Update app.py

app.py

@@ -15,7 +15,6 @@ from transformers import AutoModelForCausalLM, AutoTokenizer, LogitsProcessor, L
 
 print(f"Transformers version: {transformers.__version__}")
 
-REPO_ID = "khtsly/Luau-Coder-1.0"
 TOKENIZER_REPO_ID = "khtsly/luau-coder-1.0-preview-tokenizer"
 HF_TOKEN = os.environ.get("HF_TOKEN")
 
@@ -23,10 +22,11 @@ HF_TOKEN = os.environ.get("HF_TOKEN")
 # /models endpoint, so adding a new entry here is enough to offer it.
 # Only one variant is kept loaded at a time; switching unloads the previous.
 MODEL_VARIANTS = {
-    "0.5B-Base": {"repo_id": f"{REPO_ID}-0.5B-Base", "enabled": True},
-    "1.4B-Base (Stage1)": {"repo_id": f"{REPO_ID}-1.4B-Base-Stage1", "enabled": True},
+    "Nalu 1.0 (Base)": {"repo_id": f"khtsly/Luau-Coder-1.0-0.5B-Base", "enabled": True},
+    "Moana 1.0 (Base)": {"repo_id": f"khtsly/Luau-Coder-1.0-1.4B-Base-Stage1", "enabled": True},
+    "Moana 1.5": {"repo_id": "", "enabled": False},
 }
-DEFAULT_VARIANT = "0.5B-Base"
+DEFAULT_VARIANT = "Nalu 1.0 (Base)"
 
 
 def _repair_checkpoint_keys(weights_path: Path) -> bool:
@@ -70,7 +70,7 @@ def _repair_checkpoint_keys(weights_path: Path) -> bool:
 def _prepare_model_source(variant: str) -> str:
     """Return a local model directory whose safetensors keys are loadable."""
     repo_id = MODEL_VARIANTS[variant]["repo_id"]
-    cache_dir = Path(tempfile.gettempdir()) / f"luau-coder-1.0-{variant}"
+    cache_dir = Path(tempfile.gettempdir()) / f"model-{variant}"
 
     # Local override for the default variant (checked-in model.safetensors).
     if variant == DEFAULT_VARIANT:
@@ -100,7 +100,7 @@ def _prepare_model_source(variant: str) -> str:
 tokenizer = AutoTokenizer.from_pretrained(
     TOKENIZER_REPO_ID,
     trust_remote_code=True,
-    token=HF_TOKEN
+    token=HF_TOKEN,
 )
 
 _model = None
@@ -200,9 +200,10 @@ def predict(
     temperature: float = 0.6,
     top_p: float = 0.95,
     top_k: int = 50,
+    min_p: float = 0.0,
     repetition_penalty: float = 1.0,
     presence_penalty: float = 0.0,
-    model_id: str = "0.5B-Base",
+    model_id: str = "Nalu 1.0 (Base)",
 ) -> Generator[str, None, None]:
     # History is accepted for front-end chat compatibility; generation uses
     # the current message only to preserve base-model continuation behavior.
@@ -238,12 +239,13 @@ def predict(
         max_new_tokens=int(max_tokens),
         pad_token_id=tokenizer.eos_token_id,
         use_cache=True,
-        generation_mode=True
+        generation_mode=True,
     )
 
     temperature = float(temperature)
     top_p = float(top_p)
     top_k = int(top_k)
+    min_p = float(min_p)
     repetition_penalty = float(repetition_penalty)
     presence_penalty = float(presence_penalty)
 
@@ -253,6 +255,8 @@ def predict(
             generation_kwargs["top_p"] = top_p
         if top_k > 0:
             generation_kwargs["top_k"] = top_k
+        if 0.0 < min_p <= 1.0:
+            generation_kwargs["min_p"] = min_p
     else:
         generation_kwargs.update(do_sample=False)