add attn_weights judgment and remove inference code in train_forward

huggingface · Mar 6, 2025 · bc34f7f · bc34f7f
1 parent 1dc414b
commit bc34f7f
Showing 1 changed file with 23 additions and 36 deletions.
diff --git a/optimum/habana/transformers/models/deepseek_v2/modeling_deepseek_v2.py b/optimum/habana/transformers/models/deepseek_v2/modeling_deepseek_v2.py
@@ -1019,6 +1019,7 @@ def train_forward(
         attention_mask: Optional[torch.Tensor] = None,
         position_ids: Optional[torch.LongTensor] = None,
         past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
         token_idx: Optional[torch.Tensor] = None,
         cache_position: Optional[torch.LongTensor] = None,
         attn_softmax_bf16: Optional[bool] = False,
@@ -1089,8 +1090,22 @@ def train_forward(
             key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
         # optimization
         if use_flash_attention and FusedSDPA is not None:
-            if q_len == 1:
-                # next token
+            softmax_mode = "fast" if flash_attention_fast_softmax else "None"
+            if flash_attention_causal_mask:
+                attn_output = self.fused_scaled_dot_product_attention(
+                    query_states,
+                    key_states,
+                    value_states,
+                    None,
+                    0.0,
+                    True,
+                    None,
+                    softmax_mode,
+                    flash_attention_recompute,
+                    valid_sequence_lengths,
+                    "left",
+                )
+            else:
                 attn_output = self.fused_scaled_dot_product_attention(
                     query_states,
                     key_states,
@@ -1099,43 +1114,11 @@ def train_forward(
                     0.0,
                     False,
                     None,
-                    "None",
-                    False,
+                    softmax_mode,
+                    flash_attention_recompute,
                     None,
                     "None",
                 )
-            else:
-                # first token
-                softmax_mode = "fast" if flash_attention_fast_softmax else "None"
-                if flash_attention_causal_mask:
-                    attn_output = self.fused_scaled_dot_product_attention(
-                        query_states,
-                        key_states,
-                        value_states,
-                        None,
-                        0.0,
-                        True,
-                        None,
-                        softmax_mode,
-                        flash_attention_recompute,
-                        valid_sequence_lengths,
-                        "left",
-                    )
-                else:
-                    attn_output = self.fused_scaled_dot_product_attention(
-                        query_states,
-                        key_states,
-                        value_states,
-                        attention_mask,
-                        0.0,
-                        False,
-                        None,
-                        softmax_mode,
-                        flash_attention_recompute,
-                        None,
-                        "None",
-                    )
-
         else:
             query_states, key_states, value_states, attention_mask = gaudi_deepseekv2_repeat_kv(
                 query_states, key_states, value_states, attention_mask, self.num_key_value_groups
@@ -1159,6 +1142,9 @@ def train_forward(
             attn_weights = torch.nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
             attn_output = self.matmul_av(attn_weights, value_states)
 
+        if not output_attentions:
+            attn_weights = None
+
         return attn_output, attn_weights, past_key_value
 
     def prefill_forward(
@@ -1428,6 +1414,7 @@ def forward(
                 attention_mask=attention_mask,
                 position_ids=position_ids,
                 past_key_value=past_key_value,
+                output_attentions=output_attentions,
                 token_idx=token_idx,
                 cache_position=cache_position,
                 attn_softmax_bf16=attn_softmax_bf16,