optimize internlm xcomposer2 performance (#11550)

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

@@ -42,6 +42,7 @@ from typing import Optional, Tuple, List
 import torch
 import torch.utils.checkpoint
 from torch import nn
+from ipex_llm.utils.common.log4Error import invalidInputError
 from ipex_llm.transformers.models.utils import should_use_fuse_rope, apply_rotary_pos_emb
 from ipex_llm.transformers.models.utils import apply_rotary_pos_emb_cache_freq_xpu
 from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp8_kv_cache
@@ -307,9 +308,15 @@ def pre_process_attn_and_mlp(module: torch.nn.Module):
 def add_lora(x: torch.Tensor, result: torch.Tensor,
              im_mask: torch.Tensor = None, lora_scaling: float = 0,
              Plora_A: torch.nn.Linear = None, Plora_B: torch.nn.Linear = None):
-    if im_mask is not None and torch.sum(im_mask) > 0:
-        part_x = x[im_mask]
-        result[im_mask] += Plora_B(Plora_A(part_x) * lora_scaling)
+    invalidInputError(x.dim() == 3 and result.dim() == 3,
+                      "`x` and `result` should have 3 dims")
+    if len(im_mask) == 0 or x.size(1) == 1:
+        return result
+    else:
+        for start_idx, end_idx in im_mask:
+            result[:, start_idx:end_idx, :] += Plora_B(
+                Plora_A(x[:, start_idx:end_idx, :]) * lora_scaling
+            )
         return result
 
 
@@ -457,18 +464,32 @@ def internlm_xcomposser2_chat(
     **kwargs,
 ):
     # ipex-llm changes start: fix device and dtype conversion
+    # replace im_mask with start_idx and end_idx to improve performance
     if image is None:
         inputs = self.build_inputs(tokenizer, query, history, meta_instruction)
-        im_mask = torch.zeros(inputs['input_ids'].shape[:2]).bool()
+        im_mask = []
     else:
         image = self.encode_img(image)
         inputs, im_mask = self.interleav_wrap_chat(tokenizer, query, image,
                                                    history, meta_instruction)
+        mask = im_mask.cpu().flatten().tolist()
+        length = len(mask)
+        im_mask = []
+        i = 0
+        while i < length:
+            while i < length and not mask[i]:
+                i = i + 1
+            start_idx = i
+            while i < length and mask[i]:
+                i = i + 1
+            end_idx = i
+            if start_idx != end_idx:
+                im_mask.append((start_idx, end_idx))
+
     inputs = {
         k: v.to(device=self.device, dtype=self.dtype)
         for k, v in inputs.items() if torch.is_tensor(v)
     }
-    im_mask = im_mask.to(self.device)
     # ipex-llm changes end
 
     # also add end-of-assistant token in eos token id to avoid unnecessary generation