filipe(@filicroval):i came across this llama.cpp fork that uses less VRAM and KV memory. on my GX10, Qwen 3.5-family Bonsai 27B, 60k context, median of three runs: q8_0: • 2,006 MiB KV • 792 tok/s prefill • 27.2 tok/s decode KVarN5 + 1,024-token tail: • 1,349 MiB KV • 740 tok/s prefill • 25.6 tok/s decode The trade: 32.8% less KV memory for roughly 6% lower throughput might be a solid option for people trying to squeeze in some more tok/s ! repo:

filipe

@filicroval

data eng | 1xAsus Ascent GX10 | benchmarking local models so you don't have to

Joined May 2023

214 Following 153.4K Followers

filipe@filicroval

2026.07.26 19:47

i came across this llama.cpp fork that uses less VRAM and KV memory. on my GX10, Qwen 3.5-family Bonsai 27B, 60k context, median of three runs: q8_0: • 2,006 MiB KV • 792 tok/s prefill • 27.2 tok/s decode KVarN5 + 1,024-token tail: • 1,349 MiB KV • 740 tok/s prefill • 25.6 tok/s decode The trade: 32.8% less KV memory for roughly 6% lower throughput might be a solid option for people trying to squeeze in some more tok/s ! repo: