[fix] fix misusing of is_cuda (#7790)
This commit is contained in:
@@ -913,7 +913,7 @@ class DeepseekV2AttentionMLA(nn.Module):
|
||||
and self.fused_qkv_a_proj_with_mqa.weight.dtype == torch.bfloat16
|
||||
and self.fused_qkv_a_proj_with_mqa.weight.shape[0] == 2112
|
||||
and self.fused_qkv_a_proj_with_mqa.weight.shape[1] == 7168
|
||||
and is_cuda
|
||||
and _is_cuda
|
||||
and _device_sm >= 90
|
||||
)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user