fix vispruner bugs and update holitom_merge (#431)

zhangbilang · web-flow · commit fb8f3c13e494 · 2025-08-01T13:07:41.000+08:00
* fix vispruner bugs and update holitom_merge

* set lmms_eval==0.3.0 temporarily
diff --git a/llmc/compression/token_reduction/dycoke.py b/llmc/compression/token_reduction/dycoke.py
@@ -7,8 +7,8 @@
 
 try:
     from llava.model.llava_arch import LlavaMetaForCausalLM
-except ModuleNotFoundError:
-    logger.info('LlavaMetaForCausalLM not found, if need, please install llava first.')
+except ImportError:
+    pass
 from transformers.cache_utils import Cache, DynamicCache
 
 from llmc.utils.registry_factory import TOKEN_REDUCTION_REGISTRY
diff --git a/llmc/compression/token_reduction/fastvid.py b/llmc/compression/token_reduction/fastvid.py
@@ -13,8 +13,8 @@
     from llava.model.multimodal_encoder.siglip_encoder import (
         SigLipVisionConfig, SigLipVisionModel)
     from llava.utils import rank0_print
-except ModuleNotFoundError:
-    logger.info('LlavaMetaForCausalLM not found, if need, please install llava first.')
+except ImportError:
+    pass
 
 from llmc.utils.registry_factory import TOKEN_REDUCTION_REGISTRY
 
diff --git a/llmc/compression/token_reduction/holitom.py b/llmc/compression/token_reduction/holitom.py
@@ -20,8 +20,8 @@
     from llava.utils import rank0_print
     from transformers.modeling_outputs import (BaseModelOutput,
                                                BaseModelOutputWithPooling)
-except ModuleNotFoundError:
-    logger.info('LlavaMetaForCausalLM not found, if need, please install llava first.')
+except ImportError:
+    pass
 
 from llmc.utils.registry_factory import TOKEN_REDUCTION_REGISTRY
 
diff --git a/llmc/compression/token_reduction/prunevid.py b/llmc/compression/token_reduction/prunevid.py
@@ -8,8 +8,8 @@
 
 try:
     from llava.model.llava_arch import LlavaMetaForCausalLM
-except ModuleNotFoundError:
-    logger.info('LlavaMetaForCausalLM not found, if need, please install llava first.')
+except ImportError:
+    pass
 
 from llmc.utils.registry_factory import TOKEN_REDUCTION_REGISTRY
 
diff --git a/llmc/compression/token_reduction/random.py b/llmc/compression/token_reduction/random.py
@@ -1,9 +1,7 @@
 import functools
-from functools import wraps
 from types import MethodType
 
 import torch
-from loguru import logger
 
 from llmc.utils.registry_factory import TOKEN_REDUCTION_REGISTRY
 
@@ -21,61 +19,21 @@ def __init__(self, config, model, blocks):
     def add_sparse_config(self):
 
         self.pruning_loc = self.special_config['pruning_loc']
-        self.special_config['image_token_length'] = self.model.pruning_config[
-            'image_token_length'
-        ]
-
         self.pruning_paras = self.special_config
 
     def register_reduction_modules(self):
 
-        def input_hook_llava(fn, pruning_paras):
-            @wraps(fn)
-            def wrapper(self, *args, **kwargs):
-                if len(args) == 0:
-                    return fn(*args, **kwargs)
-                input_args = args[0]
-                if hasattr(input_args[0], 'shape') and input_args[0].shape[0] == 1:
-                    return fn(*args, **kwargs)
-
-                input_ids = args[0]
-                attention_mask = args[2]
-                token_indices = input_ids[0][attention_mask[0]] == IMAGE_TOKEN_INDEX
-                pruning_paras['image_token_start_index'] = torch.where(token_indices)[
-                    0
-                ][0].item()
-
-                outputs = fn(*args, **kwargs)
-                return outputs
-
-            return wrapper
-
-        @prefill_wrapper
-        def input_hook(module, input_args, pruning_paras):
-            input_ids = input_args[0]
-            image_token_idxs = (
-                input_ids[0] == pruning_paras['vision_token_index']
-            ).nonzero(as_tuple=True)[0]
-            pruning_paras['image_token_start_index'] = image_token_idxs[0].item()
-
-            return input_args
-
         @prefill_wrapper
         def random_pruning_hook(module, args, kwargs, pruning_paras):
 
-            logger.info(' ========random_pruning_hook======== ')
-
-            rate = pruning_paras['rate']
-            image_token_start_index = pruning_paras['image_token_start_index']
-            image_token_length = pruning_paras['image_token_length']
+            rate = pruning_paras['prune_ratio']
+            image_token_start_index = pruning_paras['vision_token_start_index']
+            image_token_length = pruning_paras['vision_token_length']
 
             hidden_states = args[0]
             causal_mask = kwargs['attention_mask']
 
-            logger.info(f'before hidden_states : {hidden_states.shape}')
-
             device = hidden_states.device
-
             vision_indexes = torch.arange(
                 image_token_start_index,
                 image_token_start_index + image_token_length,
@@ -130,25 +88,169 @@ def random_pruning_hook(module, args, kwargs, pruning_paras):
             position_embeddings[0].resize_as_(new_pe0).copy_(new_pe0)
             position_embeddings[1].resize_as_(new_pe0).copy_(new_pe1)
 
-            logger.info(f'after hidden_states : {hidden_states.shape}')
             return (hidden_states,), kwargs
 
-        if self.model.__class__.__name__ == 'LlavaHf':
-            self.model.embed_tokens.register_forward_pre_hook(
-                functools.partial(input_hook, pruning_paras=self.pruning_paras)
+        @prefill_wrapper
+        def holitom_merge_hook(module, args, kwargs, pruning_paras):
+
+            rate = pruning_paras['prune_ratio']
+            image_token_start_index = pruning_paras['vision_token_start_index']
+            image_token_length = pruning_paras['vision_token_length']
+
+            hidden_states = args[0]
+            causal_mask = kwargs['attention_mask']
+
+            device = hidden_states.device
+            last_layer_attention = pruning_paras['attn_scores']
+            # compute average attention over different head
+            last_layer_attention_avg = torch.mean(
+                last_layer_attention, dim=1
+            )[0]
+            # generate new attention mask based on the average attention,
+            # sample the top ATTENTION_RANK tokens with highest attention
+            last_layer_attention_avg_last_tok = (
+                last_layer_attention_avg[-1]
+            )
+            # get the attention in image token
+            last_layer_attention_avg_last_tok_image = \
+                last_layer_attention_avg_last_tok[
+                    image_token_start_index:
+                    image_token_start_index + image_token_length
+                ]
+            # get the indexes of the top ATTENTION_RANK tokens
+            top_attention_rank_index = (
+                last_layer_attention_avg_last_tok_image.topk(
+                    round(
+                        image_token_length * (1 - rate)
+                    )
+                ).indices
+                + image_token_start_index
             )
-        elif self.model.__class__.__name__ == 'Llava':
-            from llava.constants import IMAGE_TOKEN_INDEX
 
-            hook_fn = input_hook_llava(
-                self.model.vlm_model.prepare_inputs_labels_for_multimodal,
-                self.pruning_paras,
+            all_indices = torch.arange(
+                image_token_length, device=device
             )
-            self.model.vlm_model.prepare_inputs_labels_for_multimodal = MethodType(
-                hook_fn, self.model.vlm_model
+            non_topk_mask = ~torch.isin(
+                all_indices,
+                top_attention_rank_index
+                - image_token_start_index,
+            )
+            non_topk_indices = (
+                all_indices[non_topk_mask]
+                + image_token_start_index
+            )
+            non_topk_states = hidden_states[
+                :, non_topk_indices, :
+            ]  # [batch_size, len(non_topk), hidden_size]
+            topk_states = hidden_states[
+                :, top_attention_rank_index, :
+            ]  # [batch_size, len(topk), hidden_size]
+            non_topk_norm = torch.norm(
+                non_topk_states, dim=-1, keepdim=True
+            )  # [batch_size, len(non_topk), 1]
+            topk_norm = torch.norm(
+                topk_states, dim=-1, keepdim=True
+            )  # [batch_size, len(topk), 1]
+            dot_product = torch.bmm(
+                non_topk_states, topk_states.transpose(1, 2)
+            )  # [batch_size, len(non_topk), len(topk)]
+            sim_matrix = dot_product / (
+                non_topk_norm * topk_norm.transpose(1, 2)
+            )
+            sim_max, sim_max_index = torch.max(sim_matrix, dim=-1)
+
+            batch_size = hidden_states.size(0)
+            num_topk = len(top_attention_rank_index)
+            num_non_topk = len(non_topk_indices)
+            topk_counter = torch.ones((batch_size, num_topk, 1), device=hidden_states.device)
+
+            for b in range(batch_size):
+                for i in range(num_non_topk):
+                    topk_rel_idx = sim_max_index[b, i].item()  # 这是 topk 中的相对索引
+                    topk_abs_idx = top_attention_rank_index[topk_rel_idx]  # 得到绝对索引
+                    non_topk_abs_idx = non_topk_indices[i]
+
+                    # 累加non-topk到topk token上（就地）
+                    hidden_states[b, topk_abs_idx, :] += hidden_states[b, non_topk_abs_idx, :]
+                    # 增加计数
+                    topk_counter[b, topk_rel_idx] += 1
+
+            # 平均化所有topk token（包含自己和所有被合并的）
+            for b in range(batch_size):
+                for i in range(num_topk):
+                    topk_abs_idx = top_attention_rank_index[i]
+                    hidden_states[b, topk_abs_idx, :] /= topk_counter[b, i]
+
+            keep_indexs = torch.cat(
+                (
+                    torch.arange(
+                        image_token_start_index,
+                        device=device,
+                    ),
+                    top_attention_rank_index,
+                    torch.arange(
+                        image_token_start_index
+                        + image_token_length,
+                        hidden_states.shape[1],
+                        device=device,
+                    ),
+                )
             )
 
-        self.blocks[self.pruning_loc].register_forward_pre_hook(
-            functools.partial(random_pruning_hook, pruning_paras=self.pruning_paras),
-            with_kwargs=True,
-        )
+            # sort index
+            keep_indexs = keep_indexs.sort().values
+            # filter hidden states &
+            hidden_states = hidden_states[:, keep_indexs, :]
+            # update position ids
+            position_ids = keep_indexs.unsqueeze(0)
+            # update attention mask
+            if causal_mask is not None:
+                causal_mask = causal_mask[:, :, :hidden_states.shape[1], :hidden_states.shape[1]]
+                kwargs['attention_mask'].resize_as_(causal_mask).copy_(causal_mask.clone())
+            kwargs['cache_position'].resize_as_(position_ids.squeeze(0)).copy_(
+                position_ids.squeeze(0).clone())
+            kwargs['position_ids'].resize_as_(position_ids).copy_(position_ids.clone())
+
+            position_embeddings = kwargs['position_embeddings']
+            index_dim = 1 if position_embeddings[0].dim() == 3 else 2
+            new_pe0 = position_embeddings[0].index_select(index_dim, keep_indexs).clone()
+            new_pe1 = position_embeddings[1].index_select(index_dim, keep_indexs).clone()
+            position_embeddings[0].resize_as_(new_pe0).copy_(new_pe0)
+            position_embeddings[1].resize_as_(new_pe0).copy_(new_pe1)
+
+            return (hidden_states,), kwargs
+
+        def update_output_attentions_hook(module, args, kwargs):
+            kwargs['output_attentions'] = True
+            return args, kwargs
+
+        def store_attention_hook(m, x, layer_outputs, pruning_paras):
+            layer_attention = layer_outputs[1]
+            pruning_paras['attn_scores'] = layer_attention
+
+        if self.special_config['vision_token_length'] is None:
+            if self.model.__class__.__name__ == 'Llava':
+                self.model.vlm_model.prepare_inputs_labels_for_multimodal = MethodType(
+                    self.vtoken_length_for_llava_hook(
+                        self.model.vlm_model.prepare_inputs_labels_for_multimodal,
+                        self.pruning_paras
+                    ), self.model.vlm_model
+                )
+
+        if self.special_config['metric'] == 'random':
+            self.blocks[self.pruning_loc].register_forward_pre_hook(
+                functools.partial(random_pruning_hook, pruning_paras=self.pruning_paras),
+                with_kwargs=True
+            )
+        elif self.special_config['metric'] == 'holitom_merge':
+            self.blocks[self.pruning_loc - 1].register_forward_pre_hook(
+                update_output_attentions_hook,
+                with_kwargs=True
+            )
+            self.blocks[self.pruning_loc - 1].register_forward_hook(
+                functools.partial(store_attention_hook, pruning_paras=self.pruning_paras),
+            )
+            self.blocks[self.pruning_loc].register_forward_pre_hook(
+                functools.partial(holitom_merge_hook, pruning_paras=self.pruning_paras),
+                with_kwargs=True
+            )
diff --git a/llmc/compression/token_reduction/tome.py b/llmc/compression/token_reduction/tome.py
@@ -3,8 +3,6 @@
 from typing import Callable, Optional, Tuple
 
 import torch
-import torch.nn.functional as F
-from loguru import logger
 from transformers.models.clip.modeling_clip import CLIPEncoderLayer
 
 from llmc.utils.registry_factory import TOKEN_REDUCTION_REGISTRY
@@ -20,8 +18,7 @@ def __init__(self, config, model, blocks):
         self.patch_layer()
 
     def add_sparse_config(self):
-        special_config = self.config.get('special', {})
-        r_param = special_config.get('r', 0)
+        r_param = self.special_config.get('r', 0)
         if isinstance(r_param, int) or isinstance(r_param, float):
             self.r = [max(int(r_param), 0)] * len(self.blocks)
         elif isinstance(r_param, (tuple, list)):
@@ -36,19 +33,17 @@ def add_sparse_config(self):
         else:
             raise ValueError('Invalid r format. Expected int or (start, step) tuple.')
 
-        self.pruning_paras = special_config
+        self.pruning_paras = self.special_config
 
     def patch_layer(self):
         for idx, block in enumerate(self.blocks):
             if self.r[idx] > 0:
                 block.r = self.r[idx]
                 if isinstance(block, CLIPEncoderLayer):  # llava
-                    block.self_attn.original_forward = block.self_attn.forward
                     block.self_attn.forward = types.MethodType(
                         tome_CLIPSdpaAttention_forward,
                         block.self_attn
                     )
-                    block.original_forward = block.forward
                     block.forward = types.MethodType(
                         tome_CLIPEncoderLayer_forward,
                         block
diff --git a/llmc/compression/token_reduction/utils.py b/llmc/compression/token_reduction/utils.py
@@ -9,8 +9,8 @@
 
 try:
     from llava.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX
-except Exception as e:
-    logger.debug('LLaVA is not installed. Please install LLaVA to use this model.\nError: %s' % e)
+except ImportError:
+    pass
 import random
 
 
diff --git a/llmc/compression/token_reduction/visionzip.py b/llmc/compression/token_reduction/visionzip.py
@@ -585,19 +585,8 @@ def prune_qwenv25vl_hook(module, args, kwargs, pruning_paras):
             st_idx = torch.nonzero(img_mask, as_tuple=True)[0]
 
             if st_idx.numel() > 0:
-                discontinuities = torch.where(st_idx[1:] - st_idx[:-1] != 1)[0]
-                if discontinuities.numel() > 0:
-                    raise ValueError('Visual tokens are not contiguous in input_ids!')
-                    segment_starts = [st_idx[0].item()] + [st_idx[i + 1].item() for i in discontinuities.tolist()]  # noqa
-                    segment_ends = [st_idx[i].item() for i in discontinuities.tolist()] + [st_idx[-1].item()]  # noqa
-                    offset = 0
-                    for first, last in zip(segment_starts, segment_ends):
-                        length = last - first + 1
-                        # [15 1502] [1505 3289]
-                        img_mask[first: last + 1] = ~select_mask[offset: offset + length]
-                else:
-                    first, last = st_idx[0].item(), st_idx[-1].item()
-                    img_mask[first: last + 1] = ~select_mask
+                first, last = st_idx[0].item(), st_idx[-1].item()
+                img_mask[first: last + 1] = ~select_mask
                 img_mask = ~img_mask
                 contextual_input_idx = false_pos[target_indices] + first
 
diff --git a/llmc/compression/token_reduction/vispruner.py b/llmc/compression/token_reduction/vispruner.py
diff --git a/requirements/runtime.txt b/requirements/runtime.txt
