[chmjkb]

Description

Part of #70, phase 4: a GPU test that runs Moonshine from int8 and int4 checkpoints.

test/zoo/moonshine/model.gpu.test.ts builds a Moonshine checkpoint in Hugging Face names, quantizes every Linear, and compares the quantized model against the CPU-dequantized float reference and the float model, in int8 and int4. 4 cases.

Measured on the real Hub files with runntime quantize --bits 8 --group-size 32 from the f16 files (this Mac, f16 math):

model	f16	int8 g32	15 s clip f16 → int8	WER vs reference
Tiny	54.2 MB	39.3 MB	104.2 → 107.5 ms	6.6% → 6.6%
Base	123.1 MB	81.5 MB	106.0 → 106.1 ms	7.0% → 7.7%

int4 raises Tiny's WER to 10.7% and is not faster. Moonshine runs at float speed on int8 because decode is bound by dispatch count, not by weight bytes.

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

1. pnpm typecheck
2. pnpm --filter runntime exec vitest run --project gpu test/zoo/moonshine

Related issues

Part of #70. Stacked on #96.

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

• The test helpers (cosine, quantize every Linear, dequantized reference) now exist in the MiniLM, EmbeddingGemma 2 and Moonshine tests; they move into test/core/weights/checkpoint.ts once these land.
• The quantized files are not on the Hub yet; models.ts entries come with the upload.