[update]update ver:3.0.5

Author: tusiqi1

README.md

@@ -18,7 +18,7 @@ linger 基于 PyTorch 进行量化算子的搭建，因此只要符合 PyTorch 
 
 ## 快速入门
 - [安装](doc/tutorial/install.md)：支持pip、源码、docker三种安装方式
-- [量化训练快速入门](doc/tutorial/quant_quick_strat.md): 先进行浮点网络的约束训练，再针对量化友好的浮点模型进行量化训练微调
+- [量化训练快速入门](doc/tutorial/quant_quick_start.md): 先进行浮点网络的约束训练，再针对量化友好的浮点模型进行量化训练微调
 - [量化训练进阶指导](doc/tutorial/quant_advanced_guide.md): 量化进阶配置
 - [onnx导出教程](doc/tutorial/export_onnx.md)：将量化无损的PyTorch模型导出为ONNX格式的模型
 

doc/tutorial/install.md

@@ -5,7 +5,7 @@
 
 ### 创建虚拟环境
 ```Shell
-conda create -n linger_thinker_3.x python==3.12.10
+conda create -n linger_thinker_3.x python==3.10.0
 conda activate linger_thinker_3.x
 pip install -U pip
 cat requirements.txt | xargs -n 1 pip install

doc/tutorial/linger_api.md

@@ -1,4 +1,4 @@
-## linger 约束训练接口
+## linger normalize 约束训练接口
 
 
 ## linger 量化训练接口

doc/tutorial/quant_advanced_guide.md

@@ -1,18 +1,16 @@
 ## 校准（PTQ）使用方法
 * 因为校准时会默认按照weight的clip配置进行weight的初始化，故暂不支持循环多组数据校准（仅支持一轮输入校准）
 * 校准时会创建add、bmm等小算子的module
-```python
-    @linger.register_calibrate_method('custom_calibration')
-    def test_init(self, tensor):
-        with torch.no_grad():
-            self.learning_data.fill_(torch.tensor(-999))
-            self.scale.fill_(torch.tensor(-999))
-            self.is_calibrate.fill_(True)
-    
-    
-    with linger.calibration(a_calibrate_name="custom_calibration", w_calibrate_name="custom_calibration"):
-        model = linger.init(model)
-        model(torch.load("/yrfs4/inference/sqtu2/LLM/code/linger3.0/my_linger/calibrate_input.pt"))  
+```python    
+    # 修改cfg.yaml，通过a_calibrate_name和w_calibrate_name设置校准方法，推荐使用默认配置即可；
+    # 量化配置
+    model = linger.init(model, config_file = 'cfg.yaml')
+    # 加载预训练模型
+    model.load_state_dict("./pre_train.pt")
+    with linger.calibration():  # 校准开关
+        model(torch.load("/yrfs4/inference/sqtu2/LLM/code/linger3.0/my_linger/calibrate_input.pt")) # 走一遍前向，开始校准
+
+    # 开始量化训练
 ```
 
 ## linger.init/constrain中'disable_module'使用方法
@@ -30,7 +28,6 @@
 ## linger.init中可通过yaml文件加载配置，当前配置可通过linger.config_save_to_yaml保存
 ## config.yaml 介绍
 * 基础配置
-    calibration: false  # 校准开关
     clamp_info: # 约束信息配置
         clamp_activation_value: 8   # 激活约束浮点值，8代表约束到[-8, 8]
         clamp_bias_value: null      # bias约束浮点值，默认值为None

doc/tutorial/quant_quick_strat.md doc/tutorial/quant_quick_start.mddoc/tutorial/quant_quick_strat.md renamed to doc/tutorial/quant_quick_start.md

@@ -41,7 +41,6 @@ with torch.no_grad():
 
 ## config.yaml 介绍
 * 基础配置
-    calibration: false  # 校准开关
     clamp_info: # 约束信息配置
         clamp_activation_value: 8   # 激活约束浮点值，8代表约束到[-8, 8]
         clamp_bias_value: null      # bias约束浮点值，默认值为None

doc/tutorial/release.md

@@ -10,4 +10,7 @@
 * 解决LSTM和GRU算子一致性
 # V3.0.4  2026.01.07
 * 优化导图代码
-* 解决部分算子一致性问题
+* 解决部分算子一致性问题
+# V3.0.5  2026.02.05
+* 优化导图代码，解决cat算子导图问题
+* 优化代码适配不同torch版本

linger/__version.py

@@ -8,5 +8,5 @@ def _to_int(s):
         return s
 
 
-__version__ = "3.0.4"
+__version__ = "3.0.5"
 version_info = tuple(_to_int(s) for s in __version__.split("."))

linger/config.py

@@ -133,6 +133,7 @@ def _save_to_yaml(cls, config_path: str):
         将当前配置保存到 YAML 文件
         """
         config_dict = cls._to_save_dict()
+        config_dict.pop('calibration', None)
         os.makedirs(os.path.dirname(config_path), exist_ok=True)
         with open(config_path, 'w', encoding='utf-8') as f:
             yaml.dump(config_dict, f, default_flow_style=False, indent=2, allow_unicode=True)

linger/constrain/cmodule.py

@@ -130,7 +130,7 @@ def cweight(self):
 
     @property
     def cbias(self):
-        return self.bias if self.clamp_bias is None else torch.clamp(self.bias, min = -self.clamp_bias, max = self.clamp_bias)
+        return self.bias if (self.clamp_bias is None or self.bias is None) else torch.clamp(self.bias, min = -self.clamp_bias, max = self.clamp_bias)
 
 
     def cforward(self, input: torch.Tensor) -> torch.Tensor:

linger/initialize.py

@@ -14,15 +14,48 @@
 from .constrain.cmodule import constrain_module, _CMODULE_TABLE
 from typing import Any, Dict, List, Optional, Union
 
+def fuse_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
+    
+    eps = 1e-5
+    clamp_conv_name = prefix + 'conv'
+    clamp_bn_name = prefix + 'bn'
+    conv_int_name = prefix
+    if clamp_conv_name + '.weight' in state_dict and clamp_bn_name + '.weight' in state_dict:
+        b_mean = state_dict[clamp_bn_name + '.running_mean']
+        b_var = state_dict[clamp_bn_name + '.running_var']
+        b_w = state_dict[clamp_bn_name + '.weight']
+        b_b = state_dict[clamp_bn_name + '.bias']
+        sigma = 1 / torch.sqrt(b_var + eps)
+        alpha = b_w * sigma
+        beta = b_b - b_mean * alpha
+        c_w = state_dict[clamp_conv_name + '.weight']
+        state_dict[conv_int_name +
+                   'weight'] = (c_w * alpha.view(-1, *([1]*(len(c_w.shape)-1))))
+        if clamp_conv_name + '.bias' in state_dict:
+            c_b = state_dict[clamp_conv_name + '.bias']
+            state_dict[conv_int_name + 'bias'] = (c_b * alpha + beta)
+            state_dict.pop(clamp_conv_name + '.bias')
+        else:
+            state_dict[conv_int_name + 'bias'] = beta
+        state_dict.pop(clamp_bn_name + '.running_mean')
+        state_dict.pop(clamp_bn_name + '.running_var')
+        state_dict.pop(clamp_bn_name + '.weight')
+        state_dict.pop(clamp_bn_name + '.bias')
+        state_dict.pop(clamp_bn_name + '.num_batches_tracked')
+        state_dict.pop(clamp_conv_name + '.weight')
+    else:
+        assert clamp_conv_name + '.weight' not in state_dict and clamp_bn_name + \
+            '.weight' not in state_dict, 'load quanted model but contain float clamp params'
+
 @contextmanager
-def calibration(a_calibrate_name='top_10', w_calibrate_name='abs_max'):
+def calibration():
     # 保存旧值
-    old_a_calibrate_name = QUANT_CONFIGS.quant_info.a_calibrate_name
-    old_w_calibrate_name = QUANT_CONFIGS.quant_info.w_calibrate_name
+    # old_a_calibrate_name = QUANT_CONFIGS.quant_info.a_calibrate_name
+    # old_w_calibrate_name = QUANT_CONFIGS.quant_info.w_calibrate_name
     try:
         QUANT_CONFIGS.calibration = True
-        QUANT_CONFIGS.quant_info.a_calibrate_name = a_calibrate_name
-        QUANT_CONFIGS.quant_info.w_calibrate_name = w_calibrate_name
+        # QUANT_CONFIGS.quant_info.a_calibrate_name = a_calibrate_name
+        # QUANT_CONFIGS.quant_info.w_calibrate_name = w_calibrate_name
         yield  # <<< 关键点：控制权交给 with 块
     finally:
         QUANT_CONFIGS.calibration = False
@@ -84,15 +117,20 @@ def init(model: nn.Module, config_file: str = None, disable_module=None, disable
     # traced_model = symbolic_trace(model)
     # model = _replace_ops(traced_model, q_configs)
 
+    has_replaced = []
     for name, m in model.named_modules():
         if disable_submodel is not None and any(fnmatch(name, pattern) for pattern in disable_submodel):
             continue
+        if any(name.startswith(p + ".") for p in has_replaced):
+            continue
 
         m.register_forward_pre_hook(hook_pre_forward)
         m.register_forward_hook(hook_forward)
 
-        _quantize_submodule(model, name, m, weights_cfg=q_configs.quant_info.to_dict(), activations_cfg=q_configs.quant_info.to_dict(), bias_cfg=q_configs.quant_info.to_dict(), constrain =  q_configs.clamp_info.to_dict())
-    
+        is_replaced = _quantize_submodule(model, name, m, weights_cfg=q_configs.quant_info.to_dict(), activations_cfg=q_configs.quant_info.to_dict(), bias_cfg=q_configs.quant_info.to_dict(), constrain =  q_configs.clamp_info.to_dict())
+        if is_replaced:
+            has_replaced.append(name)
+
     def quant_tensor_pre_hook(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
 
         def quant_tensor_layer(module, prefix=''):
@@ -212,13 +250,21 @@ def _quantize_submodule(
     constrain: Optional[Union[str]] = None,
 ):
     qmodule = quantize_module(module, weights_cfg=weights_cfg, activations_cfg=activations_cfg, bias_cfg = bias_cfg, dim = getattr(module, "dim", None), constrain = constrain)
+    if isinstance(module, ConvBN1d) or isinstance(module, CConvBN1d)  \
+        or isinstance(module, ConvBN2d) or isinstance(module, CConvBN2d) \
+        or isinstance(module, ConvTransposeBN1d) or isinstance(module, CConvTransposeBN1d) \
+        or isinstance(module, ConvTransposeBN2d) or isinstance(module, CConvTransposeBN2d):
+        qmodule._register_load_state_dict_pre_hook(fuse_state_dict)
+
     if qmodule is not None:
         _set_module_by_name(model, name, qmodule)
         qmodule.name = name
         for name, param in module.named_parameters():
             # Save device memory by clearing parameters
             setattr(module, name, None)
             del param
+        return True
+    return False
 
 def _constrain_submodule(
     model: torch.nn.Module,