import os import warnings import shutil from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, BitsAndBytesConfig, AutoProcessor import torch from ola.model import * from ola.model.speech_encoder.builder import build_speech_encoder def load_pretrained_model(model_path, model_type, model_base, is_lora=False, s2s=False, load_8bit=False, load_4bit=False, device="cuda", use_flash_attn=False, **kwargs): device = "cuda" if load_8bit: kwargs['load_in_8bit'] = True elif load_4bit: kwargs['load_in_4bit'] = True kwargs['quantization_config'] = BitsAndBytesConfig( load_in_4bit=True, bnb_4bit_compute_dtype=torch.float16, bnb_4bit_use_double_quant=True, bnb_4bit_quant_type='nf4' ) else: kwargs['torch_dtype'] = torch.bfloat16 if use_flash_attn: kwargs['attn_implementation'] = 'flash_attention_2' if model_type == 'ola_internvl': model_cls = OlaQwen3ForCausalLM print('Loading OlaQwen3ForCausalLM model...') else: model_cls = OlaQwenForCausalLM # Load Ola model if is_lora: assert model_base is not None, "model_base is required for LoRA models." from ola.model.language_model.ola_qwen import OlaConfigQwen lora_cfg_pretrained = OlaConfigQwen.from_pretrained(model_path) tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False) print('Loading Ola from base model...') model = model_cls.from_pretrained(model_base, low_cpu_mem_usage=False, config=lora_cfg_pretrained, **kwargs) print('Loading additional Ola weights...') if os.path.exists(os.path.join(model_path, 'non_lora_trainables.bin')): non_lora_trainables = torch.load(os.path.join(model_path, 'non_lora_trainables.bin'), map_location='cpu') non_lora_trainables = {(k[11:] if k.startswith('base_model.') else k): v for k, v in non_lora_trainables.items()} if any(k.startswith('model.model.') for k in non_lora_trainables): non_lora_trainables = {(k[6:] if k.startswith('model.') else k): v for k, v in non_lora_trainables.items()} model.load_state_dict(non_lora_trainables, strict=False, assign=True) from peft import PeftModel print('Loading LoRA weights...') model = PeftModel.from_pretrained(model, model_path) print('Merging LoRA weights...') model = model.merge_and_unload() print('Model is loaded...') elif model_base is not None: print('Loading Ola from base model...') tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False) cfg_pretrained = AutoConfig.from_pretrained(model_path) model = model_cls.from_pretrained(model_base, low_cpu_mem_usage=False, config=cfg_pretrained, **kwargs) speech_projector_weights = torch.load(os.path.join(model_path, 'speech_projector.bin'), map_location='cpu') speech_projector_weights = {k: v.to(torch.float16) for k, v in speech_projector_weights.items()} model.load_state_dict(speech_projector_weights, strict=False, assign=True) model = model.to(device=device) elif model_type == 'ola_internvl': cfg = AutoConfig.from_pretrained("/data1/cxy/plm-v/modeling/old_ola", trust_remote_code=True) # breakpoint() tokenizer = AutoTokenizer.from_pretrained("/data1/cxy/plm-v/modeling/internvl3_5-2B", use_fast=False) with torch.device("cpu"): # model = model_cls.from_pretrained("/data1/cxy/plm-v/modeling/internvl3_5-2B", low_cpu_mem_usage=False, attn_implementation="eager", config=cfg, **kwargs) # model = model_cls.from_config(config=cfg) model = model_cls(cfg) # breakpoint() # model.model.layers[1].self_attn.q_proj.weight else: tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False) with torch.device("cpu"): model = model_cls.from_pretrained( model_path, **kwargs, ) model = model.to(device=device) # model.resize_token_embeddings(len(tokenizer)) from safetensors.torch import load_file partial_state_dict = load_file(f"/data1/cxy/plm-v/modeling/internvl3_5-2B/model.safetensors") # 替换为你的部分权重路径 mapping = { "mlp1.0.weight": "model.mm_projector.layer_norm.weight", "mlp1.0.bias": "model.mm_projector.layer_norm.bias", "mlp1.1.weight": "model.mm_projector.linear_1.weight", "mlp1.1.bias": "model.mm_projector.linear_1.bias", "mlp1.3.weight": "model.mm_projector.linear_2.weight", "mlp1.3.bias": "model.mm_projector.linear_2.bias", } # 遍历 state_dict 并重命名 def remap_keys(state_dict, mapping): new_state_dict = {} for k, v in state_dict.items(): if k in mapping: new_state_dict[mapping[k]] = v else: new_state_dict[k] = v return new_state_dict # merged_state_dict = {**partial_state_dict, **partial_state_dict2} # 2. 重命名 key：multi_modal_projector -> mm_projector # breakpoint() rename_dict = {} for k in list(partial_state_dict.keys()): if k.startswith("language_model"): new_k = k.replace("language_model.", "", 1) rename_dict[k] = new_k if k.startswith("vision_model"): new_k = k.replace("vision_model", "model.vision_tower", 1) rename_dict[k] = new_k # 应用重命名 for old_k, new_k in rename_dict.items(): partial_state_dict[new_k] = partial_state_dict.pop(old_k) partial_state_dict = remap_keys(partial_state_dict, mapping) whisper_state_dict = torch.load("/data1/cxy/model/THUdyh/Ola-7b/large-v3.pt", map_location='cpu') # breakpoint() whisper_state_dict = whisper_state_dict["model_state_dict"] # Filter to keep only encoder weights whisper_encoder_dict = {} for key, value in whisper_state_dict.items(): if key.startswith('encoder.'): whisper_encoder_dict[key] = value print(f"Original Whisper keys: {len(whisper_state_dict)}") print(f"Filtered encoder keys: {len(whisper_encoder_dict)}") print("Sample encoder keys:") for i, key in enumerate(list(whisper_encoder_dict.keys())[:5]): print(f" {key}") # Create mapping for Whisper parameters to OLA format def create_whisper_mapping(): mapping = {} # Base encoder components base_mappings = { 'encoder.positional_embedding': 'model.speech_encoder.whisper_model.positional_embedding', 'encoder.conv1.weight': 'model.speech_encoder.whisper_model.conv1.weight', 'encoder.conv1.bias': 'model.speech_encoder.whisper_model.conv1.bias', 'encoder.conv2.weight': 'model.speech_encoder.whisper_model.conv2.weight', 'encoder.conv2.bias': 'model.speech_encoder.whisper_model.conv2.bias', 'encoder.ln_post.weight': 'model.speech_encoder.whisper_model.ln_post.weight', 'encoder.ln_post.bias': 'model.speech_encoder.whisper_model.ln_post.bias', } mapping.update(base_mappings) # Encoder blocks (32 blocks: 0-31) for block_idx in range(32): # Attention components attn_components = [ 'attn.query.weight', 'attn.query.bias', 'attn.key.weight', 'attn.key.bias', 'attn.value.weight', 'attn.value.bias', 'attn.out.weight', 'attn.out.bias', 'attn_ln.weight', 'attn_ln.bias' ] for component in attn_components: source_key = f'encoder.blocks.{block_idx}.{component}' target_key = f'model.speech_encoder.whisper_model.blocks.{block_idx}.{component}' mapping[source_key] = target_key # MLP components mlp_components = [ 'mlp.0.weight', 'mlp.0.bias', 'mlp.2.weight', 'mlp.2.bias', 'mlp_ln.weight', 'mlp_ln.bias' ] for component in mlp_components: source_key = f'encoder.blocks.{block_idx}.{component}' target_key = f'model.speech_encoder.whisper_model.blocks.{block_idx}.{component}' mapping[source_key] = target_key return mapping # Apply mapping to whisper_encoder_dict whisper_mapping = create_whisper_mapping() mapped_whisper_dict = {} unmapped_whisper_keys = [] for key, value in whisper_encoder_dict.items(): if key in whisper_mapping: mapped_key = whisper_mapping[key] mapped_whisper_dict[mapped_key] = value else: unmapped_whisper_keys.append(key) print(f"Warning: No mapping found for Whisper encoder key '{key}'") if unmapped_whisper_keys: print(f"Total unmapped Whisper encoder keys: {len(unmapped_whisper_keys)}") print("First 10 unmapped Whisper encoder keys:") for key in unmapped_whisper_keys[:10]: print(f" {key}") print(f"Successfully mapped {len(mapped_whisper_dict)} encoder parameters") beat_state_dict = torch.load("/data1/cxy/model/THUdyh/Ola-7b//BEATs_iter3_plus_AS2M_finetuned_on_AS2M_cpt2.pt", map_location='cpu') beat_state_dict = beat_state_dict['model'] beat_state_dict = {"model.speech_encoder.beats_model."+k: v for k, v in beat_state_dict.items()} # 处理 BEATs 模型中的参数化权重映射 (先pop后添加) keys_to_process = list(beat_state_dict.keys()) breakpoint() processed_count = 0 # for key in keys_to_process: # if 'weight_g' in key: # # pop 原始权重并添加为 weight_g # weight_tensor = beat_state_dict.pop(key) # new_key = key.replace('weight_g','parametrizations.weight.original0') # beat_state_dict[new_key] = weight_tensor # processed_count += 1 # elif 'weight_v' in key: # # pop 原始权重并添加为 weight_v # weight_tensor = beat_state_dict.pop(key) # new_key = key.replace('weight_v', 'parametrizations.weight.original1') # beat_state_dict[new_key] = weight_tensor # processed_count += 1 print(f"Processed {processed_count} parametrized weight keys in BEATs model (pop and add)") breakpoint() # breakpoint() partial_state_dict = {**partial_state_dict, **mapped_whisper_dict, **beat_state_dict} # Ensure all tensors in the state dict are on CPU and have proper device information print("Moving all state dict tensors to CPU...") for key, tensor in partial_state_dict.items(): if torch.is_tensor(tensor): # Ensure tensor has device information and move to CPU if not tensor.device.type: print(f"Warning: Tensor {key} has no device, creating on CPU") partial_state_dict[key] = torch.tensor(tensor.detach().numpy()).cpu() else: partial_state_dict[key] = tensor.cpu() # Ensure model is on CPU before loading state dict to avoid device mismatches print("Moving model to CPU before loading state dict...") model = model.cpu() print("Loading state dict...") breakpoint() missing, unexpected = model.load_state_dict(partial_state_dict, strict=False, assign=True) print("Missing keys:", missing) print("Unexpected keys:", unexpected) # Convert model to bfloat16 before saving print("Converting model to bfloat16...") model = model.to(torch.bfloat16) model = model.to("cpu") # Save model in bfloat16 format print("Saving model in bfloat16 format...") model.save_pretrained("/data1/cxy/plm-v/modeling/plm_internvl3_ola", safe_serialization=False, torch_dtype=torch.bfloat16) print("Model saved successfully in bfloat16 format!") breakpoint() # model.model.mm_projector.linear_1.weight:-0.0106 multi_modal_projector.linear_1.weight model.mm_projector.linear_2.bias # model.vision_tower.encoder.layers.7.attn.proj.bias # model.model.vision_tower.encoder.layers[0].attn.qkv.weight: -6.5613e-03 dui # # breakpoint() # model.get_model().speech_encoder.load_model("") # language_model.model.layers.9.mlp.up_proj.weight vision_model.encoder.layers # model.layers.14.self_attn.q_proj.weight model.vision_tower.encoder.layers.23.attn.proj.bias # model.get_model().speech_encoder = build_speech_encoder(model.config) # model.get_model().speech_encoder.to(device=device, dtype=torch.float16) image_processor = None model.resize_token_embeddings(len(tokenizer)) vision_tower = model.get_vision_tower() print("Loading vision tower...") # if not vision_tower.is_loaded: # vision_tower.load_model(device_map=device) # if device != "auto": # vision_tower.to(device="cuda", dtype=torch.bfloat16) # else: # vision_tower.to(device="cuda:0", dtype=torch.bfloat16) # image_processor = vision_tower.image_processor print("Loading vision tower succeeded.") if hasattr(model.config, "max_sequence_length"): context_len = model.config.max_sequence_length else: context_len = 16384 image_processor = AutoProcessor.from_pretrained("/data1/cxy/plm-v/modeling/internvl3_5-2B-HF") # breakpoint() return tokenizer, model, image_processor, context_len