| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from abc import ABC, abstractmethod |
| 5 | from collections.abc import Sequence |
| 6 | from dataclasses import dataclass |
| 7 | from typing import Generic, TypeVar |
| 8 | |
| 9 | import torch |
| 10 | |
| 11 | from vllm.model_executor.layers.fusion.quant_activation import ( |
| 12 | QuantizedActivation, |
| 13 | as_quantized_activation, |
| 14 | ) |
| 15 | from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8 |
| 16 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 17 | QuantKey, |
| 18 | ) |
| 19 | from vllm.platforms import current_platform |
| 20 | |
| 21 | from ..base import MMLinearLayerConfig |
| 22 | |
| 23 | |
| 24 | @dataclass |
| 25 | class Int8ScaledMMLinearLayerConfig(MMLinearLayerConfig): |
| 26 | # TODO: Change to QuantKey like FP8ScaledMMLinearLayerConfig |
| 27 | is_static_input_scheme: bool |
| 28 | is_channelwise: bool |
| 29 | input_symmetric: bool |
| 30 | |
| 31 | |
| 32 | @dataclass |
| 33 | class FP8ScaledMMLinearLayerConfig(MMLinearLayerConfig): |
| 34 | weight_quant_key: QuantKey |
| 35 | activation_quant_key: QuantKey |
| 36 | weight_shape: tuple[int, int] |
| 37 | input_dtype: torch.dtype |
| 38 | out_dtype: torch.dtype |
| 39 | |
| 40 | |
| 41 | _FP8ParamsT = tuple[ |
| 42 | torch.Tensor, # weight |
| 43 | torch.Tensor, # weight_scale |
| 44 | torch.Tensor | None, # input_scale, |
| 45 | torch.Tensor | None, # input_scale_ub, |
| 46 | ] |
| 47 | _Int8ParamsT = tuple[ |
| 48 | torch.Tensor, # weight |
| 49 | torch.Tensor, # weight_scale |
| 50 | torch.Tensor | None, # input_scale, |
| 51 | torch.Tensor | None, # input_zp |
| 52 | torch.Tensor | None, # azp_adj |
| 53 | ] |
| 54 | |
| 55 | _ParamsT = TypeVar("_ParamsT", _Int8ParamsT, _FP8ParamsT) |
| 56 | _ConfigT = TypeVar("_ConfigT", bound=MMLinearLayerConfig) |
| 57 | |
| 58 | |
| 59 | class ScaledMMLinearKernel(Generic[_ConfigT, _ParamsT], ABC): |
| 60 | @classmethod |
| 61 | @abstractmethod |
| 62 | def is_supported( |
| 63 | cls, compute_capability: int | None = None |
| 64 | ) -> tuple[bool, str | None]: |
| 65 | raise NotImplementedError |
| 66 | |
| 67 | @classmethod |
| 68 | @abstractmethod |
| 69 | def can_implement(cls, c: _ConfigT) -> tuple[bool, str | None]: |
| 70 | raise NotImplementedError |
| 71 | |
| 72 | def __init__(self, c: _ConfigT, layer_param_names: Sequence[str]) -> None: |
| 73 | assert self.can_implement(c)[0] |
| 74 | assert self.is_supported()[0] |
| 75 | self.config = c |
| 76 | self.layer_param_names = layer_param_names |
| 77 | |
| 78 | def input_quant_key(self) -> QuantKey | None: |
| 79 | """The activation quant key this kernel can consume pre-quantized. |
| 80 | |
| 81 | Manual fusion uses this to decide whether to hoist activation |
| 82 | quantization out of apply_weights into an upstream fused kernel. |
| 83 | Return None when the kernel needs in-kernel quantization (custom |
| 84 | padding or swizzling, dynamic scales, etc.). Kernels that return a |
| 85 | key must consume the activation via as_quantized_activation. |
| 86 | """ |
| 87 | return None |
| 88 | |
| 89 | @abstractmethod |
| 90 | def process_weights_after_loading(self, layer: torch.nn.Module) -> None: |
| 91 | raise NotImplementedError |
| 92 | |
| 93 | @abstractmethod |
| 94 | def apply_weights( |
| 95 | self, |
| 96 | layer: torch.nn.Module, |
| 97 | x: torch.Tensor, |
| 98 | bias: torch.Tensor | None = None, |
| 99 | ) -> torch.Tensor: |
| 100 | raise NotImplementedError |
| 101 | |
| 102 | # return a covariant type in the subclass |
| 103 | @abstractmethod |
| 104 | def _get_layer_params(self, layer) -> _ParamsT: |
| 105 | raise NotImplementedError |
| 106 | |
| 107 | |
| 108 | class FP8ScaledMMLinearKernel( |
| 109 | ScaledMMLinearKernel[FP8ScaledMMLinearLayerConfig, _FP8ParamsT], ABC |
| 110 | ): |
| 111 | def __init__( |
| 112 | self, c: FP8ScaledMMLinearLayerConfig, layer_param_names: Sequence[str] |
| 113 | ) -> None: |
| 114 | act_scale_descriptor = c.activation_quant_key.scale |
| 115 | self.quant_fp8 = QuantFP8( |
| 116 | static=act_scale_descriptor.static, |
| 117 | group_shape=act_scale_descriptor.group_shape, |
| 118 | num_token_padding=self.get_output_padding(), |
| 119 | ) |
| 120 | self.fp8_dtype = current_platform.fp8_dtype() |
| 121 | super().__init__(c, layer_param_names) |
| 122 | |
| 123 | def process_weights_after_loading(self, layer: torch.nn.Module) -> None: |
| 124 | pass |
| 125 | |
| 126 | def _get_layer_params(self, layer) -> _FP8ParamsT: |
| 127 | w, w_s, x_s, x_s_ub = self.layer_param_names |
| 128 | return ( |
| 129 | getattr(layer, w), |
| 130 | getattr(layer, w_s), |
| 131 | getattr(layer, x_s, None), |
| 132 | getattr(layer, x_s_ub, None), |
| 133 | ) |
| 134 | |
| 135 | def apply_weights( |
| 136 | self, |
| 137 | layer: torch.nn.Module, |
| 138 | x: torch.Tensor | QuantizedActivation, |
| 139 | bias: torch.Tensor | None = None, |
| 140 | ) -> torch.Tensor: |
| 141 | fp8_dtype = self.fp8_dtype |
| 142 | maybe_out_dtype = self.config.out_dtype |
| 143 | w, w_s, x_s, x_s_ub = self._get_layer_params(layer) |
| 144 | |
| 145 | qa = as_quantized_activation(x, self.input_quant_key()) |
| 146 | if qa is not None: |
| 147 | x_data, x_s = qa.data, qa.scale |
| 148 | orig_shape, orig_dtype = qa.orig_shape, qa.orig_dtype |
| 149 | assert x_data.dtype == fp8_dtype |
| 150 | else: |
| 151 | assert isinstance(x, torch.Tensor) |
| 152 | x_data = x |
| 153 | orig_shape, orig_dtype = x.shape, x.dtype |
| 154 | |
| 155 | x_2d = x_data.view(-1, x_data.shape[-1]) |
| 156 | output_shape = [*orig_shape[:-1], w.shape[1]] |
| 157 | out_dtype = orig_dtype if maybe_out_dtype is None else maybe_out_dtype |
| 158 | |
| 159 | x_2d_q = x_2d |
| 160 | if qa is None: |
| 161 | x_2d_q, x_s = self.quant_fp8(x_2d, x_s, x_s_ub) |
| 162 | return self.apply_scaled_mm( |
| 163 | A=x_2d_q, |
| 164 | B=w, |
| 165 | out_dtype=out_dtype, |
| 166 | As=x_s, |
| 167 | Bs=w_s, |
| 168 | bias=bias, |
| 169 | output_shape=output_shape, |
| 170 | ) |
| 171 | |
| 172 | @abstractmethod |
| 173 | def apply_scaled_mm( |
| 174 | self, |
| 175 | *, |
| 176 | A: torch.Tensor, |
| 177 | B: torch.Tensor, |
| 178 | out_dtype: torch.dtype, |
| 179 | As: torch.Tensor, |
| 180 | Bs: torch.Tensor, |
| 181 | bias: torch.Tensor | None, |
| 182 | output_shape: list, |
| 183 | ) -> torch.Tensor: |
| 184 | raise NotImplementedError |
| 185 | |
| 186 | def get_output_padding(self) -> int | None: |
| 187 | return None |
| 188 | |
| 189 | |
| 190 | class Int8ScaledMMLinearKernel( |
| 191 | ScaledMMLinearKernel[Int8ScaledMMLinearLayerConfig, _Int8ParamsT], ABC |
| 192 | ): |
| 193 | def _get_layer_params(self, layer) -> _Int8ParamsT: |
| 194 | w_q, w_s, i_s, i_zp, azp_adj = self.layer_param_names |
| 195 | return ( |
| 196 | getattr(layer, w_q), |
| 197 | getattr(layer, w_s), |
| 198 | getattr(layer, i_s, None), |
| 199 | getattr(layer, i_zp, None), |
| 200 | getattr(layer, azp_adj, None), |
| 201 | ) |