antirez / antirez/ds4

ds4-server: optional --imatrix-out to collect the imatrix from live traffic?

Open
#370 1 comment 1 reaction 0 assignees View on GitHub
Dominant language
C
Stars
22.3k
Forks
2.1k
Avg merge
1d 3h
Merged PRs (30d)
4

Description

Before opening a PR I'd rather ask, since it touches the serving path.

Proposal: `ds4-server --imatrix-out FILE` reuses the existing `ds4_imatrix_collector` and the `imatrix_collect_layer_batch` hook already in `metal_graph_prefill_layer_major`, accumulating the routed-MoE imatrix from live prefills and writing it on shutdown. No new backend code — a `ds4_session` field, a couple of public wrappers, and the prefill call-sites passing the collector. Default off is byte-identical (the field stays NULL unless `--imatrix-out` is given). The file holds only per-expert second moments + counts — no prompt text, ids, or positions.

Why on the server and not just `ds4 --imatrix-dataset`: the offline path needs you to materialize a prompt dataset on disk first. When prompts can't be written to disk or leave the machine for compliance reasons — e.g. patient data under GDPR — that path isn't usable, but collecting aggregate stats in-process is, since no prompt content ever reaches disk. That's the case where the server entry point isn't redundant with the CLI.

Cost: while enabled it forces the layer-major readback (suppresses the streaming-decode fast path), so prefill is slower; decode unaffected; opt-in.

Does this fit ds4, or would you rather keep imatrix collection CLI-only?

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.