File size: 11,253 Bytes
28404e6 | 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 | # Copyright 2025 Tencent Inc. All Rights Reserved.
#
# 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.
import copy
import logging
import re
from typing import Callable, List, Optional, Tuple, Union
import torch
import tqdm
from .modules import FP4DynamicLinear, FP8DynamicLinear, FP8WeightOnlyLinear, INT8DynamicLinear
from .quant_func import (
fp8_per_block_quant,
fp8_per_channel_quant,
fp8_per_tensor_quant,
fp8_per_token_group_quant,
int8_per_channel_quant,
mxfp4_per_tensor_quant,
mxfp6_per_tensor_quant,
mxfp8_per_tensor_quant,
nvfp4_per_tensor_quant,
)
from .utils import (
QuantType,
_ensure_deep_gemm,
_ensure_sgl_kernel,
cleanup_memory,
load_fp8_scales,
load_quantized_model,
replace_module,
save_quantized_model,
should_quantize_layer,
)
__all__ = ["DynamicDiTQuantizer"]
logger = logging.getLogger(__name__)
class DynamicDiTQuantizer:
"""
Quantizer for DiT that supports various FP8 quantization strategies.
Optimized for efficient initialization and conversion.
"""
def __init__(
self,
quant_type: str = QuantType.FP8_PER_TENSOR,
layer_filter: Optional[Callable[[str], bool]] = None,
include_patterns: Optional[List[Union[str, re.Pattern]]] = None,
exclude_patterns: Optional[List[Union[str, re.Pattern]]] = None,
native_fp8_support: Optional[bool] = None,
):
QuantType.validate(quant_type)
self.fp8_scales_map = {}
self.quant_type = quant_type
self.include_patterns = include_patterns or [
"wrapped_module",
"block",
"lin",
"img",
"txt",
]
self.exclude_patterns = exclude_patterns or ["embed"]
# Configure layer filter function
self.layer_filter = (
layer_filter
if layer_filter is not None
else lambda name: should_quantize_layer(
name, self.include_patterns, self.exclude_patterns
)
)
# Auto-detect FP8 native support, fallback to False if not present
if native_fp8_support is not None:
self.native_fp8_support = native_fp8_support
else:
self.native_fp8_support = (
torch.cuda.is_available() and torch.cuda.get_device_capability() >= (8, 9)
)
self.quantize_linear_module = self._set_quantize_linear_module()
def _set_quantize_linear_module(self) -> torch.nn.Module:
if self.quant_type in (
QuantType.INT8,
QuantType.INT8_TORCHAO,
QuantType.INT8_VLLM,
QuantType.INT8_TRITON,
QuantType.INT8_SGL,
QuantType.INT8_Q8F,
):
return INT8DynamicLinear
if self.quant_type in (QuantType.NVFP4, QuantType.MXFP4, QuantType.MXFP6, QuantType.MXFP8):
return FP4DynamicLinear
if "fp8" in self.quant_type:
if self.quant_type == QuantType.FP8_PER_TENSOR_WEIGHT_ONLY:
return FP8WeightOnlyLinear
else:
return FP8DynamicLinear
raise ValueError(f"Invalid quant_type: {self.quant_type}")
def _quantize_linear_weight(
self, linear: torch.nn.Linear
) -> Union[Tuple[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor, torch.Tensor]]:
if self.quant_type == QuantType.FP8_PER_TENSOR:
quant_weight, weight_scale = fp8_per_tensor_quant(linear.weight)
elif self.quant_type == QuantType.FP8_PER_TOKEN:
quant_weight, weight_scale = fp8_per_token_group_quant(
linear.weight, linear.weight.shape[-1]
)
weight_scale = weight_scale.t()
elif self.quant_type == QuantType.FP8_PER_TOKEN_SGL:
if self.native_fp8_support:
_ensure_sgl_kernel()
quant_weight, weight_scale = fp8_per_channel_quant(linear.weight)
elif self.quant_type == QuantType.FP8_PER_BLOCK:
if self.native_fp8_support:
_ensure_deep_gemm()
quant_weight, weight_scale = fp8_per_block_quant(linear.weight)
elif self.quant_type == QuantType.FP8_PER_CHANNEL_VLLM:
quant_weight, weight_scale = fp8_per_channel_quant(linear.weight)
elif self.quant_type == QuantType.FP8_PER_TENSOR_WEIGHT_ONLY:
quant_weight, weight_scale = fp8_per_tensor_quant(linear.weight)
elif self.quant_type in (
QuantType.INT8,
QuantType.INT8_TORCHAO,
QuantType.INT8_VLLM,
QuantType.INT8_TRITON,
QuantType.INT8_SGL,
QuantType.INT8_Q8F,
):
quant_weight, weight_scale = int8_per_channel_quant(linear.weight)
elif self.quant_type == QuantType.NVFP4:
quant_weight, weight_scale, weight_global_scale = nvfp4_per_tensor_quant(linear.weight)
return quant_weight, weight_scale, weight_global_scale
elif self.quant_type == QuantType.MXFP4:
quant_weight, weight_scale, weight_global_scale = mxfp4_per_tensor_quant(linear.weight)
return quant_weight, weight_scale, weight_global_scale
elif self.quant_type == QuantType.MXFP8:
quant_weight, weight_scale, weight_global_scale = mxfp8_per_tensor_quant(linear.weight)
return quant_weight, weight_scale, weight_global_scale
elif self.quant_type == QuantType.MXFP6:
# Mixed kernel route expects MXFP8-formatted weights.
quant_weight, weight_scale, weight_global_scale = mxfp6_per_tensor_quant(linear.weight)
return quant_weight, weight_scale, weight_global_scale
else:
raise ValueError(f"Invalid quant_type: {self.quant_type}")
return quant_weight, weight_scale
def _convert_linear_with_scale(
self, model: torch.nn.Module, scale: Union[torch.Tensor, float]
):
model.to(torch.bfloat16)
assert scale is not None, "scale is required"
self.fp8_scales_map = load_fp8_scales(scale)
converted_count = 0
for name, module in tqdm.tqdm(list(model.named_modules()), desc="converting linear"):
if isinstance(module, torch.nn.Linear) and self.layer_filter(name):
# Prefer $name.weight_scale, fallback to "$name" key if needed
s = self.fp8_scales_map.get(f"{name}.weight_scale") or self.fp8_scales_map.get(
name
)
if s is None:
continue
# import pdb; pdb.set_trace()
weight = module.weight.to(torch.float8_e4m3fn)
bias = copy.deepcopy(module.bias) if module.bias is not None else None
quant_linear = self.quantize_linear_module(
weight=weight,
weight_scale=s.float(),
bias=bias,
native_fp8_support=self.native_fp8_support,
quant_type=self.quant_type,
)
replace_module(model, name, quant_linear)
del module.weight, module.bias, module
converted_count += 1
print(f"[INIT] quantized linear layers: {converted_count}")
cleanup_memory()
def _convert_linear(self, model: torch.nn.Module):
model.to(torch.bfloat16)
named_modules = list(model.named_modules())
converted_count = 0
for name, module in tqdm.tqdm(named_modules, desc="converting linear"):
if isinstance(module, torch.nn.Linear) and self.layer_filter(name):
quantized = self._quantize_linear_weight(module)
if self.quant_type in (
QuantType.NVFP4,
QuantType.MXFP4,
QuantType.MXFP6,
QuantType.MXFP8,
):
quant_weight, weight_scale, weight_global_scale = quantized
else:
quant_weight, weight_scale = quantized
self.fp8_scales_map[f"{name}.weight_scale"] = weight_scale
bias = copy.deepcopy(module.bias) if module.bias is not None else None
if self.quant_type in (
QuantType.NVFP4,
QuantType.MXFP4,
QuantType.MXFP6,
QuantType.MXFP8,
):
self.fp8_scales_map[f"{name}.weight_global_scale"] = weight_global_scale
quant_linear = self.quantize_linear_module(
weight=quant_weight,
weight_scale=weight_scale,
weight_global_scale=weight_global_scale,
bias=bias,
native_fp8_support=self.native_fp8_support,
quant_type=self.quant_type,
)
else:
quant_linear = self.quantize_linear_module(
weight=quant_weight,
weight_scale=weight_scale,
bias=bias,
native_fp8_support=self.native_fp8_support,
quant_type=self.quant_type,
)
replace_module(model, name, quant_linear)
del module.weight, module.bias, module
converted_count += 1
print(f"[INIT] quantized linear layers: {converted_count}")
cleanup_memory()
def convert_linear(
self, model: torch.nn.Module, scale: Optional[Union[torch.Tensor, float]] = None
):
if scale is not None:
self._convert_linear_with_scale(model, scale)
else:
self._convert_linear(model)
def export_quantized_weight(self, model: torch.nn.Module, save_path: str):
assert (
self.quant_type == QuantType.FP8_PER_TENSOR
), "Currently only FP8_PER_TENSOR is supported for export"
self.convert_linear(model)
save_quantized_model(model, save_path, self.fp8_scales_map)
logger.info(f"Quantized model saved to {save_path}")
logger.info(f"Quantized scales saved to {save_path}/fp8_scales.safetensors")
@staticmethod
def load_quantized_model(model_class, save_path: str, device: str = "cpu"):
"""
Load a quantized model from save_path.
Args:
model_class: The model class to instantiate
save_path: Path to the saved model directory
device: Device to load the model on
Returns:
Loaded model with quantized weights
"""
return load_quantized_model(model_class, save_path, device)
|