Minimum model size: 50K parameters

players.py

@@ -12,6 +12,7 @@ from huggingface_hub import HfApi, hf_hub_download
 from huggingface_hub.utils import GatedRepoError, RepositoryNotFoundError
 
 MAX_PARAMS = int(os.environ.get("MAX_PARAMS", 250_000_000))
+MIN_PARAMS = int(os.environ.get("MIN_PARAMS", 50_000))
 ALLOW_REMOTE_CODE = os.environ.get("ALLOW_REMOTE_CODE", "1") == "1"
 MAX_CACHED_MODELS = int(os.environ.get("MAX_CACHED_MODELS", 6))
 BATCH_SIZE = 16
@@ -255,6 +256,8 @@ def precheck(model_id: str) -> dict:
     est = None
     if getattr(info, "safetensors", None) and getattr(info.safetensors, "total", None):
         est = int(info.safetensors.total)
+        if est < MIN_PARAMS:
+            raise ModelRejected(f"`{model_id}` has only ~{fmt_params(est)} parameters; the minimum is {fmt_params(MIN_PARAMS)}.")
     else:
         dtype = str(cfg.get("dtype") or cfg.get("torch_dtype") or "float32")
         est = int(sum(use.values()) / (2 if ("16" in dtype) else 4))
@@ -292,6 +295,9 @@ def load_player(model_id: str, meta: dict) -> LMPlayer:
     if n_params > MAX_PARAMS:
         del model
         raise ModelRejected(f"`{model_id}` has {fmt_params(n_params)} parameters; the limit is {fmt_params(MAX_PARAMS)}.")
+    if n_params < MIN_PARAMS:
+        del model
+        raise ModelRejected(f"`{model_id}` has only {fmt_params(n_params)} parameters; the minimum is {fmt_params(MIN_PARAMS)}.")
     player = LMPlayer(model_id, meta["sha"], model, tok, n_params, meta["custom_code"])
     # smoke test: one forward pass on a real prompt
     try: