mirror of
https://github.com/zenlm/grpo.git
synced 2026-10-30 14:59:43 +00:00
zeekay
f926b2fa3d
Major additions: - Code quality reward function for evaluating programming style - Programming style dataset with good/bad code examples - Daily training pipeline for continuous improvement - Code example collector from Hanzo Chat conversations - Comprehensive CLI tool with multiple commands - Docker configuration for training, development, and inference - GitHub Actions workflow for automated training pipeline - Makefile for easy command execution Features: - Collect code examples from chat interactions - Analyze code quality with multiple metrics (syntax, documentation, type hints, etc.) - Create training pairs from real-world code improvements - Automated daily fine-tuning with deployment pipeline - Support for multiple programming languages - Integration with Hanzo AI cloud platform
92 lines
2.0 KiB
Docker
92 lines
2.0 KiB
Docker
# Multi-stage build for GRPO
|
|
FROM nvidia/cuda:12.1.1-cudnn8-runtime-ubuntu22.04 AS base
|
|
|
|
# Set environment variables
|
|
ENV PYTHONUNBUFFERED=1 \
|
|
PYTHONDONTWRITEBYTECODE=1 \
|
|
DEBIAN_FRONTEND=noninteractive \
|
|
TZ=UTC
|
|
|
|
# Install system dependencies
|
|
RUN apt-get update && apt-get install -y \
|
|
python3.10 \
|
|
python3-pip \
|
|
git \
|
|
curl \
|
|
vim \
|
|
build-essential \
|
|
&& rm -rf /var/lib/apt/lists/*
|
|
|
|
# Install Python dependencies
|
|
RUN pip3 install --no-cache-dir --upgrade pip setuptools wheel
|
|
|
|
# Create app directory
|
|
WORKDIR /app
|
|
|
|
# Copy requirements first for better caching
|
|
COPY requirements.txt .
|
|
RUN pip3 install --no-cache-dir -r requirements.txt
|
|
|
|
# Copy source code
|
|
COPY . .
|
|
|
|
# Install the package
|
|
RUN pip3 install -e .
|
|
|
|
# Create necessary directories
|
|
RUN mkdir -p /app/models /app/data /app/outputs /app/logs
|
|
|
|
# Default command
|
|
CMD ["python3", "-m", "grpo.cli", "--help"]
|
|
|
|
# Development image with additional tools
|
|
FROM base AS development
|
|
|
|
# Install development dependencies
|
|
RUN pip3 install --no-cache-dir \
|
|
jupyter \
|
|
ipython \
|
|
tensorboard \
|
|
wandb
|
|
|
|
# Expose ports
|
|
EXPOSE 8888 6006
|
|
|
|
# Training image optimized for GPU
|
|
FROM base AS training
|
|
|
|
# Set up for training
|
|
ENV CUDA_VISIBLE_DEVICES=0
|
|
|
|
# Health check
|
|
HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
|
|
CMD python3 -c "import torch; print(torch.cuda.is_available())"
|
|
|
|
# Default training command
|
|
CMD ["python3", "src/train.py", "--config", "config/grpo_config.yaml"]
|
|
|
|
# Production inference image
|
|
FROM base AS inference
|
|
|
|
# Optimize for inference
|
|
ENV OMP_NUM_THREADS=1 \
|
|
TORCH_NUM_THREADS=1
|
|
|
|
# Install inference-specific dependencies
|
|
RUN pip3 install --no-cache-dir \
|
|
fastapi \
|
|
uvicorn \
|
|
prometheus-client
|
|
|
|
# Copy inference server
|
|
COPY scripts/inference_server.py /app/
|
|
|
|
# Expose API port
|
|
EXPOSE 8000
|
|
|
|
# Health check for API
|
|
HEALTHCHECK --interval=30s --timeout=10s --start-period=5s --retries=3 \
|
|
CMD curl -f http://localhost:8000/health || exit 1
|
|
|
|
# Run inference server
|
|
CMD ["uvicorn", "inference_server:app", "--host", "0.0.0.0", "--port", "8000"] |