Skip to content

ggml: bump to qvac-ext-ggml#8 (Supertonic ops + Vulkan/Metal fixes)#143

Closed
ogad-tether wants to merge 1 commit into
mainfrom
feat/ggml-from-speech
Closed

ggml: bump to qvac-ext-ggml#8 (Supertonic ops + Vulkan/Metal fixes)#143
ogad-tether wants to merge 1 commit into
mainfrom
feat/ggml-from-speech

Conversation

@ogad-tether

Copy link
Copy Markdown
Contributor

Summary

Updates the registry's ggml port pin from 05afdc59 (merge of qvac-ext-ggml#6) to 60a172e (merge of qvac-ext-ggml#8) — the head of the speech branch.

Bumps ggml port-version 7 → 8.

What's new in this bump

Between the prior pin (05afdc59) and the new one (60a172e), the speech branch picked up:

From qvac-ext-ggml#7 (merge 91676f0): Vulkan / Metal / CMake fixes cherry-picked from the 2026-01-30 branch — persistent VkPipelineCache + crash-safe eager flush, GGML_LIB_OUTPUT_PREFIX option, hybrid backend packaging, GGML_METAL_FUSE_MV_BIAS CMake option, std::filesystem pipeline-cache rename (Windows-correct overwrite).

From qvac-ext-ggml#8 (merge 60a172e): The Supertonic 2 fused custom op family — five fused Metal kernels (depthwise_1d / layer_norm_channel / pw2_residual / bias_gelu / edge_pad_1d), each with a CPU forward as parity backstop. Stride-parameterised so the same compiled Metal kernel handles both [T, C] and [C, T] activations via an op_params layout flag. New public ctors in ggml.h: ggml_supertonic_*{,_ct,_causal_ct}. GGML_OP_COUNT bumps 96 → 101.

Downstream consumer

qvac-ext-lib-whisper.cpp#15 (tts-cpp Supertonic Metal port) currently carries a local overlay-port to ship the supertonic patches. Once this lands, that PR can delete its entire tts-cpp/cmake/vcpkg-overlay-ports/ggml/ directory + the tts-cpp/vcpkg-configuration.json shim, and source the patched ggml directly from this registry.

Test plan

  • vcpkg x-add-version ggml clean (added 2026-01-30#8 to versions/g-/ggml.json + versions/baseline.json)
  • tts-cpp builds against this branch end-to-end on Apple M2 / Metal
  • supertonic-cli writes a valid 2.85 s WAV at 44.1 kHz with all _ct paths engaged
  • supertonic-bench: vocoder 13.88 ms / vec_est 63 ms / total 96 ms median (causal kernel active end-to-end via the upstream path)
  • Reviewer to confirm no regression for non-Supertonic ggml consumers — all additions on the speech branch are net-new ops; no existing code paths modified

🤖 Generated with Claude Code

Updates the pin from 05afdc59 (merge of qvac-ext-ggml#6) to 60a172e
(merge of qvac-ext-ggml#8), bringing in:

  - Vulkan / Metal / CMake fixes cherry-picked from the 2026-01-30
    branch (persistent pipeline cache, GGML_LIB_OUTPUT_PREFIX, hybrid
    backend packaging, GGML_METAL_FUSE_MV_BIAS, std::filesystem
    pipeline-cache rename) — merged via qvac-ext-ggml#7.

  - The Supertonic 2 fused custom op family — five fused Metal kernels
    with CPU forwards, stride-parameterised so the same compiled
    kernel handles both [T, C] and [C, T] activations.  Consumed by
    tts-cpp's Supertonic Metal path (qvac-ext-lib-whisper.cpp#15) —
    once this lands, that PR can drop its entire local overlay-port
    machinery and source the patched ggml directly from this registry.
@ogad-tether

Copy link
Copy Markdown
Contributor Author

Closing — the bump scope is too broad for this round.

The registry's ggml port is shared across many consumers (whisper-cpp, llama-cpp, qvac-fabric, …); bumping the pin from 05afdc59 to 60a172e brings in 9 commits (Vulkan/Metal/CMake fixes from qvac-ext-ggml#7 + the Supertonic ops from #8). Even though everything on the speech branch is additive (new ops, new CMake options), GGML_OP_COUNT changes 96 → 101 and the other tweaks could affect other consumers' builds in ways that haven't been validated here.

The downstream consumer (qvac-ext-lib-whisper.cpp#15) already uses a local overlay-port that sources directly from qvac-ext-ggml@speech HEAD 60a172e with zero patch files — the supertonic ops come in via the source commits, not via the registry. That keeps the blast radius scoped to tts-cpp.

We can revisit this bump as its own PR once other consumers have had a chance to validate against speech HEAD.

🤖 Generated with Claude Code

@GustavoA1604 GustavoA1604 deleted the feat/ggml-from-speech branch May 20, 2026 12:04
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