[Optimization] support bladedisc fp16 optimization (#1790)

xiaowan0322

2024-06-06 9a9b474e7de7cc90d2ee124dc8d6c2cfa887c059

 funasr/models/sanm/attention.py

@@ -780,7 +780,7 @@
        return q, k, v

    def forward_attention(self, value, scores, mask, ret_attn):
        scores = scores + mask
        scores = scores + mask.to(scores.device)

        self.attn = torch.softmax(scores, dim=-1)
        context_layer = torch.matmul(self.attn, value)  # (batch, head, time1, d_k)