@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 9cdba9669c
Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 16, 08:15 UTC

[Doc] Update help text for `--distributed-executor-backend` (#10231)

Signed-off-by: Russell Bryant <rbryant@redhat.com>

Russell Bryant committed Nov 11, 2024 at 20:55 UTC 9cdba9669cb32191aa0ae6782c0648be3e0e44ed
2 files changed +14 -6
vllm/config.py
+6 -3
@@ -951,9 +951,12 @@ class ParallelConfig:
951 https://docs.ray.io/en/latest/ray-observability/user-guides/profiling.html#profiling-nsight-profiler.
952 placement_group: ray distributed model workers placement group.
953 distributed_executor_backend: Backend to use for distributed model
954 - workers, either "ray" or "mp" (multiprocessing). If either
955 - pipeline_parallel_size or tensor_parallel_size is greater than 1,
956 - will default to "ray" if Ray is installed or "mp" otherwise.
954 + workers, either "ray" or "mp" (multiprocessing). If the product
955 + of pipeline_parallel_size and tensor_parallel_size is less than
956 + or equal to the number of GPUs available, "mp" will be used to
957 + keep processing on a single host. Otherwise, this will default
958 + to "ray" if Ray is installed and fail otherwise. Note that tpu
959 + and hpu only support Ray for distributed inference.
960 """
961
962 def __init__(
vllm/engine/arg_utils.py
+8 -3
@@ -369,9 +369,14 @@ class EngineArgs:
369 '--distributed-executor-backend',
370 choices=['ray', 'mp'],
371 default=EngineArgs.distributed_executor_backend,
372 - help='Backend to use for distributed serving. When more than 1 GPU '
373 - 'is used, will be automatically set to "ray" if installed '
374 - 'or "mp" (multiprocessing) otherwise.')
372 + help='Backend to use for distributed model '
373 + 'workers, either "ray" or "mp" (multiprocessing). If the product '
374 + 'of pipeline_parallel_size and tensor_parallel_size is less than '
375 + 'or equal to the number of GPUs available, "mp" will be used to '
376 + 'keep processing on a single host. Otherwise, this will default '
377 + 'to "ray" if Ray is installed and fail otherwise. Note that tpu '
378 + 'and hpu only support Ray for distributed inference.')
379 +
380 parser.add_argument(
381 '--worker-use-ray',
382 action='store_true',