optimize internvl2 vision model's attention (#12198)

python/llm/src/ipex_llm/transformers/convert.py

@@ -1580,8 +1580,12 @@ def _optimize_post(model, lightweight_bmm=False):
         model.batch_chat = MethodType(internvl_batch_chat, model)
         if model.vision_model.__class__.__name__ == "InternVisionModel":
             from ipex_llm.transformers.models.internvl import _get_pos_embed
-            vision_embedding = model.vision_model.embeddings
+            from ipex_llm.transformers.models.internvl import intern_attention_forward
+            vision_model = model.vision_model
+            vision_embedding = vision_model.embeddings
             vision_embedding._get_pos_embed = MethodType(_get_pos_embed, vision_embedding)
+            vision_module = importlib.import_module(vision_model.__class__.__module__)
+            convert_forward(vision_model, vision_module.InternAttention, intern_attention_forward)
         _optimize_post(model.language_model, lightweight_bmm=lightweight_bmm)
     elif model.config.model_type == "qwen":
         if hasattr(model.config, "visual"):

python/llm/src/ipex_llm/transformers/models/internvl.py

@@ -163,3 +163,27 @@ def internvl_batch_chat(self, tokenizer, pixel_values, questions, generation_con
     responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True)
     responses = [response.split(template.sep)[0].strip() for response in responses]
     return responses
+
+
+def intern_attention_forward(self, x: torch.Tensor) -> torch.Tensor:
+    B, N, C = x.shape
+    qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+    q, k, v = qkv.unbind(0)  # make torchscript happy (cannot use tensor as tuple)
+
+    if self.qk_normalization:
+        B_, H_, N_, D_ = q.shape
+        q = self.q_norm(q.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
+        k = self.k_norm(k.transpose(1, 2).flatten(-2, -1)).view(B_, N_, H_, D_).transpose(1, 2)
+
+    if x.device.type == "xpu":
+        import xe_addons
+        x = xe_addons.sdp_non_causal(q.contiguous(), k.contiguous(), v.contiguous(), None)
+    else:
+        attn = ((q * self.scale) @ k.transpose(-2, -1))
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+        x = attn @ v
+    x = x.transpose(1, 2).reshape(B, N, C)
+    x = self.proj(x)
+    x = self.proj_drop(x)
+    return x