LiquidAI LFM3-1.7B hits the AMD BC-250 with custom HIP kernels and Vulkan llama.cpp
Bilawal Riaz releases lfm3-bc250, bringing fast inference and 2.1 to 2.5 times faster LoRA fine-tuning for LiquidAI LFM3-1.7B to the AMD BC-250 mining board.
The AMD BC-250 board just got a new tool for local AI work. Developer bilawalriaz released lfm3-bc250, a repository that delivers fast inference and LoRA fine-tuning for LiquidAI LFM3-1.7B on the ex-mining board’s custom Cyan Skillfish APU (gfx1013).
Inference runs via llama.cpp using the Vulkan backend and the official QAD-Q4_0 GGUF. It hits 3,130 tokens per second for prompt processing and 229 tokens per second for generation on a single request, scaling up to 1,346 tokens per second total across 32 parallel requests.
For training, the repository relies on custom ROCm, PyTorch, and HIP kernels. LoRA training runs 2.1 to 2.5 times faster than stock Hugging Face and PEFT paths, letting 32k-token sequences fit in memory. Part of the speedup comes from a GPU clock governor that throttles at 90 C instead of 80 C, so the numbers assume that change.
- Read AGENTS.md in the repository first. It lists the commands for each task in order.
- Run
scripts/serve.shto start an OpenAI-compatible server on port 8090. - Use
bench_train.pyinside thebench/directory to run training benchmarks.