Fix: Add libaio-dev and reorder pip installs for Unsloth stability

Dockerfile

@@ -1,35 +1,34 @@
-# Use a stable CUDA base
-FROM nvidia/cuda:12.4.1-devel-ubuntu22.04
+# Use a pre-configured PyTorch + CUDA base to save build time and avoid 'int1' issues
+FROM pytorch/pytorch:2.5.1-cuda12.4-cudnn9-devel
 
 # Avoid prompts
 ENV DEBIAN_FRONTEND=noninteractive
 
-# Install Python 3.11 and essentials
+# Install system dependencies needed for bitsandbytes and Unsloth
 RUN apt-get update && apt-get install -y \
-    python3.11 \
-    python3-pip \
-    python3.11-dev \
     git \
+    libaio-dev \
     && rm -rf /var/lib/apt/lists/*
 
-# Set python3.11 as default
-RUN ln -s /usr/bin/python3.11 /usr/bin/python
-
 WORKDIR /app
 
 # Upgrade pip
 RUN pip install --no-cache-dir -U pip setuptools wheel
 
-# Install specific stable versions to bypass 2026 experimental version conflicts
-# Pinning to known-good versions from the 2024-2025 cycle
+# 1. Install Unsloth first (it's the most sensitive to environment)
+RUN pip install --no-cache-dir "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"
+
+# 2. Install specifically pinned versions of the RL stack
+# We use 0.12.1 for TRL and 4.48.2 for transformers - these are stable with Torch 2.5.1
 RUN pip install --no-cache-dir \
-    "torch==2.5.1" \
     "transformers==4.48.2" \
     "trl==0.12.1" \
     "peft==0.14.0" \
     "accelerate==1.2.1" \
-    "bitsandbytes==0.45.0" \
-    "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git" \
+    "bitsandbytes==0.45.0"
+
+# 3. Install remaining data and server dependencies
+RUN pip install --no-cache-dir \
     datasets \
     wandb \
     matplotlib \
@@ -41,7 +40,7 @@ RUN pip install --no-cache-dir \
 # Copy the project files
 COPY . .
 
-# Install the local package
+# Install the local package in editable mode
 RUN pip install -e .
 
 # Set environment variables
@@ -49,5 +48,5 @@ ENV MODEL_NAME="meta-llama/Llama-3.2-3B-Instruct"
 ENV OUTPUT_DIR="outputs/commitguard-llama-3b-grpo"
 ENV WANDB_PROJECT="commitguard"
 
-# Start training
+# Start training automatically
 CMD ["python", "scripts/train_grpo.py", "--samples", "200", "--max-steps", "300", "--push-to-hub"]