[ROCm] Enable the cuMem CUDA-graph pool offload on ROCm (#59523)
Signed-off-by: aoshen02 <aoshen@inferact.ai> Signed-off-by: indianspeedster <cspandey016@gmail.com> Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com> Co-authored-by: indianspeedster <cspandey016@gmail.com>
This commit is contained in:
co-authored by
Claude Fable 5.1
indianspeedster
parent
bacbbe1878
commit
a6f43e57d6
@@ -151,6 +151,7 @@ class CudaCommunicator(DeviceCommunicatorBase):
|
||||
self.aiter_ar_comm = AiterCustomAllreduce(
|
||||
group=self.cpu_group,
|
||||
device=self.device,
|
||||
register_graph_buffers=register,
|
||||
)
|
||||
|
||||
if use_custom_allreduce and self.aiter_ar_comm is None and self.world_size > 1:
|
||||
@@ -174,6 +175,7 @@ class CudaCommunicator(DeviceCommunicatorBase):
|
||||
self.aiter_ar_comm = AiterCustomAllreduce(
|
||||
group=self.cpu_group,
|
||||
device=self.device,
|
||||
register_graph_buffers=register,
|
||||
)
|
||||
if self.aiter_ar_comm.disabled:
|
||||
self.aiter_ar_comm = None
|
||||
|
||||
Reference in New Issue
Block a user