Ornith-1.0-9B-MTP-GGUF

GGUF builds of deepreinforce-ai/Ornith-1.0-9B with the KL-distilled MTP draft head from protoLabsAI/Ornith-1.0-9B-MTP baked in — so llama.cpp does lossless multi-token (self-)speculative decoding out of the box, no separate draft model required.

~1.4–1.7× single-stream decode speedup on a single RTX A6000, distribution-lossless. The head's per-token acceptance on llama.cpp matches the vLLM reference (0.766 here vs 0.762).

Just want the base model with no MTP? Use deepreinforce-ai/Ornith-1.0-9B-GGUF. These files add the nextn head on top of the same trunk.

Files

File	Form	Size	Use
ornith-9b-mtp-kl-Q8_0.gguf	bundled (trunk + head)	9.8 GB	highest quality / biggest relative speedup
ornith-9b-mtp-kl-Q6_K.gguf	bundled	7.6 GB	near-lossless quant
ornith-9b-mtp-kl-Q5_K_M.gguf	bundled	6.6 GB	balanced
ornith-9b-mtp-kl-Q4_K_M.gguf	bundled	5.8 GB	fastest k-quant
ornith-9b-mtp-kl-IQ4_XS.gguf	bundled (imatrix)	5.5 GB	low VRAM, near-Q4 quality
ornith-9b-mtp-kl-IQ3_M.gguf	bundled (imatrix)	4.7 GB	lower VRAM
ornith-9b-mtp-kl-IQ2_M.gguf	bundled (imatrix)	3.9 GB	very low VRAM (~5 GB to serve)
ornith-9b-mtp-kl-BF16.gguf	bundled (full precision)	18.4 GB	the master; re-quantize from this
mtp-ornith-9b-mtp-kl-Q8_0.gguf	standalone draft head	2.4 GB	attach to a base GGUF via --model-draft

The IQ quants are i-quants built with an importance matrix (calibrated on the trunk) for quality at low bit-rates, with the MTP nextn head pinned to Q8_0 so speculative-decode acceptance holds even on the 2-bit trunk (verified ~0.81–0.84 accept on IQ2_M–IQ4_XS, on par with the k-quants). Serve them exactly like the k-quants (--spec-type draft-mtp).

Requires llama.cpp ≥ b9616 (Qwen3.5 qwen35 arch + --spec-type draft-mtp).

Run

Bundled (recommended) — the head travels in the file:

llama-server --model ornith-9b-mtp-kl-Q4_K_M.gguf \
  --n-gpu-layers 99 --ctx-size 8192 --flash-attn on --jinja \
  --spec-type draft-mtp --spec-draft-n-max 3

Standalone draft — pair the small head with any base Ornith-9B GGUF:

llama-server --model ornith-1.0-9b-Q4_K_M.gguf \
  --model-draft mtp-ornith-9b-mtp-kl-Q8_0.gguf \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  --n-gpu-layers 99 --ctx-size 8192 --flash-attn on --jinja

--spec-draft-n-max is the draft depth: 2 maximizes acceptance, 3 maximizes throughput, 4 starts to regress. Tune per workload.

Benchmarks (RTX A6000, ctx 8192, flash-attn, greedy; 6-prompt code+general mix)

n-max sweep (Q8_0)

config	decode tok/s	acceptance	speedup
base (no MTP)	71.0	—	1.00×
MTP n-max 2	118.3	0.766	1.67×
MTP n-max 3	122.6	0.651	1.73×
MTP n-max 4	120.8	0.565	1.70×

Across quants (MTP n-max 3)

quant	base tok/s	MTP tok/s	speedup	acceptance
Q4_K_M	105.4	145.3	1.38×	0.659
Q8_0	71.0	122.6	1.73×	0.651

Acceptance is quant-stable (~0.65 @ n-max 3 even with the Q4 head). Q4_K_M is fastest in absolute terms; the relative MTP gain grows with precision (Q8's slow bandwidth-bound baseline has more to gain from the parallel verify).

"Lossless" — read this

MTP speculative decoding is distribution-lossless: every drafted token is verified against the target, so the output distribution is unchanged. It is not bitwise-identical to plain decode at greedy/temp 0 — the batched verification path computes target logits in a different floating-point reduction order than sequential decoding, which can flip a greedy argmax and fork the text. Both outputs are equally valid and equal quality; this is expected llama.cpp behavior, not a defect of these weights.

How these were built

# 1. graft the mtp.* head into the base trunk (15 tensors, 1 nextn layer)
python graft.py --donor protoLabsAI/Ornith-1.0-9B-MTP \
                --target deepreinforce-ai/Ornith-1.0-9B --out ./ornith-9b-mtp-kl
# 2. convert (the converter remaps mtp.* -> blk.<32>.nextn.* automatically)
python convert_hf_to_gguf.py ./ornith-9b-mtp-kl --outfile out/...-BF16.gguf --outtype bf16
python convert_hf_to_gguf.py ./ornith-9b-mtp-kl --outfile out/ --outtype q8_0 --mtp   # standalone draft
# 3. quantize
llama-quantize out/...-BF16.gguf out/...-Q4_K_M.gguf Q4_K_M

The graft.py recipe and the KL-distillation details live in the head repo protoLabsAI/Ornith-1.0-9B-MTP.

Common error: wrong number of tensors expected 442 got 427

(or got 426 for the smaller quants — the gap is the 15 mtp.* head tensors.)

This happens if you run convert_hf_to_gguf.py directly on the base deepreinforce-ai/Ornith-1.0-9B without grafting the head first. The base keeps mtp_num_hidden_layers: 1 in its config.json (text_config) but ships none of the mtp.* weights — so the converter writes block_count = 33 / nextn_predict_layers = 1 into the GGUF metadata (declaring the blk.32 MTP layer) while leaving those 15 tensors empty. llama.cpp then expects 442 tensors and finds 427 → load fails.

Fix: graft the head into the trunk before converting (step 1 above), then convert with no --mtp flag. Note that only 4 of the 15 head tensors are named blk.32.nextn.* (eh_proj, enorm, hnorm, shared_head_norm); the other 11 land as ordinary blk.32.* layer tensors (attn_*, ffn_*, the norms) — so grepping for nextn shows only 4, but the head is complete.

Don't want to graft? You don't have to build the bundled file at all — run the base GGUF with --model-draft mtp-ornith-9b-mtp-kl-Q8_0.gguf --spec-type draft-mtp. Functionally identical.

Provenance & license

• Base: deepreinforce-ai/Ornith-1.0-9B (MIT) — a Qwen3.5-9B hybrid (linear-attention + full-attention) fine-tune.
• MTP head: protoLabsAI/Ornith-1.0-9B-MTP (MIT) — KL-distilled against Ornith's own hidden states.
• These GGUFs are a derivative of both; MIT. Built by protoLabs.studio.