-
Notifications
You must be signed in to change notification settings - Fork 2
Expand file tree
/
Copy pathinterface.py
More file actions
1983 lines (1805 loc) · 80.7 KB
/
interface.py
File metadata and controls
1983 lines (1805 loc) · 80.7 KB
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
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
#!/usr/bin/env python3
"""
Gradio Interface for SmolLM3/GPT-OSS Fine-tuning Pipeline
This app mirrors the core flow of launch.sh with a click-and-run UI.
Tokens are read from environment variables:
- HF_WRITE_TOKEN (required)
- HF_READ_TOKEN (optional; used to switch the Trackio Space token after training)
Key steps (configurable via UI):
1) Optional HF Dataset repo setup for Trackio
2) Optional Trackio Space deployment
3) Training (SmolLM3 or GPT-OSS)
4) Push trained model to the HF Hub
5) Optional switch Trackio HF_TOKEN to read token
This uses the existing scripts in scripts/ and config/ to avoid code duplication.
"""
from __future__ import annotations
import os
import sys
import time
import json
import shlex
import traceback
import importlib.util
import re
from dataclasses import dataclass
from datetime import datetime
from pathlib import Path
from typing import Dict, Any, Generator, Optional, Tuple
# Third-party
try:
import gradio as gr # type: ignore
except Exception as _e:
raise RuntimeError(
"Gradio is required. Please install it first: pip install gradio"
) from _e
# --------------------------------------------------------------------------------------
# Utilities
# --------------------------------------------------------------------------------------
PROJECT_ROOT = Path(__file__).resolve().parent
def mask_token(token: Optional[str]) -> str:
if not token:
return "<not set>"
token = str(token)
if len(token) <= 8:
return "*" * len(token)
return f"{token[:4]}****{token[-4:]}"
def get_python() -> str:
return sys.executable or "python"
def get_username_from_token(token: str) -> Optional[str]:
try:
from huggingface_hub import HfApi # type: ignore
api = HfApi(token=token)
info = api.whoami()
if isinstance(info, dict):
return info.get("name") or info.get("username")
if isinstance(info, str):
return info
except Exception:
return None
return None
def detect_nvidia_driver() -> Tuple[bool, str]:
"""Detect NVIDIA driver/GPU presence with multiple strategies.
Returns (available, human_message).
"""
# 1) Try torch CUDA
try:
import torch # type: ignore
if torch.cuda.is_available():
try:
num = torch.cuda.device_count()
names = [torch.cuda.get_device_name(i) for i in range(num)]
return True, f"NVIDIA GPU detected: {', '.join(names)}"
except Exception:
return True, "NVIDIA GPU detected (torch.cuda available)"
except Exception:
pass
# 2) Try NVML via pynvml
try:
import pynvml # type: ignore
try:
pynvml.nvmlInit()
cnt = pynvml.nvmlDeviceGetCount()
names = []
for i in range(cnt):
h = pynvml.nvmlDeviceGetHandleByIndex(i)
names.append(pynvml.nvmlDeviceGetName(h).decode("utf-8", errors="ignore"))
drv = pynvml.nvmlSystemGetDriverVersion().decode("utf-8", errors="ignore")
pynvml.nvmlShutdown()
if cnt > 0:
return True, f"NVIDIA driver {drv}; GPUs: {', '.join(names)}"
except Exception:
pass
except Exception:
pass
# 3) Try nvidia-smi
try:
import subprocess
res = subprocess.run(["nvidia-smi", "-L"], capture_output=True, text=True, timeout=3)
if res.returncode == 0 and res.stdout.strip():
return True, res.stdout.strip().splitlines()[0]
except Exception:
pass
return False, "No NVIDIA driver/GPU detected"
def duplicate_space_hint() -> str:
space_id = os.environ.get("SPACE_ID") or os.environ.get("HF_SPACE_ID")
if space_id:
space_url = f"https://huggingface.co/spaces/{space_id}"
dup_url = f"{space_url}?duplicate=true"
return (
f"ℹ️ No NVIDIA driver detected. If you're on Hugging Face Spaces, "
f"please duplicate this Space to GPU hardware: [Duplicate this Space]({dup_url})."
)
return (
"ℹ️ No NVIDIA driver detected. To enable training, run on a machine with an NVIDIA GPU/driver "
"or duplicate this Space on Hugging Face with GPU hardware."
)
def markdown_links_to_html(text: str) -> str:
"""Convert simple Markdown links [text](url) to HTML anchors for UI rendering."""
try:
return re.sub(r"\[([^\]]+)\]\(([^)]+)\)", r'<a href="\2" target="_blank" rel="noopener noreferrer">\1</a>', text)
except Exception:
return text
def _write_generated_config(filename: str, content: str) -> Path:
"""Write a generated config under config/ and return the full path."""
cfg_dir = PROJECT_ROOT / "config"
cfg_dir.mkdir(parents=True, exist_ok=True)
path = cfg_dir / filename
with open(path, "w", encoding="utf-8") as f:
f.write(content)
return path
def generate_medical_o1_config_file(
dataset_config: str,
system_message: Optional[str],
developer_message: Optional[str],
num_train_epochs: float,
batch_size: int,
gradient_accumulation_steps: int,
learning_rate: float,
max_seq_length: int,
) -> Path:
"""Create a GPT-OSS Medical o1 SFT config file from user inputs."""
# Sanitize quotes in messages
def _q(s: Optional[str]) -> str:
if s is None or s == "":
return "None"
return repr(s)
py = f"""
from config.train_gpt_oss_custom import GPTOSSEnhancedCustomConfig
config = GPTOSSEnhancedCustomConfig(
dataset_name="FreedomIntelligence/medical-o1-reasoning-SFT",
dataset_config={repr(dataset_config)},
dataset_split="train",
dataset_format="medical_o1_sft",
# Field mapping and prefixes
input_field="Question",
target_field="Response",
question_field="Question",
reasoning_field="Complex_CoT",
response_field="Response",
reason_prefix="Reasoning: ",
answer_prefix="Final Answer: ",
# Optional context
system_message={_q(system_message)},
developer_message={_q(developer_message)},
# Training hyperparameters
num_train_epochs={num_train_epochs},
batch_size={batch_size},
gradient_accumulation_steps={gradient_accumulation_steps},
learning_rate={learning_rate},
min_lr=2e-5,
weight_decay=0.01,
warmup_ratio=0.03,
# Sequence length
max_seq_length={max_seq_length},
# Precision & performance
fp16=False,
bf16=True,
dataloader_num_workers=4,
dataloader_pin_memory=True,
dataloader_prefetch_factor=2,
group_by_length=True,
remove_unused_columns=True,
# LoRA & quantization
use_lora=True,
lora_config={
"r": 16,
"lora_alpha": 32,
"lora_dropout": 0.05,
"target_modules": "all-linear",
"target_parameters": [
"7.mlp.experts.gate_up_proj",
"7.mlp.experts.down_proj",
"15.mlp.experts.gate_up_proj",
"15.mlp.experts.down_proj",
"23.mlp.experts.gate_up_proj",
"23.mlp.experts.down_proj",
],
"bias": "none",
"task_type": "CAUSAL_LM",
},
use_quantization=True,
quantization_config={
"dequantize": True,
"load_in_4bit": False,
},
# Logging & evaluation
eval_strategy="steps",
eval_steps=100,
logging_steps=10,
save_strategy="steps",
save_steps=500,
save_total_limit=3,
metric_for_best_model="eval_loss",
greater_is_better=False,
)
"""
return _write_generated_config("_generated_gpt_oss_medical_o1_sft.py", py)
def generate_gpt_oss_custom_config_file(
dataset_name: str,
dataset_split: str,
dataset_format: str,
input_field: str,
target_field: Optional[str],
system_message: Optional[str],
developer_message: Optional[str],
model_identity: Optional[str],
max_samples: Optional[int],
min_length: int,
max_length: Optional[int],
num_train_epochs: float,
batch_size: int,
gradient_accumulation_steps: int,
learning_rate: float,
min_lr: float,
weight_decay: float,
warmup_ratio: float,
max_seq_length: int,
lora_r: int,
lora_alpha: int,
lora_dropout: float,
mixed_precision: str, # "bf16"|"fp16"|"fp32"
num_workers: int,
quantization_type: str, # "mxfp4"|"bnb4"|"none"
max_grad_norm: float,
logging_steps: int,
eval_steps: int,
save_steps: int,
) -> Path:
# Precision flags
if mixed_precision.lower() == "bf16":
fp16_flag = False
bf16_flag = True
elif mixed_precision.lower() == "fp16":
fp16_flag = True
bf16_flag = False
else:
fp16_flag = False
bf16_flag = False
# Quantization flags/config
if quantization_type == "mxfp4":
use_quant = True
quant_cfg = '{"dequantize": True, "load_in_4bit": False}'
elif quantization_type == "bnb4":
use_quant = True
quant_cfg = '{"dequantize": False, "load_in_4bit": True, "bnb_4bit_compute_dtype": "bfloat16", "bnb_4bit_use_double_quant": True, "bnb_4bit_quant_type": "nf4"}'
else:
use_quant = False
quant_cfg = '{"dequantize": False, "load_in_4bit": False}'
def _q(s: Optional[str]) -> str:
if s is None or s == "":
return "None"
return repr(s)
py = f"""
from config.train_gpt_oss_custom import GPTOSSEnhancedCustomConfig
config = GPTOSSEnhancedCustomConfig(
# Dataset
dataset_name={repr(dataset_name)},
dataset_split={repr(dataset_split)},
dataset_format={repr(dataset_format)},
input_field={repr(input_field)},
target_field={repr(target_field)} if {repr(target_field)} != 'None' else None,
system_message={_q(system_message)},
developer_message={_q(developer_message)},
max_samples={repr(max_samples)} if {repr(max_samples)} != 'None' else None,
min_length={min_length},
max_length={repr(max_length)} if {repr(max_length)} != 'None' else None,
# Training hyperparameters
num_train_epochs={num_train_epochs},
batch_size={batch_size},
gradient_accumulation_steps={gradient_accumulation_steps},
learning_rate={learning_rate},
min_lr={min_lr},
weight_decay={weight_decay},
warmup_ratio={warmup_ratio},
max_grad_norm={max_grad_norm},
# Model
max_seq_length={max_seq_length},
# Precision
fp16={str(fp16_flag)},
bf16={str(bf16_flag)},
# LoRA
lora_config={{
"r": {lora_r},
"lora_alpha": {lora_alpha},
"lora_dropout": {lora_dropout},
"target_modules": "all-linear",
"bias": "none",
"task_type": "CAUSAL_LM",
}},
# Quantization
use_quantization={str(use_quant)},
quantization_config={quant_cfg},
# Performance
dataloader_num_workers={num_workers},
dataloader_pin_memory=True,
group_by_length=True,
# Logging & eval
logging_steps={logging_steps},
eval_steps={eval_steps},
save_steps={save_steps},
# Chat template (Harmony)
chat_template_kwargs={{
"add_generation_prompt": True,
"tokenize": False,
"auto_insert_role": True,
"reasoning_effort": "medium",
"model_identity": {_q(model_identity) if _q(model_identity) != 'None' else repr('You are GPT-Tonic, a large language model trained by TonicAI.')},
"builtin_tools": [],
}},
)
"""
return _write_generated_config("_generated_gpt_oss_custom.py", py)
def generate_smollm3_custom_config_file(
model_name: str,
dataset_name: Optional[str],
max_seq_length: int,
batch_size: int,
gradient_accumulation_steps: int,
learning_rate: float,
save_steps: int,
eval_steps: int,
logging_steps: int,
filter_bad_entries: bool,
input_field: str,
target_field: str,
sample_size: Optional[int],
sample_seed: int,
trainer_type: str,
) -> Path:
# Create subclass to include dataset fields similar to other configs
def _bool(b: bool) -> str:
return "True" if b else "False"
ds_section = """
# HF Dataset configuration
dataset_name={}
dataset_split="train"
input_field={}
target_field={}
filter_bad_entries={}
bad_entry_field="bad_entry"
sample_size={}
sample_seed={}
""".format(
repr(dataset_name) if dataset_name else "None",
repr(input_field),
repr(target_field),
_bool(filter_bad_entries),
repr(sample_size) if sample_size is not None else "None",
sample_seed,
)
py = f"""
from dataclasses import dataclass
from typing import Optional
from config.train_smollm3 import SmolLM3Config
@dataclass
class SmolLM3GeneratedConfig(SmolLM3Config):
{ds_section}
config = SmolLM3GeneratedConfig(
trainer_type={repr(trainer_type.lower())},
model_name={repr(model_name)},
max_seq_length={max_seq_length},
use_flash_attention=True,
use_gradient_checkpointing=True,
batch_size={batch_size},
gradient_accumulation_steps={gradient_accumulation_steps},
learning_rate={learning_rate},
weight_decay=0.01,
warmup_steps=100,
max_iters=None,
eval_interval={eval_steps},
log_interval={logging_steps},
save_interval={save_steps},
optimizer="adamw",
beta1=0.9,
beta2=0.95,
eps=1e-8,
scheduler="cosine",
min_lr=1e-6,
fp16=True,
bf16=False,
save_steps={save_steps},
eval_steps={eval_steps},
logging_steps={logging_steps},
save_total_limit=3,
eval_strategy="steps",
metric_for_best_model="eval_loss",
greater_is_better=False,
load_best_model_at_end=True,
)
"""
return _write_generated_config("_generated_smollm3_custom.py", py)
def generate_smollm3_long_context_config_file(
model_name: str,
dataset_name: Optional[str],
input_field: str,
target_field: str,
filter_bad_entries: bool,
sample_size: Optional[int],
sample_seed: int,
max_seq_length: int,
batch_size: int,
gradient_accumulation_steps: int,
learning_rate: float,
warmup_steps: int,
max_iters: int,
save_steps: int,
eval_steps: int,
logging_steps: int,
use_chat_template: bool,
no_think_system_message: bool,
trainer_type: str,
) -> Path:
"""Create a SmolLM3 long-context config file with optional dataset fields."""
def _bool(b: bool) -> str:
return "True" if b else "False"
ds_section = """
# HF Dataset configuration
dataset_name={}
dataset_split="train"
input_field={}
target_field={}
filter_bad_entries={}
bad_entry_field="bad_entry"
sample_size={}
sample_seed={}
""".format(
repr(dataset_name) if dataset_name else "None",
repr(input_field),
repr(target_field),
_bool(filter_bad_entries),
repr(sample_size) if sample_size is not None else "None",
sample_seed,
)
py = f"""
from dataclasses import dataclass
from typing import Optional
from config.train_smollm3 import SmolLM3Config
@dataclass
class SmolLM3LongContextGeneratedConfig(SmolLM3Config):
{ds_section}
config = SmolLM3LongContextGeneratedConfig(
trainer_type={repr(trainer_type.lower())},
model_name={repr(model_name)},
max_seq_length={max_seq_length},
use_flash_attention=True,
use_gradient_checkpointing=True,
batch_size={batch_size},
gradient_accumulation_steps={gradient_accumulation_steps},
learning_rate={learning_rate},
weight_decay=0.01,
warmup_steps={warmup_steps},
max_iters={max_iters},
fp16=True,
bf16=False,
save_steps={save_steps},
eval_steps={eval_steps},
logging_steps={logging_steps},
save_total_limit=3,
eval_strategy="steps",
metric_for_best_model="eval_loss",
greater_is_better=False,
load_best_model_at_end=True,
use_chat_template={_bool(use_chat_template)},
chat_template_kwargs={{
"add_generation_prompt": True,
"no_think_system_message": {_bool(no_think_system_message)}
}}
)
"""
return _write_generated_config("_generated_smollm3_long_context.py", py)
def ensure_dataset_repo(username: str, dataset_name: str, token: str) -> Tuple[str, bool, str]:
"""Create or ensure dataset repo exists. Returns (repo_id, created_or_exists, message)."""
from huggingface_hub import create_repo # type: ignore
repo_id = f"{username}/{dataset_name}"
try:
create_repo(repo_id=repo_id, repo_type="dataset", token=token, exist_ok=True, private=False)
return repo_id, True, f"Dataset repo ready: {repo_id}"
except Exception as e:
return repo_id, False, f"Failed to create dataset repo {repo_id}: {e}"
def import_config_object(config_path: Path) -> Optional[Any]:
"""Import a config file and return its 'config' object if present, else None."""
try:
spec = importlib.util.spec_from_file_location("config_module", str(config_path))
if not spec or not spec.loader:
return None
module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(module) # type: ignore
if hasattr(module, "config"):
return getattr(module, "config")
return None
except Exception:
return None
def run_command_stream(args: list[str], env: Dict[str, str], cwd: Optional[Path] = None) -> Generator[str, None, int]:
"""Run a command and yield stdout/stderr lines as they arrive. Returns exit code at the end."""
import subprocess
yield f"$ {' '.join(shlex.quote(a) for a in ([get_python()] + args))}"
process = subprocess.Popen(
[get_python()] + args,
stdout=subprocess.PIPE,
stderr=subprocess.STDOUT,
text=True,
env=env,
cwd=str(cwd or PROJECT_ROOT),
bufsize=1,
universal_newlines=True,
)
assert process.stdout is not None
for line in iter(process.stdout.readline, ""):
yield line.rstrip()
process.stdout.close()
code = process.wait()
yield f"[exit_code={code}]"
return code
# --------------------------------------------------------------------------------------
# Configuration Mappings (mirror launch.sh)
# --------------------------------------------------------------------------------------
SMOL_CONFIGS = {
"Basic Training": {
"config_file": "config/train_smollm3.py",
"default_model": "HuggingFaceTB/SmolLM3-3B",
},
"H100 Lightweight (Rapid)": {
"config_file": "config/train_smollm3_h100_lightweight.py",
"default_model": "HuggingFaceTB/SmolLM3-3B",
},
"A100 Large Scale": {
"config_file": "config/train_smollm3_openhermes_fr_a100_large.py",
"default_model": "HuggingFaceTB/SmolLM3-3B",
},
"Multiple Passes": {
"config_file": "config/train_smollm3_openhermes_fr_a100_multiple_passes.py",
"default_model": "HuggingFaceTB/SmolLM3-3B",
},
}
GPT_OSS_CONFIGS = {
"GPT-OSS Basic Training": {
"config_file": "config/train_gpt_oss_basic.py",
"default_model": "openai/gpt-oss-20b",
},
"GPT-OSS H100 Optimized": {
"config_file": "config/train_gpt_oss_h100_optimized.py",
"default_model": "openai/gpt-oss-20b",
},
"GPT-OSS Multilingual Reasoning": {
"config_file": "config/train_gpt_oss_multilingual_reasoning.py",
"default_model": "openai/gpt-oss-20b",
},
"GPT-OSS Memory Optimized": {
"config_file": "config/train_gpt_oss_memory_optimized.py",
"default_model": "openai/gpt-oss-20b",
},
"GPT-OSS OpenHermes-FR (Recommended)": {
"config_file": "config/train_gpt_oss_openhermes_fr.py",
"default_model": "openai/gpt-oss-20b",
},
"GPT-OSS OpenHermes-FR Memory Optimized": {
"config_file": "config/train_gpt_oss_openhermes_fr_memory_optimized.py",
"default_model": "openai/gpt-oss-20b",
},
# Custom dataset and medical SFT can be added later as advanced UI panels
}
def get_config_map(family: str) -> Dict[str, Dict[str, str]]:
return SMOL_CONFIGS if family == "SmolLM3" else GPT_OSS_CONFIGS
# --------------------------------------------------------------------------------------
# Pipeline Orchestration
# --------------------------------------------------------------------------------------
@dataclass
class PipelineInputs:
model_family: str
config_choice: str
trainer_type: str # "SFT" | "DPO"
monitoring_mode: str # "both" | "trackio" | "dataset" | "none"
experiment_name: str
repo_short: str
author_name: str
model_description: str
trackio_space_name: Optional[str]
deploy_trackio_space: bool
create_dataset_repo: bool
push_to_hub: bool
switch_to_read_after: bool
scheduler_override: Optional[str]
min_lr: Optional[float]
min_lr_rate: Optional[float]
# Optional override config path generated from Advanced tab
override_config_path: Optional[str] = None
def make_defaults(model_family: str) -> Tuple[str, str]:
ts = datetime.now().strftime("%Y%m%d_%H%M%S")
family_slug = "gpt-oss" if model_family == "GPT-OSS" else "smollm3"
exp = f"smolfactory-{family_slug}_{ts}"
repo_short = f"smolfactory-{datetime.now().strftime('%Y%m%d')}"
return exp, repo_short
def run_pipeline(params: PipelineInputs) -> Generator[str, None, None]:
# Tokens from environment
write_token = os.environ.get("HF_WRITE_TOKEN") or os.environ.get("HF_TOKEN")
read_token = os.environ.get("HF_READ_TOKEN")
if not write_token:
yield "❌ HF_WRITE_TOKEN (or HF_TOKEN) is not set in the environment."
return
# Resolve username
username = get_username_from_token(write_token) or os.environ.get("HF_USERNAME")
if not username:
yield "❌ Could not resolve Hugging Face username from token."
return
yield f"✅ Authenticated as: {username}"
# Compute Trackio URL if applicable
trackio_url: Optional[str] = None
if params.monitoring_mode != "none" and params.trackio_space_name:
trackio_url = f"https://huggingface.co/spaces/{username}/{params.trackio_space_name}"
yield f"Trackio Space URL: {trackio_url}"
# Decide space deploy token per monitoring mode
space_deploy_token = write_token if params.monitoring_mode in ("both", "trackio") else (read_token or write_token)
# Dataset repo setup
dataset_repo = f"{username}/trackio-experiments"
if params.create_dataset_repo and params.monitoring_mode != "none":
yield f"Creating/ensuring dataset repo exists: {dataset_repo}"
rid, ok, msg = ensure_dataset_repo(username, "trackio-experiments", write_token)
yield ("✅ " if ok else "⚠️ ") + msg
dataset_repo = rid
# Resolve config file and model name (allow override from Advanced tab)
conf_map = get_config_map(params.model_family)
if params.override_config_path:
config_file = Path(params.override_config_path)
if not config_file.exists():
yield f"❌ Generated config file not found: {config_file}"
return
# Best-effort to infer base model from generated config
cfg_obj = import_config_object(config_file)
base_model_fallback = getattr(cfg_obj, "model_name", None) or (
conf_map.get(params.config_choice, {}).get("default_model", "")
)
else:
if params.config_choice not in conf_map:
yield f"❌ Unknown config choice: {params.config_choice}"
return
config_file = PROJECT_ROOT / conf_map[params.config_choice]["config_file"]
base_model_fallback = conf_map[params.config_choice]["default_model"]
if not config_file.exists():
yield f"❌ Config file not found: {config_file}"
return
cfg_obj = import_config_object(config_file)
base_model = getattr(cfg_obj, "model_name", base_model_fallback) if cfg_obj else base_model_fallback
dataset_name = getattr(cfg_obj, "dataset_name", None) if cfg_obj else None
batch_size = getattr(cfg_obj, "batch_size", None) if cfg_obj else None
learning_rate = getattr(cfg_obj, "learning_rate", None) if cfg_obj else None
max_seq_length = getattr(cfg_obj, "max_seq_length", None) if cfg_obj else None
# Prepare env for subprocesses
env = os.environ.copy()
env["HF_TOKEN"] = write_token
env["HUGGING_FACE_HUB_TOKEN"] = write_token
env["HF_USERNAME"] = username
env["TRACKIO_DATASET_REPO"] = dataset_repo
env["MONITORING_MODE"] = params.monitoring_mode
# Optional Trackio Space deployment
if params.deploy_trackio_space and params.monitoring_mode != "none" and params.trackio_space_name:
yield f"\n=== Deploying Trackio Space: {params.trackio_space_name} ==="
# deploy_trackio_space.py expects: space_name, token, git_email, git_name, dataset_repo
args = [
str(PROJECT_ROOT / "scripts/trackio_tonic/deploy_trackio_space.py"),
params.trackio_space_name,
space_deploy_token,
f"{username}@users.noreply.hf.co",
username,
dataset_repo,
]
for line in run_command_stream(args, env, cwd=PROJECT_ROOT / "scripts/trackio_tonic"):
yield line
# Dataset setup and Trackio configuration (mirror launch.sh) when monitoring is enabled
if params.monitoring_mode != "none":
# Ensure HF Dataset structure
yield f"\n=== Setting up HF Dataset: {dataset_repo} ==="
ds_args = [
str(PROJECT_ROOT / "scripts/dataset_tonic/setup_hf_dataset.py"),
write_token,
]
for line in run_command_stream(ds_args, env, cwd=PROJECT_ROOT / "scripts/dataset_tonic"):
yield line
# Configure Trackio Space
yield f"\n=== Configuring Trackio Space ({params.trackio_space_name or 'N/A'}) ==="
conf_args = [str(PROJECT_ROOT / "scripts/trackio_tonic/configure_trackio.py")]
# Use space deploy token (READ for dataset-only; WRITE otherwise)
conf_env = env.copy()
conf_env["HF_TOKEN"] = space_deploy_token
conf_env["HUGGING_FACE_HUB_TOKEN"] = space_deploy_token
for line in run_command_stream(conf_args, conf_env, cwd=PROJECT_ROOT / "scripts/trackio_tonic"):
yield line
# Training output directory
out_dir = PROJECT_ROOT / "outputs" / f"{params.experiment_name}_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
out_dir.mkdir(parents=True, exist_ok=True)
yield f"\nOutput directory: {out_dir}"
# Scheduler overrides (GPT-OSS only)
if params.model_family == "GPT-OSS" and params.scheduler_override:
env["GPT_OSS_SCHEDULER"] = params.scheduler_override
if params.min_lr is not None:
env["GPT_OSS_MIN_LR"] = str(params.min_lr)
if params.min_lr_rate is not None:
env["GPT_OSS_MIN_LR_RATE"] = str(params.min_lr_rate)
# Start training
yield f"\n=== Starting Training ({params.model_family}) ==="
if params.model_family == "GPT-OSS":
args = [
str(PROJECT_ROOT / "scripts/training/train_gpt_oss.py"),
"--config", str(config_file),
"--experiment-name", params.experiment_name,
"--output-dir", str(out_dir),
"--trackio-url", trackio_url or "",
"--trainer-type", params.trainer_type.lower(),
]
else:
args = [
str(PROJECT_ROOT / "scripts/training/train.py"),
"--config", str(config_file),
"--experiment-name", params.experiment_name,
"--output-dir", str(out_dir),
"--trackio-url", trackio_url or "",
"--trainer-type", params.trainer_type.lower(),
]
# Stream training logs
train_failed = False
for line in run_command_stream(args, env):
yield line
if line.strip().startswith("[exit_code=") and not line.strip().endswith("[exit_code=0]"):
train_failed = True
if train_failed:
yield "❌ Training failed. Aborting remaining steps."
return
# Push to Hub
if params.push_to_hub:
yield "\n=== Pushing Model to Hugging Face Hub ==="
repo_name = f"{username}/{params.repo_short}"
if params.model_family == "GPT-OSS":
push_args = [
str(PROJECT_ROOT / "scripts/model_tonic/push_gpt_oss_to_huggingface.py"),
str(out_dir),
repo_name,
"--token", write_token,
"--trackio-url", trackio_url or "",
"--experiment-name", params.experiment_name,
"--dataset-repo", dataset_repo,
"--author-name", params.author_name or username,
"--model-description", params.model_description,
"--training-config-type", params.config_choice,
"--model-name", base_model,
]
if dataset_name:
push_args += ["--dataset-name", str(dataset_name)]
if batch_size is not None:
push_args += ["--batch-size", str(batch_size)]
if learning_rate is not None:
push_args += ["--learning-rate", str(learning_rate)]
if max_seq_length is not None:
push_args += ["--max-seq-length", str(max_seq_length)]
push_args += ["--trainer-type", params.trainer_type]
else:
push_args = [
str(PROJECT_ROOT / "scripts/model_tonic/push_to_huggingface.py"),
str(out_dir),
repo_name,
"--token", write_token,
"--trackio-url", trackio_url or "",
"--experiment-name", params.experiment_name,
"--dataset-repo", dataset_repo,
"--author-name", params.author_name or username,
"--model-description", params.model_description,
"--training-config-type", params.config_choice,
"--model-name", base_model,
]
if dataset_name:
push_args += ["--dataset-name", str(dataset_name)]
if batch_size is not None:
push_args += ["--batch-size", str(batch_size)]
if learning_rate is not None:
push_args += ["--learning-rate", str(learning_rate)]
if max_seq_length is not None:
push_args += ["--max-seq-length", str(max_seq_length)]
push_args += ["--trainer-type", params.trainer_type]
for line in run_command_stream(push_args, env):
yield line
# Switch Space token to read-only (security)
if params.switch_to_read_after and params.monitoring_mode in ("both", "trackio") and params.trackio_space_name and read_token:
yield "\n=== Switching Trackio Space HF_TOKEN to READ token ==="
space_id = f"{username}/{params.trackio_space_name}"
sw_args = [
str(PROJECT_ROOT / "scripts/trackio_tonic/switch_to_read_token.py"),
space_id,
read_token,
write_token,
]
for line in run_command_stream(sw_args, env, cwd=PROJECT_ROOT / "scripts/trackio_tonic"):
yield line
elif params.switch_to_read_after and not read_token:
yield "⚠️ HF_READ_TOKEN not set; skipping token switch."
# Final summary
yield "\n🎉 Pipeline completed."
if params.monitoring_mode != "none" and trackio_url:
yield f"Trackio: {trackio_url}"
yield f"Model repo (if pushed): https://huggingface.co/{username}/{params.repo_short}"
yield f"Outputs: {out_dir}"
# --------------------------------------------------------------------------------------
# Gradio UI
# --------------------------------------------------------------------------------------
MODEL_FAMILIES = ["SmolLM3", "GPT-OSS"]
TRAINER_CHOICES = ["SFT", "DPO"]
MONITORING_CHOICES = ["both", "trackio", "dataset", "none"]
SCHEDULER_CHOICES = [None, "linear", "cosine", "cosine_with_min_lr", "constant"]
def ui_defaults(family: str) -> Tuple[str, str, str, str]:
exp, repo_short = make_defaults(family)
default_desc = (
"A fine-tuned GPT-OSS-20B model optimized for multilingual reasoning and instruction following."
if family == "GPT-OSS"
else "A fine-tuned SmolLM3-3B model optimized for instruction following and French language tasks."
)
trackio_space_name = f"trackio-monitoring-{datetime.now().strftime('%Y%m%d')}"
return exp, repo_short, default_desc, trackio_space_name
title_md = """
# 🙋🏻♂️ Welcome to 🌟Tonic's 🤏🏻🏭 SmolFactory !
"""
howto_md = """
### How to use
To get started: duplicate the space, select a model family and a configuration, click run.
"""
joinus_md = """
### Join us :
🌟TeamTonic🌟 is always making cool demos! Join our active builder's 🛠️community 👻 [](https://discord.gg/qdfnvSPcqP) On 🤗Huggingface:[MultiTransformer](https://huggingface.co/MultiTransformer) On 🌐Github: [Tonic-AI](https://github.com/tonic-ai) & contribute to🌟 [Build Tonic](https://git.tonic-ai.com/contribute)🤗Big thanks to Yuvi Sharma and all the folks at huggingface for the community grant 🤗
"""
# Load inline SVG to render before the Join Us section
try:
_OUTPUT_SVG_HTML = (PROJECT_ROOT / "docs" / "output.svg").read_text(encoding="utf-8")
except Exception:
_OUTPUT_SVG_HTML = ""
def on_family_change(family: str):
"""Update UI when the model family changes.
- Refresh available prebuilt configuration choices
- Reset defaults (experiment name, repo short, description, space name)
- Reveal the next step (trainer type)
"""
confs = list(get_config_map(family).keys())
exp, repo_short, desc, space = ui_defaults(family)
# Initial dataset information placeholder until a specific config is chosen
training_md = (
f"Select a training configuration for {family} to see details (dataset, batch size, etc.)."
)
# Update objects:
return (
gr.update(choices=confs, value=(confs[0] if confs else None)),
exp,
repo_short,
desc,
space,
training_md,
gr.update(choices=[], value=None),
gr.update(visible=True), # show step 2 (trainer)
gr.update(visible=True), # show step 3 immediately (default monitoring 'dataset')
gr.update(visible=True), # show step 4 immediately so users see configs
gr.update(visible=False), # GPT-OSS advanced group hidden until enabled
gr.update(visible=False), # SmolLM3 advanced group hidden until enabled
)
def on_config_change(family: str, config_choice: str):