🧠 两小时从 0 训出你的 64M 大模型。Train your own 64M LLM from scratch in 2 hours.
update export_model
gongjy13e791e51612
diff --git a/export_model.py b/export_model.pyindex fd019c1..3305299 100644--- a/export_model.py+++ b/export_model.py@@ -18,7 +18,7 @@ def export_transformers_model():lm_model = Transformer(lm_config)device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')moe_path = '_moe' if lm_config.use_moe else ''- ckpt_path = f'./out/single_chat/full_sft_{lm_config.dim}{moe_path}.pth'+ ckpt_path = f'./out/full_sft_{lm_config.dim}{moe_path}.pth'state_dict = torch.load(ckpt_path, map_location=device)unwanted_prefix = '_orig_mod.'@@ -28,13 +28,13 @@ def export_transformers_model():lm_model.load_state_dict(state_dict, strict=False)print(f'模型参数: {count_parameters(lm_model) / 1e6} 百万 = {count_parameters(lm_model) / 1e9} B (Billion)')- lm_model.save_pretrained("minimind-small", safe_serialization=False)+ lm_model.save_pretrained("minimind-small-T", safe_serialization=False)def export_tokenizer():- tokenizer = AutoTokenizer.from_pretrained('./model',+ tokenizer = AutoTokenizer.from_pretrained('./model/minimind_tokenizer',trust_remote_code=True, use_fast=False)- tokenizer.save_pretrained("minimind-small")+ tokenizer.save_pretrained("minimind-small-T")def push_to_hf():