[Feat][RL][1/2] Native Weight Syncing API: NCCL (#31943)

Signed-off-by: ahao-anyscale <ahao@anyscale.com>
Signed-off-by: Aaron Hao <ahao@anyscale.com>
Co-authored-by: SumanthRH <sumanthrh99@gmail.com>
This commit is contained in:
Aaron Hao
2026-02-05 12:13:23 -05:00
committed by GitHub
co-authored by SumanthRH
parent 82914d2ae8
commit c1858b7ec8
27 changed files with 2974 additions and 2 deletions
+14
View File
@@ -54,6 +54,7 @@ from vllm.config import (
SpeculativeConfig,
StructuredOutputsConfig,
VllmConfig,
WeightTransferConfig,
get_attr_docs,
)
from vllm.config.cache import (
@@ -581,6 +582,11 @@ class EngineArgs:
kv_offloading_backend: KVOffloadingBackend = CacheConfig.kv_offloading_backend
tokens_only: bool = False
weight_transfer_config: WeightTransferConfig | None = None
"""Configuration for weight transfer during RL training.
Accepts a JSON string or dict with backend-specific options.
Example: '{"backend": "nccl"}'"""
def __post_init__(self):
# support `EngineArgs(compilation_config={...})`
# without having to manually construct a
@@ -591,6 +597,10 @@ class EngineArgs:
self.attention_config = AttentionConfig(**self.attention_config)
if isinstance(self.eplb_config, dict):
self.eplb_config = EPLBConfig(**self.eplb_config)
if isinstance(self.weight_transfer_config, dict):
self.weight_transfer_config = WeightTransferConfig(
**self.weight_transfer_config
)
# Setup plugins
from vllm.plugins import load_general_plugins
@@ -1189,6 +1199,9 @@ class EngineArgs:
vllm_group.add_argument(
"--optimization-level", **vllm_kwargs["optimization_level"]
)
vllm_group.add_argument(
"--weight-transfer-config", **vllm_kwargs["weight_transfer_config"]
)
# Other arguments
parser.add_argument(
@@ -1765,6 +1778,7 @@ class EngineArgs:
profiler_config=self.profiler_config,
additional_config=self.additional_config,
optimization_level=self.optimization_level,
weight_transfer_config=self.weight_transfer_config,
)
return config
+14
View File
@@ -6,6 +6,10 @@ from collections.abc import AsyncGenerator, Iterable, Mapping
from typing import Any
from vllm.config import ModelConfig, VllmConfig
from vllm.distributed.weight_transfer.base import (
WeightTransferInitRequest,
WeightTransferUpdateRequest,
)
from vllm.inputs.data import PromptType, StreamingInput
from vllm.lora.request import LoRARequest
from vllm.outputs import PoolingRequestOutput, RequestOutput
@@ -191,3 +195,13 @@ class EngineClient(ABC):
async def get_supported_tasks(self) -> tuple[SupportedTask, ...]:
"""Get supported tasks"""
raise NotImplementedError
async def init_weight_transfer_engine(
self, init_request: WeightTransferInitRequest
) -> None:
"""Initialize weight transfer for RL training."""
raise NotImplementedError
async def update_weights(self, request: WeightTransferUpdateRequest) -> None:
"""Batched weight update for RL training."""
raise NotImplementedError