fix: force torch 2.5.1+cu121 after unsloth to prevent colab-new downgrade

hf_training/Dockerfile

@@ -5,24 +5,21 @@ RUN apt-get update && apt-get install -y git && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
 
-# Upgrade PyTorch to 2.5.1 (cu121) — unsloth_zoo requires torch._inductor.config
-# which only became an accessible attribute in PyTorch 2.4+
-RUN pip install --upgrade --no-cache-dir \
-    torch==2.5.1 torchvision torchaudio \
-    --index-url https://download.pytorch.org/whl/cu121
-
-# Install Unsloth and training dependencies
+# Install Unsloth (no [colab-new] extra — that extra pins old PyTorch and downgrades torch)
 RUN pip install --no-cache-dir \
-    "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git" \
+    "unsloth @ git+https://github.com/unslothai/unsloth.git" \
     && pip install --no-cache-dir --no-deps xformers \
     && pip install --no-cache-dir \
     trl peft accelerate bitsandbytes huggingface_hub safetensors
 
-# Copy training script and dataset
+# Force PyTorch 2.5.1+cu121 LAST — unsloth_zoo requires torch._inductor.config (PyTorch 2.4+)
+# Must be after all other installs to prevent any package from downgrading it
+RUN pip install --force-reinstall --no-cache-dir \
+    "torch==2.5.1" "torchvision==0.20.1" "torchaudio==2.5.1" \
+    --index-url https://download.pytorch.org/whl/cu121
+
 COPY train.py .
 COPY aegis_training_data_500.json .
 
 EXPOSE 7860
-
-# -u for unbuffered stdout so logs appear in real time in HF Space console
 CMD ["python", "-u", "train.py"]