upgraded lilylet version.

app.py

@@ -36,6 +36,7 @@ HERE = os.path.dirname(os.path.abspath(__file__))
 HF_MODEL_REPO = os.environ.get('LILYSCRIPT_MODEL_REPO', 'k-l-lambda/LilyNota')
 HF_MODEL_SUBDIR = 'onnx'		# weights + geometry + tokenizer live here in the repo
 MODEL_DIR = os.environ.get('LILYSCRIPT_MODEL_DIR')		# set -> use this local dir instead of the hub
+LOCAL_MODEL_DIR = os.path.join(HERE, 'models')		# repo-local onnx bundle; preferred over the hub when present
 ASSET_DIR = os.path.join(HERE, 'assets')
 EXAMPLES_DIR = os.path.join(HERE, 'examples')
 OUTPUT_DIR = os.path.join(HERE, 'outputs')
@@ -107,12 +108,19 @@ _init_logging()
 
 
 def resolve_model_dir ():
-	'''Where the ONNX weights live. If LILYSCRIPT_MODEL_DIR is set, use it as-is
-	(local dev). Otherwise pull the `onnx/` bundle from the HF model repo and
-	return its local snapshot path. The tokenizer is NOT pulled — it's read from
-	the app's own assets/ dir — so we only fetch the weight files.'''
+	'''Where the ONNX weights live, in priority order:
+	  1. LILYSCRIPT_MODEL_DIR (explicit override, local dev),
+	  2. the repo-local `models/` dir IF it holds the full weight bundle,
+	  3. otherwise pull the `onnx/` bundle from the HF model repo.
+	The tokenizer is NOT pulled — it's read from the app's own assets/ dir — so we
+	only fetch the weight files.'''
 	if MODEL_DIR:
 		return MODEL_DIR
+	required = ('geometry.json', 'patch_kv_int8.onnx', 'token_kv_int8.onnx', 'wte.npy')
+	if os.path.isdir(LOCAL_MODEL_DIR) and all(
+			os.path.isfile(os.path.join(LOCAL_MODEL_DIR, name)) for name in required):
+		LOG.info('using local model weights in %s', LOCAL_MODEL_DIR)
+		return LOCAL_MODEL_DIR
 	from huggingface_hub import snapshot_download
 	LOG.info('downloading model weights from hf:%s (%s/) ...', HF_MODEL_REPO, HF_MODEL_SUBDIR)
 	local = snapshot_download(

lilyscript/generator.py

@@ -159,6 +159,33 @@ class StreamingLilyletGenerator:
 	def patches_to_text (self, patches):
 		return ''.join(self.patch_to_text(p) for p in patches)
 
+	def _encode_lines (self, lines):
+		'''Encode text lines into padded patches (trailing newline per line, matching
+		the patchifier). Returns a list of patch_size-long id lists.'''
+		patches = []
+		for line in lines:
+			ids = self.tokenizer.encode(line + '\n')
+			for i in range(0, len(ids), self.patch_size):
+				chunk = ids[i:i + self.patch_size]
+				patches.append(chunk + [self.pad_id] * (self.patch_size - len(chunk)))
+		return patches
+
+	def _seed_patches (self, prompt_text):
+		'''Build the seed patch list, mirroring the training patchifier's layout:
+
+			%-prompt patches  ->  <bos> patch  ->  [header] patches
+
+		Lines starting with `%` are the unsupervised style PROMPT (go BEFORE <bos>); the
+		rest (`[field "..."]` headers) are the supervised HEADER (go AFTER <bos>). With no
+		`%` lines this reduces to `[<bos>] + header` (the legacy <bos>-at-index-0 layout).
+		'''
+		prompt_lines, header_lines = [], []
+		for line in prompt_text.splitlines():
+			(prompt_lines if line.lstrip().startswith('%') else header_lines).append(line)
+
+		bos_patch = [self.bos_id] * (self.patch_size - 1) + [self.eos_id]
+		return self._encode_lines(prompt_lines) + [bos_patch] + self._encode_lines(header_lines)
+
 	# ---- KV plumbing ----
 
 	def _empty_patch_past (self):
@@ -260,24 +287,18 @@ class StreamingLilyletGenerator:
 		'''
 		rng = np.random.default_rng(seed)
 
-		bos_patch = [self.bos_id] * (self.patch_size - 1) + [self.eos_id]
-		patches = [bos_patch]
-		if prompt_text:
-			for line in prompt_text.splitlines():
-				ids = self.tokenizer.encode(line + '\n')
-				for i in range(0, len(ids), self.patch_size):
-					chunk = ids[i:i + self.patch_size]
-					patches.append(chunk + [self.pad_id] * (self.patch_size - len(chunk)))
-
-		out_text = self.patches_to_text(patches[1:])
+		patches = self._seed_patches(prompt_text)
+
+		out_text = self.patches_to_text(patches)
 		# 0-based marker: `y` counts measures remaining AFTER this one (patchifier:
 		# y = total - i - 1), so `[r:0/{measures-1}]` yields exactly `measures` total.
 		prime_ids = self.tokenizer.encode(f'[r:0/{measures - 1}]') if measures is not None and measures >= 1 else None
 		primed = False
 
-		# seed the monitor's running context/stream from the prompt patches (if any)
-		if monitor is not None and len(patches) > 1:
-			for p in patches[1:]:
+		# seed the monitor's running context/stream from the seed patches (commit_forced
+		# skips bos/pad/eos, so passing the whole list incl. the <bos> patch is safe).
+		if monitor is not None and len(patches) > 0:
+			for p in patches:
 				for tid in p:
 					monitor.commit_forced(tid)
 

web/vendor/lilylet.bundle.js

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:51d680585d5628989b6d3c145a2b1a0cdd5010549c958fe512fca77d809f617a
-size 586799
+oid sha256:869c6da801fee553a8c8d6c1ca24b6fae3fbe65993167f9a3c0a01540cd413a8
+size 86557