lmcoleman/Ornith-1.5-35B-A3B-ROCmFPX-GGUF

Ornith-1.5-35B-A3B — ROCmFPX renders

MagicQuant's per-group scheme assignments for ornith-ai/Ornith-1.5-35B-A3B rendered into ROCmFPX tensor types (q*_0_rocmfpx), a quant family aimed at AMD ROCm inference. Same search that produced the MagicQuant GGUFs, different render target.

Pins: MagicQuant ed17991, Foundry 2f99202.

⚠ Engine requirement — read before downloading

These files use ROCmFPX tensor types that stock llama.cpp cannot load. They require a ROCmFPX fork build. A build that is too old fails with invalid ggml type on load; the types are not a stock-llama.cpp feature and no amount of rebuilding mainline will help. If you are not already running a ROCmFPX fork, you want the MagicQuant repo linked above instead.

Files

file	size
Ornith-1.5-35B-A3B-ROCMFPX-MQ-Q4.gguf	20.50 GiB
Ornith-1.5-35B-A3B-ROCMFPX-MQ-Q6.gguf	28.57 GiB
mmproj-Ornith-1.5-35B-A3B-f16.gguf	0.86 GiB

Both passed the pipeline's perplexity smoke gate (6.55 and 6.50 respectively).

Those smoke numbers are not comparable to the MagicQuant repo's PPL table. The smoke gate is a short run against a different corpus path than the search used, so the scale differs for instrument reasons, not quality reasons. No matched-instrument comparison between these ROCmFPX renders and their MagicQuant counterparts has been run, so this repo makes no claim about which is better. If that comparison matters to you, measure both yourself under one instrument.

What is deliberately absent

• No Q5. The pipeline's band guard refused it: rendering MagicQuant's Q5 configuration into ROCmFPX types predicts 26.87 GiB against a 66.13 GiB baseline — a 0.406 ratio, which is the Q6 band, not Q5. Publishing it as "Q5" would have mislabelled it, so it was not built.
• No Q6_0_ROCMFPX_AGENT variant. It built, then segfaulted under perplexity evaluation (exited -11). The pipeline quarantined it automatically and aborted before upload. It is a crash, not a quality miss, and it is not published because it does not work.

Both absences are disclosed rather than quietly dropped, which is the point.

Notes

• Vision requires the mmproj file alongside the model.
• Chat template embedded and verified present before upload; vocab metadata (tokenizer.ggml.token_type) is INT32 per spec.
• Built and measured on an AMD Strix Halo (Ryzen AI MAX+ 395, gfx1151) host.