Update Llama convert: load eot id from json

yvonwin · Apr 26, 2024 · 82c89cf · 82c89cf
1 parent 2cf8901
commit 82c89cf
Showing 1 changed file with 9 additions and 11 deletions.
diff --git a/qwen_cpp/convert.py b/qwen_cpp/convert.py
@@ -1,5 +1,5 @@
 """
-Convert Hugging Face Qwen models to GGML format
+Convert Hugging Face Qwen and LLama models to GGML format
 """
 import argparse
 import platform
@@ -229,8 +229,7 @@ def dump_model(f, model, ggml_type):
 
 class Llama3Converter:
     """
-    主要是tokenizer的区别。
-    还有qkv没有bias
+    qkv has no bias
     """
     MODEL_TYPE = ModelType.Llama3
     @classmethod
@@ -251,10 +250,10 @@ def dump_config(f, config, generation_config, tokenizer, ggml_type):
             config.num_hidden_layers,
             config.intermediate_size,
             config.max_position_embeddings,
-            config.eos_token_id,                             # eos 151645
-            128001, # pad
-            128000, # <|begin_of_text|>
-            128009, # "<|end_of_text|>"
+            config.eos_token_id,       # <|end_of_text|> 128001
+            config.eos_token_id, # llama3 no pad, so not use actually
+            config.bos_token_id, # <|begin_of_text|> 128000
+            generation_config.eos_token_id[1] if isinstance(generation_config.eos_token_id, list) else 128009 # <|eot_id|> 128009
         ]
         f.write(struct.pack("i" * len(config_values), *config_values))
 
@@ -278,7 +277,7 @@ def dump_model(f, model, ggml_type):
             "model.norm.weight",
             "lm_head.weight",
         ]
-        print(len(weight_names))
+        # print(len(weight_names)) // 8b: 291
         dump_state_dict(f, weight_names, model.state_dict(), ggml_type)
 
 class Qwen2Converter:
@@ -461,7 +460,7 @@ def convert(f: BinaryIO, model_name_or_path: str, dtype: str = "q4_0"):
     tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
     model = AutoModelForCausalLM.from_pretrained(model_name_or_path)
 
-    print(model)
+    # print(model)
     # state_dict = model.state_dict()
     # keys = state_dict.keys()
     # print(keys)
@@ -471,7 +470,6 @@ def convert(f: BinaryIO, model_name_or_path: str, dtype: str = "q4_0"):
     # print(model.config.eos_token_id) # 151645
     # print(model.generation_config)
     # print(tokenizer.model_max_length)
-    # print(list(tokenizer.added_tokens_decoder.keys())) # 1.5 only
 
     if model.config.architectures[0]=="Qwen2ForCausalLM":
         if "Code" not in model_name_or_path:
@@ -498,7 +496,7 @@ def main():
     parser.add_argument(
         "-i",
         "--model_name_or_path",
-        default="Qwen/Qwen1.5-1.8B-Chat", # Qwen/Qwen1.5-0.5B-Chat; Qwen/Qwen1.5-MoE-A2.7B-Chat
+        default="Qwen/Qwen1.5-1.8B-Chat", # Qwen/Qwen1.5-7B-Chat; meta-llama/Meta-Llama-3-8B-Instruct
         type=str,
         help="Model name or path used in AutoModel.from_pretrained",
     )