Upstream sync 2024 07 07 (#366)

Upstream sync 2024 07 07
(#355) - ties to v0.5.1 of
upstream. Release candidate

SUMMARY:
* Merge commits from
vllm-project@8e0817c
to
vllm-project@79d406e
* Note that
vllm-project@8e0817c
is NOT included in this merge.

COMPARE vs UPSTREAM:

https://github.com/neuralmagic/nm-vllm/compare/upstream-sync-2024-07-07..79d406e9183aa12cdef6f1876eb9a15385662587

---------

Signed-off-by: Thomas Parnell <[email protected]>
Signed-off-by: Xiaowei Jiang <[email protected]>
Signed-off-by: Muralidhar Andoorveedu <[email protected]>
Signed-off-by: Travis Johnson <[email protected]>
Signed-off-by: Yuan Zhou <[email protected]>
Signed-off-by: Christian Rohmann <[email protected]>
Co-authored-by: Antoni Baum <[email protected]>
Co-authored-by: Nick Hill <[email protected]>
Co-authored-by: James Whedbee <[email protected]>
Co-authored-by: Thomas Parnell <[email protected]>
Co-authored-by: Joshua Rosenkranz <[email protected]>
Co-authored-by: Alexander Matveev <[email protected]>
Co-authored-by: danieljannai21 <[email protected]>
Co-authored-by: xwjiang2010 <[email protected]>
Co-authored-by: Cyrus Leung <[email protected]>
Co-authored-by: Roger Wang <[email protected]>
Co-authored-by: Cyrus Leung <[email protected]>
Co-authored-by: Sirej Dua <[email protected]>
Co-authored-by: Sirej Dua <[email protected]>
Co-authored-by: Murali Andoorveedu <[email protected]>
Co-authored-by: Robert Shaw <rshaw@neuralmagic>
Co-authored-by: Michael Goin <[email protected]>
Co-authored-by: Qubitium-ModelCloud <[email protected]>
Co-authored-by: ZX <[email protected]>
Co-authored-by: Mor Zusman <[email protected]>
Co-authored-by: Erez Schwartz <[email protected]>
Co-authored-by: Mor Zusman <[email protected]>
Co-authored-by: tomeras91 <[email protected]>
Co-authored-by: Tomer Asida <[email protected]>
Co-authored-by: Zhuohan Li <[email protected]>
Co-authored-by: Muralidhar Andoorveedu <[email protected]>
Co-authored-by: youkaichao <[email protected]>
Co-authored-by: Xiaowei Jiang <[email protected]>
Co-authored-by: Roger Wang <[email protected]>
Co-authored-by: SangBin Cho <[email protected]>
Co-authored-by: Travis Johnson <[email protected]>
Co-authored-by: Gregory Shtrasberg <[email protected]>
Co-authored-by: Hongxia Yang <[email protected]>
Co-authored-by: Yuan <[email protected]>
Co-authored-by: Lily Liu <[email protected]>
Co-authored-by: Simon Mo <[email protected]>
Co-authored-by: Christian Rohmann <[email protected]>
Co-authored-by: jvlunteren <[email protected]>
Co-authored-by: JGSweets <[email protected]>

.buildkite/download-images.sh

@@ -8,10 +8,6 @@ set -o pipefail
 # aws s3 sync s3://air-example-data-2/vllm_opensource_llava/ images/
 mkdir -p images
 cd images
-wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/stop_sign_pixel_values.pt
-wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/stop_sign_image_features.pt
-wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/cherry_blossom_pixel_values.pt
-wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/cherry_blossom_image_features.pt
 wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/stop_sign.jpg
 wget https://air-example-data-2.s3.us-west-2.amazonaws.com/vllm_opensource_llava/cherry_blossom.jpg
 

.buildkite/lm-eval-harness/configs/Mixtral-8x22B-Instruct-v0.1-FP8-Dynamic.yaml

@@ -0,0 +1,11 @@
+# bash ./run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Mixtral-8x22B-Instruct-v0.1-FP8-dynamic -b "auto" -l 250 -f 5 -t 8
+model_name: "neuralmagic/Mixtral-8x22B-Instruct-v0.1-FP8-dynamic"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.86
+  - name: "exact_match,flexible-extract"
+    value: 0.86
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Mixtral-8x7B-Instruct-v0.1-FP8.yaml

@@ -0,0 +1,11 @@
+# bash ./run-lm-eval-gsm-vllm-baseline.sh -m neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8 -b "auto" -l 250 -f 5 -t 4
+model_name: "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.624
+  - name: "exact_match,flexible-extract"
+    value: 0.624
+limit: 250
+num_fewshot: 5

.buildkite/lm-eval-harness/configs/Qwen2-57B-A14-Instruct.yaml

@@ -0,0 +1,11 @@
+# bash ./run-lm-eval-gsm-vllm-baseline.sh -m Qwen/Qwen2-57B-A14B-Instruct -b "auto" -l 250 -f 5 -t 4
+model_name: "Qwen/Qwen2-57B-A14B-Instruct"
+tasks:
+- name: "gsm8k"
+  metrics:
+  - name: "exact_match,strict-match"
+    value: 0.792
+  - name: "exact_match,flexible-extract"
+    value: 0.824
+limit: 250
+num_fewshot: 5

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

@@ -1,2 +1,3 @@
 Meta-Llama-3-70B-Instruct.yaml
 Mixtral-8x7B-Instruct-v0.1.yaml
+Qwen2-57B-A14-Instruct.yaml

.buildkite/run-cpu-test.sh

@@ -12,8 +12,10 @@ trap remove_docker_container EXIT
 remove_docker_container
 
 # Run the image
-docker run -itd -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus=48-95 --cpuset-mems=1 --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --name cpu-test cpu-test
-docker run -itd -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus=48-95 --cpuset-mems=1 --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --name cpu-test-avx2 cpu-test-avx2
+docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus=48-95 \
+  --cpuset-mems=1 --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --name cpu-test cpu-test
+docker run -itd --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus=48-95 \
+  --cpuset-mems=1 --network host -e HF_TOKEN --env VLLM_CPU_KVCACHE_SPACE=4 --name cpu-test-avx2 cpu-test-avx2
 
 # offline inference
 docker exec cpu-test bash -c "python3 examples/offline_inference.py"
@@ -23,4 +25,4 @@ docker exec cpu-test-avx2 bash -c "python3 examples/offline_inference.py"
 docker exec cpu-test bash -c "cd tests;
   pip install pytest Pillow protobuf
   cd ../
-  pytest -v -s tests/models -m \"not vlm\" --ignore=tests/models/test_embedding.py --ignore=tests/models/test_registry.py"
+  pytest -v -s tests/models -m \"not vlm\" --ignore=tests/models/test_embedding.py --ignore=tests/models/test_registry.py --ignore=tests/models/test_jamba.py" # Mamba on CPU is not supported

.buildkite/test-pipeline.yaml

@@ -74,6 +74,17 @@ steps:
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
   - pytest -v -s spec_decode/e2e/test_integration_dist_tp4.py
 
+- label: Pipeline Parallelism Test
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 4
+  commands:
+  - TP_SIZE=2 PP_SIZE=2 EAGER_MODE=1 CHUNKED_PREFILL=1 pytest -v -s distributed/test_pipeline_parallel.py
+  - TP_SIZE=2 PP_SIZE=2 EAGER_MODE=1 CHUNKED_PREFILL=0 pytest -v -s distributed/test_pipeline_parallel.py
+  - TP_SIZE=1 PP_SIZE=3 EAGER_MODE=1 CHUNKED_PREFILL=0 pytest -v -s distributed/test_pipeline_parallel.py
+  - PP_SIZE=4 EAGER_MODE=1 CHUNKED_PREFILL=1 pytest -v -s distributed/test_pipeline_parallel.py
+  - PP_SIZE=4 EAGER_MODE=1 CHUNKED_PREFILL=0 pytest -v -s distributed/test_pipeline_parallel.py
+
+
 - label: Engine Test
   mirror_hardwares: [amd]
   command: pytest -v -s engine tokenization test_sequence.py test_config.py test_logger.py
@@ -107,12 +118,15 @@ steps:
 
 - label: Kernels Test %N
   #mirror_hardwares: [amd]
-  command: pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  commands:
+    - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.7/flashinfer-0.0.7+cu121torch2.3-cp310-cp310-linux_x86_64.whl
+    - pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 4
 
 - label: Models Test
   #mirror_hardwares: [amd]
   commands:
+    - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.7/flashinfer-0.0.7+cu121torch2.3-cp310-cp310-linux_x86_64.whl
     - pytest -v -s models -m \"not vlm\"
 
 - label: Vision Language Models Test
@@ -223,7 +237,7 @@ steps:
   - pytest -v -s distributed/test_custom_all_reduce.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
-  - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.5/flashinfer-0.0.5+cu121torch2.3-cp310-cp310-linux_x86_64.whl
+  - pip install https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.7/flashinfer-0.0.7+cu121torch2.3-cp310-cp310-linux_x86_64.whl
   - VLLM_ATTENTION_BACKEND=FLASHINFER TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - VLLM_ATTENTION_BACKEND=FLASHINFER TEST_DIST_MODEL=meta-llama/Meta-Llama-3-8B DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - pytest -v -s -x lora/test_mixtral.py

CMakeLists.txt

@@ -171,6 +171,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     "csrc/quantization/marlin/sparse/marlin_24_cuda_kernel.cu"
     "csrc/quantization/gptq_marlin/gptq_marlin.cu"
     "csrc/quantization/gptq_marlin/gptq_marlin_repack.cu"
+    "csrc/quantization/fp8/fp8_marlin.cu"
     "csrc/custom_all_reduce.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_entry.cu"
     "csrc/quantization/cutlass_w8a8/scaled_mm_c2x.cu"

Dockerfile.cpu

@@ -6,7 +6,13 @@ RUN apt-get update  -y \
     && apt-get install -y git wget vim numactl gcc-12 g++-12 python3 python3-pip libtcmalloc-minimal4 \
     && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-12 10 --slave /usr/bin/g++ g++ /usr/bin/g++-12
 
-RUN echo 'export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:$LD_PRELOAD' >> ~/.bashrc
+# https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/performance_tuning/tuning_guide.html
+# intel-openmp provides additional performance improvement vs. openmp
+# tcmalloc provides better memory allocation efficiency, e.g, holding memory in caches to speed up access of commonly-used objects.
+RUN pip install intel-openmp
+
+ENV LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4:/usr/local/lib/libiomp5.so:$LD_PRELOAD"
+
 
 RUN pip install https://intel-extension-for-pytorch.s3.amazonaws.com/ipex_dev/cpu/intel_extension_for_pytorch-2.3.100%2Bgit0eb3473-cp310-cp310-linux_x86_64.whl
 
@@ -31,4 +37,4 @@ WORKDIR /workspace/
 
 RUN ln -s /workspace/vllm/tests && ln -s /workspace/vllm/examples && ln -s /workspace/vllm/benchmarks
 
-CMD ["/bin/bash"]
+ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]

benchmarks/backend_request_func.py

@@ -225,8 +225,8 @@ async def async_request_openai_completions(
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        "v1/completions"
-    ), "OpenAI Completions API URL must end with 'v1/completions'."
+        "completions"
+    ), "OpenAI Completions API URL must end with 'completions'."
 
     async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
         assert not request_func_input.use_beam_search
@@ -304,8 +304,8 @@ async def async_request_openai_chat_completions(
 ) -> RequestFuncOutput:
     api_url = request_func_input.api_url
     assert api_url.endswith(
-        "v1/chat/completions"
-    ), "OpenAI Chat Completions API URL must end with 'v1/chat/completions'."
+        "chat/completions"
+    ), "OpenAI Chat Completions API URL must end with 'chat/completions'."
 
     async with aiohttp.ClientSession(timeout=AIOHTTP_TIMEOUT) as session:
         assert not request_func_input.use_beam_search

benchmarks/benchmark_latency.py

@@ -46,6 +46,7 @@ def main(args: argparse.Namespace):
         load_format=args.load_format,
         distributed_executor_backend=args.distributed_executor_backend,
         otlp_traces_endpoint=args.otlp_traces_endpoint,
+        enable_prefix_caching=args.enable_prefix_caching,
     )
 
     sampling_params = SamplingParams(
@@ -220,6 +221,9 @@ def run_to_completion(profile_dir: Optional[str] = None):
         action='store_true',
         help='If True, the prefill requests can be chunked based on the '
         'max_num_batched_tokens')
+    parser.add_argument("--enable-prefix-caching",
+                        action='store_true',
+                        help="Enable automatic prefix caching")
     parser.add_argument('--use-v2-block-manager', action='store_true')
     parser.add_argument(
         "--ray-workers-use-nsight",

csrc/ops.h

@@ -93,6 +93,11 @@ torch::Tensor gptq_marlin_repack(torch::Tensor& b_q_weight, torch::Tensor& perm,
                                  int64_t size_k, int64_t size_n,
                                  int64_t num_bits);
 
+torch::Tensor fp8_marlin_gemm(torch::Tensor& a, torch::Tensor& b_q_weight,
+                              torch::Tensor& b_scales, torch::Tensor& workspace,
+                              int64_t num_bits, int64_t size_m, int64_t size_n,
+                              int64_t size_k);
+
 bool cutlass_scaled_mm_supports_fp8(int64_t cuda_device_capability);
 
 void cutlass_scaled_mm(torch::Tensor& out, torch::Tensor const& a,