Spaces:
Running
on
Zero
Running
on
Zero
Implement improved attention masking for bidirectional_masked
Browse files- llama_diffusion_model.py +9 -3
llama_diffusion_model.py
CHANGED
@@ -47,10 +47,16 @@ class BidirectionalLlamaAttention(LlamaAttention):
|
|
47 |
key_states = self.repeat_kv(key, module.num_key_value_groups)
|
48 |
value_states = self.repeat_kv(value, module.num_key_value_groups)
|
49 |
|
50 |
-
attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
|
|
|
|
|
|
|
|
51 |
if attention_mask is not None:
|
52 |
-
|
53 |
-
|
|
|
|
|
54 |
|
55 |
attn_weights = nn.functional.softmax(attn_weights, dim=-1).to(query.dtype)
|
56 |
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|
|
|
47 |
key_states = self.repeat_kv(key, module.num_key_value_groups)
|
48 |
value_states = self.repeat_kv(value, module.num_key_value_groups)
|
49 |
|
50 |
+
# attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
|
51 |
+
# if attention_mask is not None:
|
52 |
+
# causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
|
53 |
+
# attn_weights = attn_weights + causal_mask
|
54 |
+
|
55 |
if attention_mask is not None:
|
56 |
+
# Convert bool -> float with -inf where masked
|
57 |
+
attn_mask = attention_mask.masked_fill(~attention_mask, float('-inf')).to(query.dtype)
|
58 |
+
attn_weights = attn_weights + attn_mask
|
59 |
+
|
60 |
|
61 |
attn_weights = nn.functional.softmax(attn_weights, dim=-1).to(query.dtype)
|
62 |
attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training)
|