Upload 2 files

Dockerfile

@@ -21,9 +21,11 @@ RUN pip install --no-cache-dir \
         pydantic==2.7.1
 
 # Install llama-cpp-python CPU-only (compiled from source)
+# 0.3.4: new enough to read current GGUF format, stable enough to avoid
+# the malloc/GGML_ASSERT corruption seen in 0.3.9 on CPU-only builds
 RUN CMAKE_ARGS="-DGGML_CUDA=OFF -DGGML_METAL=OFF -DGGML_OPENCL=OFF" \
     FORCE_CMAKE=1 \
-    pip install --no-cache-dir llama-cpp-python==0.3.9
+    pip install --no-cache-dir llama-cpp-python==0.3.4
 
 # ── App code ─────────────────────────────────────────────────
 COPY app.py .

app.py

@@ -58,6 +58,13 @@ def _download_model() -> None:
         return
 
     path.parent.mkdir(parents=True, exist_ok=True)
+
+    # Clean up any stale partial download from a previous crashed attempt
+    tmp = Path(str(MODEL_PATH) + ".part")
+    if tmp.exists():
+        logger.warning(f"Removing stale partial download: {tmp}")
+        tmp.unlink()
+
     logger.info(f"Model not found — downloading from {MODEL_URL} ...")
     logger.info("This will take a while on first boot (file is ~9 GB).")
 
@@ -67,7 +74,6 @@ def _download_model() -> None:
     if not HF_TOKEN:
         logger.warning("HF_TOKEN not set — download may fail for gated models.")
 
-    tmp = Path(str(MODEL_PATH) + ".part")
     req = urllib.request.Request(MODEL_URL, headers=headers)
 
     with urllib.request.urlopen(req) as response, open(tmp, "wb") as out: