remove torch.cuda.synchronize() from run_nvfp4_grouped_gemm (cudagraph-safe)
This commit is contained in:
@@ -344,6 +344,5 @@ def run_nvfp4_grouped_gemm(
|
||||
stream,
|
||||
global_scale_a=gsa_c, global_scale_b=gsb_c,
|
||||
)
|
||||
torch.cuda.synchronize()
|
||||
|
||||
return out
|
||||
|
||||
Reference in New Issue
Block a user