ROCmFPX quantisations of Ornith-1.5-35B-A3B for AMD Strix Halo (gfx1151), with the MTP head kept live for speculative decoding.

More variants may be added later.

Every stock ROCmFP4 preset leaves the obvious lever unused on a 256-expert MoE: they apply one type to every tensor. The hybrid splits them:

4.41 bpw · 18.21 GiB. Routed experts are sparse (8 of 256 fire per token) and tolerate 4-bit; attention and the shared expert are on every token's critical path and get 6-bit.

wikitext-2, 145 chunks @ ctx 2048, identical corpus, Vulkan, all measured here:

The three stock presets cluster within 0.8% of each other — preset choice barely matters on this architecture, because none of them differentiate by tensor class. Class-aware assignment moves 4.8% for +0.14 bpw.

Perplexity measures prose next-token prediction, not agentic capability. Use it to compare quantisations of the same weights, not to rank models.

Prefill is identical (within 0.5%) — it is compute-bound, so the FP6 weights cost nothing there. Decode pays the whole price: −28%, because FP6 attention means more bytes per generated token.

⇒ Pick HYBRID for prefill-dominated work (digesting repos/documents, long context, short answers). Pick FAST for generation-dominated work. The recipe is a quality/decode dial, not a free win.

Needle-in-a-haystack retrieval passes at 8.5K, 34.5K and 69.5K on both.

Same build, same model, same flags — only -dev changes:

The ROCm nightly is a prefill-for-decode trade: +12% prefill over HIP 7.2 but −20% decode, and −34% decode against Vulkan. Vulkan wins overall and needs no container.

The nextn.eh_proj head is often kept at Q8_0 on the theory that it determines draft acceptance. Measured on this model, dropping it to FP4 did not hurt — identical perplexity to 4 decimals (7.7749 both) and slightly better acceptance:

Requires a ROCmFPX build — mainline llama.cpp does not know the Q4_0_ROCMFP4_* / Q6_0_ROCMFPX tensor types. The MTP head is native to Ornith 1.5 (blk.40.nextn.*, nextn_predict_layers=1); no graft is needed, unlike 1.0.

Reproduce the recipe with:

attn_.*=q6_0_rocmfpx
ffn_(gate|up|down)_shexp=q6_0_rocmfpx
token_embd.weight=q6_0_rocmfpx
output.weight=q6_0_rocmfpx
nextn.*=q4_0_rocmfp4_fast

On wikitext perplexity, Ornith 1.0 scores far better — 6.19 (ROCmFP4-COHERENT) against 7.40 here, and the gap is present at BF16, so it is a property of the 1.5 weights and not of this quantisation. 1.0 also decodes faster (86.7 t/s) with higher draft acceptance (0.88).

Ornith 1.5 is chosen here for its reported agentic/SWE gains, which wikitext does not measure. If your workload is prose modelling rather than agentic coding, 1.0 may serve you better.