From 9a9b474e7de7cc90d2ee124dc8d6c2cfa887c059 Mon Sep 17 00:00:00 2001
From: xiaowan0322 <wanchen.swc@alibaba-inc.com>
Date: 星期四, 06 六月 2024 15:59:56 +0800
Subject: [PATCH] [Optimization] support bladedisc fp16 optimization (#1790)
---
funasr/models/sanm/attention.py | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)
diff --git a/funasr/models/sanm/attention.py b/funasr/models/sanm/attention.py
index 08f7dc7..c7e8a8e 100644
--- a/funasr/models/sanm/attention.py
+++ b/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)
--
Gitblit v1.9.1