
+    typename std::conditional<sizeof(T) == 2, uint16_t, uint32_t>::type retval;
+    if (we == 5 && is_half && !negative_zero_nan) {
+        retval = x << 8;
+        return reinterpret_cast<const T&>(retval);
+    }
+
+    const int exp_low_cutoff = (1 << (weo - 1)) - (1 << (we - 1)) + 1 - (negative_zero_nan ? 1 : 0);
+
+    // subnormal input
+    if (exponent == 0) {
+        // guaranteed mantissa!=0 since cases 0x0 and 0x80 are handled above
+        int sh = 1 + clz(mantissa) - (32 - wm);
+        mantissa <<= sh;
+        exponent += 1 - sh;
+        mantissa &= ((1 << wm) - 1);
+    }
+    exponent += exp_low_cutoff - 1;
+    mantissa <<= wmo - wm;
+
+    // subnormal output (occurs when T=half, we=5, negative_zero_nan=true)
+    if (exponent <= 0) {
+        mantissa |= 1 << wmo;
+        mantissa >>= 1 - exponent;
+        exponent = 0;
+    }
+
+    if (sizeof(T) == 2) {
+        retval = (sign << 15) | (exponent << 10) | mantissa;
+    } else {
+        retval = (sign << 31) | (exponent << 23) | mantissa;
+    }
+    return reinterpret_cast<const T&>(retval);
+}
+
+} // namespace hip_fp8_impl

csrc/quantization/fp8/amd_detail/quant_utils.cuh

@@ -0,0 +1,517 @@
+#pragma once
+#include "hip_float8.h"
+
+#include <hip/hip_fp16.h>
+#include <hip/hip_bf16.h>
+#include <hip/hip_bfloat16.h>
+
+#include "../../../attention/dtype_float32.cuh"
+#include "../../../attention/dtype_bfloat16.cuh"
+
+namespace vllm
+{
+namespace fp8_e4m3 {
+template <typename Tout, typename Tin>
+__inline__ __device__ Tout vec_conversion(const Tin& x)
+{
+    return x;
+}
+
+template <typename Tout, typename Tin>
+__inline__ __device__ Tout scaled_vec_conversion(const Tin& x, const float scale)
+{
+    return x;
+}
+
+// fp8 -> half
+template <>
+__inline__ __device__ uint16_t vec_conversion<uint16_t, uint8_t>(const uint8_t& a)
+{
+    hip_fp8 f8{a, hip_fp8::from_bits()};
+    __half_raw res;
+    res.data = static_cast<float>(f8);
+    return res.x;
+}
+
+// fp8x2 -> half2
+template <>
+__inline__ __device__ uint32_t vec_conversion<uint32_t, uint16_t>(const uint16_t& a)
+{
+#if defined(__HIP__MI300__) && defined(__HIP_FP8_EXPERIMENTAL_BULK_CONVERT__)
+    const auto& f2 = __builtin_amdgcn_cvt_pk_f32_fp8(a, 0);
+    union {
+        __half2_raw h2r;
+        uint32_t ui32;
+    } tmp;
+    tmp.h2r.x.data = f2[0];
+    tmp.h2r.y.data = f2[1];
+    return tmp.ui32;
+#else
+    union {
+        uint16_t u16[2];
+        uint32_t u32;
+    } tmp;
+
+    tmp.u16[0] = vec_conversion<uint16_t, uint8_t>(static_cast<uint8_t>(a));
+    tmp.u16[1] = vec_conversion<uint16_t, uint8_t>(static_cast<uint8_t>(a >> 8U));
+    return tmp.u32;
+#endif
+}
+
+// fp8x4 -> half2x2
+template <>
+__inline__ __device__ uint2 vec_conversion<uint2, uint32_t>(const uint32_t& a)
+{
+    union {
+        uint2 u32x2;
+        uint32_t u32[2];
+    } tmp;
+    tmp.u32[0] = vec_conversion<uint32_t, uint16_t>((uint16_t)a);
+    tmp.u32[1] = vec_conversion<uint32_t, uint16_t>((uint16_t)(a >> 16U));
+    return tmp.u32x2;
+}
+
+// fp8x8 -> half2x4
+template <>
+__inline__ __device__ uint4 vec_conversion<uint4, uint2>(const uint2& a)
+{
+    union {
+        uint4 u64x2;
+        uint2 u64[2];
+    } tmp;
+    tmp.u64[0] = vec_conversion<uint2, uint32_t>(a.x);
+    tmp.u64[1] = vec_conversion<uint2, uint32_t>(a.y);
+    return tmp.u64x2;
+}
+
+using __nv_bfloat16 = __hip_bfloat16;
+
+// fp8 -> __nv_bfloat16
+template <>
+__inline__ __device__ __nv_bfloat16 vec_conversion<__nv_bfloat16, uint8_t>(const uint8_t& a)
+{
+    hip_fp8 f8{a, hip_fp8::from_bits()};
+    float f{f8};
+    return __float2bfloat16(f);
+}
+
+using __nv_bfloat162 = __hip_bfloat162;
+
+// fp8x2 -> __nv_bfloat162
+template <>
+__inline__ __device__ __nv_bfloat162 vec_conversion<__nv_bfloat162, uint16_t>(const uint16_t& a)
+{
+    __nv_bfloat162 res;
+    res.x = vec_conversion<__nv_bfloat16, uint8_t>((uint8_t)a);
+    res.y = vec_conversion<__nv_bfloat16, uint8_t>((uint8_t)(a >> 8U));
+    return res;
+}
+
+// fp8x4 -> bf16_4_t
+template <>
+__inline__ __device__ bf16_4_t vec_conversion<bf16_4_t, uint32_t>(const uint32_t& a)
+{
+    bf16_4_t res;
+    res.x = vec_conversion<__nv_bfloat162, uint16_t>((uint16_t)a);
+    res.y = vec_conversion<__nv_bfloat162, uint16_t>((uint16_t)(a >> 16U));
+    return res;
+}
+
+// fp8x8 -> bf16_8_t
+template <>
+__inline__ __device__ bf16_8_t vec_conversion<bf16_8_t, uint2>(const uint2& a)
+{
+    bf16_4_t tmp1, tmp2;
+    tmp1 = vec_conversion<bf16_4_t, uint32_t>(a.x);
+    tmp2 = vec_conversion<bf16_4_t, uint32_t>(a.y);
+    bf16_8_t res;
+    res.x = tmp1.x;
+    res.y = tmp1.y;
+    res.z = tmp2.x;
+    res.w = tmp2.y;
+    return res;
+}
+
+// fp8 -> float
+template <>
+__inline__ __device__ float vec_conversion<float, uint8_t>(const uint8_t& a)
+{
+    hip_fp8 fp8{a, hip_fp8::from_bits()};
+    return static_cast<float>(fp8);
+}
+
+// fp8x2 -> float2
+template <>
+__inline__ __device__ float2 vec_conversion<float2, uint16_t>(const uint16_t& a)
+{
+#if defined(__HIP__MI300__) && defined(__HIP_FP8_EXPERIMENTAL_BULK_CONVERT__)
+    float2 res;
+    const auto& f2 = __builtin_amdgcn_cvt_pk_f32_fp8(a, 0);
+    res.x = f2[0];
+    res.y = f2[1];
+    return res;
+#else
+    float2 res;
+    res.x = vec_conversion<float, uint8_t>(static_cast<uint8_t>(a));
+    res.y = vec_conversion<float, uint8_t>(static_cast<uint8_t>(a >> 8U));
+    return res;
+#endif
+}
+
+// fp8x4 -> float4
+template <>
+__inline__ __device__ Float4_ vec_conversion<Float4_, uint32_t>(const uint32_t& a)
+{
+    Float4_ res;
+    res.x = vec_conversion<float2, uint16_t>((uint16_t)a);
+    res.y = vec_conversion<float2, uint16_t>((uint16_t)(a >> 16U));
+    return res;
+}
+
+// fp8x8 -> float8
+template <>
+__inline__ __device__ Float8_ vec_conversion<Float8_, uint2>(const uint2& a)
+{
+    Float4_ tmp1, tmp2;
+    tmp1 = vec_conversion<Float4_, uint32_t>(a.x);
+    tmp2 = vec_conversion<Float4_, uint32_t>(a.y);
+    Float8_ res;
+    res.x = tmp1.x;
+    res.y = tmp1.y;
+    res.z = tmp2.x;
+    res.w = tmp2.y;
+    return res;
+}
+
+// half -> fp8
+template <>
+__inline__ __device__ uint8_t vec_conversion<uint8_t, uint16_t>(const uint16_t& a)
+{
+    __half_raw tmp;
+    tmp.x = a;
+
+    hip_fp8 f8{static_cast<float>(tmp.data)};
+    return f8.data;
+}
+
+// bf16 -> fp8
+template <>
+__inline__ __device__ uint8_t vec_conversion<uint8_t, __nv_bfloat16>(const __nv_bfloat16& a)
+{
+    hip_fp8 res{__bfloat162float(a)};
+    return res.data;
+}
+
+// float -> fp8
+template <>
+__inline__ __device__ uint8_t vec_conversion<uint8_t, float>(const float& a)
+{
+    hip_fp8 f8(a);
+    return f8.data;
+}
+
+// fp8x4 -> float4
+template <>
+__inline__ __device__ float4 vec_conversion<float4, uint32_t>(const uint32_t& a)
+{
+    Float4_ tmp = vec_conversion<Float4_, uint32_t>(a);
+    float4 res = make_float4(tmp.x.x, tmp.x.y, tmp.y.x, tmp.y.y);
+    return res;
+}
+
+// float2 -> half2
+template <>
+__inline__ __device__ uint32_t vec_conversion<uint32_t, float2>(const float2& a)
+{
+    union {
+        half2 float16;
+        uint32_t uint32;
+    };
+
+    float16 = __float22half2_rn(a);
+    return uint32;
+}
+
+// Float4 -> half2x2
+template <>
+__inline__ __device__ uint2 vec_conversion<uint2, Float4_>(const Float4_& a)
+{
+    uint2 b;
+    float2 val;
+    val.x = a.x.x;
+    val.y = a.x.y;
+    b.x = vec_conversion<uint32_t, float2>(val);
+
+    val.x = a.y.x;
+    val.y = a.y.y;
+    b.y = vec_conversion<uint32_t, float2>(val);
+    return b;
+}
+
+// Float4 -> float4
+template <>
+__inline__ __device__ float4 vec_conversion<float4, Float4_>(const Float4_& a)
+{
+    float4 b;
+    b.x = a.x.x;
+    b.y = a.x.y;
+    b.z = a.y.x;
+    b.w = a.y.y;
+    return b;
+}
+
+// Float8 -> half2x4
+template <>
+__inline__ __device__ uint4 vec_conversion<uint4, Float8_>(const Float8_& a)
+{
+    uint4 b;
+    b.x = vec_conversion<uint32_t, float2>(a.x);
+    b.y = vec_conversion<uint32_t, float2>(a.y);
+    b.z = vec_conversion<uint32_t, float2>(a.z);
+    b.w = vec_conversion<uint32_t, float2>(a.w);
+    return b;
+}
+
+// float2 -> bfloat162
+template <>
+__inline__ __device__ __nv_bfloat162 vec_conversion<__nv_bfloat162, float2>(const float2& a)
+{
+    __nv_bfloat162 b = __float22bfloat162_rn(a);
+    return b;
+}
+
+// Float4 -> bfloat162x2
+template <>
+__inline__ __device__ bf16_4_t vec_conversion<bf16_4_t, Float4_>(const Float4_& a)
+{
+    bf16_4_t b;
+    b.x = __float22bfloat162_rn(a.x);
+    b.y = __float22bfloat162_rn(a.y);
+    return b;
+}
+
+// Float8 -> bfloat162x4
+template <>
+__inline__ __device__ bf16_8_t vec_conversion<bf16_8_t, Float8_>(const Float8_& a)
+{
+    bf16_8_t b;
+    b.x = __float22bfloat162_rn(a.x);
+    b.y = __float22bfloat162_rn(a.y);
+    b.z = __float22bfloat162_rn(a.z);
+    b.w = __float22bfloat162_rn(a.w);
+    return b;
+}
+
+
+/* Scaled and vectorized conversions, for data exchange between high and low precision domains
+
+   Convention of the scale in API, e.g: FP8_data = Quantization( High_Precision_data / scale )
+   s.t.
+     Quantize(HP / scale) => FP8
+     Dequant(FP8) * scale =>  HP
+
+ */
+
+// fp8 -> half
+template <>
+__inline__ __device__ uint16_t scaled_vec_conversion<uint16_t, uint8_t>(const uint8_t& a, const float scale)
+{
+    hip_fp8 f8{a, hip_fp8::from_bits()};
+    __half_raw res;
+    res.data = static_cast<float>(f8) * scale;
+    return res.x;
+}
+
+// fp8x2 -> half2
+template <>
+__inline__ __device__ uint32_t scaled_vec_conversion<uint32_t, uint16_t>(const uint16_t& a, const float scale)
+{
+#if defined(__HIP__MI300__) && defined(__HIP_FP8_EXPERIMENTAL_BULK_CONVERT__)
+    const auto& f2 = __builtin_amdgcn_cvt_pk_f32_fp8(a, 0);
+    union {
+        __half2_raw h2r;
+        uint32_t ui32;
+    } tmp;
+    tmp.h2r.x.data = f2[0] * scale;
+    tmp.h2r.y.data = f2[1] * scale;
+    return tmp.ui32;
+#else
+    union {
+        uint16_t u16[2];
+        uint32_t u32;
+    } tmp;
+
+    tmp.u16[0] = scaled_vec_conversion<uint16_t, uint8_t>(static_cast<uint8_t>(a), scale);
+    tmp.u16[1] = scaled_vec_conversion<uint16_t, uint8_t>(static_cast<uint8_t>(a >> 8U), scale);
+    return tmp.u32;
+#endif
+}
+
+// fp8x4 -> half2x2
+template <>
+__inline__ __device__ uint2 scaled_vec_conversion<uint2, uint32_t>(const uint32_t& a, const float scale)
+{
+    union {
+        uint2 u32x2;
+        uint32_t u32[2];
+    } tmp;
+    tmp.u32[0] = scaled_vec_conversion<uint32_t, uint16_t>((uint16_t)a, scale);
+    tmp.u32[1] = scaled_vec_conversion<uint32_t, uint16_t>((uint16_t)(a >> 16U), scale);
+    return tmp.u32x2;
+}
+
+// fp8x8 -> half2x4
+template <>
+__inline__ __device__ uint4 scaled_vec_conversion<uint4, uint2>(const uint2& a, const float scale)
+{
+    union {
+        uint4 u64x2;
+        uint2 u64[2];
+    } tmp;
+    tmp.u64[0] = scaled_vec_conversion<uint2, uint32_t>(a.x, scale);
+    tmp.u64[1] = scaled_vec_conversion<uint2, uint32_t>(a.y, scale);
+    return tmp.u64x2;
+}
+
+using __nv_bfloat16 = __hip_bfloat16;
+
+// fp8 -> __nv_bfloat16
+template <>
+__inline__ __device__ __nv_bfloat16 scaled_vec_conversion<__nv_bfloat16, uint8_t>(const uint8_t& a, const float scale)
+{
+    hip_fp8 f8{a, hip_fp8::from_bits()};
+    float f{f8};
+    return __float2bfloat16(f * scale);
+}
+
+using __nv_bfloat162 = __hip_bfloat162;
+
+// fp8x2 -> __nv_bfloat162
+template <>
+__inline__ __device__ __nv_bfloat162 scaled_vec_conversion<__nv_bfloat162, uint16_t>(const uint16_t& a, const float scale)
+{
+    __nv_bfloat162 res;
+    res.x = scaled_vec_conversion<__nv_bfloat16, uint8_t>((uint8_t)a, scale);
+    res.y = scaled_vec_conversion<__nv_bfloat16, uint8_t>((uint8_t)(a >> 8U), scale);
+    return res;
+}
+
+// fp8x4 -> bf16_4_t
+template <>
+__inline__ __device__ bf16_4_t scaled_vec_conversion<bf16_4_t, uint32_t>(const uint32_t& a, const float scale)
+{
+    bf16_4_t res;
+    res.x = scaled_vec_conversion<__nv_bfloat162, uint16_t>((uint16_t)a, scale);
+    res.y = scaled_vec_conversion<__nv_bfloat162, uint16_t>((uint16_t)(a >> 16U), scale);
+    return res;
+}
+
+// fp8x8 -> bf16_8_t
+template <>
+__inline__ __device__ bf16_8_t scaled_vec_conversion<bf16_8_t, uint2>(const uint2& a, const float scale)
+{
+    bf16_4_t tmp1, tmp2;
+    tmp1 = scaled_vec_conversion<bf16_4_t, uint32_t>(a.x, scale);
+    tmp2 = scaled_vec_conversion<bf16_4_t, uint32_t>(a.y, scale);
+    bf16_8_t res;
+    res.x = tmp1.x;
+    res.y = tmp1.y;
+    res.z = tmp2.x;
+    res.w = tmp2.y;
+    return res;
+}
+
+// fp8 -> float
+template <>
+__inline__ __device__ float scaled_vec_conversion<float, uint8_t>(const uint8_t& a, const float scale)
+{
+    hip_fp8 fp8{a, hip_fp8::from_bits()};
+    return static_cast<float>(fp8) * scale;
+}
+
+// fp8x2 -> float2
+template <>
+__inline__ __device__ float2 scaled_vec_conversion<float2, uint16_t>(const uint16_t& a, const float scale)
+{
+#if defined(__HIP__MI300__) && defined(__HIP_FP8_EXPERIMENTAL_BULK_CONVERT__)
+    float2 res;
+    const auto& f2 = __builtin_amdgcn_cvt_pk_f32_fp8(a, 0);
+    res.x = f2[0] * scale;
+    res.y = f2[1] * scale;
+    return res;
+#else
+    float2 res;
+    res.x = scaled_vec_conversion<float, uint8_t>(static_cast<uint8_t>(a), scale);
+    res.y = scaled_vec_conversion<float, uint8_t>(static_cast<uint8_t>(a >> 8U), scale);
+    return res;
+#endif
+}
+
+// fp8x4 -> float4
+template <>
+__inline__ __device__ Float4_ scaled_vec_conversion<Float4_, uint32_t>(const uint32_t& a, const float scale)
+{
+    Float4_ res;
+    res.x = scaled_vec_conversion<float2, uint16_t>((uint16_t)a, scale);
+    res.y = scaled_vec_conversion<float2, uint16_t>((uint16_t)(a >> 16U), scale);
+    return res;
+}
+
+// fp8x8 -> float8
+template <>
+__inline__ __device__ Float8_ scaled_vec_conversion<Float8_, uint2>(const uint2& a, const float scale)
+{
+    Float4_ tmp1, tmp2;
+    tmp1 = scaled_vec_conversion<Float4_, uint32_t>(a.x, scale);
+    tmp2 = scaled_vec_conversion<Float4_, uint32_t>(a.y, scale);
+    Float8_ res;
+    res.x = tmp1.x;
+    res.y = tmp1.y;
+    res.z = tmp2.x;
+    res.w = tmp2.y;
+    return res;
+}
+
+
+/* Quantize(HP / scale) => FP8 */
+
+// TODO(Hai): vectorized to add
+
+// half -> fp8
+template <>
+__inline__ __device__ uint8_t scaled_vec_conversion<uint8_t, uint16_t>(const uint16_t& a, const float scale)
+{
+    __half_raw tmp;
+    tmp.x = a;
+
+    hip_fp8 f8{static_cast<float>(tmp.data)/scale};
+    return f8.data;
+}
+
+// bf16 -> fp8
+template <>
+__inline__ __device__ uint8_t scaled_vec_conversion<uint8_t, __nv_bfloat16>(const __nv_bfloat16& a, const float scale)
+{
+    hip_fp8 res{__bfloat162float(a)/scale};
+    return res.data;
+}
+
+// float -> fp8
+template <>
+__inline__ __device__ uint8_t scaled_vec_conversion<uint8_t, float>(const float& a, const float scale)
+{
+    hip_fp8 f8(a/scale);
+    return f8.data;
+}
+
+// fp8x4 -> float4
+template <>
+__inline__ __device__ float4 scaled_vec_conversion<float4, uint32_t>(const uint32_t& a, const float scale)
+{
+    Float4_ tmp = scaled_vec_conversion<Float4_, uint32_t>(a, scale);
+    float4 res = make_float4(tmp.x.x, tmp.x.y, tmp.y.x, tmp.y.y);
+    return res;
+}
+
+}
+} // namespace vllm

csrc/quantization/fp8/fp8_cuda_kernels.cu

@@ -0,0 +1,103 @@
+#include <ATen/cuda/CUDAContext.h>
+#include <torch/extension.h>
+#include <c10/cuda/CUDAGuard.h>
+
+#include <cmath>
+
+#include "cuda_compat.h"
+#include "dispatch_utils.h"
+
+namespace vllm {
+
+__device__ __forceinline__ float atomicMaxFloat(float* addr, float value) {
+    float old;
+    old = (value >= 0) ? __int_as_float(atomicMax((int*)addr, __float_as_int(value))) :
+         __uint_as_float(atomicMin((unsigned int*)addr, __float_as_uint(value)));
+
+    return old;
+}
+
+// Compute the absolute maximum m of the input tensor and store
+// m / float8_e4m3::max() in *scale. Each thread block performs a
+// reduction tree and the memory in scale is atomically updated.
+// So to get the right answer, *scale needs to be initialized to
+// a value <= 0.0 and we need to wait for all thread blocks to
+// finish before consuming *scale.
+template<typename scalar_t>
+__global__ void segmented_max_reduction(
+  float* __restrict__ scale,
+  const scalar_t* __restrict__ input,
+  int64_t num_elems) {
+  __shared__ float cache[1024];
+  int i = blockDim.x * blockIdx.x + threadIdx.x;
+
+  // First store maximum for all values processes by
+  // the current thread in cache[threadIdx.x]
+  scalar_t tmp = 0.0;
+  while (i < num_elems) {
+    float x = static_cast<float>(input[i]);
+    tmp = max(tmp, fabs(x));
+    i += blockDim.x * gridDim.x;
+  }
+  cache[threadIdx.x] = tmp;
+
+  __syncthreads();
+
+  // Now perform parallel reduction within the thread block
+  int ib = blockDim.x / 2;
+  while (ib != 0) {
+    if (threadIdx.x < ib && cache[threadIdx.x + ib] > cache[threadIdx.x]) {
+        cache[threadIdx.x] = cache[threadIdx.x + ib];
+    }
+    __syncthreads();
+    ib /= 2;
+  }
+  // Finally, since cache[0] contains the maximum for this thread block,
+  // atomically write the max to the target location
+  if (threadIdx.x == 0) {
+    atomicMaxFloat(scale, cache[0] / std::numeric_limits<c10::Float8_e4m3fn>::max());
+  }
+}
+
+template<typename scalar_t>
+__global__ void scaled_fp8_quant_kernel(
+  c10::Float8_e4m3fn* __restrict__ out,
+  const scalar_t* __restrict__ input,
+  const float* __restrict__ scale,
+  int64_t num_elems) {
+  int i = blockDim.x * blockIdx.x + threadIdx.x;
+  while (i < num_elems) {
+    out[i] = static_cast<c10::Float8_e4m3fn>(input[i] / *scale);
+    i += blockDim.x * gridDim.x;
+  }
+}
+
+} // namespace vllm
+
+void scaled_fp8_quant(
+  torch::Tensor& out,      // [..., d]
+  torch::Tensor& input,    // [..., d]
+  torch::Tensor& scale)    // [1]
+{
+  int64_t num_tokens = input.numel() / input.size(-1);
+  int64_t num_elems = input.numel();
+  dim3 grid(num_tokens);
+  dim3 block(1024);
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
+  const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+  VLLM_DISPATCH_FLOATING_TYPES(
+    input.scalar_type(),
+    "scaled_fp8_quant_kernel",
+    [&] {
+      vllm::segmented_max_reduction<scalar_t><<<grid, block, 0, stream>>>(
+        scale.data_ptr<float>(),
+        input.data_ptr<scalar_t>(),
+        num_elems);
+      vllm::scaled_fp8_quant_kernel<scalar_t><<<grid, block, 0, stream>>>(
+        out.data_ptr<c10::Float8_e4m3fn>(),
+        input.data_ptr<scalar_t>(),
+        scale.data_ptr<float>(),
+        num_elems);
+      });
+}
+

csrc/quantization/fp8_e5m2_kvcache/quant_utils.cuh

@@ -9,7 +9,6 @@
 #include "../../attention/dtype_float16.cuh"
 #include "../../attention/dtype_bfloat16.cuh"
 
-#pragma once
 
 namespace vllm {
 #ifdef ENABLE_FP8_E5M2

csrc/quantization/gptq/matrix_view.cuh

@@ -146,6 +146,129 @@ public:
     __device__ __forceinline__ const uint32_t* item_uint32_ptr(int row, int column) { return &data[row / 8 * width + column]; }
 };
 
+class MatrixView_q2_row
+{
+public:
+    const uint32_t* data;
+    const int height;
+    const int width;
+
+    __device__ __forceinline__ MatrixView_q2_row(const uint32_t* data, const int height, const int width)
+        : data(data), height(height), width(width)
+    { }
+
+    __device__ __forceinline__ int item(int row, int column) const
+    {
+        int shift = (column & 0x0f) * 2;
+        return (data[row * width / 16 + column / 16] >> shift) & 0x03;
+    }
+
+    __device__ __forceinline__ void item2(int (&items)[2], int row, int column) const
+    {
+        int shift = (column & 0x0f) * 2;
+        uint32_t d = data[row * width / 16 + column / 16] >> shift;
+        items[0] = d & 0x03;
+        items[1] = (d >> 2) & 0x03;
+    }
+
+    __device__ __forceinline__ void item4(int (&items)[4], int row, int column) const
+    {
+        int shift = (column & 0x0f) * 2;
+        uint32_t d = data[row * width / 16 + column / 16] >> shift;
+        items[0] = d & 0x03;
+        items[1] = (d >> 2) & 0x03;
+        items[2] = (d >> 4) & 0x03;
+        items[3] = (d >> 6) & 0x03;
+    }
+};
+
+class MatrixView_q3_row
+{
+public:
+    const uint32_t* data;
+    const int height;
+    const int width;
+
+    __device__ __forceinline__ MatrixView_q3_row(const uint32_t* data, const int height, const int width)
+        : data(data), height(height), width(width)
+    { }
+
+    __device__ __forceinline__ int item(int row, int column) const
+    {
+        int z_w = column * 3 / 32;
+        int z_mod =  column & 0x1f;
+
+        if (z_mod == 10) {
+            return (data[row * width * 3 / 32 + z_w] >> 30) | ((data[row * width * 3 / 32 + (z_w + 1)] << 2) & 0x4);
+        } else if (z_mod == 21) {
+            return (data[row * width * 3 / 32 + z_w] >> 31) | ((data[row * width * 3 / 32 + (z_w + 1)] << 1) & 0x6);
+        } else if (z_mod < 10) {
+            return (data[row * width * 3 / 32 + z_w] >> (z_mod * 3)) & 0x07;
+        } else if (z_mod < 21) {
+            return (data[row * width * 3 / 32 + z_w] >> (z_mod * 3  - 32)) & 0x07;
+        } else {
+            return (data[row * width * 3 / 32 + z_w] >> (z_mod * 3  - 64)) & 0x07;
+        }
+    }
+
+    __device__ __forceinline__ void item4(int (&items)[4], int row, int column) const
+    {
+        int shift = (column & 0x1f);
+        uint32_t d;
+        if (shift <= 4) {
+            d = data[row * width / 32 * 3 + column * 3 / 32] >> (shift * 3);
+        } else if (shift == 8) {
+            d = (data[row * width / 32 * 3 + column * 3 / 32] >> 24) | ((data[row * width / 32 * 3 + column * 3 / 32 + 1] & 0x0f) << 8);
+        } else if (shift <= 16) {
+            d = data[row * width / 32 * 3 + column * 3 / 32] >> (shift * 3 - 32);
+        } else if (shift == 20) {
+            d = (data[row * width / 32 * 3 + column * 3 / 32] >> 28) | ((data[row * width / 32 * 3 + column * 3 / 32 + 1] & 0xff) << 4);
+        } else {
+            d = data[row * width / 32 * 3 + column * 3 / 32] >> (shift * 3 - 64);
+        }
+        items[0] = d & 0x07;
+        items[1] = (d >> 3) & 0x07;
+        items[2] = (d >> 6) & 0x07;
+        items[3] = (d >> 9) & 0x07;
+    }
+};
+
+class MatrixView_q8_row
+{
+public:
+    const uint32_t* data;
+    const int height;
+    const int width;
+
+    __device__ __forceinline__ MatrixView_q8_row(const uint32_t* data, const int height, const int width)
+        : data(data), height(height), width(width)
+    { }
+
+    __device__ __forceinline__ int item(int row, int column) const
+    {
+        int shift = (column & 0x03) * 8;
+        return (data[row * width / 4 + column / 4] >> shift) & 0xff;
+    }
+
+    __device__ __forceinline__ void item2(int (&items)[2], int row, int column) const
+    {
+        int shift = (column & 0x03) * 8;
+        uint32_t d = data[row * width / 4 + column / 4] >> shift;
+        items[0] = d & 0xff;
+        items[1] = (d >> 8) & 0xff;
+    }
+
+    __device__ __forceinline__ void item4(int (&items)[4], int row, int column) const
+    {
+        int shift = (column & 0x03) * 2;
+        uint32_t d = data[row * width / 4 + column / 4] >> shift;
+        items[0] = d & 0xff;
+        items[1] = (d >> 8) & 0xff;
+        items[2] = (d >> 16) & 0xff;
+        items[3] = (d >> 24) & 0xff;
+    }
+};
+
 }  // namespace gptq
 }  // namespace vllm
 #endif

csrc/quantization/gptq/qdq_2.cuh

@@ -0,0 +1,87 @@
+/*
+Copied from https://github.com/turboderp/exllamav2
+*/
+
+#ifndef _qdq_2_cuh
+#define _qdq_2_cuh
+
+#include "qdq_util.cuh"
+
+namespace vllm {
+namespace gptq {
+
+// Permutation:
+//
+// ffddbb99 77553311  eeccaa88 66442200
+
+__forceinline__ __device__ void shuffle_2bit_16
+(
+    uint32_t* q,
+    int stride
+)
+{
+    uint32_t qa = q[0];
+    uint32_t qb = 0;
+
+    #pragma unroll
+    for (int i = 0; i < 8; i++)
+    {
+        uint32_t qa0 = qa & 0x03;
+        uint32_t qa1 = (qa & 0x0c) >> 2;
+        qa >>= 4;
+        qb |= (qa1 << (i * 2 + 16));
+        qb |= (qa0 << (i * 2));
+    }
+    q[0] = qb;
+}
+
+__forceinline__ __device__ void dequant_2bit_16
+(
+    const uint32_t q_0,
+    half2 (&dq)[8],
+    int stride,
+    const uint32_t zero
+)
+{
+    const uint32_t c0 = 0x64006400;
+    const half y4_  = __float2half_rn(1.0f /  4.0f);
+    const half y16_ = __float2half_rn(1.0f / 16.0f);
+    const half y64_ = __float2half_rn(1.0f / 64.0f);
+    const half2 y4  = __halves2half2(y4_,  y4_);
+    const half2 y16 = __halves2half2(y16_, y16_);
+    const half2 y64 = __halves2half2(y64_, y64_);
+
+    const half_uint16 z1_(0xe400 | zero); // half(-1024.0f - zero);
+    const half z4_ = __hsub(__int2half_rn(-256), __int2half_rn(zero));
+    const half z16_ = __hsub(__int2half_rn(-64), __int2half_rn(zero));
+    const half z64_ = __hsub(__int2half_rn(-16), __int2half_rn(zero));
+    const half2 z1 = __half2half2(z1_.as_half);
+    const half2 z4 = __half2half2(z4_);
+    const half2 z16 = __half2half2(z16_);
+    const half2 z64 = __half2half2(z64_);
+
+    uint32_t qa = q_0;
+    half2_uint32 q0((qa & 0x00030003) | c0); // half2(q[ 0], q[ 1])      + 1024
+    half2_uint32 q1((qa & 0x000c000c) | c0); // half2(q[ 2], q[ 3]) *  4 + 1024
+    half2_uint32 q2((qa & 0x00300030) | c0); // half2(q[ 4], q[ 5]) * 16 + 1024
+    half2_uint32 q3((qa & 0x00c000c0) | c0); // half2(q[ 6], q[ 7]) * 64 + 1024
+    qa >>= 8;
+    half2_uint32 q4((qa & 0x00030003) | c0); // half2(q[ 8], q[ 8])      + 1024
+    half2_uint32 q5((qa & 0x000c000c) | c0); // half2(q[10], q[11]) *  4 + 1024
+    half2_uint32 q6((qa & 0x00300030) | c0); // half2(q[12], q[13]) * 16 + 1024
+    half2_uint32 q7((qa & 0x00c000c0) | c0); // half2(q[14], q[15]) * 64 + 1024
+
+    dq[0] = __hadd2(q0.as_half2, z1);
+    dq[1] = __hfma2(q1.as_half2, y4,  z4);
+    dq[2] = __hfma2(q2.as_half2, y16, z16);
+    dq[3] = __hfma2(q3.as_half2, y64, z64);
+    dq[4] = __hadd2(q4.as_half2, z1);
+    dq[5] = __hfma2(q5.as_half2, y4,  z4);
+    dq[6] = __hfma2(q6.as_half2, y16, z16);
+    dq[7] = __hfma2(q7.as_half2, y64, z64);
+}
+
+}  // namespace gptq
+}  // namespace vllm
+
+#endif

csrc/quantization/gptq/qdq_3.cuh

@@ -0,0 +1,141 @@
+#ifndef _qdq_3_cuh
+#define _qdq_3_cuh
+
+#include "qdq_util.cuh"
+
+namespace vllm {
+namespace gptq {
+// Permutation:
+//
+// v9997775 55333111  u8886664 44222000  (u, v lsb)
+// vjjjhhhf ffdddbbb  uiiiggge eecccaaa
+// vtttrrrp ppnnnlll  usssqqqo oommmkkk
+
+__forceinline__ __device__ void shuffle_3bit_32
+(
+    uint32_t* q,
+    int stride
+)
+{
+    uint32_t qa = q[0 * stride];
+    uint32_t qb = q[1 * stride];
+    uint32_t qc = q[2 * stride];
+
+    // qa: aa999888 77766655  54443332 22111000
+    // qb: lkkkjjji iihhhggg  fffeeedd dcccbbba
+    // qc: vvvuuutt tsssrrrq  qqpppooo nnnmmmll
+
+    uint32_t qd = qc >> 26;
+    qc <<= 4;
+    qc |= qb >> 28;
+    qb <<= 2;
+    qb |= qa >> 30;
+
+    // qa: ..999888 77766655  54443332 22111000
+    // qb: ..jjjiii hhhgggff  feeedddc ccbbbaaa
+    // qc: ..tttsss rrrqqqpp  pooonnnm mmlllkkk
+    // qd:                               vvvuuu
+
+    uint32_t za = 0;
+    uint32_t zb = 0;
+    uint32_t zc = 0;
+
+    for (int i = 0; i < 5; i++) { uint32_t t0 = qa & 0x07; uint32_t t1 = (qa & 0x38) >> 3; qa >>= 6; za |= (t0 << (i * 3)); za |= (t1 << (i * 3 + 16)); }
+    for (int i = 0; i < 5; i++) { uint32_t t0 = qb & 0x07; uint32_t t1 = (qb & 0x38) >> 3; qb >>= 6; zb |= (t0 << (i * 3)); zb |= (t1 << (i * 3 + 16)); }
+    for (int i = 0; i < 5; i++) { uint32_t t0 = qc & 0x07; uint32_t t1 = (qc & 0x38) >> 3; qc >>= 6; zc |= (t0 << (i * 3)); zc |= (t1 << (i * 3 + 16)); }
+
+    // za:  9997775 55333111   8886664 44222000
+    // zb:  jjjhhhf ffdddbbb   iiiggge eecccaaa
+    // zc:  tttrrrp ppnnnlll   sssqqqo oommmkkk
+    // qd:                               vvvuuu
+
+    za |= ((qd & 0x01) >> 0) << 15;
+    zb |= ((qd & 0x02) >> 1) << 15;
+    zc |= ((qd & 0x04) >> 2) << 15;
+    za |= ((qd & 0x08) >> 3) << 31;
+    zb |= ((qd & 0x10) >> 4) << 31;
+    zc |= ((qd & 0x20) >> 5) << 31;
+
+    // za: v9997775 55333111  u8886664 44222000  (u, v lsb)
+    // zb: vjjjhhhf ffdddbbb  uiiiggge eecccaaa
+    // zc: vtttrrrp ppnnnlll  usssqqqo oommmkkk
+
+    q[0 * stride] = za;
+    q[1 * stride] = zb;
+    q[2 * stride] = zc;
+}
+
+__forceinline__ __device__ void dequant_3bit_32
+(
+    const uint32_t q_0,
+    const uint32_t q_1,
+    const uint32_t q_2,
+    half2 (&dq)[16],
+    int stride,
+    const uint32_t zero
+)
+{
+    const uint32_t c0 = 0x64006400;
+    const half y8_  = __float2half_rn(1.0f /  8.0f);
+    const half y64_ = __float2half_rn(1.0f / 64.0f);
+    const half2 y8  = __halves2half2(y8_,  y8_);
+    const half2 y64 = __halves2half2(y64_, y64_);
+    const half_uint16 z1_(0xe400 | zero); // half(-1024.0f - zero);
+    const half z8_ = __hsub(__int2half_rn(-128), __int2half_rn(zero));
+    const half z64_ = __hsub(__int2half_rn(-16), __int2half_rn(zero));
+    const half2 z1  = __halves2half2(z1_.as_half,  z1_.as_half);
+    const half2 z8  = __halves2half2(z8_,  z8_);
+    const half2 z64 = __halves2half2(z64_, z64_);
+
+    uint32_t qa = q_0;
+    uint32_t qb = q_1;
+    uint32_t qc = q_2;
+
+    half2_uint32 q0((qa & 0x00070007) | c0); // half2(q[ 0], q[ 1])      + 1024
+    half2_uint32 q1((qa & 0x00380038) | c0); // half2(q[ 2], q[ 3]) *  8 + 1024
+    qa >>= 6;
+    half2_uint32 q2((qa & 0x00070007) | c0); // half2(q[ 4], q[ 5])      + 1024
+    half2_uint32 q3((qa & 0x00380038) | c0); // half2(q[ 6], q[ 7]) *  8 + 1024
+    half2_uint32 q4((qa & 0x01c001c0) | c0); // half2(q[ 8], q[ 9]) * 64 + 1024
+    qa >>= 9;
+    qa &= 0x00010001;
+    half2_uint32 q5((qb & 0x00070007) | c0); // half2(q[10], q[11])      + 1024
+    half2_uint32 q6((qb & 0x00380038) | c0); // half2(q[12], q[13]) *  8 + 1024
+    qb >>= 6;
+    half2_uint32 q7((qb & 0x00070007) | c0); // half2(q[14], q[15])      + 1024
+    half2_uint32 q8((qb & 0x00380038) | c0); // half2(q[16], q[17]) *  8 + 1024
+    half2_uint32 q9((qb & 0x01c001c0) | c0); // half2(q[18], q[19]) * 64 + 1024
+    qb >>= 8;
+    qb &= 0x00020002;
+    half2_uint32 q10((qc & 0x00070007) | c0); // half2(q[20], q[21])      + 1024
+    half2_uint32 q11((qc & 0x00380038) | c0); // half2(q[22], q[23]) *  8 + 1024
+    qc >>= 6;
+    half2_uint32 q12((qc & 0x00070007) | c0); // half2(q[24], q[25])      + 1024
+    half2_uint32 q13((qc & 0x00380038) | c0); // half2(q[26], q[27]) *  8 + 1024
+    half2_uint32 q14((qc & 0x01c001c0) | c0); // half2(q[28], q[29]) * 64 + 1024
+    qc >>= 7;
+    qc &= 0x00040004;
+    half2_uint32 q15((qa | qb | qc) | c0);
+
+    dq[ 0] = __hadd2( q0.as_half2, z1);
+    dq[ 1] = __hfma2( q1.as_half2, y8,  z8);
+    dq[ 2] = __hadd2( q2.as_half2, z1);
+    dq[ 3] = __hfma2( q3.as_half2, y8,  z8);
+    dq[ 4] = __hfma2( q4.as_half2, y64, z64);
+    dq[ 5] = __hadd2( q5.as_half2, z1);
+    dq[ 6] = __hfma2( q6.as_half2, y8,  z8);
+    dq[ 7] = __hadd2( q7.as_half2, z1);
+    dq[ 8] = __hfma2( q8.as_half2, y8,  z8);
+    dq[ 9] = __hfma2( q9.as_half2, y64, z64);
+    dq[10] = __hadd2(q10.as_half2, z1);
+    dq[11] = __hfma2(q11.as_half2, y8,  z8);
+    dq[12] = __hadd2(q12.as_half2, z1);
+    dq[13] = __hfma2(q13.as_half2, y8,  z8);
+    dq[14] = __hfma2(q14.as_half2, y64, z64);
+    dq[15] = __hadd2(q15.as_half2, z1);
+}
+
+}  // namespace gptq
+}  // namespace vllm
+
+#endif

csrc/quantization/gptq/qdq_4.cuh

@@ -38,16 +38,17 @@ __forceinline__ __device__ void dequant_4bit_8
 (
     const uint32_t q_0,
     half2 (&dq)[4],
-    int stride
+    int stride,
+    const uint32_t zero
 )
 {
     const uint32_t c0 = 0x64006400;
     const half y16_ = __float2half_rn(1.0f / 16.0f);
     const half2 y16 = __halves2half2(y16_, y16_);
-    const half z1_  = __float2half_rn(-1024.0f         - 8.0f);
-    const half z16_ = __float2half_rn(-1024.0f / 16.0f - 8.0f);
-    const half2 z1  = __halves2half2(z1_,  z1_);
-    const half2 z16 = __halves2half2(z16_, z16_);
+    const half_uint16 z1_(0xe400 | zero); // half(-1024.0f - zero);
+    const half z16_ = __hsub(__int2half_rn(-64), __int2half_rn(zero));
+    const half2 z1 = __half2half2(z1_.as_half);
+    const half2 z16 = __half2half2(z16_);
 
     uint32_t qa = q_0;
     half2_uint32 q0((qa & 0x000f000f) | c0); // half2(q[ 0], q[ 1])      + 1024
@@ -143,93 +144,4 @@ __forceinline__ __device__ void dequant_4bit_8_gptq
 }  // namespace gptq
 }  // namespace vllm
 
-#else
-
-namespace vllm {
-namespace gptq {
-__forceinline__ __device__ void shuffle_4bit_8
-(
-    uint32_t* q,
-    int stride
-)
-{
-}
-
-__forceinline__ __device__ void dequant_4bit_8
-(
-    const uint32_t q_0,
-    half2 (&dq)[4],
-    int stride
-)
-{
-    half dqh[8];
-    for (int i = 0; i < 8; i++) dqh[i] = dq_ns(exb(q_0, i * 4, 0x0f), 8);
-
-    for (int i = 0; i < 4; i++) dq[i] = __halves2half2(dqh[i * 2], dqh[i * 2 + 1]);
-}
-
-__forceinline__ __device__ void dequant_4bit_8_prep_zero_scale
-(
-    const uint32_t zero,
-    const half scale,
-    half2 (&z1)[2],
-    half2 (&y1)[2]
-)
-{
-    half z = __int2half_rn(-((int)zero));
-    z = __hmul(z, scale);
-    z1[0] = __half2half2(z);
-    y1[0] = __half2half2(scale);
-}
-
-__forceinline__ __device__ void dequant_4bit_8_prep_zero
-(
-    const uint32_t zero,
-    half2(&z1)[2],
-    half2(&y1)[2]
-)
-{
-    half z = __int2half_rn(-((int)zero));
-    z1[0] = __half2half2(z);
-}
-
-__forceinline__ __device__ void dequant_4bit_8_gptq
-(
-    const uint32_t q_0,
-    half2 (&dq)[4],
-    half2 (&z1)[2],
-    half2 (&y1)[2],
-    int stride,
-    bool scaled
-)
-{
-    half2 dqh2[8];
-
-    uint32_t qa = q_0;
-    for (int i = 0; i < 4; i++)
-    {
-        half d0 = __int2half_rn(qa & 0x0f); qa >>= 4;
-        half d1 = __int2half_rn(qa & 0x0f); qa >>= 4;
-        dqh2[i] = __halves2half2(d0, d1);
-    }
-
-    if (scaled)
-    {
-        dq[0] = __hfma2(dqh2[0], y1[0], z1[0]);
-        dq[1] = __hfma2(dqh2[1], y1[0], z1[0]);
-        dq[2] = __hfma2(dqh2[2], y1[0], z1[0]);
-        dq[3] = __hfma2(dqh2[3], y1[0], z1[0]);
-    }
-    else
-    {
-        dq[0] = __hadd2(dqh2[0], z1[0]);
-        dq[1] = __hadd2(dqh2[1], z1[0]);
-        dq[2] = __hadd2(dqh2[2], z1[0]);
-        dq[3] = __hadd2(dqh2[3], z1[0]);
-    }
-}
-
-}  // namespace gptq
-}  // namespace vllm
-
 #endif

csrc/quantization/gptq/qdq_8.cuh

@@ -0,0 +1,40 @@
+/*
+Copied from https://github.com/turboderp/exllamav2
+*/
+
+#ifndef _qdq_8_cuh
+#define _qdq_8_cuh
+
+#include "qdq_util.cuh"
+
+namespace vllm {
+namespace gptq {
+
+__forceinline__ __device__ void shuffle_8bit_4
+(
+    uint32_t* q,
+    int stride
+)
+{
+}
+
+__forceinline__ __device__ void dequant_8bit_8
+(
+    const uint32_t q_0,
+    const uint32_t q_1,
+    half2 (&dq)[4],
+    int stride,
+    const uint32_t zero
+)
+{
+    half dqh[8];
+    for (int i = 0; i < 4; i++) dqh[i    ] = dq_ns(exb(q_0, i * 8, 0xff), zero);
+    for (int i = 0; i < 4; i++) dqh[i + 4] = dq_ns(exb(q_1, i * 8, 0xff), zero);
+
+    for (int i = 0; i < 4; i++) dq[i] = __halves2half2(dqh[i * 2], dqh[i * 2 + 1]);
+}
+
+}  // namespace gptq
+}  // namespace vllm
+
+#endif