Sync experiment for HF Jobs Docker image
Browse files- .hf_space_build_stamp +4 -4
- train.py +21 -2
.hf_space_build_stamp
CHANGED
|
@@ -1,4 +1,4 @@
|
|
| 1 |
-
utc=2026-08-
|
| 2 |
-
github_sha=
|
| 3 |
-
run_id=
|
| 4 |
-
nonce=
|
|
|
|
| 1 |
+
utc=2026-08-03T07:26:13.440492+00:00
|
| 2 |
+
github_sha=21d4354ada3f3263c3e0c64120c98a0c39886bcc
|
| 3 |
+
run_id=30793571239
|
| 4 |
+
nonce=c3950d1e3a51495683e65df6cb2438ac
|
train.py
CHANGED
|
@@ -105,7 +105,7 @@ def run_test_time_training(model, tokenizer, eval_set, fmt_opts, hp):
|
|
| 105 |
The model never sees the answers — only the demo pairs and the input grids.
|
| 106 |
"""
|
| 107 |
from datasets import Dataset
|
| 108 |
-
from unsloth import FastLanguageModel, UnslothTrainer as Trainer
|
| 109 |
from unsloth import UnslothTrainingArguments as TrainingArguments
|
| 110 |
from unsloth import is_bfloat16_supported
|
| 111 |
from model_tools import InputMaskingDataCollator
|
|
@@ -160,7 +160,12 @@ def run_test_time_training(model, tokenizer, eval_set, fmt_opts, hp):
|
|
| 160 |
report_to="none",
|
| 161 |
),
|
| 162 |
)
|
| 163 |
-
unsloth_train(trainer)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 164 |
return model
|
| 165 |
|
| 166 |
|
|
@@ -230,6 +235,20 @@ def main() -> None:
|
|
| 230 |
print(f"Experiment: {os.environ.get('EXPERIMENT_ID', '<unset>')}", flush=True)
|
| 231 |
print(f"Hyperparameters: {json.dumps(hp)}", flush=True)
|
| 232 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 233 |
# Freeze the environment before anything heavy runs, so we capture it even if inference fails.
|
| 234 |
lock_path = capture_lockfile()
|
| 235 |
|
|
|
|
| 105 |
The model never sees the answers — only the demo pairs and the input grids.
|
| 106 |
"""
|
| 107 |
from datasets import Dataset
|
| 108 |
+
from unsloth import FastLanguageModel, UnslothTrainer as Trainer
|
| 109 |
from unsloth import UnslothTrainingArguments as TrainingArguments
|
| 110 |
from unsloth import is_bfloat16_supported
|
| 111 |
from model_tools import InputMaskingDataCollator
|
|
|
|
| 160 |
report_to="none",
|
| 161 |
),
|
| 162 |
)
|
| 163 |
+
# Upstream calls unsloth_train(trainer) here, but that symbol is absent from the
|
| 164 |
+
# unsloth version the authors themselves pin (2024.9.post4, installed from PyPI by their
|
| 165 |
+
# own download notebook) — their run_evaluation script imports a name its pinned unsloth
|
| 166 |
+
# never shipped. UnslothTrainer subclasses trl.SFTTrainer and does not override train(),
|
| 167 |
+
# so trainer.train() is the equivalent (and only available) training path.
|
| 168 |
+
trainer.train()
|
| 169 |
return model
|
| 170 |
|
| 171 |
|
|
|
|
| 235 |
print(f"Experiment: {os.environ.get('EXPERIMENT_ID', '<unset>')}", flush=True)
|
| 236 |
print(f"Hyperparameters: {json.dumps(hp)}", flush=True)
|
| 237 |
|
| 238 |
+
# GPU-side preflight: the CPU Space-build guard can't import unsloth (it probes for a
|
| 239 |
+
# GPU), so unsloth's own API surface is unchecked until runtime. The TTT path imports
|
| 240 |
+
# several unsloth symbols lazily inside run_test_time_training(), and a missing one
|
| 241 |
+
# otherwise only surfaces after the model has already been downloaded and loaded (this is
|
| 242 |
+
# how the unsloth_train import mismatch first burned GPU setup time). Resolve those
|
| 243 |
+
# imports here, first thing, so a bad symbol fails in seconds rather than after minutes.
|
| 244 |
+
if hp["mode"] == "with_ttt":
|
| 245 |
+
import unsloth # GPU-only import; safe here because this is the GPU job
|
| 246 |
+
for _sym in ("FastLanguageModel", "UnslothTrainer", "UnslothTrainingArguments",
|
| 247 |
+
"is_bfloat16_supported"):
|
| 248 |
+
if not hasattr(unsloth, _sym):
|
| 249 |
+
raise SystemExit(f"TTT preflight: unsloth is missing '{_sym}'")
|
| 250 |
+
print("TTT preflight: unsloth trainer symbols present.", flush=True)
|
| 251 |
+
|
| 252 |
# Freeze the environment before anything heavy runs, so we capture it even if inference fails.
|
| 253 |
lock_path = capture_lockfile()
|
| 254 |
|