#6252 [Other] Intel Arc (xe driver): the installer fallback always uses --stream-experts, even with enough host RAM for a resident arena

open · @jhorology · 2 コメント · GitHub で見る

BenchmarksSetup & installModels & quants

本文

`terminal/strata_cli/launch/adopt.py`'s Intel-card fallback (used when setup.py can't set up the card and prepared model files already exist) picks `--stream-experts` for every non-i915 Arc card ("xe": B-series), regardless of how much host RAM is available:

```python
if intel_driver() == "i915":
    args += ["--ple-io", "ram", "--vram-reserve-mib", reserve]
else:
    args += ["--stream-experts", "--vram-reserve-mib", reserve]
```

`--stream-experts`'s own `--help` text says it's meant for "machines with less RAM than experts." On an Arc Pro B70 with 63 GB RAM running Swift 1.5 IQ3_S (~47 GB of experts), this path is taken unconditionally, giving noticeably lower throughput than simply omitting `--stream-experts` and letting the default resident arena + `--expert-cache auto` handle it (measured on this box: Code-gen decode ~56 -> ~65-78 tok/s, short prompts ~21 -> ~30 tok/s, same model/context/quant).

Proposal: gate the `else` branch on available RAM (e.g. check `/proc/meminfo` or similar against the expert-set size) the same way `--ple-io ram` is offered for i915, instead of hard-coding `--stream-experts` for every xe card.

関連リンク

インストール・モデル・リリースへのサイト内リンク。