Merge branch 'main' of github.com:cduk/vllm-pascal

.buildkite/check-wheel-size.py

@@ -1,7 +1,7 @@
 import os
 import zipfile
 
-MAX_SIZE_MB = 150
+MAX_SIZE_MB = 200
 
 
 def print_top_10_largest_files(zip_file):

.buildkite/run-amd-test.sh

@@ -5,6 +5,34 @@ set -ex
 echo "--- ROCm info"
 rocminfo
 
+# cleanup older docker images
+cleanup_docker() {
+  # Get Docker's root directory
+  docker_root=$(docker info -f '{{.DockerRootDir}}')
+  if [ -z "$docker_root" ]; then
+    echo "Failed to determine Docker root directory."
+    exit 1
+  fi
+  echo "Docker root directory: $docker_root"
+  # Check disk usage of the filesystem where Docker's root directory is located
+  disk_usage=$(df "$docker_root" | tail -1 | awk '{print $5}' | sed 's/%//')
+  # Define the threshold
+  threshold=70
+  if [ "$disk_usage" -gt "$threshold" ]; then
+    echo "Disk usage is above $threshold%. Cleaning up Docker images and volumes..."
+    # Remove dangling images (those that are not tagged and not used by any container)
+    docker image prune -f
+    # Remove unused volumes
+    docker volume prune -f
+    echo "Docker images and volumes cleanup completed."
+  else
+    echo "Disk usage is below $threshold%. No cleanup needed."
+  fi
+}
+
+# Call the cleanup docker function
+cleanup_docker
+
 echo "--- Resetting GPUs"
 
 echo "reset" > /opt/amdgpu/etc/gpu_state

.buildkite/run-cpu-test.sh

@@ -11,4 +11,4 @@ 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
+docker run --network host --env VLLM_CPU_KVCACHE_SPACE=1 --name cpu-test cpu-test python3 vllm/examples/offline_inference.py

.buildkite/test-pipeline.yaml

@@ -37,7 +37,6 @@ steps:
   working_dir: "/vllm-workspace/tests"
   num_gpus: 2
   commands:
-  - pytest -v -s distributed/test_pynccl_library.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_basic_distributed_correctness.py
   - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
@@ -63,9 +62,9 @@ steps:
   mirror_hardwares: [amd]
 
   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
+  - pytest -v -s test_inputs.py
+  - pytest -v -s entrypoints -m llm
+  - pytest -v -s entrypoints -m openai
 
 - label: Examples Test
   working_dir: "/vllm-workspace/examples"
@@ -110,6 +109,9 @@ steps:
   mirror_hardwares: [amd]
   command: pytest -v -s test_logits_processor.py
 
+- label: Utils Test
+  command: pytest -v -s test_utils.py
+
 - label: Worker Test
   mirror_hardwares: [amd]
   command: pytest -v -s worker

.clang-format

@@ -0,0 +1,26 @@
+BasedOnStyle: Google
+UseTab: Never
+IndentWidth: 2
+ColumnLimit: 80
+
+# Force pointers to the type for C++.
+DerivePointerAlignment: false
+PointerAlignment: Left
+
+# Reordering #include statements can (and currently will) introduce errors
+SortIncludes: false
+
+# Style choices
+AlignConsecutiveAssignments: false
+AlignConsecutiveDeclarations: false
+IndentPPDirectives: BeforeHash
+
+IncludeCategories:
+  - Regex:           '^<'
+    Priority:        4
+  - Regex:           '^"(llvm|llvm-c|clang|clang-c|mlir|mlir-c)/'
+    Priority:        3
+  - Regex:           '^"(qoda|\.\.)/'
+    Priority:        2
+  - Regex:           '.*'
+    Priority:        1

.github/ISSUE_TEMPLATE/400-bug report.yml

@@ -59,6 +59,8 @@ body:
 
       Please also paste or describe the results you observe instead of the expected results. If you observe an error, please paste the error message including the **full** traceback of the exception. It may be relevant to wrap error messages in ```` ```triple quotes blocks``` ````.
 
+      Please set the environment variable `export VLLM_LOGGING_LEVEL=DEBUG` to turn on more logging to help debugging potential issues.
+
       If you experienced crashes or hangs, it would be helpful to run vllm with `export VLLM_TRACE_FUNCTION=1` . All the function calls in vllm will be recorded. Inspect these log files, and tell which function crashes or hangs.
     placeholder: |
       A clear and concise description of what the bug is.

.github/workflows/clang-format.yml

@@ -0,0 +1,42 @@
+name: clang-format
+
+on:
+  # Trigger the workflow on push or pull request,
+  # but only for the main branch
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+
+jobs:
+  clang-format:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.11"]
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install clang-format==18.1.5
+    - name: Running clang-format
+      run: |
+        EXCLUDES=(
+            'csrc/moe/topk_softmax_kernels.cu'
+            'csrc/punica/bgmv/bgmv_bf16_bf16_bf16.cu'
+            'csrc/punica/bgmv/bgmv_config.h'
+            'csrc/punica/bgmv/bgmv_impl.cuh'
+            'csrc/punica/bgmv/vec_dtypes.cuh'
+            'csrc/punica/punica_ops.cu'
+            'csrc/punica/type_convert.h'
+        )
+        find csrc/ \( -name '*.h' -o -name '*.cpp' -o -name '*.cu' -o -name '*.cuh' \) -print \
+            | grep -vFf <(printf "%s\n" "${EXCLUDES[@]}") \
+            | xargs clang-format --dry-run --Werror

CMakeLists.txt

@@ -167,6 +167,7 @@ set(VLLM_EXT_SRC
   "csrc/layernorm_kernels.cu"
   "csrc/quantization/squeezellm/quant_cuda_kernel.cu"
   "csrc/quantization/gptq/q_gemm.cu"
+  "csrc/quantization/compressed_tensors/int8_quant_kernels.cu"
   "csrc/quantization/fp8/common.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/moe_align_block_size_kernels.cu"
@@ -176,7 +177,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   include(FetchContent)
   SET(CUTLASS_ENABLE_HEADERS_ONLY=ON)
   FetchContent_Declare(
-        cutlass 
+        cutlass
         GIT_REPOSITORY https://github.com/nvidia/cutlass.git
         # CUTLASS 3.5.0
         GIT_TAG 7d49e6c7e2f8896c47f586706e67e1fb215529dc
@@ -199,11 +200,13 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
   # The CUTLASS kernels for Hopper require sm90a to be enabled.
   # This is done via the below gencode option, BUT that creates kernels for both sm90 and sm90a.
   # That adds an extra 17MB to compiled binary, so instead we selectively enable it.
-  set_source_files_properties(
-      "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu"
-      PROPERTIES
-      COMPILE_FLAGS
-      "-gencode arch=compute_90a,code=sm_90a")
+  if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER 12.0)
+    set_source_files_properties(
+          "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu"
+          PROPERTIES
+          COMPILE_FLAGS
+          "-gencode arch=compute_90a,code=sm_90a")
+  endif()
 
 endif()
 

Dockerfile

@@ -79,12 +79,6 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
 COPY .buildkite/check-wheel-size.py check-wheel-size.py
 RUN python3 check-wheel-size.py 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 ####################
 
 #################### vLLM installation IMAGE ####################

Dockerfile.cpu

@@ -17,4 +17,6 @@ RUN pip install -v -r requirements-cpu.txt --extra-index-url https://download.py
 
 RUN VLLM_TARGET_DEVICE=cpu python3 setup.py install
 
+WORKDIR /workspace/
+
 CMD ["/bin/bash"]

README.md

@@ -101,6 +101,7 @@ vLLM is a community project. Our compute resources for development and testing a
 - Crusoe Cloud
 - Databricks
 - DeepInfra
+- Dropbox
 - Lambda Lab
 - NVIDIA
 - Replicate

benchmarks/backend_request_func.py

@@ -89,6 +89,9 @@ async def async_request_tgi(
                     output.latency = most_recent_timestamp - st
                     output.success = True
                     output.generated_text = data["generated_text"]
+                else:
+                    output.error = response.reason or ""
+                    output.success = False
         except Exception:
             output.success = False
             exc_info = sys.exc_info()
@@ -276,6 +279,9 @@ async def async_request_openai_completions(
                     output.generated_text = generated_text
                     output.success = True
                     output.latency = latency
+                else:
+                    output.error = response.reason or ""
+                    output.success = False
         except Exception:
             output.success = False
             exc_info = sys.exc_info()

benchmarks/benchmark_latency.py

@@ -3,13 +3,14 @@
 import json
 import time
 from pathlib import Path
-from typing import Optional
+from typing import List, Optional
 
 import numpy as np
 import torch
 from tqdm import tqdm
 
 from vllm import LLM, SamplingParams
+from vllm.inputs import PromptStrictInputs
 from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
 
 
@@ -34,7 +35,8 @@ def main(args: argparse.Namespace):
               use_v2_block_manager=args.use_v2_block_manager,
               enable_chunked_prefill=args.enable_chunked_prefill,
               download_dir=args.download_dir,
-              block_size=args.block_size)
+              block_size=args.block_size,
+              gpu_memory_utilization=args.gpu_memory_utilization)
 
     sampling_params = SamplingParams(
         n=args.n,
@@ -48,7 +50,9 @@ def main(args: argparse.Namespace):
     dummy_prompt_token_ids = np.random.randint(10000,
                                                size=(args.batch_size,
                                                      args.input_len))
-    dummy_prompt_token_ids = dummy_prompt_token_ids.tolist()
+    dummy_inputs: List[PromptStrictInputs] = [{
+        "prompt_token_ids": batch
+    } for batch in dummy_prompt_token_ids.tolist()]
 
     def run_to_completion(profile_dir: Optional[str] = None):
         if profile_dir:
@@ -59,13 +63,13 @@ def run_to_completion(profile_dir: Optional[str] = None):
                     ],
                     on_trace_ready=torch.profiler.tensorboard_trace_handler(
                         str(profile_dir))) as p:
-                llm.generate(prompt_token_ids=dummy_prompt_token_ids,
+                llm.generate(dummy_inputs,
                              sampling_params=sampling_params,
                              use_tqdm=False)
             print(p.key_averages())
         else:
             start_time = time.perf_counter()
-            llm.generate(prompt_token_ids=dummy_prompt_token_ids,
+            llm.generate(dummy_inputs,
                          sampling_params=sampling_params,
                          use_tqdm=False)
             end_time = time.perf_counter()
@@ -153,15 +157,13 @@ def run_to_completion(profile_dir: Optional[str] = None):
                         action='store_true',
                         help='enforce eager mode and disable CUDA graph')
     parser.add_argument(
-        "--kv-cache-dtype",
+        '--kv-cache-dtype',
         type=str,
-        choices=['auto', 'fp8'],
-        default='auto',
-        help=
-        '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.')
+        choices=['auto', 'fp8', 'fp8_e5m2', 'fp8_e4m3'],
+        default="auto",
+        help='Data type for kv cache storage. If "auto", will use model '
+        'data type. CUDA 11.8+ supports fp8 (=fp8_e4m3) and fp8_e5m2. '
+        'ROCm (AMD GPU) supports fp8 (=fp8_e4m3)')
     parser.add_argument(
         '--quantization-param-path',
         type=str,
@@ -213,5 +215,11 @@ def run_to_completion(profile_dir: Optional[str] = None):
         type=str,
         default=None,
         help='Path to save the latency results in JSON format.')
+    parser.add_argument('--gpu-memory-utilization',
+                        type=float,
+                        default=0.9,
+                        help='the fraction of GPU memory to be used for '
+                        'the model executor, which can range from 0 to 1.'
+                        'If unspecified, will use the default value of 0.9.')
     args = parser.parse_args()
     main(args)

benchmarks/benchmark_serving.py

@@ -215,6 +215,11 @@ def calculate_metrics(
         else:
             actual_output_lens.append(0)
 
+    if completed == 0:
+        warnings.warn(
+            "All requests failed. This is likely due to a misconfiguration "
+            "on the benchmark arguments.",
+            stacklevel=2)
     metrics = BenchmarkMetrics(
         completed=completed,
         total_input=total_input,
@@ -226,9 +231,9 @@ def calculate_metrics(
         1000,  # ttfts is empty if streaming is not supported by backend
         median_ttft_ms=np.median(ttfts or 0) * 1000,
         p99_ttft_ms=np.percentile(ttfts or 0, 99) * 1000,
-        mean_tpot_ms=np.mean(tpots) * 1000,
-        median_tpot_ms=np.median(tpots) * 1000,
-        p99_tpot_ms=np.percentile(tpots, 99) * 1000,
+        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,
     )
 
     return metrics, actual_output_lens
@@ -250,6 +255,24 @@ async def benchmark(
     else:
         raise ValueError(f"Unknown backend: {backend}")
 
+    print("Starting initial single prompt test run...")
+    test_prompt, test_prompt_len, test_output_len = input_requests[0]
+    test_input = RequestFuncInput(
+        model=model_id,
+        prompt=test_prompt,
+        api_url=api_url,
+        prompt_len=test_prompt_len,
+        output_len=test_output_len,
+        best_of=best_of,
+        use_beam_search=use_beam_search,
+    )
+    test_output = await request_func(request_func_input=test_input)
+    if not test_output.success:
+        raise ValueError(
+            "Initial test run failed - Please make sure benchmark arguments "
+            f"are correctly specified. Error: {test_output.error}")
+    else:
+        print("Initial test run completed. Starting main benchmark run...")
     print(f"Traffic request rate: {request_rate}")
 
     pbar = None if disable_tqdm else tqdm(total=len(input_requests))