[OOT] Support sync_model_loading for OOT (#25126)

Signed-off-by: Chendi Xue <Chendi.Xue@intel.com>
This commit is contained in:
Chendi.Xue
2025-09-19 05:41:53 +00:00
committed by GitHub
parent 6c8a3c099b
commit a6149aa587
4 changed files with 33 additions and 17 deletions
+23
View File
@@ -594,6 +594,29 @@ class Platform:
"""
return False
@classmethod
def use_sync_weight_loader(cls) -> bool:
"""
Returns if the current platform needs to sync weight loader.
"""
return False
@classmethod
def make_synced_weight_loader(cls, original_weight_loader):
"""
Wrap the original weight loader to make it synced.
"""
if not cls.use_sync_weight_loader():
return original_weight_loader
def _synced_weight_loader(param, *args, **kwargs):
out = original_weight_loader(param, *args, **kwargs)
if param.device != torch.device("cpu"):
torch._sync(param)
return out
return _synced_weight_loader
class UnspecifiedPlatform(Platform):
_enum = PlatformEnum.UNSPECIFIED
+4
View File
@@ -226,6 +226,10 @@ class TpuPlatform(Platform):
torch.ops.xla.dynamo_set_buffer_donor_(src_cache, True)
dst_cache[dst_block_indices] = src_cache[src_block_indices].cpu()
@classmethod
def use_sync_weight_loader(cls) -> bool:
return True
try:
from tpu_commons.platforms import TpuPlatform as TpuCommonsPlatform