gpu-optimizer

What problem does it solve?

Provides evidence-based optimization patterns to reduce out-of-memory errors, improve training throughput, and migrate CPU-bound workflows to NVIDIA GPUs for consumer-class cards (8–24GB VRAM).

Core Features & Use Cases

• XGBoost GPU acceleration: QuantileDMatrix, gpu_hist, and device-aware training for faster boosting on CUDA-enabled builds.
• PyTorch mixed precision & compilation: BF16/FP16 selection, GradScaler fallbacks, fused optimizers, and torch.compile guidance to boost performance.
• VRAM management & diagnostics: Gradient checkpointing, accumulation strategies, peak memory monitoring, and profiling to locate bottlenecks.
• CuPy / cuDF migrations: Patterns for NumPy→CuPy and Pandas→cuDF transitions including zero-copy interop with PyTorch.
• Practical trade-offs & anti-patterns: Clear guidance on when to apply each pattern and what to avoid to preserve GPU pipeline efficiency.

Quick Start

Optimize your model training on an NVIDIA GPU with 8–24GB VRAM by enabling BF16 or FP16, turning on gradient checkpointing, and running the diagnostics to validate memory and performance.