Merge branch 'main' into woosuk-tpu

.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/test-pipeline.yaml

@@ -34,7 +34,10 @@ steps:
   command: pytest -v -s engine tokenization test_sequence.py test_config.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: Examples Test
   working_dir: "/vllm-workspace/examples"
@@ -90,7 +93,7 @@ steps:
   - bash run-benchmarks.sh
 
 - label: Documentation Build
-  working_dir: "/vllm-workspace/docs"
+  working_dir: "/vllm-workspace/test_docs/docs"
   no_gpu: True
   commands:
   - pip install -r requirements-docs.txt

.buildkite/test-template.j2

@@ -8,6 +8,9 @@ steps:
       queue: amd
     command: bash .buildkite/run-amd-test.sh
 
+  - label: "CPU Test"
+    command: bash .buildkite/run-cpu-test.sh
+
   - label: ":docker: build image"
     commands:
       - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."

.github/workflows/publish.yml

@@ -49,7 +49,7 @@ 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:

.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

.gitignore

@@ -181,6 +181,7 @@ _build/
 # hip files generated by PyTorch
 *.hip
 *_hip*
+hip_compat.h
 
 # Benchmark dataset
 *.json

CMakeLists.txt

@@ -2,7 +2,10 @@ 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)
 
@@ -16,7 +19,7 @@ set(PYTHON_SUPPORTED_VERSIONS "3.8" "3.9" "3.10" "3.11")
 set(CUDA_SUPPORTED_ARCHS "7.0;7.5;8.0;8.6;8.9;9.0")
 
 # Supported AMD GPU architectures.
-set(HIP_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;gfx1100")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100")
 
 #
 # Supported/expected torch versions for CUDA/ROCm.
@@ -28,7 +31,7 @@ set(HIP_SUPPORTED_ARCHS "gfx908;gfx90a;gfx942;gfx1100")
 # requirements.txt files and should be kept consistent.  The ROCm torch
 # versions are derived from Dockerfile.rocm
 #
-set(TORCH_SUPPORTED_VERSION_CUDA "2.1.2")
+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")
 
@@ -76,6 +79,19 @@ find_package(Torch REQUIRED)
 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.

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

Dockerfile

@@ -2,6 +2,7 @@
 # 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 \
@@ -16,18 +17,26 @@ 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
@@ -38,18 +47,16 @@ RUN --mount=type=cache,target=/root/.cache/pip \
 # install compiler cache to speed up compilation leveraging local or remote caching
 RUN apt-get update -y && apt-get install -y ccache
 
-# copy input files
+# files and directories related to build wheels
 COPY csrc csrc
 COPY setup.py setup.py
 COPY cmake cmake
 COPY CMakeLists.txt CMakeLists.txt
-COPY requirements.txt requirements.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}
@@ -61,7 +68,15 @@ ENV VLLM_INSTALL_PUNICA_KERNELS=1
 
 ENV CCACHE_DIR=/root/.cache/ccache
 RUN --mount=type=cache,target=/root/.cache/ccache \
-    python3 setup.py build_ext --inplace
+    --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 ####################
@@ -81,57 +96,59 @@ RUN pip --verbose wheel flash-attn==${FLASH_ATTN_VERSION} \
 
 #################### 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/
-# Install flash attention (from pre-built wheel)
-RUN --mount=type=bind,from=flash-attn-builder,src=/usr/src/flash-attention-v2,target=/usr/src/flash-attention-v2 \
-    pip install /usr/src/flash-attention-v2/*.whl --no-cache-dir
-# ignore build dependencies installation because we are using pre-complied extensions
-RUN rm pyproject.toml
-RUN --mount=type=cache,target=/root/.cache/pip VLLM_USE_PRECOMPILED=1 pip install . --verbose
-#################### TEST IMAGE ####################
 
-
-#################### RUNTIME BASE IMAGE ####################
-# We used base cuda image because pytorch installs its own cuda libraries.
-# However pynccl depends on cuda libraries so we had to switch to the runtime image
-# In the future it would be nice to get a container with pytorch and cuda without duplicating cuda
-FROM nvidia/cuda:12.1.0-runtime-ubuntu22.04 AS vllm-base
-
-# libnccl required for ray
-RUN apt-get update -y \
-    && apt-get install -y python3-pip
-
-WORKDIR /workspace
-COPY requirements.txt requirements.txt
+# install development dependencies (for testing)
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install -r requirements.txt
+    pip install -r requirements-dev.txt
 
-# Install flash attention (from pre-built wheel)
-RUN --mount=type=bind,from=flash-attn-builder,src=/usr/src/flash-attention-v2,target=/usr/src/flash-attention-v2 \
-    pip install /usr/src/flash-attention-v2/*.whl --no-cache-dir
-
-#################### RUNTIME BASE IMAGE ####################
+# 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 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"]

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.rocm

@@ -23,6 +23,9 @@ RUN echo "FA_BRANCH is $FA_BRANCH"
 # In that case, we need to use the python reference attention implementation in vllm
 ARG BUILD_FA="1"
 
+# whether to build triton on rocm
+ARG BUILD_TRITON="1"
+
 # Install some basic utilities
 RUN apt-get update && apt-get install python3 python3-pip -y
 
@@ -75,6 +78,17 @@ RUN if [ "$BUILD_FA" = "1" ]; then \
 RUN if [ "$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

MANIFEST.in

@@ -1,5 +1,6 @@
 include LICENSE
-include requirements.txt
+include requirements-common.txt
+include requirements-cuda.txt
 include CMakeLists.txt
 
 recursive-include cmake *

README.md

@@ -14,18 +14,8 @@ Easy, fast, and cheap LLM serving for everyone
 
 </p>
 
----
-
-**The Third vLLM Bay Area Meetup (April 2nd 6pm-8:30pm PT)**
-
-We are thrilled to announce our third vLLM Meetup!
-The vLLM team will share recent updates and roadmap.
-We will also have vLLM collaborators from Roblox coming up to the stage to discuss their experience in deploying LLMs with vLLM.
-Please register [here](https://robloxandvllmmeetup2024.splashthat.com/) 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.
@@ -80,6 +70,7 @@ vLLM seamlessly supports many Hugging Face models, including the following archi
 - 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 (`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.)
 - MPT (`mosaicml/mpt-7b`, `mosaicml/mpt-30b`, etc.)
@@ -88,7 +79,7 @@ vLLM seamlessly supports many Hugging Face models, including the following archi
 - 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.)

benchmarks/backend_request_func.py

@@ -334,7 +334,8 @@ async def async_request_openai_chat_completions(
                             timestamp = time.perf_counter()
                             data = json.loads(chunk)
 
-                            if "content" in data["choices"][0]["delta"]:
+                            delta = data["choices"][0]["delta"]
+                            if delta.get("content", None):
                                 # First token
                                 if ttft == 0:
                                     ttft = time.perf_counter() - st
@@ -345,8 +346,7 @@ async def async_request_openai_chat_completions(
                                     output.itl.append(timestamp -
                                                       most_recent_timestamp)
 
-                                generated_text += data["choices"][0]["delta"][
-                                    "content"]
+                                generated_text += delta["content"]
 
                             most_recent_timestamp = timestamp
 

benchmarks/benchmark_latency.py

@@ -24,6 +24,7 @@ def main(args: argparse.Namespace):
               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,
@@ -67,7 +68,8 @@ 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
@@ -83,7 +85,12 @@ def main(args: argparse.Namespace):
     latencies = []
     for _ in tqdm(range(args.num_iters), desc="Profiling iterations"):
         latencies.append(run_to_completion(profile_dir=None))
+    latencies = np.array(latencies)
+    percentages = [10, 25, 50, 75, 90]
+    percentiles = np.percentile(latencies, percentages)
     print(f'Avg latency: {np.mean(latencies)} seconds')
+    for percentage, percentile in zip(percentages, percentiles):
+        print(f'{percentage}% percentile latency: {percentile} seconds')
 
 
 if __name__ == '__main__':
@@ -105,9 +112,13 @@ 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',
@@ -127,10 +138,23 @@ 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',
@@ -145,8 +169,8 @@ if __name__ == '__main__':
         "--device",
         type=str,
         default="cuda",
-        choices=["cuda"],
-        help='device type for vLLM execution, supporting CUDA only currently.')
+        choices=["cuda", "cpu"],
+        help='device type for vLLM execution, supporting CUDA and CPU.')
     parser.add_argument('--block-size',
                         type=int,
                         default=16,

benchmarks/benchmark_serving.py

@@ -110,7 +110,9 @@ def sample_sonnet_requests(
     prefix_len: int,
     tokenizer: PreTrainedTokenizerBase,
 ) -> List[Tuple[str, str, int, int]]:
-    assert input_len > prefix_len, "input_len must be greater than prefix_len."
+    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:
@@ -131,8 +133,9 @@ def sample_sonnet_requests(
         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.input-len' higher than {base_prompt_offset}."
+    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)
 
@@ -140,7 +143,7 @@ def sample_sonnet_requests(
     # prompt are fixed poem lines.
     assert (
         prefix_len > base_prompt_offset
-    ), f"Please set 'args.prefix-len' higher than {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)
@@ -373,9 +376,9 @@ def main(args: argparse.Namespace):
             input_requests = sample_sonnet_requests(
                 dataset_path=args.dataset_path,
                 num_requests=args.num_prompts,
-                input_len=args.input_len,
-                output_len=args.output_len,
-                prefix_len=args.prefix_len,
+                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)
@@ -388,9 +391,9 @@ def main(args: argparse.Namespace):
             input_requests = sample_sonnet_requests(
                 dataset_path=args.dataset_path,
                 num_requests=args.num_prompts,
-                input_len=args.input_len,
-                output_len=args.output_len,
-                prefix_len=args.prefix_len,
+                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)

benchmarks/benchmark_throughput.py

@@ -29,22 +29,23 @@ 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 +54,7 @@ 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 +71,7 @@ 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,
     gpu_memory_utilization: float = 0.9,
@@ -89,6 +89,7 @@ def run_vllm(
               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)
@@ -217,7 +218,8 @@ def main(args: argparse.Namespace):
                                 args.seed, args.n, args.use_beam_search,
                                 args.trust_remote_code, args.dtype,
                                 args.max_model_len, args.enforce_eager,
-                                args.kv_cache_dtype, args.device,
+                                args.kv_cache_dtype,
+                                args.quantization_param_path, args.device,
                                 args.enable_prefix_caching,
                                 args.gpu_memory_utilization, args.download_dir)
     elif args.backend == "hf":
@@ -306,16 +308,29 @@ 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(
         "--device",
         type=str,
         default="cuda",
-        choices=["cuda"],
-        help='device type for vLLM execution, supporting CUDA only currently.')
+        choices=["cuda", "cpu"],
+        help='device type for vLLM execution, supporting CUDA and CPU.')
     parser.add_argument(
         "--enable-prefix-caching",
         action='store_true',

benchmarks/kernels/benchmark_paged_attention.py

@@ -97,6 +97,9 @@ def main(
             torch.cuda.cudart().cudaProfilerStart()
         start_time = time.perf_counter()
 
+        # Using default kv_scale
+        kv_scale = 1.0
+
         for _ in range(num_iters):
             if version == "v1":
                 ops.paged_attention_v1(
@@ -112,6 +115,7 @@ def main(
                     max_context_len,
                     alibi_slopes,
                     kv_cache_dtype,
+                    kv_scale,
                 )
             elif version == "v2":
                 ops.paged_attention_v2(
@@ -130,6 +134,7 @@ def main(
                     max_context_len,
                     alibi_slopes,
                     kv_cache_dtype,
+                    kv_scale,
                 )
             else:
                 raise ValueError(f"Invalid version: {version}")
@@ -179,11 +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.')
-    parser.add_argument("--device", type=str, choices=["cuda"], default="cuda")
+        '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)
 

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/utils.cmake

@@ -100,6 +100,8 @@ function (get_torch_gpu_compiler_flags OUT_GPU_FLAGS GPU_LANG)
 
     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__"
@@ -117,6 +119,7 @@ function (get_torch_gpu_compiler_flags OUT_GPU_FLAGS GPU_LANG)
 
     list(APPEND GPU_FLAGS
       "-DUSE_ROCM"
+      "-DENABLE_FP8_E4M3"
       "-U__HIP_NO_HALF_CONVERSIONS__"
       "-U__HIP_NO_HALF_OPERATORS__"
       "-fno-gpu-rdc")

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

@@ -22,12 +22,26 @@
 
 #include "attention_dtypes.h"
 #include "attention_utils.cuh"
-#ifdef ENABLE_FP8_E5M2
+
+#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))
@@ -78,7 +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]
@@ -95,7 +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;
@@ -142,7 +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
 
@@ -208,11 +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
@@ -292,7 +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;
@@ -328,11 +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
@@ -423,7 +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]
@@ -437,11 +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).
@@ -451,7 +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]
@@ -468,11 +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).
@@ -579,9 +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,                                                                            \
@@ -594,14 +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,
@@ -613,7 +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);
@@ -677,8 +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,                                                               \
@@ -688,20 +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);     \
@@ -720,7 +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);
@@ -731,7 +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) {
@@ -748,7 +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,                                                                           \
@@ -764,7 +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,                                                                                  \
@@ -778,7 +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(
@@ -794,7 +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);
@@ -864,8 +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,                                                                  \
@@ -878,20 +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);           \
@@ -913,7 +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);
@@ -924,7 +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

@@ -21,9 +21,10 @@ void reshape_and_cache(
   torch::Tensor& key_cache,
   torch::Tensor& value_cache,
   torch::Tensor& slot_mapping,
-  const std::string& kv_cache_dtype);
+  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

@@ -4,8 +4,10 @@
 
 #include "cuda_compat.h"
 #include "dispatch_utils.h"
-#ifdef ENABLE_FP8_E5M2
+#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>
@@ -151,7 +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]
@@ -163,7 +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) {
@@ -195,10 +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
@@ -211,8 +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()),                                              \
@@ -223,7 +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]
@@ -231,7 +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);
@@ -254,7 +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) {
@@ -270,15 +278,17 @@ void reshape_and_cache(
 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
@@ -287,16 +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);
 
@@ -305,16 +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/layernorm_kernels.cu

@@ -59,6 +59,8 @@ __global__ void rms_norm_kernel(
 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;
@@ -85,8 +87,8 @@ struct _typeConvert<c10::BFloat16> {
   __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
-
+#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.

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,

csrc/pybind.cpp

@@ -91,9 +91,9 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
     &reshape_and_cache,
     "Reshape the key and value tensors and cache them");
   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/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

docs/requirements-docs.txt

@@ -7,4 +7,6 @@ sphinx-argparse
 # packages to install to build the documentation
 pydantic
 -f https://download.pytorch.org/whl/cpu
-torch
+torch
+py-cpuinfo
+transformers

docs/source/conf.py

@@ -11,13 +11,10 @@
 # documentation root, use os.path.abspath to make it absolute, like shown here.
 
 import logging
-import os
 import sys
 
 from sphinx.ext import autodoc
 
-sys.path.insert(0, os.path.abspath(os.path.join('..', '..')))
-
 logger = logging.getLogger(__name__)
 
 # -- Project information -----------------------------------------------------
@@ -75,6 +72,7 @@ html_theme_options = {
 
 # Mock out external dependencies here.
 autodoc_mock_imports = [
+    "cpuinfo",
     "torch",
     "transformers",
     "psutil",

docs/source/getting_started/cpu-installation.rst

@@ -0,0 +1,87 @@
+.. _installation_cpu:
+
+Installation with CPU
+========================
+
+vLLM initially supports basic model inferencing and serving on x86 CPU platform, with data types FP32 and BF16.
+
+Table of contents:
+
+#. :ref:`Requirements <cpu_backend_requirements>`
+#. :ref:`Quick start using Dockerfile <cpu_backend_quick_start_dockerfile>`
+#. :ref:`Build from source <build_cpu_backend_from_source>`
+#. :ref:`Performance tips <cpu_backend_performance_tips>`
+
+.. _cpu_backend_requirements:
+
+Requirements
+------------
+
+* OS: Linux
+* Compiler: gcc/g++>=12.3.0 (recommended)
+* Instruction set architecture (ISA) requirement: AVX512 is required.
+
+.. _cpu_backend_quick_start_dockerfile:
+
+Quick start using Dockerfile
+----------------------------
+
+.. code-block:: console
+
+    $ docker build -f Dockerfile.cpu -t vllm-cpu-env --shm-size=4g .
+    $ docker run -it \
+                 --rm \
+                 --network=host \
+                 --cpuset-cpus=<cpu-id-list, optional> \
+                 --cpuset-mems=<memory-node, optional> \
+                 vllm-cpu-env
+
+.. _build_cpu_backend_from_source:
+
+Build from source
+-----------------
+
+- First, install required compiler. We recommend to use ``gcc/g++ >= 12.3.0`` as the default compiler to avoid potential problems. For example, on Ubuntu 22.4, you can run:
+
+.. code-block:: console
+
+    $ sudo apt-get update  -y
+    $ sudo apt-get install -y gcc-12 g++-12
+    $ sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
+
+- Second, install Python packages for vLLM CPU backend building:
+
+.. code-block:: console
+
+    $ pip install --upgrade pip
+    $ pip install wheel packaging ninja setuptools>=49.4.0 numpy
+    $ pip install -v -r requirements-cpu.txt --extra-index-url https://download.pytorch.org/whl/cpu
+
+- Finally, build and install vLLM CPU backend: 
+
+.. code-block:: console
+
+    $ VLLM_TARGET_DEVICE=cpu python setup.py install
+
+.. note::
+    - BF16 is the default data type in the current CPU backend (that means the backend will cast FP16 to BF16), and is compatible will all CPUs with AVX512 ISA support. 
+
+    - AVX512_BF16 is an extension ISA provides native BF16 data type conversion and vector product instructions, will brings some performance improvement compared with pure AVX512. The CPU backend build script will check the host CPU flags to determine whether to enable AVX512_BF16. 
+    
+    - If you want to force enable AVX512_BF16 for the cross-compilation, please set environment variable VLLM_CPU_AVX512BF16=1 before the building.    
+
+.. _cpu_backend_performance_tips:
+
+Performance tips
+-----------------
+
+- vLLM CPU backend uses environment variable ``VLLM_CPU_KVCACHE_SPACE`` to specify the KV Cache size (e.g, ``VLLM_CPU_KVCACHE_SPACE=40`` means 40 GB space for KV cache), larger setting will allow vLLM running more requests in parallel. This parameter should be set based on the hardware configuration and memory management pattern of users.
+
+- vLLM CPU backend uses OpenMP for thread-parallel computation. If you want the best performance on CPU, it will be very critical to isolate CPU cores for OpenMP threads with other thread pools (like web-service event-loop), to avoid CPU oversubscription. 
+
+- If using vLLM CPU backend on a bare-metal machine, it is recommended to disable the hyper-threading.
+
+- If using vLLM CPU backend on a multi-socket machine with NUMA, be aware to set CPU cores and memory nodes, to avoid the remote memory node access. ``numactl`` is an useful tool for CPU core and memory binding on NUMA platform. Besides, ``--cpuset-cpus`` and ``--cpuset-mems`` arguments of ``docker run`` are also useful.
+
+
+

docs/source/index.rst

@@ -63,6 +63,7 @@ Documentation
    getting_started/installation
    getting_started/amd-installation
    getting_started/neuron-installation
+   getting_started/cpu-installation
    getting_started/quickstart
 
 .. toctree::
@@ -90,7 +91,8 @@ Documentation
    :caption: Quantization
 
    quantization/auto_awq
-   quantization/fp8_e5m2_kv_cache
+   quantization/fp8_e5m2_kvcache
+   quantization/fp8_e4m3_kvcache
 
 .. toctree::
    :maxdepth: 2

docs/source/models/adding_model.rst

@@ -21,6 +21,8 @@ This document provides a high-level guide on integrating a `HuggingFace Transfor
 Start by forking our `GitHub`_ repository and then :ref:`build it from source <build_from_source>`.
 This gives you the ability to modify the codebase and test your model.
 
+.. tip::
+    If you don't want to fork the repository and modify vLLM's codebase, please refer to the "Out-of-Tree Model Integration" section below.
 
 1. Bring your model code
 ------------------------
@@ -94,3 +96,28 @@ This method should load the weights from the HuggingFace's checkpoint file and a
 ----------------------
 
 Finally, include your :code:`*ForCausalLM` class in `vllm/model_executor/models/__init__.py <https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/__init__.py>`_ and register it to the :code:`_MODEL_REGISTRY` in `vllm/model_executor/model_loader.py <https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/model_loader.py>`_.
+
+6. Out-of-Tree Model Integration
+--------------------------------------------
+
+We also provide a way to integrate a model without modifying the vLLM codebase. Step 2, 3, 4 are still required, but you can skip step 1 and 5.
+
+Just add the following lines in your code:
+
+.. code-block:: python
+
+    from vllm import ModelRegistry
+    from your_code import YourModelForCausalLM
+    ModelRegistry.register_model("YourModelForCausalLM", YourModelForCausalLM)
+
+If you are running api server with `python -m vllm.entrypoints.openai.api_server args`, you can wrap the entrypoint with the following code:
+
+.. code-block:: python
+
+    from vllm import ModelRegistry
+    from your_code import YourModelForCausalLM
+    ModelRegistry.register_model("YourModelForCausalLM", YourModelForCausalLM)
+    import runpy
+    runpy.run_module('vllm.entrypoints.openai.api_server', run_name='__main__')
+
+Save the above code in a file and run it with `python your_file.py args`.

docs/source/models/engine_args.rst

@@ -118,3 +118,19 @@ Below, you can find an explanation of every engine argument for vLLM:
 .. option:: --quantization (-q) {awq,squeezellm,None}
 
     Method used to quantize the weights.
+
+Async Engine Arguments
+----------------------
+Below are the additional arguments related to the asynchronous engine:
+
+.. option:: --engine-use-ray
+
+    Use Ray to start the LLM engine in a separate process as the server process.
+
+.. option:: --disable-log-requests
+
+    Disable logging requests.
+    
+.. option:: --max-log-len
+
+    Max number of prompt characters or prompt ID numbers being printed in log. Defaults to unlimited.

docs/source/models/supported_models.rst

@@ -83,6 +83,10 @@ Alongside each architecture, we include some popular models that use it.
     - LLaMA, LLaMA-2, Vicuna, Alpaca, Yi
     - :code:`meta-llama/Llama-2-13b-hf`, :code:`meta-llama/Llama-2-70b-hf`, :code:`openlm-research/open_llama_13b`, :code:`lmsys/vicuna-13b-v1.3`, :code:`01-ai/Yi-6B`, :code:`01-ai/Yi-34B`, etc.
     - ✅︎
+  * - :code:`MiniCPMForCausalLM`
+    - MiniCPM
+    - :code:`openbmb/MiniCPM-2B-sft-bf16`, :code:`openbmb/MiniCPM-2B-dpo-bf16`, etc.
+    -
   * - :code:`MistralForCausalLM`
     - Mistral, Mistral-Instruct
     - :code:`mistralai/Mistral-7B-v0.1`, :code:`mistralai/Mistral-7B-Instruct-v0.1`, etc.

docs/source/quantization/fp8_e4m3_kvcache.rst

@@ -0,0 +1,49 @@
+.. _fp8_e4m3_kvcache:
+
+FP8 E4M3 KV Cache
+==================
+
+Quantizing the KV cache to FP8 reduces its memory footprint. This increases the number of tokens that can be stored in the cache, 
+improving throughput. OCP (Open Compute Project www.opencompute.org) specifies two common 8-bit floating point data formats: E5M2 
+(5 exponent bits and 2 mantissa bits) and E4M3FN (4 exponent bits and 3 mantissa bits), often shortened as E4M3. One benefit of 
+the E4M3 format over E5M2 is that floating point numbers are represented in higher precision. However, the small dynamic range of 
+FP8 E4M3 (±240.0 can be represented) typically necessitates the use of a higher-precision (typically FP32) scaling factor alongside 
+each quantized tensor. For now, only per-tensor (scalar) scaling factors are supported. Development is ongoing to support scaling 
+factors of a finer granularity (e.g. per-channel).
+
+These scaling factors can be specified by passing an optional quantization param JSON to the LLM engine at load time. If 
+this JSON is not specified, scaling factors default to 1.0. These scaling factors are typically obtained when running an 
+unquantized model through a quantizer tool (e.g. AMD quantizer or NVIDIA AMMO). 
+
+To install AMMO (AlgorithMic Model Optimization):
+
+.. code-block:: console
+
+        $ pip install --no-cache-dir --extra-index-url https://pypi.nvidia.com nvidia-ammo
+
+Studies have shown that FP8 E4M3 quantization typically only minimally degrades inference accuracy. The most recent silicon 
+offerings e.g. AMD MI300, NVIDIA Hopper or later support native hardware conversion to and from fp32, fp16, bf16, etc. 
+Thus, LLM inference is greatly accelerated with minimal accuracy loss.
+
+
+Here is an example of how to enable this feature:
+
+.. code-block:: python
+
+        # two float8_e4m3fn kv cache scaling factor files are provided under tests/fp8_kv, please refer to 
+        # https://github.com/vllm-project/vllm/blob/main/examples/fp8/README.md to generate kv_cache_scales.json of your own.
+
+        from vllm import LLM, SamplingParams
+        sampling_params = SamplingParams(temperature=1.3, top_p=0.8)
+        llm = LLM(model="meta-llama/Llama-2-7b-chat-hf",
+                  kv_cache_dtype="fp8",
+                  quantization_param_path="./tests/fp8_kv/llama2-7b-fp8-kv/kv_cache_scales.json")
+        prompt = "London is the capital of"
+        out = llm.generate(prompt, sampling_params)[0].outputs[0].text
+        print(out)
+
+        # output w/ scaling factors:  England, the United Kingdom, and one of the world's leading financial,
+        # output w/o scaling factors:  England, located in the southeastern part of the country. It is known 
+
+Note, current prefix caching doesn't work with FP8 KV cache enabled, forward_prefix kernel should handle different KV and cache type.
+

docs/source/quantization/fp8_e5m2_kvcache.rst

@@ -1,4 +1,4 @@
-.. _fp8_e5m2_kv_cache:
+.. _fp8_kv_cache:
 
 FP8 E5M2 KV Cache
 ==================
@@ -21,7 +21,7 @@ Here is an example of how to enable this feature:
     # Create a sampling params object.
     sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
     # Create an LLM.
-    llm = LLM(model="facebook/opt-125m", kv_cache_dtype="fp8_e5m2")
+    llm = LLM(model="facebook/opt-125m", kv_cache_dtype="fp8")
     # Generate texts from the prompts. The output is a list of RequestOutput objects
     # that contain the prompt, generated text, and other information.
     outputs = llm.generate(prompts, sampling_params)
@@ -31,3 +31,6 @@ Here is an example of how to enable this feature:
         generated_text = output.outputs[0].text
         print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
 
+
+Note, current prefix caching doesn't work with FP8 KV cache enabled, forward_prefix kernel should handle different KV and cache type.
+

examples/fp8/README.md

@@ -0,0 +1,96 @@
+# FP8 KV Cache 
+
+This utility extracts the KV cache scaling factors from a quantized HF (Hugging Face) model. The extracted scaling factors are saved to a JSON file, which can later be used by vLLM (variable-length language model) during runtime. This tool is particularly useful when the KV cache data type is FP8 and is intended for use on ROCm (AMD GPU) platforms.
+
+## Prerequisites
+
+- Python 3.x
+- PyTorch
+- NumPy
+- Hugging Face Transformers
+- Hugging Face Hub
+- AMMO 
+
+Before incorporating the FP8 datatype for inference workloads, you must adhere to the following steps:
+1. Install all necessary prerequisites and dependencies. 
+2. Convert HF model into a quantized HF model. 
+3. Extract KV Cache Scaling Factors from quantized HF model.
+4. Load KV Cache Scaling Factors into VLLM.
+
+### 2. Convert HF model into a quantized HF model.
+Note: The following steps are adapted from the [TensorRT-LLM repository](https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/quantization/README.md).
+
+`quantize.py` (examples/fp8/quantizer/quantize.py) uses the quantization toolkit  (AMMO) to calibrate the PyTorch models and export TensorRT-LLM checkpoints. Each TensorRT-LLM checkpoint contains a config file (in .json format) and one or several rank weight files (in .safetensors format).
+
+The detailed quantization toolkit (AMMO) conversion guide for FP8 can be found at `examples/fp8/quantizer/README.md`.
+
+### 3. Extract KV Cache Scaling Factors from quantized HF model.
+`extract_scales.py` (examples/fp8/extract_scales.py) can be utilized to extract the KV cache scaling factors from your quantized HF model, however at the moment, this tool exclusively supports Llama 2 models. It is also important to note the following:
+1. **File Structure**: The utility operates under the assumption that all parameters, including KV cache scaling factors, corresponding to a particular Tensor Parallelism (TP) rank are stored in a single file. These files must adhere to a specific naming convention where the TP rank is immediately identified after a specific keyword (e.g., "rank") in the filename.
+
+2. **TP Decomposition**: The utility assumes consistency between the TP decomposition employed by the quantizer tool and that used by vLLM.
+
+3. **AMMO Compatibility**: Currently, the generated KV cache scaling factors for AMMO remain uniform across all TP ranks.
+
+```python
+# prerequisites:
+# - Quantized HF LLaMa 2 model 
+python3 examples/fp8/extract_scales.py --help
+Usage: extract_scales.py [-h] --quantized_model QUANTIZED_MODEL [--load_format {auto,safetensors,npz,pt}] [--output_dir OUTPUT_DIR] [--output_name OUTPUT_NAME] [--tp_size TP_SIZE]
+
+KV Scale Extraction Example
+
+optional arguments:
+--quantized_model: Specify either the local path to, or name of, a quantized HF model. It is expected that the quantization format is FP8_E4M3, for use on ROCm (AMD GPU).
+Optional arguments:
+--cache_dir: Specify a cache directory to use in the event of a HF model download. (Default: None)
+--load_format: Specify the format of the model's tensor files containing the KV cache scaling factors. (Choices: auto, safetensors, npz, pt; Default: auto)
+--revision: Specify the model's revision number. (Default: None)
+--output_dir: Specify the output directory. By default the KV cache scaling factors will be saved in the model directory. (Default: None)
+--output_name: Specify the output filename. (Default: kv_cache_scales.json)
+--tp_size: Specify the tensor-parallel (TP) size that the quantized model should correspond to. If specified, during KV cache scaling factor extraction the observed TP size will be checked against this and an error will be raised if there is a mismatch. (Default: None)
+```
+```python
+Example:
+python3 examples/fp8/extract_scales.py --quantized_model <QUANTIZED_MODEL_DIR> --tp_size <TENSOR_PARALLEL_SIZE> --output_dir <PATH_TO_OUTPUT_DIR>
+```
+### 4. Load KV Cache Scaling Factors into VLLM.
+This script evaluates the inference throughput of language models using various backends such as vLLM. It measures the time taken to process a given number of prompts and generate sequences for each prompt. The recently generated KV cache scaling factors are now integrated into the benchmarking process and allow for KV cache scaling factors to be utilized for FP8.
+```python
+# prerequisites:
+# -  LLaMa 2 kv_cache_scales.json file
+
+python3 benchmarks/benchmark_throughput.py --help 
+usage: benchmark_throughput.py [-h] [--backend {vllm,hf,mii}] [--dataset DATASET] [--input-len INPUT_LEN] [--output-len OUTPUT_LEN] [--model MODEL]
+                               [--tokenizer TOKENIZER] [--quantization {awq,gptq,squeezellm,None}] [--tensor-parallel-size TENSOR_PARALLEL_SIZE] [--n N]
+                               [--use-beam-search] [--num-prompts NUM_PROMPTS] [--seed SEED] [--hf-max-batch-size HF_MAX_BATCH_SIZE] [--trust-remote-code]
+                               [--max-model-len MAX_MODEL_LEN] [--dtype {auto,half,float16,bfloat16,float,float32}] [--enforce-eager] [--kv-cache-dtype {auto,fp8}]
+                               [--quantization-param-path KV_CACHE_quantization_param_path]
+
+Benchmark Throughput Example  
+optional arguments:
+  -h, --help  show this help message and exit
+  --backend {vllm,hf,mii}
+  --dataset DATASET  Path to the dataset.
+  --input-len INPUT_LEN  Input prompt length for each request
+  --output-len OUTPUT_LEN  Output length for each request. Overrides the output length from the dataset.
+  --model MODEL
+  --tokenizer TOKENIZER
+  --quantization {awq,gptq,squeezellm,None}, -q {awq,gptq,squeezellm,None}
+  --tensor-parallel-size TENSOR_PARALLEL_SIZE, -tp TENSOR_PARALLEL_SIZE
+  --n N  Number of generated sequences per prompt.
+  --use-beam-search
+  --num-prompts NUM_PROMPTS  Number of prompts to process.
+  --seed SEED
+  --hf-max-batch-size HF_MAX_BATCH_SIZE   Maximum batch size for HF backend.
+  --trust-remote-code trust remote code from huggingface
+  --max-model-len MAX_MODEL_LEN  Maximum length of a sequence (including prompt and output). If None, will be derived from the model.
+  --dtype {auto,half,float16,bfloat16,float,float32}  data type for model weights and activations. The "auto" option will use FP16 precision for FP32 and FP16 models, and BF16 precision for BF16 models.
+  --enforce-eager  enforce eager execution
+  --kv-cache-dtype {auto,fp8} 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.
+  --quantization-param-path QUANT_PARAM_JSON 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.
+```
+```
+Example:
+python3 benchmarks/benchmark_throughput.py --input-len <INPUT_LEN> --output-len <OUTPUT_LEN> -tp <TENSOR_PARALLEL_SIZE> --kv-cache-dtype fp8 --quantization-param-path <path/to/kv_cache_scales.json> --model <path-to-llama2>
+```python

examples/fp8/extract_scales.py

@@ -0,0 +1,367 @@
+import argparse
+import glob
+import json
+import os
+from typing import Any, Callable, Dict, Iterable, List, Optional, Tuple
+
+import numpy as np
+import torch
+from safetensors.torch import safe_open
+
+from vllm.model_executor.layers.quantization.schema import QuantParamSchema
+
+
+# Adapted from vllm/model_executor/weight_utils.py
+# The main differences are that we add the NPZ format and simplify
+# its functionality drastically for our purposes (e.g. we assume that
+# the quantized model exists locally and there is no need to download it)
+def _prepare_hf_weights(
+    quantized_model_dir: str,
+    load_format: str = "auto",
+    fall_back_to_pt: bool = True,
+) -> Tuple[str, List[str], bool]:
+    if not os.path.isdir(quantized_model_dir):
+        raise FileNotFoundError(
+            f"The quantized model directory `{quantized_model_dir}` "
+            "does not exist.")
+    use_safetensors = False
+    # Some quantized models use .pt files for storing the weights.
+    if load_format == "auto":
+        allow_patterns = ["*.safetensors", "*.bin"]
+    elif load_format == "safetensors":
+        use_safetensors = True
+        allow_patterns = ["*.safetensors"]
+    elif load_format == "pt":
+        allow_patterns = ["*.pt"]
+    elif load_format == "npz":
+        allow_patterns = ["*.npz"]
+    else:
+        raise ValueError(f"Unknown load_format: {load_format}")
+    if fall_back_to_pt:
+        allow_patterns += ["*.pt"]
+
+    hf_weights_files: List[str] = []
+    for pattern in allow_patterns:
+        hf_weights_files += glob.glob(
+            os.path.join(quantized_model_dir, pattern))
+        if len(hf_weights_files) > 0:
+            if pattern == "*.safetensors":
+                use_safetensors = True
+            break
+
+    if not use_safetensors:
+        # Exclude files that are not needed for inference.
+        # https://github.com/huggingface/transformers/blob/v4.34.0/src/transformers/trainer.py#L227-L233
+        blacklist = [
+            "training_args.bin",
+            "optimizer.bin",
+            "optimizer.pt",
+            "scheduler.pt",
+            "scaler.pt",
+        ]
+        hf_weights_files = [
+            f for f in hf_weights_files
+            if not any(f.endswith(x) for x in blacklist)
+        ]
+
+    if len(hf_weights_files) == 0:
+        raise RuntimeError(
+            f"Cannot find any model weights with `{quantized_model_dir}`")
+
+    return hf_weights_files, use_safetensors
+
+
+# Adapted from vllm/model_executor/weight_utils.py
+def _hf_tensorfile_iterator(filename: str, load_format: str,
+                            use_safetensors: bool):
+    if load_format == "npz":
+        assert not use_safetensors
+        with np.load(filename) as data:
+            for name in data.files:
+                param = torch.from_numpy(data[name])
+                yield name, param
+    elif use_safetensors:
+        with safe_open(filename, framework="pt") as f:
+            for name in f.keys():  # NOQA: SIM118
+                param = f.get_tensor(name)
+                yield name, param
+    else:
+        state = torch.load(filename, map_location="cpu")
+        for name, param in state.items():
+            yield name, param
+        del state
+        torch.cuda.empty_cache()
+
+
+def _kv_scales_extractor(
+        hf_tensor_files: Iterable[str],
+        use_safetensors: bool,
+        rank_keyword: str = "rank",
+        expected_tp_size: Optional[int] = None) -> Dict[int, Dict[int, float]]:
+    """
+    Given a list of files containing tensor data, attempt to extract KV cache
+    scales from these files. Intended as a helper function taking in the output
+    from _prepare_hf_weights.
+    Args:
+    rank_keyword        Matches the number immediately after this keyword in the
+                        tensor filename to determine the TP rank corresponding
+                        to said tensor file
+    expected_tp_size    If specified, the TP size of the tensor files is checked
+                        against this and an error is raised if they don't match.
+    Returns a dictionary mapping TP ranks to their relevant KV cache scales.
+    The per-rank scales are themselves represented as a dictionary of layer
+    indices to the respective per-layer scale.
+    """
+    for char in rank_keyword:
+        assert not char.isdecimal(
+        ), f"Rank keyword {rank_keyword} contains a numeric character!"
+    rank_scales_map = {}
+    for tensor_file in hf_tensor_files:
+        try:
+            rank_idx = tensor_file.find(rank_keyword)
+            if rank_idx != -1:
+                start_idx = rank_idx + len(rank_keyword)
+                stop_idx = start_idx
+                while stop_idx < len(
+                        tensor_file) and tensor_file[stop_idx].isdecimal():
+                    stop_idx += 1
+                if stop_idx == start_idx:
+                    raise RuntimeError("Did not find rank # in filename.")
+                rank = int(tensor_file[start_idx:stop_idx])
+            elif len(hf_tensor_files) == 1:
+                # Since there is only one tensor file, we can assume
+                # that it's intended for TP rank 0
+                rank = 0
+            else:
+                raise RuntimeError(
+                    f"Filename does not contain '{rank_keyword}'.")
+        except RuntimeError:
+            print("Unable to determine TP rank "
+                  f"corresponding to file '{tensor_file}'")
+            raise
+
+        if rank not in rank_scales_map:
+            layer_scales_map = {}
+            rank_scales_map[rank] = layer_scales_map
+        else:
+            raise RuntimeError(
+                f"Tensor file '{tensor_file}' shares TP rank {rank} "
+                "with another tensor file.")
+
+        module_delimiter = ":" if args.load_format == "npz" else "."
+        for name, param in _hf_tensorfile_iterator(tensor_file,
+                                                   args.load_format,
+                                                   use_safetensors):
+            if "kv_cache_scaling_factor" in name:
+                nums = [
+                    int(s) for s in name.split(module_delimiter)
+                    if s.isdecimal()
+                ]
+                assert len(
+                    nums) == 1, f"Could not determine layer idx for {name}"
+                layer_idx = nums[0]
+                assert layer_idx not in layer_scales_map, f"Duplicate scaling"\
+                    f" factor corresponding to layer {layer_idx}"
+                try:
+                    layer_scales_map[layer_idx] = param.item()
+                except RuntimeError:
+                    print(
+                        "This utility supports only per-tensor scalar scales "
+                        f"for now. The tensor\n {name} = {param} \nis an "
+                        "invalid scale factor.")
+                    raise
+
+    if all(
+            len(layer_scales_map) == 0
+            for layer_scales_map in rank_scales_map.values()):
+        # Note: this is true even if the rank_scales_map is empty
+        print("WARNING: No KV cache scale factors found. No output saved.")
+        return None
+    empirical_tp_world_size = max(rank_scales_map.keys()) + 1
+    if expected_tp_size is not None:
+        assert expected_tp_size == empirical_tp_world_size, \
+            f"User expected TP world size = {expected_tp_size} " \
+            "from model but tool is expecting TP world size = " \
+            f"{empirical_tp_world_size} from model instead."
+    for i in range(empirical_tp_world_size):
+        assert i in rank_scales_map, "Expected TP world size = "\
+            f"{empirical_tp_world_size} but did not find KV " \
+            f"cache scaling factors for TP rank {i}"
+    print(f"Found TP world size = {empirical_tp_world_size} "
+          "when extracting KV cache scales!")
+    return rank_scales_map
+
+
+def _metadata_extractor(quantized_model_dir: str,
+                        metadata_extract_fns: \
+                        Dict[str, Callable[[Dict[str, Any]], Any]]) \
+                        -> Dict[str, Any]:
+    """
+    Given a directory containing quantized model files, this function
+    aims to extract metadata from the JSON files within this directory.
+    Each JSON file is expected to represent a dictionary in JSON
+    format (referred to as a "JSON-dictionary"). Metadata extraction is
+    defined by a dictionary called metadata_extract_fns, where each
+    metadata field name is mapped to an extraction function.
+
+    These extraction functions are designed to take a JSON-dictionary
+    as their only argument  and return the corresponding metadata.
+    While extraction functions are permitted to raise  exceptions, they
+    should only raise a KeyError or ValueError if the metadata field
+    cannot  be extracted from the current JSON-dictionary, yet there's
+    a possibility of finding it in another JSON-dictionary.
+
+    The function returns a dictionary that maps metadata fields to
+    their extracted data. The keys of this dictionary correspond exactly
+    to those in metadata_extract_fns. If any fields fail to be extracted,
+    their corresponding values are set to None, and a warning is printed.
+    """
+    if not os.path.isdir(quantized_model_dir):
+        raise FileNotFoundError(
+            f"The quantized model directory `{quantized_model_dir}` "
+            "does not exist.")
+    metadata_files = glob.glob(os.path.join(quantized_model_dir, "*.json"))
+
+    result = {}
+    for file in metadata_files:
+        with open(file) as f:
+            try:
+                metadata = json.load(f)
+            except json.JSONDecodeError:
+                print(f"Could not parse `{file}` as a valid metadata file,"
+                      " skipping it.")
+                continue
+            if not isinstance(metadata, dict):
+                print(f"The file `{file}` does not correspond to a "
+                      "JSON-serialized dictionary, skipping it.")
+                continue
+            for metadata_name, extract_fn in metadata_extract_fns.items():
+                try:
+                    metadata_info = extract_fn(metadata)
+                    if metadata_name not in result:
+                        result[metadata_name] = metadata_info
+                    elif metadata_info != result[metadata_name]:
+                        raise RuntimeError(
+                            "Metadata mismatch! Originally found "
+                            f"{metadata_name} = {result[metadata_name]} but "
+                            f"now found {metadata_name} = {metadata_info} in "
+                            f"`{file}`")
+                except KeyError:
+                    # It is possible that a given file does not contain some
+                    # of our selected metadata as it could be located in some
+                    # other metadata file.
+                    # 'EFINAE': extract_fn failure is not an error.
+                    pass
+                except ValueError:
+                    # See above.
+                    pass
+
+    # Warn if we cannot find any of the requested metadata
+    for metadata_name in metadata_extract_fns:
+        if metadata_name not in result:
+            print("WARNING: Unable to find requested metadata field "
+                  f"`{metadata_name}`, setting it to None.")
+            result[metadata_name] = None
+
+    return result
+
+
+def main(args):
+    metadata_extract_fns = {
+        "model_type": lambda json_dict: json_dict["layers"][0]["decoder_type"],
+        "tp_size": lambda json_dict: int(json_dict["tensor_parallel"]),
+        "model_dtype": lambda json_dict: json_dict["dtype"]
+    }
+    recovered_metadata = _metadata_extractor(args.quantized_model,
+                                             metadata_extract_fns)
+    if args.tp_size is not None:
+        metadata_tp_size = recovered_metadata["tp_size"]
+        if metadata_tp_size is not None:
+            assert args.tp_size == metadata_tp_size, \
+              f"User expected TP world size = {args.tp_size} " \
+              f"but found TP world size = {metadata_tp_size} from metadata!"
+    expected_tp_size = args.tp_size or recovered_metadata["tp_size"]
+    rank_keyword = "rank"
+    hf_tensor_files, use_safetensors = _prepare_hf_weights(
+        args.quantized_model, args.load_format)
+    rank_scales_map = _kv_scales_extractor(hf_tensor_files, use_safetensors,
+                                           rank_keyword, expected_tp_size)
+    # Postprocess: formatting to the current schema. Consider pulling it
+    # out into a dedicated function should it ever become more complicated.
+    rank_scales_map = {
+        rank: {k: scale[k]
+               for k in sorted(scale.keys())}
+        for rank, scale in rank_scales_map.items()
+    }
+    # TODO: Expand this with activation and weights scaling factors when
+    # they are used in the future
+    schema = QuantParamSchema(
+        model_type=recovered_metadata["model_type"],
+        kv_cache={
+            "dtype": ("float8_e4m3fn" if len(rank_scales_map) > 0 else
+                      recovered_metadata["model_dtype"]),
+            "scaling_factor":
+            rank_scales_map
+        },
+    )
+
+    if args.output_dir is None:
+        output_file = os.path.join(args.quantized_model, args.output_name)
+    else:
+        if not os.path.isdir(args.output_dir):
+            os.makedirs(args.output_dir, exist_ok=True)
+        output_file = os.path.join(args.output_dir, args.output_name)
+
+    with open(output_file, 'w') as f:
+        f.write(schema.model_dump_json(indent=4))
+        print(f"Completed! KV cache scaling factors saved to {output_file}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        description="This simple utility extracts the "
+        "KV cache scaling factors from a quantized HF model "
+        "and saves them to a JSON file compatible with later "
+        "use by vLLM (pass this file to the appropriate "
+        "runtime typically using the argument "
+        "--quantization-param-path <filename>). This is only used "
+        "if the KV cache dtype is FP8 and on ROCm (AMD GPU).")
+    parser.add_argument(
+        "--quantized_model",
+        help="Specify the directory containing a single quantized HF model. "
+        "It is expected that the quantization format is FP8_E4M3, for use "
+        "on ROCm (AMD GPU).",
+        required=True)
+    parser.add_argument(
+        "--load_format",
+        help="Optionally specify the format of the model's tensor files "
+        "containing the KV cache scaling factors.",
+        choices=["auto", "safetensors", "npz", "pt"],
+        default="auto")
+    parser.add_argument(
+        "--output_dir",
+        help="Optionally specify the output directory. By default the "
+        "KV cache scaling factors will be saved in the model directory, "
+        "however you can override this behavior here.",
+        default=None)
+    parser.add_argument(
+        "--output_name",
+        help="Optionally specify the output filename.",
+        # TODO: Change this once additional scaling factors are enabled
+        default="kv_cache_scales.json")
+    parser.add_argument(
+        "--tp_size",
+        help="Optionally specify the tensor-parallel (TP) size that the "
+        "quantized model should correspond to. If specified, during KV "
+        "cache scaling factor extraction the observed TP size will be "
+        "checked against this and an error will be raised if there is "
+        "a mismatch. If not specified, the quantized model's expected "
+        "TP size is instead inferred from the largest TP rank observed. "
+        "The expected TP size is cross-checked against the TP ranks "
+        "observed in the quantized model and an error is raised if any "
+        "discrepancies are found.",
+        default=None,
+        type=int)
+    args = parser.parse_args()
+
+    main(args)

examples/fp8/quantizer/README.md

@@ -0,0 +1,32 @@
+### Quantizer Utilities
+`quantize.py`: NVIDIA Quantization utilities using AMMO, ported from TensorRT-LLM:
+`https://github.com/NVIDIA/TensorRT-LLM/blob/main/examples/quantization/quantize.py`
+
+### Prerequisite
+
+#### AMMO (AlgorithMic Model Optimization) Installation: nvidia-ammo 0.7.1 or later
+`pip install --no-cache-dir --extra-index-url https://pypi.nvidia.com nvidia-ammo` 
+
+#### AMMO Download (code and docs)
+`https://developer.nvidia.com/downloads/assets/cuda/files/nvidia-ammo/nvidia_ammo-0.5.0.tar.gz`
+`https://developer.nvidia.com/downloads/assets/cuda/files/nvidia-ammo/nvidia_ammo-0.7.1.tar.gz`
+
+### Usage
+
+#### Run on H100 system for speed if FP8; number of GPUs depends on the model size
+
+#### Example: quantize Llama2-7b model from HF to FP8 with FP8 KV Cache:
+`python quantize.py --model_dir ./ll2-7b --dtype float16 --qformat fp8 --kv_cache_dtype fp8 --output_dir ./ll2_7b_fp8 --calib_size 512 --tp_size 1`
+
+Outputs: model structure, quantized model & parameters (with scaling factors) are in JSON and Safetensors (npz is generated only for the reference)
+```
+# ll ./ll2_7b_fp8/
+total 19998244
+drwxr-xr-x 2 root root        4096 Feb  7 01:08 ./
+drwxrwxr-x 8 1060 1061        4096 Feb  7 01:08 ../
+-rw-r--r-- 1 root root      176411 Feb  7 01:08 llama_tp1.json
+-rw-r--r-- 1 root root 13477087480 Feb  7 01:09 llama_tp1_rank0.npz
+-rw-r--r-- 1 root root  7000893272 Feb  7 01:08 rank0.safetensors
+#
+```
+

examples/fp8/quantizer/quantize.py

@@ -0,0 +1,367 @@
+# SPDX-FileCopyrightText: Copyright (c) 2022-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved. # noqa: E501
+# 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.
+"""
+Adapted from examples/quantization/hf_ptq.py
+"""
+
+import argparse
+import copy
+import json
+import random
+import time
+
+import ammo.torch.quantization as atq
+import numpy as np
+import torch
+from ammo.torch.export import export_model_config
+from datasets import load_dataset
+from torch.utils.data import DataLoader
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+RAND_SEED = 1234
+MAX_SEQ_LEN = 2048
+
+EMPTY_CFG = {
+    "quant_cfg": {
+        "*weight_quantizer": {
+            "enable": False,
+        },
+        "*input_quantizer": {
+            "enable": False
+        },
+        "*lm_head*": {
+            "enable": False
+        },
+        "*output_layer*": {
+            "enable": False
+        },
+        "default": {
+            "enable": False
+        },
+    },
+    "algorithm": "max",
+}
+
+KV_CACHE_CFG = {
+    "*.query_key_value.output_quantizer": {
+        "num_bits": 8,
+        "axis": None,
+        "enable": True
+    },
+    "*.Wqkv.output_quantizer": {
+        "num_bits": 8,
+        "axis": None,
+        "enable": True
+    },
+    "*.W_pack.output_quantizer": {
+        "num_bits": 8,
+        "axis": None,
+        "enable": True
+    },
+    "*.c_attn.output_quantizer": {
+        "num_bits": 8,
+        "axis": None,
+        "enable": True
+    },
+    "*.k_proj.output_quantizer": {
+        "num_bits": 8,
+        "axis": None,
+        "enable": True
+    },
+    "*.v_proj.output_quantizer": {
+        "num_bits": 8,
+        "axis": None,
+        "enable": True
+    },
+}
+
+QUANT_CFG_CHOICES = {
+    "int8_sq": atq.INT8_SMOOTHQUANT_CFG,
+    "fp8": atq.FP8_DEFAULT_CFG,
+    "int4_awq": atq.INT4_AWQ_CFG,
+    "w4a8_awq": atq.W4A8_AWQ_BETA_CFG,
+    "int8_wo": EMPTY_CFG,
+    "int4_wo": EMPTY_CFG,
+    "full_prec": EMPTY_CFG,
+}
+
+MODEL_NAME_PATTERN_MAP = {
+    "GPT2": "gpt2",
+    "Xverse": "llama",
+    "Llama": "llama",
+    "Mistral": "llama",
+    "GPTJ": "gptj",
+    "FalconForCausalLM": "falcon",
+    "RWForCausalLM": "falcon",
+    "baichuan": "baichuan",
+    "MPT": "mpt",
+    "Bloom": "bloom",
+    "ChatGLM": "chatglm",
+    "QWen": "qwen",
+}
+
+
+def get_tokenizer(ckpt_path, max_seq_len=MAX_SEQ_LEN, model_type=None):
+    print(f"Initializing tokenizer from {ckpt_path}")
+    tokenizer = AutoTokenizer.from_pretrained(
+        ckpt_path,
+        model_max_length=max_seq_len,
+        padding_side="left",
+        trust_remote_code=True,
+    )
+    if model_type and model_type == "qwen":
+        # qwen use token id 151643 as pad and eos tokens
+        tokenizer.pad_token = tokenizer.convert_ids_to_tokens(151643)
+        tokenizer.eos_token = tokenizer.convert_ids_to_tokens(151643)
+
+    # can't set attribute 'pad_token' for "<unk>"
+    if tokenizer.pad_token != "<unk>":
+        tokenizer.pad_token = tokenizer.eos_token
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    assert (tokenizer.pad_token
+            is not None), f"Pad token for {model_type} cannot be set!"
+
+    return tokenizer
+
+
+def get_model(ckpt_path, dtype="fp16", device="cuda"):
+    print(f"Initializing model from {ckpt_path}")
+    if dtype == "bf16" or dtype == "bfloat16":
+        dtype = torch.bfloat16
+    elif dtype == "fp16" or dtype == "float16":
+        dtype = torch.float16
+    elif dtype == "fp32" or dtype == "float32":
+        dtype = torch.float32
+    else:
+        raise NotImplementedError(f"Unknown dtype {dtype}")
+
+    # model_kwargs = {"torch_dtype": dtype}
+    model_kwargs = {"torch_dtype": "auto"}
+
+    model = AutoModelForCausalLM.from_pretrained(ckpt_path,
+                                                 device_map="auto",
+                                                 **model_kwargs,
+                                                 trust_remote_code=True)
+    model.eval()
+
+    model_dtype = next(model.parameters()).dtype
+    if dtype != model_dtype:
+        print("[TensorRT-LLM][WARNING] The manually set model data type is "
+              f"{dtype}, but the data type of the HuggingFace model is "
+              f"{model_dtype}.")
+
+    return model
+
+
+def get_model_type(model):
+    for k, v in MODEL_NAME_PATTERN_MAP.items():
+        if k.lower() in type(model).__name__.lower():
+            return v
+    return None
+
+
+def get_calib_dataloader(data="cnn_dailymail",
+                         tokenizer=None,
+                         batch_size=1,
+                         calib_size=512,
+                         block_size=512,
+                         device=None):
+    print("Loading calibration dataset")
+    if data == "pileval":
+        dataset = load_dataset(
+            "json",
+            data_files="https://the-eye.eu/public/AI/pile/val.jsonl.zst",
+            split="train")
+        dataset = dataset["text"][:calib_size]
+    elif data == "cnn_dailymail":
+        dataset = load_dataset("cnn_dailymail", name="3.0.0", split="train")
+        dataset = dataset["article"][:calib_size]
+    else:
+        raise NotImplementedError
+
+    batch_encoded = tokenizer.batch_encode_plus(dataset,
+                                                return_tensors="pt",
+                                                padding="max_length",
+                                                truncation=True,
+                                                max_length=block_size)
+    if device:
+        batch_encoded = batch_encoded.to(device)
+    batch_encoded = batch_encoded["input_ids"]
+
+    calib_dataloader = DataLoader(batch_encoded,
+                                  batch_size=batch_size,
+                                  shuffle=False)
+
+    return calib_dataloader
+
+
+def quantize_model(model, quant_cfg, calib_dataloader=None):
+
+    def calibrate_loop():
+        if calib_dataloader is None:
+            return
+        """Adjusts weights and scaling factors based on selected algorithms."""
+        for idx, data in enumerate(calib_dataloader):
+            print(f"Calibrating batch {idx}")
+            model(data)
+
+    print("Starting quantization...")
+    start_time = time.time()
+    atq.quantize(model, quant_cfg, forward_loop=calibrate_loop)
+    end_time = time.time()
+    print("Quantization done. Total time used: {:.2f} s.".format(end_time -
+                                                                 start_time))
+
+    return model
+
+
+def main(args):
+    if not torch.cuda.is_available():
+        raise EnvironmentError("GPU is required for inference.")
+
+    random.seed(RAND_SEED)
+    np.random.seed(RAND_SEED)
+
+    model = get_model(args.model_dir, args.dtype, args.device)
+    model_type = get_model_type(model)
+    tokenizer = get_tokenizer(args.model_dir, model_type=model_type)
+
+    if args.qformat in ["full_prec", "int8_wo", "int4_wo"
+                        ] and args.kv_cache_dtype is None:
+        print(f"No quantization applied, export {args.dtype} model")
+    else:
+        if "awq" in args.qformat:
+            if args.calib_size > 32:
+                print("AWQ calibration could take longer with calib_size = "
+                      f"{args.calib_size}, Using calib_size=32 instead")
+                args.calib_size = 32
+            print("\nAWQ calibration could take longer than other calibration "
+                  "methods. Please increase the batch size to speed up the "
+                  "calibration process. Batch size can be set by adding the "
+                  "argument --batch_size <batch_size> to the command line.\n")
+
+        calib_dataloader = get_calib_dataloader(
+            tokenizer=tokenizer,
+            batch_size=args.batch_size,
+            calib_size=args.calib_size,
+            device=args.device,
+        )
+
+        if args.qformat in QUANT_CFG_CHOICES:
+            quant_cfg = QUANT_CFG_CHOICES[args.qformat]
+        else:
+            raise ValueError(
+                f"Unsupported quantization format: {args.qformat}")
+
+        if "awq" in args.qformat:
+            quant_cfg = copy.deepcopy(QUANT_CFG_CHOICES[args.qformat])
+            weight_quantizer = quant_cfg["quant_cfg"][
+                "*weight_quantizer"]  # type: ignore
+            if isinstance(weight_quantizer, list):
+                weight_quantizer = weight_quantizer[0]
+            weight_quantizer["block_sizes"][-1] = args.awq_block_size
+
+        if args.kv_cache_dtype is not None:
+            if args.kv_cache_dtype == "fp8":
+                for value in KV_CACHE_CFG.values():
+                    value.update({"num_bits": (4, 3)})  # type: ignore
+            quant_cfg["quant_cfg"].update(KV_CACHE_CFG)  # type: ignore
+
+        print(quant_cfg)
+
+        model = quantize_model(model, quant_cfg, calib_dataloader)
+
+    with torch.inference_mode():
+        if model_type is None:
+            print(f"Unknown model type {type(model).__name__}. Continue "
+                  "exporting...")
+            model_type = f"unknown:{type(model).__name__}"
+
+        export_path = args.output_dir
+        start_time = time.time()
+
+        if args.qformat == "int4_awq" and model_type == "qwen":
+            torch.save(model.state_dict(), export_path)
+        else:
+            export_npz = (model_type not in [
+                'gptj', 'falcon', 'chatglm', 'mpt', 'llama', 'baichuan'
+            ])
+
+            # export safetensors
+            export_model_config(
+                model,
+                model_type,
+                getattr(torch, args.dtype),
+                export_dir=export_path,
+                inference_tensor_parallel=args.tp_size,
+                inference_pipeline_parallel=args.pp_size,
+                # export_tensorrt_llm_config=(not export_npz),
+                export_tensorrt_llm_config=False,
+                export_npz=export_npz)
+
+            # Workaround for wo quantization
+            if args.qformat in ["int8_wo", "int4_wo", "full_prec"]:
+                with open(f"{export_path}/config.json", 'r') as f:
+                    tensorrt_llm_config = json.load(f)
+                if args.qformat == "int8_wo":
+                    tensorrt_llm_config["quantization"]["quant_algo"] = 'W8A16'
+                elif args.qformat == "int4_wo":
+                    tensorrt_llm_config["quantization"]["quant_algo"] = 'W4A16'
+                else:
+                    tensorrt_llm_config["quantization"]["quant_algo"] = None
+                with open(f"{export_path}/config.json", "w") as f:
+                    json.dump(tensorrt_llm_config, f, indent=4)
+
+        end_time = time.time()
+        print("Quantized model exported to {} \nTotal time used {:.2f} s.".
+              format(export_path, end_time - start_time))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--model_dir",
+                        help="Specify where the HuggingFace model is",
+                        required=True)
+    parser.add_argument("--device", default="cuda")
+    parser.add_argument("--dtype", help="Model data type.", default="float16")
+    parser.add_argument(
+        "--qformat",
+        help="Quantization format.",
+        default="full_prec",
+        choices=[
+            "fp8", "int8_sq", "int4_awq", "w4a8_awq", "int8_wo", "int4_wo",
+            "full_prec"
+        ],
+    )
+    parser.add_argument("--batch_size",
+                        help="Batch size for calibration.",
+                        type=int,
+                        default=1)
+    parser.add_argument("--calib_size",
+                        help="Number of samples for calibration.",
+                        type=int,
+                        default=512)
+    parser.add_argument("--output_dir", default="exported_model")
+    parser.add_argument("--tp_size", type=int, default=1)
+    parser.add_argument("--pp_size", type=int, default=1)
+    parser.add_argument("--awq_block_size", type=int, default=128)
+    parser.add_argument("--kv_cache_dtype",
+                        help="KV Cache dtype.",
+                        default=None,
+                        choices=["int8", "fp8", None])
+    args = parser.parse_args()
+
+    main(args)

pyproject.toml

@@ -5,7 +5,7 @@ requires = [
     "ninja",
     "packaging",
     "setuptools >= 49.4.0",
-    "torch == 2.1.2",
+    "torch == 2.2.1",
     "wheel",
 ]
 build-backend = "setuptools.build_meta"
@@ -13,6 +13,10 @@ build-backend = "setuptools.build_meta"
 [tool.ruff]
 # Allow lines to be as long as 80.
 line-length = 80
+exclude = [
+    # External file, leaving license intact
+    "examples/fp8/quantizer/quantize.py"
+]
 
 [tool.ruff.lint]
 select = [

requirements-build.txt

@@ -3,5 +3,5 @@ cmake>=3.21
 ninja
 packaging
 setuptools>=49.4.0
-torch==2.1.2
+torch==2.2.1
 wheel

requirements-common.txt

@@ -1,20 +1,14 @@
-cmake>=3.21
+cmake >= 3.21
 ninja  # For faster builds.
 psutil
-ray >= 2.9
 sentencepiece  # Required for LLaMA tokenizer.
 numpy
-torch == 2.1.2
 requests
-psutil
 py-cpuinfo
 transformers >= 4.39.1  # Required for StarCoder2 & Llava.
-xformers == 0.0.23.post1  # Required for CUDA 12.1.
 fastapi
 uvicorn[standard]
 pydantic >= 2.0  # Required for OpenAI server.
 prometheus_client >= 0.18.0
-pynvml == 11.5.0
-triton >= 2.1.0
-outlines == 0.0.34
-tiktoken == 0.6.0 # Required for DBRX tokenizer
+tiktoken == 0.6.0  # Required for DBRX tokenizer
+outlines == 0.0.34 # Requires torch >= 2.1.0

requirements-cpu.txt

@@ -0,0 +1,6 @@
+# Common dependencies
+-r requirements-common.txt
+
+# Dependencies for x86_64 CPUs
+torch == 2.2.1+cpu
+triton >= 2.1.0  # FIXME(woosuk): This is a hack to avoid import error.

requirements-cuda.txt

@@ -0,0 +1,10 @@
+# Common dependencies
+-r requirements-common.txt
+
+# Dependencies for NVIDIA GPUs
+ray >= 2.9
+pynvml == 11.5.0
+vllm-nccl-cu12>=2.18,<2.19  # for downloading nccl library
+torch == 2.2.1
+xformers == 0.0.25  # Requires PyTorch 2.2.1
+triton >= 2.1.0

requirements-neuron.txt

@@ -1,12 +1,7 @@
-sentencepiece  # Required for LLaMA tokenizer.
-numpy
+# Common dependencies
+-r requirements-common.txt
+
+# Dependencies for Neuron devices
 transformers-neuronx >= 0.9.0
 torch-neuronx >= 2.1.0
 neuronx-cc
-fastapi
-uvicorn[standard]
-pydantic >= 2.0  # Required for OpenAI server.
-prometheus_client >= 0.18.0
-requests
-psutil
-py-cpuinfo

requirements-rocm.txt

@@ -1,17 +1,5 @@
-cmake>=3.21
-ninja  # For faster builds.
-typing-extensions>=4.8.0
-starlette
-requests
-py-cpuinfo
-psutil
+# Common dependencies
+-r requirements-common.txt
+
+# Dependencies for AMD GPUs
 ray == 2.9.3
-sentencepiece  # Required for LLaMA tokenizer.
-numpy
-tokenizers>=0.15.0
-transformers >= 4.39.1  # Required for StarCoder2 & Llava.
-fastapi
-uvicorn[standard]
-pydantic >= 2.0  # Required for OpenAI server.
-prometheus_client >= 0.18.0
-outlines == 0.0.34

setup.py

@@ -15,6 +15,8 @@ from torch.utils.cpp_extension import CUDA_HOME
 
 ROOT_DIR = os.path.dirname(__file__)
 logger = logging.getLogger(__name__)
+# Target device of vLLM, supporting [cuda (by default), rocm, neuron, cpu]
+VLLM_TARGET_DEVICE = os.getenv("VLLM_TARGET_DEVICE", "cuda")
 
 # vLLM only supports Linux platform
 assert sys.platform.startswith(
@@ -112,6 +114,7 @@ class cmake_build_ext(build_ext):
             '-DCMAKE_BUILD_TYPE={}'.format(cfg),
             '-DCMAKE_LIBRARY_OUTPUT_DIRECTORY={}'.format(outdir),
             '-DCMAKE_ARCHIVE_OUTPUT_DIRECTORY={}'.format(self.build_temp),
+            '-DVLLM_TARGET_DEVICE={}'.format(VLLM_TARGET_DEVICE),
         ]
 
         verbose = bool(int(os.getenv('VERBOSE', '0')))
@@ -186,11 +189,14 @@ class cmake_build_ext(build_ext):
 
 def _is_cuda() -> bool:
     has_cuda = torch.version.cuda is not None
-    return has_cuda and not (_is_neuron() or _is_tpu())
+    return (VLLM_TARGET_DEVICE == "cuda"
+            and has_cuda
+            and not (_is_neuron() or _is_tpu()))
 
 
 def _is_hip() -> bool:
-    return torch.version.hip is not None
+    return (VLLM_TARGET_DEVICE == "cuda"
+            or VLLM_TARGET_DEVICE == "rocm") and torch.version.hip is not None
 
 
 def _is_neuron() -> bool:
@@ -206,8 +212,12 @@ def _is_tpu() -> bool:
     return True  # FIXME
 
 
+def _is_cpu() -> bool:
+    return VLLM_TARGET_DEVICE == "cpu"
+
+
 def _build_custom_ops() -> bool:
-    return _is_cuda() or _is_hip()
+    return _is_cuda() or _is_hip() or _is_cpu()
 
 
 def _install_punica() -> bool:
@@ -307,6 +317,8 @@ def get_vllm_version() -> str:
             version += f"+neuron{neuron_version_str}"
     elif _is_tpu():
         version += "+tpu"
+    elif _is_cpu():
+        version += "+cpu"
     else:
         raise RuntimeError("Unknown runtime environment")
 
@@ -324,22 +336,40 @@ def read_readme() -> str:
 
 def get_requirements() -> List[str]:
     """Get Python package dependencies from requirements.txt."""
+
+    def _read_requirements(filename: str) -> List[str]:
+        with open(get_path(filename)) as f:
+            requirements = f.read().strip().split("\n")
+        resolved_requirements = []
+        for line in requirements:
+            if line.startswith("-r "):
+                resolved_requirements += _read_requirements(line.split()[1])
+            else:
+                resolved_requirements.append(line)
+        return resolved_requirements
+
     if _is_cuda():
-        with open(get_path("requirements.txt")) as f:
-            requirements = f.read().strip().split("\n")
+        requirements = _read_requirements("requirements-cuda.txt")
+        cuda_major = torch.version.cuda.split(".")[0]
+        modified_requirements = []
+        for req in requirements:
+            if "vllm-nccl-cu12" in req:
+                modified_requirements.append(
+                    req.replace("vllm-nccl-cu12", f"vllm-nccl-cu{cuda_major}"))
+            else:
+                modified_requirements.append(req)
+        requirements = modified_requirements
     elif _is_hip():
-        with open(get_path("requirements-rocm.txt")) as f:
-            requirements = f.read().strip().split("\n")
+        requirements = _read_requirements("requirements-rocm.txt")
     elif _is_neuron():
-        with open(get_path("requirements-neuron.txt")) as f:
-            requirements = f.read().strip().split("\n")
+        requirements = _read_requirements("requirements-neuron.txt")
     elif _is_tpu():
-        with open(get_path("requirements-tpu.txt")) as f:
-            requirements = f.read().strip().split("\n")
+        requirements = _read_requirements("requirements-tpu.txt")
+    elif _is_cpu():
+        requirements = _read_requirements("requirements-cpu.txt")
     else:
         raise ValueError(
-            "Unsupported platform, please use CUDA, ROCM or Neuron.")
-
+            "Unsupported platform, please use CUDA, ROCm, Neuron, or CPU.")
     return requirements
 
 

tests/conftest.py

@@ -55,10 +55,24 @@ def cleanup():
     torch.cuda.empty_cache()
 
 
+@pytest.fixture()
+def should_do_global_cleanup_after_test(request) -> bool:
+    """Allow subdirectories to skip global cleanup by overriding this fixture.
+    This can provide a ~10x speedup for non-GPU unit tests since they don't need
+    to initialize torch.
+    """
+
+    if request.node.get_closest_marker("skip_global_cleanup"):
+        return False
+
+    return True
+
+
 @pytest.fixture(autouse=True)
-def cleanup_fixture():
+def cleanup_fixture(should_do_global_cleanup_after_test: bool):
     yield
-    cleanup()
+    if should_do_global_cleanup_after_test:
+        cleanup()
 
 
 @pytest.fixture(scope="session")

tests/core/block/conftest.py

@@ -0,0 +1,12 @@
+import pytest
+
+
+@pytest.fixture()
+def should_do_global_cleanup_after_test() -> bool:
+    """Disable the global cleanup fixture for tests in this directory. This
+    provides a ~10x speedup for unit tests that don't load a model to GPU.
+
+    This requires that tests in this directory clean up after themselves if they
+    use the GPU.
+    """
+    return False

tests/core/block/e2e/conftest.py

@@ -1,25 +1,10 @@
-import contextlib
-import gc
-
 import pytest
-import ray
-import torch
 
+from tests.conftest import cleanup
 from vllm import LLM
-from vllm.model_executor.parallel_utils.parallel_state import (
-    destroy_model_parallel)
 from vllm.model_executor.utils import set_random_seed
 
 
-def cleanup():
-    destroy_model_parallel()
-    with contextlib.suppress(AssertionError):
-        torch.distributed.destroy_process_group()
-    gc.collect()
-    torch.cuda.empty_cache()
-    ray.shutdown()
-
-
 @pytest.fixture
 def baseline_llm_generator(common_llm_kwargs, per_test_common_llm_kwargs,
                            baseline_llm_kwargs, seed):

tests/core/block/e2e/test_correctness.py

@@ -16,7 +16,7 @@ from vllm import SamplingParams
 
         # Allow only 5 sequences of ~1024 tokens in worst case.
         "block_size": 16,
-        "forced_num_gpu_blocks": 5 * (64 + 1),
+        "num_gpu_blocks_override": 5 * (64 + 1),
     }])
 @pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
 @pytest.mark.parametrize("baseline_llm_kwargs", [{
@@ -77,6 +77,159 @@ def test_v1_v2_greedy_equality_with_preemption(baseline_llm_generator,
     assert baseline_token_ids == test_token_ids
 
 
+@pytest.mark.parametrize(
+    "common_llm_kwargs",
+    [{
+        # Use a small model for a fast test.
+        "model": "facebook/opt-125m",
+
+        # skip cuda graph creation for fast test.
+        "enforce_eager": True,
+
+        # Use a large block size to trigger more copy-on-writes.
+        "block_size": 32,
+    }])
+@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
+@pytest.mark.parametrize("baseline_llm_kwargs", [{
+    "use_v2_block_manager": False
+}])
+@pytest.mark.parametrize("test_llm_kwargs", [{"use_v2_block_manager": True}])
+@pytest.mark.parametrize("batch_size", [10])
+@pytest.mark.parametrize("seed", [1])
+def test_v1_v2_greedy_equality_with_cow(baseline_llm_generator,
+                                        test_llm_generator, batch_size):
+    """Verify beam search equality with block manager v1 and v2.
+
+    This requires copy-on-writes; if the v1 and v2 output is the same, then
+    we have some confidence cow is working.
+    """
+    output_len = 128
+    temperature = 0.0
+
+    prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+
+    prompts = [prompt for prompt, _ in zip(cycle(prompts), range(batch_size))]
+
+    sampling_params = SamplingParams(
+        max_tokens=output_len,
+        ignore_eos=True,
+        temperature=temperature,
+        use_beam_search=True,
+        best_of=2,
+    )
+
+    print('Getting token ids from block manager v1')
+    baseline_token_ids = get_token_ids_from_llm_generator(
+        baseline_llm_generator, prompts, sampling_params)
+
+    print('Getting token ids from block manager v2')
+    test_token_ids = get_token_ids_from_llm_generator(test_llm_generator,
+                                                      prompts, sampling_params)
+
+    for expected_token_ids, actual_token_ids in zip(baseline_token_ids,
+                                                    test_token_ids):
+        assert expected_token_ids == actual_token_ids
+
+    assert baseline_token_ids == test_token_ids
+
+
+@pytest.mark.parametrize(
+    "common_llm_kwargs",
+    [{
+        # Use a small model for a fast test.
+        "model": "facebook/opt-125m",
+
+        # Our prompts will generate 128 tokens; since the prompts themselves are
+        # small, we don't need much KV space beyond 128.
+        "max_model_len": 160,
+
+        # skip cuda graph creation for fast test.
+        "enforce_eager": True,
+
+        # Lookahead scheduling only supported in v2 block manager.
+        "use_v2_block_manager": True,
+    }])
+@pytest.mark.parametrize(
+    "per_test_common_llm_kwargs",
+    [
+        {
+            "block_size": 16,
+
+            # Allow only 2 sequences of ~128 tokens in worst case.
+            # Note 8 = 128/block_size
+            "num_gpu_blocks_override": 2 * (8 + 1),
+        },
+        {
+            "block_size": 8,
+
+            # Allow only 2 sequences of ~128 tokens in worst case.
+            # Note 16 = 128/block_size
+            "num_gpu_blocks_override": 2 * (16 + 1),
+        }
+    ])
+@pytest.mark.parametrize("baseline_llm_kwargs", [{
+    "num_lookahead_slots": 0,
+}])
+@pytest.mark.parametrize(
+    "test_llm_kwargs",
+    [{
+        # We run one test with block_size < lookahead_slots, one test with
+        # block_size > lookahead_slots
+        "num_lookahead_slots": 10,
+    }])
+@pytest.mark.parametrize("batch_size", [4])
+@pytest.mark.parametrize("seed", [1])
+def test_lookahead_greedy_equality_with_preemption(baseline_llm_generator,
+                                                   test_llm_generator,
+                                                   batch_size):
+    """Verify vLLM produces the same output with greedy sampling, when lookahead
+    scheduling is used vs. not.
+
+    Lookahead scheduling is not expected to modify the output, as it simply
+    allocates empty slots ahead of the known token ids in a sliding fashion.
+
+    This test constrains the total number of blocks to force preemption. It also
+    varies the block size so that the lookahead size is less than and greater
+    than the block size.
+    """
+    output_len = 128
+    temperature = 0.0
+
+    prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+
+    prompts = [prompt for prompt, _ in zip(cycle(prompts), range(batch_size))]
+
+    sampling_params = SamplingParams(
+        max_tokens=output_len,
+        ignore_eos=True,
+        temperature=temperature,
+    )
+
+    print('Getting token ids without lookahead scheduling')
+    baseline_token_ids = get_token_ids_from_llm_generator(
+        baseline_llm_generator, prompts, sampling_params)
+
+    print('Getting token ids with lookahead scheduling')
+    test_token_ids = get_token_ids_from_llm_generator(test_llm_generator,
+                                                      prompts, sampling_params)
+
+    for expected_token_ids, actual_token_ids in zip(baseline_token_ids,
+                                                    test_token_ids):
+        assert expected_token_ids == actual_token_ids
+
+    assert baseline_token_ids == test_token_ids
+
+
 def get_token_ids_from_llm_generator(llm_generator, prompts, sampling_params):
     for llm in llm_generator:
         outputs = llm.generate(prompts, sampling_params, use_tqdm=True)

tests/core/block/test_block_manager_v2.py

@@ -0,0 +1,103 @@
+import pytest
+
+from vllm.core.block_manager_v2 import BlockSpaceManagerV2
+from vllm.core.interfaces import AllocStatus
+from vllm.sequence import Logprob, SequenceStatus
+from vllm.utils import chunk_list
+
+from ..utils import create_seq_group
+
+
+@pytest.mark.parametrize("block_size", [16])
+@pytest.mark.parametrize("num_gpu_blocks", [8, 40, 80])
+@pytest.mark.parametrize("num_seqs_per_group", [1, 4])
+@pytest.mark.parametrize("watermark", [0.0, 0.5])
+def test_can_allocate_seq_group(block_size: int, num_seqs_per_group: int,
+                                num_gpu_blocks: int, watermark: float):
+    block_manager = BlockSpaceManagerV2(
+        block_size=block_size,
+        num_gpu_blocks=num_gpu_blocks,
+        num_cpu_blocks=1024,
+        watermark=watermark,
+    )
+    num_watermark_blocks = int(watermark * num_gpu_blocks)
+
+    num_output_blocks_per_seq = 1
+
+    # NOTE: This should be num_output_blocks_per_seq * num_seqs_per_group, but
+    # the current implementation assumes all seqs are new prompts / don't have
+    # different output lens.
+    num_output_blocks = num_output_blocks_per_seq
+
+    for num_prompt_blocks in range(1, num_gpu_blocks - num_output_blocks):
+        seq_group = create_seq_group(
+            seq_prompt_len=block_size * num_prompt_blocks,
+            seq_output_lens=[
+                block_size * num_output_blocks_per_seq
+                for _ in range(num_seqs_per_group)
+            ],
+        )
+
+        assert num_prompt_blocks + num_output_blocks <= num_gpu_blocks
+
+        can_allocate_result = block_manager.can_allocate(seq_group)
+
+        num_required_blocks = num_prompt_blocks + num_output_blocks
+
+        if num_gpu_blocks - num_required_blocks < num_watermark_blocks:
+            assert can_allocate_result == AllocStatus.NEVER
+        elif num_gpu_blocks >= num_required_blocks:
+            assert can_allocate_result == AllocStatus.OK
+        else:
+            assert can_allocate_result == AllocStatus.LATER
+
+
+@pytest.mark.parametrize("block_size", [1, 8])
+@pytest.mark.parametrize("prompt_len", [1, 7, 8])
+@pytest.mark.parametrize("num_slots_to_append", [1, 8, 129])
+@pytest.mark.parametrize("num_lookahead_slots", [0, 10])
+def test_append_slots(block_size, prompt_len, num_slots_to_append,
+                      num_lookahead_slots):
+    """Verify append_slots consumes the correct number of blocks from the block
+    table.
+    """
+
+    num_gpu_blocks = 1024
+    watermark = 0.1
+    block_manager = BlockSpaceManagerV2(
+        block_size=block_size,
+        num_gpu_blocks=num_gpu_blocks,
+        num_cpu_blocks=0,
+        watermark=watermark,
+    )
+
+    seq_group = create_seq_group(
+        seq_prompt_len=prompt_len,
+        seq_output_lens=[0],
+    )
+
+    # Allocate seq
+    assert block_manager.can_allocate(seq_group)
+    block_manager.allocate(seq_group)
+
+    # Seq seq to RUNNING
+    seq = seq_group.get_seqs()[0]
+    seq.status = SequenceStatus.RUNNING
+
+    # Append tokens to the sequeqnce
+    for token_id in range(num_slots_to_append):
+        seq.append_token_id(token_id, {token_id: Logprob(0.0)})
+
+    # Append slots for new tokens and lookahead slots.
+    free_blocks_before_append = block_manager.get_num_free_gpu_blocks()
+    block_manager.append_slots(seq, num_lookahead_slots)
+    num_consumed_blocks = (free_blocks_before_append -
+                           block_manager.get_num_free_gpu_blocks())
+
+    # Expect consumed blocks to be new blocks required to support the new slots.
+    expected_consumed_blocks = len(
+        chunk_list(
+            list(
+                range(prompt_len + num_slots_to_append + num_lookahead_slots)),
+            block_size)) - len(chunk_list(list(range(prompt_len)), block_size))
+    assert num_consumed_blocks == expected_consumed_blocks

tests/core/block/test_block_space_manager.py

@@ -1,50 +0,0 @@
-import pytest
-
-from vllm.core.block_manager_v2 import BlockSpaceManagerV2
-from vllm.core.interfaces import AllocStatus
-
-from ..utils import create_seq_group
-
-
-@pytest.mark.parametrize("block_size", [16])
-@pytest.mark.parametrize("num_gpu_blocks", [8, 40, 80])
-@pytest.mark.parametrize("num_seqs_per_group", [1, 4])
-@pytest.mark.parametrize("watermark", [0.0, 0.5])
-def test_can_allocate_seq_group(block_size: int, num_seqs_per_group: int,
-                                num_gpu_blocks: int, watermark: float):
-    block_manager = BlockSpaceManagerV2(
-        block_size=block_size,
-        num_gpu_blocks=num_gpu_blocks,
-        num_cpu_blocks=1024,
-        watermark=watermark,
-    )
-    num_watermark_blocks = int(watermark * num_gpu_blocks)
-
-    num_output_blocks_per_seq = 1
-
-    # NOTE: This should be num_output_blocks_per_seq * num_seqs_per_group, but
-    # the current implementation assumes all seqs are new prompts / don't have
-    # different output lens.
-    num_output_blocks = num_output_blocks_per_seq
-
-    for num_prompt_blocks in range(1, num_gpu_blocks - num_output_blocks):
-        seq_group = create_seq_group(
-            seq_prompt_lens=block_size * num_prompt_blocks,
-            seq_output_lens=[
-                block_size * num_output_blocks_per_seq
-                for _ in range(num_seqs_per_group)
-            ],
-        )
-
-        assert num_prompt_blocks + num_output_blocks <= num_gpu_blocks
-
-        can_allocate_result = block_manager.can_allocate(seq_group)
-
-        num_required_blocks = num_prompt_blocks + num_output_blocks
-
-        if num_gpu_blocks - num_required_blocks < num_watermark_blocks:
-            assert can_allocate_result == AllocStatus.NEVER
-        elif num_gpu_blocks >= num_required_blocks:
-            assert can_allocate_result == AllocStatus.OK
-        else:
-            assert can_allocate_result == AllocStatus.LATER

tests/core/block/test_block_table.py

@@ -498,3 +498,78 @@ def test_cow_lookahead_simple(block_size: int, sequence_len: int,
 
     # After free, expect all blocks to be freed.
     assert allocator.get_num_free_blocks(Device.GPU) == num_gpu_blocks
+
+
+@pytest.mark.parametrize("block_size", [1, 8])
+@pytest.mark.parametrize("sequence_len", [1, 16, 129])
+@pytest.mark.parametrize("num_new_tokens", [1, 16, 129])
+@pytest.mark.parametrize("num_lookahead_slots", [1, 7, 8])
+@pytest.mark.parametrize("allocator_type", ["naive", "prefix_caching"])
+def test_num_blocks_touched_by_append_slots(block_size: int, sequence_len: int,
+                                            num_new_tokens: int,
+                                            num_lookahead_slots: int,
+                                            allocator_type: str):
+    """Verify correct calculation of get_num_blocks_touched_by_append_slots.
+
+    This is done by using copy-on-write, which requires any modified block to
+    be copied before write if the refcount > 1. We set the refcount>1 by forking
+    a sequence, then measure the free blocks before and after an append. If the
+    number of consumed blocks equals what `get_num_blocks_touched_by_append_
+    slots` returns, then the calculation is correct.
+    """
+
+    num_gpu_blocks = 1024
+
+    allocator = CpuGpuBlockAllocator.create(
+        allocator_type=allocator_type,
+        num_gpu_blocks=num_gpu_blocks,
+        num_cpu_blocks=0,
+        block_size=block_size,
+    )
+
+    token_ids = list(range(sequence_len))
+    token_ids_to_append = list(range(num_new_tokens))
+
+    block_table = BlockTable(
+        block_size=block_size,
+        block_allocator=allocator,
+    )
+
+    block_table.allocate(token_ids=token_ids, device=Device.GPU)
+
+    # Add lookahead before fork so both sequences have the same lookahead
+    # blocks.
+    block_table.ensure_num_empty_slots(num_empty_slots=num_lookahead_slots)
+
+    # Fork sequence so that every block has refcount > 1.
+    _ = block_table.fork()
+
+    # Determine how many blocks should be touched.
+    expected_num_touched_blocks = (
+        block_table.get_num_blocks_touched_by_append_slots(
+            token_ids=token_ids_to_append,
+            num_lookahead_slots=num_lookahead_slots))
+
+    # Measure how many blocks are touched by measuring num_free_blocks before
+    # and after the append.
+    #
+    # We expect append_token_ids to CoW all mutated blocks that have refcount>1.
+    num_free_blocks_before_append = allocator.get_num_free_blocks(Device.GPU)
+    block_table.append_token_ids(token_ids_to_append, num_lookahead_slots)
+    num_consumed_blocks = (num_free_blocks_before_append -
+                           allocator.get_num_free_blocks(Device.GPU))
+
+    # TODO(cade) ensure equality when num_lookahead_slots > 0.
+    # The reason we have < is because lookahead blocks are not copied eagerly;
+    # they are copied on first write. This will cause issues for beam search +
+    # speculative decoding. This is acceptable for now as it is a large effort
+    # to combine the two. To fix this, we can ensure single sequence ownership
+    # of lookahead blocks by appending empty slots to each block, which will
+    # trigger the CoW.
+    #
+    # Until then, we can accept that the consumed tokens are <= the expected
+    # tokens when appending with lookahead.
+    if num_lookahead_slots > 0:
+        assert num_consumed_blocks <= expected_num_touched_blocks
+    else:
+        assert num_consumed_blocks == expected_num_touched_blocks

tests/core/test_block_manager.py

@@ -103,9 +103,9 @@ def test_append_slot_single_seq():
     block_manager.allocate(seq_group)
 
     # Nothing to append. Sequence has no new logical blocks.
-    assert block_manager.can_append_slot(seq_group)
+    assert block_manager.can_append_slots(seq_group)
     before_blocks = block_manager.get_num_free_gpu_blocks()
-    assert not block_manager.append_slot(prompt)
+    assert not block_manager.append_slots(prompt)
     after_blocks = block_manager.get_num_free_gpu_blocks()
     assert before_blocks == after_blocks
 
@@ -114,9 +114,9 @@ def test_append_slot_single_seq():
         token_id = i + 5
         prompt.append_token_id(token_id, {token_id: Logprob(0.0)})
 
-    assert block_manager.can_append_slot(seq_group)
+    assert block_manager.can_append_slots(seq_group)
     before_blocks = block_manager.get_num_free_gpu_blocks()
-    assert not block_manager.append_slot(prompt)
+    assert not block_manager.append_slots(prompt)
     after_blocks = block_manager.get_num_free_gpu_blocks()
     assert before_blocks - after_blocks == 1
 
@@ -150,13 +150,13 @@ def test_append_slot_cow():
     child.append_token_id(token_id, {token_id: Logprob(0.0)})
     block_manager.fork(prompt, child)
 
-    assert block_manager.can_append_slot(seq_group)
+    assert block_manager.can_append_slots(seq_group)
     before_blocks = block_manager.get_num_free_gpu_blocks()
 
-    maybe_src_dst_block = block_manager.append_slot(child)
-    assert maybe_src_dst_block is not None
-    src_block, dst_block = maybe_src_dst_block
-    assert src_block != dst_block
+    cows = block_manager.append_slots(child)
+    assert cows
+    for src_block, dst_blocks in cows.items():
+        assert src_block not in dst_blocks
 
     after_blocks = block_manager.get_num_free_gpu_blocks()
     assert before_blocks - after_blocks == 1
@@ -184,7 +184,7 @@ def test_fork():
     token_id = 4
     # Append token to child. Block is shared so copy on write occurs.
     child.append_token_id(token_id, {token_id: Logprob(0.0)})
-    block_manager.append_slot(child)
+    block_manager.append_slots(child)
     assert block_manager.get_block_table(
         prompt) != block_manager.get_block_table(child)
 
@@ -325,7 +325,7 @@ def test_sliding_window_multi_seq():
     token_id = 4
     # Append token to child. Block is shared so copy on write occurs.
     child.append_token_id(token_id, {token_id: Logprob(0.0)})
-    block_manager.append_slot(child)
+    block_manager.append_slots(child)
 
     # assert the number of blocks allocated is correct
     # we will use now one block more. Each seq will use 2 blocks,
@@ -335,7 +335,7 @@ def test_sliding_window_multi_seq():
 
     token_id = 5
     parent.append_token_id(token_id, {token_id: Logprob(0.0)})
-    block_manager.append_slot(parent)
+    block_manager.append_slots(parent)
 
     # assert the number of blocks allocated is correct
     # no change, because both sequences are still just sharing one block

tests/core/test_chunked_prefill_scheduler.py

@@ -0,0 +1,563 @@
+from typing import List
+from unittest.mock import MagicMock
+
+import pytest  # noqa
+
+from vllm.config import CacheConfig, SchedulerConfig
+from vllm.core.scheduler import Scheduler
+from vllm.sequence import Logprob, SequenceGroup
+
+from .utils import create_dummy_prompt
+
+
+def get_sequence_groups(scheduler_output):
+    return [s.seq_group for s in scheduler_output.scheduled_seq_groups]
+
+
+def append_new_token(seq_group, token_id: int):
+    for seq in seq_group.get_seqs():
+        seq.append_token_id(token_id, {token_id: Logprob(token_id)})
+
+
+def schedule_and_update_computed_tokens(scheduler):
+    metas, out = scheduler.schedule()
+    for s, meta in zip(out.scheduled_seq_groups, metas):
+        s.seq_group.update_num_computed_tokens(meta.token_chunk_size)
+    return metas, out
+
+
+def test_simple():
+    """Verify basic scheduling works."""
+    block_size = 4
+    num_seq_group = 4
+    max_model_len = 16
+    max_num_batched_tokens = 64
+    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+                                       num_seq_group,
+                                       max_model_len,
+                                       enable_chunked_prefill=True)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+    running: List[SequenceGroup] = []
+
+    # Add seq groups to scheduler.
+    for i in range(num_seq_group):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=block_size)
+        scheduler.add_seq_group(seq_group)
+        running.append(seq_group)
+
+    # Schedule seq groups prompts.
+    num_tokens = block_size * num_seq_group
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert set(get_sequence_groups(out)) == set(running)
+    assert out.num_batched_tokens == num_tokens
+    assert (not out.blocks_to_copy and not out.blocks_to_swap_in
+            and not out.blocks_to_swap_out)
+    assert len(seq_group_meta) == num_seq_group
+    for s in running:
+        append_new_token(s, 1)
+
+    # Schedule seq groups generation.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert set(get_sequence_groups(out)) == set(running)
+    assert out.num_batched_tokens == num_seq_group
+    assert (not out.blocks_to_copy and not out.blocks_to_swap_in
+            and not out.blocks_to_swap_out)
+    assert len(seq_group_meta) == num_seq_group
+
+
+def test_chunk():
+    """Verify prefills are chunked properly."""
+    block_size = 4
+    max_seqs = 60
+    max_model_len = 80
+    max_num_batched_tokens = 64
+    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+                                       max_seqs,
+                                       max_model_len,
+                                       enable_chunked_prefill=True)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+    running: List[SequenceGroup] = []
+
+    # Add seq groups to scheduler.
+    for i in range(2):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+        scheduler.add_seq_group(seq_group)
+        running.append(seq_group)
+
+    # Verify the second request is chunked.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert set(get_sequence_groups(out)) == set(running)
+    assert seq_group_meta[0].token_chunk_size == 60
+    # Verify it is chunked.
+    assert seq_group_meta[1].token_chunk_size == 4
+    assert out.num_prefill_groups == 2
+    assert out.num_batched_tokens == 64
+    # Only the first seq group has a new token appended.
+    append_new_token(running[0], 1)
+
+    # One chunked prefill, and one decoding.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert set(get_sequence_groups(out)) == set(running)
+    # The first one is decoding.
+    assert seq_group_meta[0].token_chunk_size == 1
+    # The second one is a chunked prefill.
+    assert seq_group_meta[1].token_chunk_size == 56
+    assert out.num_prefill_groups == 1
+    assert out.num_batched_tokens == 57
+
+
+def test_complex():
+    block_size = 4
+    max_seqs = 60
+    max_model_len = 80
+    max_num_batched_tokens = 64
+    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+                                       max_seqs,
+                                       max_model_len,
+                                       enable_chunked_prefill=True)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+    running: List[SequenceGroup] = []
+
+    # Add seq groups to scheduler.
+    for i in range(2):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+        scheduler.add_seq_group(seq_group)
+        running.append(seq_group)
+        assert seq_group.is_prefill()
+
+    # Verify the second request is chunked.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+
+    assert set(get_sequence_groups(out)) == set(running)
+    assert seq_group_meta[0].token_chunk_size == 60
+    # Verify it is chunked.
+    assert seq_group_meta[1].token_chunk_size == 4
+    assert not running[0].is_prefill()
+    assert running[1].is_prefill()
+    assert out.num_prefill_groups == 2
+    assert out.num_batched_tokens == 64
+    # Only the first seq group has a new token appended.
+    append_new_token(running[0], 1)
+
+    # Add 2 more requsets.
+    for i in range(2, 4):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+        scheduler.add_seq_group(seq_group)
+        running.append(seq_group)
+
+    # Decoding & chunked prefill & first chunk of 3rd request is scheduled.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(get_sequence_groups(out)) == 3
+    # The first one is decoding.
+    assert seq_group_meta[0].token_chunk_size == 1
+    # The second one is a chunked prefill.
+    assert seq_group_meta[1].token_chunk_size == 56
+    # The third one is also chunked.
+    assert seq_group_meta[2].token_chunk_size == 7
+    # Two of them are in chunked prefill.
+    assert out.num_prefill_groups == 2
+    assert out.num_batched_tokens == 64
+    # The first 2 requests are now in decodine phase.
+    append_new_token(running[0], 1)
+    assert not running[0].is_prefill()
+    append_new_token(running[1], 1)
+    assert not running[1].is_prefill()
+    # The third request is still in prefill stage.
+    assert running[2].is_prefill()
+
+
+def test_maximal_decoding():
+    """Verify decoding requests are prioritized."""
+    block_size = 4
+    max_seqs = 2
+    max_model_len = 2
+    max_num_batched_tokens = 2
+    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+                                       max_seqs,
+                                       max_model_len,
+                                       enable_chunked_prefill=True)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+    running: List[SequenceGroup] = []
+
+    # Add seq groups to scheduler.
+    for i in range(2):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=2)
+        scheduler.add_seq_group(seq_group)
+        running.append(seq_group)
+        assert seq_group.is_prefill()
+
+    # The first prefill is scheduled.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(get_sequence_groups(out)) == 1
+    assert seq_group_meta[0].token_chunk_size == 2
+    assert not running[0].is_prefill()
+    assert running[1].is_prefill()
+    assert out.num_prefill_groups == 1
+    assert out.num_batched_tokens == 2
+    # Only the first seq group has a new token appended.
+    append_new_token(running[0], 1)
+
+    # Create one more seq_group.
+    _, seq_group = create_dummy_prompt("3", prompt_length=2)
+    scheduler.add_seq_group(seq_group)
+    running.append(seq_group)
+    assert seq_group.is_prefill()
+    # The first decoding + second chunk is scheduled.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(get_sequence_groups(out)) == 2
+    assert seq_group_meta[0].token_chunk_size == 1
+    assert seq_group_meta[1].token_chunk_size == 1
+    assert not running[0].is_prefill()
+    assert running[1].is_prefill()
+    assert running[2].is_prefill()
+    assert out.num_prefill_groups == 1
+    assert out.num_batched_tokens == 2
+    append_new_token(running[0], 1)
+
+    # Decoding + running prefill is prioritized.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(get_sequence_groups(out)) == 2
+    assert seq_group_meta[0].token_chunk_size == 1
+    assert seq_group_meta[1].token_chunk_size == 1
+    assert not running[0].is_prefill()
+    assert not running[1].is_prefill()
+    assert out.num_prefill_groups == 1
+    assert out.num_batched_tokens == 2
+    append_new_token(running[0], 1)
+    append_new_token(running[1], 1)
+
+    # Only decoding is prioritized.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(get_sequence_groups(out)) == 2
+    assert seq_group_meta[0].token_chunk_size == 1
+    assert seq_group_meta[1].token_chunk_size == 1
+    assert not running[0].is_prefill()
+    assert not running[1].is_prefill()
+    assert out.num_prefill_groups == 0
+    assert out.num_batched_tokens == 2
+    append_new_token(running[0], 1)
+    append_new_token(running[1], 1)
+
+    # After aborting the decoding request, the fcfs new prefill is prioritized.
+    scheduler.abort_seq_group(running[0].request_id)
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(get_sequence_groups(out)) == 2
+    assert seq_group_meta[0].token_chunk_size == 1
+    assert seq_group_meta[1].token_chunk_size == 1
+    assert not running[1].is_prefill()
+    assert running[2].is_prefill()
+    assert out.num_prefill_groups == 1
+    assert out.num_batched_tokens == 2
+
+
+def test_prompt_limit():
+    """Verify max_num_batched_tokens < max_model_len is possible."""
+    block_size = 4
+    max_seqs = 32
+    max_model_len = 64
+    max_num_batched_tokens = 32
+    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+                                       max_seqs,
+                                       max_model_len,
+                                       enable_chunked_prefill=True)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+    running: List[SequenceGroup] = []
+
+    _, seq_group = create_dummy_prompt("1", prompt_length=48)
+    scheduler.add_seq_group(seq_group)
+    running.append(seq_group)
+    assert seq_group.is_prefill()
+
+    # The prompt length > max_num_batched_tokens should be still scheduled.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(get_sequence_groups(out)) == 1
+    assert seq_group_meta[0].token_chunk_size == 32
+    assert running[0].is_prefill()
+    assert out.num_prefill_groups == 1
+    assert out.num_batched_tokens == 32
+
+
+def test_prompt_limit_exceed():
+    block_size = 4
+    max_seqs = 64
+    max_model_len = 32
+    max_num_batched_tokens = 64
+    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+                                       max_seqs,
+                                       max_model_len,
+                                       enable_chunked_prefill=True)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+    running: List[SequenceGroup] = []
+
+    _, seq_group = create_dummy_prompt("2", prompt_length=48)
+    scheduler.add_seq_group(seq_group)
+    running.append(seq_group)
+    assert seq_group.is_prefill()
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.ignored_seq_groups) == 1
+    assert out.ignored_seq_groups[0] == seq_group
+
+
+def test_swap():
+    """Verify swapping works with chunked prefill requests"""
+    block_size = 4
+    max_seqs = 30
+    max_model_len = 200
+    max_num_batched_tokens = 30
+    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+                                       max_seqs,
+                                       max_model_len,
+                                       enable_chunked_prefill=True)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+
+    _, seq_group = create_dummy_prompt("1", prompt_length=60, best_of=2)
+    scheduler.add_seq_group(seq_group)
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    # The request is chunked.
+    # prefill scheduled now.
+    assert len(out.scheduled_seq_groups) == 1
+    assert out.num_prefill_groups == 1
+    assert seq_group.is_prefill()
+    assert out.num_batched_tokens == max_num_batched_tokens
+
+    # The last request should be swapped out.
+    scheduler.block_manager.can_append_slots = MagicMock()
+
+    def cannot_append_second_group(seq_group, num_lookahead_slots):
+        return seq_group.request_id != "1"
+
+    scheduler.block_manager.can_append_slots.side_effect = (
+        cannot_append_second_group)
+
+    # The running prefill is now swapped.
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 0
+    assert out.num_batched_tokens == 0
+    assert out.blocks_to_swap_out != {}
+    assert out.blocks_to_swap_in == {}
+
+    # Add 1 more task. Swap should be prioritized over new prefill.
+    _, seq_group = create_dummy_prompt("2", prompt_length=60)
+    scheduler.add_seq_group(seq_group)
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 1
+    # 3 decodes. It is swapped in.
+    assert out.num_batched_tokens == 30
+    assert out.blocks_to_swap_in != {}
+    assert out.blocks_to_swap_out == {}
+
+
+def test_running_prefill_prioritized_over_swap():
+    block_size = 4
+    max_seqs = 30
+    max_model_len = 200
+    max_num_batched_tokens = 30
+    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+                                       max_seqs,
+                                       max_model_len,
+                                       enable_chunked_prefill=True)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+
+    _, seq_group = create_dummy_prompt("1", prompt_length=60, best_of=2)
+    scheduler.add_seq_group(seq_group)
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    # The request is chunked.
+    # prefill scheduled now.
+    assert len(out.scheduled_seq_groups) == 1
+    assert out.num_prefill_groups == 1
+    assert seq_group.is_prefill()
+    assert out.num_batched_tokens == max_num_batched_tokens
+
+    # The request should be swapped out.
+    scheduler.block_manager.can_append_slots = MagicMock()
+
+    def cannot_append_second_group(seq_group, num_lookahead_slots):
+        return seq_group.request_id != "1"
+
+    scheduler.block_manager.can_append_slots.side_effect = (
+        cannot_append_second_group)
+
+    # The running prefill is now swapped.
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 0
+    assert out.num_batched_tokens == 0
+    assert out.blocks_to_swap_out != {}
+    assert out.blocks_to_swap_in == {}
+
+    # Add 1 more task. Swap is not possible, so prefill is running.
+    scheduler.block_manager.can_swap_in = MagicMock()
+    scheduler.block_manager.can_swap_in.return_value = False
+
+    _, seq_group2 = create_dummy_prompt("2", prompt_length=60)
+    scheduler.add_seq_group(seq_group2)
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 1
+    # 3 decodes. It is swapped in.
+    assert out.num_batched_tokens == 30
+    assert out.blocks_to_swap_in == {}
+    assert out.blocks_to_swap_out == {}
+    assert out.scheduled_seq_groups[0].seq_group == seq_group2
+
+    # Now although swap is possible, running prefill is prioritized.
+    scheduler.block_manager.can_swap_in.return_value = True
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 1
+    # 3 decodes. It is swapped in.
+    assert out.num_batched_tokens == 30
+    assert out.blocks_to_swap_in == {}
+    assert out.blocks_to_swap_out == {}
+    assert not seq_group2.is_prefill()
+    assert out.scheduled_seq_groups[0].seq_group == seq_group2
+    append_new_token(seq_group2, 1)
+
+    # Decoding is prioritized.
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 1
+    # 3 decodes. It is swapped in.
+    assert out.num_batched_tokens == 1
+    assert out.blocks_to_swap_in == {}
+    assert out.blocks_to_swap_out == {}
+    assert not seq_group2.is_prefill()
+    assert out.scheduled_seq_groups[0].seq_group == seq_group2
+    append_new_token(seq_group2, 1)
+
+    # Since we abort the sequence group, we can finally swap.
+    scheduler.abort_seq_group(seq_group2.request_id)
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 1
+    assert out.num_batched_tokens == 30
+    assert out.blocks_to_swap_in != {}
+    assert out.blocks_to_swap_out == {}
+
+
+def test_chunked_prefill_preempt():
+    """Verify preempt works with chunked prefill requests"""
+    block_size = 4
+    max_seqs = 30
+    max_model_len = 200
+    max_num_batched_tokens = 30
+    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+                                       max_seqs,
+                                       max_model_len,
+                                       enable_chunked_prefill=True)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+
+    _, seq_group = create_dummy_prompt("1", prompt_length=60)
+    scheduler.add_seq_group(seq_group)
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    # The request is chunked.
+    # prefill scheduled now.
+    assert len(out.scheduled_seq_groups) == 1
+    assert out.num_prefill_groups == 1
+    assert seq_group.is_prefill()
+    assert out.num_batched_tokens == max_num_batched_tokens
+
+    # The request should be preempted.
+    scheduler.block_manager.can_append_slots = MagicMock()
+
+    def cannot_append_second_group(seq_group, num_lookahead_slots):
+        return seq_group.request_id != "1"
+
+    scheduler.block_manager.can_append_slots.side_effect = (
+        cannot_append_second_group)
+
+    # The running prefill is now preempted.
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 0
+    assert out.num_batched_tokens == 0
+    assert out.blocks_to_swap_out == {}
+    assert out.blocks_to_swap_in == {}
+
+    # Make sure we can reschedule preempted request.
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 1
+    assert out.num_prefill_groups == 1
+    assert seq_group.is_prefill()
+    assert out.num_batched_tokens == max_num_batched_tokens
+    assert seq_group.get_num_uncomputed_tokens() == 30
+
+    # We should be able to run prefill twice as it is chunked.
+    def cannot_append_second_group(seq_group, num_lookahead_slots):
+        return True
+
+    scheduler.block_manager.can_append_slots.side_effect = (
+        cannot_append_second_group)
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 1
+    assert out.num_prefill_groups == 1
+    assert not seq_group.is_prefill()
+    assert out.num_batched_tokens == max_num_batched_tokens
+
+
+def test_chunked_prefill_max_seqs():
+    block_size = 4
+    max_seqs = 2
+    max_model_len = 80
+    max_num_batched_tokens = 64
+    scheduler_config = SchedulerConfig(max_num_batched_tokens,
+                                       max_seqs,
+                                       max_model_len,
+                                       enable_chunked_prefill=True)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+    running = []
+
+    _, seq_group = create_dummy_prompt("1", prompt_length=65)
+    scheduler.add_seq_group(seq_group)
+    running.append(seq_group)
+    # The first prefill is chunked.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert seq_group_meta[0].token_chunk_size == max_num_batched_tokens
+    assert len(get_sequence_groups(out)) == 1
+
+    # Add new requests.
+    for i in range(4):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=65)
+        scheduler.add_seq_group(seq_group)
+        running.append(seq_group)
+
+    # Make sure only 2 requests are scheduled.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert out.num_batched_tokens == max_num_batched_tokens
+    assert len(get_sequence_groups(out)) == 2
+    assert not running[0].is_prefill()
+    assert running[1].is_prefill()
+    append_new_token(running[0], 1)
+
+    # Although we have enough token budget, we can only schedule max_seqs.
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert seq_group_meta[0].token_chunk_size == 2
+    assert seq_group_meta[1].token_chunk_size == 1
+    assert out.num_batched_tokens == 3
+    assert len(get_sequence_groups(out)) == max_seqs
+    assert not running[0].is_prefill()
+    assert not running[1].is_prefill()

tests/core/test_scheduler.py

@@ -1,11 +1,16 @@
 import time
+from collections import deque
 from typing import List
+from unittest.mock import MagicMock
 
 import pytest  # noqa
 
-from vllm.config import CacheConfig, SchedulerConfig
-from vllm.core.scheduler import Scheduler
-from vllm.sequence import Logprob, SequenceGroup
+from vllm.config import CacheConfig, LoRAConfig, SchedulerConfig
+from vllm.core.interfaces import AllocStatus
+from vllm.core.policy import PolicyFactory
+from vllm.core.scheduler import Scheduler, SchedulingBudget
+from vllm.lora.request import LoRARequest
+from vllm.sequence import Logprob, SequenceGroup, SequenceStatus
 
 from .utils import create_dummy_prompt
 
@@ -14,6 +19,26 @@ def get_sequence_groups(scheduler_output):
     return [s.seq_group for s in scheduler_output.scheduled_seq_groups]
 
 
+def append_new_token(out, token_id: int):
+    seq_groups = get_sequence_groups(out)
+    for seq_group in seq_groups:
+        for seq in seq_group.get_seqs():
+            seq.append_token_id(token_id, {token_id: Logprob(token_id)})
+
+
+def schedule_and_update_computed_tokens(scheduler):
+    metas, out = scheduler.schedule()
+    for s, meta in zip(out.scheduled_seq_groups, metas):
+        s.seq_group.update_num_computed_tokens(meta.token_chunk_size)
+    return metas, out
+
+
+def append_new_token_seq_group(token_chunk_size, seq_group, token_id: int):
+    seq_group.update_num_computed_tokens(token_chunk_size)
+    for seq in seq_group.get_seqs():
+        seq.append_token_id(token_id, {token_id: Logprob(token_id)})
+
+
 def test_scheduler_add_seq_group():
     block_size = 4
     scheduler_config = SchedulerConfig(100, 64, 1)
@@ -71,20 +96,52 @@ def test_scheduler_schedule_simple():
 
     # Schedule seq groups prompts.
     num_tokens = block_size * num_seq_group
-    seq_group_meta, out = scheduler.schedule()
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
     assert set(get_sequence_groups(out)) == set(running)
     assert out.num_batched_tokens == num_tokens
     assert (not out.blocks_to_copy and not out.blocks_to_swap_in
             and not out.blocks_to_swap_out)
     assert len(seq_group_meta) == num_seq_group
+    append_new_token(out, 1)
 
     # Schedule seq groups generation.
-    seq_group_meta, out = scheduler.schedule()
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
     assert set(get_sequence_groups(out)) == set(running)
     assert out.num_batched_tokens == num_seq_group
     assert (not out.blocks_to_copy and not out.blocks_to_swap_in
             and not out.blocks_to_swap_out)
     assert len(seq_group_meta) == num_seq_group
+    append_new_token(out, 1)
+
+
+def test_scheduler_prefill_prioritized():
+    """Verify running batched tokens are not applied to prefill requests."""
+    block_size = 4
+    max_model_len = 30
+    max_batched_num_tokens = 30
+    scheduler_config = SchedulerConfig(max_batched_num_tokens, 2,
+                                       max_model_len)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 2
+    cache_config.num_gpu_blocks = 2
+    scheduler = Scheduler(scheduler_config, cache_config, None)
+
+    # Add seq groups to scheduler.
+    _, seq_group_a = create_dummy_prompt("1", 1)
+    scheduler.add_seq_group(seq_group_a)
+
+    # Schedule seq groups prompts.
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert get_sequence_groups(out) == [seq_group_a]
+
+    # Add a new prefill request B.
+    _, seq_group_b = create_dummy_prompt("2", 30)
+    scheduler.add_seq_group(seq_group_b)
+
+    # Verify prefill requests are prioritized. Since max_batched_num_tokens
+    # is 1, new prefill request has to be scheduled first.
+    _, out = schedule_and_update_computed_tokens(scheduler)
+    assert get_sequence_groups(out) == [seq_group_b]
 
 
 def test_scheduler_schedule_preempt_abort():
@@ -103,7 +160,7 @@ def test_scheduler_schedule_preempt_abort():
     scheduler.add_seq_group(seq_group_b)
 
     # Schedule seq groups prompts.
-    seq_group_meta, out = scheduler.schedule()
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
     assert get_sequence_groups(out) == [seq_group_a, seq_group_b]
     assert out.num_batched_tokens == block_size * 2  # seq_a and seq_b
     assert (not out.blocks_to_copy and not out.blocks_to_swap_in
@@ -113,12 +170,10 @@ def test_scheduler_schedule_preempt_abort():
 
     # Append "generated" tokens, allowing the sequence to mark prompt tokens as
     # processed.
-    token_id = 0
-    seq_a.append_token_id(token_id, {token_id: Logprob(0.0)})
-    seq_b.append_token_id(token_id, {token_id: Logprob(0.0)})
+    append_new_token(out, 1)
 
     # Schedule seq groups generation and preempt seq group b.
-    seq_group_meta, out = scheduler.schedule()
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
     assert get_sequence_groups(out) == [seq_group_a]
     assert out.num_batched_tokens == 1
     assert (not out.blocks_to_copy and not out.blocks_to_swap_in
@@ -128,7 +183,7 @@ def test_scheduler_schedule_preempt_abort():
 
     # Abort seq group a. Re-schedule seq group b prompt with recomputation.
     scheduler.abort_seq_group("1")
-    seq_group_meta, out = scheduler.schedule()
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
     assert get_sequence_groups(out) == [seq_group_b]
     assert out.num_batched_tokens == 5  # 4 prompt + 1 generation.
     assert (not out.blocks_to_copy and not out.blocks_to_swap_in
@@ -158,12 +213,14 @@ def test_scheduler_max_seqs():
     scheduler.add_seq_group(all_seq_groups[0])
 
     # Schedule seq groups prompts.
-    _, out = scheduler.schedule()
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
     assert set(get_sequence_groups(out)) == set([all_seq_groups[0]])
+    append_new_token(out, 1)
 
     # Schedule seq groups generation.
-    _, out = scheduler.schedule()
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
     assert set(get_sequence_groups(out)) == set([all_seq_groups[0]])
+    append_new_token(out, 1)
 
     # Append 2 more seq group
     scheduler.add_seq_group(all_seq_groups[1])
@@ -172,12 +229,11 @@ def test_scheduler_max_seqs():
     # Schedule seq groups prompts.
     # Only 1 seq group should be scheduled since max_seq_group is 2
     # and one is prompting.
-    _, out = scheduler.schedule()
+    _, out = schedule_and_update_computed_tokens(scheduler)
     assert set(get_sequence_groups(out)) == set([all_seq_groups[1]])
 
 
 def test_scheduler_delay_factor():
-
     block_size = 4
     scheduler_config = SchedulerConfig(100, 64, 16, delay_factor=0.5)
     cache_config = CacheConfig(block_size, 1.0, 1, "auto")
@@ -186,24 +242,630 @@ def test_scheduler_delay_factor():
     scheduler = Scheduler(scheduler_config, cache_config, None)
 
     # schedule first prompt
-    _, seq_group = create_dummy_prompt("0", prompt_length=block_size)
+    seq_group_meta, seq_group = create_dummy_prompt("0",
+                                                    prompt_length=block_size)
     scheduler.add_seq_group(seq_group)
-    seq_group_meta, out = scheduler.schedule()
-    assert out.prompt_run
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert out.num_prefill_groups > 0
     assert seq_group_meta[0].request_id == '0'
+    append_new_token(out, 1)
 
     # wait for a second before scheduling next prompt
     time.sleep(1)
-    _, seq_group = create_dummy_prompt("1", prompt_length=block_size)
+    seq_group_meta, seq_group = create_dummy_prompt("1",
+                                                    prompt_length=block_size)
     scheduler.add_seq_group(seq_group)
 
     # second prompt should *not* be scheduled
-    seq_group_meta, out = scheduler.schedule()
-    assert not out.prompt_run
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert out.num_prefill_groups == 0
     assert seq_group_meta[0].request_id == '0'
+    append_new_token(out, 1)
 
     # wait for more than 0.5 second and try again
     time.sleep(0.6)
-    seq_group_meta, out = scheduler.schedule()
-    assert out.prompt_run
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert out.num_prefill_groups > 0
     assert seq_group_meta[0].request_id == '1'
+    append_new_token(out, 1)
+
+
+def test_swapped_out_prioritized():
+    scheduler = initialize_scheduler(max_num_seqs=6)
+    # best_of=2 * 3 == 6 sequences.
+    for i in range(3):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60, best_of=2)
+        scheduler.add_seq_group(seq_group)
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    # prefill scheduled now.
+    assert len(out.scheduled_seq_groups) == 3
+    append_new_token(out, 1)
+
+    # The last request should be swapped out.
+    scheduler.block_manager.can_append_slots = MagicMock()
+
+    def cannot_append_second_group(seq_group, num_lookahead_slots):
+        return seq_group.request_id != "2"
+
+    scheduler.block_manager.can_append_slots.side_effect = (
+        cannot_append_second_group)
+
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    assert len(out.scheduled_seq_groups) == 2
+    assert out.num_batched_tokens == 2
+    assert out.blocks_to_swap_out != {}
+    assert out.blocks_to_swap_in == {}
+    append_new_token(out, 1)
+
+    # Add 1 more task. Swap should be prioritized over prefill.
+    _, seq_group = create_dummy_prompt(str(i), prompt_length=60, best_of=2)
+    scheduler.add_seq_group(seq_group)
+    seq_group_meta, out = schedule_and_update_computed_tokens(scheduler)
+    append_new_token(out, 1)
+    assert len(out.scheduled_seq_groups) == 3
+    # 3 decodes. It is swapped in.
+    assert out.num_batched_tokens == 3
+    assert out.blocks_to_swap_in != {}
+    assert out.blocks_to_swap_out == {}
+
+
+def initialize_scheduler(*,
+                         max_num_seqs=1000,
+                         max_token_budget=1000,
+                         max_model_len=1000,
+                         lora_config=None):
+    block_size = 4
+    scheduler_config = SchedulerConfig(max_token_budget, max_num_seqs,
+                                       max_model_len)
+    cache_config = CacheConfig(block_size, 1.0, 1, "auto")
+    cache_config.num_cpu_blocks = 8
+    cache_config.num_gpu_blocks = 8
+    scheduler = Scheduler(scheduler_config, cache_config, lora_config)
+    return scheduler
+
+
+def create_token_budget(token_budget: int = 10000,
+                        max_num_seqs: int = 10000) -> SchedulingBudget:
+    return SchedulingBudget(
+        token_budget=token_budget,
+        max_num_seqs=max_num_seqs,
+    )
+
+
+def add_token_budget(budget: SchedulingBudget,
+                     num_batched_tokens: int = 0,
+                     num_curr_seqs: int = 0):
+    mock_seq_group = create_dummy_prompt('10', prompt_length=60)[1]
+    budget.add_num_batched_tokens(mock_seq_group.request_id,
+                                  num_batched_tokens)
+    budget.add_num_seqs(mock_seq_group.request_id, num_curr_seqs)
+
+
+def test_prefill_schedule_max_prompt_len():
+    """
+    Test prompt longer than max_prompt_len is aborted.
+    """
+    scheduler = initialize_scheduler(max_model_len=30)
+    _, seq_group = create_dummy_prompt(0, prompt_length=60)
+    waiting = deque([seq_group])
+    budget = create_token_budget()
+    remaining_waiting, output = scheduler._schedule_prefills(
+        waiting, budget, None)
+    assert len(output.ignored_seq_groups) == 1
+    assert len(output.seq_groups) == 0
+    assert budget.num_batched_tokens == 0
+    assert budget.num_curr_seqs == 0
+    assert len(remaining_waiting) == 0
+
+
+def test_prefill_schedule_token_budget():
+    """
+    Test token budget respected.
+    """
+    scheduler = initialize_scheduler()
+    waiting = deque()
+    budget = create_token_budget(token_budget=0)
+    for i in range(2):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+        waiting.append(seq_group)
+
+    # 0 token budget == nothing is scheduled.
+    remaining_waiting, output = scheduler._schedule_prefills(
+        waiting, budget, None)
+    assert len(output.ignored_seq_groups) == 0
+    assert len(output.seq_groups) == 0
+    assert budget.num_batched_tokens == 0
+    assert budget.num_curr_seqs == 0
+    assert len(remaining_waiting) == 2
+
+    # 60 token budget == 1 request scheduled.
+    budget = create_token_budget(token_budget=60)
+    remaining_waiting, output = scheduler._schedule_prefills(
+        waiting, budget, None)
+    assert len(output.ignored_seq_groups) == 0
+    assert len(output.seq_groups) == 1
+    assert budget.num_batched_tokens == 60
+    assert budget.num_curr_seqs == 1
+    assert len(remaining_waiting) == 1
+
+    # Test when current_batched_tokens respected.
+    scheduler = initialize_scheduler()
+    waiting = deque()
+    budget = create_token_budget(token_budget=60)
+    add_token_budget(budget, 30, 0)
+    _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+    # Cannot schedule a prompt that doesn't fit the budget.
+    waiting.append(seq_group)
+    remaining_waiting, output = scheduler._schedule_prefills(
+        waiting, budget, None)
+    assert len(output.ignored_seq_groups) == 0
+    assert len(output.seq_groups) == 0
+    assert budget.num_batched_tokens == 30
+    assert budget.num_curr_seqs == 0
+    assert len(remaining_waiting) == 1
+    budget = create_token_budget(token_budget=90)
+    add_token_budget(budget, 30, 0)
+    remaining_waiting, output = scheduler._schedule_prefills(
+        waiting, budget, None)
+    assert len(output.seq_groups) == 1
+    assert budget.num_batched_tokens == 90
+    assert budget.num_curr_seqs == 1
+    assert len(remaining_waiting) == 0
+
+
+def test_prefill_schedule_max_seqs():
+    """
+    Test max seq respected.
+    """
+    scheduler = initialize_scheduler()
+    waiting = deque()
+    budget = create_token_budget(max_num_seqs=2)
+    for i in range(3):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+        waiting.append(seq_group)
+    remaining_waiting, output = scheduler._schedule_prefills(
+        waiting, budget, None)
+    assert len(output.ignored_seq_groups) == 0
+    assert len(output.seq_groups) == 2
+    assert budget.num_batched_tokens == 120
+    assert budget.num_curr_seqs == 2
+    assert len(remaining_waiting) == 1
+
+    # Verify curr_num_seqs respected.
+    waiting = deque()
+    budget = create_token_budget(max_num_seqs=2)
+    add_token_budget(budget, 0, 2)
+    _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+    waiting.append(seq_group)
+    remaining_waiting, output = scheduler._schedule_prefills(
+        waiting, budget, None)
+    assert len(output.ignored_seq_groups) == 0
+    assert len(output.seq_groups) == 0
+    assert budget.num_batched_tokens == 0
+    assert budget.num_curr_seqs == 2
+    assert len(remaining_waiting) == 1
+
+
+def test_prefill_schedule_max_lora():
+    """
+    Test max lora is respected and prioritized.
+    """
+    lora_config = LoRAConfig(max_lora_rank=8, max_loras=1)
+    scheduler = initialize_scheduler(lora_config=lora_config)
+    waiting = deque()
+    budget = create_token_budget(token_budget=120)
+    curr_loras = set()
+    for i in range(2):
+        _, seq_group = create_dummy_prompt(str(i),
+                                           prompt_length=60,
+                                           lora_request=LoRARequest(
+                                               lora_name=str(i),
+                                               lora_int_id=i + 1,
+                                               lora_local_path="abc"))
+        waiting.append(seq_group)
+    # Add two more requests to verify lora is prioritized.
+    # 0: Lora, 1: Lora, 2: regular, 3: regular
+    # In the first iteration, index 0, 2 is scheduled.
+    # If a request is not scheduled because it hits max lora, it is
+    # prioritized. Verify that.
+    for i in range(2, 4):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+        waiting.append(seq_group)
+    # Schedule 2 requests (0 and 2)
+    remaining_waiting, output = scheduler._schedule_prefills(
+        waiting, budget, curr_loras)
+    assert len(output.ignored_seq_groups) == 0
+    assert len(output.seq_groups) == 2
+    assert budget.num_batched_tokens == 120
+    assert budget.num_curr_seqs == 2
+    assert len(remaining_waiting) == 2
+    assert len(curr_loras) == 1
+    # The second lora request is scheduled next as FCFS policy.
+    # Reset curr_loras so that it can be scheduled.
+    curr_loras = set()
+    budget = create_token_budget(token_budget=60)
+    remaining_waiting, output = scheduler._schedule_prefills(
+        remaining_waiting, budget, curr_loras)
+    assert len(output.seq_groups) == 1
+    assert output.seq_groups[0].seq_group.request_id == "1"
+    assert len(remaining_waiting) == 1
+    assert len(curr_loras) == 1
+    assert budget.num_batched_tokens == 60
+
+
+def test_prefill_schedule_no_block_manager_capacity():
+    """
+    Test sequence cannot be scheduled due to block manager has no capacity.
+    """
+    scheduler = initialize_scheduler()
+    waiting = deque()
+    budget = create_token_budget()
+    for i in range(3):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+        waiting.append(seq_group)
+    scheduler.block_manager.can_allocate = MagicMock()
+    scheduler.block_manager.can_allocate.return_value = AllocStatus.LATER
+    remainig_waiting, output = scheduler._schedule_prefills(
+        waiting, budget, None)
+    assert len(output.ignored_seq_groups) == 0
+    assert len(output.seq_groups) == 0
+    assert budget.num_batched_tokens == 0
+    assert budget.num_curr_seqs == 0
+    assert len(remainig_waiting) == 3
+
+    scheduler = initialize_scheduler()
+    waiting = deque()
+    budget = create_token_budget()
+    for i in range(3):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+        waiting.append(seq_group)
+    scheduler.block_manager.can_allocate = MagicMock()
+    scheduler.block_manager.can_allocate.return_value = AllocStatus.NEVER
+    remaining_waiting, output = scheduler._schedule_prefills(
+        waiting, budget, None)
+    assert len(output.ignored_seq_groups) == 3
+    assert len(output.seq_groups) == 0
+    assert budget.num_batched_tokens == 0
+    assert budget.num_curr_seqs == 0
+    assert len(remaining_waiting) == 0
+
+
+def test_decode_schedule_preempted():
+    """
+    Test decodes cannot be scheduled and preempted.
+    """
+    scheduler = initialize_scheduler()
+    running = deque()
+    policy = PolicyFactory.get_policy(policy_name="fcfs")
+    curr_loras = None
+    for i in range(3):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+        scheduler._allocate_and_set_running(seq_group, 60)
+        append_new_token_seq_group(60, seq_group, 1)
+        running.append(seq_group)
+    scheduler.block_manager.can_append_slots = MagicMock()
+
+    def cannot_append_second_group(seq_group, num_lookahead_slots):
+        return seq_group.request_id != "1"
+
+    scheduler.block_manager.can_append_slots.side_effect = (
+        cannot_append_second_group)
+
+    # 1 cannot be scheduled, and the lowest priority (request 2)
+    # should be preempted. 1 will also be preempted.
+    budget = create_token_budget()
+    remainig_running, output = scheduler._schedule_running(
+        running, budget, curr_loras, policy)
+    assert len(remainig_running) == 0
+    assert len(output.decode_seq_groups) == 1
+    assert len(output.prefill_seq_groups) == 0
+    assert output.decode_seq_groups[0].seq_group.request_id == "0"
+    assert len(output.preempted) == 2
+    # Verify budgets are updated.
+    assert budget.num_batched_tokens == 1
+    assert budget.num_curr_seqs == 1
+    # Both should be preempted, not swapped.
+    assert output.blocks_to_swap_out == {}
+    # Nothing is copied.
+    assert output.blocks_to_copy == {}
+
+
+def test_decode_swap_beam_search():
+    """
+    Test best_of > 1 swap out blocks
+    """
+    scheduler = initialize_scheduler()
+    running = deque()
+    policy = PolicyFactory.get_policy(policy_name="fcfs")
+    curr_loras = None
+    budget = create_token_budget()
+    for i in range(3):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60, best_of=2)
+        scheduler._allocate_and_set_running(seq_group, 60)
+        running.append(seq_group)
+        append_new_token_seq_group(60, seq_group, 1)
+        budget.add_num_seqs(seq_group.request_id,
+                            seq_group.get_max_num_running_seqs())
+        budget.add_num_batched_tokens(
+            seq_group.request_id, seq_group.num_seqs(SequenceStatus.RUNNING))
+
+    # The last request should be swapped out.
+    scheduler.block_manager.can_append_slots = MagicMock()
+
+    def cannot_append_second_group(seq_group, num_lookahead_slots):
+        return seq_group.request_id != "2"
+
+    scheduler.block_manager.can_append_slots.side_effect = (
+        cannot_append_second_group)
+    scheduler.block_manager.swap_out = MagicMock()
+    expected_swap_mapping = {"5": "7"}
+    scheduler.block_manager.swap_out.return_value = expected_swap_mapping
+
+    remainig_running, output = scheduler._schedule_running(
+        running, budget, curr_loras, policy)
+    assert len(remainig_running) == 0
+    assert len(output.decode_seq_groups) == 2
+    assert len(output.prefill_seq_groups) == 0
+    assert output.decode_seq_groups[0].seq_group.request_id == "0"
+    assert output.decode_seq_groups[1].seq_group.request_id == "1"
+    assert len(output.preempted) == 0
+    assert len(output.swapped_out) == 1
+    # Budget should refledct preempted requests.
+    assert budget.num_batched_tokens == 2
+    # since there are 2 sequences, 2 should be subtracted.
+    assert budget.num_curr_seqs == 4
+    # Both should be preempted, not swapped.
+    assert output.blocks_to_swap_out == expected_swap_mapping
+    # Nothing is copied.
+    assert output.blocks_to_copy == {}
+
+
+def test_schedule_decode_blocks_to_copy_update():
+    """
+    Verify blocks_to_copy is updated.
+    """
+    scheduler = initialize_scheduler()
+    _, seq_group = create_dummy_prompt("1", prompt_length=60, best_of=2)
+    running = deque()
+    policy = PolicyFactory.get_policy(policy_name="fcfs")
+    curr_loras = None
+    scheduler._allocate_and_set_running(seq_group, 60)
+    append_new_token_seq_group(60, seq_group, 1)
+    running.append(seq_group)
+
+    # The last request should be swapped out.
+    scheduler.block_manager.append_slots = MagicMock()
+    scheduler.block_manager.append_slots.return_value = {2: [3]}
+
+    budget = create_token_budget()
+    remaining_running, output = scheduler._schedule_running(
+        running, budget, curr_loras, policy)
+    assert len(remaining_running) == 0
+    assert len(output.decode_seq_groups) == 1
+    assert len(output.prefill_seq_groups) == 0
+    assert len(output.preempted) == 0
+    assert len(output.swapped_out) == 0
+    # Nothing is preempted.
+    assert output.blocks_to_swap_out == {}
+    # Since append_slot returns the source -> dist mapping, it should
+    # applied.
+    assert output.blocks_to_copy == {2: [3]}
+
+
+def test_schedule_swapped_simple():
+    scheduler = initialize_scheduler()
+    swapped = deque()
+    policy = PolicyFactory.get_policy(policy_name="fcfs")
+    curr_loras = None
+    blocks_to_swap_out = {}
+    _, seq_group = create_dummy_prompt("1", prompt_length=60, best_of=2)
+    scheduler._allocate_and_set_running(seq_group, 60)
+    append_new_token_seq_group(60, seq_group, 1)
+    scheduler._swap_out(seq_group, blocks_to_swap_out)
+    swapped.append(seq_group)
+
+    budget = create_token_budget()
+    remaining_swapped, output = scheduler._schedule_swapped(
+        swapped, budget, curr_loras, policy)
+    assert len(remaining_swapped) == 0
+    assert budget.num_batched_tokens == 1
+    assert budget.num_curr_seqs == 2
+    assert len(output.decode_seq_groups) == 1
+    assert len(output.prefill_seq_groups) == 0
+    # swap in is the reverse of swap out
+    blocks_to_swap_in_reverse = {}
+    for swapin, swapout in output.blocks_to_swap_in.items():
+        blocks_to_swap_in_reverse[swapout] = swapin
+    assert blocks_to_swap_out == blocks_to_swap_in_reverse
+
+
+def test_schedule_swapped_max_token_budget():
+    scheduler = initialize_scheduler()
+    swapped = deque()
+    policy = PolicyFactory.get_policy(policy_name="fcfs")
+    curr_loras = None
+    blocks_to_swap_out = {}
+    for _ in range(2):
+        _, seq_group = create_dummy_prompt("1", prompt_length=60, best_of=2)
+        scheduler._allocate_and_set_running(seq_group, 60)
+        append_new_token_seq_group(60, seq_group, 1)
+        scheduler._swap_out(seq_group, blocks_to_swap_out)
+        swapped.append(seq_group)
+
+    budget = create_token_budget(token_budget=1)
+    remaining_swapped, output = scheduler._schedule_swapped(
+        swapped, budget, curr_loras, policy)
+    assert len(remaining_swapped) == 1
+    assert budget.num_batched_tokens == 1
+    assert budget.num_curr_seqs == 2
+    assert len(output.decode_seq_groups) == 1
+    assert len(output.prefill_seq_groups) == 0
+
+    # Verify num_batched_tokens are respected.
+    budget = create_token_budget(token_budget=1)
+    add_token_budget(budget, 1, 0)
+    remaining_swapped, output = scheduler._schedule_swapped(
+        remaining_swapped, budget, curr_loras, policy)
+    assert len(remaining_swapped) == 1
+    assert budget.num_batched_tokens == 1
+    assert budget.num_curr_seqs == 0
+    assert len(output.decode_seq_groups) == 0
+    assert len(output.prefill_seq_groups) == 0
+
+
+def test_schedule_swapped_max_seqs():
+    scheduler = initialize_scheduler()
+    swapped = deque()
+    policy = PolicyFactory.get_policy(policy_name="fcfs")
+    curr_loras = None
+    blocks_to_swap_out = {}
+    for i in range(4):
+        _, seq_group = create_dummy_prompt(str(i), prompt_length=60)
+        scheduler._allocate_and_set_running(seq_group, 60)
+        append_new_token_seq_group(60, seq_group, 1)
+        scheduler._swap_out(seq_group, blocks_to_swap_out)
+        swapped.append(seq_group)
+
+    budget = create_token_budget(max_num_seqs=2)
+    remaining_swapped, output = scheduler._schedule_swapped(
+        swapped, budget, curr_loras, policy)
+    assert len(remaining_swapped) == 2
+    assert budget.num_batched_tokens == 2
+    assert budget.num_curr_seqs == 2
+    assert len(output.decode_seq_groups) == 2
+    assert len(output.prefill_seq_groups) == 0
+
+    # Verify num_curr_seqs are respected.
+    remaining_swapped, output = scheduler._schedule_swapped(
+        remaining_swapped, budget, curr_loras, policy)
+    assert len(remaining_swapped) == 2
+    assert budget.num_batched_tokens == 2
+    assert budget.num_curr_seqs == 2
+    assert len(output.decode_seq_groups) == 0
+    assert len(output.prefill_seq_groups) == 0
+
+
+def test_schedule_swapped_max_loras():
+    lora_config = LoRAConfig(max_lora_rank=8, max_loras=1)
+    scheduler = initialize_scheduler(lora_config=lora_config)
+    swapped = deque()
+    policy = PolicyFactory.get_policy(policy_name="fcfs")
+    curr_loras = set()
+    blocks_to_swap_out = {}
+    for i in range(2):
+        _, seq_group = create_dummy_prompt(str(i),
+                                           prompt_length=60,
+                                           lora_request=LoRARequest(
+                                               lora_name=str(i),
+                                               lora_int_id=i + 1,
+                                               lora_local_path="abc"))
+        scheduler._allocate_and_set_running(seq_group, 60)
+        append_new_token_seq_group(60, seq_group, 1)
+        scheduler._swap_out(seq_group, blocks_to_swap_out)
+        swapped.append(seq_group)
+
+    budget = create_token_budget()
+    remaining_swapped, output = scheduler._schedule_swapped(
+        swapped, budget, curr_loras, policy)
+    assert len(remaining_swapped) == 1
+    assert budget.num_batched_tokens == 1
+    assert budget.num_curr_seqs == 1
+    assert len(output.decode_seq_groups) == 1
+    assert len(output.prefill_seq_groups) == 0
+    assert len(curr_loras) == 1
+
+
+def test_schedule_swapped_cannot_swap_in():
+    scheduler = initialize_scheduler()
+    swapped = deque()
+    policy = PolicyFactory.get_policy(policy_name="fcfs")
+    curr_loras = None
+    blocks_to_swap_out = {}
+    for _ in range(2):
+        _, seq_group = create_dummy_prompt("1", prompt_length=60, best_of=2)
+        scheduler._allocate_and_set_running(seq_group, 60)
+        append_new_token_seq_group(60, seq_group, 1)
+        scheduler._swap_out(seq_group, blocks_to_swap_out)
+        swapped.append(seq_group)
+
+    # The last request should be swapped out.
+    scheduler.block_manager.can_swap_in = MagicMock()
+    scheduler.block_manager.can_swap_in.return_value = False
+    # Since we cannot swap in, none of the requests are swapped in.
+    budget = create_token_budget()
+    remaining_swapped, output = scheduler._schedule_swapped(
+        swapped, budget, curr_loras, policy)
+    assert len(remaining_swapped) == 2
+    assert budget.num_batched_tokens == 0
+    assert budget.num_curr_seqs == 0
+    assert len(output.decode_seq_groups) == 0
+    assert len(output.prefill_seq_groups) == 0
+
+
+def test_schedule_swapped_blocks_to_copy():
+    scheduler = initialize_scheduler()
+    swapped = deque()
+    policy = PolicyFactory.get_policy(policy_name="fcfs")
+    curr_loras = None
+    _, seq_group = create_dummy_prompt("1", prompt_length=60, best_of=2)
+    scheduler._allocate_and_set_running(seq_group, 60)
+    append_new_token_seq_group(60, seq_group, 1)
+    blocks_to_swap_out = {}
+    scheduler._swap_out(seq_group, blocks_to_swap_out)
+    swapped.append(seq_group)
+
+    # The last request should be swapped out.
+    scheduler.block_manager.append_slots = MagicMock()
+    scheduler.block_manager.append_slots.return_value = {2: [3]}
+
+    budget = create_token_budget()
+    remaining_swapped, output = scheduler._schedule_swapped(
+        swapped, budget, curr_loras, policy)
+    assert len(remaining_swapped) == 0
+    assert len(output.decode_seq_groups) == 1
+    assert len(output.prefill_seq_groups) == 0
+    assert output.blocks_to_copy == {2: [3]}
+
+
+def test_scheduling_budget():
+    TOKEN_BUDGET = 4
+    MAX_SEQS = 4
+    budget = SchedulingBudget(token_budget=TOKEN_BUDGET, max_num_seqs=MAX_SEQS)
+    assert budget.can_schedule(num_new_tokens=1, num_new_seqs=1)
+    assert budget.can_schedule(num_new_tokens=4, num_new_seqs=4)
+    assert not budget.can_schedule(num_new_tokens=1, num_new_seqs=5)
+    assert not budget.can_schedule(num_new_tokens=5, num_new_seqs=1)
+    assert not budget.can_schedule(num_new_tokens=5, num_new_seqs=5)
+    assert budget.remaining_token_budget() == TOKEN_BUDGET
+
+    # Verify add/subtract num batched tokens.
+    _, seq_group = create_dummy_prompt("1", 3)
+    budget.add_num_batched_tokens(seq_group.request_id, 2)
+    assert budget.remaining_token_budget() == 2
+    assert budget.num_batched_tokens == 2
+    assert budget.can_schedule(num_new_tokens=2, num_new_seqs=1)
+    assert not budget.can_schedule(num_new_tokens=3, num_new_seqs=1)
+    # Verify adding another seq group is no-op.
+    budget.add_num_batched_tokens(seq_group.request_id, 2)
+    assert budget.remaining_token_budget() == 2
+    assert budget.num_batched_tokens == 2
+    budget.subtract_num_batched_tokens(seq_group.request_id, 2)
+    assert budget.remaining_token_budget() == 4
+    assert budget.num_batched_tokens == 0
+    budget.subtract_num_batched_tokens(seq_group.request_id, 2)
+    assert budget.remaining_token_budget() == 4
+    assert budget.num_batched_tokens == 0
+
+    # Verify add/subtract max seqs.
+    _, seq_group = create_dummy_prompt("1", 3)
+    budget.add_num_seqs(seq_group.request_id, 2)
+    assert budget.can_schedule(num_new_tokens=1, num_new_seqs=2)
+    assert not budget.can_schedule(num_new_tokens=1, num_new_seqs=3)
+    assert budget.num_curr_seqs == 2
+    # Verify adding another seq group is no-op.
+    budget.add_num_seqs(seq_group.request_id, 2)
+    assert budget.num_curr_seqs == 2
+    budget.subtract_num_seqs(seq_group.request_id, 2)
+    assert budget.num_curr_seqs == 0
+    budget.subtract_num_seqs(seq_group.request_id, 2)
+    assert budget.num_curr_seqs == 0

tests/core/utils.py

@@ -1,14 +1,19 @@
 import time
-from typing import Tuple
+from typing import Optional, Tuple
 
 from vllm import SamplingParams
+from vllm.lora.request import LoRARequest
 from vllm.sequence import Logprob, Sequence, SequenceGroup
 
 
 def create_dummy_prompt(
-        request_id: str,
-        prompt_length: int,
-        block_size: int = None) -> Tuple[Sequence, SequenceGroup]:
+    request_id: str,
+    prompt_length: int,
+    block_size: Optional[int] = None,
+    lora_request: Optional[LoRARequest] = None,
+    use_beam_search: bool = False,
+    best_of: int = 1,
+) -> Tuple[Sequence, SequenceGroup]:
     if not block_size:
         block_size = prompt_length
 
@@ -17,14 +22,16 @@ def create_dummy_prompt(
     prompt_tokens = list(range(prompt_length))
     prompt_str = " ".join([str(t) for t in prompt_tokens])
     prompt = Sequence(int(request_id), prompt_str, prompt_tokens, block_size)
-    seq_group = SequenceGroup(request_id, [prompt], SamplingParams(),
-                              time.time(), None)
+    seq_group = SequenceGroup(
+        request_id, [prompt],
+        SamplingParams(use_beam_search=use_beam_search, best_of=best_of),
+        time.time(), lora_request)
 
     return prompt, seq_group
 
 
 def create_seq_group(
-    seq_prompt_lens=1024,
+    seq_prompt_len=1024,
     seq_output_lens=(128, ),
     request_id='0',
     seq_id_start=0,
@@ -32,7 +39,7 @@ def create_seq_group(
 
     assert len(seq_output_lens) > 0
 
-    prompt_token_ids = [0] * seq_prompt_lens
+    prompt_token_ids = [0] * seq_prompt_len
 
     seqs = []
     for seq_id_offset, output_len in enumerate(seq_output_lens):

tests/engine/test_detokenization.py

@@ -0,0 +1,32 @@
+import pytest
+
+from vllm.entrypoints.llm import LLM
+from vllm.sampling_params import SamplingParams
+
+
+@pytest.mark.parametrize("model", ["facebook/opt-125m"])
+def test_computed_prefix_blocks(model: str):
+    # This test checks if the engine generates completions both with and
+    # without optional detokenization, that detokenization includes text
+    # and no-detokenization doesn't, and that both completions have the same
+    # token_ids.
+    prompt = (
+        "You are a helpful assistant. How do I build a car from cardboard and "
+        "paper clips? Is there an easy to follow video tutorial available "
+        "online for free?")
+
+    llm = LLM(model=model)
+    sampling_params = SamplingParams(max_tokens=10,
+                                     temperature=0.0,
+                                     detokenize=False)
+
+    outputs_no_detokenization = llm.generate(prompt,
+                                             sampling_params)[0].outputs[0]
+    sampling_params.detokenize = True
+    outputs_with_detokenization = llm.generate(prompt,
+                                               sampling_params)[0].outputs[0]
+
+    assert outputs_no_detokenization.text == ''
+    assert outputs_with_detokenization.text != ''
+    assert outputs_no_detokenization.token_ids == \
+        outputs_with_detokenization.token_ids

tests/entrypoints/test_server_oot_registration.py

@@ -0,0 +1,66 @@
+import multiprocessing
+import sys
+import time
+
+import torch
+from openai import OpenAI, OpenAIError
+
+from vllm import ModelRegistry
+from vllm.model_executor.models.opt import OPTForCausalLM
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.utils import get_open_port
+
+
+class MyOPTForCausalLM(OPTForCausalLM):
+
+    def compute_logits(self, hidden_states: torch.Tensor,
+                       sampling_metadata: SamplingMetadata) -> torch.Tensor:
+        # this dummy model always predicts the first token
+        logits = super().compute_logits(hidden_states, sampling_metadata)
+        logits.zero_()
+        logits[:, 0] += 1.0
+        return logits
+
+
+def server_function(port):
+    # register our dummy model
+    ModelRegistry.register_model("OPTForCausalLM", MyOPTForCausalLM)
+    sys.argv = ["placeholder.py"] + \
+        ("--model facebook/opt-125m --dtype"
+        f" float32 --api-key token-abc123 --port {port}").split()
+    import runpy
+    runpy.run_module('vllm.entrypoints.openai.api_server', run_name='__main__')
+
+
+def test_oot_registration_for_api_server():
+    port = get_open_port()
+    server = multiprocessing.Process(target=server_function, args=(port, ))
+    server.start()
+    client = OpenAI(
+        base_url=f"http://localhost:{port}/v1",
+        api_key="token-abc123",
+    )
+    while True:
+        try:
+            completion = client.chat.completions.create(
+                model="facebook/opt-125m",
+                messages=[{
+                    "role": "system",
+                    "content": "You are a helpful assistant."
+                }, {
+                    "role": "user",
+                    "content": "Hello!"
+                }],
+                temperature=0,
+            )
+            break
+        except OpenAIError as e:
+            if "Connection error" in str(e):
+                time.sleep(3)
+            else:
+                raise e
+    server.kill()
+    generated_text = completion.choices[0].message.content
+    # make sure only the first token is generated
+    rest = generated_text.replace("<s>", "")
+    assert rest == ""

tests/fp8_kv/llama2-70b-fp8-kv/kv_cache_scales.json

@@ -0,0 +1,90 @@
+{
+    "model_type": "llama",
+    "kv_cache": {
+        "dtype": "float8_e4m3fn",
+        "scaling_factor": {
+            "0": {
+                "0": 0.0230364128947258,
+                "1": 0.01979283057153225,
+                "2": 0.0241350457072258,
+                "3": 0.0308314748108387,
+                "4": 0.0430733822286129,
+                "5": 0.0370396226644516,
+                "6": 0.0306222103536129,
+                "7": 0.0357491634786129,
+                "8": 0.0358189195394516,
+                "9": 0.0443289652466774,
+                "10": 0.0433175228536129,
+                "11": 0.0416782945394516,
+                "12": 0.0366908498108387,
+                "13": 0.0432477705180645,
+                "14": 0.0410505048930645,
+                "15": 0.0457589291036129,
+                "16": 0.0418526791036129,
+                "17": 0.0432477705180645,
+                "18": 0.0469447560608387,
+                "19": 0.0514787957072258,
+                "20": 0.0541294664144516,
+                "21": 0.0587681382894516,
+                "22": 0.0625,
+                "23": 0.0585588738322258,
+                "24": 0.0600237175822258,
+                "25": 0.0588030144572258,
+                "26": 0.0531180277466774,
+                "27": 0.06396484375,
+                "28": 0.0603027381002903,
+                "29": 0.0582101047039032,
+                "30": 0.0625348836183548,
+                "31": 0.0585588738322258,
+                "32": 0.0582798570394516,
+                "33": 0.0575125589966774,
+                "34": 0.0590820349752903,
+                "35": 0.0614188089966774,
+                "36": 0.0631975457072258,
+                "37": 0.0615931935608387,
+                "38": 0.0601283498108387,
+                "39": 0.0571986623108387,
+                "40": 0.0670340433716774,
+                "41": 0.0523507259786129,
+                "42": 0.0547223798930645,
+                "43": 0.0631975457072258,
+                "44": 0.0663713738322258,
+                "45": 0.0603376142680645,
+                "46": 0.0652204304933548,
+                "47": 0.0734514519572258,
+                "48": 0.0693708211183548,
+                "49": 0.0725446492433548,
+                "50": 0.0627790242433548,
+                "51": 0.0691266804933548,
+                "52": 0.0688825398683548,
+                "53": 0.068429134786129,
+                "54": 0.0605119988322258,
+                "55": 0.0799386203289032,
+                "56": 0.0853097140789032,
+                "57": 0.0661969929933548,
+                "58": 0.0689871683716774,
+                "59": 0.0724051371216774,
+                "60": 0.0541643425822258,
+                "61": 0.0626743882894516,
+                "62": 0.0628487765789032,
+                "63": 0.0607212632894516,
+                "64": 0.0589076466858387,
+                "65": 0.0451660193502903,
+                "66": 0.0453055277466774,
+                "67": 0.0414341539144516,
+                "68": 0.0385044664144516,
+                "69": 0.0414341539144516,
+                "70": 0.0466308631002903,
+                "71": 0.0399693101644516,
+                "72": 0.0437011756002903,
+                "73": 0.0434221550822258,
+                "74": 0.0428989976644516,
+                "75": 0.0401785746216774,
+                "76": 0.0431082621216774,
+                "77": 0.0484444759786129,
+                "78": 0.0417829267680645,
+                "79": 0.0418178029358387
+            }
+        }
+    }
+}

tests/fp8_kv/llama2-7b-fp8-kv/kv_cache_scales.json

@@ -0,0 +1,42 @@
+{
+    "model_type": "llama",
+    "kv_cache": {
+        "dtype": "float8_e4m3fn",
+        "scaling_factor": {
+            "0": {
+                "0": 0.0152239128947258,
+                "1": 0.0188860222697258,
+                "2": 0.0354178324341774,
+                "3": 0.0376674123108387,
+                "4": 0.0418526791036129,
+                "5": 0.0433175228536129,
+                "6": 0.0397600457072258,
+                "7": 0.0424455925822258,
+                "8": 0.0415387861430645,
+                "9": 0.0408412404358387,
+                "10": 0.0395856611430645,
+                "11": 0.0377371683716774,
+                "12": 0.0400739423930645,
+                "13": 0.040771484375,
+                "14": 0.0393415205180645,
+                "15": 0.0369001142680645,
+                "16": 0.03857421875,
+                "17": 0.0387486070394516,
+                "18": 0.0403180830180645,
+                "19": 0.0396205373108387,
+                "20": 0.0375627800822258,
+                "21": 0.0407366082072258,
+                "22": 0.0432477705180645,
+                "23": 0.0377022884786129,
+                "24": 0.0399693101644516,
+                "25": 0.0374581478536129,
+                "26": 0.0413295216858387,
+                "27": 0.0442243330180645,
+                "28": 0.0424804724752903,
+                "29": 0.0456891767680645,
+                "30": 0.0409109964966774,
+                "31": 0.0482352152466774
+            }
+        }
+    }
+}

tests/kernels/test_attention.py

@@ -32,7 +32,7 @@ HEAD_SIZES = [64, 80, 96, 112, 128, 256
 
 BLOCK_SIZES = [16, 32]
 USE_ALIBI = [False, True]
-KV_CACHE_DTYPE = ["auto", "fp8_e5m2"]
+KV_CACHE_DTYPE = ["auto", "fp8"]
 SEEDS = [0]
 CUDA_DEVICES = [
     f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)
@@ -172,6 +172,9 @@ def test_paged_attention(
                                                 device)
     key_cache, value_cache = key_caches[0], value_caches[0]
 
+    # Using default kv_scale
+    kv_scale = 1.0
+
     # Call the paged attention kernel.
     output = torch.empty_like(query)
     if version == "v1":
@@ -188,6 +191,7 @@ def test_paged_attention(
             max_context_len,
             alibi_slopes,
             kv_cache_dtype,
+            kv_scale,
         )
     elif version == "v2":
         num_partitions = ((max_context_len + PARTITION_SIZE - 1) //
@@ -219,12 +223,13 @@ def test_paged_attention(
             max_context_len,
             alibi_slopes,
             kv_cache_dtype,
+            kv_scale,
         )
     else:
         raise AssertionError(f"Unknown version: {version}")
 
     # Run the reference implementation.
-    if kv_cache_dtype == "fp8_e5m2":
+    if kv_cache_dtype == "fp8":
         # Convert cache data back to dtype.
         x = 16 // torch.tensor([], dtype=dtype).element_size()
         key_cache_shape = (NUM_BLOCKS, num_kv_heads, head_size // x,
@@ -232,14 +237,14 @@ def test_paged_attention(
         dequantized_key_cache = torch.empty(size=key_cache_shape,
                                             dtype=dtype,
                                             device=device)
-        cache_ops.convert_fp8_e5m2(key_cache, dequantized_key_cache)
+        cache_ops.convert_fp8(key_cache, dequantized_key_cache)
         key_cache = dequantized_key_cache
 
         value_cache_shape = value_cache.shape
         dequantized_value_cache = torch.empty(size=value_cache_shape,
                                               dtype=dtype,
                                               device=device)
-        cache_ops.convert_fp8_e5m2(value_cache, dequantized_value_cache)
+        cache_ops.convert_fp8(value_cache, dequantized_value_cache)
         value_cache = dequantized_value_cache
 
     ref_output = torch.empty_like(query)
@@ -263,7 +268,8 @@ def test_paged_attention(
 
     # NOTE(zhaoyang): FP8 KV Cache will introduce quantization error,
     # so we use a relaxed tolerance for the test.
-    if kv_cache_dtype == "fp8_e5m2":
+    atol, rtol = 1e-3, 1e-5
+    if kv_cache_dtype == "fp8":
         atol, rtol = 1e-2, 1e-5
     assert torch.allclose(output, ref_output, atol=atol, rtol=rtol)
 

tests/kernels/test_cache.py

@@ -5,6 +5,7 @@ import pytest
 import torch
 
 from vllm._C import cache_ops
+from vllm.utils import is_hip
 
 COPYING_DIRECTION = [('cuda', 'cpu'), ('cuda', 'cuda'), ('cpu', 'cuda')]
 DTYPES = [torch.half, torch.bfloat16, torch.float]
@@ -23,7 +24,7 @@ SEEDS = [0]
 CUDA_DEVICES = [
     f"cuda:{i}" for i in range(1 if torch.cuda.device_count() == 1 else 2)
 ]
-KV_CACHE_DTYPE = ["auto", "fp8_e5m2"]
+KV_CACHE_DTYPE = ["auto", "fp8"]
 
 
 @pytest.mark.parametrize("num_mappings", NUM_MAPPINGS)
@@ -105,6 +106,7 @@ def test_copy_blocks(
 @pytest.mark.parametrize("dtype", DTYPES)
 @pytest.mark.parametrize("seed", SEEDS)
 @pytest.mark.parametrize("device", CUDA_DEVICES)
+@pytest.mark.parametrize("kv_cache_dtype", KV_CACHE_DTYPE)
 @torch.inference_mode()
 def test_reshape_and_cache(
     kv_cache_factory,
@@ -116,7 +118,10 @@ def test_reshape_and_cache(
     dtype: torch.dtype,
     seed: int,
     device: str,
+    kv_cache_dtype: str,
 ) -> None:
+    if not is_hip() and kv_cache_dtype == "fp8":
+        pytest.skip()  # This test is not tuned for e5m2 cuda precision
     random.seed(seed)
     torch.random.manual_seed(seed)
     if torch.cuda.is_available():
@@ -132,17 +137,33 @@ def test_reshape_and_cache(
 
     # Create the KV caches.
     key_caches, value_caches = kv_cache_factory(num_blocks, block_size, 1,
-                                                num_heads, head_size, dtype,
-                                                None, seed, device)
+                                                num_heads, head_size,
+                                                kv_cache_dtype, dtype, seed,
+                                                device)
     key_cache, value_cache = key_caches[0], value_caches[0]
 
     # Clone the KV caches.
-    cloned_key_cache = key_cache.clone()
-    cloned_value_cache = value_cache.clone()
+    if kv_cache_dtype == "fp8":
+        cloned_key_cache = torch.empty_like(key_cache, dtype=torch.float16)
+        cache_ops.convert_fp8(key_cache, cloned_key_cache)
+        cloned_value_cache = torch.empty_like(value_cache, dtype=torch.float16)
+        cache_ops.convert_fp8(value_cache, cloned_value_cache)
+    else:
+        cloned_key_cache = key_cache.clone()
+        cloned_value_cache = value_cache.clone()
+
+    # Using default kv_scale
+    kv_scale = 1.0
 
     # Call the reshape_and_cache kernel.
     cache_ops.reshape_and_cache(key, value, key_cache, value_cache,
-                                slot_mapping, "auto")
+                                slot_mapping, kv_cache_dtype, kv_scale)
+
+    if kv_cache_dtype == "fp8":
+        result_key_cache = torch.empty_like(key_cache, dtype=torch.float16)
+        cache_ops.convert_fp8(key_cache, result_key_cache)
+        result_value_cache = torch.empty_like(value_cache, dtype=torch.float16)
+        cache_ops.convert_fp8(value_cache, result_value_cache)
 
     # Run the reference implementation.
     reshaped_key = key.reshape(num_tokens, *key_cache[0, :, :, 0, :].shape)
@@ -156,8 +177,18 @@ def test_reshape_and_cache(
         cloned_key_cache[block_idx, :, :, block_offset, :] = reshaped_key[i]
         cloned_value_cache[block_idx, :, :, block_offset] = value[i]
 
-    assert torch.allclose(key_cache, cloned_key_cache)
-    assert torch.allclose(value_cache, cloned_value_cache)
+    if kv_cache_dtype == "fp8":
+        assert torch.allclose(result_key_cache,
+                              cloned_key_cache,
+                              atol=0.001,
+                              rtol=0.1)
+        assert torch.allclose(result_value_cache,
+                              cloned_value_cache,
+                              atol=0.001,
+                              rtol=0.1)
+    else:
+        assert torch.allclose(key_cache, cloned_key_cache)
+        assert torch.allclose(value_cache, cloned_value_cache)
 
 
 @pytest.mark.parametrize("direction", COPYING_DIRECTION)
@@ -169,6 +200,7 @@ def test_reshape_and_cache(
 @pytest.mark.parametrize("dtype", DTYPES)
 @pytest.mark.parametrize("seed", SEEDS)
 @pytest.mark.parametrize("device", CUDA_DEVICES)
+@pytest.mark.parametrize("kv_cache_dtype", KV_CACHE_DTYPE)
 @torch.inference_mode()
 def test_swap_blocks(
     kv_cache_factory,
@@ -181,7 +213,12 @@ def test_swap_blocks(
     dtype: torch.dtype,
     seed: int,
     device: str,
+    kv_cache_dtype: str,
 ) -> None:
+    if kv_cache_dtype == "fp8" and "cpu" in direction:
+        pytest.skip()
+    if not is_hip() and kv_cache_dtype == "fp8":
+        pytest.skip()  # This test is not tuned for e5m2 cuda precision
     random.seed(seed)
     torch.random.manual_seed(seed)
     if torch.cuda.is_available():
@@ -202,13 +239,13 @@ def test_swap_blocks(
 
     # Create the KV caches on the first device.
     src_key_caches, src_value_caches = kv_cache_factory(
-        num_blocks, block_size, 1, num_heads, head_size, dtype, None, seed,
-        src_device)
+        num_blocks, block_size, 1, num_heads, head_size, kv_cache_dtype, dtype,
+        seed, src_device)
 
     # Create the KV caches on the second device.
     dist_key_caches, dist_value_caches = kv_cache_factory(
-        num_blocks, block_size, 1, num_heads, head_size, dtype, None, seed,
-        dst_device)
+        num_blocks, block_size, 1, num_heads, head_size, kv_cache_dtype, dtype,
+        seed, dst_device)
 
     src_key_caches_clone = src_key_caches[0].clone()
     src_value_caches_clone = src_value_caches[0].clone()
@@ -223,3 +260,40 @@ def test_swap_blocks(
                               dist_key_caches[0][dst].cpu())
         assert torch.allclose(src_value_caches_clone[src].cpu(),
                               dist_value_caches[0][dst].cpu())
+
+
+@pytest.mark.skipif(not is_hip(), reason="FP8 conversion test requires e4m3")
+@pytest.mark.parametrize("num_heads", NUM_HEADS)
+@pytest.mark.parametrize("head_size", HEAD_SIZES)
+@pytest.mark.parametrize("block_size", BLOCK_SIZES)
+@pytest.mark.parametrize("num_blocks", NUM_BLOCKS)
+@pytest.mark.parametrize("dtype", DTYPES)
+@pytest.mark.parametrize("seed", SEEDS)
+@pytest.mark.parametrize("device", CUDA_DEVICES)
+@torch.inference_mode()
+def test_fp8_conversion(
+    num_heads: int,
+    head_size: int,
+    block_size: int,
+    num_blocks: int,
+    dtype: torch.dtype,
+    seed: int,
+    device: str,
+) -> None:
+    random.seed(seed)
+    torch.random.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+
+    low = -224.0
+    high = 224.0
+    shape = (num_blocks, num_heads, head_size, block_size)
+    cache = torch.empty(shape, dtype=dtype, device=device)
+    cache.uniform_(low, high)
+
+    cache_fp8 = torch.empty_like(cache, dtype=torch.uint8)
+    cache_ops.convert_fp8(cache, cache_fp8)
+
+    converted_cache = torch.empty_like(cache)
+    cache_ops.convert_fp8(cache_fp8, converted_cache)
+
+    assert torch.allclose(cache, converted_cache, atol=0.001, rtol=0.1)

tests/lora/test_lora_checkpoints.py

@@ -0,0 +1,40 @@
+import pytest
+
+from vllm.lora.models import LoRAModel
+from vllm.model_executor.models.baichuan import BaiChuanBaseForCausalLM
+
+
+@pytest.mark.parametrize("lora_name", ["baichuan7B", "chatglm3-6b"])
+def test_load_checkpoints(lora_name, chatglm3_lora_files, baichuan_lora_files):
+    supported_lora_modules = BaiChuanBaseForCausalLM.supported_lora_modules
+    packed_modules_mapping = BaiChuanBaseForCausalLM.packed_modules_mapping
+    embedding_modules = BaiChuanBaseForCausalLM.embedding_modules
+    embed_padding_modules = BaiChuanBaseForCausalLM.embedding_padding_modules
+    expected_lora_modules = []
+    for module in supported_lora_modules:
+        if module in packed_modules_mapping:
+            expected_lora_modules.extend(packed_modules_mapping[module])
+        else:
+            expected_lora_modules.append(module)
+    if lora_name == "baichuan7B":
+        # For the baichuan7B model, load it's LoRA,
+        # and the test should pass.
+        LoRAModel.from_local_checkpoint(
+            baichuan_lora_files,
+            expected_lora_modules,
+            lora_model_id=1,
+            device="cpu",
+            embedding_modules=embedding_modules,
+            embedding_padding_modules=embed_padding_modules)
+    else:
+        # For the baichuan7B model, load chatglm3-6b's LoRA,
+        # and the test should raise the following error.
+        expected_error = "Please verify that the loaded LoRA module is correct"  # noqa: E501
+        with pytest.raises(ValueError, match=expected_error):
+            LoRAModel.from_local_checkpoint(
+                chatglm3_lora_files,
+                expected_lora_modules,
+                lora_model_id=1,
+                device="cpu",
+                embedding_modules=embedding_modules,
+                embedding_padding_modules=embed_padding_modules)

tests/lora/test_worker.py

@@ -3,8 +3,8 @@ import random
 import tempfile
 from unittest.mock import patch
 
-from vllm.config import (DeviceConfig, LoRAConfig, ModelConfig, ParallelConfig,
-                         SchedulerConfig)
+from vllm.config import (CacheConfig, DeviceConfig, LoRAConfig, ModelConfig,
+                         ParallelConfig, SchedulerConfig)
 from vllm.lora.models import LoRAMapping
 from vllm.lora.request import LoRARequest
 from vllm.worker.worker import Worker
@@ -27,6 +27,10 @@ def test_worker_apply_lora(sql_lora_files):
         parallel_config=ParallelConfig(1, 1, False),
         scheduler_config=SchedulerConfig(32, 32, 32),
         device_config=DeviceConfig("cuda"),
+        cache_config=CacheConfig(block_size=16,
+                                 gpu_memory_utilization=1.,
+                                 swap_space=0,
+                                 cache_dtype="auto"),
         local_rank=0,
         rank=0,
         lora_config=LoRAConfig(max_lora_rank=8, max_cpu_loras=32,

tests/model_executor/weight_utils.py

@@ -0,0 +1,26 @@
+import os
+
+import huggingface_hub.constants
+import pytest
+
+from vllm.model_executor.weight_utils import enable_hf_transfer
+
+
+def test_hf_transfer_auto_activation():
+    if "HF_HUB_ENABLE_HF_TRANSFER" in os.environ:
+        # in case it is already set, we can't test the auto activation
+        pytest.skip(
+            "HF_HUB_ENABLE_HF_TRANSFER is set, can't test auto activation")
+    enable_hf_transfer()
+    try:
+        # enable hf hub transfer if available
+        import hf_transfer  # type: ignore # noqa
+        HF_TRANFER_ACTIVE = True
+    except ImportError:
+        HF_TRANFER_ACTIVE = False
+    assert (huggingface_hub.constants.HF_HUB_ENABLE_HF_TRANSFER ==
+            HF_TRANFER_ACTIVE)
+
+
+if __name__ == "__main__":
+    test_hf_transfer_auto_activation()

tests/models/test_oot_registration.py

@@ -0,0 +1,32 @@
+import torch
+
+from vllm import LLM, ModelRegistry, SamplingParams
+from vllm.model_executor.models.opt import OPTForCausalLM
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+
+
+class MyOPTForCausalLM(OPTForCausalLM):
+
+    def compute_logits(self, hidden_states: torch.Tensor,
+                       sampling_metadata: SamplingMetadata) -> torch.Tensor:
+        # this dummy model always predicts the first token
+        logits = super().compute_logits(hidden_states, sampling_metadata)
+        logits.zero_()
+        logits[:, 0] += 1.0
+        return logits
+
+
+def test_oot_registration():
+    # register our dummy model
+    ModelRegistry.register_model("OPTForCausalLM", MyOPTForCausalLM)
+    prompts = ["Hello, my name is", "The text does not matter"]
+    sampling_params = SamplingParams(temperature=0)
+    llm = LLM(model="facebook/opt-125m")
+    first_token = llm.get_tokenizer().decode(0)
+    outputs = llm.generate(prompts, sampling_params)
+
+    for output in outputs:
+        generated_text = output.outputs[0].text
+        # make sure only the first token is generated
+        rest = generated_text.replace(first_token, "")
+        assert rest == ""

tests/quantization/test_autogptq_marlin_configs.py

@@ -0,0 +1,68 @@
+"""Tests whether Marlin models can be loaded from the autogptq config.
+
+Run `pytest tests/quantization/test_autogptq_marlin_configs.py --forked`.
+"""
+
+from dataclasses import dataclass
+
+import pytest
+
+from vllm.config import ModelConfig
+
+
+@dataclass
+class ModelPair:
+    model_marlin: str
+    model_gptq: str
+
+
+# Model Id // Expected Kernel
+MODELS_QUANT_TYPE = [
+    # compat: autogptq <=0.7.1 is_marlin_format: bool
+    ("neuralmagic/TinyLlama-1.1B-Chat-v1.0-marlin", "marlin"),
+    ("TheBloke/Llama-2-7B-Chat-GPTQ", "gptq"),
+    # compat: autogptq >=0.8.0 use checkpoint_format: str
+    ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-Marlin-4bit", "marlin"),
+    ("LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit", "gptq")
+]
+
+
+@pytest.mark.parametrize("model_quant_type", MODELS_QUANT_TYPE)
+def test_auto_gptq(model_quant_type: str, ) -> None:
+    model_path, quant_type = model_quant_type
+
+    model_config_no_quant_arg = ModelConfig(
+        model_path,
+        model_path,
+        tokenizer_mode="auto",
+        trust_remote_code=False,
+        download_dir=None,
+        load_format="dummy",
+        seed=0,
+        dtype="float16",
+        revision=None,
+        quantization=None  # case 1
+    )
+
+    model_config_quant_arg = ModelConfig(
+        model_path,
+        model_path,
+        tokenizer_mode="auto",
+        trust_remote_code=False,
+        download_dir=None,
+        load_format="dummy",
+        seed=0,
+        dtype="float16",
+        revision=None,
+        quantization="gptq"  # case 2
+    )
+
+    assert model_config_no_quant_arg.quantization == quant_type, (
+        f"Expected quant_type == {quant_type} for {model_path}, "
+        f"but found {model_config_no_quant_arg.quantization} "
+        "for no --quantization None case")
+
+    assert model_config_quant_arg.quantization == quant_type, (
+        f"Expected quant_type == {quant_type} for {model_path}, "
+        f"but found {model_config_quant_arg.quantization} "
+        "for --quantization gptq case")

tests/samplers/test_logits_processor.py

@@ -0,0 +1,62 @@
+import pytest
+import torch
+
+from vllm import SamplingParams
+
+MODELS = ["facebook/opt-125m"]
+
+
+@pytest.mark.parametrize("model", MODELS)
+@pytest.mark.parametrize("dtype", ["half"])
+def test_logits_processor_force_generate(
+    vllm_runner,
+    example_prompts,
+    model: str,
+    dtype: str,
+) -> None:
+    vllm_model = vllm_runner(model, dtype=dtype)
+    tokenizer = vllm_model.model.get_tokenizer()
+    repeat_times = 2
+    enforced_answers = " vLLM"
+    vllm_token_ids = tokenizer.encode(enforced_answers,
+                                      add_special_tokens=False)
+    max_tokens = len(vllm_token_ids) * repeat_times
+
+    def pick_vllm(token_ids, logits):
+        token_id = vllm_token_ids[len(token_ids) % len(vllm_token_ids)]
+        logits[token_id] = torch.finfo(logits.dtype).max
+        return logits
+
+    params_with_logprobs = SamplingParams(
+        logits_processors=[pick_vllm],
+        prompt_logprobs=3,
+        max_tokens=max_tokens,
+    )
+
+    # test logits_processors when prompt_logprobs is not None
+    vllm_model.model._add_request(
+        prompt=example_prompts[0],
+        sampling_params=params_with_logprobs,
+        prompt_token_ids=None,
+    )
+
+    # test prompt_logprobs is not None
+    vllm_model.model._add_request(
+        prompt=example_prompts[1],
+        sampling_params=SamplingParams(
+            prompt_logprobs=3,
+            max_tokens=max_tokens,
+        ),
+        prompt_token_ids=None,
+    )
+
+    # test grouped requests
+    vllm_model.model._add_request(
+        prompt=example_prompts[2],
+        sampling_params=SamplingParams(max_tokens=max_tokens),
+        prompt_token_ids=None,
+    )
+
+    outputs = vllm_model.model._run_engine(False)
+
+    assert outputs[0].outputs[0].text == enforced_answers * repeat_times

tests/spec_decode/e2e/conftest.py

@@ -0,0 +1,41 @@
+import pytest
+
+from tests.conftest import cleanup
+from vllm import LLM
+from vllm.model_executor.utils import set_random_seed
+
+
+@pytest.fixture
+def baseline_llm_generator(common_llm_kwargs, per_test_common_llm_kwargs,
+                           baseline_llm_kwargs, seed):
+    return create_llm_generator(common_llm_kwargs, per_test_common_llm_kwargs,
+                                baseline_llm_kwargs, seed)
+
+
+@pytest.fixture
+def test_llm_generator(common_llm_kwargs, per_test_common_llm_kwargs,
+                       test_llm_kwargs, seed):
+    return create_llm_generator(common_llm_kwargs, per_test_common_llm_kwargs,
+                                test_llm_kwargs, seed)
+
+
+def create_llm_generator(common_llm_kwargs, per_test_common_llm_kwargs,
+                         distinct_llm_kwargs, seed):
+    kwargs = {
+        **common_llm_kwargs,
+        **per_test_common_llm_kwargs,
+        **distinct_llm_kwargs,
+    }
+
+    def generator_inner():
+        llm = LLM(**kwargs)
+
+        set_random_seed(seed)
+
+        yield llm
+        del llm
+        cleanup()
+
+    for llm in generator_inner():
+        yield llm
+        del llm

tests/spec_decode/e2e/test_correctness.py

@@ -0,0 +1,50 @@
+import pytest
+
+from vllm import SamplingParams
+
+
+@pytest.mark.parametrize(
+    "common_llm_kwargs",
+    [{
+        # Use a small model for a fast test.
+        "model": "facebook/opt-125m",
+        "speculative_model": "facebook/opt-125m",
+        "num_speculative_tokens": 5,
+
+        # Required for spec decode.
+        "use_v2_block_manager": True
+    }])
+@pytest.mark.parametrize("per_test_common_llm_kwargs", [{}])
+@pytest.mark.parametrize("test_llm_kwargs", [{}])
+@pytest.mark.parametrize("seed", [1])
+def test_spec_decode_config(test_llm_generator):
+    output_len = 1024
+    temperature = 0.0
+
+    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(
+        max_tokens=output_len,
+        ignore_eos=True,
+        temperature=temperature,
+    )
+
+    with pytest.raises(
+            AssertionError,
+            match="Speculative decoding not yet supported for GPU backend"):
+        get_token_ids_from_llm_generator(test_llm_generator, prompts,
+                                         sampling_params)
+
+
+def get_token_ids_from_llm_generator(llm_generator, prompts, sampling_params):
+    for llm in llm_generator:
+        outputs = llm.generate(prompts, sampling_params, use_tqdm=True)
+        token_ids = [output.outputs[0].token_ids for output in outputs]
+        del llm
+
+    return token_ids

tests/spec_decode/test_batch_expansion.py

@@ -7,6 +7,7 @@ from .utils import create_seq_group_metadata_from_prompts, mock_worker
 
 
 @pytest.mark.parametrize('num_target_seq_ids', [100])
+@pytest.mark.skip_global_cleanup
 def test_create_target_seq_id_iterator(num_target_seq_ids: int):
     """Verify all new sequence ids are greater than all input
     seq ids.
@@ -27,6 +28,7 @@ def test_create_target_seq_id_iterator(num_target_seq_ids: int):
 
 
 @pytest.mark.parametrize('k', [1, 2, 6])
+@pytest.mark.skip_global_cleanup
 def test_get_token_ids_to_score(k: int):
     """Verify correct tokens are selected for scoring.
     """
@@ -53,6 +55,7 @@ def test_get_token_ids_to_score(k: int):
 
 
 @pytest.mark.parametrize('k', [1, 2, 6])
+@pytest.mark.skip_global_cleanup
 def test_create_single_target_seq_group_metadata(k: int):
     """Verify correct creation of a batch-expanded seq group metadata.
     """

tests/spec_decode/test_spec_decode_worker.py

@@ -487,7 +487,7 @@ def test_empty_input_batch(k: int, batch_size: int):
         **execute_model_data.to_dict())
 
 
-@torch.inference_mode()
+@pytest.mark.skip_global_cleanup
 def test_init_device():
     """Verify SpecDecodeWorker invokes proposer/scorer worker init_device, as
     well as other GPU initialization.
@@ -512,8 +512,8 @@ def test_init_device():
 
 
 @torch.inference_mode()
-def test_init_cache_engine():
-    """Verify SpecDecodeWorker invokes init_cache_engine on proposer/scorer
+def test_initialize_cache():
+    """Verify SpecDecodeWorker invokes initialize_cache on proposer/scorer
     workers.
     """
     draft_worker = mock_worker(cls=MultiStepWorker)
@@ -525,23 +525,22 @@ def test_init_cache_engine():
     worker = SpecDecodeWorker(draft_worker, target_worker, rejection_sampler,
                               metrics_collector)
 
-    cache_config = MagicMock()
+    kwargs = {"num_gpu_blocks": 1024, "num_cpu_blocks": 1023}
+    worker.initialize_cache(**kwargs)
 
-    worker.init_cache_engine(cache_config)
-
-    draft_worker.init_cache_engine.assert_called_once_with(cache_config)
-    target_worker.init_cache_engine.assert_called_once_with(cache_config)
+    draft_worker.initialize_cache.assert_called_once_with(**kwargs)
+    target_worker.initialize_cache.assert_called_once_with(**kwargs)
 
 
 @pytest.mark.parametrize('available_gpu_blocks', [1, 1024])
 @pytest.mark.parametrize('available_cpu_blocks', [500])
 @pytest.mark.parametrize('target_cache_block_size_bytes', [2 * 2 * 4096])
 @pytest.mark.parametrize('draft_kv_size_bytes', [0, 2 * 2 * 768, 2 * 2 * 4096])
-@torch.inference_mode()
-def test_profile_num_available_blocks(available_gpu_blocks: int,
-                                      available_cpu_blocks: int,
-                                      target_cache_block_size_bytes: int,
-                                      draft_kv_size_bytes: int):
+@pytest.mark.skip_global_cleanup
+def test_determine_num_available_blocks(available_gpu_blocks: int,
+                                        available_cpu_blocks: int,
+                                        target_cache_block_size_bytes: int,
+                                        draft_kv_size_bytes: int):
     """Verify SpecDecodeWorker correctly profiles num available GPU blocks.
     Specifically, it should run profiling in the scorer worker, and then evenly
     split the blocks between proposer and scorer worker.
@@ -552,7 +551,7 @@ def test_profile_num_available_blocks(available_gpu_blocks: int,
     rejection_sampler.token_id_dtype = torch.int64
     metrics_collector = MagicMock(spec=AsyncMetricsCollector)
 
-    target_worker.profile_num_available_blocks.return_value = (
+    target_worker.determine_num_available_blocks.return_value = (
         available_gpu_blocks, available_cpu_blocks)
     target_worker.get_cache_block_size_bytes.return_value = (
         target_cache_block_size_bytes)
@@ -561,17 +560,9 @@ def test_profile_num_available_blocks(available_gpu_blocks: int,
     worker = SpecDecodeWorker(draft_worker, target_worker, rejection_sampler,
                               metrics_collector)
 
-    # These values do not directly impact the adjusted block size calculation,
-    # so they can be fixed.
-    gpu_memory_utilization = 0.9
-    cpu_swap_space = 100
-    block_size = 16
+    num_gpu_blocks, num_cpu_blocks = worker.determine_num_available_blocks()
 
-    num_gpu_blocks, num_cpu_blocks = worker.profile_num_available_blocks(
-        block_size, gpu_memory_utilization, cpu_swap_space, cache_dtype="auto")
-
-    target_worker.profile_num_available_blocks.assert_called_once_with(
-        block_size, gpu_memory_utilization, cpu_swap_space, "auto")
+    target_worker.determine_num_available_blocks.assert_called_once()
     assert num_cpu_blocks == available_cpu_blocks
 
     assert num_gpu_blocks == split_num_cache_blocks_evenly(
@@ -584,7 +575,7 @@ def test_profile_num_available_blocks(available_gpu_blocks: int,
 @pytest.mark.parametrize('target_cache_block_size_bytes',
                          [2 * 2 * 4096, 2 * 2 * 8192])
 @pytest.mark.parametrize('draft_kv_size_bytes', [0, 2 * 2 * 768, 2 * 2 * 4096])
-@torch.inference_mode()
+@pytest.mark.skip_global_cleanup
 def test_split_num_cache_blocks_evenly(available_gpu_blocks: int,
                                        target_cache_block_size_bytes: int,
                                        draft_kv_size_bytes: int):

tests/spec_decode/utils.py

@@ -107,18 +107,17 @@ def create_worker(cls: type,
         block_size=block_size,
         enforce_eager=enforce_eager,
     )
-
-    (model_config, cache_config, parallel_config, scheduler_config,
-     device_config, _, _) = engine_args.create_engine_configs()
+    engine_config = engine_args.create_engine_config()
 
     distributed_init_method = get_distributed_init_method(
         get_ip(), get_open_port())
 
     worker = cls(
-        model_config=model_config,
-        parallel_config=parallel_config,
-        scheduler_config=scheduler_config,
-        device_config=device_config,
+        model_config=engine_config.model_config,
+        parallel_config=engine_config.parallel_config,
+        scheduler_config=engine_config.scheduler_config,
+        device_config=engine_config.device_config,
+        cache_config=engine_config.cache_config,
         local_rank=0,
         rank=0,
         distributed_init_method=distributed_init_method,
@@ -128,10 +127,11 @@ def create_worker(cls: type,
     worker.init_device()
     worker.load_model()
 
-    cache_config.num_gpu_blocks = num_gpu_blocks
-    cache_config.num_cpu_blocks = 0
-    worker.init_cache_engine(cache_config)
-    worker.warm_up_model()
+    engine_config.cache_config.num_gpu_blocks = num_gpu_blocks
+    engine_config.cache_config.num_cpu_blocks = 0
+    worker.initialize_cache(
+        num_gpu_blocks=engine_config.cache_config.num_gpu_blocks,
+        num_cpu_blocks=engine_config.cache_config.num_cpu_blocks)
 
     return worker
 

tests/test_sequence.py

@@ -1,7 +1,36 @@
+import time
+from typing import Optional
+
 import pytest
 
-from vllm.sequence import (SamplerOutput, SequenceData, SequenceGroupOutput,
-                           SequenceOutput)
+from vllm import SamplingParams
+from vllm.lora.request import LoRARequest
+from vllm.sequence import (SamplerOutput, Sequence, SequenceData,
+                           SequenceGroup, SequenceGroupOutput, SequenceOutput)
+
+
+def create_dummy_prompt(
+    request_id: str,
+    prompt_length: int,
+    block_size: Optional[int] = None,
+    lora_request: Optional[LoRARequest] = None,
+    use_beam_search: bool = False,
+    best_of: int = 1,
+) -> SequenceGroup:
+    if not block_size:
+        block_size = prompt_length
+
+    # Create dummy prompt sequence with tokens 0...block_size-1
+    # and prompt "0 ... block_size".
+    prompt_tokens = list(range(prompt_length))
+    prompt_str = " ".join([str(t) for t in prompt_tokens])
+    prompt = Sequence(int(request_id), prompt_str, prompt_tokens, block_size)
+    seq_group = SequenceGroup(
+        request_id, [prompt],
+        SamplingParams(use_beam_search=use_beam_search, best_of=best_of),
+        time.time(), lora_request)
+
+    return seq_group
 
 
 @pytest.fixture
@@ -67,6 +96,29 @@ def test_sequence_data_prefill():
 
     # append tokens and reset, simulating recompute
     seq_data.append_token_id(1, logprob=0.0)
-    seq_data.reset_num_computed_tokens()
+    seq_data.reset_state_for_recompute()
     assert seq_data.get_num_uncomputed_tokens() == 5
     assert seq_data.get_num_computed_tokens() == 0
+
+
+def test_sequence_group_stage():
+    seq_group = create_dummy_prompt("1", 12)
+    assert seq_group.is_prefill() is True
+    seq_group.update_num_computed_tokens(6)
+    assert seq_group.is_prefill() is True
+    seq_group.update_num_computed_tokens(5)
+    assert seq_group.is_prefill() is True
+    seq_group.update_num_computed_tokens(1)
+    assert seq_group.is_prefill() is False
+    seqs = seq_group.get_seqs()
+    assert len(seqs) == 1
+    seqs[0].data.append_token_id(1, logprob=0.0)
+    for seq in seq_group.get_seqs():
+        seq.reset_state_for_recompute()
+    assert seq_group.is_prefill() is True
+    seq_group.update_num_computed_tokens(5)
+    assert seq_group.is_prefill() is True
+    seq_group.update_num_computed_tokens(7)
+    assert seq_group.is_prefill() is True
+    seq_group.update_num_computed_tokens(1)
+    assert seq_group.is_prefill() is False

tests/tokenization/test_detokenize.py

@@ -4,8 +4,8 @@ import pytest
 from transformers import AutoTokenizer
 
 from vllm.sequence import Logprob, SamplingParams, Sequence, SequenceGroup
-from vllm.transformers_utils.detokenizer import Detokenizer
-from vllm.transformers_utils.tokenizer import detokenize_incrementally
+from vllm.transformers_utils.detokenizer import (Detokenizer,
+                                                 detokenize_incrementally)
 from vllm.transformers_utils.tokenizer_group import get_tokenizer_group
 
 TRUTH = [

tests/worker/test_swap.py

@@ -10,19 +10,19 @@ def test_swap() -> None:
     engine_args = EngineArgs(model="facebook/opt-125m",
                              dtype="half",
                              load_format="dummy")
-    (model_config, cache_config, parallel_config, scheduler_config,
-     device_config, _, _) = engine_args.create_engine_configs()
-    cache_config.num_gpu_blocks = 100
-    cache_config.num_cpu_blocks = 100
+    engine_config = engine_args.create_engine_config()
+    engine_config.cache_config.num_gpu_blocks = 1000
+    engine_config.cache_config.num_cpu_blocks = 1000
 
     # Create the worker.
     distributed_init_method = get_distributed_init_method(
         get_ip(), get_open_port())
     worker = Worker(
-        model_config=model_config,
-        parallel_config=parallel_config,
-        scheduler_config=scheduler_config,
-        device_config=device_config,
+        model_config=engine_config.model_config,
+        parallel_config=engine_config.parallel_config,
+        scheduler_config=engine_config.scheduler_config,
+        device_config=engine_config.device_config,
+        cache_config=engine_config.cache_config,
         local_rank=0,
         rank=0,
         distributed_init_method=distributed_init_method,
@@ -32,8 +32,9 @@ def test_swap() -> None:
     # Initialize the worker.
     worker.init_device()
     worker.load_model()
-    worker.init_cache_engine(cache_config)
-    worker.warm_up_model()
+    worker.initialize_cache(
+        num_gpu_blocks=engine_config.cache_config.num_gpu_blocks,
+        num_cpu_blocks=engine_config.cache_config.num_cpu_blocks)
 
     # Randomly initialize the cache.
     gpu_cache = worker.cache_engine.gpu_cache

vllm/__init__.py

@@ -5,13 +5,15 @@ from vllm.engine.async_llm_engine import AsyncLLMEngine
 from vllm.engine.llm_engine import LLMEngine
 from vllm.engine.ray_utils import initialize_ray_cluster
 from vllm.entrypoints.llm import LLM
+from vllm.model_executor.models import ModelRegistry
 from vllm.outputs import CompletionOutput, RequestOutput
 from vllm.sampling_params import SamplingParams
 
-__version__ = "0.4.0"
+__version__ = "0.4.0.post1"
 
 __all__ = [
     "LLM",
+    "ModelRegistry",
     "SamplingParams",
     "RequestOutput",
     "CompletionOutput",

vllm/attention/backends/abstract.py

@@ -81,5 +81,6 @@ class AttentionImpl(ABC):
         value: torch.Tensor,
         kv_cache: torch.Tensor,
         attn_metadata: AttentionMetadata,
+        kv_scale: float,
     ) -> torch.Tensor:
         raise NotImplementedError

vllm/attention/backends/flash_attn.py

@@ -156,6 +156,7 @@ class FlashAttentionImpl(AttentionImpl):
         value: torch.Tensor,
         kv_cache: torch.Tensor,
         attn_metadata: FlashAttentionMetadata,
+        kv_scale: float,
     ) -> torch.Tensor:
         """Forward pass with FlashAttention and PagedAttention.
 
@@ -184,7 +185,8 @@ class FlashAttentionImpl(AttentionImpl):
             PagedAttention.write_to_paged_cache(key, value, key_cache,
                                                 value_cache,
                                                 attn_metadata.slot_mapping,
-                                                attn_metadata.kv_cache_dtype)
+                                                attn_metadata.kv_cache_dtype,
+                                                kv_scale)
 
         if attn_metadata.is_prompt:
             # Prompt run.
@@ -207,6 +209,9 @@ class FlashAttentionImpl(AttentionImpl):
                 )
             else:
                 # prefix-enabled attention
+                # TODO(Hai) this triton kernel has regression issue (broke) to
+                # deal with different data types between KV and FP8 KV cache,
+                # to be addressed separately.
                 output = PagedAttention.forward_prefix(
                     query,
                     key,
@@ -233,6 +238,7 @@ class FlashAttentionImpl(AttentionImpl):
                 self.num_kv_heads,
                 self.scale,
                 self.alibi_slopes,
+                kv_scale,
             )
 
         # Reshape the output tensor.

vllm/attention/backends/rocm_flash_attn.py

@@ -0,0 +1,348 @@
+"""Attention layer ROCm GPUs."""
+import os
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Type
+
+import torch
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata)
+from vllm.attention.ops.paged_attn import (PagedAttention,
+                                           PagedAttentionMetadata)
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class ROCmFlashAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_impl_cls() -> Type["ROCmFlashAttentionImpl"]:
+        return ROCmFlashAttentionImpl
+
+    @staticmethod
+    def make_metadata(*args, **kwargs) -> "ROCmFlashAttentionMetadata":
+        return ROCmFlashAttentionMetadata(*args, **kwargs)
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: Dict[int, List[int]],
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class ROCmFlashAttentionMetadata(AttentionMetadata, PagedAttentionMetadata):
+    """Metadata for FlashAttentionBackend.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+    # Currently, input sequences can only contain all prompts
+    # or all decoding. True if all sequences are prompts.
+    is_prompt: bool
+    # (batch_size,). The prompt length per sequence. None if it is a decoding.
+    prompt_lens: Optional[List[int]]
+    # prompt_lens stored as a tensor.
+    prompt_lens_tensor: Optional[torch.Tensor]
+    # The number of prompt tokens. Doesn't include padding.
+    num_prompt_tokens: int
+    # The number of generation tokens. Doesn't include padding.
+    num_generation_tokens: int
+
+    # NOTE(sang): Definition of context_len, subquery_len, and seqlen.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seqlen ----------------------|
+    #                                   |- subquery_len -|
+
+    # WARNING(sang): context_len has different definition depending on if it is
+    # prefill vs decoding. When it is prefill, it doesn't include new tokens.
+    # When it is for decoding, it includes a new token.
+
+    # Maximum subquery length in the batch.
+    max_subquery_len: Optional[int]
+    # Maximum prompt length in the batch.
+    max_prompt_len: Optional[int]
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    subquery_start_loc: Optional[torch.Tensor]
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor]
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+    use_cuda_graph: bool
+
+
+class ROCmFlashAttentionImpl(AttentionImpl):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prompt_tokens -------------->|
+    |<--prompt_0-->|<--prompt_1-->|...|<--prompt_N-1-->|
+
+    Otherwise, the layout is as follows:
+    |<------------------ num_generation_tokens (M) ----------------->|
+    |<--generation_0-->|..........|<--generation_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+    """
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: Optional[int] = None,
+        alibi_slopes: Optional[List[float]] = None,
+        sliding_window: Optional[int] = None,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.sliding_window = ((sliding_window, sliding_window)
+                               if sliding_window is not None else (-1, -1))
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        suppored_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in suppored_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {suppored_head_sizes}.")
+
+        self.use_naive_attn = torch.cuda.get_device_capability()[0] != 9
+        # NOTE: Allow for switching between Triton and CK. Defaulting to triton.
+        self.use_triton_flash_attn = (os.environ.get(
+            "VLLM_USE_TRITON_FLASH_ATTN", "True").lower() in ("true", "1"))
+        if self.use_naive_attn:
+            # AMD Radeon 7900 series (gfx1100) currently does not support
+            # xFormers nor FlashAttention. As a temporary workaround, we use
+            # naive PyTorch implementation of attention.
+            self.attn_fuc = _naive_attention()
+            logger.debug("Using naive attention in ROCmBackend")
+        elif self.use_triton_flash_attn:
+            from vllm.attention.ops.triton_flash_attention import (  # noqa: F401
+                triton_attention)
+            self.attn_func = triton_attention
+            logger.debug("Using Triton FA in ROCmBackend")
+        else:
+            from flash_attn import flash_attn_varlen_func  # noqa: F401
+            self.attn_func = flash_attn_varlen_func
+            logger.debug("Using CK FA in ROCmBackend")
+
+    def repeat_kv(self, x: torch.Tensor, n_rep: int) -> torch.Tensor:
+        """torch.repeat_interleave(x, dim=1, repeats=n_rep)"""
+        tokens, n_kv_heads, head_dim = x.shape
+        return (x[:, :,
+                  None, :].expand(tokens, n_kv_heads, n_rep,
+                                  head_dim).reshape(tokens, n_kv_heads * n_rep,
+                                                    head_dim))
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: ROCmFlashAttentionMetadata,
+        kv_scale: float = 1.0,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention and PagedAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        num_tokens, hidden_size = query.shape
+        # Reshape the query, key, and value tensors.
+        query = query.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
+
+        if kv_cache is not None:
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            # Reshape the input keys and values and store them in the cache.
+            # If kv_cache is not provided, the new key and value tensors are
+            # not cached. This happens during the initial memory profiling run.
+            PagedAttention.write_to_paged_cache(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                attn_metadata.slot_mapping,
+                attn_metadata.kv_cache_dtype,
+                kv_scale,
+            )
+
+        if attn_metadata.is_prompt:
+            # Prompt run.
+            if kv_cache is None or attn_metadata.block_tables.numel() == 0:
+                # triton attention
+                # When block_tables are not filled, it means q and k are the
+                # prompt, and they have the same length.
+                if self.use_naive_attn or self.use_triton_flash_attn:
+                    if self.num_kv_heads != self.num_heads:
+                        # Interleave for MQA workaround.
+                        key = self.repeat_kv(key, self.num_queries_per_kv)
+                        value = self.repeat_kv(value, self.num_queries_per_kv)
+                    if self.use_naive_attn:
+                        output = self.attn_fuc(
+                            query,
+                            key,
+                            value,
+                            attn_metadata.prompt_lens,
+                            self.scale,
+                        )
+                    else:
+                        output, _ = self.attn_func(
+                            query,
+                            key,
+                            value,
+                            None,
+                            attn_metadata.seq_start_loc,
+                            attn_metadata.seq_start_loc,
+                            attn_metadata.max_prompt_len,
+                            attn_metadata.max_prompt_len,
+                            True,
+                            self.scale,
+                        )
+                else:
+                    output = self.attn_func(
+                        q=query,
+                        k=key,
+                        v=value,
+                        cu_seqlens_q=attn_metadata.seq_start_loc,
+                        cu_seqlens_k=attn_metadata.seq_start_loc,
+                        max_seqlen_q=attn_metadata.max_prompt_len,
+                        max_seqlen_k=attn_metadata.max_prompt_len,
+                        softmax_scale=self.scale,
+                        causal=True,
+                    )
+
+            else:
+                # prefix-enabled attention
+                output = PagedAttention.forward_prefix(
+                    query,
+                    key,
+                    value,
+                    key_cache,
+                    value_cache,
+                    attn_metadata.block_tables,
+                    attn_metadata.subquery_start_loc,
+                    attn_metadata.prompt_lens_tensor,
+                    attn_metadata.context_lens,
+                    attn_metadata.max_subquery_len,
+                    self.alibi_slopes,
+                )
+        else:
+            # Decoding run.
+            output = PagedAttention.forward_decode(
+                query,
+                key_cache,
+                value_cache,
+                attn_metadata.block_tables,
+                attn_metadata.context_lens,
+                attn_metadata.max_context_len,
+                attn_metadata.kv_cache_dtype,
+                self.num_kv_heads,
+                self.scale,
+                self.alibi_slopes,
+                kv_scale,
+            )
+
+        # Reshape the output tensor.
+        return output.view(num_tokens, hidden_size)
+
+
+def _naive_attention(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    prompt_lens: List[int],
+    scale: float,
+) -> torch.Tensor:
+    num_tokens = query.shape[0]
+    output = torch.empty_like(query)
+    start = 0
+    for _, prompt_len in enumerate(prompt_lens):
+        end = start + prompt_len
+        out = _naive_masked_attention(
+            query[None, start:end],
+            key[None, start:end],
+            value[None, start:end],
+            scale,
+        )
+        # TODO(woosuk): Unnecessary copy. Optimize.
+        output[start:end].copy_(out)
+        start += prompt_len
+
+    # Using view got RuntimeError: view size is not compatible
+    # with input tensor's size and stride (at least one
+    # dimension spans across two contiguous subspaces).
+    # Use reshape instead.
+    return output.reshape(num_tokens, -1)
+
+
+def _naive_masked_attention(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    scale: float,
+) -> torch.Tensor:
+    seq_len, _, _ = query.shape
+    attn_mask = torch.triu(torch.ones(seq_len,
+                                      seq_len,
+                                      dtype=query.dtype,
+                                      device=query.device),
+                           diagonal=1)
+    attn_mask = attn_mask * torch.finfo(query.dtype).min
+
+    attn_weights = scale * torch.einsum("qhd,khd->hqk", query, key).float()
+    attn_weights = attn_weights + attn_mask.float()
+    attn_weights = torch.softmax(attn_weights, dim=-1).to(value.dtype)
+    out = torch.einsum("hqk,khd->qhd", attn_weights, value)
+    return out

vllm/attention/backends/torch_sdpa.py

@@ -0,0 +1,256 @@
+""" Attention layer with torch scaled_dot_product_attention
+    and PagedAttention."""
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Type
+
+import torch
+from torch.nn.functional import scaled_dot_product_attention
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata)
+from vllm.attention.ops.paged_attn import (PagedAttention,
+                                           PagedAttentionMetadata)
+
+
+class TorchSDPABackend(AttentionBackend):
+
+    @staticmethod
+    def get_impl_cls() -> Type["TorchSDPABackendImpl"]:
+        return TorchSDPABackendImpl
+
+    @staticmethod
+    def make_metadata(*args, **kwargs) -> "TorchSDPAMetadata":
+        return TorchSDPAMetadata(*args, **kwargs)
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: Dict[int, List[int]],
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class TorchSDPAMetadata(AttentionMetadata, PagedAttentionMetadata):
+    """Metadata for TorchSDPABackend.
+    """
+    # Currently, input sequences can only contain all prompts
+    # or all decoding. True if all sequences are prompts.
+    is_prompt: bool
+    slot_mapping: torch.Tensor
+    prompt_lens: Optional[List[int]]
+    prompt_lens_tensor: Optional[torch.Tensor]
+    num_prompt_tokens: int
+    num_generation_tokens: int
+
+    max_subquery_len: Optional[int] = None
+    max_prompt_len: Optional[int] = None
+    subquery_start_loc: Optional[torch.Tensor] = None
+    seq_start_loc: Optional[torch.Tensor] = None
+    use_cuda_graph: bool = False
+
+    def __post_init__(self):
+        # Set during the execution of the first attention op.
+        # It is a list because it is needed to set per prompt
+        # when alibi slopes is used. It is because of the limitation
+        # from xformer API.
+        # will not appear in the __repr__ and __init__
+        self.attn_bias: Optional[List[torch.Tensor]] = None
+
+
+class TorchSDPABackendImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: Optional[int] = None,
+        alibi_slopes: Optional[List[float]] = None,
+        sliding_window: Optional[int] = None,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.sliding_window = sliding_window
+        if alibi_slopes is not None:
+            assert len(alibi_slopes) == num_heads
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        self.need_mask = (self.alibi_slopes is not None
+                          or self.sliding_window is not None)
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+        suppored_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in suppored_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {suppored_head_sizes}.")
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: Optional[torch.Tensor],
+        attn_metadata: TorchSDPAMetadata,
+        kv_scale: float,
+    ) -> torch.Tensor:
+        """Forward pass with torch SDPA and PagedAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        num_tokens, hidden_size = query.shape
+        # Reshape the query, key, and value tensors.
+        query = query.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
+
+        if kv_cache is not None:
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+            PagedAttention.write_to_paged_cache(key, value, key_cache,
+                                                value_cache,
+                                                attn_metadata.slot_mapping,
+                                                attn_metadata.kv_cache_dtype,
+                                                kv_scale)
+
+        if attn_metadata.is_prompt:
+            if (kv_cache is None or attn_metadata.block_tables.numel() == 0):
+                if self.num_kv_heads != self.num_heads:
+                    key = key.repeat_interleave(self.num_queries_per_kv, dim=1)
+                    value = value.repeat_interleave(self.num_queries_per_kv,
+                                                    dim=1)
+
+                if attn_metadata.attn_bias is None:
+                    if self.alibi_slopes is not None:
+                        att_masks = _make_alibi_bias(
+                            self.alibi_slopes, query.dtype,
+                            attn_metadata.prompt_lens)  # type: ignore
+                    elif self.sliding_window is not None:
+                        att_masks = _make_sliding_window_bias(
+                            attn_metadata.prompt_lens, self.sliding_window,
+                            query.dtype)  # type: ignore
+                    else:
+                        att_masks = [None] * len(attn_metadata.prompt_lens)
+                    attn_metadata.attn_bias = att_masks
+
+                query = query.movedim(0, query.dim() - 2)
+                key = key.movedim(0, key.dim() - 2)
+                value = value.movedim(0, value.dim() - 2)
+
+                start = 0
+                output = torch.empty(
+                    (num_tokens, self.num_heads, self.head_size),
+                    dtype=query.dtype)
+                for prompt_len, mask in zip(attn_metadata.prompt_lens,
+                                            attn_metadata.attn_bias):
+                    end = start + prompt_len
+                    sub_out = scaled_dot_product_attention(
+                        query[:, start:end, :],
+                        key[:, start:end, :],
+                        value[:, start:end, :],
+                        attn_mask=mask,
+                        dropout_p=0.0,
+                        is_causal=not self.need_mask,
+                        scale=self.scale).movedim(query.dim() - 2, 0)
+                    output[start:end, :, :] = sub_out
+                    start = end
+            else:
+                # prefix-enabled attention
+                raise RuntimeError(
+                    "Torch SDPA backend doesn't support prefix decoding.")
+
+        else:
+            # Decoding run.
+            output = PagedAttention.forward_decode(
+                query,
+                key_cache,
+                value_cache,
+                attn_metadata.block_tables,
+                attn_metadata.context_lens,
+                attn_metadata.max_context_len,
+                attn_metadata.kv_cache_dtype,
+                self.num_kv_heads,
+                self.scale,
+                self.alibi_slopes,
+                kv_scale,
+            )
+
+        # Reshape the output tensor.
+        return output.view(-1, self.num_heads * self.head_size)
+
+
+def _make_alibi_bias(
+    alibi_slopes: torch.Tensor,
+    dtype: torch.dtype,
+    prompt_lens: List[int],
+) -> List[torch.Tensor]:
+    attn_biases = []
+    for prompt_len in prompt_lens:
+        bias = torch.arange(prompt_len, dtype=dtype)
+        # NOTE(zhuohan): HF uses
+        #     `bias = bias[None, :].repeat(prompt_len, 1)`
+        # here. We find that both biases give the same results, but
+        # the bias below more accurately follows the original ALiBi
+        # paper.
+        bias = bias[None, :] - bias[:, None]
+
+        num_heads = alibi_slopes.shape[0]
+        bias = bias[None, :].expand(num_heads, prompt_len, prompt_len)
+        bias.mul_(alibi_slopes[:, None, None])
+        inf_mask = torch.empty(
+            (1, prompt_len, prompt_len),
+            dtype=bias.dtype).fill_(-torch.inf).triu_(diagonal=1)
+        attn_biases.append((bias + inf_mask).to(dtype))
+
+    return attn_biases
+
+
+def _make_sliding_window_bias(
+    prompt_lens: List[int],
+    window_size: Optional[int],
+    dtype: torch.dtype,
+) -> List[torch.Tensor]:
+    attn_biases = []
+    for prompt_len in prompt_lens:
+        tensor = torch.full(
+            (1, prompt_len, prompt_len),
+            dtype=dtype,
+            fill_value=1,
+        )
+        shift = 0
+        mask = torch.tril(tensor, diagonal=shift).to(dtype)  # type: ignore
+        if window_size is not None:
+            mask = torch.triu(mask, diagonal=shift - window_size + 1)
+        mask = torch.log(mask)
+        attn_biases.append(mask.to(dtype))
+
+    return attn_biases

vllm/attention/backends/xformers.py

@@ -1,5 +1,4 @@
 """Attention layer with xFormers and PagedAttention."""
-import importlib
 from dataclasses import dataclass
 from typing import Dict, List, Optional, Tuple, Type
 
@@ -14,7 +13,6 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
 from vllm.attention.ops.paged_attn import (PagedAttention,
                                            PagedAttentionMetadata)
 from vllm.logger import init_logger
-from vllm.utils import is_hip
 
 logger = init_logger(__name__)
 
@@ -166,11 +164,6 @@ class XFormersImpl(AttentionImpl):
                 f"Head size {head_size} is not supported by PagedAttention. "
                 f"Supported head sizes are: {suppored_head_sizes}.")
 
-        # AMD Radeon 7900 series (gfx1100) currently does not support xFormers
-        # nor FlashAttention. As a temporary workaround, we use naive PyTorch
-        # implementation of attention.
-        self.use_naive_attention = _check_use_naive_attention()
-
     def forward(
         self,
         query: torch.Tensor,
@@ -178,6 +171,7 @@ class XFormersImpl(AttentionImpl):
         value: torch.Tensor,
         kv_cache: Optional[torch.Tensor],
         attn_metadata: XFormersMetadata,
+        kv_scale: float,
     ) -> torch.Tensor:
         """Forward pass with xFormers and PagedAttention.
 
@@ -205,7 +199,8 @@ class XFormersImpl(AttentionImpl):
             PagedAttention.write_to_paged_cache(key, value, key_cache,
                                                 value_cache,
                                                 attn_metadata.slot_mapping,
-                                                attn_metadata.kv_cache_dtype)
+                                                attn_metadata.kv_cache_dtype,
+                                                kv_scale)
 
         if attn_metadata.is_prompt:
             # Prompt run.
@@ -231,34 +226,13 @@ class XFormersImpl(AttentionImpl):
                                                   self.num_queries_per_kv,
                                                   value.shape[-1])
 
-                if self.use_naive_attention:
-                    output = torch.empty_like(query)
-                    start = 0
-                    for _, prompt_len in enumerate(attn_metadata.prompt_lens):
-                        end = start + prompt_len
-                        out = _naive_masked_attention(
-                            query[None, start:end],
-                            key[None, start:end],
-                            value[None, start:end],
-                            self.num_heads,
-                            self.num_kv_heads,
-                            self.head_size,
-                            self.scale,
-                        )
-                        # TODO(woosuk): Unnecessary copy. Optimize.
-                        output[start:end].copy_(out)
-                        start += prompt_len
-
-                    # Using view got RuntimeError: view size is not compatible
-                    # with input tensor's size and stride (at least one
-                    # dimension spans across two contiguous subspaces).
-                    # Use reshape instead.
-                    return output.reshape(num_tokens, hidden_size)
-
                 output = self._run_memory_efficient_xformers_forward(
                     query, key, value, attn_metadata)
             else:
                 # prefix-enabled attention
+                # TODO(Hai) this triton kernel has regression issue (broke) to
+                # deal with different data types between KV and FP8 KV cache,
+                # to be addressed separately.
                 output = PagedAttention.forward_prefix(
                     query,
                     key,
@@ -285,6 +259,7 @@ class XFormersImpl(AttentionImpl):
                 self.num_kv_heads,
                 self.scale,
                 self.alibi_slopes,
+                kv_scale,
             )
 
         # Reshape the output tensor.
@@ -323,8 +298,6 @@ class XFormersImpl(AttentionImpl):
                     self.alibi_slopes, self.num_kv_heads, query.dtype,
                     attn_metadata.prompt_lens)
 
-        op = xops.fmha.MemoryEfficientAttentionFlashAttentionOp[0] if (
-            is_hip()) else None
         # No alibi slopes.
         # TODO(woosuk): Too many view operations. Let's try to reduce
         # them in the future for code readability.
@@ -338,8 +311,7 @@ class XFormersImpl(AttentionImpl):
                 value,
                 attn_bias=attn_metadata.attn_bias[0],
                 p=0.0,
-                scale=self.scale,
-                op=op)
+                scale=self.scale)
 
             return out.view_as(query)
 
@@ -357,8 +329,7 @@ class XFormersImpl(AttentionImpl):
                 value[None, start:end],
                 attn_bias=attn_metadata.attn_bias[i],
                 p=0.0,
-                scale=self.scale,
-                op=op)
+                scale=self.scale)
             # TODO(woosuk): Unnecessary copy. Optimize.
             output[start:end].copy_(out.squeeze(0))
             start += prompt_len
@@ -399,42 +370,3 @@ def _make_alibi_bias(
         attn_biases.append(LowerTriangularMaskWithTensorBias(bias))
 
     return attn_biases
-
-
-def _check_use_naive_attention() -> bool:
-    if not is_hip():
-        return False
-    # For ROCm, check whether flash attention is installed or not.
-    use_naive_attention = importlib.util.find_spec("flash_attn") is None
-    if use_naive_attention:
-        logger.warning("flash_attn is not installed. Using naive attention. "
-                       "This will take significantly more GPU memory.")
-        return True
-    return False
-
-
-def _naive_masked_attention(
-    query: torch.Tensor,
-    key: torch.Tensor,
-    value: torch.Tensor,
-    num_heads: int,
-    num_kv_heads: int,
-    head_size: int,
-    scale: float,
-) -> torch.Tensor:
-    query = query.view(-1, num_heads, head_size)
-    key = key.view(-1, num_kv_heads, head_size)
-    value = value.view(-1, num_kv_heads, head_size)
-    seq_len, _, _ = query.shape
-    attn_mask = torch.triu(torch.ones(seq_len,
-                                      seq_len,
-                                      dtype=query.dtype,
-                                      device=query.device),
-                           diagonal=1)
-    attn_mask = attn_mask * torch.finfo(query.dtype).min
-
-    attn_weights = scale * torch.einsum("qhd,khd->hqk", query, key).float()
-    attn_weights = attn_weights + attn_mask.float()
-    attn_weights = torch.softmax(attn_weights, dim=-1).to(value.dtype)
-    out = torch.einsum("hqk,khd->qhd", attn_weights, value)
-    return out

vllm/attention/layer.py

@@ -42,5 +42,7 @@ class Attention(nn.Module):
         value: torch.Tensor,
         kv_cache: Optional[torch.Tensor],
         attn_metadata: AttentionMetadata,
+        kv_scale: float = 1.0,
     ) -> torch.Tensor:
-        return self.impl.forward(query, key, value, kv_cache, attn_metadata)
+        return self.impl.forward(query, key, value, kv_cache, attn_metadata,
+                                 kv_scale)

vllm/attention/ops/paged_attn.py

@@ -73,6 +73,7 @@ class PagedAttention:
         value_cache: torch.Tensor,
         slot_mapping: torch.Tensor,
         kv_cache_dtype: str,
+        kv_scale: float,
     ) -> None:
         cache_ops.reshape_and_cache(
             key,
@@ -81,6 +82,7 @@ class PagedAttention:
             value_cache,
             slot_mapping.flatten(),
             kv_cache_dtype,
+            kv_scale,
         )
 
     @staticmethod
@@ -95,6 +97,7 @@ class PagedAttention:
         num_kv_heads: int,
         scale: float,
         alibi_slopes: Optional[torch.Tensor],
+        kv_scale: float,
     ) -> torch.Tensor:
         output = torch.empty_like(query)
 
@@ -126,6 +129,7 @@ class PagedAttention:
                 max_context_len,
                 alibi_slopes,
                 kv_cache_dtype,
+                kv_scale,
             )
         else:
             # Run PagedAttention V2.
@@ -157,6 +161,7 @@ class PagedAttention:
                 max_context_len,
                 alibi_slopes,
                 kv_cache_dtype,
+                kv_scale,
             )
         return output
 

vllm/attention/ops/triton_flash_attention.py

@@ -0,0 +1,809 @@
+#!/usr/bin/env python
+"""
+Fused Attention
+===============
+
+This is a Triton implementation of the Flash Attention v2 algorithm from Tri Dao
+(https://tridao.me/publications/flash2/flash2.pdf)
+Credits: OpenAI kernel team, AMD ML Frameworks Triton team
+
+Features supported:
+
+1) Fwd with causal masking
+2) Any sequence lengths without padding (currently fwd kernel only)
+3) Support for different sequence lengths for q and k
+4) Nested tensor API currently does not support dropout or bias.
+
+Not currently supported:
+
+1) Non power of two head dims
+
+"""
+
+import torch
+import triton
+import triton.language as tl
+
+torch_dtype: tl.constexpr = torch.float16
+
+
+@triton.jit
+def cdiv_fn(x, y):
+    return (x + y - 1) // y
+
+
+@triton.jit
+def max_fn(x, y):
+    return tl.math.max(x, y)
+
+
+@triton.jit
+def dropout_offsets(philox_seed, philox_offset, dropout_p, m, n, stride):
+    ms = tl.arange(0, m)
+    ns = tl.arange(0, n)
+    return philox_offset + ms[:, None] * stride + ns[None, :]
+
+
+@triton.jit
+def dropout_rng(philox_seed, philox_offset, dropout_p, m, n, stride):
+    rng_offsets = dropout_offsets(philox_seed, philox_offset, dropout_p, m, n,
+                                  stride).to(tl.uint32)
+    # TODO: use tl.randint for better performance
+    return tl.rand(philox_seed, rng_offsets)
+
+
+@triton.jit
+def dropout_mask(philox_seed, philox_offset, dropout_p, m, n, stride):
+    rng_output = dropout_rng(philox_seed, philox_offset, dropout_p, m, n,
+                             stride)
+    rng_keep = rng_output > dropout_p
+    return rng_keep
+
+
+@triton.jit
+def load_fn(block_ptr, first, second, pad):
+    if first and second:
+        tensor = tl.load(block_ptr, boundary_check=(0, 1), padding_option=pad)
+    elif first:
+        tensor = tl.load(block_ptr, boundary_check=(0, ), padding_option=pad)
+    elif second:
+        tensor = tl.load(block_ptr, boundary_check=(1, ), padding_option=pad)
+    else:
+        tensor = tl.load(block_ptr)
+    return tensor
+
+
+@triton.jit
+def _attn_fwd_inner(
+    acc,
+    l_i,
+    m_i,
+    q,
+    K_block_ptr,
+    V_block_ptr,
+    start_m,
+    actual_seqlen_k,
+    dropout_p,
+    philox_seed,
+    batch_philox_offset,
+    encoded_softmax_block_ptr,
+    block_min,
+    block_max,
+    offs_n_causal,
+    masked_blocks,
+    n_extra_tokens,
+    bias_ptr,
+    IS_CAUSAL: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    OFFS_M: tl.constexpr,
+    OFFS_N: tl.constexpr,
+    PRE_LOAD_V: tl.constexpr,
+    MASK_STEPS: tl.constexpr,
+    ENABLE_DROPOUT: tl.constexpr,
+    RETURN_ENCODED_SOFTMAX: tl.constexpr,
+    PADDED_HEAD: tl.constexpr,
+):
+    # loop over k, v, and update accumulator
+    for start_n in range(block_min, block_max, BLOCK_N):
+        # For padded blocks, we will overrun the tensor size if
+        # we load all BLOCK_N. For others, the blocks are all within range.
+        k = load_fn(
+            K_block_ptr,
+            PADDED_HEAD,
+            MASK_STEPS and (n_extra_tokens != 0),
+            "zero",
+        )
+        if PRE_LOAD_V:
+            v = load_fn(
+                V_block_ptr,
+                MASK_STEPS and (n_extra_tokens != 0),
+                PADDED_HEAD,
+                "zero",
+            )
+        qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+        # We start from end of seqlen_k so only the first iteration would need
+        # to be checked for padding if it is not a multiple of block_n
+        # TODO: This can be optimized to only be true for the padded block.
+        if MASK_STEPS:  # noqa: SIM102
+            # If this is the last block / iteration, we want to
+            # mask if the sequence length is not a multiple of block size
+            # a solution is to always do BLOCK_M // BLOCK_N + 1 steps
+            # if not is_modulo_mn. last step might get wasted but that is okay.
+            # check if this masking works for that case.
+            if (start_n + BLOCK_N == block_max) and (n_extra_tokens != 0):
+                boundary_m = tl.full([BLOCK_M],
+                                     actual_seqlen_k,
+                                     dtype=tl.int32)
+                size_n = start_n + OFFS_N[None, :]
+                mask = size_n < boundary_m[:, None]
+                qk = tl.where(mask, qk, float("-inf"))
+        if IS_CAUSAL:
+            causal_boundary = start_n + offs_n_causal
+            causal_mask = OFFS_M[:, None] >= causal_boundary[None, :]
+            qk = tl.where(causal_mask, qk, float("-inf"))
+        # -- compute qk ----
+        qk += tl.dot(q, k)
+        if bias_ptr is not None:
+            bias = load_fn(bias_ptr, False, MASK_STEPS
+                           and (n_extra_tokens != 0), "zero")
+            # While bias is added after multiplying qk with sm_scale, our
+            # optimization to use 2^x instead of e^x results in an additional
+            # scale factor of log2(e) which we must also multiply the bias with.
+            qk += bias * 1.44269504089
+        m_ij = tl.maximum(m_i, tl.max(qk, 1))
+        qk = qk - m_ij[:, None]
+        p = tl.math.exp2(qk)
+
+        # CAVEAT: Must update l_ij before applying dropout
+        l_ij = tl.sum(p, 1)
+        if ENABLE_DROPOUT:
+            philox_offset = (batch_philox_offset +
+                             start_m * BLOCK_M * actual_seqlen_k + start_n -
+                             BLOCK_N)
+            keep = dropout_mask(
+                philox_seed,
+                philox_offset,
+                dropout_p,
+                BLOCK_M,
+                BLOCK_N,
+                actual_seqlen_k,
+            )
+            if RETURN_ENCODED_SOFTMAX:
+                tl.store(
+                    encoded_softmax_block_ptr,
+                    tl.where(keep, p,
+                             -p).to(encoded_softmax_block_ptr.type.element_ty),
+                )
+            p = tl.where(keep, p, 0.0)
+        elif RETURN_ENCODED_SOFTMAX:
+            tl.store(
+                encoded_softmax_block_ptr,
+                p.to(encoded_softmax_block_ptr.type.element_ty),
+            )
+        # -- update output accumulator --
+        alpha = tl.math.exp2(m_i - m_ij)
+        acc = acc * alpha[:, None]
+        if not PRE_LOAD_V:
+            v = load_fn(
+                V_block_ptr,
+                MASK_STEPS and (n_extra_tokens != 0),
+                PADDED_HEAD,
+                "zero",
+            )
+        # -- update m_i and l_i
+        l_i = l_i * alpha + l_ij
+        # update m_i and l_i
+        m_i = m_ij
+        acc += tl.dot(p.to(V_block_ptr.type.element_ty), v)
+        V_block_ptr = tl.advance(V_block_ptr, (BLOCK_N, 0))
+        K_block_ptr = tl.advance(K_block_ptr, (0, BLOCK_N))
+        if bias_ptr is not None:
+            bias_ptr = tl.advance(bias_ptr, (0, BLOCK_N))
+        if RETURN_ENCODED_SOFTMAX:
+            encoded_softmax_block_ptr = tl.advance(encoded_softmax_block_ptr,
+                                                   (0, BLOCK_N))
+    return acc, l_i, m_i
+
+
+@triton.autotune(
+    configs=[
+        triton.Config(
+            {
+                "BLOCK_M": 256,
+                "BLOCK_N": 64,
+                "waves_per_eu": 2,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=8,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 128,
+                "BLOCK_N": 128,
+                "waves_per_eu": 2,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 256,
+                "BLOCK_N": 128,
+                "waves_per_eu": 2,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=8,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 128,
+                "BLOCK_N": 64,
+                "waves_per_eu": 3,
+                "PRE_LOAD_V": True,
+            },
+            num_stages=1,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 128,
+                "BLOCK_N": 64,
+                "waves_per_eu": 3,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 64,
+                "BLOCK_N": 64,
+                "waves_per_eu": 4,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=8,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 32,
+                "BLOCK_N": 32,
+                "waves_per_eu": 4,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=8,
+        ),
+        # TODO: This config fails with head_size not pow2 with data mismatches.
+        #    triton.Config({'BLOCK_M': 32, 'BLOCK_N': 16, 'waves_per_eu': 1,
+        #                   'PRE_LOAD_V': False}, num_stages=1, num_warps=4),
+        triton.Config(
+            {
+                "BLOCK_M": 16,
+                "BLOCK_N": 16,
+                "waves_per_eu": 1,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=4,
+        ),
+    ],
+    key=["hq", "hk", "IS_CAUSAL", "dropout_p", "BLOCK_DMODEL"],
+)
+@triton.jit
+def attn_fwd(
+    Q,
+    K,
+    V,
+    bias,
+    sm_scale,
+    L,
+    Out,
+    stride_qz,
+    stride_qh,
+    stride_qm,
+    stride_qk,
+    stride_kz,
+    stride_kh,
+    stride_kn,
+    stride_kk,
+    stride_vz,
+    stride_vh,
+    stride_vk,
+    stride_vn,
+    stride_oz,
+    stride_oh,
+    stride_om,
+    stride_on,
+    stride_bz,
+    stride_bh,
+    stride_bm,
+    stride_bn,
+    cu_seqlens_q,
+    cu_seqlens_k,
+    dropout_p,
+    philox_seed,
+    philox_offset_base,
+    encoded_softmax,
+    hq,
+    hk,
+    ACTUAL_BLOCK_DMODEL: tl.constexpr,
+    MAX_SEQLENS_Q: tl.constexpr,
+    MAX_SEQLENS_K: tl.constexpr,
+    VARLEN: tl.constexpr,
+    IS_CAUSAL: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    PRE_LOAD_V: tl.constexpr,
+    BIAS_TYPE: tl.constexpr,
+    ENABLE_DROPOUT: tl.constexpr,
+    RETURN_ENCODED_SOFTMAX: tl.constexpr,
+):
+    start_m = tl.program_id(0)
+    off_h_q = tl.program_id(1)
+    off_z = tl.program_id(2)
+    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    offs_n = tl.arange(0, BLOCK_N)
+    if VARLEN:
+        cu_seqlens_q_start = tl.load(cu_seqlens_q + off_z)
+        cu_seqlens_q_end = tl.load(cu_seqlens_q + off_z + 1)
+        seqlen_q = cu_seqlens_q_end - cu_seqlens_q_start
+        # We have a one-size-fits-all grid in id(0). Some seqlens might be too
+        # small for all start_m so for those we return early.
+        if start_m * BLOCK_M > seqlen_q:
+            return
+        cu_seqlens_k_start = tl.load(cu_seqlens_k + off_z)
+        cu_seqlens_k_end = tl.load(cu_seqlens_k + off_z + 1)
+        seqlen_k = cu_seqlens_k_end - cu_seqlens_k_start
+    else:
+        cu_seqlens_q_start = 0
+        cu_seqlens_k_start = 0
+        seqlen_q = MAX_SEQLENS_Q
+        seqlen_k = MAX_SEQLENS_K
+
+    # Now we compute whether we need to exit early due to causal masking.
+    # This is because for seqlen_q > seqlen_k, M rows of the attn scores
+    # are completely masked, resulting in 0s written to the output, and
+    # inf written to LSE. We don't need to do any GEMMs in this case.
+    # This block of code determines what N is, and if this WG is operating
+    # on those M rows.
+    n_blocks = cdiv_fn(seqlen_k, BLOCK_N)
+    if IS_CAUSAL:
+        # If seqlen_q == seqlen_k, the attn scores are a square matrix.
+        # If seqlen_q != seqlen_k, attn scores are rectangular which means
+        # the causal mask boundary is bottom right aligned, and ends at either
+        # the top edge (seqlen_q < seqlen_k) or left edge.
+        # This captures the decrease in n_blocks if we have a rectangular attn
+        # matrix
+        n_blocks_seqlen = cdiv_fn(
+            (start_m + 1) * BLOCK_M + seqlen_k - seqlen_q, BLOCK_N)
+        # This is what adjusts the block_max for the current WG, only
+        # if IS_CAUSAL. Otherwise we want to always iterate through all n_blocks
+        n_blocks = min(n_blocks, n_blocks_seqlen)
+        # If we have no blocks after adjusting for seqlen deltas, this WG is
+        # part of the blocks that are all 0. We exit early.
+        if n_blocks <= 0:
+            o_offset = (off_z * stride_oz + cu_seqlens_q_start * stride_om +
+                        off_h_q * stride_oh)
+            O_block_ptr = tl.make_block_ptr(
+                base=Out + o_offset,
+                shape=(seqlen_q, BLOCK_DMODEL),
+                strides=(stride_om, stride_on),
+                offsets=(start_m * BLOCK_M, 0),
+                block_shape=(BLOCK_M, BLOCK_DMODEL),
+                order=(1, 0),
+            )
+            acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=Out.type.element_ty)
+            # We still need to write 0s to the result
+            # tl.store(O_block_ptr,
+            # acc.to(Out.type.element_ty), boundary_check=(0,1))
+            # l_ptrs = L + off_z * hq * MAX_SEQLENS_Q + off_h_q * MAX_SEQLENS_Q
+            #          + offs_m
+            # We store inf to LSE, not -inf because in the bwd pass,
+            # we subtract this
+            # from qk which makes it -inf, such that exp(qk - inf) = 0
+            # for these masked blocks.
+            # l = tl.full([BLOCK_M], value=float("inf"), dtype=tl.float32)
+            # tl.store(l_ptrs, l)
+            # TODO: Should dropout and return encoded softmax be handled here?
+            return
+
+    is_mqa = hq != hk
+    off_h_k = off_h_q % hk if is_mqa else off_h_q
+    n_extra_tokens = 0
+    if seqlen_k < BLOCK_N:
+        n_extra_tokens = BLOCK_N - seqlen_k
+    elif seqlen_k % BLOCK_N:
+        n_extra_tokens = seqlen_k % BLOCK_N
+    padded_head = ACTUAL_BLOCK_DMODEL != BLOCK_DMODEL
+
+    # Compute pointers for all the tensors used in this kernel.
+    q_offset = (off_z * stride_qz + off_h_q * stride_qh +
+                cu_seqlens_q_start * stride_qm)
+    Q_block_ptr = tl.make_block_ptr(
+        base=Q + q_offset,
+        shape=(seqlen_q, ACTUAL_BLOCK_DMODEL),
+        strides=(stride_qm, stride_qk),
+        offsets=(start_m * BLOCK_M, 0),
+        block_shape=(BLOCK_M, BLOCK_DMODEL),
+        order=(1, 0),
+    )
+    k_offset = (off_z * stride_kz + off_h_k * stride_kh +
+                cu_seqlens_k_start * stride_kn)
+    K_block_ptr = tl.make_block_ptr(
+        base=K + k_offset,
+        shape=(ACTUAL_BLOCK_DMODEL, seqlen_k),
+        strides=(stride_kk, stride_kn),
+        offsets=(0, 0),
+        block_shape=(BLOCK_DMODEL, BLOCK_N),
+        order=(0, 1),
+    )
+    v_offset = (off_z * stride_vz + off_h_k * stride_vh +
+                cu_seqlens_k_start * stride_vk)
+    V_block_ptr = tl.make_block_ptr(
+        base=V + v_offset,
+        shape=(seqlen_k, ACTUAL_BLOCK_DMODEL),
+        strides=(stride_vk, stride_vn),
+        offsets=(0, 0),
+        block_shape=(BLOCK_N, BLOCK_DMODEL),
+        order=(1, 0),
+    )
+    if BIAS_TYPE != 0:
+        bias_ptr = tl.make_block_ptr(
+            base=bias + off_h_q * stride_bh,
+            shape=(seqlen_q, seqlen_k),
+            strides=(stride_bm, stride_bn),
+            offsets=(start_m * BLOCK_M, 0),
+            block_shape=(BLOCK_M, BLOCK_N),
+            order=(1, 0),
+        )
+    else:
+        bias_ptr = None
+    if ENABLE_DROPOUT:
+        batch_philox_offset = philox_offset_base \
+                              + (off_z * hq + off_h_q) \
+                              * seqlen_q * seqlen_k
+    else:
+        batch_philox_offset = 0
+    # We can ask to return the dropout mask without actually doing any dropout.
+    # In this case, we return an invalid pointer so indicate the mask is not i
+    # valid.
+    # TODO: Fix encoded softmax. It currently uses just h_q in the base offset.
+    if RETURN_ENCODED_SOFTMAX:
+        encoded_softmax_block_ptr = tl.make_block_ptr(
+            base=encoded_softmax + off_h_q * seqlen_q * seqlen_k,
+            shape=(seqlen_q, seqlen_k),
+            strides=(seqlen_k, 1),
+            offsets=(start_m * BLOCK_M, 0),
+            block_shape=(BLOCK_M, BLOCK_N),
+            order=(1, 0),
+        )
+    else:
+        encoded_softmax_block_ptr = 0
+    # initialize pointer to m and l
+    m_i = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
+    l_i = tl.full([BLOCK_M], 1.0, dtype=tl.float32)
+    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
+    # scale sm_scale by log_2(e) and use 2^x in the loop as we do not
+    # have native e^x support in HW.
+    qk_scale = sm_scale * 1.44269504089
+    # Q is loaded once at the beginning and shared by all N blocks.
+    q = load_fn(Q_block_ptr, True, padded_head, "zero")
+    q = (q * qk_scale).to(Q_block_ptr.type.element_ty)
+
+    # Here we compute how many full and masked blocks we have.
+    padded_block_k = n_extra_tokens != 0
+    is_modulo_mn = not padded_block_k and (seqlen_q % BLOCK_M == 0)
+    if IS_CAUSAL:
+        # There are always at least BLOCK_M // BLOCK_N masked blocks.
+        # Additionally there might be one more due to dissimilar seqlens.
+        masked_blocks = BLOCK_M // BLOCK_N + (not is_modulo_mn)
+    else:
+        # Padding on Q does not need to be masked in the FA loop.
+        masked_blocks = padded_block_k
+    # if IS_CAUSAL, not is_modulo_mn does not always result in an additional
+    # block. In this case we might exceed n_blocks so pick the min.
+    masked_blocks = min(masked_blocks, n_blocks)
+    n_full_blocks = n_blocks - masked_blocks
+    block_min = 0
+    block_max = n_blocks * BLOCK_N
+    # Compute for full blocks. Here we set causal to false regardless of its
+    # value because there is no masking. Similarly we do not need padding.
+    if n_full_blocks > 0:
+        block_max = (n_blocks - masked_blocks) * BLOCK_N
+        acc, l_i, m_i = _attn_fwd_inner(
+            acc,
+            l_i,
+            m_i,
+            q,
+            K_block_ptr,
+            V_block_ptr,
+            start_m,
+            seqlen_k,
+            dropout_p,
+            philox_seed,
+            batch_philox_offset,
+            encoded_softmax_block_ptr,
+            # _, _, offs_n_causal, masked_blocks, n_extra_tokens, _
+            block_min,
+            block_max,
+            0,
+            0,
+            0,
+            bias_ptr,
+            # IS_CAUSAL, ....
+            False,
+            BLOCK_M,
+            BLOCK_DMODEL,
+            BLOCK_N,
+            offs_m,
+            offs_n,
+            # _, MASK_STEPS, ...
+            PRE_LOAD_V,
+            False,
+            ENABLE_DROPOUT,
+            RETURN_ENCODED_SOFTMAX,
+            padded_head,
+        )
+        block_min = block_max
+        block_max = n_blocks * BLOCK_N
+
+    tl.debug_barrier()
+    # Remaining blocks, if any, are full / not masked.
+    if masked_blocks > 0:
+        offs_n_causal = offs_n + (seqlen_q - seqlen_k) if IS_CAUSAL else 0
+        K_block_ptr = tl.advance(K_block_ptr, (0, n_full_blocks * BLOCK_N))
+        V_block_ptr = tl.advance(V_block_ptr, (n_full_blocks * BLOCK_N, 0))
+        if bias_ptr is not None:
+            bias_ptr = tl.advance(bias_ptr, (0, n_full_blocks * BLOCK_N))
+        if RETURN_ENCODED_SOFTMAX:
+            encoded_softmax_block_ptr = tl.advance(encoded_softmax_block_ptr,
+                                                   (0, n_full_blocks))
+        acc, l_i, m_i = _attn_fwd_inner(
+            acc,
+            l_i,
+            m_i,
+            q,
+            K_block_ptr,
+            V_block_ptr,
+            start_m,
+            seqlen_k,
+            dropout_p,
+            philox_seed,
+            batch_philox_offset,
+            encoded_softmax_block_ptr,
+            block_min,
+            block_max,
+            offs_n_causal,
+            masked_blocks,
+            n_extra_tokens,
+            bias_ptr,
+            IS_CAUSAL,
+            BLOCK_M,
+            BLOCK_DMODEL,
+            BLOCK_N,
+            offs_m,
+            offs_n,
+            # _, MASK_STEPS, ...
+            PRE_LOAD_V,
+            True,
+            ENABLE_DROPOUT,
+            RETURN_ENCODED_SOFTMAX,
+            padded_head,
+        )
+    # epilogue
+    acc = acc / l_i[:, None]
+    if ENABLE_DROPOUT:
+        acc = acc / (1 - dropout_p)
+    # If seqlen_q > seqlen_k but the delta is not a multiple of BLOCK_M,
+    # then we have one block with a row of all NaNs which come from computing
+    # softmax over a row of all -infs (-inf - inf = NaN). We check for that here
+    # and store 0s where there are NaNs as these rows should've been zeroed out.
+    end_m_idx = (start_m + 1) * BLOCK_M
+    start_m_idx = start_m * BLOCK_M
+    causal_start_idx = seqlen_q - seqlen_k
+    acc = acc.to(Out.type.element_ty)
+    if IS_CAUSAL:  # noqa: SIM102
+        if causal_start_idx > start_m_idx and causal_start_idx < end_m_idx:
+            out_mask_boundary = tl.full((BLOCK_DMODEL, ),
+                                        causal_start_idx,
+                                        dtype=tl.int32)
+            mask_m_offsets = start_m_idx + tl.arange(0, BLOCK_M)
+            out_ptrs_mask = (mask_m_offsets[:, None] >=
+                             out_mask_boundary[None, :])
+            z = 0.0
+            acc = tl.where(out_ptrs_mask, acc, z.to(acc.type.element_ty))
+    # write back LSE
+    # l_ptrs = L + off_z * hq * MAX_SEQLENS_Q + off_h_q * MAX_SEQLENS_Q + offs_m
+    # If seqlen_q not multiple of BLOCK_M, we need to mask out the last
+    # few rows. This is only true for the last M block. For others,
+    # overflow_size will be -ve
+    # overflow_size = end_m_idx - seqlen_q
+    # if overflow_size > 0:
+    #    boundary = tl.full((BLOCK_M,), BLOCK_M - overflow_size, dtype=tl.int32)
+    #    # This is a > check because mask being 0 blocks the store.
+    #    l_ptrs_mask = boundary > tl.arange(0, BLOCK_M)
+    #    tl.store(l_ptrs, m_i + tl.math.log2(l_i), mask=l_ptrs_mask)
+    # else:
+    #    tl.store(l_ptrs, m_i + tl.math.log2(l_i))
+
+    # write back O
+    o_offset = (off_z * stride_oz + cu_seqlens_q_start * stride_om +
+                off_h_q * stride_oh)
+    O_block_ptr = tl.make_block_ptr(
+        base=Out + o_offset,
+        shape=(seqlen_q, ACTUAL_BLOCK_DMODEL),
+        strides=(stride_om, stride_on),
+        offsets=(start_m * BLOCK_M, 0),
+        block_shape=(BLOCK_M, BLOCK_DMODEL),
+        order=(1, 0),
+    )
+    # Need boundary check on this to make sure the padding from the
+    # Q and KV tensors in both dims are not part of what we store back.
+    # TODO: Do the boundary check optionally.
+    tl.store(O_block_ptr, acc, boundary_check=(0, 1))
+
+
+def check_args(
+    q,
+    k,
+    v,
+    o,
+    varlen=True,
+    max_seqlens=None,
+    cu_seqlens_q=None,
+    cu_seqlens_k=None,
+):
+    assert q.dim() == k.dim() and q.dim() == v.dim()
+    if varlen:
+        assert q.dim() == 3
+        total_q, nheads_q, head_size = q.shape
+        total_k, nheads_k, _ = k.shape
+        assert cu_seqlens_q is not None
+        assert cu_seqlens_k is not None
+        assert len(cu_seqlens_q) == len(cu_seqlens_k)
+    else:
+        assert q.dim() == 4
+        batch, nheads_q, seqlen_q, head_size = q.shape
+        _, nheads_k, seqlen_k, _ = k.shape
+        assert max_seqlens > 0
+    assert k.shape == v.shape
+    assert q.shape[-1] == k.shape[-1] and q.shape[-1] == v.shape[-1]
+    # TODO: Change assert if we support qkl f8 and v f16
+    assert q.dtype == k.dtype and q.dtype == v.dtype
+    # TODO: Fix assert to check head size <=256 once supported
+    assert head_size <= 128
+    assert o.shape == q.shape
+    assert (nheads_q % nheads_k) == 0
+
+
+class _attention(torch.autograd.Function):
+
+    @staticmethod
+    def forward(
+        ctx,
+        q,
+        k,
+        v,
+        o,
+        cu_seqlens_q,
+        cu_seqlens_k,
+        max_seqlens_q,
+        max_seqlens_k,
+        causal=False,
+        sm_scale=1.0,
+        bias=None,
+    ):
+        if o is None:
+            o = torch.empty_like(q, dtype=v.dtype)
+
+        check_args(
+            q,
+            k,
+            v,
+            o,
+            varlen=True,
+            cu_seqlens_q=cu_seqlens_q,
+            cu_seqlens_k=cu_seqlens_k,
+        )
+        if True:  # varlen
+            total_q, nheads_q, head_size = q.shape
+            total_k, nheads_k, _ = k.shape
+            batch = len(cu_seqlens_q) - 1
+            q_strides = (0, q.stride(1), q.stride(0), q.stride(2))
+            k_strides = (0, k.stride(1), k.stride(0), k.stride(2))
+            v_strides = (0, v.stride(1), v.stride(0), v.stride(2))
+            o_strides = (0, o.stride(1), o.stride(0), o.stride(2))
+        else:
+            batch, seqlen_q, nheads_q, head_size = q.shape
+            _, seqlen_k, nheads_k, _ = k.shape
+            q_strides = (q.stride(0), q.stride(2), q.stride(1), q.stride(3))
+            k_strides = (k.stride(0), k.stride(2), k.stride(1), k.stride(3))
+            v_strides = (v.stride(0), v.stride(2), v.stride(1), v.stride(3))
+            o_strides = (o.stride(0), o.stride(2), o.stride(1), o.stride(3))
+
+        # Get closest power of 2 over or equal to 32.
+        unpadded_head_dims = {32, 64, 128}
+        if head_size not in unpadded_head_dims:
+            padded_d_model = None
+            for i in unpadded_head_dims:
+                if i > head_size:
+                    padded_d_model = i
+                    break
+            assert padded_d_model is not None
+        else:
+            padded_d_model = head_size
+
+        grid = lambda META: (
+            triton.cdiv(max_seqlens_q, META["BLOCK_M"]),
+            nheads_q,
+            batch,
+        )
+
+        encoded_softmax = None
+
+        # Seed the RNG so we get reproducible results for testing.
+        philox_seed = 0x1BF52
+        philox_offset = 0x1D4B42
+
+        if bias is not None:
+            bias_strides = (
+                bias.stride(0),
+                bias.stride(1),
+                bias.stride(2),
+                bias.stride(3),
+            )
+        else:
+            bias_strides = (0, 0, 0, 0)
+
+        attn_fwd[grid](
+            q,
+            k,
+            v,
+            bias,
+            sm_scale,
+            None,
+            o,
+            *q_strides,
+            *k_strides,
+            *v_strides,
+            *o_strides,
+            *bias_strides,
+            cu_seqlens_q,
+            cu_seqlens_k,
+            dropout_p=0.0,
+            philox_seed=philox_seed,
+            philox_offset_base=philox_offset,
+            encoded_softmax=encoded_softmax,
+            hq=nheads_q,
+            hk=nheads_k,
+            ACTUAL_BLOCK_DMODEL=head_size,
+            MAX_SEQLENS_Q=max_seqlens_q,
+            MAX_SEQLENS_K=max_seqlens_k,
+            IS_CAUSAL=causal,
+            VARLEN=True,
+            BLOCK_DMODEL=padded_d_model,
+            BIAS_TYPE=0 if bias is None else 1,
+            ENABLE_DROPOUT=False,
+            RETURN_ENCODED_SOFTMAX=False,
+        )
+
+        ctx.grid = grid
+        ctx.sm_scale = sm_scale
+        ctx.BLOCK_DMODEL = head_size
+        ctx.causal = causal
+        ctx.dropout_p = 0.0
+        ctx.philox_seed = philox_seed
+        ctx.philox_offset = philox_offset
+        ctx.encoded_softmax = encoded_softmax
+        ctx.return_encoded_softmax = False
+        return o, encoded_softmax
+
+
+triton_attention = _attention.apply

vllm/attention/selector.py

@@ -1,3 +1,4 @@
+import enum
 from functools import lru_cache
 from typing import Type
 
@@ -5,50 +6,81 @@ import torch
 
 from vllm.attention.backends.abstract import AttentionBackend
 from vllm.logger import init_logger
-from vllm.utils import is_hip
+from vllm.utils import is_cpu, is_hip, is_tpu
 
 logger = init_logger(__name__)
 
 
+class _Backend(enum.Enum):
+    FLASH_ATTN = enum.auto()
+    XFORMERS = enum.auto()
+    ROCM_FLASH = enum.auto()
+    TORCH_SDPA = enum.auto()
+    PALLAS = enum.auto()
+
+
 @lru_cache(maxsize=None)
 def get_attn_backend(dtype: torch.dtype) -> Type[AttentionBackend]:
-    if True:
-        logger.info("Using PallasAttention backend.")
-        from vllm.attention.backends.pallas import (  # noqa: F401
-            PallasAttentionBackend)
-        return PallasAttentionBackend
-    elif _can_use_flash_attn(dtype):
+    backend = _which_attn_to_use(dtype)
+    if backend == _Backend.FLASH_ATTN:
         logger.info("Using FlashAttention backend.")
         from vllm.attention.backends.flash_attn import (  # noqa: F401
             FlashAttentionBackend)
         return FlashAttentionBackend
-    else:
+    elif backend == _Backend.XFORMERS:
         logger.info("Using XFormers backend.")
         from vllm.attention.backends.xformers import (  # noqa: F401
             XFormersBackend)
         return XFormersBackend
+    elif backend == _Backend.ROCM_FLASH:
+        logger.info("Using ROCmFlashAttention backend.")
+        from vllm.attention.backends.rocm_flash_attn import (  # noqa: F401
+            ROCmFlashAttentionBackend)
+        return ROCmFlashAttentionBackend
+    elif backend == _Backend.TORCH_SDPA:
+        logger.info("Using Torch SDPA backend.")
+        from vllm.attention.backends.torch_sdpa import TorchSDPABackend
+        return TorchSDPABackend
+    elif backend == _Backend.PALLAS:
+        logger.info("Using PallasAttention backend.")
+        from vllm.attention.backends.pallas import (  # noqa: F401
+            PallasAttentionBackend)
+        return PallasAttentionBackend
+    else:
+        raise ValueError("Invalid attention backend.")
 
 
-def _can_use_flash_attn(dtype: torch.dtype) -> bool:
+def _which_attn_to_use(dtype: torch.dtype) -> _Backend:
+    """Returns which flash attention backend to use."""
+    if is_tpu():
+        return _Backend.PALLAS
+    if is_cpu():
+        return _Backend.TORCH_SDPA
+
     if is_hip():
         # AMD GPUs.
-        logger.info("Cannot use FlashAttention backend for AMD GPUs.")
-        return False
+        if torch.cuda.get_device_capability()[0] != 9:
+            # not Instinct series GPUs.
+            logger.info("flash_atten is not supported on NAVI GPUs.")
+        return _Backend.ROCM_FLASH
+
+    # NVIDIA GPUs.
     if torch.cuda.get_device_capability()[0] < 8:
         # Volta and Turing NVIDIA GPUs.
         logger.info("Cannot use FlashAttention backend for Volta and Turing "
                     "GPUs.")
-        return False
+        return _Backend.XFORMERS
+
     if dtype not in (torch.float16, torch.bfloat16):
         logger.info("Cannot use FlashAttention backend for dtype other than "
                     "torch.float16 or torch.bfloat16.")
-        return False
+        return _Backend.XFORMERS
 
     try:
         import flash_attn  # noqa: F401
     except ImportError:
         logger.info(
-            "Cannot use FlashAttention because the package is not found. "
-            "Please install it for better performance.")
-        return False
-    return True
+            "Cannot use FlashAttention backend because the flash_attn package "
+            "is not found. Please install it for better performance.")
+        return _Backend.XFORMERS
+    return _Backend.FLASH_ATTN