Fix lora bench (#6302)
This commit is contained in:
@@ -170,9 +170,7 @@ class LoRAManager:
|
||||
dim=0,
|
||||
out=self.cuda_graph_batch_info.seg_indptr[1 : bs + 1],
|
||||
)
|
||||
self.cuda_graph_batch_info.max_len = int(
|
||||
torch.max(self.cuda_graph_batch_info.seg_lens[:bs])
|
||||
)
|
||||
self.cuda_graph_batch_info.max_len = 1
|
||||
|
||||
for i, lora_path in enumerate(forward_batch.lora_paths):
|
||||
self.cuda_graph_batch_info.weight_indices[i] = (
|
||||
|
||||
Reference in New Issue
Block a user