Breakmodel's CUDA stream should be on primary device
This commit is contained in:
parent
aa59f8b4b2
commit
3649ba9fa4
|
@ -387,7 +387,7 @@ def new_forward(
|
||||||
all_hidden_states = () if output_hidden_states else None
|
all_hidden_states = () if output_hidden_states else None
|
||||||
|
|
||||||
if breakmodel and ram_blocks:
|
if breakmodel and ram_blocks:
|
||||||
copystream = torch.cuda.Stream(device=0,priority = -1)
|
copystream = torch.cuda.Stream(device=primary_device, priority=-1)
|
||||||
|
|
||||||
for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
|
for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
|
||||||
|
|
||||||
|
|
Loading…
Reference in New Issue