|
| 1 | +from typing import Iterable, List, Literal, Optional, Union |
| 2 | + |
| 3 | +from compressed_tensors.transform import ( |
| 4 | + TransformArgs, |
| 5 | + TransformConfig, |
| 6 | + TransformScheme, |
| 7 | + apply_transform_config, |
| 8 | +) |
| 9 | +from pydantic import Field, ValidationInfo, field_validator |
| 10 | + |
| 11 | +from llmcompressor.core import Event, EventType, State |
| 12 | +from llmcompressor.modifiers import Modifier |
| 13 | + |
| 14 | +__all__ = ["QuIPModifier"] |
| 15 | + |
| 16 | + |
| 17 | +class QuIPModifier(Modifier): |
| 18 | + """ |
| 19 | + Implements the transforms according to |
| 20 | + [QuIP#: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks](https://arxiv.org/pdf/2402.04396) # noqa: E501 |
| 21 | + [QuIP: 2-Bit Quantization of Large Language Models With Guarantees](https://arxiv.org/abs/2307.13304) # noqa: E501 |
| 22 | +
|
| 23 | + Transforms (rotations) are extra layers added to a model which reduce the accuracy |
| 24 | + loss induced by quantization. This is achived through "rotating" weights and |
| 25 | + activations into a space with a smaller dynamic range of values, thus decreasing |
| 26 | + the range of scales required for quantization. |
| 27 | + |
| 28 | + QuIP and QuIP# apply transforms to every linear layer, two of which are fused into |
| 29 | + the model weights and two of which remain as online rotations computed at runtime. |
| 30 | +
|
| 31 | + :param transform_type: The type of transform to apply to the model. |
| 32 | + `"hadamard"` has the least performance cost but only supports sizes which are |
| 33 | + powers of power of two. |
| 34 | + `"random-matrix"` has more performance cost, but supports a much larger set of |
| 35 | + sizes. |
| 36 | + `"random-matrix"` has the greatest performance cost, but supports any size |
| 37 | + :param randomize: If true, create distinct transforms for each application |
| 38 | + :param learnable: If true, attach gradients to transform weights for training |
| 39 | + :param ignore: Modules to ignore when attaching transforms |
| 40 | + :param transform_config: Optional transform config for overriding provided arguments |
| 41 | + """ |
| 42 | + |
| 43 | + transform_type: Literal["hadamard", "random-hadamard", "random-matrix"] = Field( |
| 44 | + default="hadamard", exclude=True |
| 45 | + ) |
| 46 | + randomize: bool = Field(default=False, exclude=True) |
| 47 | + learnable: bool = Field(default=False, exclude=True) |
| 48 | + ignore: Union[str, List[str]] = Field(default="lm_head", exclude=True) |
| 49 | + |
| 50 | + # optional override for more fine-grained control |
| 51 | + # also included in recipe serialization |
| 52 | + transform_config: Optional[TransformConfig] = Field(default=None, repr=False) |
| 53 | + |
| 54 | + @field_validator("randomize", "learnable", mode="before") |
| 55 | + def validate_not_implemented(cls, value, info: ValidationInfo): |
| 56 | + raise NotImplementedError(f"{info.field_name} is not supported right now") |
| 57 | + |
| 58 | + def on_initialize(self, state: State, **kwargs) -> bool: |
| 59 | + if self.transform_config is not None: |
| 60 | + return True |
| 61 | + |
| 62 | + self.transform_config = self._create_config() |
| 63 | + return True |
| 64 | + |
| 65 | + def on_start(self, state: State, event: Event, **kwargs): |
| 66 | + self.started_ = True |
| 67 | + |
| 68 | + apply_transform_config(state.model, self.transform_config) |
| 69 | + |
| 70 | + def on_event(self, state: State, event: Event, **kwargs): |
| 71 | + if event.type_ == EventType.CALIBRATION_EPOCH_START: |
| 72 | + if not self.started_: |
| 73 | + self.on_start(state, None) |
| 74 | + |
| 75 | + elif event.type_ == EventType.SEQUENTIAL_EPOCH_END: |
| 76 | + pass |
| 77 | + |
| 78 | + elif event.type_ == EventType.CALIBRATION_EPOCH_END: |
| 79 | + if not self.ended_: |
| 80 | + self.on_end(state, None) |
| 81 | + |
| 82 | + def on_end(self, state: State, event: Event, **kwargs): |
| 83 | + self.ended_ = True |
| 84 | + |
| 85 | + def on_finalize(self, state: State, **kwargs) -> bool: |
| 86 | + if not self.ended_: |
| 87 | + self.on_end(state, None) |
| 88 | + |
| 89 | + return True |
| 90 | + |
| 91 | + def _create_config(self) -> TransformConfig: |
| 92 | + return TransformConfig( |
| 93 | + config_groups={ |
| 94 | + "v": TransformScheme( |
| 95 | + type=self.transform_type, |
| 96 | + apply=[ |
| 97 | + TransformArgs( |
| 98 | + targets=["Linear"], |
| 99 | + location="input", # non-mergable |
| 100 | + ignore=self.ignore, |
| 101 | + ), |
| 102 | + TransformArgs( |
| 103 | + targets=["Linear"], |
| 104 | + location="weight_input", |
| 105 | + inverse=True, |
| 106 | + ignore=self.ignore, |
| 107 | + ), |
| 108 | + ], |
| 109 | + randomize=self.randomize, |
| 110 | + requires_grad=self.learnable, |
| 111 | + ), |
| 112 | + "u": TransformScheme( |
| 113 | + type=self.transform_type, |
| 114 | + apply=[ |
| 115 | + TransformArgs( |
| 116 | + targets=["Linear"], |
| 117 | + location="weight_output", |
| 118 | + ignore=self.ignore, |
| 119 | + ), |
| 120 | + TransformArgs( |
| 121 | + targets=["Linear"], |
| 122 | + location="output", # non-mergable |
| 123 | + inverse=True, |
| 124 | + ignore=self.ignore, |
| 125 | + ), |
| 126 | + ], |
| 127 | + randomize=self.randomize, |
| 128 | + requires_grad=self.learnable, |
| 129 | + ), |
| 130 | + } |
| 131 | + ) |
0 commit comments