[QeRL] Fix online quantized reloading (#38442)

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>
This commit is contained in:
Kyle Sayers
2026-03-29 14:56:41 -06:00
committed by GitHub
parent 995dea1354
commit d28d86e8a3
9 changed files with 104 additions and 62 deletions
@@ -94,14 +94,15 @@ def restore_layer_on_meta(layer: torch.nn.Module, info: LayerReloadingInfo):
layer.register_buffer(name, buffer)
def materialize_layer(layer: torch.nn.Module) -> None:
def materialize_layer(layer: torch.nn.Module, info: LayerReloadingInfo):
"""Materialize all meta tensors in a layer to actual tensors."""
if layer.__class__.__name__ in SKIP_MODULES:
return
for name, tensor in get_layer_tensors(layer).items():
if name not in SKIP_TENSORS:
setattr(layer, name, materialize_meta_tensor(tensor))
with info.restore_device:
for name, tensor in get_layer_tensors(layer).items():
if name not in SKIP_TENSORS:
setattr(layer, name, materialize_meta_tensor(tensor))
class CopyCounter(TorchDispatchMode):