AlexsJones/llmfit v1.1.20 on GitHub

1.1.20 (2026-11-03)

NVFP4 and MXFP4 safetensors repos are no longer offered as llama.cpp fits, so they don't top the list on ROCm and CPU machines (#1230)
39 large models now have real attention-head counts from their config.json (including Llama-3.1-405B and MiniMax-M1), so --cluster can check their tensor-parallel splits (#1230)
diff and the TUI compare view no longer put a model that doesn't fit, or whose cluster split is refused, ahead of one that fits (#1230)
cluster: the banner shows the VRAM pool actually used when --memory is set (#1230)
cluster: Ray discovery sets backend from the GPU vendor where Ray reports it, and otherwise names the nodes that need backend set instead of assuming CUDA (#1230)
cluster: --json errors from cluster init, cluster status, cluster clear and --cluster; "1 node" instead of "1 nodes"; cpu_cores taken from the machine's core count, not Ray's CPU resource (#1230)

fix: v1.1.19 Framework/ROCm findings: NVFP4 routing, head backfill, cluster diff/banner/Ray backend by @AlexsJones in #1230
chore(main): release 1.1.20 by @AlexsJones in #1231

Full Changelog: v1.1.19...v1.1.20