lazy import attn backends (#4200)
This commit is contained in:
@@ -302,7 +302,7 @@ class CudaGraphRunner:
|
||||
self.stream = graph_capture_context.stream
|
||||
# Reverse the order to enable better memory sharing across cuda graphs.
|
||||
capture_range = (
|
||||
tqdm.tqdm(reversed(self.capture_bs))
|
||||
tqdm.tqdm(list(reversed(self.capture_bs)))
|
||||
if get_tensor_model_parallel_rank() == 0
|
||||
else reversed(self.capture_bs)
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user