updated directory config.

.gitignore

@@ -1,5 +1,5 @@
-# Large int8 ONNX weights — loaded from a local dir for now (LFS / HF hub later).
-models/*.onnx
+# Model weights + geometry — loaded from a local dir for now (LFS / HF hub later).
+models/
 
 # Generated outputs written at runtime.
 outputs/
@@ -7,3 +7,5 @@ outputs/
 __pycache__/
 *.pyc
 .DS_Store
+
+*.local

lilyscript/generator.py

@@ -58,7 +58,7 @@ class StreamingLilyletGenerator:
 	def __init__ (self, model_dir, asset_dir, threads=None):
 		from .tokenizer import LilyletTokenizer
 
-		geo = json.load(open(os.path.join(asset_dir, 'geometry.json')))
+		geo = json.load(open(os.path.join(model_dir, 'geometry.json')))
 		self.patch_size = geo['patch_size']
 		self.pad_id = geo['pad_id']
 		self.bos_id = geo['bos_id']
@@ -73,7 +73,7 @@ class StreamingLilyletGenerator:
 		self.t_head_dim = geo['token']['head_dim']
 
 		self.tokenizer = LilyletTokenizer(os.path.join(asset_dir, 'lilylet-tokenizer.json'))
-		self.wte = np.load(os.path.join(asset_dir, 'wte.npy'))		# [vocab, hidden]
+		self.wte = np.load(os.path.join(model_dir, 'wte.npy'))		# [vocab, hidden] — model weight, lives with the onnx
 
 		so = ort.SessionOptions()
 		if threads: