import os
import argparse
from safetensors import safe_open
from safetensors.torch import save_file

def split_sdxl(input_path):
    if not os.path.exists(input_path):
        print(f"错误: 找不到输入文件 '{input_path}'")
        return

    base_dir = os.path.dirname(input_path)
    base_name = os.path.splitext(os.path.basename(input_path))[0]
    
    clipL_path = os.path.join(base_dir, f"{base_name}_clipL.safetensors")
    clipG_path = os.path.join(base_dir, f"{base_name}_clipG.safetensors")
    vae_path = os.path.join(base_dir, f"{base_name}_VAE.safetensors")
    unet_path = os.path.join(base_dir, f"{base_name}_Diffusion.safetensors")

    clipL_tensors = {}
    clipG_tensors = {}
    vae_tensors = {}
    diffusion_tensors = {}

    print(f"📦 正在执行 SDXL 纯净切分与权重重组: {input_path} ...")

    with safe_open(input_path, framework="pt", device="cpu") as f:
        all_keys = f.keys()
        
        # 动态判断 CLIP-G 是否为原生 OpenCLIP 格式
        # 如果有 resblocks 或是 model.transformer.resblocks，说明需要转换
        is_openclip_g = any("resblocks." in k for k in all_keys if k.startswith("conditioner.embedders.1."))

        for key in all_keys:
            tensor = f.get_tensor(key)
            
            # 1. 提取并清洗 CLIP L (768 维)
            if key.startswith("conditioner.embedders.0."):
                clean_key = key.replace("conditioner.embedders.0.", "", 1)
                if clean_key.startswith("transformer."):
                    clean_key = clean_key.replace("transformer.", "", 1)
                
                # 过滤 position_ids
                if "position_ids" in clean_key:
                    continue
                    
                clipL_tensors[clean_key] = tensor

            # 2. 提取并清洗 CLIP G (1280 维)
            elif key.startswith("conditioner.embedders.1."):
                clean_key = key.replace("conditioner.embedders.1.", "", 1)
                
                # 过滤全局标量
                if clean_key.endswith("logit_scale") or "position_ids" in clean_key:
                    continue

                if is_openclip_g:
                    # === OpenCLIP (如 Juggernaut) 格式转换 ===
                    # 剥离所有可能的前缀
                    openclip_key = clean_key.replace("model.", "", 1).replace("transformer.", "", 1)
                    
                    if "resblocks." in openclip_key:
                        new_key = openclip_key.replace("resblocks.", "text_model.encoder.layers.", 1)
                        new_key = new_key.replace(".ln_1.", ".layer_norm1.", 1)
                        new_key = new_key.replace(".ln_2.", ".layer_norm2.", 1)
                        new_key = new_key.replace(".mlp.c_fc.", ".mlp.fc1.", 1)
                        new_key = new_key.replace(".mlp.c_proj.", ".mlp.fc2.", 1)
                    elif openclip_key == "positional_embedding":
                        new_key = "text_model.embeddings.position_embedding.weight"
                    elif openclip_key == "token_embedding.weight":
                        new_key = "text_model.embeddings.token_embedding.weight"
                    elif openclip_key == "ln_final.weight":
                        new_key = "text_model.final_layer_norm.weight"
                    elif openclip_key == "ln_final.bias":
                        new_key = "text_model.final_layer_norm.bias"
                    elif openclip_key == "text_projection":
                        new_key = "text_projection.weight"
                    else:
                        new_key = openclip_key # 其他键保持原样

                    # 核心：处理复合权重 (in_proj)
                    if ".attn.in_proj_weight" in openclip_key:
                        base_layer = new_key.replace(".attn.in_proj_weight", ".self_attn", 1)
                        q, k, v = tensor.chunk(3, dim=0)
                        clipG_tensors[f"{base_layer}.q_proj.weight"] = q
                        clipG_tensors[f"{base_layer}.k_proj.weight"] = k
                        clipG_tensors[f"{base_layer}.v_proj.weight"] = v
                    elif ".attn.in_proj_bias" in openclip_key:
                        base_layer = new_key.replace(".attn.in_proj_bias", ".self_attn", 1)
                        q, k, v = tensor.chunk(3, dim=0)
                        clipG_tensors[f"{base_layer}.q_proj.bias"] = q
                        clipG_tensors[f"{base_layer}.k_proj.bias"] = k
                        clipG_tensors[f"{base_layer}.v_proj.bias"] = v
                    elif ".attn.out_proj.weight" in openclip_key:
                        new_key = new_key.replace(".attn.out_proj.weight", ".self_attn.out_proj.weight", 1)
                        clipG_tensors[new_key] = tensor
                    elif ".attn.out_proj.bias" in openclip_key:
                        new_key = new_key.replace(".attn.out_proj.bias", ".self_attn.out_proj.bias", 1)
                        clipG_tensors[new_key] = tensor
                    else:
                        clipG_tensors[new_key] = tensor
                else:
                    # === 已经是标准 Transformers 格式 ===
                    if clean_key.startswith("transformer."):
                        clean_key = clean_key.replace("transformer.", "", 1)
                    clipG_tensors[clean_key] = tensor

            # 3. 提取并清洗 VAE
            elif key.startswith("first_stage_model."):
                vae_tensors[key.replace("first_stage_model.", "", 1)] = tensor
                
            # 4. 提取并清洗 UNet (Diffusion)
            elif key.startswith("model.diffusion_model."):
                diffusion_tensors[key.replace("model.diffusion_model.", "", 1)] = tensor

    print("\n💾 正在保存组件...")
    
    if clipL_tensors:
        save_file(clipL_tensors, clipL_path)
        print(f"  --> CLIP-L 保存成功 (键数: {len(clipL_tensors)})")
    if clipG_tensors:
        save_file(clipG_tensors, clipG_path)
        print(f"  --> CLIP-G 保存成功 (键数: {len(clipG_tensors)})") # 正常应该输出 517
    if vae_tensors: 
        save_file(vae_tensors, vae_path)
        print(f"  --> VAE 保存成功 (键数: {len(vae_tensors)})")
    if diffusion_tensors: 
        save_file(diffusion_tensors, unet_path)
        print(f"  --> UNet 保存成功 (键数: {len(diffusion_tensors)})")

    print("\n🎉 SDXL 切分任务完成！")

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="纯净 SDXL 模型切分与格式转换工具")
    parser.add_argument("input_file", type=str, help="输入的 SDXL 大模型文件路径")
    args = parser.parse_args()
    split_sdxl(args.input_file)