Skip to content

Fix CosyVoice3 Metal flow layout - #455

Merged
0xShug0 merged 1 commit into
mainfrom
fix/cosyvoice3-metal-flow-layout
Sep 4, 2026
Merged

Fix CosyVoice3 Metal flow layout#455
0xShug0 merged 1 commit into
mainfrom
fix/cosyvoice3-metal-flow-layout

Conversation

@0xShug0

@0xShug0 0xShug0 commented Sep 4, 2026

Copy link
Copy Markdown
Owner

Summary

  • Materialize the CosyVoice3 pre-lookahead conditioning tensor after transpose before ggml_pad_ext() on Metal.
  • Reuse the original token embedding as the residual instead of rebuilding the same embedding node.
  • Remove the temporary CosyVoice3 Metal backend block.

Validation

Built audiocpp_cli from build/debug, then ran the CosyVoice3 official path case against this branch and a fresh main baseline.

Backend Path case Result Compare
CUDA cosyvoice3_official_paths Pass artifacts/text match, 0 difference(s)
Vulkan cosyvoice3_official_paths Pass artifacts/text match, 0 difference(s)

Timing movement for the short path case was small: CUDA +1.87%, Vulkan +1.78%.

@0xShug0
0xShug0 merged commit a2a4d4f into main Sep 4, 2026
6 checks passed
@0xShug0
0xShug0 deleted the fix/cosyvoice3-metal-flow-layout branch September 4, 2026 16:47
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