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 5, 18:30 UTC
DOC: TPU mention fix (#38129)
Signed-off-by: Mateusz Sokół <mat646@gmail.com>
Mateusz Sokół committed
Mar 31, 2026 at 12:27 UTC
8278825b577675fa5fabfc3a38a48d55296a55b9
1 file changed
+6
-2
vllm/config/parallel.py
+6
-2
@@ -214,14 +214,18 @@ class ParallelConfig:
214
distributed_executor_backend: (
215
str | DistributedExecutorBackend | type[Executor] | None
216
) = None
217
- """Backend to use for distributed model workers, either "ray" or "mp"
217
+ """
218
+ Backend to use for distributed model workers, either "ray" or "mp"
219
(multiprocessing). If the product of pipeline_parallel_size and tensor_parallel_size
220
is less than or equal to the number of GPUs available, "mp" will be used to
221
keep processing on a single host. Otherwise, an error will be raised. To use "mp"
222
you must also set nnodes, and to use "ray" you must manually set
223
distributed_executor_backend to "ray".
224
224
- Note that tpu only support Ray for distributed inference."""
225
+ Note:
226
+ [TPU](https://docs.vllm.ai/projects/tpu/en/latest/) platform only supports Ray
227
+ for distributed inference.
228
+ """
229
230
worker_cls: str = "auto"
231
"""The full name of the worker class to use. If "auto", the worker class