Repository navigation
Home
CUDA backend for rmems/agoge-forger — local LLM fine-tuning and training on the RTX 5080 (Blackwell / sm_120 / ~16 GB VRAM, CUDA 13.3, driver 610.x).
This repo owns GPU kernels and engine CUDA-path measurements. The forge owns the trainer, datasets, and the SFT / QLoRA / post-training ladder. agoge-forger/cuda/ is a stub that points here.
| Tagged so far |
v0.1.0 (M0 closed) |
| Release target | v0.2.0 — measured first-party CUDA operator consumed by Agoge on both MiniCPM5 and Granite 4.1 (docs/RELEASE_V0_2_0.md) |
| Tip (main) |
853a2fb — feat: add bounded low-overhead GPU sampling for Agoge runs (#53) (#65)
|
| Host | ShipOfTheseus · RTX 5080 · ~16 GB GDDR7 · CUDA 13.3 |
| License | see LICENSE
|
| Repo | https://github.com/rmems/blackwell-kernel-lab |
Hard host rules: target sm_120 only (no FA4 / TMEM / MIG); keep ≥ 2 GiB VRAM free before measurement or train; train and GPU CI serialize — never both at once.
New @ 853a2fb (#65): F0 per-run GPU sampler (tools/f0_gpu_sampler.py) records NVML samples for one Agoge run, and a WSL2 Ubuntu 24.04 setup path (scripts/setup-wsl-gpu-runner.sh, recipes/wsl-gpu-runner.md) registers the ShipOfTheseus runner with the same labels.
Existing smoke tests, telemetry, and synthetic L1/L2/L3 benchmarks are foundations; they do not yet establish the v0.2.0 training-win outcome.
| Page | What it covers |
|---|---|
| Getting-Started | Host sanity, CMake build, smoke/benches, recipes |
| Architecture | L1/L2/L3 layering, kernels tree, forge boundary, F0, CI |
| F0-GPU-Sampler | Per-run NVML sampler for one Agoge training run (#65): CLI, outputs, overhead budget, headroom refusal |
| Glossary | sm_120, Green Context, F0, headroom, L1–L3 |
Forge integration without copying .cu: docs/FORGE_CONSUME.md · boundary: docs/FORGE_BOUNDARY.md.
Refreshed 2026-10-22 by Wiki Keeper (Grok Bot) from code at main @ 853a2fb (tagged v0.1.0; release target v0.2.0).