from typing import Dict import torch import torch.distributed as dist from torch import nn, Tensor import torch.nn.functional as F # 如果文件顶部没引入的话 from transformers import PreTrainedModel, AutoModelForCausalLM, AutoConfig from peft import LoraConfig, get_peft_model, PeftModel from src.model.processor import QWEN2_5_VL_TOKENSELECTION from src.arguments import ModelArguments, TrainingArguments from src.model.processor import LLAVA_NEXT, QWEN2_VL, PHI3V, get_backbone_name, print_master, QWEN2_5_VL, \ backbone2model, QWEN2_VL_TOKENSELECTION, QWEN2_5_VL_TOKENSELECTION, E5_V from src.arguments import ModelArguments from src.model.processor import LLAVA_NEXT, QWEN2_VL, PHI3V, get_backbone_name, print_master, QWEN2_5_VL, INTERNVIDEO2, \ QWEN2_VL_TOKENSELECTION, backbone2model, GME, VLM_IMAGE_TOKENS, LamRA, LamRA_QWEN2_5, COLPALI from src.model.baseline_backbone.colpali import ColPali from src.model.baseline_backbone.gme.gme_inference import GmeQwen2VL from src.model.baseline_backbone.lamra.lamra_inference import LamRAQwen2VL from src.model.baseline_backbone.lamra.lamra_qwen25_inference import LamRAQwen25VL from src.model.baseline_backbone.phi3_v.modeling_phi3_v import Phi3VForCausalLM from src.model.baseline_backbone.llava_next import LlavaNextForConditionalGeneration from transformers import modeling_utils if not hasattr(modeling_utils, "ALL_PARALLEL_STYLES") or modeling_utils.ALL_PARALLEL_STYLES is None: modeling_utils.ALL_PARALLEL_STYLES = ["tp", "none", "colwise", 'rowwise'] class MMEBModel(nn.Module): TRANSFORMER_CLS = AutoModelForCausalLM def __init__(self, encoder: PreTrainedModel, pooling: str = 'last', normalize: bool = False, temperature: float = 0.02, ): super().__init__() self.config = encoder.config self.encoder = encoder self.pooling = pooling self.normalize = normalize self.temperature = temperature self.cross_entropy = nn.CrossEntropyLoss(reduction='mean') self.is_ddp = dist.is_initialized() if self.is_ddp: self.process_rank = dist.get_rank() self.world_size = dist.get_world_size() self.layer_indices = [20, -1] self.dual_layer_idx = 20 # query 的第20层 self.dual_alpha = 0.15 # 两个 CE 的加权系数 def _encode_query_dual(self, input): """ 返回 [B, 2, D]: 第20层与最后一层的池化向量。 对不支持 hidden_states 的 backbone，回退为两份相同的最后一层。 """ mb = getattr(self, "model_backbone", None) def norm(x): return F.normalize(x, p=2, dim=-1) if self.normalize else x # 支持 hidden_states 的分支（LLAVA_NEXT + 默认HF） if mb not in [GME, LamRA, LamRA_QWEN2_5, INTERNVIDEO2, COLPALI]: out = self.encoder(**input, return_dict=True, output_hidden_states=True) hs = out.hidden_states # [emb, layer1, ..., layerL] idx20 = self.dual_layer_idx # 安全检查与边界保护 if idx20 < 0: idx20 = len(hs) + idx20 # 允许负索引 idx20 = max(1, min(idx20, len(hs) - 1)) # 1..L rep20 = self._pooling(hs[idx20], input['attention_mask']) replast = self._pooling(hs[-1], input['attention_mask']) rep20, replast = norm(rep20), norm(replast) reps = torch.stack([rep20, replast], dim=1) # [B, 2, D] return reps # LLAVA_NEXT 特殊：保持和原来一致地取 hidden_states if mb == LLAVA_NEXT: input = dict(input) # 避免原地修改 input['pixel_values'] = input['pixel_values'].squeeze(dim=1) input['image_sizes'] = input['image_sizes'].squeeze(dim=1) out = self.encoder(**input, return_dict=True, output_hidden_states=True) hs = out.hidden_states idx20 = self.dual_layer_idx if idx20 < 0: idx20 = len(hs) + idx20 idx20 = max(1, min(idx20, len(hs) - 1)) rep20 = self._pooling(hs[idx20], input['attention_mask']) replast = self._pooling(hs[-1], input['attention_mask']) rep20, replast = norm(rep20), norm(replast) reps = torch.stack([rep20, replast], dim=1) return reps # 其他不支持中间层的backbone：回退（两份最后一层） last = self.encode_input(input) # [B, D]，已有归一化 reps = torch.stack([last, last], dim=1) return reps # def encode_input(self, input): def encode_input(self, input, layer_indices=None): if getattr(self, "model_backbone", None) == INTERNVIDEO2: if "input_ids" in input.keys(): # text side text_output = self.encoder.get_text_encoder()( input["input_ids"], attention_mask=input["attention_mask"], return_dict=True, mode="text", ) text_embeds = text_output.last_hidden_state pooled_text_embeds = text_embeds[:, 0] pooled_output = self.encoder.text_proj(pooled_text_embeds) pooled_output /= pooled_output.norm(dim=-1, keepdim=True) return pooled_output else: _, vfeat = self.encoder.encode_vision(input["pixel_values"], test=True) vfeat = self.encoder.vision_proj(vfeat) vfeat /= vfeat.norm(dim=-1, keepdim=True) return vfeat elif getattr(self, "model_backbone", None) in [GME, LamRA, LamRA_QWEN2_5]: # pooled_output = self.encoder(**input, return_dict=True, output_hidden_states=True) texts = [text.replace(VLM_IMAGE_TOKENS[QWEN2_VL] + '\n', '') for text in input["texts"]] # we are actually passing video queries so this should not happen images = [] for imgs in input['images']: # if multi images are given, select the middle frame only if isinstance(imgs, list): imgs = imgs[len(imgs) // 2] assert not isinstance(imgs, list) # make sure we have extracted the middle frame and it is no longer a list images.append(imgs) else: images.append(imgs) pooled_output = self.encoder.get_fused_embeddings(texts=texts, images=images) return pooled_output elif getattr(self, "model_backbone", None) == COLPALI: pooled_output = self.encoder(**input, return_dict=True, output_hidden_states=True) return pooled_output elif getattr(self, "model_backbone", None) == LLAVA_NEXT: input['pixel_values'] = input['pixel_values'].squeeze(dim=1) input['image_sizes'] = input['image_sizes'].squeeze(dim=1) hidden_states = self.encoder(**input, return_dict=True, output_hidden_states=True) hidden_states = hidden_states.hidden_states[-1] pooled_output = self._pooling(hidden_states, input['attention_mask']) return pooled_output else: # hidden_states = self.encoder(**input, return_dict=True, output_hidden_states=True) # # hidden_states = self.encoder(**input, compression_rate=compression_rate, return_dict=True, output_hidden_states=True) # hidden_states = hidden_states.hidden_states[-1] # pooled_output = self._pooling(hidden_states, input['attention_mask']) # return pooled_output # 默认HF模型：支持 hidden_states out = self.encoder(**input, return_dict=True, output_hidden_states=True) hs_list = out.hidden_states if layer_indices is None or isinstance(layer_indices, int): h = hs_list[-1] if layer_indices is None else hs_list[layer_indices] reps = self._pooling(h, input['attention_mask']) return reps else: reps_list = [] for idx in layer_indices: h = hs_list[idx] r = self._pooling(h, input['attention_mask']) reps_list.append(r) reps = torch.stack(reps_list, dim=1) # [B, L, D] return reps def _pooling(self, last_hidden_state, attention_mask): if self.pooling == 'last' or self.pooling == 'eos': left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0]) batch_size = last_hidden_state.shape[0] if left_padding: # Get the vectors at the last position reps = last_hidden_state[torch.arange(batch_size), -1, :] else: # Calculate last 1 position in the original tensor eos_indices = attention_mask.sum(dim=1) - 1 # Get the vectors at the last 1 position of each attention mask reps = last_hidden_state[ torch.arange(batch_size, device=last_hidden_state.device), eos_indices] else: raise NotImplementedError if self.normalize: reps = torch.nn.functional.normalize(reps, p=2, dim=-1) return reps @classmethod def build(cls, model_args: ModelArguments, **kwargs): config = AutoConfig.from_pretrained(model_args.model_name, trust_remote_code=True) variant = getattr(config, "backbone_variant", None) if variant == "layerprune": model_backbone = "QWEN2_VL_LayerPrune" else: model_backbone = get_backbone_name(hf_config=config) print_master(f'Loading backbone [{model_backbone}] from {model_args.model_name}') # Loading the base model if model_backbone == PHI3V: config._attn_implementation = "eager" config.padding_side = "right" config.use_cache = False base_model = Phi3VForCausalLM.from_pretrained( model_args.model_name, config=config, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, ) elif model_backbone == LLAVA_NEXT: config.use_cache = False config.padding_side = "left" base_model = LlavaNextForConditionalGeneration.from_pretrained( model_args.model_name, config=config, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, ) elif model_backbone in [QWEN2_VL, QWEN2_5_VL]: config._attn_implementation = "flash_attention_2" config.padding_side = "left" config.use_cache = False base_model = backbone2model[model_backbone].from_pretrained( model_args.model_name, config=config, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, ) elif model_backbone in ["QWEN2_VL_LayerPrune"]: config._attn_implementation = "flash_attention_2" config.padding_side = "left" config.use_cache = False base_model = backbone2model[model_backbone].from_pretrained( model_args.model_name, config=config, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, ) elif model_backbone in [QWEN2_VL_TOKENSELECTION, QWEN2_5_VL_TOKENSELECTION]: config._attn_implementation = "flash_attention_2" config.padding_side = "left" config.use_cache = False from .utils import parse_layer_type lm_qwen_layer = 28 vis_qwen_layer = 32 lm_skip_layer = parse_layer_type(model_args.lm_skip_layer, lm_qwen_layer) vis_skip_layer = parse_layer_type(model_args.vis_skip_layer, vis_qwen_layer) base_model = backbone2model[model_backbone].from_pretrained( model_args.model_name, config=config, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, lm_skip_layer=lm_skip_layer, vis_skip_layer=vis_skip_layer, ) else: config.use_cache = False base_model = cls.TRANSFORMER_CLS.from_pretrained( model_args.model_name, **kwargs, config=config, attn_implementation="flash_attention_2", torch_dtype=torch.bfloat16, trust_remote_code=True) if model_args.lora: print_master(f'Loading lora adapter from {base_model}') lora_config = LoraConfig( r=model_args.lora_r, lora_alpha=model_args.lora_alpha, target_modules=model_args.lora_target_modules.split(','), lora_dropout=model_args.lora_dropout, init_lora_weights="gaussian", use_dora=True, inference_mode=False ) lora_model = get_peft_model(base_model, lora_config) model = cls( encoder=lora_model, pooling=model_args.pooling, normalize=model_args.normalize, temperature=model_args.temperature ) else: model = cls( encoder=base_model, pooling=model_args.pooling, normalize=model_args.normalize, temperature=model_args.temperature ) # 注入属性（供 _encode_query_dual 使用） setattr(model, 'model_backbone', model_backbone) setattr(model, 'dual_layer_idx', getattr(model_args, 'dual_layer_idx', 20)) setattr(model, 'dual_alpha', getattr(model_args, 'dual_alpha', 0.15)) setattr(model, 'layer_indices', [model.dual_layer_idx, -1]) return model @classmethod def load(cls, model_args: ModelArguments, is_trainable=True, **kwargs): # Loading the base model model_name_or_path = model_args.checkpoint_path if model_args.checkpoint_path else model_args.model_name config = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=True) if not hasattr(model_args, "model_backbone") or not model_args.model_backbone: model_backbone = get_backbone_name(hf_config=config, model_type=model_args.model_type) setattr(model_args, 'model_backbone', model_backbone) print_master(f'Loading backbone [{model_args.model_backbone}] from {model_name_or_path}') if model_args.model_backbone in {LLAVA_NEXT, QWEN2_VL, QWEN2_5_VL, QWEN2_VL_TOKENSELECTION, QWEN2_5_VL_TOKENSELECTION, E5_V}: config = AutoConfig.from_pretrained(model_args.model_name, trust_remote_code=True) config._attn_implementation = "flash_attention_2" config.vision_config._attn_implementation = "flash_attention_2" base_model = backbone2model[model_args.model_backbone].from_pretrained( model_args.model_name, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, config=config ) elif model_args.model_backbone == PHI3V: config = AutoConfig.from_pretrained(model_args.model_name, trust_remote_code=True) config.use_cache = False config.padding_side = "right" base_model = Phi3VForCausalLM.from_pretrained(model_args.model_name, **kwargs, config=config, torch_dtype=torch.bfloat16, trust_remote_code=True) base_model.padding_side = "right" elif model_args.model_backbone == INTERNVIDEO2: print_master(f'Loading backbone [{model_args.model_backbone}] from {"src/model/vlm_backbone/internvideo2/"}') config = AutoConfig.from_pretrained("src/model/vlm_backbone/internvideo2/", trust_remote_code=True) base_model = backbone2model[model_args.model_backbone].from_pretrained("src/model/vlm_backbone/internvideo2/", config=config, trust_remote_code=True) elif model_args.model_backbone == GME: base_model = GmeQwen2VL(model_args.model_name, processor=kwargs['processor']) setattr(base_model, 'config', config) elif model_args.model_backbone == LamRA: base_model = LamRAQwen2VL(model_args.model_name) setattr(base_model, 'config', config) elif model_args.model_backbone == LamRA_QWEN2_5: base_model = LamRAQwen25VL(model_args.model_name) setattr(base_model, 'config', config) elif model_args.model_backbone == COLPALI: base_model = ColPali.from_pretrained(model_args.model_name) setattr(base_model, 'config', config) else: # Loading external base model from HF config = AutoConfig.from_pretrained(model_args.model_name, trust_remote_code=True) config.use_cache = False base_model = cls.TRANSFORMER_CLS.from_pretrained( model_name_or_path, **kwargs, config=config, torch_dtype=torch.bfloat16, trust_remote_code=True) # Building the model on top of the base if model_args.lora: print_master(f'Loading LoRA from {model_name_or_path}') lora_config = LoraConfig.from_pretrained(model_name_or_path) lora_model = PeftModel.from_pretrained(base_model, model_name_or_path, config=lora_config, is_trainable=is_trainable) lora_model.load_adapter(model_name_or_path, lora_model.active_adapter, is_trainable=is_trainable) if not is_trainable: lora_model = lora_model.merge_and_unload() model = cls( encoder=lora_model, pooling=model_args.pooling, normalize=model_args.normalize, temperature=model_args.temperature ) else: model = cls( encoder=base_model, pooling=model_args.pooling, normalize=model_args.normalize, temperature=model_args.temperature ) model.model_backbone = model_args.model_backbone return model def save(self, output_dir: str): self.encoder.save_pretrained(output_dir) # def forward(self, qry: Dict[str, Tensor] = None, tgt: Dict[str, Tensor] = None, *args, **kwargs): # qry_reps, tgt_reps = None, None # if qry is not None: # qry_reps = self.encode_input(qry, layer_indices = self.layer_indices) # 仅qry侧用裁层 # if tgt is not None: # tgt_reps = self.encode_input(tgt) # cand侧保持全层（或按你设置） # # print('qry_reps:', qry_reps) # # print('tgt:', tgt) # # print('self.layer_indices:', self.layer_indices) # # exit() # # 只编码一侧时，按你之前的返回约定原样返回 # if qry_reps is None or tgt_reps is None: # return {"qry_reps": qry_reps, "tgt_reps": tgt_reps} # if self.is_ddp: # all_qry_reps = self._dist_gather_tensor(qry_reps) # all_tgt_reps = self._dist_gather_tensor(tgt_reps) # else: # all_qry_reps = qry_reps # all_tgt_reps = tgt_reps # scores = self.compute_similarity(all_qry_reps, all_tgt_reps) # scores = scores.view(all_qry_reps.size(0), -1) # target = torch.arange(scores.size(0), device=scores.device, dtype=torch.long) # target = target * (all_qry_reps.size(0) // all_tgt_reps.size(0)) # loss = self.cross_entropy(scores / self.temperature, target) # if self.is_ddp: # loss = loss * self.world_size # return loss def forward(self, qry: Dict[str, Tensor] = None, tgt: Dict[str, Tensor] = None, *args, **kwargs): # GradCache：只给一侧，返回表示 if qry is not None and tgt is None: qry_reps = self._encode_query_dual(qry) # [B, 2, D] return {"qry_reps": qry_reps, "tgt_reps": None} if tgt is not None and qry is None: tgt_reps = self.encode_input(tgt) # [B, D] return {"qry_reps": None, "tgt_reps": tgt_reps} # 非 GradCache：两侧同时给，直接算双 CE qry_dual = self._encode_query_dual(qry) # [B, 2, D] tgt_last = self.encode_input(tgt) # [B, D] # DDP 全局收集 if self.is_ddp: q20_all = self._dist_gather_tensor(qry_dual[:, 0, :]) qlast_all= self._dist_gather_tensor(qry_dual[:, 1, :]) p_all = self._dist_gather_tensor(tgt_last) else: q20_all, qlast_all, p_all = qry_dual[:, 0, :], qry_dual[:, 1, :], tgt_last # 计算两个 logits 并分别做 CE scores20 = torch.matmul(q20_all, p_all.transpose(0, 1)) scoreslast = torch.matmul(qlast_all, p_all.transpose(0, 1)) scores20 = scores20 / self.temperature scoreslast = scoreslast / self.temperature B = scores20.size(0) target = torch.arange(B, device=scores20.device, dtype=torch.long) # 如遇 Nq != Nt，可参考原逻辑修正（通常相等） # target = target * (q20_all.size(0) // p_all.size(0)) loss20 = self.cross_entropy(scores20, target) print('loss20:', loss20) losslast = self.cross_entropy(scoreslast, target) print('losslast:', losslast) alpha = getattr(self, "dual_alpha", 0.15) loss = alpha * loss20 + (1.0 - alpha) * losslast if self.is_ddp: loss = loss * self.world_size # 与 Trainer 的 /world_size 抵消 return loss def _dist_gather_tensor(self, t: Tensor): t = t.contiguous() all_tensors = [torch.empty_like(t) for _ in range(self.world_size)] dist.all_gather(all_tensors, t) all_tensors[self.process_rank] = t all_tensors = torch.cat(all_tensors, dim=0) return all_tensors def compute_similarity(self, q_reps, p_reps): return torch.matmul(q_reps, p_reps.transpose(0, 1))