Update modeling_phi.py
Browse files- modeling_phi.py +4 -4
modeling_phi.py
CHANGED
@@ -302,6 +302,9 @@ class PhiAttention(nn.Module):
|
|
302 |
else:
|
303 |
raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
|
304 |
|
|
|
|
|
|
|
305 |
def forward(
|
306 |
self,
|
307 |
hidden_states: torch.Tensor,
|
@@ -359,10 +362,7 @@ class PhiAttention(nn.Module):
|
|
359 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
360 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
361 |
|
362 |
-
|
363 |
-
attn_weights = torch.matmul(
|
364 |
-
query_states.to(torch.float32), key_states.to(torch.float32).transpose(2, 3)
|
365 |
-
) / math.sqrt(self.head_dim)
|
366 |
|
367 |
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
368 |
raise ValueError(
|
|
|
302 |
else:
|
303 |
raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
|
304 |
|
305 |
+
# Phi-2 has an attention overflow issue (with FP16) and requires autocast to be disabled
|
306 |
+
@torch.autocast("cpu", enabled=False)
|
307 |
+
@torch.autocast("cuda", enabled=False)
|
308 |
def forward(
|
309 |
self,
|
310 |
hidden_states: torch.Tensor,
|
|
|
362 |
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
363 |
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
364 |
|
365 |
+
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
|
|
|
|
|
|
366 |
|
367 |
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
368 |
raise ValueError(
|