fix: infinite spinner on question gen, 4-bit quant for low VRAM

- new_question returns error text in question_box (was empty string)
so JS polling can detect completion even on failure
- JS polling exits spinner on error instead of waiting forever
- @spaces.GPU timeout raised 60s -> 180s
- LLM loads in 4-bit (bitsandbytes) when VRAM < 17GB, bfloat16 otherwise
- Pre-load libnvJitLink.so.13 so bitsandbytes works locally with cu130
- Add bitsandbytes to requirements.txt

app.py

@@ -267,17 +267,18 @@ def load_pdf(pdf_file, language):
     return chunks, s["loaded"](len(chunks)), 0, 0
 
 
-@spaces.GPU(duration=60)
+@spaces.GPU(duration=180)
 def new_question(chunks, language):
     s = UI[language]
     if not chunks:
-        return "", "", s["no_pdf"]
+        return s["no_pdf"], "", s["no_pdf"]
     try:
         chunk = random.choice(chunks)
         question = generate_question(chunk, language=language)
         return question, chunk, ""
     except Exception as e:
-        return "", "", f"Erreur : {e}"
+        err = f"Erreur : {e}"
+        return err, "", err
 
 
 @spaces.GPU(duration=120)
@@ -526,7 +527,19 @@ BRIDGE_JS = """() => {
 
     if (S.waitingQ) {
       const qEl = document.querySelector('#hidden-question-output textarea');
-      if (qEl && qEl.value && qEl.value !== S.prevQuestion) showQuestion(qEl.value);
+      if (qEl && qEl.value && qEl.value !== S.prevQuestion) {
+        const val = qEl.value;
+        const isErr = val.startsWith('Erreur') || val.startsWith('⚠️');
+        if (isErr) {
+          S.waitingQ = false;
+          const ql=$('q-loading'), nq=$('new-q-btn');
+          if(ql) ql.classList.add('hidden');
+          if(nq) nq.disabled=false;
+          applyStatus(val);
+        } else {
+          showQuestion(val);
+        }
+      }
     }
 
     if (S.waitingFB) {

model/llm.py

@@ -22,13 +22,39 @@ Public API:
 """
 
 import os
+import ctypes
 import torch
 from functools import lru_cache
-from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
+from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline, BitsAndBytesConfig
 
 DEFAULT_MODEL_ID = "openbmb/MiniCPM4-8B"
 DEFAULT_MAX_NEW_TOKENS = 512
 
+# Pre-load libnvJitLink.so.13 bundled with the nvidia-cu13 wheel so that
+# bitsandbytes can find it when it calls dlopen internally.
+def _preload_nvjitlink() -> None:
+    try:
+        import site
+        for sp in site.getsitepackages():
+            candidate = os.path.join(sp, "nvidia", "cu13", "lib", "libnvJitLink.so.13")
+            if os.path.exists(candidate):
+                ctypes.CDLL(candidate)
+                return
+    except Exception:
+        pass
+
+_preload_nvjitlink()
+
+
+def _build_quantization_config(vram_gb: float):
+    if vram_gb < 17:
+        try:
+            import bitsandbytes  # noqa: F401
+            return BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16)
+        except Exception:
+            pass
+    return None
+
 
 class LLM:
     """Thin wrapper around a HuggingFace text-generation pipeline."""
@@ -36,9 +62,17 @@ class LLM:
     def __init__(self, model_id: str, device: str):
         self.model_id = model_id
         self._tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
+
+        vram_gb = 0.0
+        if torch.cuda.is_available():
+            vram_gb = torch.cuda.get_device_properties(0).total_memory / 1024**3
+        quant_cfg = _build_quantization_config(vram_gb)
+        print(f"[LLM] VRAM={vram_gb:.1f}GB — {'4-bit quant' if quant_cfg else 'bfloat16'}")
+
         model = AutoModelForCausalLM.from_pretrained(
             model_id,
-            dtype=torch.bfloat16,
+            quantization_config=quant_cfg,
+            torch_dtype=torch.bfloat16 if quant_cfg is None else None,
             device_map=device,
             trust_remote_code=True,
         )

requirements.txt

@@ -16,3 +16,4 @@ transformers==4.57.1
 accelerate>=0.31.0
 sentencepiece>=0.2.0
 protobuf>=4.25.0
+bitsandbytes>=0.43.0