Spaces:
Running
on
Zero
Running
on
Zero
File size: 11,704 Bytes
3de1ebe f9abc90 3de1ebe f9abc90 3de1ebe 92d8df6 f9abc90 92d8df6 f9abc90 3de1ebe |
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 |
from collections import OrderedDict
from PIL import Image
from torchao import quantize_
from torchao.quantization import Float8DynamicActivationFloat8WeightConfig, Float8WeightOnlyConfig, Int4WeightOnlyConfig, Int8DynamicActivationInt8WeightConfig, Int8WeightOnlyConfig, ModuleFqnToConfig, PerRow
from torchao.utils import get_model_size_in_bytes
from qwenimage.debug import ftimed, print_first_param
from qwenimage.experiments.experiments_qwen import ExperimentRegistry, QwenBaseExperiment
from qwenimage.models.attention_processors import QwenDoubleStreamAttnProcessorFA3
from qwenimage.optimization import optimize_pipeline_
# ModuleFqnToConfig
# @ExperimentRegistry.register(name="qwen_fa3_aot")
# class Qwen_FA3_AoT(QwenBaseExperiment):
# @ftimed
# def optimize(self):
# self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
# optimize_pipeline_(
# self.pipe,
# cache_compiled=self.config.cache_compiled,
# quantize=False,
# suffix="_fa3",
# pipe_kwargs={
# "image": [Image.new("RGB", (1024, 1024))],
# "prompt":"prompt",
# "num_inference_steps":4
# }
# )
@ExperimentRegistry.register(name="qwen_fa3_aot_fp8wo")
class Qwen_FA3_AoT_fp8wo(QwenBaseExperiment):
@ftimed
def optimize(self):
self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
optimize_pipeline_(
self.pipe,
cache_compiled=self.config.cache_compiled,
quantize=True,
quantize_config=Float8WeightOnlyConfig(),
suffix="_fp8wo_fa3",
pipe_kwargs={
"image": [Image.new("RGB", (1024, 1024))],
"prompt":"prompt",
"num_inference_steps":4
}
)
@ExperimentRegistry.register(name="qwen_fa3_aot_int8wo")
class Qwen_FA3_AoT_int8wo(QwenBaseExperiment):
@ftimed
def optimize(self):
self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
optimize_pipeline_(
self.pipe,
cache_compiled=self.config.cache_compiled,
quantize=True,
quantize_config=Int8WeightOnlyConfig(),
suffix="_int8wo_fa3",
pipe_kwargs={
"image": [Image.new("RGB", (1024, 1024))],
"prompt":"prompt",
"num_inference_steps":4
}
)
@ExperimentRegistry.register(name="qwen_fa3_aot_fp8da")
class Qwen_FA3_AoT_fp8da(QwenBaseExperiment):
@ftimed
def optimize(self):
self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
optimize_pipeline_(
self.pipe,
cache_compiled=self.config.cache_compiled,
quantize=True,
quantize_config=Float8DynamicActivationFloat8WeightConfig(),
suffix="_fp8da_fa3",
pipe_kwargs={
"image": [Image.new("RGB", (1024, 1024))],
"prompt":"prompt",
"num_inference_steps":4
}
)
@ExperimentRegistry.register(name="qwen_fa3_aot_int8da")
class Qwen_FA3_AoT_int8da(QwenBaseExperiment):
@ftimed
def optimize(self):
self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
optimize_pipeline_(
self.pipe,
cache_compiled=self.config.cache_compiled,
quantize=True,
suffix="_int8da_fa3",
quantize_config=Int8DynamicActivationInt8WeightConfig(),
pipe_kwargs={
"image": [Image.new("RGB", (1024, 1024))],
"prompt":"prompt",
"num_inference_steps":4
}
)
@ExperimentRegistry.register(name="qwen_fa3_aot_fp8darow")
class Qwen_FA3_AoT_fp8darow(QwenBaseExperiment):
@ftimed
def optimize(self):
self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
optimize_pipeline_(
self.pipe,
cache_compiled=self.config.cache_compiled,
quantize=True,
suffix="_fp8dqrow_fa3",
quantize_config=Float8DynamicActivationFloat8WeightConfig(granularity=PerRow()),
pipe_kwargs={
"image": [Image.new("RGB", (1024, 1024))],
"prompt":"prompt",
"num_inference_steps":4
}
)
ATTENTION_QKV_REGEX = "re:^transformer_blocks\\.\\d+\\.attn\\.(to_q|to_k|to_v|to_qkv|to_added_qkv|add_q_proj|add_k_proj|add_v_proj)$"
ATTENTION_QKV_REGEX = r"re:^transformer_blocks\.\d+\.attn\.(to_q|to_k|to_v|to_qkv|to_added_qkv|add_q_proj|add_k_proj|add_v_proj)$"
# Attention QKV projections (all Linear)
# Attention output projections (Linear)
ATTENTION_OUT_REGEX = r"re:^transformer_blocks\.\d+\.attn\.to_out\.0$"
ATTENTION_ADD_OUT_REGEX = r"re:^transformer_blocks\.\d+\.attn\.to_add_out$"
# Image modulation Linear layer
IMG_MOD_LINEAR_REGEX = r"re:^transformer_blocks\.\d+\.img_mod\.1$"
# Image MLP Linear layers
IMG_MLP_LINEAR1_REGEX = r"re:^transformer_blocks\.\d+\.img_mlp\.net\.0\.proj$"
IMG_MLP_LINEAR2_REGEX = r"re:^transformer_blocks\.\d+\.img_mlp\.net\.2$"
# Text modulation Linear layer
TXT_MOD_LINEAR_REGEX = r"re:^transformer_blocks\.\d+\.txt_mod\.1$"
# Text MLP Linear layers
TXT_MLP_LINEAR1_REGEX = r"re:^transformer_blocks\.\d+\.txt_mlp\.net\.0\.proj$"
TXT_MLP_LINEAR2_REGEX = r"re:^transformer_blocks\.\d+\.txt_mlp\.net\.2$"
# Top-level Linear layers (these were already fine)
IMG_IN_REGEX = r"re:^img_in$"
TXT_IN_REGEX = r"re:^txt_in$"
PROJ_OUT_REGEX = r"re:^proj_out$"
ATTN_LAST_LAYER = r"re:^transformer_blocks\.59\..*$"
ATTN_FIRST_LAYER = r"re:^transformer_blocks\.0\..*$"
@ExperimentRegistry.register(name="qwen_fa3_aot_qkvint4oint8")
class Qwen_FA3_AoT_qkvint4oint8(QwenBaseExperiment):
@ftimed
def optimize(self):
self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
module_fqn_to_config = ModuleFqnToConfig(
OrderedDict([
(ATTENTION_QKV_REGEX,Int4WeightOnlyConfig(),),
("_default",Int8WeightOnlyConfig(),),
])
)
optimize_pipeline_(
self.pipe,
cache_compiled=self.config.cache_compiled,
quantize=True,
suffix="_qkvint4oint8_fa3",
quantize_config=module_fqn_to_config,
pipe_kwargs={
"image": [Image.new("RGB", (1024, 1024))],
"prompt":"prompt",
"num_inference_steps":4
}
)
@ExperimentRegistry.register(name="qwen_fa3_aot_qkvfp8oint8")
class Qwen_FA3_AoT_qkvfp8oint8(QwenBaseExperiment):
@ftimed
def optimize(self):
self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
module_fqn_to_config = ModuleFqnToConfig(
OrderedDict([
(ATTENTION_QKV_REGEX,Float8DynamicActivationFloat8WeightConfig(),),
("_default",Int8WeightOnlyConfig(),),
])
)
optimize_pipeline_(
self.pipe,
cache_compiled=self.config.cache_compiled,
quantize=True,
suffix="_qkvfp8oint8_fa3",
quantize_config=module_fqn_to_config,
pipe_kwargs={
"image": [Image.new("RGB", (1024, 1024))],
"prompt":"prompt",
"num_inference_steps":4
}
)
@ExperimentRegistry.register(name="qwen_fa3_aot_fp8darow_nolast")
class Qwen_FA3_AoT_fp8darow_nolast(QwenBaseExperiment):
@ftimed
def optimize(self):
self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
module_fqn_to_config = ModuleFqnToConfig(
OrderedDict([
(ATTN_LAST_LAYER, None),
("_default",Float8DynamicActivationFloat8WeightConfig(granularity=PerRow()),),
])
)
optimize_pipeline_(
self.pipe,
cache_compiled=self.config.cache_compiled,
quantize=True,
suffix="_fp8darow_nolast_fa3",
quantize_config=module_fqn_to_config,
pipe_kwargs={
"image": [Image.new("RGB", (1024, 1024))],
"prompt":"prompt",
"num_inference_steps":4
}
)
def quantize_transformer_fp8da_nolast(model):
module_fqn_to_config = ModuleFqnToConfig(
OrderedDict([
(ATTN_LAST_LAYER, None),
("_default",Float8DynamicActivationFloat8WeightConfig(),),
])
)
print(f"original model size: {get_model_size_in_bytes(model) / 1024 / 1024} MB")
quantize_(model, module_fqn_to_config)
print_first_param(model)
print(f"quantized model size: {get_model_size_in_bytes(model) / 1024 / 1024} MB")
def quantize_transformer_fp8darow_nolast(model):
module_fqn_to_config = ModuleFqnToConfig(
OrderedDict([
(ATTN_LAST_LAYER, None),
("_default",Float8DynamicActivationFloat8WeightConfig(granularity=PerRow()),),
])
)
print(f"original model size: {get_model_size_in_bytes(model) / 1024 / 1024} MB")
quantize_(model, module_fqn_to_config)
print_first_param(model)
print(f"quantized model size: {get_model_size_in_bytes(model) / 1024 / 1024} MB")
def conf_fp8darow_nolast():
module_fqn_to_config = ModuleFqnToConfig(
OrderedDict([
(ATTN_LAST_LAYER, None),
("_default",Float8DynamicActivationFloat8WeightConfig(granularity=PerRow()),),
])
)
return module_fqn_to_config
@ExperimentRegistry.register(name="qwen_fa3_aot_fp8darow_nofirstlast")
class Qwen_FA3_AoT_fp8darow_nofirstlast(QwenBaseExperiment):
@ftimed
def optimize(self):
self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
module_fqn_to_config = ModuleFqnToConfig(
OrderedDict([
(ATTN_LAST_LAYER, None),
(ATTN_FIRST_LAYER, None),
("_default",Float8DynamicActivationFloat8WeightConfig(granularity=PerRow()),),
])
)
optimize_pipeline_(
self.pipe,
cache_compiled=self.config.cache_compiled,
quantize=True,
suffix="_fp8darow_nofirstlast_fa3",
quantize_config=module_fqn_to_config,
pipe_kwargs={
"image": [Image.new("RGB", (1024, 1024))],
"prompt":"prompt",
"num_inference_steps":4
}
)
@ExperimentRegistry.register(name="qwen_fa3_aot_fp8darow_nolast_cint8")
class Qwen_FA3_AoT_fp8darow_nolast_cint8(QwenBaseExperiment):
@ftimed
def optimize(self):
self.pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
module_fqn_to_config = ModuleFqnToConfig(
OrderedDict([
(ATTN_LAST_LAYER, None),
(IMG_IN_REGEX, Int8WeightOnlyConfig()),
(TXT_IN_REGEX, Int8WeightOnlyConfig()),
(PROJ_OUT_REGEX, Int8WeightOnlyConfig()),
("_default",Float8DynamicActivationFloat8WeightConfig(granularity=PerRow()),),
])
)
optimize_pipeline_(
self.pipe,
cache_compiled=self.config.cache_compiled,
quantize=True,
suffix="_fp8darow_nolast_cint8_fa3",
quantize_config=module_fqn_to_config,
pipe_kwargs={
"image": [Image.new("RGB", (1024, 1024))],
"prompt":"prompt",
"num_inference_steps":4
}
)
|