[Bug Fix] Fix broken sgemm_lora_a_graph_fwd due to invalid torch.mm() call (#24760)
This commit is contained in:
@@ -65,7 +65,7 @@ def sgemm_lora_a_graph_fwd(
|
||||
x_seq = torch.where(batch_token_mask, inputs, 0)
|
||||
w_seq = weights[lora_idx]
|
||||
|
||||
output.add_(scaling_tensor[lora_idx] * torch.mm(x_seq, w_seq.t(), 0))
|
||||
output.add_(scaling_tensor[lora_idx] * torch.mm(x_seq, w_seq.t()))
|
||||
|
||||
return output
|
||||
|
||||
|
||||
Reference in New Issue
Block a user