Skip to content

feat(zoo): load the Higgs Audio v2 codec decoder from int8 and int4 checkpoints - #102

Merged
xxpopielxx merged 2 commits into
mainfrom
feat/quant-codec
Nov 13, 2026
Merged

xxpopielxx merged 2 commits into
mainfrom
feat/quant-codec

Conversation

@xxpopielxx

Copy link
Copy Markdown
Member

Description

Stacked on #97 (base feat/quant-conv)

The codec decoder now loads from a quantized checkpoint. Its convs load quantized through Conv1d and ConvTranspose1d from #97. The load folds project_out and fc2 into a CPU table, so a quantized project_out or fc2 now folds as its ints times their scales.

Introduces a breaking change?

  • Yes
  • No

Type of change

  • Bug fix (change which fixes an issue)
  • New feature (change which adds functionality)
  • Documentation update (improves or adds clarity to existing documentation)
  • Performance (kernel tuning, graph building, anything related to performance)
  • Other (chores, tests, code style improvements etc.)

Testing instructions

  • pnpm --filter runntime exec vitest run --project unit test/zoo/higgs: the fold of an int8 and int4 project_out and fc2, and every conv loading quantized.
  • pnpm --filter runntime exec vitest run --project gpu test/zoo/higgs/decoder.gpu.test.ts: the decoder from an int8 and int4 file in f32 and f16 against the same decoder loaded with the dequantized floats.

Related issues

Depends on #97.

Checklist

  • I have performed a self-review of my code
  • I have commented my code, particularly in hard-to-understand areas
  • I have updated the documentation accordingly
  • My changes generate no new warnings

Additional notes

@xxpopielxx xxpopielxx added core Engine: graph ops, dispatch, buffer pool, weight loading enhancement New feature or request zoo Task runners and model loaders in src/zoo and removed core Engine: graph ops, dispatch, buffer pool, weight loading labels Nov 5, 2026
@xxpopielxx
xxpopielxx force-pushed the feat/quant-codec branch 2 times, most recently from fd8b7b0 to 28d202e Compare November 10, 2026 17:12
@xxpopielxx
xxpopielxx requested a review from chmjkb November 10, 2026 17:13
@xxpopielxx
xxpopielxx added this pull request to stack #114 November 12, 2026 13:21
Base automatically changed from feat/quant-conv to main November 12, 2026 13:28
…heckpoints

The CPU fold reads a quantized project_out or fc2 as its ints times their scales. The convs
load quantized through Conv1d and ConvTranspose1d. Unit and GPU tests compare against the
dequantized floats
…loats

A quantized conv_t1 stored as [out, in·kernel] with the kernel reversed loads as its
dequantized torch weight. The test checkpoints quantize every conv in torch layout.
@xxpopielxx
xxpopielxx merged commit 7170a20 into main Nov 13, 2026
2 checks passed
@xxpopielxx
xxpopielxx deleted the feat/quant-codec branch November 13, 2026 14:07
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

enhancement New feature or request zoo Task runners and model loaders in src/zoo

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants