Update convert_hf_cpm.py

quick_start_clean/convert_hf_cpm.py

@@ -17,20 +17,34 @@ from collections import OrderedDict
 
 import torch
 import argparse
+import os
 
 parser = argparse.ArgumentParser(description='Load and save model weights with specified paths.')
 parser.add_argument('--model_path', type=str, required=True, help='Path to the model directory.')
 parser.add_argument('--output_path', type=str, required=True, help='Path to save the new weights.')
-parser.add_argument('--layer_num', type=int, required=True, help='The layers of model')
-
+parser.add_argument('--model_type',type=str,default='fm9g',help='The model type need to be one of "fm9g" or "9g-8b"')
+parser.add_argument('--task',type=str,default='pt2bin',help='The task need to be one of "pt2bin" or "bin2pt"')
+# parser.add_argument('--layer_num', type=int, required=True, help='The layers of model')
 
 args = parser.parse_args()
 
 src_path = args.model_path
-dst_path = args.output_path
-layer_num = args.layer_num
+dst_path = args.output_path if args.output_path.endswith('/') else args.output_path + ('/')
+model_type = args.model_type
+task = args.task
+
+assert model_type in ['fm9g'], 'The "model_type" must be one of "fm9g"!'
+assert task in ['pt2bin','bin2pt'], 'The task need to be one of "pt2bin" or "bin2pt"!'
+
+if model_type == 'fm9g':
+    layer_num = 40
+
+if not os.path.exists(dst_path):
+    os.makedirs(dst_path)
+
 
 def convert_hf_to_fm9g():
+    # 2B模型转换bin2pt
     ckpt = torch.load(src_path)
     new_ckpt = OrderedDict()
 
@@ -48,14 +62,16 @@ def convert_hf_to_fm9g():
         new_ckpt[f"encoder.layers.{i}.ffn.ffn.w_in.w_1.weight"] = ckpt[f'model.layers.{i}.mlp.up_proj.weight']
         new_ckpt[f"encoder.layers.{i}.ffn.ffn.w_out.weight"] = ckpt[f'model.layers.{i}.mlp.down_proj.weight']
 
-    torch.save(new_ckpt, dst_path)
+    torch.save(new_ckpt, f"{dst_path}fm9g.pt")
 
 def convert_fm9g_to_hf():
+    #2B模型转换pt2bin
     state = torch.load(src_path)
+
     new_state = {}
     new_state["model.embed_tokens.weight"] = state["input_embedding.weight"]
     new_state["model.norm.weight"] = state["encoder.output_layernorm.weight"]
-    for lid in range(40):
+    for lid in range(layer_num):
         print(lid)
         new_state[f"model.layers.{lid}.self_attn.q_proj.weight"] = state[f"encoder.layers.{lid}.self_att.self_attention.project_q.weight"]
         new_state[f"model.layers.{lid}.self_attn.k_proj.weight"] = state[f"encoder.layers.{lid}.self_att.self_attention.project_k.weight"]
@@ -70,9 +86,13 @@ def convert_fm9g_to_hf():
         new_state[f"model.layers.{lid}.post_attention_layernorm.weight"] = state[f"encoder.layers.{lid}.ffn.layernorm_before_ffn.weight"]
     del state
     state = None
-    torch.save(new_state, f"{dst_path}pytorch_model.bin")
-    
+    torch.save(new_state, f"{dst_path}fm9g.bin")  
+
 
 if __name__ == "__main__":
-    convert_hf_to_fm9g()
- 
+    if model_type == 'fm9g' and task == 'bin2pt':
+        convert_hf_to_fm9g()
+    elif model_type == 'fm9g' and task == 'pt2bin':
+        convert_fm9g_to_hf()
+    else:
+        raise ValueError('Please check the model type and task!')