Fix: upgrade to Python 3.11-slim + CPU PyTorch >=2.4

Dockerfile

@@ -4,7 +4,10 @@
 # No local Docker needed — HF handles the build and hosting.
 # ──────────────────────────────────────────────────────────────────────────
 
-FROM pytorch/pytorch:2.2.0-cuda12.1-cudnn8-runtime
+# Use plain Python slim — no CUDA needed on the free CPU tier,
+# and it avoids the old PyTorch 2.2 in the pytorch/pytorch base image
+# (GLM-OCR's transformers requires PyTorch >= 2.4).
+FROM python:3.11-slim
 
 # ── System deps ────────────────────────────────────────────────────────────
 RUN apt-get update && apt-get install -y --no-install-recommends \
@@ -16,7 +19,15 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 # ── Workdir ────────────────────────────────────────────────────────────────
 WORKDIR /app
 
-# ── Python deps ────────────────────────────────────────────────────────────
+# ── Install CPU PyTorch first (must come before requirements.txt so pip
+#    doesn't pull the CUDA build from PyPI and stays on >=2.4) ─────────────
+RUN pip install --no-cache-dir \
+    "torch>=2.4.0" \
+    "torchvision>=0.19.0" \
+    --index-url https://download.pytorch.org/whl/cpu
+
+# ── Install remaining dependencies ─────────────────────────────────────────
+# pip will see torch/torchvision already satisfy their constraints and skip them
 COPY requirements.txt ./
 RUN pip install --no-cache-dir -r requirements.txt
 
@@ -24,10 +35,6 @@ RUN pip install --no-cache-dir -r requirements.txt
 COPY main.py ocr_engine.py ./
 COPY frontend/ ./frontend/
 
-# ── Model weights download at first startup (not baked into image) ─────────
-# HF Spaces caches ~/.cache/huggingface across restarts on paid tiers.
-# On the free tier the model (~1-2 GB) re-downloads on each cold start.
-
 # ── Expose & run (HF Spaces requires port 7860) ────────────────────────────
 EXPOSE 7860
 CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "7860"]