Explain unsupported GGUF and BitConfig models

README.md

@@ -39,3 +39,5 @@ This MVP intentionally targets standard `transformers` +
 - A model must be downloaded before it can be loaded or used.
 - Large models may exceed ZeroGPU memory or take a long time to load.
 - Remote model code is disabled in this first version for safety and stability.
+- GGUF-only, vision, and other non-causal checkpoints are detected and reported
+  as unsupported instead of producing an opaque Transformers traceback.

app.py

@@ -13,7 +13,12 @@ import spaces
 import torch
 import gradio as gr
 
-from model_manager import ModelCache, TransformersCausalLMRuntime, validate_model_id
+from model_manager import (
+    ModelCache,
+    TransformersCausalLMRuntime,
+    UnsupportedModelError,
+    validate_model_id,
+)
 
 
 logging.basicConfig(level=logging.INFO)
@@ -47,8 +52,13 @@ def download_model(model_id: str) -> tuple[str, str]:
     try:
         model_id = validate_model_id(model_id)
         snapshot_path = cache.download(model_id)
+        try:
+            cache.ensure_transformers_checkpoint(model_id)
+            status = f"Downloaded on CPU: `{model_id}`"
+        except UnsupportedModelError as exc:
+            status = f"Downloaded on CPU, but this MVP cannot Load it: {exc}"
         return (
-            f"Downloaded on CPU: `{model_id}`",
+            status,
             f"Disk cache: ready ({snapshot_path.name}).",
         )
     except Exception as exc:
@@ -62,7 +72,6 @@ def load_model_on_gpu(model_id: str) -> tuple[str, str, str]:
 
     try:
         model_id = validate_model_id(model_id)
-        cache.cached_snapshot(model_id)  # local-only check; never downloads here
         active = runtime.ensure_loaded(model_id)
         return (
             f"Loaded on ZeroGPU: `{active}`",
@@ -87,7 +96,6 @@ def chat_with_model(
 
     try:
         model_id = validate_model_id(model_id)
-        cache.cached_snapshot(model_id)  # a chat never performs a download
         tokens, temp, nucleus = _safe_generation_settings(
             max_new_tokens, temperature, top_p
         )

model_manager.py

@@ -8,6 +8,7 @@ weights are loaded during Space startup.
 from __future__ import annotations
 
 import gc
+import json
 import logging
 import os
 import re
@@ -34,6 +35,10 @@ def validate_model_id(model_id: str) -> str:
     return normalized
 
 
+class UnsupportedModelError(RuntimeError):
+    """Raised when a repository is outside the MVP's Transformers LM scope."""
+
+
 class ModelCache:
     """A dedicated Hugging Face cache for downloaded model snapshots."""
 
@@ -70,6 +75,39 @@ class ModelCache:
             ) from exc
         return Path(snapshot_path)
 
+    def ensure_transformers_checkpoint(self, model_id: str) -> Path:
+        """Validate that a cached repo looks like a standard Transformers LM."""
+
+        snapshot_path = self.cached_snapshot(model_id)
+        files = [path for path in snapshot_path.rglob("*") if path.is_file()]
+        has_gguf = any(path.name.lower().endswith(".gguf") for path in files)
+        has_standard_weights = any(
+            path.name.lower().endswith((".safetensors", ".bin", ".pt", ".pth"))
+            or path.name.lower() in {"model.safetensors.index.json", "pytorch_model.bin.index.json"}
+            for path in files
+        )
+
+        if has_gguf and not has_standard_weights:
+            raise UnsupportedModelError(
+                "This Space currently supports standard Transformers checkpoints only. "
+                "The selected repository contains GGUF/quantized files; use a repo with "
+                "safetensors or PyTorch weights, or wait for a GGUF backend."
+            )
+
+        config_path = snapshot_path / "config.json"
+        if config_path.is_file():
+            try:
+                config = json.loads(config_path.read_text(encoding="utf-8"))
+            except (OSError, json.JSONDecodeError):
+                config = {}
+            if config.get("model_type") == "bit":
+                raise UnsupportedModelError(
+                    "The selected repository declares model_type=bit, which is a vision "
+                    "backbone and not a causal language model."
+                )
+
+        return snapshot_path
+
     def describe(self, model_id: str) -> str:
         """Return a short cache status for the UI."""
 
@@ -135,14 +173,13 @@ class TransformersCausalLMRuntime:
 
         model_id = validate_model_id(model_id)
         with self._lock:
+            snapshot_path = self.cache.ensure_transformers_checkpoint(model_id)
             if self._model is not None and self._model_id == model_id:
                 return model_id
 
             # Switching models always releases the old object before reading
             # the new checkpoint, keeping the one-model invariant explicit.
             self.unload()
-            snapshot_path = self.cache.cached_snapshot(model_id)
-
             tokenizer = AutoTokenizer.from_pretrained(
                 str(snapshot_path),
                 local_files_only=True,