[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
@@ -216,6 +216,11 @@ def _layerwise_process(layer: torch.nn.Module, info: LayerReloadingInfo):
# Materialize layer tensors onto device
materialize_layer(layer)
# Reset FP8 online quantization flag so process_weights_after_loading
# will run again during reload
if hasattr(layer, "_already_called_process_weights_after_loading"):
delattr(layer, "_already_called_process_weights_after_loading")
# Unwrap layerwise loading wrappers
for param in get_layer_tensors(layer).values():
param.weight_loader = _get_original_loader(param)