Contributing 17 verified .pte models to executorch-community

[john-rocky]

Thanks for keeping the executorch-community org open on Hugging Face — I just sent a join request (HF username: mlboydaisuke) and wanted to add context beyond the bare notification.

I've converted 17 models to XNNPACK .pte, each shipping with a verification card: parity vs torch fp32 (max_abs_diff + correlation), measured latency, exact pre/post-processing spec, and license. I'd like to contribute them to the org, where they're easier to find than on a personal account — the one ask is to consider the join request, or point me at whichever contribution route you prefer.

• Vision (14): SAM2.1-hiera-tiny (two-file split, E2E mask IoU 1.0000 vs PyTorch), RT-DETRv2-S, D-FINE-S, YOLOX-s, Depth-Anything-V2-Small, CLIP, MODNet, and more.
• LLM (3): LFM2.5-350M / 1.2B and Qwen3.5-0.8B, with iPhone 17 Pro decode numbers on the cards (LFM2.5-1.2B: 55–81 tok/s).
• Conversion is reproducible: one shared harness plus a thin script per model, all public.

Models with cards: https://huggingface.co/mlboydaisuke
Catalog + conversion scripts: https://github.com/john-rocky/executorch-models
Related threads filed while converting: #21805 (iOS tokenizer lookahead abort), and a repro + workaround comment on #21480 (XNNPACK PReLU).

Happy to reshape any of this to the org's conventions — and thanks for ExecuTorch itself; the XNNPACK path has been a pleasure to build on.