Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 4.7
parent
3d31ac2672
commit
6575aea128
@@ -1003,9 +1003,7 @@ class ModelRunner(ModelRunnerKVCacheMixin):
|
|||||||
sv = torch.empty(0, dtype=torch.uint8, device=t.device).set_(
|
sv = torch.empty(0, dtype=torch.uint8, device=t.device).set_(
|
||||||
t.untyped_storage()
|
t.untyped_storage()
|
||||||
)
|
)
|
||||||
if sv.data_ptr() not in {
|
if sv.data_ptr() not in {v.data_ptr() for v in model_tensors.values()}:
|
||||||
v.data_ptr() for v in model_tensors.values()
|
|
||||||
}:
|
|
||||||
model_tensors[f"{name}.__storage"] = sv
|
model_tensors[f"{name}.__storage"] = sv
|
||||||
|
|
||||||
nixl_metadata = nixl_mgr.register_tensors(model_tensors)
|
nixl_metadata = nixl_mgr.register_tensors(model_tensors)
|
||||||
|
|||||||
@@ -2408,9 +2408,7 @@ class RemoteInstanceModelLoader(BaseModelLoader):
|
|||||||
|
|
||||||
# --- Transport-specific transfer ---
|
# --- Transport-specific transfer ---
|
||||||
if transport == "nixl":
|
if transport == "nixl":
|
||||||
self._transfer_via_nixl(
|
self._transfer_via_nixl(model, nixl_mgr, source_worker, tp_rank)
|
||||||
model, nixl_mgr, source_worker, tp_rank
|
|
||||||
)
|
|
||||||
else:
|
else:
|
||||||
self._transfer_via_transfer_engine(
|
self._transfer_via_transfer_engine(
|
||||||
model, transfer_engine, source_worker, tp_rank
|
model, transfer_engine, source_worker, tp_rank
|
||||||
@@ -2555,8 +2553,7 @@ class RemoteInstanceModelLoader(BaseModelLoader):
|
|||||||
)
|
)
|
||||||
|
|
||||||
logger.info(
|
logger.info(
|
||||||
"ModelExpress [nixl]: transferred %d tensors, "
|
"ModelExpress [nixl]: transferred %d tensors, " "%.2f GB in %.2fs",
|
||||||
"%.2f GB in %.2fs",
|
|
||||||
matched,
|
matched,
|
||||||
total_bytes / 1e9,
|
total_bytes / 1e9,
|
||||||
duration,
|
duration,
|
||||||
|
|||||||
@@ -7109,7 +7109,10 @@ class ServerArgs:
|
|||||||
if self.remote_instance_weight_loader_start_seed_via_transfer_engine:
|
if self.remote_instance_weight_loader_start_seed_via_transfer_engine:
|
||||||
return True
|
return True
|
||||||
# ModelExpress source mode needs TransferEngine init only if transport is transfer_engine.
|
# ModelExpress source mode needs TransferEngine init only if transport is transfer_engine.
|
||||||
if self.modelexpress_source and self.modelexpress_transport == "transfer_engine":
|
if (
|
||||||
|
self.modelexpress_source
|
||||||
|
and self.modelexpress_transport == "transfer_engine"
|
||||||
|
):
|
||||||
return True
|
return True
|
||||||
# Use TransferEngine as client backend.
|
# Use TransferEngine as client backend.
|
||||||
elif (
|
elif (
|
||||||
|
|||||||
Reference in New Issue
Block a user