surogate

Why we included this project

Surogate is a training engine for LLMs built on a native C++/CUDA core with a Python wrapper, for teams that want to pre-train, fine-tune, or run reinforcement learning on NVIDIA GPUs, whether on their own hardware or in the cloud. The precision options are the most interesting part: BF16, FP8, and NVFP4 recipes, plus QLoRA with online or pre-quantized weights, so you can trade accuracy against speed and VRAM depending on whether you are on a consumer RTX 50-series card or a B200/B300 datacenter part. Native CPU offloading is pitched as a way to fit larger models on limited memory without dropping to QLoRA, and the package also covers multi-GPU and multi-node DDP training, GRPO/DPO support, and adaptive monitoring with early stopping and learning-rate management. Developers who want quick experiments with Qwen, Llama, Gemma, and other supported architectures get pre-built recipes and a CLI to get a run going, while teams chasing maximum throughput on Blackwell hardware have the FP4 path. It is a training framework rather than an inference server.