Create copy_weights.py

copy_weights.py

@@ -0,0 +1,45 @@
+import torch
+from JiRackTernary_new import JiRackConfig, JiRackTernary1B
+from transformers import AutoTokenizer
+import os
+
+print("🚀 Copying embeddings and lm_head...")
+
+old_model_path = "." 
+new_tokenizer_path = "./jirack_code_tokenizer_fixed"
+save_path = "./JiRack_init_new_vocab"
+
+os.makedirs(save_path, exist_ok=True)
+
+# Load new tokenizer
+tokenizer = AutoTokenizer.from_pretrained(new_tokenizer_path)
+new_vocab_size = len(tokenizer)
+
+# Create new model with updated vocab size
+config = JiRackConfig()
+model = JiRackTernary1B(config)
+
+# Load old model weights
+old_state = torch.load(f"{old_model_path}/pytorch_model.bin", map_location="cpu")
+
+old_vocab_size = 128256
+
+with torch.no_grad():
+    # Copy old weights
+    model.token_emb.weight[:old_vocab_size] = old_state['token_emb.weight'][:old_vocab_size]
+    model.lm_head.weight[:old_vocab_size] = old_state['lm_head.weight'][:old_vocab_size]
+    
+    # Initialize the new 3 tokens (FIM markers) with mean value
+    mean_emb = old_state['token_emb.weight'].mean(dim=0)
+    model.token_emb.weight[old_vocab_size:] = mean_emb
+    model.lm_head.weight[old_vocab_size:] = mean_emb
+
+print(f"✅ Copied {old_vocab_size} tokens")
+print(f"✅ Initialized {new_vocab_size - old_vocab_size} new tokens")
+
+# Save
+torch.save(model.state_dict(), f"{save_path}/pytorch_model.bin")
+tokenizer.save_pretrained(save_path)
+
+print(f"\n🎉 Done! New model saved to: {save_path}")
+print("Use this folder as the starting weights for training from scratch.")