Update modeling_phi.py

modeling_phi.py

@@ -302,6 +302,9 @@ class PhiAttention(nn.Module):
             else:
                 raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
 
+    # Phi-2 has an attention overflow issue (with FP16) and requires autocast to be disabled
+    @torch.autocast("cpu", enabled=False)
+    @torch.autocast("cuda", enabled=False)
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -359,10 +362,7 @@ class PhiAttention(nn.Module):
         key_states = repeat_kv(key_states, self.num_key_value_groups)
         value_states = repeat_kv(value_states, self.num_key_value_groups)
 
-        # Queries and keys upcast to fp32 is required by Phi-2 to avoid overflow
-        attn_weights = torch.matmul(
-            query_states.to(torch.float32), key_states.to(torch.float32).transpose(2, 3)
-        ) / math.sqrt(self.head_dim)
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
 
         if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
             raise ValueError(