Repository navigation
0.19.10 release plan: iPhone VAE watchdog, Web browser coverage, LiteRT-LM and ASR device qualification #1995
Description
Goal
Ship core 0.19.10 from the patch-level P2 backlog: the iPhone full-VAE watchdog (#1047), Web browser coverage (#1025), LiteRT-LM strict-output device evidence (#174), and Web Qwen3-ASR validation (#329). The 0.19.9 GPU leftovers (#1189, plus #174's L4 run) run whenever GCE has capacity. Tracker #1993 is filler.
The maintainer confirmed this scope on 2026-11-13, after a review of the open P2s.
Baseline: 0.19.9 (prep #1994, plan #1969).
Status — November 14
Published: owner releases stable-diffusion-native v0.3.0-4 (prerelease), litert-lm-native v0.19.1-5 (prerelease) and llama-web-bridge v0.1.85. Main pins LiteRT-LM v0.19.1-5, stable-diffusion v0.3.0-4 and bridge v0.1.85.
| Lane | Item | State |
|---|---|---|
| 1 | #1047 iPhone full-VAE watchdog | Done: pin bump #2005 merged (v0.3.0-4, re-admits companion v0.3.0-3); iPhone 16 Pro and 17 Pro passed on DDP. #1047 closed. Known limit #2004 (iOS 26 retains about 2 GB after the second dispose). |
| 2 | #1025 Safari/Firefox/NVIDIA WebGPU | Repin #2001 and the Safari/Firefox harness #2002 are merged. Optional: Safari GPU smoke (needs an unlocked screen) and a real Firefox run. |
| 3 | #174 LiteRT-LM strict output | Android arm64 GPU admitted (#2006). Linux L4 run passed (evidence); gate #2007 admits Linux x64 GPU (high-risk, audited; delta re-audit of a doc fix pending, then the maintainer's merge). Windows x64 and Linux arm64 GPU stay unqualified. |
| 3a | #1998 Android WebGPU route regression | Fixed by repin #2003 (v0.19.1-5); the S24 qualifier passed 20/20. Closed. |
| 4 | #329 Web Qwen3-ASR | #1997 and harness #1999 merged. Remaining: a human mic pass. |
| 5 | #1189 GCE runs | Run B covered by #1025. Run A passed: native CUDA and Vulkan media vectors on an L4 with published 0.19.9 (evidence). Android GPU, Windows and WebGPU-hardware vectors remain. |
| 6 | #1993 filler | Audit P3s from this cycle added; not started. |
Release prep: bump the llamadart_litert_lm_flutter and llamadart_stable_diffusion_flutter companions, and raise their core floors to >=0.19.10 <0.20.0.
Excluded
- New API or features that need a minor release: Add streaming per-token log-probabilities to generation #694, Define and implement truthful native video-input capability #320, Web image generation: SDXS and SD-Turbo on WebGPU through a stable-diffusion.cpp core in llama-web-bridge #969.
- Blocked elsewhere: Observability: remaining LiteRT usage, token accounting and optional OTel companion #696 (litert-lm-native#91), Support text document context in Genkit model requests #1210 and Support media-bearing Genkit tool results after runtime qualification #1212 (Genkit upstream), CI: investigate Flutter Chrome test shutdown hanging after passing tests #1702 (Flutter upstream), gpuDevices: validate multi-GPU layer and row split on real 2-GPU hardware #1528 (2-GPU quota denied 3 times).
Process
AGENTS.md applies. Owner-repo work lands in the owner repo first. GCE and Firebase Test Lab run on the maintainer's personal project. The maintainer decides high-risk merges and the 0.19.10 release.
Activity