gpu_poor

Why we included this project

This is a browser-based estimator for the question everyone hits before pulling down a big model: will it actually run on my GPU, and how fast? Raw model size alone won't tell you, since the KV cache can eat a gigabyte on a 7B model at a modest context length, so the tool splits total memory need into its parts and shows you where the budget goes. It also projects token-per-second throughput and rough fine-tuning time per iteration, which is exactly what you want when choosing between GGML, bitsandbytes, or QLoRA quantization, or deciding whether full, LoRA, or QLoRA training fits your hardware. The estimates have been checked against real GPUs and land within about half a gigabyte, so you get a sane planning range rather than a guess. For anyone sizing a workstation, a rented cloud box, or an inference rig, that removes a lot of trial and error before you commit to a download.