Merge branch 'main' into omer/run-loader

.buildkite/run-gh200-test.sh

@@ -0,0 +1,25 @@
+#!/bin/bash
+
+# This script build the GH200 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_BUILDKIT=1 docker build . \
+  --target vllm-openai \
+  --platform "linux/arm64" \
+  -t gh200-test \
+  --build-arg max_jobs=66 \
+  --build-arg nvcc_threads=2 \
+  --build-arg torch_cuda_arch_list="9.0+PTX" \
+  --build-arg vllm_fa_cmake_gpu_arches="90-real"
+
+# Setup cleanup
+remove_docker_container() { docker rm -f gh200-test || true; }
+trap remove_docker_container EXIT
+remove_docker_container
+
+# Run the image and test offline inference
+docker run --name gh200-test --gpus=all --entrypoint="" gh200-test bash -c '
+    python3 examples/offline_inference.py
+'

.buildkite/test-pipeline.yaml

@@ -181,14 +181,14 @@ steps:
   commands:
     - VLLM_USE_V1=1 pytest -v -s v1
 
-- label: Examples Test # 15min
+- label: Examples Test # 25min
   working_dir: "/vllm-workspace/examples"
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/entrypoints
   - examples/
   commands:
-    - pip install awscli tensorizer # for llava example and tensorizer test
+    - pip install tensorizer # for tensorizer test
     - python3 offline_inference.py
     - python3 cpu_offload.py
     - python3 offline_inference_chat.py
@@ -198,10 +198,13 @@ steps:
     - python3 offline_inference_vision_language_multi_image.py
     - python3 tensorize_vllm_model.py --model facebook/opt-125m serialize --serialized-directory /tmp/ --suffix v1 && python3 tensorize_vllm_model.py --model facebook/opt-125m deserialize --path-to-tensors /tmp/vllm/facebook/opt-125m/v1/model.tensors
     - python3 offline_inference_encoder_decoder.py
-    - python3 offline_profile.py --model facebook/opt-125m
+    - python3 offline_inference_classification.py
+    - python3 offline_inference_embedding.py
+    - python3 offline_inference_scoring.py
+    - python3 offline_profile.py --model facebook/opt-125m run_num_steps --num-steps 2
 
 - label: Prefix Caching Test # 9min
-  #mirror_hardwares: [amd]
+  mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
   - tests/prefix_caching
@@ -321,7 +324,7 @@ steps:
 
 #####  models test  #####
 
-- label: Basic Models Test # 30min
+- label: Basic Models Test # 24min
   source_file_dependencies:
   - vllm/
   - tests/models
@@ -331,7 +334,7 @@ steps:
     - pytest -v -s models/test_registry.py
     - pytest -v -s models/test_initialization.py
 
-- label: Language Models Test (Standard) # 42min
+- label: Language Models Test (Standard) # 32min
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
@@ -342,7 +345,7 @@ steps:
     - pytest -v -s models/decoder_only/language -m 'core_model or quant_model'
     - pytest -v -s models/embedding/language -m core_model
 
-- label: Language Models Test (Extended) # 50min
+- label: Language Models Test (Extended) # 1h10min
   optional: true
   source_file_dependencies:
   - vllm/
@@ -353,7 +356,7 @@ steps:
     - pytest -v -s models/decoder_only/language -m 'not core_model and not quant_model'
     - pytest -v -s models/embedding/language -m 'not core_model'
 
-- label: Multi-Modal Models Test (Standard) # 26min
+- label: Multi-Modal Models Test (Standard) # 28min
   #mirror_hardwares: [amd]
   source_file_dependencies:
   - vllm/
@@ -369,7 +372,7 @@ steps:
     - pytest -v -s models/encoder_decoder/language -m core_model
     - pytest -v -s models/encoder_decoder/vision_language -m core_model
 
-- label: Multi-Modal Models Test (Extended) # 1h15m
+- label: Multi-Modal Models Test (Extended) 1 # 1h16m
   optional: true
   source_file_dependencies:
   - vllm/
@@ -380,14 +383,24 @@ steps:
   commands:
     - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
     - pytest -v -s models/decoder_only/audio_language -m 'not core_model and not quant_model'
+    - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=0) and not core_model and not quant_model'
     # HACK - run phi3v tests separately to sidestep this transformers bug
     # https://github.com/huggingface/transformers/issues/34307
     - pytest -v -s models/decoder_only/vision_language/test_phi3v.py
-    - pytest -v -s --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
+    - pytest -v -s --ignore models/decoder_only/vision_language/test_models.py --ignore models/decoder_only/vision_language/test_phi3v.py models/decoder_only/vision_language -m 'not core_model and not quant_model'
     - pytest -v -s models/embedding/vision_language -m 'not core_model'
     - pytest -v -s models/encoder_decoder/language -m 'not core_model'
     - pytest -v -s models/encoder_decoder/vision_language -m 'not core_model'
 
+- label: Multi-Modal Models Test (Extended) 2 # 38m
+  optional: true
+  source_file_dependencies:
+  - vllm/
+  - tests/models/decoder_only/vision_language
+  commands:
+    - pip install git+https://github.com/TIGER-AI-Lab/Mantis.git
+    - pytest -v -s models/decoder_only/vision_language/test_models.py -m 'split(group=1) and not core_model and not quant_model'
+
 # This test is used only in PR development phase to test individual models and should never run on main
 - label: Custom Models Test
   optional: true
@@ -422,11 +435,11 @@ steps:
   - tests/distributed/
   commands:
   - # the following commands are for the first node, with ip 192.168.10.10 (ray environment already set up)
-    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep -q 'Same node test passed'
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed'
     - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_multi_node_assignment.py
     - VLLM_MULTI_NODE=1 pytest -v -s distributed/test_pipeline_parallel.py
   - # the following commands are for the second node, with ip 192.168.10.11 (ray environment already set up)
-    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep -q 'Same node test passed'
+    - VLLM_TEST_SAME_HOST=0 torchrun --nnodes 2 --nproc-per-node=2 --rdzv_backend=c10d --rdzv_endpoint=192.168.10.10 distributed/test_same_node.py | grep 'Same node test passed'
 
 - label: Distributed Tests (2 GPUs) # 40min
   #mirror_hardwares: [amd]
@@ -445,12 +458,12 @@ steps:
   commands:
   - pytest -v -s ./compile/test_basic_correctness.py
   - pytest -v -s ./compile/test_wrapper.py
-  - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep -q 'Same node test passed'
-  - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m distributed_2_gpus
+  - VLLM_TEST_SAME_HOST=1 torchrun --nproc-per-node=4 distributed/test_same_node.py | grep 'Same node test passed'
+  - TARGET_TEST_SUITE=L4 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
   # Avoid importing model tests that cause CUDA reinitialization error
-  - pytest models/encoder_decoder/language/test_bart.py -v -s -m distributed_2_gpus
-  - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m distributed_2_gpus
-  - pytest models/decoder_only/vision_language/test_models.py -v -s -m distributed_2_gpus
+  - pytest models/encoder_decoder/language/test_bart.py -v -s -m 'distributed(num_gpus=2)'
+  - pytest models/encoder_decoder/vision_language/test_broadcast.py -v -s -m 'distributed(num_gpus=2)'
+  - pytest models/decoder_only/vision_language/test_models.py -v -s -m 'distributed(num_gpus=2)'
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp2.py
   - pip install -e ./plugins/vllm_add_dummy_model
   - pytest -v -s distributed/test_distributed_oot.py
@@ -540,7 +553,7 @@ steps:
   # see https://github.com/vllm-project/vllm/pull/5689 for details
   - pytest -v -s distributed/test_custom_all_reduce.py
   - torchrun --nproc_per_node=2 distributed/test_ca_buffer_sharing.py
-  - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m distributed_2_gpus
+  - TARGET_TEST_SUITE=A100 pytest basic_correctness/ -v -s -m 'distributed(num_gpus=2)'
   - pytest -v -s -x lora/test_mixtral.py
 
 - label: LM Eval Large Models # optional

CMakeLists.txt

@@ -196,6 +196,7 @@ set(VLLM_EXT_SRC
   "csrc/quantization/gptq/q_gemm.cu"
   "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
   "csrc/quantization/fp8/common.cu"
+  "csrc/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
   "csrc/quantization/gguf/gguf_kernel.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/prepare_inputs/advance_step.cu"
@@ -300,7 +301,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   #
   # For the cutlass_scaled_mm kernels we want to build the c2x (CUTLASS 2.x)
   # kernels for the remaining archs that are not already built for 3x.
-  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS 
+  cuda_archs_loose_intersection(SCALED_MM_2X_ARCHS
     "7.5;8.0;8.6;8.7;8.9;9.0" "${CUDA_ARCHS}")
   # subtract out the archs that are already built for 3x
   list(REMOVE_ITEM SCALED_MM_2X_ARCHS ${SCALED_MM_3X_ARCHS})

Dockerfile

@@ -11,6 +11,7 @@ ARG CUDA_VERSION=12.4.1
 FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu20.04 AS base
 ARG CUDA_VERSION=12.4.1
 ARG PYTHON_VERSION=3.12
+ARG TARGETPLATFORM
 ENV DEBIAN_FRONTEND=noninteractive
 
 # Install Python and other dependencies
@@ -46,9 +47,14 @@ WORKDIR /workspace
 # install build and runtime dependencies
 COPY requirements-common.txt requirements-common.txt
 COPY requirements-cuda.txt requirements-cuda.txt
+COPY requirements-cuda-arm64.txt requirements-cuda-arm64.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install -r requirements-cuda.txt
 
+RUN --mount=type=cache,target=/root/.cache/pip \
+    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
+        python3 -m pip install -r requirements-cuda-arm64.txt; \
+    fi
 
 # cuda arch list used by torch
 # can be useful for both `dev` and `test`
@@ -63,13 +69,19 @@ ENV VLLM_FA_CMAKE_GPU_ARCHES=${vllm_fa_cmake_gpu_arches}
 
 #################### WHEEL BUILD IMAGE ####################
 FROM base AS build
+ARG TARGETPLATFORM
 
 # install build dependencies
 COPY requirements-build.txt requirements-build.txt
 
 RUN --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install -r requirements-build.txt
 
+RUN --mount=type=cache,target=/root/.cache/pip \
+    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
+        python3 -m pip install -r requirements-cuda-arm64.txt; \
+    fi
+
 COPY . .
 ARG GIT_REPO_CHECK=0
 RUN --mount=type=bind,source=.git,target=.git \
@@ -134,15 +146,18 @@ COPY requirements-test.txt requirements-test.txt
 COPY requirements-dev.txt requirements-dev.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install -r requirements-dev.txt
-
 #################### DEV IMAGE ####################
+
 #################### vLLM installation IMAGE ####################
 # image with vLLM installed
 FROM nvidia/cuda:${CUDA_VERSION}-base-ubuntu22.04 AS vllm-base
 ARG CUDA_VERSION=12.4.1
 ARG PYTHON_VERSION=3.12
 WORKDIR /vllm-workspace
 ENV DEBIAN_FRONTEND=noninteractive
+ARG TARGETPLATFORM
+
+COPY requirements-cuda-arm64.txt requirements-cuda-arm64.txt
 
 RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
     echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
@@ -168,18 +183,25 @@ RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
 # or future versions of triton.
 RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 
-# install vllm wheel first, so that torch etc will be installed
+# 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 \
     python3 -m pip install dist/*.whl --verbose
 
 RUN --mount=type=cache,target=/root/.cache/pip \
-    . /etc/environment && \
-    python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.6/flashinfer-0.1.6+cu121torch2.4-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl
+    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
+        pip uninstall -y torch && \
+        python3 -m pip install -r requirements-cuda-arm64.txt; \
+    fi
+
+RUN --mount=type=cache,target=/root/.cache/pip \
+. /etc/environment && \
+if [ "$TARGETPLATFORM" != "linux/arm64" ]; then \
+    python3 -m pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.6/flashinfer-0.1.6+cu121torch2.4-cp${PYTHON_VERSION_STR}-cp${PYTHON_VERSION_STR}-linux_x86_64.whl; \
+fi
 COPY examples examples
 #################### vLLM installation IMAGE ####################
 
-
 #################### TEST IMAGE ####################
 # image to run unit testing suite
 # note that this uses vllm installed by `pip`
@@ -209,7 +231,6 @@ COPY vllm/v1 /usr/local/lib/python3.12/dist-packages/vllm/v1
 RUN mkdir test_docs
 RUN mv docs test_docs/
 RUN mv vllm test_docs/
-
 #################### TEST IMAGE ####################
 
 #################### OPENAI API SERVER ####################
@@ -218,8 +239,11 @@ FROM vllm-base AS vllm-openai
 
 # install additional dependencies for openai api server
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.44.0' timm==0.9.10 boto3 runai-model-streamer runai-model-streamer[s3]
-
+    if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
+        pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.42.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
+    else \
+        pip install accelerate hf_transfer 'modelscope!=1.15.0' 'bitsandbytes>=0.45.0' 'timm==0.9.10' boto3 runai-model-streamer runai-model-streamer[s3]; \
+    fi
 ENV VLLM_USAGE_SOURCE production-docker-image
 
 ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

README.md

@@ -134,3 +134,7 @@ If you use vLLM for your research, please cite our [paper](https://arxiv.org/abs
 * For coordinating contributions and development, please use Slack.
 * For security disclosures, please use Github's security advisory feature.
 * For collaborations and partnerships, please contact us at vllm-questions AT lists.berkeley.edu.
+
+## Media Kit
+
+* If you wish to use vLLM's logo, please refer to [our media kit repo](https://github.com/vllm-project/media-kit).

benchmarks/benchmark_serving.py

@@ -781,6 +781,7 @@ def main(args: argparse.Namespace):
     backend = args.backend
     model_id = args.model
     tokenizer_id = args.tokenizer if args.tokenizer is not None else args.model
+    tokenizer_mode = args.tokenizer_mode
 
     if args.base_url is not None:
         api_url = f"{args.base_url}{args.endpoint}"
@@ -790,6 +791,7 @@ def main(args: argparse.Namespace):
         base_url = f"http://{args.host}:{args.port}"
 
     tokenizer = get_tokenizer(tokenizer_id,
+                              tokenizer_mode=tokenizer_mode,
                               trust_remote_code=args.trust_remote_code)
 
     if args.dataset is not None:
@@ -1210,5 +1212,15 @@ def main(args: argparse.Namespace):
         "from the sampled HF dataset.",
     )
 
+    parser.add_argument(
+        '--tokenizer-mode',
+        type=str,
+        default="auto",
+        choices=['auto', 'slow', 'mistral'],
+        help='The tokenizer mode.\n\n* "auto" will use the '
+        'fast tokenizer if available.\n* "slow" will '
+        'always use the slow tokenizer. \n* '
+        '"mistral" will always use the `mistral_common` tokenizer.')
+
     args = parser.parse_args()
     main(args)