Upload modal_notebook_options.md with huggingface_hub

modal_notebook_options.md

@@ -93,15 +93,31 @@ print("pushed to https://huggingface.co/CodeIsAbstract/sanskrit-sandhi-boundary-
 
 ## Cell 5 (optional) — quick sanity check on a phrase
 
+`decode_best.py` needs the full local Sanskrit stack (sanskrit_parser, vidyut,
+SanskritTokenizer) that only exists on your Mac — don't import it here.
+This lightweight check only needs what Cell 2 already downloaded:
+
 ```python
 import os, torch
-from huggingface_hub import hf_hub_download
-hf_hub_download("CodeIsAbstract/sanskrit-sandhi-boundaries-v2", "decode_best.py",
-                repo_type="dataset", local_dir=".")
-from decode_best import load_model, predict
-
-model, delay = load_model("checkpoints/delayed_model_save_fixed_delay2", torch.device("cuda"))
-print(predict(model, "satyamevajayatenAnftaM", delay=delay))
+from model import HybridTimeScaleConfig
+from train_split_model import BinaryBoundaryModel, char_to_idx, PAD_ID
+from safetensors.torch import load_file
+
+ckpt = "checkpoints/delayed_model_save_fixed_delay2"
+cfg = HybridTimeScaleConfig.from_pretrained(ckpt)   # config.json written by the trainer
+model = BinaryBoundaryModel(cfg)
+model.load_state_dict(load_file(os.path.join(ckpt, "model.safetensors")), strict=False)
+model.eval().to(torch.device("cuda"))
+
+text = "satyamevajayatenAnftaM"
+ids = torch.tensor([[char_to_idx.get(c, PAD_ID) for c in text] + [PAD_ID]*2],
+                   dtype=torch.long, device="cuda")
+mask = torch.ones(1, len(text)+2, dtype=torch.long, device="cuda")
+with torch.no_grad():
+    p = torch.sigmoid(model(input_ids=ids, attention_mask=mask)["logits"].squeeze(0)).cpu()
+
+print("boundary probs:", [f"{p[i]:.2f}" for i in range(len(text))])
+print("predicted split:", "".join(c + (" " if p[i] > 0.5 else "") for i, c in enumerate(text)))
 ```
 
 ---