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 8, 17:45 UTC
Update `WeightTransferConfig` to be more standard like the others (#33989)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Feb 6, 2026 at 13:15 UTC
51a7bda6252dfff7aa2ab9e82a86914430034563
2 files changed
+4
-6
vllm/config/weight_transfer.py
-2
@@ -1,13 +1,11 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
-from dataclasses import dataclass
3
from typing import Literal
4
5
from vllm.config.utils import config
6
7
8
@config
10
-@dataclass
9
class WeightTransferConfig:
10
"""Configuration for weight transfer during RL training."""
11
vllm/engine/arg_utils.py
+4
-4
@@ -582,10 +582,10 @@ class EngineArgs:
582
kv_offloading_backend: KVOffloadingBackend = CacheConfig.kv_offloading_backend
583
tokens_only: bool = False
584
585
- weight_transfer_config: WeightTransferConfig | None = None
586
- """Configuration for weight transfer during RL training.
587
- Accepts a JSON string or dict with backend-specific options.
588
- Example: '{"backend": "nccl"}'"""
585
+ weight_transfer_config: WeightTransferConfig | None = get_field(
586
+ VllmConfig,
587
+ "weight_transfer_config",
588
+ )
589
590
def __post_init__(self):
591
# support `EngineArgs(compilation_config={...})`