Skip to content

armgit / laya

Public Imported from GitHub
0

Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.

fix(notebooks): clamp fit_one_temp to common TEMP_MIN and TEMP_MAX bounds

Sarthak Pandey1580c398df46
diff --git a/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb b/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb
index 33ee17d8..29ae152d 100644
--- a/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb
+++ b/notebooks/laya_finetune_typed_decisions_2xT4_kaggle.ipynb
@@ -4,7 +4,7 @@
"cell_type": "markdown",
"metadata": {},
"source": [
- "# Fine-Tuning Laya on `LocalLLaMA/typed-decisions` (Kaggle GPU T4 \u00d72 DDP)\n",
+ "# Fine-Tuning Laya on `LocalLLaMA/typed-decisions` (Kaggle GPU T4 ×2 DDP)\n",
"\n",
"[![PyPI version](https://img.shields.io/pypi/v/laya.svg)](https://pypi.org/project/laya/)\n",
"[![Hugging Face Model](https://img.shields.io/badge/%F0%9F%A4%97%20Model-convaiinnovations%2Flaya-blue)](https://huggingface.co/convaiinnovations/laya)\n",
@@ -14,7 +14,7 @@
"\n",
"---\n",
"\n",
- "### \u26a0\ufe0f Kaggle Notebook Settings:\n",
+ "### ⚠️ Kaggle Notebook Settings:\n",
"In the right-hand sidebar panel under **Notebook options**:\n",
"* **Accelerator:** Select **`GPU T4 x2`** (ensure both GPUs are allocated)\n",
"* **Internet:** Set to **`On`**\n",
@@ -171,7 +171,7 @@
"from torch.nn.parallel import DistributedDataParallel as DDP\n",
"from safetensors.torch import load_file, save_file\n",
"from transformers import AutoTokenizer\n",
- "from laya.common import build_model, proper_reward, QTYPES\n",
+ "from laya.common import build_model, proper_reward, QTYPES, TEMP_MIN, TEMP_MAX\n",
"\n",
"def collate_train_batch(items, pad_id):\n",
" n, L = len(items), max(len(it[\"ids\"]) for it in items)\n",
@@ -215,7 +215,7 @@
" loss.backward()\n",
" return loss\n",
" opt.step(closure)\n",
- " return float(torch.clamp(log_t.exp(), 0.1, 10.0).item())\n",
+ " return float(torch.clamp(log_t.exp(), TEMP_MIN, TEMP_MAX).item())\n",
"\n",
"def main():\n",
" dist.init_process_group(\"nccl\")\n",
@@ -431,7 +431,7 @@
" dist.destroy_process_group()\n",
"\n",
"if __name__ == \"__main__\":\n",
- " main()\n"
+ " main()"
],
"execution_count": null,
"outputs": []
diff --git a/tests/test_calibration_persistence.py b/tests/test_calibration_persistence.py
index 0b6292ac..c2187965 100644
--- a/tests/test_calibration_persistence.py
+++ b/tests/test_calibration_persistence.py
@@ -25,7 +25,7 @@ from tokenizers.models import WordLevel # noqa: E402
from transformers import BertConfig, BertModel, PreTrainedTokenizerFast # noqa: E402
from laya import load # noqa: E402
-from laya.common import DecisionModel, QTYPES # noqa: E402
+from laya.common import DecisionModel, QTYPES, TEMP_MIN, TEMP_MAX # noqa: E402
def export_notebook_config(cfg, fitted_temps, output_dir):
@@ -164,6 +164,25 @@ class CalibrationPersistenceTests(unittest.TestCase):
self.write_config()
self.assert_inference_temperatures({(t, 2): 1.0 for t in QTYPES})
+ def test_notebook_fit_one_temp_clamps_to_common_bounds(self):
+ notebook = Path(__file__).resolve().parents[1] / "notebooks" / (
+ "laya_finetune_typed_decisions_2xT4_kaggle.ipynb"
+ )
+ cells = json.loads(notebook.read_text(encoding="utf-8"))["cells"]
+ script, = ["".join(c["source"]) for c in cells
+ if "".join(c["source"]).startswith("%%writefile ")]
+ start = script.index("def fit_one_temp(sel):")
+ end = script.index("\ndef main():", start)
+ fn_code = ast.parse(textwrap.dedent(script[start:end]))
+ scope = {"torch": torch, "TEMP_MIN": TEMP_MIN, "TEMP_MAX": TEMP_MAX}
+ exec(compile(fn_code, str(notebook), "exec"), scope)
+ fit_one_temp = scope["fit_one_temp"]
+
+ high_sel = [([10.0, 0.0], [0.5, 0.5]) for _ in range(20)]
+ t_high = fit_one_temp(high_sel)
+ self.assertLessEqual(t_high, TEMP_MAX)
+ self.assertGreaterEqual(t_high, TEMP_MIN)
+
if __name__ == "__main__":
torch.set_num_threads(1)