optimize qwen2_vl multiple image input or video input (#12487)

python/llm/src/ipex_llm/transformers/models/qwen2_vl.py

@@ -191,37 +191,66 @@ def qwen2_vision_attention_forward(
                                               ).permute(1, 0, 2, 3).unbind(0)
     q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
     k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb).squeeze(0)
+    # q, k, v: [seq_length, num_heads, head_dim]
 
-    q = q.transpose(0, 1)
-    k = k.transpose(0, 1)
-    v = v.transpose(0, 1)
+    seq_lens = cu_seqlens.tolist()
+    invalidInputError(seq_lens[0] == 0 and seq_lens[-1] == seq_length,
+                      "unexpected input")
 
-    if len(cu_seqlens) == 2 and cu_seqlens.tolist() == [0, seq_length]:
-        attention_mask = None
+    if use_sdp_non_causal(self.head_dim, q.device, q.dtype):
+        import xe_addons
+        image_num = len(seq_lens) - 1
+        image_size = seq_lens[1] - seq_lens[0]
+        guessed_seq_lens = torch.arange(0, (image_num + 1) * image_size, image_size,
+                                        dtype=cu_seqlens.dtype, device=cu_seqlens.device)
+        if (guessed_seq_lens == cu_seqlens).all():
+            q = q.view(image_num, image_size, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
+            k = k.view(image_num, image_size, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
+            v = v.view(image_num, image_size, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
+            # q, k, v: [image_num, num_heads, image_size, head_dim]
+
+            attn_output = xe_addons.sdp_non_causal(q, k.contiguous(), v.contiguous(), None)
+            attn_output = attn_output.permute(0, 2, 1, 3).contiguous()
+            attn_output = attn_output.view(seq_length, self.num_heads, self.head_dim)
+            # attn_output: [seq_length, num_heads, head_dim]
+        else:
+            q = q.transpose(0, 1).unsqueeze(0)
+            k = k.transpose(0, 1).unsqueeze(0).contiguous()
+            v = v.transpose(0, 1).unsqueeze(0).contiguous()
+            # q, k, v: [1, num_heads, seq_length, head_dim]
+
+            attn_outputs = []
+            for i in range(image_num):
+                start_idx = seq_lens[i]
+                end_idx = seq_lens[i + 1]
+                tmp_q = q[:, :, start_idx:end_idx, :]
+                tmp_k = k[:, :, start_idx:end_idx, :]
+                tmp_v = v[:, :, start_idx:end_idx, :]
+                attn_output = xe_addons.sdp_non_causal(tmp_q, tmp_k, tmp_v, None)
+                attn_output = attn_output.permute(0, 2, 1, 3)
+                # attn_output: [1, seq_length, num_heads, head_dim]
+                attn_outputs.append(attn_output)
+            attn_output = torch.cat(attn_outputs, dim=1).squeeze(0)
+            # attn_output: [seq_length, num_heads, head_dim]
     else:
         attention_mask = torch.full(
             [1, seq_length, seq_length], torch.finfo(q.dtype).min, device=q.device, dtype=q.dtype
         )
-        for i in range(1, len(cu_seqlens)):
-            attention_mask[..., cu_seqlens[i - 1]:cu_seqlens[i],
-                           cu_seqlens[i - 1]:cu_seqlens[i]] = 0
+        for i in range(1, len(seq_lens)):
+            attention_mask[..., seq_lens[i - 1]:seq_lens[i], seq_lens[i - 1]:seq_lens[i]] = 0
+
+        q = q.transpose(0, 1)
+        k = k.transpose(0, 1)
+        v = v.transpose(0, 1)
+        # q, k, v: [num_heads, seq_length, head_dim]
 
-    if use_sdp_non_causal(self.head_dim, q.device, q.dtype):
-        import xe_addons
-        q = q.unsqueeze(0)
-        k = k.unsqueeze(0)
-        v = v.unsqueeze(0)
-        if attention_mask is not None:
-            attention_mask = attention_mask.unsqueeze(0)
-        attn_output = xe_addons.sdp_non_causal(q, k.contiguous(), v.contiguous(), attention_mask)
-        attn_output = attn_output.squeeze(0)
-    else:
         attn_weights = torch.matmul(q, k.transpose(1, 2)) / math.sqrt(self.head_dim)
-        if attention_mask is not None:
-            attn_weights = attn_weights + attention_mask
+        attn_weights = attn_weights + attention_mask
         attn_weights = attention_softmax(attn_weights)
         attn_output = torch.matmul(attn_weights, v)
-    attn_output = attn_output.transpose(0, 1)
+        attn_output = attn_output.transpose(0, 1)
+        # attn_output: [seq_length, num_heads, head_dim]
+
     attn_output = attn_output.reshape(seq_length, -1)
     attn_output = self.proj(attn_output)
     return attn_output