Add routed expert loading for gpt-oss (#52209)

Signed-off-by: Wyett <wyettzeng@gmail.com>
Signed-off-by: wyett <wyettzeng@gmail.com>
This commit is contained in:
Wyett
2026-08-23 17:59:08 +08:00
committed by GitHub
parent e25c586b90
commit b26039b09f
4 changed files with 370 additions and 3 deletions
@@ -124,11 +124,14 @@ def restore_layer_on_meta(layer: torch.nn.Module, info: LayerReloadingInfo):
return
non_persistent = set(layer._non_persistent_buffers_set)
for name in get_layer_tensors(layer):
if name not in SKIP_TENSORS:
restore_params, restore_buffers = info.restore_metadata
tensor_names = (
get_layer_tensors(layer).keys() | restore_params.keys() | restore_buffers.keys()
)
for name in tensor_names:
if name not in SKIP_TENSORS and hasattr(layer, name):
delattr(layer, name)
restore_params, restore_buffers = info.restore_metadata
for name, param in restore_params.items():
if name not in SKIP_TENSORS:
param = restore_layer_refs(param, layer)