Merge branch 'upstream' into openai-logprobs

.buildkite/nightly-benchmarks/kickoff-pipeline.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/env bash
+
+set -euo pipefail
+
+# Install system packages
+apt update
+apt install -y curl jq
+
+# Install minijinja for templating
+curl -sSfL https://github.com/mitsuhiko/minijinja/releases/latest/download/minijinja-cli-installer.sh | sh
+source $HOME/.cargo/env
+
+# If BUILDKITE_PULL_REQUEST != "false", then we check the PR labels using curl and jq
+if [ "$BUILDKITE_PULL_REQUEST" != "false" ]; then
+  PR_LABELS=$(curl -s "https://api.github.com/repos/vllm-project/vllm/pulls/$BUILDKITE_PULL_REQUEST" | jq -r '.labels[].name')
+
+  if [[ $PR_LABELS == *"perf-benchmarks"* ]]; then
+    echo "This PR has the 'perf-benchmarks' label. Proceeding with the nightly benchmarks."
+  else
+    echo "This PR does not have the 'perf-benchmarks' label. Skipping the nightly benchmarks."
+    exit 0
+  fi
+fi
+
+# Upload sample.yaml
+buildkite-agent pipeline upload .buildkite/nightly-benchmarks/sample.yaml

.buildkite/nightly-benchmarks/sample.yaml

@@ -0,0 +1,39 @@
+steps:
+  # NOTE(simon): You can create separate blocks for different jobs
+  - label: "A100: NVIDIA SMI"
+    agents:
+      queue: A100
+    plugins:
+    - kubernetes:
+        podSpec:
+          containers:
+          # - image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:$BUILDKITE_COMMIT
+          # TODO(simon): check latest main branch or use the PR image.
+          - image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:45c35f0d58f4508bf43bd6af1d3d0d0ec0c915e6
+            command:
+            - bash -c 'nvidia-smi && nvidia-smi topo -m && pwd && ls'
+            resources:
+              limits:
+                nvidia.com/gpu: 8
+            volumeMounts:
+            - name: devshm
+              mountPath: /dev/shm
+          nodeSelector:
+            nvidia.com/gpu.product: NVIDIA-A100-SXM4-80GB
+          volumes:
+          - name: devshm
+            emptyDir:
+              medium: Memory
+  # TODO(simon): bring H100 online
+  # - label: "H100: NVIDIA SMI"
+  #   agents:
+  #     queue: H100
+  #   plugins:
+  #   - docker#v5.11.0:
+  #       image: us-central1-docker.pkg.dev/vllm-405802/vllm-ci-test-repo/vllm-test:45c35f0d58f4508bf43bd6af1d3d0d0ec0c915e6
+  #       command:
+  #       - bash -c 'nvidia-smi && nvidia-smi topo -m'
+  #       propagate-environment: true
+  #       ipc: host
+  #       gpus: all
+

.buildkite/run-benchmarks.sh

@@ -50,7 +50,7 @@ echo "### Serving Benchmarks" >> benchmark_results.md
 sed -n '1p' benchmark_serving.txt >> benchmark_results.md # first line
 echo "" >> benchmark_results.md
 echo '```' >> benchmark_results.md
-tail -n 20 benchmark_serving.txt >> benchmark_results.md # last 20 lines
+tail -n 24 benchmark_serving.txt >> benchmark_results.md # last 24 lines
 echo '```' >> benchmark_results.md
 
 # if the agent binary is not found, skip uploading the results, exit 0

.buildkite/test-pipeline.yaml

@@ -45,7 +45,8 @@ steps:
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_chunked_prefill_distributed.py
-  - pytest -v -s spec_decode/e2e/test_integration_dist.py 
+  - pytest -v -s spec_decode/e2e/test_integration_dist.py
+  - CUDA_VISIBLE_DEVICES=0,1 pytest -v -s test_sharded_state_loader.py
 
 - label: Distributed Tests (Multiple Groups)
   #mirror_hardwares: [amd]
@@ -93,14 +94,13 @@ steps:
 - label: Models Test
   #mirror_hardwares: [amd]
   commands:
-    - bash ../.buildkite/download-images.sh
-    - pytest -v -s models --ignore=models/test_llava.py
+    - pytest -v -s models -m \"not llava\"
 
 - label: Llava Test
   mirror_hardwares: [amd]
   commands:
     - bash ../.buildkite/download-images.sh
-    - pytest -v -s models/test_llava.py
+    - pytest -v -s models -m llava
 
 - label: Prefix Caching Test
   mirror_hardwares: [amd]
@@ -124,7 +124,10 @@ steps:
 
 - label: Speculative decoding tests
   #mirror_hardwares: [amd]
-  command: pytest -v -s spec_decode
+  commands:
+    # See https://github.com/vllm-project/vllm/issues/5152
+    - export VLLM_ATTENTION_BACKEND=XFORMERS
+    - pytest -v -s spec_decode
 
 - label: LoRA Test %N
   #mirror_hardwares: [amd]
@@ -136,14 +139,7 @@ steps:
   num_gpus: 4
   # This test runs llama 13B, so it is required to run on 4 GPUs.
   commands:
-    # Temporarily run this way because we cannot clean up GPU mem usage
-    # for multi GPU tests.
-    # TODO(sang): Fix it.
-    - pytest -v -s lora/test_long_context.py::test_rotary_emb_replaced
-    - pytest -v -s lora/test_long_context.py::test_batched_rope_kernel
-    - pytest -v -s lora/test_long_context.py::test_self_consistency
-    - pytest -v -s lora/test_long_context.py::test_quality
-    - pytest -v -s lora/test_long_context.py::test_max_len
+    - pytest -v -s -x lora/test_long_context.py
 
 - label: Tensorizer Test
   #mirror_hardwares: [amd]

CMakeLists.txt

@@ -66,19 +66,6 @@ endif()
 #
 find_package(Torch REQUIRED)
 
-#
-# Normally `torch.utils.cpp_extension.CUDAExtension` would add
-# `libtorch_python.so` for linking against an extension. Torch's cmake
-# configuration does not include this library (presumably since the cmake
-# config is used for standalone C++ binaries that link against torch).
-# The `libtorch_python.so` library defines some of the glue code between
-# torch/python via pybind and is required by VLLM extensions for this
-# reason. So, add it by manually with `find_library` using torch's
-# installed library path.
-#
-find_library(torch_python_LIBRARY torch_python PATHS
-  "${TORCH_INSTALL_PREFIX}/lib")
-
 #
 # Forward the non-CUDA device extensions to external CMake scripts.
 #
@@ -171,7 +158,7 @@ set(VLLM_EXT_SRC
   "csrc/quantization/fp8/common.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/moe_align_block_size_kernels.cu"
-  "csrc/pybind.cpp")
+  "csrc/torch_bindings.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
   include(FetchContent)
@@ -218,14 +205,15 @@ define_gpu_extension_target(
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
   INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
+  USE_SABI 3
   WITH_SOABI)
 
 #
 # _moe_C extension
 #
 
 set(VLLM_MOE_EXT_SRC
-  "csrc/moe/moe_ops.cpp"
+  "csrc/moe/torch_bindings.cpp"
   "csrc/moe/topk_softmax_kernels.cu")
 
 define_gpu_extension_target(
@@ -235,6 +223,7 @@ define_gpu_extension_target(
   SOURCES ${VLLM_MOE_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
+  USE_SABI 3
   WITH_SOABI)
 
 #
@@ -249,7 +238,7 @@ set(VLLM_PUNICA_EXT_SRC
   "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
   "csrc/punica/punica_ops.cu"
-  "csrc/punica/punica_pybind.cpp")
+  "csrc/punica/torch_bindings.cpp")
 
 #
 # Copy GPU compilation flags+update for punica
@@ -286,6 +275,7 @@ if (VLLM_PUNICA_GPU_ARCHES)
     SOURCES ${VLLM_PUNICA_EXT_SRC}
     COMPILE_FLAGS ${VLLM_PUNICA_GPU_FLAGS}
     ARCHITECTURES ${VLLM_PUNICA_GPU_ARCHES}
+    USE_SABI 3
     WITH_SOABI)
 else()
   message(WARNING "Unable to create _punica_C target because none of the "

Dockerfile.rocm

@@ -106,9 +106,9 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     pip install -U -r requirements-rocm.txt \
     && patch /opt/rocm/include/hip/amd_detail/amd_hip_bf16.h ./rocm_patch/rocm_bf16.patch \
     && python3 setup.py install \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_C.cpython-39-x86_64-linux-gnu.so vllm/ \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_punica_C.cpython-39-x86_64-linux-gnu.so vllm/ \
-    && cp build/lib.linux-x86_64-cpython-39/vllm/_moe_C.cpython-39-x86_64-linux-gnu.so vllm/ \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_C.abi3.so vllm/ \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_punica_C.abi3.so vllm/ \
+    && cp build/lib.linux-x86_64-cpython-39/vllm/_moe_C.abi3.so vllm/ \
     && cd ..
 
 

README.md

@@ -16,6 +16,13 @@ Easy, fast, and cheap LLM serving for everyone
 
 ---
 
+**Ray Summit CPF is Open (June 4th to June 20th)!**
+
+There will be a track for vLLM at the Ray Summit (09/30-10/02, SF) this year!
+If you have cool projects related to vLLM or LLM inference, we would love to see your proposals.
+This will be a great chance for everyone in the community to get together and learn.
+Please submit your proposal [here](https://raysummit.anyscale.com/flow/anyscale/raysummit2024/landing/page/eventsite)
+
 **The Fourth vLLM Bay Area Meetup (June 11th 5:30pm-8pm PT)**
 
 We are thrilled to announce our fourth vLLM Meetup!
@@ -104,6 +111,7 @@ vLLM is a community project. Our compute resources for development and testing a
 - Dropbox
 - Lambda Lab
 - NVIDIA
+- Sequoia Capital
 - Replicate
 - Roblox
 - RunPod

benchmarks/benchmark_latency.py

@@ -36,7 +36,8 @@ def main(args: argparse.Namespace):
               enable_chunked_prefill=args.enable_chunked_prefill,
               download_dir=args.download_dir,
               block_size=args.block_size,
-              gpu_memory_utilization=args.gpu_memory_utilization)
+              gpu_memory_utilization=args.gpu_memory_utilization,
+              distributed_executor_backend=args.distributed_executor_backend)
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -221,5 +222,12 @@ def run_to_completion(profile_dir: Optional[str] = None):
                         help='the fraction of GPU memory to be used for '
                         'the model executor, which can range from 0 to 1.'
                         'If unspecified, will use the default value of 0.9.')
+    parser.add_argument(
+        '--distributed-executor-backend',
+        choices=['ray', 'mp'],
+        default=None,
+        help='Backend to use for distributed serving. When more than 1 GPU '
+        'is used, will be automatically set to "ray" if installed '
+        'or "mp" (multiprocessing) otherwise.')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_serving.py

@@ -56,6 +56,9 @@ class BenchmarkMetrics:
     mean_tpot_ms: float
     median_tpot_ms: float
     p99_tpot_ms: float
+    mean_itl_ms: float
+    median_itl_ms: float
+    p99_itl_ms: float
 
 
 def sample_sharegpt_requests(
@@ -200,16 +203,24 @@ def calculate_metrics(
     actual_output_lens = []
     total_input = 0
     completed = 0
+    itls = []
     tpots = []
     ttfts = []
     for i in range(len(outputs)):
         if outputs[i].success:
-            output_len = len(tokenizer(outputs[i].generated_text).input_ids)
+            # We use the tokenizer to count the number of output tokens for all
+            # serving backends instead of looking at len(outputs[i].itl) since
+            # multiple output tokens may be bundled together
+            # Note: this may inflate the output token count slightly
+            output_len = len(
+                tokenizer(outputs[i].generated_text,
+                          add_special_tokens=False).input_ids)
             actual_output_lens.append(output_len)
             total_input += input_requests[i][1]
             if output_len > 1:
                 tpots.append(
                     (outputs[i].latency - outputs[i].ttft) / (output_len - 1))
+            itls += outputs[i].itl
             ttfts.append(outputs[i].ttft)
             completed += 1
         else:
@@ -234,6 +245,9 @@ def calculate_metrics(
         mean_tpot_ms=np.mean(tpots or 0) * 1000,
         median_tpot_ms=np.median(tpots or 0) * 1000,
         p99_tpot_ms=np.percentile(tpots or 0, 99) * 1000,
+        mean_itl_ms=np.mean(itls or 0) * 1000,
+        median_itl_ms=np.median(itls or 0) * 1000,
+        p99_itl_ms=np.percentile(itls or 0, 99) * 1000,
     )
 
     return metrics, actual_output_lens
@@ -333,6 +347,10 @@ async def benchmark(
     print("{:<40} {:<10.2f}".format("Median TPOT (ms):",
                                     metrics.median_tpot_ms))
     print("{:<40} {:<10.2f}".format("P99 TPOT (ms):", metrics.p99_tpot_ms))
+    print("{s:{c}^{n}}".format(s='Inter-token Latency', n=50, c='-'))
+    print("{:<40} {:<10.2f}".format("Mean ITL (ms):", metrics.mean_itl_ms))
+    print("{:<40} {:<10.2f}".format("Median ITL (ms):", metrics.median_itl_ms))
+    print("{:<40} {:<10.2f}".format("P99 ITL (ms):", metrics.p99_itl_ms))
     print("=" * 50)
 
     result = {
@@ -349,6 +367,9 @@ async def benchmark(
         "mean_tpot_ms": metrics.mean_tpot_ms,
         "median_tpot_ms": metrics.median_tpot_ms,
         "p99_tpot_ms": metrics.p99_tpot_ms,
+        "mean_itl_ms": metrics.mean_itl_ms,
+        "median_itl_ms": metrics.median_itl_ms,
+        "p99_itl_ms": metrics.p99_itl_ms,
         "input_lens": [output.prompt_len for output in outputs],
         "output_lens": actual_output_lens,
         "ttfts": [output.ttft for output in outputs],

benchmarks/benchmark_throughput.py

@@ -78,6 +78,7 @@ def run_vllm(
     enable_prefix_caching: bool,
     enable_chunked_prefill: bool,
     max_num_batched_tokens: int,
+    distributed_executor_backend: Optional[str],
     gpu_memory_utilization: float = 0.9,
     download_dir: Optional[str] = None,
 ) -> float:
@@ -100,6 +101,7 @@ def run_vllm(
         download_dir=download_dir,
         enable_chunked_prefill=enable_chunked_prefill,
         max_num_batched_tokens=max_num_batched_tokens,
+        distributed_executor_backend=distributed_executor_backend,
     )
 
     # Add the requests to the engine.
@@ -225,8 +227,8 @@ def main(args: argparse.Namespace):
             args.enforce_eager, args.kv_cache_dtype,
             args.quantization_param_path, args.device,
             args.enable_prefix_caching, args.enable_chunked_prefill,
-            args.max_num_batched_tokens, args.gpu_memory_utilization,
-            args.download_dir)
+            args.max_num_batched_tokens, args.distributed_executor_backend,
+            args.gpu_memory_utilization, args.download_dir)
     elif args.backend == "hf":
         assert args.tensor_parallel_size == 1
         elapsed_time = run_hf(requests, args.model, tokenizer, args.n,
@@ -368,6 +370,13 @@ def main(args: argparse.Namespace):
         type=str,
         default=None,
         help='Path to save the throughput results in JSON format.')
+    parser.add_argument(
+        '--distributed-executor-backend',
+        choices=['ray', 'mp'],
+        default=None,
+        help='Backend to use for distributed serving. When more than 1 GPU '
+        'is used, will be automatically set to "ray" if installed '
+        'or "mp" (multiprocessing) otherwise.')
     args = parser.parse_args()
     if args.tokenizer is None:
         args.tokenizer = args.model

benchmarks/kernels/benchmark_moe.py

@@ -255,7 +255,8 @@ def main(args: argparse.Namespace):
 
     if args.batch_size is None:
         batch_sizes = [
-            1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024, 1536, 2048, 3072, 4096
+            1, 2, 4, 8, 16, 24, 32, 48, 64, 96, 128, 256, 512, 1024, 1536,
+            2048, 3072, 4096
         ]
     else:
         batch_sizes = [args.batch_size]