Skip to content

Glossary

Jump to bottom
Wiki Keeper (Grok Bot) edited this page Oct 22, 2026 · 3 revisions

Glossary

Hardware and toolchain

  • Blackwell — NVIDIA GPU architecture family. This lab targets the consumer Blackwell card on this host; datacenter Blackwell (sm_100) is out of scope.
  • sm_120 — compute capability 12.0 (RTX 5080). Hard rule: sm_120 ≠ sm_100 — no FA4, TMEM, tcgen05, or MIG assumptions.
  • ShipOfTheseus / RTX 5080 — lab host: ~16 GB VRAM, CUDA 13.3, driver 610.x. Self-hosted Actions labels: self-hosted, Linux, X64, CUDA.
  • Headroom rule — keep ≥ 2 GiB free on GPU 0 before measurement, training, or GPU CI. Train and GPU CI serialize.
  • Green Context — CUDA driver-API SM partitioning (not MIG). Used in L2 (bkl_green_ctx_bench). This host reports 84 SMs, eight-SM minimum partition.

Kernel layering

  • L1 — measure engine CUDA paths (FA, graphs, quant, prefix). Start here.
  • L2 — host scheduling (Green Contexts, queues).
  • L3 — first-party .cu / CUTLASS under kernels/, only after L1 proves a gap.
  • Proven gap — measured L1 shortfall that justifies writing device code.

Engines and measurements

  • Engine FA — fused attention inside the engine under test (llama.cpp / Ollama). Not datacenter FA4.
  • CUDA graph — captured GPU work replayed with one launch (bkl_graph_launch_bench).
  • Prefix / KV reuse — reuse KV when the prompt shares a byte-stable prefix; keep stable system/tool material first.
  • Q4 / NVFP4 / MXFP4 — 4-bit weight paths measured as L1 levers.
  • v0.2.0 — active release target (not yet tagged): measured first-party operator with training-speed gains on MiniCPM5 and Granite. Tagged so far: v0.1.0.
  • F0 correlation — bkl.f0_correlation.v1 JSONL join of Agoge markers and BKL GPU samples.
  • F0 GPU sampler — tools/f0_gpu_sampler.py (#65): per-run NVML sampler that writes bkl_gpu_sample rows for one agoge_run_id, then exits. Not a daemon. See F0-GPU-Sampler.
  • Overhead budget — sampler overhead_ok: median collect_duration_ns ≤ 50 ms and ≤ 5% of cadence_ms (default 1000 ms).
  • agoge-forger (the forge) — sibling training repo; consumes this lab via docs/FORGE_CONSUME.md.
  • ci-gpu — self-hosted GPU workflow; waits ≤10 min for ≥2048 MiB free; 30-minute job timeout.

Refreshed 2026-10-22 by Wiki Keeper (Grok Bot) from code at main @ 853a2fb (tagged v0.1.0; release target v0.2.0).

Clone this wiki locally