Test report: M4 Pro 24GB — memory ceiling, 720p packing constraint, vae/config.json gap, sharpness data

[AideYu]

Test report: MiniMax-H3 on a 24 GB M4 Pro with a TB5 external SSD

Measured results from running v0.1.22 on a 24 GB M4 Pro (macOS 26.6.1) with models on an external TB5 SSD. Highlights:

1. Memory ceiling on 24 GB: 1344×768 and 1280×736 are refused by the preflight (needs ~7.2–7.8 GB scratch vs ~7.5 GB GPU working set); practical ceiling is 960×544. --memory-cap-mb 2048 does not change it.
2. Height must be a multiple of 32, not 16: patch_size [1,2,2] + 16× VAE means 1280×720 fails with "could not pack a 37x45x80 latent" (latent H=45 is odd). Worth adding to the MINIMAX-H3 docs.
3. Official MiniMaxAI shards quantize directly via model-quantize (no Comfy-Org download needed). DiT tensor structure is identical between the two sources (535 tensors); text encoders differ (64-layer Qwen3-VL vs 50-layer Comfy-Org file).
4. FL2VA against a Comfy-Org-quantized model: vae-encode goes inert ("no readable checkpoint") because resolve_vae_dir/vae_family_ need a config.json the quantization output doesn't produce under vae/. Copying the official video_vae/config.json into vae/config.json fixes it.
5. Sharpness: Laplacian variance on the same frame is 227 (vpipe 4-bit) vs 422 (mlx-serve 4-bit) — identical between official-shard and Comfy-Org quantizations, so it's engine-level precision, not configuration.

Full report (with the 5-way elimination) is in a PR-ready doc: https://github.com/AideYu/vpipe/blob/docs/m4-pro-24gb-test-report/docs/M4-PRO-24GB-TEST-REPORT.md

(Also trying to open a PR — GitHub rejected the API call with a permission error; hence the issue as a fallback.)

[tgo-app-dev]

Thanks for the thorough test report!

The memory limitation on 24GB should be a bug. I can generate 1376x768 on M5 16GB. I'll investigate the memory headroom accounting.

[AideYu]

Thanks for looking into it! Curious how 1376×768 fits in 16GB with weight streaming — that makes the 960×544 ceiling on my 24GB M4 Pro clearly worth revisiting. Happy to re-run the 768p test once the headroom accounting fix lands and report back. Also curious whether you have plans around the sharpness gap (Laplacian 227 vs 422) — that's the one thing that keeps vpipe from being my quality-tier engine.