Spaces:
Running
on
Zero
Running
on
Zero
File size: 17,320 Bytes
d711508 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 |
# Copyright 2023-present the HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# Reference code: https://github.com/yxli2123/LoftQ/blob/main/utils.py
# Reference paper: https://arxiv.org/abs/2310.08659
from __future__ import annotations
import logging
import os
from typing import Callable, Optional, Union
import torch
from huggingface_hub import snapshot_download
from huggingface_hub.utils import LocalEntryNotFoundError
from safetensors import SafetensorError, safe_open
from transformers.utils import cached_file
from transformers.utils.hub import get_checkpoint_shard_files
from peft.import_utils import is_bnb_4bit_available, is_bnb_available
class NFQuantizer:
def __init__(self, num_bits=2, device="cuda", method="normal", block_size=64, *args, **kwargs):
super().__init__(*args, **kwargs)
self.num_bits = num_bits
self.device = device
self.method = method
self.block_size = block_size
if self.method == "normal":
self.norm_lookup_table = self.create_normal_map(num_bits=self.num_bits)
self.norm_lookup_table = self.norm_lookup_table.to(device)
elif self.method == "uniform":
self.norm_lookup_table = self.create_uniform_map(num_bits=self.num_bits)
self.norm_lookup_table = self.norm_lookup_table.to(device)
else:
raise NotImplementedError("Other quantization methods not supported yet.")
@staticmethod
def create_uniform_map(symmetric=False, num_bits=4):
if symmetric:
# print("symmetric uniform quantization")
negative = torch.linspace(-1, 0, 2 ** (num_bits - 1))
positive = torch.linspace(0, 1, 2 ** (num_bits - 1))
table = torch.cat([negative, positive[1:]])
else:
# print("asymmetric uniform quantization")
table = torch.linspace(-1, 1, 2**num_bits)
return table
@staticmethod
def create_normal_map(offset=0.9677083, symmetric=False, num_bits=2):
try:
from scipy.stats import norm
except ImportError:
raise ImportError("The required package 'scipy' is not installed. Please install it to continue.")
variations = 2**num_bits
if symmetric:
v = norm.ppf(torch.linspace(1 - offset, offset, variations + 1)).tolist()
values = []
for index in range(len(v) - 1):
values.append(0.5 * v[index] + 0.5 * v[index + 1])
v = values
else:
# one more positive value, this is an asymmetric type
v1 = norm.ppf(torch.linspace(offset, 0.5, variations // 2 + 1)[:-1]).tolist()
v2 = [0]
v3 = (-norm.ppf(torch.linspace(offset, 0.5, variations // 2)[:-1])).tolist()
v = v1 + v2 + v3
values = torch.Tensor(v)
values = values.sort().values
values /= values.max()
return values
def quantize_tensor(self, weight):
max_abs = torch.abs(weight).max()
weight_normed = weight / max_abs
weight_normed_expanded = weight_normed.unsqueeze(-1)
# Reshape L to have the same number of dimensions as X_expanded
L_reshaped = torch.tensor(self.norm_lookup_table).reshape(1, -1)
# Calculate the absolute difference between X_expanded and L_reshaped
abs_diff = torch.abs(weight_normed_expanded - L_reshaped)
# Find the index of the minimum absolute difference for each element
qweight = torch.argmin(abs_diff, dim=-1)
return qweight, max_abs
def dequantize_tensor(self, qweight, max_abs):
qweight_flatten = qweight.flatten()
weight_normed = self.norm_lookup_table[qweight_flatten]
weight = weight_normed * max_abs
weight = weight.reshape(qweight.shape)
return weight
def quantize_block(self, weight):
if len(weight.shape) != 2:
raise ValueError(f"Only support 2D matrix, but your input has {len(weight.shape)} dimensions.")
if weight.shape[0] * weight.shape[1] % self.block_size != 0:
raise ValueError(
f"Weight with shape ({weight.shape[0]} x {weight.shape[1]}) "
f"is not dividable by block size {self.block_size}."
)
M, N = weight.shape
device = weight.device
# Quantization
weight_flatten = weight.flatten() # (M*N, )
weight_block = weight_flatten.reshape(-1, self.block_size) # (L, B), L = M * N / B
if self.method == "normal":
weight_max = weight_block.abs().max(dim=-1)[0] # (L, 1)
elif self.method == "uniform":
weight_max = weight_block.mean(dim=-1) + 2.5 * weight_block.std(dim=-1)
else:
raise NotImplementedError("Method not supported yet.")
weight_max = weight_max.unsqueeze(-1)
weight_divabs = weight_block / weight_max # (L, B)
weight_divabs = weight_divabs.unsqueeze(-1) # (L, B, 1)
L_reshaped = self.norm_lookup_table.reshape(1, -1) # (1, 2**K)
abs_diff = torch.abs(weight_divabs - L_reshaped) # (L, B, 2**K)
qweight = torch.argmin(abs_diff, dim=-1) # (L, B)
# Pack multiple k-bit into uint8
qweight = qweight.reshape(-1, 8 // self.num_bits)
qweight_pack = torch.zeros((M * N // 8 * self.num_bits, 1), dtype=torch.uint8, device=device)
# data format example:
# [1, 0, 3, 2] or [01, 00, 11, 10] -> [10110001], LIFO
for i in range(8 // self.num_bits):
qweight[:, i] = qweight[:, i] << i * self.num_bits
qweight_pack[:, 0] |= qweight[:, i]
return qweight_pack, weight_max, weight.shape
def dequantize_block(self, qweight, weight_max, weight_shape):
# unpack weight
device = qweight.device
weight = torch.zeros((qweight.shape[0], 8 // self.num_bits), dtype=torch.float32, device=device)
for i in range(8 // self.num_bits):
lookup_table_idx = qweight.to(torch.long) % 2**self.num_bits # get the most right 2 bits
lookup_table_idx = lookup_table_idx.to(torch.long)
weight[:, i] = self.norm_lookup_table[lookup_table_idx].squeeze()
qweight = qweight >> self.num_bits # right shift 2 bits of the original data
weight_block = weight.reshape(-1, self.block_size)
weight = weight_block * weight_max
weight = weight.reshape(weight_shape)
return weight
def _low_rank_decomposition(weight, reduced_rank=32):
"""
:param weight: The matrix to decompose, of shape (H, W) :param reduced_rank: the final rank :return:
"""
matrix_dimension = len(weight.size())
if matrix_dimension != 2:
raise ValueError(f"Only support 2D matrix, but your input has {matrix_dimension} dimensions.")
# Use SVD to decompose a matrix, default full_matrices is False to save parameters
U, S, Vh = torch.linalg.svd(weight, full_matrices=False)
L = U @ (torch.sqrt(torch.diag(S)[:, 0:reduced_rank]))
R = torch.sqrt(torch.diag(S)[0:reduced_rank, :]) @ Vh
return {"L": L, "R": R, "U": U, "S": S, "Vh": Vh, "reduced_rank": reduced_rank}
@torch.no_grad()
def loftq_init(weight: Union[torch.Tensor, torch.nn.Parameter], num_bits: int, reduced_rank: int, num_iter=1):
if is_bnb_available():
import bitsandbytes as bnb
else:
raise ValueError("bitsandbytes is not available, please install it to use LoftQ.")
if num_bits not in [2, 4, 8]:
raise ValueError("Only support 2, 4, 8 bits quantization")
if num_iter <= 0:
raise ValueError("Number of iterations must be greater than 0")
out_feature, in_feature = weight.size()
device = weight.device
dtype = weight.dtype
logging.info(
f"Weight: ({out_feature}, {in_feature}) | Rank: {reduced_rank} "
f"| Num Iter: {num_iter} | Num Bits: {num_bits}"
)
if not is_bnb_4bit_available() or num_bits in [2, 8]:
quantizer = NFQuantizer(num_bits=num_bits, device=device, method="normal", block_size=64)
compute_device = device
else:
compute_device = "cuda"
weight = weight.to(device=compute_device, dtype=torch.float32)
res = weight.clone()
for i in range(num_iter):
torch.cuda.empty_cache()
# Quantization
if num_bits == 4 and is_bnb_4bit_available():
qweight = bnb.nn.Params4bit(
res.to("cpu"), requires_grad=False, compress_statistics=False, quant_type="nf4"
).to(compute_device)
dequantized_weight = bnb.functional.dequantize_4bit(qweight.data, qweight.quant_state)
else:
quantized_weight, max_abs, shape = quantizer.quantize_block(res)
dequantized_weight = quantizer.dequantize_block(quantized_weight, max_abs, shape)
res = weight - dequantized_weight
# Decompose the residual by SVD
output = _low_rank_decomposition(res, reduced_rank=reduced_rank)
L, R, reduced_rank = output["L"], output["R"], output["reduced_rank"]
res = weight - torch.mm(L, R)
lora_A, lora_B = R, L
return dequantized_weight.to(device=device, dtype=dtype), lora_A, lora_B
@torch.no_grad()
def _loftq_init_new(qweight, weight, num_bits: int, reduced_rank: int):
import bitsandbytes as bnb
if num_bits != 4:
raise ValueError("Only 4 bit quantization supported at the moment.")
if not is_bnb_4bit_available():
raise ValueError("bitsandbytes 4bit quantization is not available.")
compute_device = "cuda"
dequantized_weight = bnb.functional.dequantize_4bit(qweight.data, qweight.quant_state)
weight = weight.to(device=compute_device, dtype=torch.float32)
residual = weight - dequantized_weight
torch.cuda.empty_cache()
# Decompose the residualidual by SVD
output = _low_rank_decomposition(residual, reduced_rank=reduced_rank)
L, R, reduced_rank = output["L"], output["R"], output["reduced_rank"]
return R, L
class _SafetensorLoader:
"""
Simple utility class that loads tensors with safetensors from a single file or sharded files.
Takes care of file name normalization etc.
"""
def __init__(self, peft_model, model_path):
if model_path is None:
try:
model_path = snapshot_download(peft_model.base_model.config._name_or_path, local_files_only=True)
except AttributeError as exc:
raise ValueError(
"The provided model does not appear to be a transformers model. In this case, you must pass the "
"model_path to the safetensors file."
) from exc
except LocalEntryNotFoundError as exc:
raise ValueError(
"The model.safetensors file must be present on disk, but it could not be found."
) from exc
suffix = "model.safetensors"
if not model_path.endswith(suffix):
model_path = os.path.join(model_path, suffix)
self.model_path = model_path
self.base_model_prefix = getattr(peft_model.get_base_model(), "base_model_prefix", None)
self.prefix = "base_model.model."
self.is_sharded = False
self.weight_map = None
if not os.path.exists(model_path):
# check if the file is sharded
par_dir = model_path.rpartition(os.path.sep)[0]
try:
resolved_archive_file, sharded_metadata = get_checkpoint_shard_files(
par_dir, cached_file(par_dir, "model.safetensors.index.json")
)
except OSError as exc:
raise FileNotFoundError(
f"Could not find file for {model_path}, ensure that there is a (sharded) safetensors file of the model."
) from exc
self.is_sharded = True
# maps from 'model-X-of-Y.safetensors' to full file path
file_map = {k.rpartition(os.path.sep)[-1]: k for k in resolved_archive_file}
self.weight_map = {k: file_map[v] for k, v in sharded_metadata["weight_map"].items()}
def get_tensor(self, name):
if not self.is_sharded:
file_path = self.model_path
else:
file_path = self.weight_map[name]
with safe_open(file_path, framework="pt", device="cpu") as f:
try:
tensor = f.get_tensor(name)
except SafetensorError as exc:
# no matching key found, we probably need to remove the base model prefix
if self.base_model_prefix:
# remove 1 extra character for "."
name = name[len(self.base_model_prefix) + 1 :]
tensor = f.get_tensor(name)
else:
raise exc
return tensor
@torch.no_grad()
def replace_lora_weights_loftq(
peft_model,
model_path: Optional[str] = None,
adapter_name: str = "default",
callback: Optional[Callable[[torch.nn.Module, str], bool]] = None,
):
"""
Replace the LoRA weights of a model quantized with bitsandbytes, using the LoftQ technique.
The replacement is done on the fly by loading in the non-quantized weights from a locally stored safetensors model
file and initializing the LoRA weights such that the quantization error between the original and quantized weights
is minimized.
As lazy loading is not possible with pickle, normal PyTorch checkpoint files cannot be supported.
Depending on the model size, calling this function may take some time to finish.
Args:
peft_model (`PeftModel`):
The model to replace the weights of. Must be a quantized PEFT model with LoRA layers.
model_path (`Optional[str]`):
The path to the model safetensors file. If the model is a Hugging Face model, this will be inferred from
the model's config. Otherwise, it must be provided.
adapter_name (`str`):
The name of the adapter to replace the weights of. The default adapter name is "default".
callback (`Optional[Callable[[PeftModel, str], bool]]`):
A callback function that will be called after each module is replaced. The callback function should take
the model and the name of the current module as input and return a boolean indicating whether the
replacement should be kept. If the callback returns False, the replacement will be rolled back. This can be
very useful to confirm that the LoftQ initialization actually decreases the quantization error of the
model. As an example, this callback could generate logits for given input and compare it with the logits
from the original, non-quanitzed model with the same input, and only return `True` if there is an
improvement. As this is a greedy optimization, it's possible that calling this function multiple times
yields incremental improvements.
"""
if not is_bnb_4bit_available():
raise ValueError("bitsandbytes must be installed and the model must be quantized in 4bits.")
from peft.tuners.lora import Linear4bit
# model_path = _check_model_path_loftq(model_path, peft_model)
prefix = "base_model.model."
any_match = False
safetensor_loader = _SafetensorLoader(peft_model, model_path)
# if too slow, consider adding tqdm as an option
for name, module in peft_model.named_modules():
if not isinstance(module, Linear4bit):
continue
if not name.startswith(prefix):
raise TypeError("The passed model does not appear to be a valid PeftModel")
any_match = True
name = name[len(prefix) :]
tensor = safetensor_loader.get_tensor(name + ".weight")
reduced_rank = module.r[adapter_name]
lora_A, lora_B = _loftq_init_new(module.weight, tensor, num_bits=4, reduced_rank=reduced_rank)
if not callback:
module.lora_A[adapter_name].weight.data = lora_A
module.lora_B[adapter_name].weight.data = lora_B
continue
lora_A_before = module.lora_A[adapter_name].weight.data
lora_B_before = module.lora_B[adapter_name].weight.data
module.lora_A[adapter_name].weight.data = lora_A
module.lora_B[adapter_name].weight.data = lora_B
should_replace = callback(peft_model, name)
if not should_replace:
# roll back
module.lora_A[adapter_name].weight.data = lora_A_before
module.lora_B[adapter_name].weight.data = lora_B_before
del lora_A_before, lora_B_before
if not any_match:
raise ValueError("No bnb LoRA module found on the model")
|