[Hardware][Intel CPU] Adding intel openmp tunings in Docker file (#6008)

Signed-off-by: Yuan Zhou <yuan.zhou@intel.com>

.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 --entrypoint /bin/bash -v ~/.cache/huggingface:/root/.cache/huggingface --cpuset-cpus=48-95 \
-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
+  --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"

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"]

vllm/utils.py

@@ -398,6 +398,27 @@ def update_environment_variables(envs: Dict[str, str]):
         os.environ[k] = v
 
 
+def init_kmp_env():
+    if not is_cpu():
+        return
+
+    ld_prealod_str = os.getenv("LD_PRELOAD", "")
+    if "libiomp5.so" not in ld_prealod_str:
+        return
+
+    # The time(milliseconds) that a thread should wait after completing the
+    # execution of a parallel region, before sleeping.
+    os.environ['KMP_BLOCKTIME'] = "1"
+    # dump settings on start up
+    os.environ['KMP_SETTINGS'] = "1"
+    # Prevents the CPU to run into low performance state
+    os.environ['KMP_TPAUSE'] = "0"
+    # Provides fine granularity parallelism
+    os.environ['KMP_FORKJOIN_BARRIER_PATTERN'] = "dist,dist"
+    os.environ['KMP_PLAIN_BARRIER_PATTERN'] = "dist,dist"
+    os.environ['KMP_REDUCTION_BARRIER_PATTERN'] = "dist,dist"
+
+
 def chunk_list(lst: List[T], chunk_size: int) -> List[List[T]]:
     """Yield successive chunk_size chunks from lst."""
     return [lst[i:i + chunk_size] for i in range(0, len(lst), chunk_size)]

vllm/worker/cpu_worker.py

@@ -13,7 +13,7 @@ from vllm.distributed import (ensure_model_parallel_initialized,
 from vllm.logger import init_logger
 from vllm.model_executor import set_random_seed
 from vllm.sequence import ExecuteModelRequest
-from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE
+from vllm.utils import STR_DTYPE_TO_TORCH_DTYPE, init_kmp_env
 from vllm.worker.cpu_model_runner import CPUModelRunner
 from vllm.worker.worker_base import (LocalOrDistributedWorkerBase,
                                      LoraNotSupportedWorkerBase, WorkerInput)
@@ -150,6 +150,9 @@ class CPUWorker(LoraNotSupportedWorkerBase, LocalOrDistributedWorkerBase):
         if self.is_driver_worker:
             assert self.rank == 0, "The driver worker must have rank 0."
 
+        # try to initialize intel openmp optimized tunings
+        init_kmp_env()
+
         if self.model_config.trust_remote_code:
             # note: lazy import to avoid importing torch before initializing
             from vllm.utils import init_cached_hf_modules