Skip to content

Refactor model variant selection into explicit routing tables - #1256

Merged
0xShug0 merged 1 commit into
mainfrom
refactor/model-implementation-table
Nov 10, 2026
Merged

0xShug0 merged 1 commit into
mainfrom
refactor/model-implementation-table

Conversation

@0xShug0

@0xShug0 0xShug0 commented Nov 10, 2026

Copy link
Copy Markdown
Owner

Model-family and variant selection currently uses separate conditionals and local lookup code. This change adds ModelRouteTable<Key, Impl> and explicit model-local registration tables, retaining one selected descriptor for execution and family identity.

Migrates Parakeet/Phonon2/Bark3/Haas, PocketTTS/Gradium Phonon, VibeVoice ASR, RoFormer, Audio Flamingo, Irodori RF/MF, and SeedVC. Irodori selects sampler and properties together; RoFormer registrations use named policy assignments. Routing stays at load, session, or request boundaries. Model specs, graph math, backend dispatch, and public options are unchanged.

Validation

  • Debug build and four focused tests passed, covering route selection, invalid registrations, Parakeet variants, Bark3 sessions, and Irodori scheduling.
  • All 25 real-package CUDA end-to-end cases passed; text and audio artifacts matched the main baseline exactly, including offline and streaming cases where applicable.
  • Sequential A→B→B→A comparisons, followed by reverse-order checks for positive timing differences, found no established performance regression. Offline measurements excluded session warmup; streaming CLI measurements used fresh sessions.
  • Irodori RF showed a 1.5–3.6% unpinned timing increase. A repeat with identical CPU affinity reversed it to 1.0% faster, supporting scheduling variability rather than an established routing regression.
  • All 96 repeated-output comparisons passed.

@0xShug0
0xShug0 merged commit 44ec808 into main Nov 10, 2026
7 checks passed
@0xShug0
0xShug0 deleted the refactor/model-implementation-table branch November 10, 2026 20:55
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