src/model/model_multi_layer_distill_AOP_pooling.py

from typing import Dict
import os
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_multi_layer 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.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']
from contextlib import contextmanager

class _AOPSwitch:
    """
    Temporarily toggle encoder.aop_prune_config.enabled for one forward call.
    """
    def __init__(self, module: nn.Module, enable: bool):
        self.module = module
        self.enable = bool(enable)
        self._old = getattr(module, "aop_prune_config", None)

    def __enter__(self):
        # if no config set, nothing to do
        if self._old is None:
            return self
        if not self.enable:
            # disable only for this scope
            if isinstance(self._old, dict):
                cfg = dict(self._old)
                cfg["enabled"] = False
                setattr(self.module, "aop_prune_config", cfg)
            else:
                setattr(self.module, "aop_prune_config", None)
        # if enable=True, keep as is
        return self

    def __exit__(self, exc_type, exc, tb):
        # restore original
        setattr(self.module, "aop_prune_config", self._old)
        return False

class _VPoolSwitch:
    """
    Temporarily toggle encoder.vision_pooling_config.enabled for one forward call.
    """
    def __init__(self, module: nn.Module, enable: bool):
        self.module = module
        self.enable = bool(enable)
        self._old = getattr(module, "vision_pooling_config", None)
    def __enter__(self):
        if self._old is None: return self
        if not self.enable:
            if isinstance(self._old, dict):
                cfg = dict(self._old); cfg["enabled"] = False
                setattr(self.module, "vision_pooling_config", cfg)
            else:
                setattr(self.module, "vision_pooling_config", None)
        return self
    def __exit__(self, exc_type, exc, tb):
        setattr(self.module, "vision_pooling_config", self._old)
        return False

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]
        # 默认：一个或多个中间层 + 最后一层
        self.supervise_layers = [20, -1]   # -1 必须表示最后一层
        self.supervise_weights = [0.15, 0.85]  # 与 supervise_layers 对齐
    
    @property
    def device(self) -> torch.device:
        try:
            return next(self.parameters()).device
        except StopIteration:
            # 没有参数时兜底 CPU
            return torch.device("cpu")
        
    def _want_prune_for(self, side: str) -> bool:
        """
        side: "qry" or "tgt"
        """
        cfg = getattr(self.encoder, "aop_prune_config", None)
        if not isinstance(cfg, dict) or not cfg.get("enabled", False):
            return False
        apply_to = str(cfg.get("apply_to", "both")).lower()
        return (apply_to == "both") or (apply_to == side.lower())

    def _want_pool_for(self, side: str) -> bool:
        cfg = getattr(self.encoder, "vision_pooling_config", None)
        if not isinstance(cfg, dict) or not cfg.get("enabled", False): return False
        apply_to = str(cfg.get("apply_to", "both")).lower()
        return (apply_to == "both") or (apply_to == side.lower())
        
    def _normalize_layers(self, hs_len: int, layers: list[int]) -> list[int]:
        Lmax = hs_len - 1
        out = []
        for idx in layers:
            if idx < 0:
                idx = hs_len + idx
            idx = max(1, min(idx, Lmax))
            out.append(idx)
        if (hs_len - 1) not in out:
            out.append(hs_len - 1)
        return out

    def _encode_multi(self, input):
        """
        通用多层编码：返回 [B, K, D]，K=len(self.supervise_layers，经规范化且包含最后一层)。
        """
        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 的通用分支（Qwen2-VL 等）
        if mb not in [GME, LamRA, LamRA_QWEN2_5, INTERNVIDEO2, COLPALI]:
            out = self.encoder(
                **input,
                return_dict=True,
                output_hidden_states=True,
            )
            hs_list = out.hidden_states  # tuple/list, len = num_layers + 1（进入每层前的快照 + 最后norm）
            # 剪裁后 attention_mask（若未剪裁则为 None）
            post_mask = getattr(out, "attention_mask", None)    # [B, L_post] or None
            pre_mask  = input['attention_mask']                 # [B, L_pre]

            # 规范化 supervise_layers，并确保包含最后一层
            idxs = self._normalize_layers(len(hs_list), list(dict.fromkeys(self.supervise_layers)))

            # 读取训练时的剪裁层（1-based）。剪裁发生在进入 cut_layer 前，所以 idx >= cut_layer+1 才会看到 post 形状
            aop_cfg = getattr(self.encoder, "aop_prune_config", None)
            cut_layer = None
            if isinstance(aop_cfg, dict) and aop_cfg.get("enabled", False):
                try:
                    cut_layer = int(aop_cfg.get("layer_idx") or 0)
                    if cut_layer <= 0:
                        cut_layer = None
                except Exception:
                    cut_layer = None
            vpool_cfg = getattr(self.encoder, "vision_pooling_config", None)
            pool_layer = None
            if isinstance(vpool_cfg, dict) and vpool_cfg.get("enabled", False):
                try:
                    pool_layer = int(vpool_cfg.get("layer_idx") or 0)
                    if pool_layer <= 0:
                        pool_layer = None
                except Exception:
                    pool_layer = None

            reps = []
            for idx in idxs:
                h = hs_list[idx]  # [B, L_idx, D]
                use_post = False
                if post_mask is not None:
                    if (cut_layer is not None and idx >= cut_layer + 1) or (pool_layer is not None and idx >= pool_layer + 1):
                        use_post = True

                mask_this = post_mask if use_post else pre_mask

                if mask_this is not None and h.size(1) != mask_this.size(1):
                    if pre_mask is not None and pre_mask.size(1) == h.size(1):
                        mask_this = pre_mask
                    elif post_mask is not None and post_mask.size(1) == h.size(1):
                        mask_this = post_mask
                    else:
                        mask_this = torch.ones(h.size(0), h.size(1), dtype=torch.long, device=h.device)

                r = self._pooling(h, mask_this)
                reps.append(F.normalize(r, p=2, dim=-1) if self.normalize else r)

            return torch.stack(reps, dim=1)

    # 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:
            # 默认HF模型：支持 hidden_states（含 AOP 剪裁）
            out = self.encoder(**input, return_dict=True, output_hidden_states=True)
            hs_list = out.hidden_states
            post_mask = getattr(out, "attention_mask", None)    # [B, L_post] or None
            pre_mask  = input['attention_mask']                 # [B, L_pre]

            # === AOP_MONITOR：观测每个样本的剪枝前/后长度与有效保留率 ===
            if os.getenv("AOP_MONITOR", "0") == "1":
                try:
                    B = pre_mask.size(0) if pre_mask is not None else hs_list[-1].size(0)
                    # 全局长度
                    pre_len = pre_mask.sum(dim=1).detach().cpu().tolist() if pre_mask is not None else [hs_list[-1].size(1)] * B
                    post_len = post_mask.sum(dim=1).detach().cpu().tolist() if post_mask is not None else pre_len

                    # 最近一次采样到的 keep_ratio（trainer 写入 cfg）仅用于参考打印
                    aop_cfg = getattr(self.encoder, "aop_prune_config", None)
                    kr_t = aop_cfg.get("_last_sampled_keep_ratio_text") if isinstance(aop_cfg, dict) else None
                    kr_v = aop_cfg.get("_last_sampled_keep_ratio_vision") if isinstance(aop_cfg, dict) else None

                    # 文本/视觉细分（可选）
                    pre_txt_cnt = pre_vis_cnt = post_txt_cnt = post_vis_cnt = None
                    input_ids = input.get("input_ids", None)
                    if input_ids is not None and pre_mask is not None:
                        cfg = self.encoder.config
                        valid_pre = pre_mask.bool()
                        vis_pre = (input_ids == getattr(cfg, "image_token_id", -999))
                        if hasattr(cfg, "video_token_id") and cfg.video_token_id is not None and cfg.video_token_id >= 0:
                            vis_pre = vis_pre | (input_ids == cfg.video_token_id)
                        special_pre = torch.zeros_like(input_ids, dtype=torch.bool)
                        for name in ["bos_token_id", "eos_token_id", "pad_token_id"]:
                            tid = getattr(cfg, name, None)
                            if tid is not None and tid >= 0:
                                special_pre |= (input_ids == tid)
                        pre_vis_cnt = (vis_pre & valid_pre).sum(dim=1).detach().cpu().tolist()
                        pre_txt_cnt = (valid_pre & (~vis_pre) & (~special_pre)).sum(dim=1).detach().cpu().tolist()

                    vis_post_mask = getattr(out, "image_token_bool_masks", None)
                    txt_post_mask = getattr(out, "text_token_bool_masks", None)
                    if vis_post_mask is not None:
                        post_vis_cnt = vis_post_mask.sum(dim=1).detach().cpu().tolist()
                    if txt_post_mask is not None:
                        post_txt_cnt = txt_post_mask.sum(dim=1).detach().cpu().tolist()

                    # 限制打印批次数，避免刷屏
                    if not hasattr(self, "_aop_mon_prints"):
                        self._aop_mon_prints = 0
                    if self._aop_mon_prints < 3:  # 仅前3个batch打印
                        print(f"[AOP][monitor] B={B} sampled: kr_text={kr_t}, kr_vision={kr_v}")
                        for b in range(min(B, 8)):  # 仅打印前8条样本
                            preL = int(pre_len[b]); postL = int(post_len[b]); keep = (postL / (preL + 1e-9))
                            msg = f"  b={b}: pre_len={preL}, post_len={postL}, keep={keep:.3f}"
                            if pre_txt_cnt is not None and post_txt_cnt is not None:
                                kt = (post_txt_cnt[b] / (pre_txt_cnt[b] + 1e-9)) if pre_txt_cnt[b] > 0 else float('nan')
                                msg += f", txt_keep={kt:.3f}"
                            if pre_vis_cnt is not None and post_vis_cnt is not None:
                                kv = (post_vis_cnt[b] / (pre_vis_cnt[b] + 1e-9)) if pre_vis_cnt[b] > 0 else float('nan')
                                msg += f", vis_keep={kv:.3f}"
                            print(msg)
                        self._aop_mon_prints += 1
                except Exception as e:
                    # 避免影响训练流程
                    print(f"[AOP][monitor] warn: monitor failed with error: {e}")

    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
            if model_backbone == QWEN2_5_VL:
                # 默认：优先走本代码的 mldaop 变体，失败回退 vanilla
                try:
                    from src.model.vlm_backbone.qwen2_5_vl_mldaop_pooling.modeling_qwen2_5_vl import (
                        Qwen2_5_VLForConditionalGeneration as Qwen2_5VL_Variant
                    )
                    base_model = Qwen2_5VL_Variant.from_pretrained(
                        model_args.model_name,
                        config=config,
                        torch_dtype=torch.bfloat16,
                        low_cpu_mem_usage=True,
                    )
                    print_master("[Backbone] Using qwen2_5_vl_mldaop_pooling variant (default).")
                except Exception as e:
                    print_master(f"[Backbone] mldaop_pooling import failed, fallback to vanilla. err={e}")
                    base_model = backbone2model[model_backbone].from_pretrained(
                        model_args.model_name,
                        config=config,
                        torch_dtype=torch.bfloat16,
                        low_cpu_mem_usage=True,
                    )
            else:  # QWEN2_VL
                try:
                    from src.model.vlm_backbone.qwen2_vl_mldaop_pooling.modeling_qwen2_vl import (
                        Qwen2VLForConditionalGeneration as Qwen2VL_Variant
                    )
                    base_model = Qwen2VL_Variant.from_pretrained(
                        model_args.model_name,
                        config=config,
                        torch_dtype=torch.bfloat16,
                        low_cpu_mem_usage=True,
                    )
                    print_master("[Backbone] Using qwen2_vl_mldaop_pooling variant (default).")
                except Exception as e:
                    print_master(f"[Backbone] qwen2_vl_mldaop_pooling import failed, fallback to vanilla. err={e}")
                    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
            )
        # 在 build(...) 末尾（return model 前）添加
        def _parse_list(val, tp=float):
            if val is None: return None
            if isinstance(val, (list, tuple)): return [tp(x) for x in val]
            s = str(val).strip()
            if s == "": return None
            return [tp(v.strip()) for v in s.split(",") if v.strip() != ""]

        layers = _parse_list(getattr(model_args, "supervise_layers", None), tp=int)
        weights = _parse_list(getattr(model_args, "supervise_weights", None), tp=float)

        if layers is None:
            # fallback 到旧的二层设置
            layers = [getattr(model_args, 'dual_layer_idx', 20), -1]
        if -1 not in layers:
            layers = list(layers) + [-1]  # 强制包含最后一层

        if weights is None or len(weights) != len(layers):
            # 若未提供或长度不匹配，则做一个合理默认：最后一层占大头
            K = len(layers)
            base = [1.0/(K-1)]*(K-1) if K>1 else [1.0]
            weights = base + [max(0.0, 1.0 - sum(base))]

        # 归一化
        s = sum(max(0.0, w) for w in weights)
        weights = [max(0.0, w)/s for w in weights]

        setattr(model, 'supervise_layers', layers)
        setattr(model, 'supervise_weights', weights)
        # 兼容旧参数
        setattr(model, 'dual_layer_idx', layers[0] if len(layers)>1 else layers[0])
        setattr(model, 'dual_alpha', weights[0] if len(weights)>1 else 1.0)
        setattr(model, 'layer_indices', layers)
        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"
            if hasattr(config, "vision_config") and getattr(config, "vision_config") is not None:
                config.vision_config._attn_implementation = "flash_attention_2"
            if model_args.model_backbone == QWEN2_5_VL:
                try:
                    from src.model.vlm_backbone.qwen2_5_vl_mldaop_pooling.modeling_qwen2_5_vl import (
                        Qwen2_5_VLForConditionalGeneration as Qwen2_5VL_Variant
                    )
                    base_model = Qwen2_5VL_Variant.from_pretrained(
                        model_args.model_name,
                        torch_dtype=torch.bfloat16,
                        low_cpu_mem_usage=True,
                        config=config
                    )
                    print_master("[Backbone:load] Using qwen2_5_vl_mldaop_pooling variant (default).")
                except Exception as e:
                    print_master(f"[Backbone:load] mldaop_pooling import failed, fallback to vanilla. err={e}")
                    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 == QWEN2_VL:
                try:
                    from src.model.vlm_backbone.qwen2_vl_mldaop_pooling.modeling_qwen2_vl import (
                        Qwen2VLForConditionalGeneration as Qwen2VL_Variant
                    )
                    base_model = Qwen2VL_Variant.from_pretrained(
                        model_args.model_name,
                        torch_dtype=torch.bfloat16,
                        low_cpu_mem_usage=True,
                        config=config
                    )
                    print_master("[Backbone:load] Using qwen2_vl_mldaop_pooling variant (default).")
                except Exception as e:
                    print_master(f"[Backbone:load] qwen2_vl_mldaop_pooling import failed, fallback to vanilla. err={e}")
                    base_model = backbone2model[model_args.model_backbone].from_pretrained(
                        model_args.model_name,
                        torch_dtype=torch.bfloat16,
                        low_cpu_mem_usage=True,
                        config=config
                    )
            else:
                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):
        # GradCache：只给一侧 -> 返回多层表示
        if qry is not None and tgt is None:
            with _AOPSwitch(self.encoder, self._want_prune_for("qry")):
                with _VPoolSwitch(self.encoder, self._want_pool_for("qry")):
                    qry_reps = self._encode_multi(qry)
            return {"qry_reps": qry_reps, "tgt_reps": None}
        if tgt is not None and qry is None:
            with _AOPSwitch(self.encoder, self._want_prune_for("tgt")):
                with _VPoolSwitch(self.encoder, self._want_pool_for("tgt")):
                    tgt_reps = self._encode_multi(tgt)
            return {"qry_reps": None, "tgt_reps": tgt_reps}

        with _AOPSwitch(self.encoder, self._want_prune_for("qry")):
            with _VPoolSwitch(self.encoder, self._want_pool_for("qry")):
                q_multi = self._encode_multi(qry)
        with _AOPSwitch(self.encoder, self._want_prune_for("tgt")):
            with _VPoolSwitch(self.encoder, self._want_pool_for("tgt")):
                p_multi = self._encode_multi(tgt)

        # DDP gather
        if self.is_ddp:
            q_multi_all = self._dist_gather_tensor(q_multi)  # [B*, K, D]
            p_multi_all = self._dist_gather_tensor(p_multi)  # [B*, K, D]
        else:
            q_multi_all, p_multi_all = q_multi, p_multi

        Bglob, K, D = q_multi_all.shape
        assert p_multi_all.shape[:2] == (Bglob, K), f"Shape mismatch: q {q_multi_all.shape}, p {p_multi_all.shape}"
        target = torch.arange(Bglob, device=q_multi_all.device, dtype=torch.long)

        w = torch.tensor(self.supervise_weights, dtype=torch.float32, device=q_multi_all.device)
        w = torch.clamp(w, min=0)
        w = w / max(w.sum().item(), 1e-8)

        loss = 0.0
        for k in range(K):
            # 逐层配对（k ↔ k）
            logits_k = torch.matmul(q_multi_all[:, k, :], p_multi_all[:, k, :].transpose(0, 1)) / self.temperature
            loss_k = self.cross_entropy(logits_k, target)
            loss = loss + w[k] * loss_k

        if self.is_ddp:
            loss = loss * self.world_size

        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))