[Frontend][Core] Add sparse NCCL weight transfer support for in-place updates (#40096)

Signed-off-by: Siddharth Bedekar <bedeksid@gmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Siddharth Bedekar
2026-06-01 15:37:30 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 182c67daf1
commit 266b9d9c64
12 changed files with 1423 additions and 75 deletions
+2 -11
View File
@@ -873,14 +873,7 @@ class LLM(BeamSearchOfflineMixin, PoolingOfflineMixin, OfflineInferenceMixin):
)
def start_weight_update(self, is_checkpoint_format: bool = True) -> None:
"""
Start a new weight update.
Args:
is_checkpoint_format: Whether incoming weights are in checkpoint
format (need layerwise processing) or kernel format (direct
copy).
"""
"""Start a new weight update."""
self.llm_engine.collective_rpc(
"start_weight_update",
kwargs={"is_checkpoint_format": is_checkpoint_format},
@@ -902,9 +895,7 @@ class LLM(BeamSearchOfflineMixin, PoolingOfflineMixin, OfflineInferenceMixin):
)
def finish_weight_update(self) -> None:
"""
Finish the current weight update.
"""
"""Finish the current weight update."""
self.llm_engine.collective_rpc("finish_weight_update")
def __repr__(self) -> str: