space: load the Warden shipped in the repo (no boot download)

.gitignore

@@ -28,3 +28,6 @@ finetune/_nemo_src/
 # OS / editor
 .DS_Store
 *.swp
+
+# the Space ships its model via hf upload, never via local git
+model/

space/app.py

@@ -57,7 +57,15 @@ STATIC = Path(__file__).parent / "static"
 
 # ------------------------------------------------------------ the Warden brain
 
-MODEL_ID = os.environ.get("WARDEN_MODEL", "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16")
+# Prefer the Warden shipped inside the Space repo (pre-quantized nf4, no
+# download at boot); fall back to the hub for self-hosters who clone without
+# the LFS weights.
+_SHIPPED = REPO_ROOT / "model"
+MODEL_ID = os.environ.get("WARDEN_MODEL") or (
+    str(_SHIPPED)
+    if (_SHIPPED / "config.json").exists()
+    else "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
+)
 INTERNAL_KEY = os.environ.get("SCRYPT_INTERNAL_KEY") or secrets.token_hex(16)
 
 tok = None
@@ -127,16 +135,27 @@ if spaces is not None:
 
         tok = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
         # 4-bit fits the `large` (48GB) slice and matches the Q4 GGUF the
-        # local game ships — quality parity, not a downgrade.
+        # local game ships — quality parity, not a downgrade. The shipped
+        # model/ dir is already nf4 (config carries quantization_config);
+        # only quantize on the fly when loading raw BF16 from the hub.
+        _cfg = Path(MODEL_ID) / "config.json"
+        _already_quant = _cfg.exists() and "quantization_config" in _cfg.read_text()
+        _quant_kwargs = (
+            {}
+            if _already_quant
+            else {
+                "quantization_config": BitsAndBytesConfig(
+                    load_in_4bit=True,
+                    bnb_4bit_quant_type="nf4",
+                    bnb_4bit_compute_dtype=torch.bfloat16,
+                )
+            }
+        )
         model = AutoModelForCausalLM.from_pretrained(
             MODEL_ID,
             trust_remote_code=True,
-            quantization_config=BitsAndBytesConfig(
-                load_in_4bit=True,
-                bnb_4bit_quant_type="nf4",
-                bnb_4bit_compute_dtype=torch.bfloat16,
-            ),
             device_map="cuda",
+            **_quant_kwargs,
         )
         WARDEN_ERR = ""
     except Exception as err:  # model is optional: the game survives without it