Ternary-Bonsai-27B-antidoom-DSpark

A DSpark-accelerated + antidoom (FTPO anti-repetition) build of prism-ml/Ternary-Bonsai-27B — the ternary (Q2_0, 2.125 bpw g128) quantization of the Qwen3.5-27B hybrid-attention model.

Runs in the PrismML-Eng/llama.cpp fork (CUDA Q2_0 hybrid-attention kernels). Built and validated on a single RTX 5090 (32 GB, Blackwell sm_120).

Companion model: Danny-Dasilva/Bonsai-27B-antidoom-1bit-DSpark (1-bit).

Contents

Run with DSpark (fastest)

build/bin/llama-server \
  -m Ternary-Bonsai-27B-antidoom-Q2_0.gguf \
  -md Ternary-Bonsai-27B-dspark-Q4_1.gguf \
  --spec-type draft-dspark --spec-draft-n-max 4 \
  -ngl 999 -ngld 999 -fa on -c 16384 -np 1

Speed comparison — all four models (RTX 5090, 32 GB)

Single stream, greedy, 256 new tokens, PrismML llama.cpp fork (CUDA 13 / sm_120). Native = llama-bench tg256 (fa=1, r=2). DSpark = llama-speculative-simple, --spec-draft-n-max 4, code prompt (high-acceptance workload); all four runs measured back-to-back in the same session.

tok/s (code prompt, bs=1)                 native ▒   +DSpark █
Bonsai 1-bit base      ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ 166.0
                       ████████████████████████████████ 208.3
Bonsai 1-bit antidoom  ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ 153.7
                       █████████████████████████████ 192.4
Ternary base           ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ 136.5
                       ████████████████████████████ 185.9
Ternary antidoom       ▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒▒ 129.2
                       ████████████████████████████ 184.6

Notes:

• DSpark's gain is workload-dependent: on low-acceptance prose (~51%) the 1-bit target can be slower than native because the tiny weights make drafting overhead relatively expensive.
• The antidoom variants are ~5–8% slower natively because re-quantization stored the output tensor at q6_K (prism's originals pack it smaller); with DSpark the gap mostly closes.
• FTPO does not hurt the DSpark drafter (accept 74.6–78.6% 1-bit, 69.2% ternary, unchanged base vs antidoom for ternary).
• All numbers beat prism-ml's published H100 figures (104.8→143.8 1-bit, 98.0→131.8 ternary): these models are memory-bandwidth-bound and the 5090 has more bandwidth headroom at bs=1.

Antidoom (FTPO) pass — what was done

Bonsai models are unusually doom-loop resistant (they self-correct repetition), so runaway-repetition pairs were surfaced by priming: contexts that already began repeating a phrase, sent to the raw completion endpoint, where the model continues the loop. Antidoom's own detector + chosen-token sampler extracted 40 FTPO preference pairs (rejected = the loop-continuation token, chosen = coherent escapes). FTPO trained a LoRA (r=64, q/k/v/o + gate/up/down + lm_head, 12 epochs, lr 3e-5) on the FP16 unpacked model:

• chosen-win (prefers the coherent escape over continuing the loop): 0.02 → 0.53 (early-stopped ≥0.4).
• LoRA merged (CPU) into FP16, re-quantized to Q2_0. The aggressive 2-bit format partially dilutes the LoRA's fine adjustments; the anti-repetition effect is strongest at higher precision.

Provenance

Base weights, DSpark drafter and mmproj are from prism-ml (model weights re-quantized after the antidoom LoRA merge; drafter/mmproj copied unchanged). Built with antidoom and the PrismML llama.cpp fork.