Merge remote-tracking branch 'IBM/main' into sync-with-ibm

Author: dtrifiro

.buildkite/check-wheel-size.py

@@ -1,7 +1,7 @@
 import os
 import zipfile
 
-MAX_SIZE_MB = 100
+MAX_SIZE_MB = 150
 
 
 def print_top_10_largest_files(zip_file):

.buildkite/run-amd-test.sh

@@ -1,4 +1,4 @@
-# This script build the ROCm docker image and runs test inside it.
+# This script runs test inside the corresponding ROCm docker container.
 set -ex
 
 # Print ROCm version
@@ -19,15 +19,16 @@ done
 
 echo "--- Building container"
 sha=$(git rev-parse --short HEAD)
-container_name=rocm_${sha}
+image_name=rocm_${sha}
+container_name=rocm_${sha}_$(tr -dc A-Za-z0-9 < /dev/urandom | head -c 10; echo)
 docker build \
-        -t ${container_name} \
+        -t ${image_name} \
         -f Dockerfile.rocm \
         --progress plain \
         .
 
 remove_docker_container() {
-   docker rm -f ${container_name} || docker image rm -f ${container_name} || true
+   docker rm -f ${container_name} || docker image rm -f ${image_name} || true
 }
 trap remove_docker_container EXIT
 
@@ -39,6 +40,6 @@ docker run \
         --rm \
         -e HF_TOKEN \
         --name ${container_name} \
-        ${container_name} \
-        /bin/bash -c $(echo $1 | sed "s/^'//" | sed "s/'$//")
+        ${image_name} \
+        /bin/bash -c "${@}"
 

.buildkite/run-benchmarks.sh

@@ -9,10 +9,10 @@ cd "$(dirname "${BASH_SOURCE[0]}")/.."
 (which wget && which curl) || (apt-get update && apt-get install -y wget curl)
 
 # run python-based benchmarks and upload the result to buildkite
-python3 benchmarks/benchmark_latency.py 2>&1 | tee benchmark_latency.txt
+python3 benchmarks/benchmark_latency.py --output-json latency_results.json 2>&1 | tee benchmark_latency.txt
 bench_latency_exit_code=$?
 
-python3 benchmarks/benchmark_throughput.py --input-len 256 --output-len 256 2>&1 | tee benchmark_throughput.txt
+python3 benchmarks/benchmark_throughput.py --input-len 256 --output-len 256 --output-json throughput_results.json 2>&1 | tee benchmark_throughput.txt
 bench_throughput_exit_code=$?
 
 # run server-based benchmarks and upload the result to buildkite
@@ -74,4 +74,5 @@ if [ $bench_serving_exit_code -ne 0 ]; then
     exit $bench_serving_exit_code
 fi
 
-/workspace/buildkite-agent artifact upload openai-*.json
+rm ShareGPT_V3_unfiltered_cleaned_split.json
+/workspace/buildkite-agent artifact upload "*.json"

.buildkite/test-pipeline.yaml

@@ -5,13 +5,16 @@
 
 steps:
 - label: Regression Test
+  mirror_hardwares: [amd]
   command: pytest -v -s test_regression.py
   working_dir: "/vllm-workspace/tests" # optional
 
 - label: AsyncEngine Test
+  #mirror_hardwares: [amd]
   command: pytest -v -s async_engine
 
 - label: Basic Correctness Test
+  mirror_hardwares: [amd]
   commands:
   - VLLM_ATTENTION_BACKEND=XFORMERS pytest -v -s basic_correctness/test_basic_correctness.py
   - VLLM_ATTENTION_BACKEND=FLASH_ATTN pytest -v -s basic_correctness/test_basic_correctness.py
@@ -24,34 +27,40 @@ steps:
   command: pytest -v -s core
 
 - label: Distributed Comm Ops Test
-  command: pytest -v -s test_comm_ops.py
-  working_dir: "/vllm-workspace/tests/distributed"
+  #mirror_hardwares: [amd]
+  command: pytest -v -s distributed/test_comm_ops.py
+  working_dir: "/vllm-workspace/tests"
   num_gpus: 2
 
 - label: Distributed Tests
-  working_dir: "/vllm-workspace/tests/distributed"
-
-  num_gpus: 2 # only support 1 or 2 for now.
   mirror_hardwares: [amd]
-
+  working_dir: "/vllm-workspace/tests"
+  num_gpus: 2
   commands:
-  - pytest -v -s test_pynccl_library.py
-  - TEST_DIST_MODEL=facebook/opt-125m pytest -v -s test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf pytest -v -s test_basic_distributed_correctness.py
-  - TEST_DIST_MODEL=facebook/opt-125m pytest -v -s test_chunked_prefill_distributed.py
-  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf pytest -v -s test_chunked_prefill_distributed.py
+  - 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
+  - TEST_DIST_MODEL=meta-llama/Llama-2-7b-hf DISTRIBUTED_EXECUTOR_BACKEND=ray pytest -v -s distributed/test_chunked_prefill_distributed.py
+  - TEST_DIST_MODEL=facebook/opt-125m DISTRIBUTED_EXECUTOR_BACKEND=mp pytest -v -s distributed/test_basic_distributed_correctness.py
+  - 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 
 
 - label: Distributed Tests (Multiple Groups)
-  working_dir: "/vllm-workspace/tests/distributed"
+  #mirror_hardwares: [amd]
+  working_dir: "/vllm-workspace/tests"
   num_gpus: 4
   commands:
-  - pytest -v -s test_pynccl.py
+  - pytest -v -s distributed/test_pynccl.py
 
 - label: Engine Test
   mirror_hardwares: [amd]
   command: pytest -v -s engine tokenization test_sequence.py test_config.py test_logger.py
 
 - label: Entrypoints Test
+  #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
@@ -62,21 +71,24 @@ steps:
   mirror_hardwares: [amd]
   commands:
     # install aws cli for llava_example.py
-    - pip install awscli
+    # install tensorizer for tensorize_vllm_model.py
+    - pip install awscli tensorizer
     - python3 offline_inference.py
     - python3 offline_inference_with_prefix.py
     - python3 llm_engine_example.py
     - python3 llava_example.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
 
 - label: Kernels Test %N
+  #mirror_hardwares: [amd]
   command: pytest -v -s kernels --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
   parallelism: 4
 
 - label: Models Test
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   commands:
     - bash ../.buildkite/download-images.sh
-    - pytest -v -s models --ignore=models/test_llava.py --ignore=models/test_mistral.py
+    - pytest -v -s models --ignore=models/test_llava.py
 
 - label: Llava Test
   mirror_hardwares: [amd]
@@ -90,6 +102,7 @@ steps:
     - pytest -v -s prefix_caching
 
 - label: Samplers Test
+  #mirror_hardwares: [amd]
   command: pytest -v -s samplers
 
 - label: LogitsProcessor Test
@@ -101,20 +114,38 @@ steps:
   command: pytest -v -s worker
 
 - label: Speculative decoding tests
-  mirror_hardwares: [amd]
+  #mirror_hardwares: [amd]
   command: pytest -v -s spec_decode
 
 - label: LoRA Test %N
-  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT
+  #mirror_hardwares: [amd]
+  command: pytest -v -s lora --shard-id=$$BUILDKITE_PARALLEL_JOB --num-shards=$$BUILDKITE_PARALLEL_JOB_COUNT --ignore=lora/test_long_context.py
   parallelism: 4
 
+- label: LoRA Long Context (Distributed)
+  #mirror_hardwares: [amd]
+  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
+
 - label: Tensorizer Test
+  #mirror_hardwares: [amd]
   command: apt-get install curl libsodium23 && pytest -v -s tensorizer_loader
 
 - label: Metrics Test
+  mirror_hardwares: [amd]
   command: pytest -v -s metrics
 
 - label: Quantization Test
+  #mirror_hardwares: [amd]
   command: pytest -v -s quantization
 
 - label: Benchmarks

.buildkite/test-template.j2

@@ -3,9 +3,8 @@
 {% set default_working_dir = "/vllm-workspace/tests" %}
 
 steps:
-
   - label: ":docker: build image"
-    commands:
+    commands: 
       - "docker build --build-arg max_jobs=16 --tag {{ docker_image }} --target test --progress plain ."
       - "docker push {{ docker_image }}"
     env:
@@ -14,6 +13,8 @@ steps:
       automatic:
         - exit_status: -1  # Agent was lost
           limit: 5
+        - exit_status: -10  # Agent was lost
+          limit: 5
   - wait
 
   - group: "AMD Tests"
@@ -24,7 +25,7 @@ steps:
       - label: "AMD: {{ step.label }}"
         agents:
           queue: amd
-        command: bash .buildkite/run-amd-test.sh "'cd {{ (step.working_dir or default_working_dir) | safe  }} && {{ step.command  or (step.commands | join(' && ')) | safe }}'"
+        command: bash .buildkite/run-amd-test.sh "cd {{ (step.working_dir or default_working_dir) | safe  }} ; {{ step.command  or (step.commands | join(" ; ")) | safe }}"
         env:
           DOCKER_BUILDKIT: "1"
     {% endif %}
@@ -53,6 +54,8 @@ steps:
       automatic:
         - exit_status: -1  # Agent was lost
           limit: 5
+        - exit_status: -10  # Agent was lost
+          limit: 5
     plugins:
       - kubernetes:
           podSpec:

CMakeLists.txt

@@ -167,19 +167,44 @@ set(VLLM_EXT_SRC
   "csrc/layernorm_kernels.cu"
   "csrc/quantization/squeezellm/quant_cuda_kernel.cu"
   "csrc/quantization/gptq/q_gemm.cu"
-  "csrc/quantization/fp8/fp8_cuda_kernels.cu"
+  "csrc/quantization/fp8/common.cu"
   "csrc/cuda_utils_kernels.cu"
   "csrc/moe_align_block_size_kernels.cu"
   "csrc/pybind.cpp")
 
 if(VLLM_GPU_LANG STREQUAL "CUDA")
+  include(FetchContent)
+  SET(CUTLASS_ENABLE_HEADERS_ONLY=ON)
+  FetchContent_Declare(
+        cutlass 
+        GIT_REPOSITORY https://github.com/nvidia/cutlass.git
+        # CUTLASS 3.5.0
+        GIT_TAG 7d49e6c7e2f8896c47f586706e67e1fb215529dc
+  )
+  FetchContent_MakeAvailable(cutlass)
+
   list(APPEND VLLM_EXT_SRC
     "csrc/quantization/aqlm/gemm_kernels.cu"
     "csrc/quantization/awq/gemm_kernels.cu"
-    "csrc/quantization/marlin/marlin_cuda_kernel.cu"
+    "csrc/quantization/marlin/dense/marlin_cuda_kernel.cu"
+    "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/custom_all_reduce.cu")
+    "csrc/custom_all_reduce.cu"
+    "csrc/quantization/cutlass_w8a8/scaled_mm_dq_entry.cu"
+    "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c2x.cu"
+    "csrc/quantization/cutlass_w8a8/scaled_mm_dq_c3x.cu")
+
+  #
+  # 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")
+
 endif()
 
 define_gpu_extension_target(
@@ -189,6 +214,7 @@ define_gpu_extension_target(
   SOURCES ${VLLM_EXT_SRC}
   COMPILE_FLAGS ${VLLM_GPU_FLAGS}
   ARCHITECTURES ${VLLM_GPU_ARCHES}
+  INCLUDE_DIRECTORIES ${CUTLASS_INCLUDE_DIR};${CUTLASS_TOOLS_UTIL_INCLUDE_DIR}
   WITH_SOABI)
 
 #
@@ -219,7 +245,8 @@ set(VLLM_PUNICA_EXT_SRC
   "csrc/punica/bgmv/bgmv_fp16_fp32_fp16.cu"
   "csrc/punica/bgmv/bgmv_fp32_bf16_bf16.cu"
   "csrc/punica/bgmv/bgmv_fp32_fp16_fp16.cu"
-  "csrc/punica/punica_ops.cc")
+  "csrc/punica/punica_ops.cu"
+  "csrc/punica/punica_pybind.cpp")
 
 #
 # Copy GPU compilation flags+update for punica
@@ -243,6 +270,9 @@ if (${VLLM_GPU_LANG} STREQUAL "CUDA")
     endif()
   endforeach()
   message(STATUS "Punica target arches: ${VLLM_PUNICA_GPU_ARCHES}")
+elseif(${VLLM_GPU_LANG} STREQUAL "HIP")
+  set(VLLM_PUNICA_GPU_ARCHES ${VLLM_GPU_ARCHES})
+  message(STATUS "Punica target arches: ${VLLM_PUNICA_GPU_ARCHES}")
 endif()
 
 if (VLLM_PUNICA_GPU_ARCHES)
@@ -277,11 +307,6 @@ add_custom_target(default)
 if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
   message(STATUS "Enabling C extension.")
   add_dependencies(default _C)
-endif()
-
-if(VLLM_GPU_LANG STREQUAL "CUDA")
-  message(STATUS "Enabling moe extension.")
-  add_dependencies(default _moe_C)
 
   # Enable punica if -DVLLM_INSTALL_PUNICA_KERNELS=ON or
   # VLLM_INSTALL_PUNICA_KERNELS is set in the environment and
@@ -292,3 +317,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     add_dependencies(default _punica_C)
   endif()
 endif()
+
+if(VLLM_GPU_LANG STREQUAL "CUDA")
+  message(STATUS "Enabling moe extension.")
+  add_dependencies(default _moe_C)
+endif()

Dockerfile

@@ -87,23 +87,6 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     pip cache remove vllm_nccl*
 #################### EXTENSION Build IMAGE ####################
 
-#################### FLASH_ATTENTION Build IMAGE ####################
-FROM dev as flash-attn-builder
-# max jobs used for build
-ARG max_jobs=2
-ENV MAX_JOBS=${max_jobs}
-# flash attention version
-ARG flash_attn_version=v2.5.8
-ENV FLASH_ATTN_VERSION=${flash_attn_version}
-
-WORKDIR /usr/src/flash-attention-v2
-
-# Download the wheel or build it if a pre-compiled release doesn't exist
-RUN pip --verbose wheel flash-attn==${FLASH_ATTN_VERSION} \
-    --no-build-isolation --no-deps --no-cache-dir
-
-#################### FLASH_ATTENTION Build IMAGE ####################
-
 #################### vLLM installation IMAGE ####################
 # image with vLLM installed
 FROM nvidia/cuda:12.4.1-base-ubuntu22.04 AS vllm-base
@@ -122,10 +105,6 @@ RUN ldconfig /usr/local/cuda-12.4/compat/
 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 ####################
 
 

Dockerfile.rocm

@@ -94,6 +94,9 @@ COPY . .
 
 RUN python3 -m pip install --upgrade pip numba
 
+# make sure punica kernels are built (for LoRA)
+ENV VLLM_INSTALL_PUNICA_KERNELS=1
+
 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 \