File size: 7,791 Bytes
2051791 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 | import os
import torch
import torch.nn as nn
import numpy as np
from transformers import AutoTokenizer, AutoModelForCausalLM
import tqdm
import json
import math
import torch.nn.functional as F
from datasets import load_dataset
@torch.no_grad()
def quantize_weight_per_channel_absmax(w, n_bits=8):
# w: (out_features, in_features)
scales = w.abs().max(dim=-1, keepdim=True)[0]
q_max = 2 ** (n_bits - 1) - 1
scales.clamp_(min=1e-5).div_(q_max)
w.div_(scales).round_().mul_(scales)
return w
@torch.no_grad()
def quantize_weight_per_tensor_absmax(w, n_bits=8):
# w: (out_features, in_features)
scales = w.abs().max()
q_max = 2 ** (n_bits - 1) - 1
scales.clamp_(min=1e-5).div_(q_max)
w.div_(scales).round_().mul_(scales)
return w
class W8A16Linear(nn.Module):
def __init__(
self,
# bit_width,
in_features,
out_features,
bias=True,
quantize_output=False,
):
super().__init__()
# self.bit_width = bit_width
self.in_features = in_features
self.out_features = out_features
self.register_buffer(
"weight",
torch.randn(
self.out_features,
self.in_features,
dtype=torch.float16,
requires_grad=False,
),
)
if bias:
self.register_buffer(
"bias",
torch.zeros(
(1, self.out_features), dtype=torch.float16, requires_grad=False
),
)
else:
self.register_buffer("bias", None)
def to(self, *args, **kwargs):
super(W8A16Linear, self).to(*args, **kwargs)
self.weight = self.weight.to(*args, **kwargs)
if self.bias is not None:
self.bias = self.bias.to(*args, **kwargs)
return self
@torch.no_grad()
def forward(self, x):
y = torch.functional.F.linear(x, self.weight, self.bias)
return y
@staticmethod
def from_float(
bit, module, weight_quant="per_channel", quantize_output=False
):
assert isinstance(module, torch.nn.Linear)
new_module = W8A16Linear(
# bit,
module.in_features,
module.out_features,
module.bias is not None,
quantize_output=quantize_output,
)
if weight_quant == "per_channel":
new_module.weight = quantize_weight_per_channel_absmax(module.weight, bit)
elif weight_quant == "per_tensor":
new_module.weight = quantize_weight_per_tensor_absmax(module.weight, bit)
else:
raise ValueError(f"Invalid weight_quant: {weight_quant}")
new_module.weight_quant_name = weight_quant
if module.bias is not None:
new_module.bias = module.bias
return new_module
def __repr__(self):
return f"W8A16Linear({self.in_features}, {self.out_features}, bias={self.bias is not None}, weight_quant={self.weight_quant_name})"
def quantize_llama_like(
model, mlp_quant, self_attn_quant, low_bit, bit_layers, weight_quant="per_channel", quantize_bmm_input=False
):
from transformers.models.llama.modeling_llama import (
LlamaAttention,
LlamaMLP,
)
for name, m in model.model.named_modules():
# print(name)
if isinstance(m, LlamaMLP):
if low_bit == 0:
continue
else:
if name in mlp_quant:
# bit = low_bit
layer_id = int(name.split('.')[1])
bit = bit_layers[layer_id]
print(f'{name} {bit} bit quant ')
else:
bit = low_bit
print(f'{name} {bit} bit quant ')
m.gate_proj = W8A16Linear.from_float(
bit, m.gate_proj, weight_quant=weight_quant
)
m.up_proj = W8A16Linear.from_float(
bit, m.up_proj, weight_quant=weight_quant
)
m.down_proj = W8A16Linear.from_float(
bit, m.down_proj, weight_quant=weight_quant
)
elif isinstance(m, LlamaAttention):
if low_bit == 0:
continue
else:
if name in self_attn_quant:
# bit = low_bit
layer_id = int(name.split('.')[1])
bit = bit_layers[layer_id]
print(f'{name} {bit} bit quant ')
else:
bit = low_bit
print(f'{name} {bit} bit quant ')
m.q_proj = W8A16Linear.from_float(
bit,
m.q_proj,
weight_quant=weight_quant,
quantize_output=quantize_bmm_input,
)
m.k_proj = W8A16Linear.from_float(
bit,
m.k_proj,
weight_quant=weight_quant,
quantize_output=quantize_bmm_input,
)
m.v_proj = W8A16Linear.from_float(
bit,
m.v_proj,
weight_quant=weight_quant,
quantize_output=quantize_bmm_input,
)
m.o_proj = W8A16Linear.from_float(
bit, m.o_proj, weight_quant=weight_quant
)
return model
def quantize_qwen_like(
model, mlp_quant, self_attn_quant, low_bit, bit_layers, weight_quant="per_channel", quantize_bmm_input=False
):
from transformers.models.qwen2.modeling_qwen2 import (
Qwen2Attention,
Qwen2MLP,
)
for name, m in model.model.named_modules():
# print(name)
if isinstance(m, Qwen2MLP):
if low_bit == 0:
continue
else:
if name in mlp_quant:
# bit = low_bit
layer_id = int(name.split('.')[1])
bit = bit_layers[layer_id]
print(f'{name} {bit} bit quant ')
else:
bit = low_bit
print(f'{name} {bit} bit quant ')
m.gate_proj = W8A16Linear.from_float(
bit, m.gate_proj, weight_quant=weight_quant
)
m.up_proj = W8A16Linear.from_float(
bit, m.up_proj, weight_quant=weight_quant
)
m.down_proj = W8A16Linear.from_float(
bit, m.down_proj, weight_quant=weight_quant
)
elif isinstance(m, Qwen2Attention):
if low_bit == 0:
continue
else:
if name in self_attn_quant:
# bit = low_bit
layer_id = int(name.split('.')[1])
bit = bit_layers[layer_id]
print(f'{name} {bit} bit quant ')
else:
bit = low_bit
print(f'{name} {bit} bit quant ')
m.q_proj = W8A16Linear.from_float(
bit,
m.q_proj,
weight_quant=weight_quant,
quantize_output=quantize_bmm_input,
)
m.k_proj = W8A16Linear.from_float(
bit,
m.k_proj,
weight_quant=weight_quant,
quantize_output=quantize_bmm_input,
)
m.v_proj = W8A16Linear.from_float(
bit,
m.v_proj,
weight_quant=weight_quant,
quantize_output=quantize_bmm_input,
)
m.o_proj = W8A16Linear.from_float(
bit, m.o_proj, weight_quant=weight_quant
)
return model
|