kingjones777/Ornith-1.5-35B-A3B-ROCmFP4-FAST-GGUF

Ornith-1.5-35B-A3B — FAST — ROCmFP4 / ROCmFPX + MTP GGUF

The speed-optimised 4-bit of ornith-ai/Ornith-1.5-35B-A3B — ftype 104 Q4_0_ROCMFP4_FAST_COHERENT, single-scale layout, Q6_K token embeddings and a Q6_K output head.

Built for AMD Strix Halo (gfx1151) — Ryzen AI MAX+ 395, 128 GB unified memory, ROCm 7.2.4 — using the ROCmFPX llama.cpp fork, which adds AMD-native FP4/FP8 tensor types that mainline llama.cpp does not have.

⚠️ These files require a ROCmFPX-capable llama.cpp build. They will not load in stock llama.cpp / Ollama / LM Studio — the Q4_0_ROCMFP4_* and Q*_0_ROCMFPX* tensor types are not in mainline.

Variants in this repo

file	ftype	size	BPW	token_embd	output.weight	decode	decode +MTP
Ornith-1.5-35B-A3B-Q4_0_ROCMFP4_FAST_COHERENT.gguf	104	17.93 GiB	4.34	q6_K	q6_K	62.71 t/s	55.59 t/s (0.89×)

Which to pick: FAST at 62.71 t/s no-drafter, statistically tied with LEAN, and it keeps Q6_K token embeddings where LEAN uses Q5_K. Under MTP at n-max 5 it is the fastest of the three. Take FAST for embedding precision, LEAN for the smaller file.

Head protection — verified in the file, not assumed

tie_word_embeddings is false on this model, so output.weight is a real standalone tensor and --output-tensor-type genuinely bites. Every artifact here was re-opened after quantization and its header read back:

ftype	token_embd.weight	output.weight
102	q6_K	q6_K
114	q8_0	q8_0
111	q8_0	q8_0
115	q8_0	q8_0

⚠️ For contrast, the other public ROCmFP4 build of this model (julianmb/Ornith-1.5-35B-A3B-ROCmFP4-GGUF, ftype 106) ships Q5_K token embeddings and a 4-bit output.weight, while its card states "FP16 embedding/norm preservation". We read both headers with two independent parsers. Head protection is not implied by an ftype name — it has to be requested and then verified in the file.

Measured — not estimated

Hardware: AMD Ryzen AI MAX+ 395 (Strix Halo, gfx1151), 128 GB unified, ROCm 7.2.4. Idle box, 2 warm-ups discarded, median of 5, 300 tokens, identical prompt across every sample.

quant	run 1	run 2	run 3	run 4	run 5	median
104	62.79	62.76	62.71	62.71	62.70	62.71

Speculative decoding (MTP)

This model ships its draft head inside the base weights — qwen35moe.nextn_predict_layers = 1, tensors under blk.40.nextn.*. Enable it with --spec-type draft-mtp --spec-draft-ngl 999 and no --model-draft.

⚠️ Do not pass mudler/Ornith-1.5-35B-A3B-APEX-MTP-GGUF as a draft model. It is the same 753 tensors with 32 bytes of extra metadata — loading it as a drafter loads a second full 35B.

⛔ MTP measured a net loss on every tier here (0.89–0.94×). The flags are documented so you can turn it on; we are not selling it as faster. Best n-max on the 4-bit was 3 (56.05 t/s) — note that n=2 had higher acceptance (0.927 vs 0.913) and was slower, so rank on t/s, not acceptance.

Source — byte-verified, not re-converted

Quantized from ornith-ai/Ornith-1.5-35B-A3B-GGUF → Ornith-1.5-35B-BF16.gguf, 71,066,994,240 bytes, sha256 a3ee48dd8f05d10f529aa8ca8b9e080082c38910909638d226001b74e3307591. The vision projector mmproj-Ornith-1.5-35B-BF16.gguf (902,822,016 bytes, sha256 d9ce31026d1cb1f3f8d5152e2e2a014d9d2b302b6c93a7dc07bb0a0487f52837) is included.

Both were byte-verified against the Hub before quantization. No re-conversion from safetensors.

Verification

Every artifact: loaded at -ngl 999 -c 4096 -fit off -fa on, 3/3 correctness (17×23 → 391, capital of Japan → Tokyo, days in 2024 → 366) asserted on both content and reasoning_content with finish_reason=stop, and 4/4 vision on a four-quadrant colour image via the mmproj.

File sizes were checked against --dry-run projections: the header delta is a constant 10.48 MiB across all artifacts (spread 0.005 MiB), which is the signature of complete, untruncated files.

⚠️ Bandwidth note: this is a 256-expert MoE with ~3B active. Effective bandwidth must be computed against the active weight (~1.72 GB at 4.58 BPW), not the 20.3 GB file. At 60.34 t/s that is ~104 GB/s — the file size is not the bus.