…ier guess was wrong The note committed an hour ago guessed that the vocoder rewrite had removed the fused pattern. It cannot have: both of those measurements run current code, so two drivers agreeing says nothing about whether the driver or the graph is what changed. The reasoning did not survive being questioned. The chain predict-woo#8 blamed is still in the tree, untouched - one ggml_add per codebook at audio_tokenizer_decoder.cpp:1108, and git diff f775bbf^ f775bbf does not touch it; that commit rewrote conv_transpose, a different part of the decoder. So there is no mechanism by which our code removed it. pacman.log dates what did change: this host ran mesa 26.1.7 on 2026-08-19 when predict-woo#8 was found, 26.1.8 from the 22nd, 26.2.1 from the 23rd. Absent on 26.2.1 here and on 25.0.7 in the image, present on 26.1.7 - a regression that landed in the 26.1 series and was fixed by 26.2 fits every observation. Also records the loose end this exposed: predict-woo#13 says the vocoder did not reach the GPU until 2026-08-24, which sits badly with a Vulkan fusion over vocoder tensors mattering on the 19th. One of the two entries is describing something narrower than its title. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GLnvmbfzUShz5PQT3GEG4Z

Is voice design (control instruction prompting) supported? #8

Description

Activity

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions