Update eagle/model/modeling_minicpm_kv.py

eagle/model/modeling_minicpm_kv.py

@@ -1589,9 +1589,9 @@ class MiniCPMDecoderLayer(nn.Module):
             self.self_attn = MiniCPMInfLLMv2Attention(config=config, layer_idx=layer_idx)
         else:
             # <mod>
-            # self.self_attn = MINICPM_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
+            self.self_attn = MINICPM_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
             # <before-after-mod> -------------------------------------------------
-            self.self_attn = MINICPM_ATTENTION_CLASSES["eager"](config=config, layer_idx=layer_idx)
+            # self.self_attn = MINICPM_ATTENTION_CLASSES["eager"](config=config, layer_idx=layer_idx)
             # </mod>
 
         self.mlp = MiniCPMMLP(config)
@@ -1797,7 +1797,7 @@ class MiniCPMModel(MiniCPMPreTrainedModel):
         self.layers = nn.ModuleList(
             [MiniCPMDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
         )
-        # <mod> dev: support sdpa only
+        # <mod>
         # self._use_sdpa = config._attn_implementation == 'sdpa'
         # self._use_flash_attention_2 = config._attn_implementation == 'flash_attention_2'
         # <before-after-mod> -------------------------------------------------