fix: v.to(q.dtype) prima di SDPA
Browse files- modeling_quark.py +1 -0
modeling_quark.py
CHANGED
|
@@ -77,6 +77,7 @@ class GroupedQueryAttention(nn.Module):
|
|
| 77 |
if self.n_groups > 1:
|
| 78 |
k = k.repeat_interleave(self.n_groups, dim=1)
|
| 79 |
v = v.repeat_interleave(self.n_groups, dim=1)
|
|
|
|
| 80 |
out = F.scaled_dot_product_attention(
|
| 81 |
q, k, v, attn_mask=None,
|
| 82 |
dropout_p=self.drop if self.training else 0.0,
|
|
|
|
| 77 |
if self.n_groups > 1:
|
| 78 |
k = k.repeat_interleave(self.n_groups, dim=1)
|
| 79 |
v = v.repeat_interleave(self.n_groups, dim=1)
|
| 80 |
+
v = v.to(q.dtype) # allinea dtype dopo RoPE (cos/sin sono float32)
|
| 81 |
out = F.scaled_dot_product_attention(
|
| 82 |
q, k, v, attn_mask=None,
|
| 83 |
dropout_p=self.drop if self.training else 0.0,
|