[Qwen2] fix lora_demo

1. uint64_t header_size
2. fix load_and_infer_embedding

models/Qwen2/lora_demo/chat.cpp

@@ -107,7 +107,7 @@ class Qwen {
   int stage_idx;
   bool enable_lora_embedding;
   bool make_in_tensors_flag;
-  int header_size;
+  uint64_t header_size;
 
   // model
   int hidden_bytes;
@@ -663,15 +663,15 @@ Qwen::load_and_infer_embedding(const std::vector<int> &tokens) {
   }
 
   file.seekg(0, std::ios::end);
-  long long file_size = static_cast<long long>(file.tellg());
-  if (file_size < header_size) {
+  uint64_t file_size_in_disk = file.tellg();
+  if (file_size_in_disk < header_size) {
     throw std::runtime_error("file is too small\n");
   }
 
-  std::vector<uint32_t> header(header_size / sizeof(uint32_t));
+  uint64_t file_size_in_header = 0;
   file.seekg(0, std::ios::beg);
-  file.read(reinterpret_cast<char *>(&header[0]), sizeof(uint32_t));
-  if (file_size != header[0]) {
+  file.read(reinterpret_cast<char *>(&file_size_in_header), sizeof(uint64_t));
+  if (file_size_in_disk != file_size_in_header) {
     throw std::runtime_error("Error: file size is not equal to file size in header\n");
   }
 

models/Qwen2/lora_demo/export_onnx.py

@@ -413,8 +413,9 @@ def convert_lora_embedding_to_bit(lora_model, lora_config, args):
         else:
             raise NotImplementedError
 
-    # 由于在final.mlir中，weight的权重排列顺序是[lora_A, lora_B]的形式
-    # 所以需要把B排列在前面
+    # 由于在final.mlir中，weight的权重排列顺序是[lora_B, lora_A]的形式
+    # 但是在加载时，是按照算子调用逻辑来调用的，lora_A先走先调，lora_B后跑后调
+    # 所以需要把A排列在前面
     for a, b in zip(lora_A_weight_list, lora_B_weight_list):
         lora_weight_list.append(a)
         lora_weight_list.append(b)
@@ -435,11 +436,11 @@ def convert_lora_embedding_to_bit(lora_model, lora_config, args):
     return lora_weights_uint8
 
 def make_header(size, header_size = 64):
-    if header_size < 4:
+    if header_size < 8:
         raise ValueError("Header size must be at least 4 bytes to store the size.")
     header = np.zeros(header_size, dtype=np.uint8)
-    size_bytes = struct.pack('<I', header_size + size)
-    header[:4] = np.frombuffer(size_bytes, dtype=np.uint8)
+    size_bytes = struct.pack('<Q', header_size + size)
+    header[:8] = np.frombuffer(size_bytes, dtype=np.uint8)
     return header
 
 def convert_total_lora_to_bit(encrypt_path, origin_model, args):