Skip to content

serve(model = "ltx25"): serve LTX-2.5 with Gemma 4 - #77

Merged
TroyHernandez merged 2 commits into
mainfrom
feat/serve-ltx25
Oct 29, 2026
Merged

TroyHernandez merged 2 commits into
mainfrom
feat/serve-ltx25

Conversation

@TroyHernandez

Copy link
Copy Markdown
Contributor

serve(model = "ltx25") serves LTX-2.5 on /v1/videos.

  • Loads the ltx2.5-nf4-full and gemma4-nf4 artifacts and places them by recommend("ltx25"): block streaming on 8-12 GB cards, Gemma 4 layer streaming below 12 GB, pin and resident_from from host RAM.
  • A missing or partial artifact stops startup with the build call to run; no downloads.
  • The connector-embed LRU and the artifact validity check are now shared helpers used by both LTX servers; the LTX-2.3 path is otherwise unchanged.
  • The encoder and render pipeline share one process (~31 GB host RAM). Below 32 GB available the server logs a warning; the split text phase doesn't fit a persistent server.

Live test on the 16 GB card under a 34 GiB cap, 512x320x49: 32.9 GB host high-water, 35 s first request (new prompt), 26 s with the prompt cached. Output checked (video + audio).

Tests: full suite 1649 passed, 0 failed (new: embed-cache LRU, artifact validity, missing-artifact error). Version 0.2.2.20.

Loads the ltx2.5-nf4-full and gemma4-nf4 artifacts and places them by
recommend("ltx25"): resident_gb block streaming and Gemma 4 layer
streaming on small cards, resident_from and pin from host RAM. The
connector-embed LRU and the artifact validity check move into shared
helpers used by both LTX servers.
@TroyHernandez
TroyHernandez merged commit fa97b0f into main Oct 29, 2026
3 of 4 checks passed
@TroyHernandez
TroyHernandez deleted the feat/serve-ltx25 branch October 29, 2026 15:40
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant