|
| 1 | +""" |
| 2 | +IMatrixGatherer: collects importance-weighted activation statistics |
| 3 | +(E[x²] per input channel) on Linear modules via forward pre-hooks. |
| 4 | +
|
| 5 | +Stores ``module._imatrix_importance`` as a 1-D float32 tensor of |
| 6 | +shape ``(in_features,)`` on each target module. Does **not** quantize |
| 7 | +or modify weights in any way. |
| 8 | +
|
| 9 | +The downstream ``imatrix_mse`` observer reads this attribute during |
| 10 | +its grid search to weight quantization error by channel importance. |
| 11 | +
|
| 12 | +Example recipe:: |
| 13 | +
|
| 14 | + recipe: |
| 15 | + - IMatrixGatherer: |
| 16 | + ignore: ["lm_head"] |
| 17 | + - QuantizationModifier: |
| 18 | + config_groups: |
| 19 | + group_0: |
| 20 | + targets: ["Linear"] |
| 21 | + weights: |
| 22 | + observer: imatrix_mse |
| 23 | +
|
| 24 | +Or composed with AWQ and GPTQ:: |
| 25 | +
|
| 26 | + recipe: |
| 27 | + - AWQModifier(...) |
| 28 | + - IMatrixGatherer: |
| 29 | + ignore: ["lm_head"] |
| 30 | + - GPTQModifier: |
| 31 | + config_groups: |
| 32 | + group_0: |
| 33 | + targets: ["Linear"] |
| 34 | + weights: |
| 35 | + observer: imatrix_mse |
| 36 | +
|
| 37 | +.. note:: |
| 38 | + Auto-prepend (inserting the gatherer automatically when |
| 39 | + ``imatrix_mse`` is detected in a recipe) is planned for a |
| 40 | + follow-up PR. |
| 41 | +
|
| 42 | +.. note:: |
| 43 | + Unlike AWQModifier, this gatherer does not use IntermediatesCache |
| 44 | + because it only stores a single accumulated 1-D tensor per layer |
| 45 | + (not full batch activations). A simple CPU offload at |
| 46 | + CALIBRATION_EPOCH_END is sufficient. |
| 47 | +""" |
| 48 | + |
| 49 | +from typing import Dict, List, Optional, Union |
| 50 | + |
| 51 | +import torch |
| 52 | +from compressed_tensors.utils import match_named_modules |
| 53 | +from loguru import logger |
| 54 | +from pydantic import Field |
| 55 | +from torch.nn import Module |
| 56 | + |
| 57 | +from llmcompressor.core import Event, EventType, State |
| 58 | +from llmcompressor.modifiers import Modifier |
| 59 | + |
| 60 | +__all__ = ["IMatrixGatherer"] |
| 61 | + |
| 62 | + |
| 63 | +class IMatrixGatherer(Modifier): |
| 64 | + """ |
| 65 | + Collects importance-weighted activation statistics (E[x²]) |
| 66 | + for each targeted module via forward pre-hooks. |
| 67 | +
|
| 68 | + Stores ``module._imatrix_importance`` as a 1-D float32 tensor |
| 69 | + of shape ``(in_features,)`` on each target module. |
| 70 | +
|
| 71 | + Does NOT quantize. Does NOT modify weights. |
| 72 | +
|
| 73 | + Statistics are kept on GPU during calibration for speed, then |
| 74 | + offloaded to CPU at CALIBRATION_EPOCH_END to free GPU memory |
| 75 | + before quantization begins. |
| 76 | +
|
| 77 | + :param ignore: layer name patterns to skip (default: ``["lm_head"]``) |
| 78 | + :param targets: module types to instrument (default: ``["Linear"]``) |
| 79 | + """ |
| 80 | + |
| 81 | + ignore: Union[str, List[str]] = Field( |
| 82 | + default_factory=lambda: ["lm_head"], |
| 83 | + ) |
| 84 | + targets: Union[str, List[str]] = Field( |
| 85 | + default_factory=lambda: ["Linear"], |
| 86 | + ) |
| 87 | + |
| 88 | + # -- internal state (excluded from serialisation) -- |
| 89 | + _target_names: Optional[List[str]] = None |
| 90 | + _sums: Optional[Dict[str, torch.Tensor]] = None |
| 91 | + _counts: Optional[Dict[str, int]] = None |
| 92 | + |
| 93 | + # ------------------------------------------------------------------ # |
| 94 | + # Lifecycle |
| 95 | + # ------------------------------------------------------------------ # |
| 96 | + |
| 97 | + def on_initialize(self, state: State, **kwargs) -> bool: |
| 98 | + if self.end and self.end != -1: |
| 99 | + raise ValueError( |
| 100 | + f"{self.__class__.__name__} can only be applied " |
| 101 | + f"during one-shot. Expected end to be None or " |
| 102 | + f"-1, got {self.end}" |
| 103 | + ) |
| 104 | + if self.start and self.start != -1: |
| 105 | + raise ValueError( |
| 106 | + f"{self.__class__.__name__} can only be applied " |
| 107 | + f"during one-shot. Expected start to be None " |
| 108 | + f"or -1, got {self.start}" |
| 109 | + ) |
| 110 | + |
| 111 | + self._resolve_targets(state.model) |
| 112 | + return True |
| 113 | + |
| 114 | + def on_start(self, state: State, event: Event, **kwargs): |
| 115 | + self.started_ = True |
| 116 | + self._register_accumulation_hooks(state.model) |
| 117 | + |
| 118 | + def on_event(self, state: State, event: Event, **kwargs): |
| 119 | + if event.type_ == EventType.CALIBRATION_EPOCH_START: |
| 120 | + if not self.started_: |
| 121 | + self.on_start(state, None) |
| 122 | + |
| 123 | + if event.type_ == EventType.SEQUENTIAL_EPOCH_END: |
| 124 | + self._compute_and_attach(state.model) |
| 125 | + |
| 126 | + if event.type_ == EventType.CALIBRATION_EPOCH_END: |
| 127 | + self._compute_and_attach(state.model, offload_to_cpu=True) |
| 128 | + |
| 129 | + if not self.ended_: |
| 130 | + self.on_end(state, None) |
| 131 | + |
| 132 | + def on_end(self, state: State, event: Event, **kwargs): |
| 133 | + self.ended_ = True |
| 134 | + self.remove_hooks() |
| 135 | + |
| 136 | + def on_finalize(self, state: State, **kwargs) -> bool: |
| 137 | + if not self.ended_: |
| 138 | + self.on_end(state, None) |
| 139 | + |
| 140 | + self._sums = None |
| 141 | + self._counts = None |
| 142 | + self._target_names = None |
| 143 | + return True |
| 144 | + |
| 145 | + # ------------------------------------------------------------------ # |
| 146 | + # Target resolution |
| 147 | + # ------------------------------------------------------------------ # |
| 148 | + |
| 149 | + def _resolve_targets(self, model: Module): |
| 150 | + """Identify target modules using compressed_tensors matching.""" |
| 151 | + self._target_names = [] |
| 152 | + self._sums = {} |
| 153 | + self._counts = {} |
| 154 | + |
| 155 | + for name, module in match_named_modules(model, self.targets, self.ignore): |
| 156 | + if not hasattr(module, "in_features"): |
| 157 | + continue |
| 158 | + |
| 159 | + self._target_names.append(name) |
| 160 | + self._sums[name] = torch.zeros(module.in_features, dtype=torch.float32) |
| 161 | + self._counts[name] = 0 |
| 162 | + |
| 163 | + logger.info(f"IMatrixGatherer: targeting {len(self._target_names)}" f" modules") |
| 164 | + |
| 165 | + # ------------------------------------------------------------------ # |
| 166 | + # Hook registration |
| 167 | + # ------------------------------------------------------------------ # |
| 168 | + |
| 169 | + def _register_accumulation_hooks(self, model: Module): |
| 170 | + """Attach a forward-pre hook to every target module.""" |
| 171 | + |
| 172 | + def _create_hook_fn(layer_name: str): |
| 173 | + """Closure captures layer_name.""" |
| 174 | + |
| 175 | + def _hook(module: Module, args): |
| 176 | + x = args[0] if not isinstance(args, torch.Tensor) else args |
| 177 | + if isinstance(x, tuple): |
| 178 | + x = x[0] |
| 179 | + if not isinstance(x, torch.Tensor): |
| 180 | + return |
| 181 | + |
| 182 | + # Per-token accumulation |
| 183 | + x_f = x.detach().float() |
| 184 | + n_tokens = x_f[..., 0].numel() |
| 185 | + token_sum = x_f.pow(2).sum(dim=list(range(x_f.dim() - 1))) |
| 186 | + |
| 187 | + device = self._sums[layer_name].device |
| 188 | + if device != token_sum.device: |
| 189 | + self._sums[layer_name] = self._sums[layer_name].to(token_sum.device) |
| 190 | + |
| 191 | + self._sums[layer_name].add_(token_sum) |
| 192 | + self._counts[layer_name] += n_tokens |
| 193 | + |
| 194 | + return _hook |
| 195 | + |
| 196 | + for name, module in match_named_modules(model, self.targets, self.ignore): |
| 197 | + if name in self._sums: |
| 198 | + self.register_hook( |
| 199 | + module, |
| 200 | + _create_hook_fn(name), |
| 201 | + "forward_pre", |
| 202 | + ) |
| 203 | + |
| 204 | + # ------------------------------------------------------------------ # |
| 205 | + # Compute & attach |
| 206 | + # ------------------------------------------------------------------ # |
| 207 | + |
| 208 | + def _compute_and_attach(self, model: Module, offload_to_cpu: bool = False): |
| 209 | + """ |
| 210 | + Compute E[x²] and store on each module. |
| 211 | +
|
| 212 | + :param model: model whose modules receive importance data |
| 213 | + :param offload_to_cpu: if True, move importance tensors to CPU |
| 214 | + after attaching. Set at CALIBRATION_EPOCH_END to free |
| 215 | + GPU memory before quantization. |
| 216 | + """ |
| 217 | + attached = 0 |
| 218 | + for name, module in match_named_modules(model, self.targets, self.ignore): |
| 219 | + if name not in self._sums: |
| 220 | + continue |
| 221 | + |
| 222 | + count = self._counts[name] |
| 223 | + if count == 0: |
| 224 | + continue |
| 225 | + |
| 226 | + importance = self._sums[name] / count |
| 227 | + |
| 228 | + if offload_to_cpu: |
| 229 | + importance = importance.to("cpu") |
| 230 | + # also free the accumulator |
| 231 | + del self._sums[name] |
| 232 | + |
| 233 | + module._imatrix_importance = importance |
| 234 | + |
| 235 | + attached += 1 |
| 236 | + logger.debug( |
| 237 | + f"iMatrix {name}: " |
| 238 | + f"mean={importance.mean():.4f}, " |
| 239 | + f"max={importance.max():.4f}, " |
| 240 | + f"ratio=" |
| 241 | + f"{importance.max() / (importance.mean() + 1e-10):.1f}" |
| 242 | + ) |
| 243 | + |
| 244 | + logger.info( |
| 245 | + f"IMatrixGatherer: attached importance to " |
| 246 | + f"{attached} modules" + (" (offloaded to CPU)" if offload_to_cpu else "") |
| 247 | + ) |
0 commit comments