optimize minicpm v 2_6 firs token perf (#11770)

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

@@ -748,6 +748,8 @@ def _optimize_pre(model, qtype=None):
         from ipex_llm.transformers.models.llama import merge_qkv
         model.apply(merge_qkv)
     if model.config.model_type == "minicpmv":
+        from ipex_llm.transformers.models.minicpmv import merge_qkv
+        model.apply(merge_qkv)
         if model.config.hidden_size == 3584 and model.config.vocab_size == 151666:
             model.llm.config.model_type = "qwen2"
             _optimize_pre(model.llm, qtype=qtype)
@@ -1763,4 +1765,9 @@ def safe_bmm_fwd(*args, **kwargs):
         minicpmv_generate = minicpmv_generate_wrapper(module.MiniCPMV.generate)
         model.generate = MethodType(minicpmv_generate, model)
 
+        modeling_module_name = model.vpm.__class__.__module__
+        module = importlib.import_module(modeling_module_name)
+        from ipex_llm.transformers.models.minicpmv import siglip_attention_forward
+        convert_forward(model, module.SiglipAttention, siglip_attention_forward)
+
     return model

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

@@ -37,7 +37,10 @@ def merge_linear(linears: List[torch.nn.Linear]) -> torch.nn.Linear:
 
 
 def merge_qkv_base(module: torch.nn.Module, attention_class):
-    if isinstance(module, attention_class):
+    if (
+        isinstance(attention_class, str) and module.__class__.__name__ == attention_class
+        or not isinstance(attention_class, str) and isinstance(module, attention_class)
+    ):
         qkv_proj = merge_linear([
             module.q_proj,
             module.k_proj,

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

@@ -16,9 +16,45 @@
 
 
 import torch
+from typing import Optional
+from ipex_llm.transformers.models.common import merge_qkv_base
 from transformers.generation.logits_process import RepetitionPenaltyLogitsProcessor
 
 
+def merge_qkv(module: torch.nn.Module):
+    return merge_qkv_base(module, "SiglipAttention")
+
+
+def siglip_attention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    output_attentions: Optional[bool] = False,
+):
+    bsz, q_len, _ = hidden_states.size()
+
+    qkv = self.qkv_proj(hidden_states)
+    qkv = qkv.view(bsz, q_len, self.num_heads * 3, self.head_dim)
+    qkv = qkv.transpose(1, 2)
+    query_states, key_states, value_states = qkv.chunk(3, dim=1)
+
+    attn_weights = torch.matmul(query_states * self.scale, key_states.transpose(2, 3))
+    if attention_mask is not None:
+        attn_weights = attn_weights + attention_mask
+
+    # upcast attention to fp32
+    attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
+    attn_weights = torch.nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+    attn_output = torch.matmul(attn_weights, value_states)
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    attn_output = attn_output.reshape(bsz, q_len, self.embed_dim)
+
+    attn_output = self.out_proj(attn_output)
+
+    return attn_output, attn_weights
+
+
 def patched_repetition_penalty_call(self, input_ids: torch.LongTensor, scores: torch.FloatTensor):
     if scores.device.type == "xpu":
         import xe_addons