Update chatNT.py
Browse files
chatNT.py
CHANGED
|
@@ -925,7 +925,11 @@ class TorchGptGroupedQueryAttention(nn.Module):
|
|
| 925 |
)
|
| 926 |
|
| 927 |
attention_weights = nn.functional.softmax(attention_logits, dim=-1)
|
|
|
|
| 928 |
|
|
|
|
|
|
|
|
|
|
| 929 |
values = torch.einsum("bhtT,bThd->bthd", attention_weights, values)
|
| 930 |
values = values.contiguous().view(batch_size, seq_len, -1)
|
| 931 |
|
|
@@ -1334,8 +1338,6 @@ class MultiHeadAttention(nn.Module):
|
|
| 1334 |
else:
|
| 1335 |
attention_weights = F.softmax(attention_weights, dim=-1)
|
| 1336 |
|
| 1337 |
-
print(f"Attention weights : {attention_weights.dtype}")
|
| 1338 |
-
print(f"Value heads : {value_heads.dtype}")
|
| 1339 |
value_out = torch.einsum(
|
| 1340 |
"...htT, ...Thd->...thd", attention_weights, value_heads
|
| 1341 |
)
|
|
|
|
| 925 |
)
|
| 926 |
|
| 927 |
attention_weights = nn.functional.softmax(attention_logits, dim=-1)
|
| 928 |
+
attention_weights = attention_weights.to(values.dtype)
|
| 929 |
|
| 930 |
+
print(f"Attention weights type : ", attention_weights.dtype)
|
| 931 |
+
print(f"Values type : ", values.dtype)
|
| 932 |
+
|
| 933 |
values = torch.einsum("bhtT,bThd->bthd", attention_weights, values)
|
| 934 |
values = values.contiguous().view(batch_size, seq_len, -1)
|
| 935 |
|
|
|
|
| 1338 |
else:
|
| 1339 |
attention_weights = F.softmax(attention_weights, dim=-1)
|
| 1340 |
|
|
|
|
|
|
|
| 1341 |
value_out = torch.einsum(
|
| 1342 |
"...htT, ...Thd->...thd", attention_weights, value_heads
|
| 1343 |
)
|