fix vlm quant bug(qwen25vl merger) (#445)

zhangbilang · web-flow · commit ff3090044c0d · 2025-11-05T20:05:04.000+08:00
diff --git a/configs/quantization/methods/GPTQ/gptq_w_only_vlm.yml b/configs/quantization/methods/GPTQ/gptq_w_only_vlm.yml
@@ -25,22 +25,38 @@ eval:
     bs: 1
     inference_per_block: False
 quant:
-    method: GPTQ
-    quant_objects: [vision, language] # default is [language]
-    weight:
-        bit: 4
-        symmetric: False
-        granularity: per_group
-        group_size: 128
-        # calib_algo: mse
-        # mse_b_num: 2
-    special:
-        actorder: True
-        static_groups: False
-        percdamp: 0.01
-        blocksize: 128
-        true_sequential: True
-    quant_out: True
+    vision:
+        method: GPTQ
+        weight:
+            bit: 4
+            symmetric: False
+            granularity: per_group
+            group_size: 128
+            # calib_algo: mse
+            # mse_b_num: 2
+        special:
+            actorder: True
+            static_groups: False
+            percdamp: 0.01
+            blocksize: 128
+            true_sequential: True
+        quant_out: True
+    language:
+        method: GPTQ
+        weight:
+            bit: 4
+            symmetric: False
+            granularity: per_group
+            group_size: 128
+            # calib_algo: mse
+            # mse_b_num: 2
+        special:
+            actorder: True
+            static_groups: False
+            percdamp: 0.01
+            blocksize: 128
+            true_sequential: True
+        quant_out: True
 save:
     save_fake: False
     save_path: /path/to/save/
diff --git a/configs/quantization/methods/RTN/rtn_w_a_vlm.yml b/configs/quantization/methods/RTN/rtn_w_a_vlm.yml
@@ -13,17 +13,28 @@ eval:
     bs: 1
     inference_per_block: False
 quant:
-    method: RTN
-    quant_objects: [vision, language] # default is [language]
-    weight:
-        bit: 8
-        symmetric: True
-        granularity: per_channel
-        group_size: -1
-    act:
-        bit: 8
-        symmetric: True
-        granularity: per_token
+    vision:
+        method: RTN
+        weight:
+            bit: 8
+            symmetric: True
+            granularity: per_channel
+            group_size: -1
+        act:
+            bit: 8
+            symmetric: True
+            granularity: per_token
+    language:
+        method: RTN
+        weight:
+            bit: 8
+            symmetric: True
+            granularity: per_channel
+            group_size: -1
+        act:
+            bit: 8
+            symmetric: True
+            granularity: per_token
 save:
     save_fake: False
     save_path: /path/to/save/
diff --git a/configs/quantization/methods/SmoothQuant/smoothquant_w_a_vlm.yml b/configs/quantization/methods/SmoothQuant/smoothquant_w_a_vlm.yml
@@ -25,18 +25,30 @@ eval:
     bs: 1
     inference_per_block: False
 quant:
-    method: SmoothQuant
-    quant_objects: [vision, language]
-    weight:
-        bit: 8
-        symmetric: True
-        granularity: per_channel
-    act:
-        bit: 8
-        symmetric: True
-        granularity: per_token
-    special:
-        alpha: 0.8
+    vision:
+        method: SmoothQuant
+        weight:
+            bit: 8
+            symmetric: True
+            granularity: per_channel
+        act:
+            bit: 8
+            symmetric: True
+            granularity: per_token
+        special:
+            alpha: 0.8
+    language:
+        method: SmoothQuant
+        weight:
+            bit: 8
+            symmetric: True
+            granularity: per_channel
+        act:
+            bit: 8
+            symmetric: True
+            granularity: per_token
+        special:
+            alpha: 0.8
 save:
     save_trans: False
     save_fake: False
diff --git a/llmc/compression/quantization/base_blockwise_quantization.py b/llmc/compression/quantization/base_blockwise_quantization.py
@@ -949,20 +949,11 @@ def deploy(self, quant_format, keep_device=False):
             self.set_no_quant_layer()
 
         module = module_mapping[quant_format]
-        if self.modality == 'vision':
-            self.model.replace_vision_module_all(
-                module,
-                self.get_replacement_params(mode=quant_format, w_only=self.w_only),
-                keep_device=keep_device,
-            )
-        if self.modality == 'language':
-            self.model.replace_language_module_all(
-                module,
-                self.get_replacement_params(mode=quant_format, w_only=self.w_only),
-                keep_device=keep_device,
-            )
-        if self.modality == 'video_gen':
-            self.model.replace_video_gen_module_all(
+
+        self.model.set_modality(self.modality)
+        logger.info(f'set modality: {self.modality}')
+        if self.modality in ('vision', 'language', 'video_gen'):
+            self.model.replace_module_all(
                 module,
                 self.get_replacement_params(mode=quant_format, w_only=self.w_only),
                 keep_device=keep_device,
diff --git a/llmc/compression/quantization/llmint8.py b/llmc/compression/quantization/llmint8.py
@@ -66,7 +66,7 @@ def deploy(self, quant_format):
         logger.info(f'-- deploy_{quant_format}_model start --')
         logger.info(f'quant_config : {self.quant_config}')
 
-        self.model.replace_language_module_all(
+        self.model.replace_module_all(
             FakeQuantLinear,
             self.get_replacement_params(
                 mode='fake_quant', w_only=self.w_only, name=None
diff --git a/llmc/compression/quantization/tesseraq.py b/llmc/compression/quantization/tesseraq.py
@@ -3,7 +3,6 @@
 import gc
 import math
 import os
-import pdb
 import random
 from contextlib import nullcontext
 from math import inf
@@ -268,7 +267,6 @@ def tesseraq_train(self, block):
 
                     if not math.isfinite(loss.item()):
                         logger.info('Loss is NAN, stopping training')
-                        pdb.set_trace()
 
                     optimizer.zero_grad()
 
diff --git a/llmc/compression/token_reduction/holitom.py b/llmc/compression/token_reduction/holitom.py
@@ -594,7 +594,6 @@ def prepare_inputs_labels_for_multimodal(
         if isinstance(modalities, str):
             modalities = [modalities]
 
-        # import pdb; pdb.set_trace()
         if type(images) is list or images.ndim == 5:
             mm_patch_merge_type = getattr(self.config, 'mm_patch_merge_type', 'flat')
             image_aspect_ratio = getattr(self.config, 'image_aspect_ratio', 'square')
@@ -733,7 +732,7 @@ def prepare_inputs_labels_for_multimodal(
                     # currently image_feature is a tensor of shape (4, num_patches, hidden_size)
                     # we want to first unflatten it to (2, 2, h, w, hidden_size)
                     # rank0_print("At least we are reaching here")
-                    # import pdb; pdb.set_trace()
+
                     if image_idx in video_idx_in_batch:  # video operations
                         # rank0_print("Video")
                         if mm_newline_position == 'grid':
@@ -1032,7 +1031,6 @@ def prepare_inputs_labels_for_multimodal(
 
             cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds]
 
-            # import pdb; pdb.set_trace()
             cur_new_input_embeds = torch.cat(cur_new_input_embeds)
             cur_new_labels = torch.cat(cur_new_labels)
 
@@ -1157,7 +1155,7 @@ def prepare_inputs_labels_for_multimodal(
             right_add = random.randint(left_add, self.config.pos_skipping_range)
             position_ids[:, :split_position] += left_add
             position_ids[:, split_position:] += right_add
-        # import pdb; pdb.set_trace()
+
         # rank0_print("Finish preparing")
         return (
             None,
diff --git a/llmc/models/base_model.py b/llmc/models/base_model.py
@@ -378,40 +378,7 @@ def get_extra_modules(self, block):
     def get_moe_gate(self, block):
         return None
 
-    def replace_vision_module_all(self, module, params_dict, keep_device=False):
-        vision_model_linears = self.get_block_linears(self.vision_model)
-        for name, m in vision_model_linears.items():
-            M = module.new(m, **params_dict)
-
-            name_tmp = name.rsplit('.', 1)
-            if len(name_tmp) == 2:
-                parent_name = name_tmp[0]
-                parent = self.vision_model.get_submodule(parent_name)
-                child_name = name_tmp[1]
-            elif len(name_tmp) == 1:
-                parent = self.vision_model
-                child_name = name_tmp[0]
-
-            setattr(parent, child_name, M)
-
-        gc.collect()
-        torch.cuda.empty_cache()
-        logger.info(f'The Replaced vision_model: {self.vision_model}')
-
-    def replace_language_module_all(self, module, params_dict, keep_device=False):
-        for block_idx in range(len(self.blocks)):
-            logger.info(f'Replace block index: {block_idx}/{len(self.blocks)}')
-            if keep_device:
-                self.replace_module_block(module, self.blocks[block_idx], block_idx, params_dict)
-            else:
-                self.blocks[block_idx].cuda()
-                self.replace_module_block(module, self.blocks[block_idx], block_idx, params_dict)
-                self.blocks[block_idx].cpu()
-            gc.collect()
-            torch.cuda.empty_cache()
-        logger.info(f'The Replaced model: {self.model}')
-
-    def replace_video_gen_module_all(self, module, params_dict, keep_device=False):
+    def replace_module_all(self, module, params_dict, keep_device=False):
         for block_idx in range(len(self.blocks)):
             logger.info(f'Replace block index: {block_idx}/{len(self.blocks)}')
             if keep_device:
@@ -422,7 +389,6 @@ def replace_video_gen_module_all(self, module, params_dict, keep_device=False):
                 self.blocks[block_idx].cpu()
             gc.collect()
             torch.cuda.empty_cache()
-        logger.info(f'The Replaced model: {self.model}')
 
     def replace_module_block(self, module, block, block_idx, params_dict):
         if module in _LLMC_LN_TYPES_ + _TRANSFORMERS_LN_TYPES_:
diff --git a/llmc/models/internvl3_5.py b/llmc/models/internvl3_5.py
@@ -157,7 +157,7 @@ def __new__(cls, config, device_map=None, use_cache=False):
         if language_part == 'qwen3':
             from .qwen3 import Qwen3
 
-            class NewClass(InternVL2SharedBehavior, Qwen3):
+            class NewClass(InternVL3_5SharedBehavior, Qwen3):
                 def __init__(self, config, device_map=None, use_cache=False):
                     super().__init__(config, device_map, use_cache)
                     setattr(
@@ -170,7 +170,7 @@ def __init__(self, config, device_map=None, use_cache=False):
         return NewClass(config, device_map, use_cache)
 
 
-class InternVL2SharedBehavior():
+class InternVL3_5SharedBehavior():
     def build_model(self):
         self.eval_name = 'InternVL3_5Eval'
         self.vlm_model_config = AutoConfig.from_pretrained(
diff --git a/llmc/models/qwen2.py b/llmc/models/qwen2.py
@@ -22,7 +22,6 @@ def find_embed_layers(self):
 
     def find_block_name(self):
         self.block_name_prefix = 'model.layers'
-        self.pairs = {'q_proj': 'qkv', 'o_proj': 'out', 'up_proj': 'fc1'}
 
     def get_embed_layers(self):
         return [self.embed_tokens]
diff --git a/llmc/models/qwen2_5vl.py b/llmc/models/qwen2_5vl.py
@@ -2,7 +2,6 @@
 from typing import Optional, Union
 
 import torch
-import torch.nn as nn
 from accelerate import Accelerator, DistributedType
 from loguru import logger
 from transformers import AutoConfig, AutoProcessor, AutoTokenizer
@@ -15,18 +14,16 @@
         'If you need it, please upgrade transformers.'
     )
 
-try:
-    from qwen_vl_utils import process_vision_info
-except Exception:
-    logger.warning(
-        'Can not import qwen_vl_utils. '
-        'If you need it, please pip install qwen-vl-utils'
-    )
-
 from llmc.utils.registry_factory import MODEL_REGISTRY
 
 from .qwen2vl import Qwen2VL
 
+#  settings for qwen2_5_vl:
+#  pip install transformers==4.51.3
+#  pip install git+https://github.com/EvolvingLMMs-Lab/lmms-eval.git@v0.3.5
+#  And you should add the following at line 92 of llmc/eval/eval_vqa.py:
+#  import argparse; cli_args = argparse.Namespace(process_with_media=True)
+
 
 @MODEL_REGISTRY
 class Qwen2_5VL(Qwen2VL):
@@ -76,7 +73,6 @@ def build_model(self):
         }
         self.first_turn_question = True
 
-    # todo: check
     def get_subsets_in_block(self, block):
         if self.get_modality() == 'language':
             return super().get_subsets_in_block(block)
@@ -152,14 +148,6 @@ def __init__(
             # Do not use kwargs for now
             assert kwargs == {}, f'Unexpected kwargs: {kwargs}'
 
-            # Validate attention implementation
-            valid_attn_implementations = [None, 'flash_attention_2', 'sdpa', 'eager']
-            if attn_implementation not in valid_attn_implementations:
-                raise ValueError(
-                    f'attn_implementation must be one of {valid_attn_implementations}, \
-                    got {attn_implementation}'
-                )
-
             self.use_custom_video_loader = use_custom_video_loader
             self.fps = fps
             # if self.fps and not self.use_custom_video_loader:
@@ -178,16 +166,6 @@ def __init__(
                 self._device = torch.device(device)
                 self.device_map = device_map if device_map else device
 
-            # Prepare model loading arguments
-            model_kwargs = {
-                'torch_dtype': 'auto',
-                'device_map': self.device_map,
-            }
-
-            # Add attention implementation if specified
-            if attn_implementation is not None:
-                model_kwargs['attn_implementation'] = attn_implementation
-
             self._model = llmc_model.eval().cuda()
             self.max_pixels = max_pixels
             self.min_pixels = min_pixels
diff --git a/llmc/models/qwen2vl.py b/llmc/models/qwen2vl.py
@@ -187,27 +187,6 @@ def get_subsets_in_block(self, block):
         else:
             raise Exception(f'Qwen2VL do not support {self.get_modality()} modality.')
 
-    def get_catcher(self, first_block_input):
-        class Catcher(nn.Module):
-            def __init__(self, module):
-                super().__init__()
-                self.module = module
-                self.mlp = self.module.mlp
-                self.signature = inspect.signature(module.forward)
-
-            def forward(self, *args, **kwargs):
-                params = list(self.signature.parameters.keys())
-                for i, arg in enumerate(args):
-                    if i > 0:
-                        kwargs[params[i]] = arg
-                first_block_input['data'].append(args[0])
-                if 'output_router_logits' in kwargs:
-                    assert kwargs['output_router_logits'] is False
-                    kwargs.pop('output_router_logits')
-                first_block_input['kwargs'].append(kwargs)
-                raise ValueError
-        return Catcher
-
 
 try:
     from lmms_eval.api.model import lmms
diff --git a/tools/quant_analysis.py b/tools/quant_analysis.py
