Update README.md

README.md

@@ -1,3 +1,100 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+---
+# maywell/EXAONE-3.0-7.8B-Instruct-Llamafied
+
+LG에서 동일 라이센스 재배포조차 막아버린 관계로 모델을 공유할 수 없게 되었습니다.
+vLLM, 추론 및 기타 활용으로 Llamafied 모델이 필요하다면 아래 스크립트를 실행해서 사용해주시면 감사하겠습니다.
+
+```python
+import torch
+from transformers import LlamaConfig, LlamaForCausalLM
+from modeling_exaone import ExaoneForCausalLM
+import gc
+
+def load_model(model_path, model_class, torch_dtype=torch.bfloat16):
+    """Load and return a model given its path and class."""
+    return model_class.from_pretrained(model_path, torch_dtype=torch_dtype, device_map="auto")
+
+def unload_model(model):
+    """Clear memory by deleting a model and calling the garbage collector."""
+    del model
+    gc.collect()
+    torch.cuda.empty_cache()
+
+def create_llama_config(exaone_config):
+    """Create and return a LLaMA configuration based on EXAONE config."""
+    return LlamaConfig(
+        vocab_size=exaone_config.vocab_size,
+        hidden_size=exaone_config.hidden_size,
+        intermediate_size=exaone_config.intermediate_size,
+        num_hidden_layers=exaone_config.num_layers,
+        num_attention_heads=exaone_config.num_attention_heads,
+        max_position_embeddings=exaone_config.max_position_embeddings,
+        rms_norm_eps=exaone_config.layer_norm_epsilon,
+        num_key_value_heads=exaone_config.num_key_value_heads,
+        rope_theta=exaone_config.rope_theta,
+        attention_bias=False,
+    )
+
+def copy_embedding_weights(llama_model, exaone_model):
+    """Copy embedding weights from EXAONE to LLaMA model."""
+    llama_model.model.embed_tokens.weight.data = exaone_model.transformer.wte.weight.data.to(llama_model.device)
+
+def copy_layer_weights(llama_layer, exaone_layer, device):
+    """Copy weights for a single layer from EXAONE to LLaMA model."""
+    # Self-attention
+    llama_layer.self_attn.q_proj.weight.data = exaone_layer.attn.attention.q_proj.weight.data.to(device)
+    llama_layer.self_attn.k_proj.weight.data = exaone_layer.attn.attention.k_proj.weight.data.to(device)
+    llama_layer.self_attn.v_proj.weight.data = exaone_layer.attn.attention.v_proj.weight.data.to(device)
+    llama_layer.self_attn.o_proj.weight.data = exaone_layer.attn.attention.out_proj.weight.data.to(device)
+    # MLP
+    llama_layer.mlp.gate_proj.weight.data = exaone_layer.mlp.c_fc_0.weight.data.to(device)
+    llama_layer.mlp.up_proj.weight.data = exaone_layer.mlp.c_fc_1.weight.data.to(device)
+    llama_layer.mlp.down_proj.weight.data = exaone_layer.mlp.c_proj.weight.data.to(device)
+    # Layer Norms
+    llama_layer.input_layernorm.weight.data = exaone_layer.ln_1.weight.data.to(device)
+    llama_layer.post_attention_layernorm.weight.data = exaone_layer.ln_2.weight.data.to(device)
+
+def copy_final_weights(llama_model, exaone_model):
+    """Copy final layer norm and LM head weights from EXAONE to LLaMA model."""
+    llama_model.model.norm.weight.data = exaone_model.transformer.ln_f.weight.data.to(llama_model.device)
+    llama_model.lm_head.weight.data = exaone_model.lm_head.weight.data.to(llama_model.device)
+
+def port_exaone_to_llama(exaone_model_path, llama_model_path):
+    print("Loading EXAONE model...")
+    exaone_model = load_model(exaone_model_path, ExaoneForCausalLM)
+    exaone_config = exaone_model.config
+
+    print("Creating LLaMA configuration...")
+    llama_config = create_llama_config(exaone_config)
+
+    print("Initializing LLaMA model...")
+    llama_model = LlamaForCausalLM(llama_config)
+    llama_model.to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
+
+    print("Copying weights...")
+    copy_embedding_weights(llama_model, exaone_model)
+
+    for i in range(exaone_config.num_layers):
+        print(f"Copying weights for layer {i+1}/{exaone_config.num_layers}")
+        copy_layer_weights(llama_model.model.layers[i], exaone_model.transformer.h[i], llama_model.device)
+
+    copy_final_weights(llama_model, exaone_model)
+
+    print("Unloading EXAONE model to free memory...")
+    unload_model(exaone_model)
+
+    print(f"Saving ported LLaMA model to {llama_model_path}")
+    llama_model.save_pretrained(llama_model_path, safe_serialization=True, max_shard_size="5GB")
+
+    print("Unloading LLaMA model...")
+    unload_model(llama_model)
+
+    print(f"EXAONE model successfully ported to LLaMA format and saved at {llama_model_path}")
+
+if __name__ == "__main__":
+    exaone_model_path = "LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct"
+    llama_model_path = "./exa_llamafied"
+    port_exaone_to_llama(exaone_model_path, llama_model_path)
+```