Kernels:
Trusted publisher
File size: 5,039 Bytes
5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 e0d830e 5855430 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 | # Copyright (c) 2023-2026, Songlin Yang, Yu Zhang, Zhiyuan Li
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# For a list of all contributors, visit:
# https://github.com/fla-org/flash-linear-attention/graphs/contributors
# BitLinear with fused normalization and quantized linear projection.
# [The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits](https://arxiv.org/abs/2402.17764)
# [Scalable MatMul-free Language Modeling](https://arxiv.org/abs/2406.02528)
#
# adapted from https://github.com/ridgerchu/matmulfreellm/
from __future__ import annotations
import torch
import torch.nn as nn
import torch.nn.functional as F
from ..modules.norm import RMSNorm
from ..modules.norm.layernorm_quant import LayerNormLinearQuantFunction, weight_quant
def activation_quant(x: torch.Tensor) -> torch.Tensor:
"""Quantize activations to 8 bits per token and return dequantized values."""
scale = 127.0 / x.abs().max(dim=-1, keepdim=True).values.clamp_(min=1e-5)
y = (x * scale).round().clamp_(min=-128, max=127) / scale
return y
def layer_norm_linear_quant(
x: torch.Tensor,
norm_weight: torch.Tensor | None,
norm_bias: torch.Tensor | None,
linear_weight: torch.Tensor,
linear_bias: torch.Tensor | None,
residual: torch.Tensor | None = None,
eps: float = 1e-6,
prenorm: bool = False,
residual_in_fp32: bool = False,
is_rms_norm: bool = False,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
return LayerNormLinearQuantFunction.apply(
x,
norm_weight,
norm_bias,
linear_weight,
linear_bias,
residual,
eps,
prenorm,
residual_in_fp32,
is_rms_norm,
)
def rms_norm_linear_quant(
x: torch.Tensor,
norm_weight: torch.Tensor | None,
norm_bias: torch.Tensor | None,
linear_weight: torch.Tensor,
linear_bias: torch.Tensor | None,
residual: torch.Tensor | None = None,
eps: float = 1e-5,
prenorm: bool = False,
residual_in_fp32: bool = False,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
return layer_norm_linear_quant(
x=x,
norm_weight=norm_weight,
norm_bias=norm_bias,
linear_weight=linear_weight,
linear_bias=linear_bias,
residual=residual,
eps=eps,
prenorm=prenorm,
residual_in_fp32=residual_in_fp32,
is_rms_norm=True,
)
def bit_linear(
x: torch.Tensor,
weight: torch.Tensor,
bias: torch.Tensor | None = None,
norm_weight: torch.Tensor | None = None,
norm_bias: torch.Tensor | None = None,
eps: float = 1e-8,
) -> torch.Tensor:
"""Apply RMS normalization and a linear projection with quantized activations and weights."""
return layer_norm_linear_quant(
x=x,
norm_weight=norm_weight,
norm_bias=norm_bias,
linear_weight=weight,
linear_bias=bias,
eps=eps,
is_rms_norm=True,
)
class BitLinear(nn.Linear):
"""
RMS-normalized linear layer with 8-bit activations and ternary weights.
Quantization uses a straight-through estimator during training.
Efficient deployment requires specialized kernels.
Args:
in_features (int):
Size of each input sample.
out_features (int):
Size of each output sample.
bias (bool, Optional):
Whether to allocate an additive bias. Default: `False`.
norm_eps (float, Optional):
Epsilon for RMS normalization. Default: 1e-8.
"""
def __init__(self, in_features: int, out_features: int, bias: bool = False, norm_eps: float = 1e-8) -> None:
super().__init__(in_features=in_features, out_features=out_features, bias=bias)
self.norm = RMSNorm(hidden_size=in_features, eps=norm_eps, dtype=torch.float32)
def __repr__(self) -> str:
return f"{self.__class__.__name__}({super().extra_repr()}, norm_eps={self.norm.eps})"
def forward(self, x: torch.Tensor) -> torch.Tensor:
w = self.weight
x_norm = self.norm(x=x)
# straight-through gradients bypass quantization
x_quant = x_norm + (activation_quant(x=x_norm) - x_norm).detach()
w_quant = w + (weight_quant(w=w) - w).detach()
y = F.linear(input=x_quant, weight=w_quant)
return y
class FusedBitLinear(BitLinear):
"""BitLinear with fused RMS normalization and activation quantization."""
def __init__(self, in_features: int, out_features: int, bias: bool = False) -> None:
super().__init__(in_features=in_features, out_features=out_features, bias=bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return layer_norm_linear_quant(
x=x,
norm_weight=self.norm.weight,
norm_bias=self.norm.bias,
linear_weight=self.weight,
linear_bias=self.bias,
eps=self.norm.eps,
is_rms_norm=True,
)
|