def create_model(source_model_id: str, source_revision: str, save_folder: str) -> str: import json from pathlib import Path import torch from huggingface_hub import hf_hub_download from safetensors.torch import save_file from transformers import AutoConfig, AutoModelForSequenceClassification, set_seed folder = Path(save_folder) folder.mkdir(parents=True, exist_ok=True) config = json.loads( Path(hf_hub_download(source_model_id, "config.json", revision=source_revision)).read_text(encoding="utf-8") ) text = config["text_config"] text.update( hidden_size=16, intermediate_size=32, head_dim=256, num_hidden_layers=8, num_attention_heads=8, num_key_value_heads=2, linear_key_head_dim=128, linear_value_head_dim=128, linear_num_key_heads=16, linear_num_value_heads=16, layer_types=["linear_attention", "linear_attention", "linear_attention", "full_attention"] * 2, ) config["vision_config"].update( hidden_size=128, intermediate_size=256, depth=4, num_heads=4, out_hidden_size=16, ) config["hidden_size"] = 16 config["num_labels"] = 2 config["architectures"] = ["Qwen3_5ForSequenceClassification"] config["problem_type"] = "single_label_classification" (folder / "config.json").write_text(json.dumps(config, indent=2) + "\n", encoding="utf-8") for name in ( "chat_template.jinja", "generation_config.json", "preprocessor_config.json", "processor_config.json", "tokenizer.json", "tokenizer_config.json", ): (folder / name).write_bytes(Path(hf_hub_download(source_model_id, name, revision=source_revision)).read_bytes()) previous_dtype = torch.get_default_dtype() try: torch.set_default_dtype(torch.bfloat16) set_seed(42) model = AutoModelForSequenceClassification.from_config(AutoConfig.from_pretrained(folder)) with torch.no_grad(): for name, parameter in sorted(model.named_parameters()): if name.endswith("A_log"): parameter.zero_() elif name.endswith("dt_bias"): parameter.fill_(1) elif name.endswith("weight") and "norm" in name: parameter.fill_(1) elif name.endswith("bias"): parameter.zero_() else: parameter.normal_(mean=0.0, std=0.02) state = {name: value.detach().cpu().contiguous() for name, value in model.state_dict().items()} save_file(state, folder / "model.safetensors", metadata={"format": "pt"}) finally: torch.set_default_dtype(previous_dtype) (folder / "config.json").write_text(json.dumps(config, indent=2) + "\n", encoding="utf-8") (folder / "generation_config.json").write_bytes( Path(hf_hub_download(source_model_id, "generation_config.json", revision=source_revision)).read_bytes() ) return str(model) create_model( source_model_id="openbmb/MiniCPM-OCR-Jev", source_revision="bdb03f86182a2824d2632d129c719cdf780502e2", save_folder="/tmp/minicpm-ocr-jev", )