Skip to content

armgit / Soup

Public Imported from GitHub
0

Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.

NameLast commit date
..
evidence
harness
results
bench_validator.py
gate-1227-ab-burn-in.md
gate-1265-ab-nig.md
gate-1418-ab-cs-accept.md
gate-1702-arena-packing.md
gate-1892-nf4-scale-cache.md
gate-395-second-stack-vram.md
gate-655-pinned-store-accounting.md
gate-674-quest-mixed-route.md
gate-674-quest-w4a4-sft.md
gate-725-lorafa-operating-point.md
gate-836-bench-train-contract.md
gate-901-14b-on-8gb.md
gate-971-async-nvme-source.md
gate-974-disk-tier-direct-io.md
gate-expert-streaming.md
gate-fp8-shard.md
gate-h100-validation.md
gate-kimi-k2-admission.md
gate-qwen4-ple-m4-max.md
gate-stream-recompute.md
gate-two-drive-striping.md
gate-v0.72.0-layer-streaming.md
gate-v0.72.2-nf4.md
gate-v0.72.3-breadth.md
gate-v0.72.4-preference-losses.md
gate-v0.73.1-measured-vram-fit.md
gate-v0.73.2-leg2-scoring.md
gate-v0.76.0-arithmetic-discrimination.md
gate-v0.76.0-common-sense-discrimination.md
gate-v0.76.0-cuda-graph-decode.md
gate-v0.76.0-mmlu-discrimination.md
gate-v0.76.0-tool-call-discrimination.md
gpu-test-runs.md
note-k2-layer-on-24-32gb-cards.md
probe-moe-expert-coverage.md
probe-rtx5070-batch-control.md
probe-rtx5070-drive-throttle.md
probe-rtx5070-giant-moe-seq-scaling.md
probe-rtx5070-giant-moe-step.md
probe-rtx5070-l2l-step0.md
probe-rtx5070-moonlight-16b.md
probe-rtx5070-moonlight-fp8-source.md
probe-rtx5070-two-drive-read.md
probe-rtx5070-two-drive-sustained.md
probe-rtx5070-what-bounds-streaming.md
probe-v0.73.0-what-bounds-streaming.md
quest-674-development-followup.md
quest-674-postfit-development-results.md
README.md
run-m1-8gb-mlx-sft.md
run-t4-colab-free-tier.md