Pull requests / #1551
#1551 engine --help: the 16 options it accepts without a line of their own
open · @stuchapin909 · 0 comentarios · En GitHub
Setup & installMulti-GPUNVIDIA / CUDAModels & quantsDocumentationWindows
Descripción
## Summary The engine parses 16 user-facing options that `strata --help` never lists on a line of their own: some appear only in passing inside another option's text (`--layer-split`, `--adapt-every`, `--adapt-swaps`, `--mtp`), the others not at all. Several are in every run config setup writes (`--pcie-frac`, `--vram-reserve-mib`, `--mtp`) or in the docs (`--split-device`, `--vram-reserve-later-mib`, `--split-skip-if-fits`, `--resident-budget-gib`, `--prompt-cache-root`), so they are the ones a user tries to look up. Same idea as #1198 (`--spec`, `--spec-min-p`). Text only; no behaviour change. ## What changed `src/program/generate.cpp`, `usage()` only, 31 lines added, each option next to its neighbours: - after `--max-context`: `--stop-eos`, `--eos-ids` - before `--pipeline-windows`: `--layer-split`, `--split-device`, `--split-skip-if-fits` - after `--prompt-cache-tail`: `--prompt-cache-root` - before `--mtp-max-t`: `--mtp`, `--mtp-window` - after `--expert-cache N`: `--vram-reserve-mib`, `--vram-reserve-later-mib`, `--adapt-every`, `--adapt-swaps`, `--adapt-decay`, `--pcie-frac`, `--pcie-mode` - after `--resident-experts`: `--resident-budget-gib` The defaults are `Options`' (adapt 4 / 96 / 0.7, reserve 700, `mtp_window` 32768, `prompt_cache_root` 2048, EOS 248044,248046; `--pcie-frac` 0.55 scaled by the probed link below 20 GB/s, 0.2 for the Q2_0 pack); the wording follows the `Options` comments, DETAILS.md and MULTI_GPU.md. `--pcie-mode`: `auto` and `kernel` are the staging copy kernel (mode 2), `dma` the copy engine, `direct` the mapped arena. Left out on purpose: the A/B and debug switches the parser also accepts (`--no-fast-attn`, `--no-fast-select`, `--no-fused-gdn`, `--no-kv-grow`, `--no-prefill-borrow`, `--no-publish-kernel`, `--no-spec-split`, `--no-split-rows`, `--spec-split`, `--spec-corrupt`, `--spec-oracle`, `--dump-final-r`, `--prefill-until`). ## Checked Built on Windows (MSVC, CUDA 13). `strata --help` exits 0; 256 lines (225 before); the widest line is still 105 characters; each of the 16 options now starts a line; every line of the old output is unchanged (diff of the old lines: none). 🤖 Generated with [Claude Code](https://claude.com/claude-code)
En el sitio
Enlaces a install, modelos, releases.