改动 / #2549

#2549 docs: SECOND_GPU.md names the engine options that exist, not setup options that never did

open · @qiuzan2001 · 0 评论 · 去 GitHub 看

Setup & installServer & APINVIDIA / CUDADocumentationWindows

说明

**What it changes.** `docs/SECOND_GPU.md` told the user to run

```
./setup.sh --setup --gpu1-experts 5000 --gpu2-experts 5000 --gpu3-experts 5000
.\START-HERE.bat --setup --gpu1-experts 5000 ...
```

and to add `--gpu-placement layer|stripe` to that command. Setup has no such options: they came with the original patch (109a209e), which put only the engine's options in the tree (`git log -S gpu1-experts -- setup.py setup.sh START-HERE.bat` is empty). Both commands stop with

```
setup.py: error: unrecognized arguments: --gpu1-experts 5000
```

The page now says what works, from the code:
- put `--expert-cache-device1..3 N|auto` and `--expert-cache-remote-placement stripe|layer` (default `stripe`) in the run config's `"args"` (`src/program/generate.cpp` parses them; `serve/server.py` `engine_args` passes `"args"` on), and start Strata again;
- on a PC with several cards setup writes `"gpu"` as the one card it runs on, and `child_env` then sets `CUDA_VISIBLE_DEVICES` to that card alone: the helper cards go in the config's `"env"` (`"CUDA_VISIBLE_DEVICES": "0,1,2,3"`, applied after `"gpu"`, as `tools/test_calibrate.py` does). A `"gpu"` list of several cards is a layer split, every card runs a stage, and the engine stops ("--expert-cache-remote with a layer split needs a GPU that runs no stage");
- a `--setup` run leaves hand-added `"args"` out (`carry_over` / `args_dropped`) and keeps `"env"`.

`docs/MULTI_GPU.md` named the same caches `--expert-cache-remote`, which is no option either; it now says `--expert-cache-device1..3`.

**What it leaves alone.** The rest of both pages, including every measurement. No code changes.

**How it was checked.** Each statement against the code lines above, and the setup error by running `python setup.py --check --gpu1-experts 5000` (with `HOME` pointed at an empty folder). Links resolve (`INSTALL.md#where-things-are-stored`, `MULTI_GPU.md`). I have one GPU, so I did not run a helper-cache config; a sub-agent reviewer checked every claim against the source.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

本站相关内容

相关页面的快捷入口。