Spaces:
Running
Running
MekkCyber
commited on
Commit
·
70dd883
1
Parent(s):
b5887d5
updating to use AoBaseConfig
Browse files
app.py
CHANGED
|
@@ -12,12 +12,15 @@ from torchao.quantization import (
|
|
| 12 |
Int8WeightOnlyConfig,
|
| 13 |
Int8DynamicActivationInt8WeightConfig,
|
| 14 |
Float8WeightOnlyConfig,
|
|
|
|
| 15 |
)
|
| 16 |
|
| 17 |
MAP_QUANT_TYPE_TO_NAME = {
|
| 18 |
"int4_weight_only": "int4wo",
|
| 19 |
"int8_weight_only": "int8wo",
|
| 20 |
-
"int8_dynamic_activation_int8_weight": "
|
|
|
|
|
|
|
| 21 |
"autoquant": "autoquant",
|
| 22 |
}
|
| 23 |
MAP_QUANT_TYPE_TO_CONFIG = {
|
|
@@ -25,6 +28,7 @@ MAP_QUANT_TYPE_TO_CONFIG = {
|
|
| 25 |
"int8_weight_only": Int8WeightOnlyConfig,
|
| 26 |
"int8_dynamic_activation_int8_weight": Int8DynamicActivationInt8WeightConfig,
|
| 27 |
"float8_weight_only": Float8WeightOnlyConfig,
|
|
|
|
| 28 |
}
|
| 29 |
|
| 30 |
|
|
@@ -164,16 +168,30 @@ It's quantized using the TorchAO library using the [torchao-my-repo](https://hug
|
|
| 164 |
|
| 165 |
|
| 166 |
def quantize_model(
|
| 167 |
-
model_name, quantization_type, group_size=128, auth_token=None, username=None
|
| 168 |
):
|
| 169 |
print(f"Quantizing model: {quantization_type}")
|
|
|
|
| 170 |
if (
|
| 171 |
-
quantization_type == "
|
| 172 |
-
or quantization_type == "int8_weight_only"
|
| 173 |
):
|
| 174 |
-
|
| 175 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 176 |
quantization_config = TorchAoConfig(quantization_type)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 177 |
model = AutoModel.from_pretrained(
|
| 178 |
model_name,
|
| 179 |
torch_dtype="auto",
|
|
@@ -181,7 +199,7 @@ def quantize_model(
|
|
| 181 |
device_map="cpu",
|
| 182 |
use_auth_token=auth_token.token,
|
| 183 |
)
|
| 184 |
-
|
| 185 |
return model
|
| 186 |
|
| 187 |
|
|
@@ -193,7 +211,10 @@ def save_model(
|
|
| 193 |
username=None,
|
| 194 |
auth_token=None,
|
| 195 |
quantized_model_name=None,
|
|
|
|
|
|
|
| 196 |
):
|
|
|
|
| 197 |
print("Saving quantized model")
|
| 198 |
with tempfile.TemporaryDirectory() as tmpdirname:
|
| 199 |
# Load and save the tokenizer
|
|
@@ -203,10 +224,11 @@ def save_model(
|
|
| 203 |
tokenizer.save_pretrained(tmpdirname, use_auth_token=auth_token.token)
|
| 204 |
|
| 205 |
# Save the model
|
|
|
|
| 206 |
model.save_pretrained(
|
| 207 |
tmpdirname, safe_serialization=False, use_auth_token=auth_token.token
|
| 208 |
)
|
| 209 |
-
|
| 210 |
if quantized_model_name:
|
| 211 |
repo_name = f"{username}/{quantized_model_name}"
|
| 212 |
else:
|
|
@@ -217,19 +239,21 @@ def save_model(
|
|
| 217 |
repo_name = f"{username}/{model_name.split('/')[-1]}-ao-{MAP_QUANT_TYPE_TO_NAME[quantization_type.lower()]}-gs{group_size}"
|
| 218 |
else:
|
| 219 |
repo_name = f"{username}/{model_name.split('/')[-1]}-ao-{MAP_QUANT_TYPE_TO_NAME[quantization_type.lower()]}"
|
| 220 |
-
|
| 221 |
model_card = create_model_card(model_name, quantization_type, group_size)
|
| 222 |
with open(os.path.join(tmpdirname, "README.md"), "w") as f:
|
| 223 |
f.write(model_card)
|
| 224 |
# Push to Hub
|
| 225 |
api = HfApi(token=auth_token.token)
|
| 226 |
-
api.create_repo(repo_name, exist_ok=True)
|
|
|
|
| 227 |
api.upload_folder(
|
| 228 |
folder_path=tmpdirname,
|
| 229 |
repo_id=repo_name,
|
| 230 |
repo_type="model",
|
| 231 |
)
|
| 232 |
-
|
|
|
|
| 233 |
import io
|
| 234 |
from contextlib import redirect_stdout
|
| 235 |
import html
|
|
@@ -273,6 +297,7 @@ def quantize_and_save(
|
|
| 273 |
quantization_type,
|
| 274 |
group_size,
|
| 275 |
quantized_model_name,
|
|
|
|
| 276 |
):
|
| 277 |
if oauth_token is None:
|
| 278 |
return """
|
|
@@ -332,8 +357,10 @@ def quantize_and_save(
|
|
| 332 |
profile.username,
|
| 333 |
oauth_token,
|
| 334 |
quantized_model_name,
|
|
|
|
| 335 |
)
|
| 336 |
except Exception as e:
|
|
|
|
| 337 |
return str(e)
|
| 338 |
|
| 339 |
|
|
@@ -464,24 +491,44 @@ with gr.Blocks(theme=gr.themes.Ocean(), css=css) as demo:
|
|
| 464 |
"int4_weight_only",
|
| 465 |
"int8_weight_only",
|
| 466 |
"int8_dynamic_activation_int8_weight",
|
|
|
|
|
|
|
| 467 |
"autoquant",
|
| 468 |
],
|
| 469 |
value="int8_weight_only",
|
| 470 |
filterable=False,
|
| 471 |
show_label=False,
|
| 472 |
)
|
|
|
|
| 473 |
group_size = gr.Textbox(
|
| 474 |
info="Group Size (only for int4_weight_only and int8_weight_only)",
|
| 475 |
value="128",
|
| 476 |
-
interactive=
|
| 477 |
show_label=False,
|
| 478 |
)
|
| 479 |
-
|
| 480 |
-
|
| 481 |
-
|
| 482 |
-
|
| 483 |
-
|
| 484 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 485 |
|
| 486 |
with gr.Column():
|
| 487 |
quantize_button = gr.Button(
|
|
@@ -517,11 +564,10 @@ with gr.Blocks(theme=gr.themes.Ocean(), css=css) as demo:
|
|
| 517 |
- int8_weight_only typically reduces size by about 50%
|
| 518 |
"""
|
| 519 |
)
|
| 520 |
-
|
| 521 |
# Keep existing click handler
|
| 522 |
quantize_button.click(
|
| 523 |
fn=quantize_and_save,
|
| 524 |
-
inputs=[model_name, quantization_type, group_size, quantized_model_name],
|
| 525 |
outputs=[output_link],
|
| 526 |
)
|
| 527 |
|
|
|
|
| 12 |
Int8WeightOnlyConfig,
|
| 13 |
Int8DynamicActivationInt8WeightConfig,
|
| 14 |
Float8WeightOnlyConfig,
|
| 15 |
+
Float8DynamicActivationFloat8WeightConfig,
|
| 16 |
)
|
| 17 |
|
| 18 |
MAP_QUANT_TYPE_TO_NAME = {
|
| 19 |
"int4_weight_only": "int4wo",
|
| 20 |
"int8_weight_only": "int8wo",
|
| 21 |
+
"int8_dynamic_activation_int8_weight": "int8da8w8",
|
| 22 |
+
"float8_weight_only": "float8wo",
|
| 23 |
+
"float8_dynamic_activation_float8_weight": "float8da8w8",
|
| 24 |
"autoquant": "autoquant",
|
| 25 |
}
|
| 26 |
MAP_QUANT_TYPE_TO_CONFIG = {
|
|
|
|
| 28 |
"int8_weight_only": Int8WeightOnlyConfig,
|
| 29 |
"int8_dynamic_activation_int8_weight": Int8DynamicActivationInt8WeightConfig,
|
| 30 |
"float8_weight_only": Float8WeightOnlyConfig,
|
| 31 |
+
"float8_dynamic_activation_float8_weight": Float8DynamicActivationFloat8WeightConfig,
|
| 32 |
}
|
| 33 |
|
| 34 |
|
|
|
|
| 168 |
|
| 169 |
|
| 170 |
def quantize_model(
|
| 171 |
+
model_name, quantization_type, group_size=128, auth_token=None, username=None, progress=gr.Progress()
|
| 172 |
):
|
| 173 |
print(f"Quantizing model: {quantization_type}")
|
| 174 |
+
progress(0, desc="Preparing Quantization")
|
| 175 |
if (
|
| 176 |
+
quantization_type == "int8_weight_only"
|
|
|
|
| 177 |
):
|
| 178 |
+
quant_config = MAP_QUANT_TYPE_TO_CONFIG[quantization_type](
|
| 179 |
+
group_size=group_size
|
| 180 |
+
)
|
| 181 |
+
quantization_config = TorchAoConfig(quant_config)
|
| 182 |
+
elif quantization_type == "int4_weight_only":
|
| 183 |
+
from torchao.dtypes import Int4CPULayout
|
| 184 |
+
|
| 185 |
+
quant_config = MAP_QUANT_TYPE_TO_CONFIG[quantization_type](
|
| 186 |
+
group_size=group_size, layout=Int4CPULayout()
|
| 187 |
+
)
|
| 188 |
+
quantization_config = TorchAoConfig(quant_config)
|
| 189 |
+
elif quantization_type == "autoquant":
|
| 190 |
quantization_config = TorchAoConfig(quantization_type)
|
| 191 |
+
else:
|
| 192 |
+
quant_config = MAP_QUANT_TYPE_TO_CONFIG[quantization_type]()
|
| 193 |
+
quantization_config = TorchAoConfig(quant_config)
|
| 194 |
+
progress(0.10, desc="Quantizing model")
|
| 195 |
model = AutoModel.from_pretrained(
|
| 196 |
model_name,
|
| 197 |
torch_dtype="auto",
|
|
|
|
| 199 |
device_map="cpu",
|
| 200 |
use_auth_token=auth_token.token,
|
| 201 |
)
|
| 202 |
+
progress(0.45, desc="Quantization completed")
|
| 203 |
return model
|
| 204 |
|
| 205 |
|
|
|
|
| 211 |
username=None,
|
| 212 |
auth_token=None,
|
| 213 |
quantized_model_name=None,
|
| 214 |
+
public=True,
|
| 215 |
+
progress=gr.Progress(),
|
| 216 |
):
|
| 217 |
+
progress(0.50, desc="Preparing to push")
|
| 218 |
print("Saving quantized model")
|
| 219 |
with tempfile.TemporaryDirectory() as tmpdirname:
|
| 220 |
# Load and save the tokenizer
|
|
|
|
| 224 |
tokenizer.save_pretrained(tmpdirname, use_auth_token=auth_token.token)
|
| 225 |
|
| 226 |
# Save the model
|
| 227 |
+
progress(0.60, desc="Saving model")
|
| 228 |
model.save_pretrained(
|
| 229 |
tmpdirname, safe_serialization=False, use_auth_token=auth_token.token
|
| 230 |
)
|
| 231 |
+
|
| 232 |
if quantized_model_name:
|
| 233 |
repo_name = f"{username}/{quantized_model_name}"
|
| 234 |
else:
|
|
|
|
| 239 |
repo_name = f"{username}/{model_name.split('/')[-1]}-ao-{MAP_QUANT_TYPE_TO_NAME[quantization_type.lower()]}-gs{group_size}"
|
| 240 |
else:
|
| 241 |
repo_name = f"{username}/{model_name.split('/')[-1]}-ao-{MAP_QUANT_TYPE_TO_NAME[quantization_type.lower()]}"
|
| 242 |
+
progress(0.70, desc="Creating model card")
|
| 243 |
model_card = create_model_card(model_name, quantization_type, group_size)
|
| 244 |
with open(os.path.join(tmpdirname, "README.md"), "w") as f:
|
| 245 |
f.write(model_card)
|
| 246 |
# Push to Hub
|
| 247 |
api = HfApi(token=auth_token.token)
|
| 248 |
+
api.create_repo(repo_name, exist_ok=True, private=not public)
|
| 249 |
+
progress(0.80, desc="Pushing to Hub")
|
| 250 |
api.upload_folder(
|
| 251 |
folder_path=tmpdirname,
|
| 252 |
repo_id=repo_name,
|
| 253 |
repo_type="model",
|
| 254 |
)
|
| 255 |
+
progress(1.00, desc="Pushing to Hub completed")
|
| 256 |
+
|
| 257 |
import io
|
| 258 |
from contextlib import redirect_stdout
|
| 259 |
import html
|
|
|
|
| 297 |
quantization_type,
|
| 298 |
group_size,
|
| 299 |
quantized_model_name,
|
| 300 |
+
public,
|
| 301 |
):
|
| 302 |
if oauth_token is None:
|
| 303 |
return """
|
|
|
|
| 357 |
profile.username,
|
| 358 |
oauth_token,
|
| 359 |
quantized_model_name,
|
| 360 |
+
public,
|
| 361 |
)
|
| 362 |
except Exception as e:
|
| 363 |
+
# raise e
|
| 364 |
return str(e)
|
| 365 |
|
| 366 |
|
|
|
|
| 491 |
"int4_weight_only",
|
| 492 |
"int8_weight_only",
|
| 493 |
"int8_dynamic_activation_int8_weight",
|
| 494 |
+
"float8_weight_only",
|
| 495 |
+
"float8_dynamic_activation_float8_weight",
|
| 496 |
"autoquant",
|
| 497 |
],
|
| 498 |
value="int8_weight_only",
|
| 499 |
filterable=False,
|
| 500 |
show_label=False,
|
| 501 |
)
|
| 502 |
+
|
| 503 |
group_size = gr.Textbox(
|
| 504 |
info="Group Size (only for int4_weight_only and int8_weight_only)",
|
| 505 |
value="128",
|
| 506 |
+
interactive=(quantization_type.value == "int4_weight_only" or quantization_type.value == "int8_weight_only"),
|
| 507 |
show_label=False,
|
| 508 |
)
|
| 509 |
+
|
| 510 |
+
gr.Markdown(
|
| 511 |
+
"""
|
| 512 |
+
### 💾 Saving Settings
|
| 513 |
+
"""
|
| 514 |
)
|
| 515 |
+
with gr.Row():
|
| 516 |
+
quantized_model_name = gr.Textbox(
|
| 517 |
+
label="✏️ Model Name",
|
| 518 |
+
info="Model Name (optional : to override default)",
|
| 519 |
+
value="",
|
| 520 |
+
interactive=True,
|
| 521 |
+
elem_classes="model-name-textbox",
|
| 522 |
+
show_label=False,
|
| 523 |
+
)
|
| 524 |
+
with gr.Row():
|
| 525 |
+
public = gr.Checkbox(
|
| 526 |
+
label="🌐 Make model public",
|
| 527 |
+
info="If checked, the model will be publicly accessible",
|
| 528 |
+
value=True,
|
| 529 |
+
interactive=True,
|
| 530 |
+
show_label=True,
|
| 531 |
+
)
|
| 532 |
|
| 533 |
with gr.Column():
|
| 534 |
quantize_button = gr.Button(
|
|
|
|
| 564 |
- int8_weight_only typically reduces size by about 50%
|
| 565 |
"""
|
| 566 |
)
|
|
|
|
| 567 |
# Keep existing click handler
|
| 568 |
quantize_button.click(
|
| 569 |
fn=quantize_and_save,
|
| 570 |
+
inputs=[model_name, quantization_type, group_size, quantized_model_name, public],
|
| 571 |
outputs=[output_link],
|
| 572 |
)
|
| 573 |
|