A DSpark-accelerated + antidoom
(FTPO anti-repetition) build of prism-ml/Bonsai-27B —
the 1-bit (Q1_0, 1.125 bpw g128) quantization of the Qwen3.5-27B hybrid-attention model.
Runs in the PrismML-Eng/llama.cpp fork (CUDA, the
Q1_0 hybrid-attention kernels). Benchmarked on a single RTX 5090 (32 GB, Blackwell sm_120).
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-27B is unusually doom-loop resistant (it self-corrects repetition and stays coherent even
under forced long generation), so runaway-repetition pairs were surfaced by priming: contexts that
have already begun repeating a phrase, sent to the raw completion endpoint, where the model continues
the loop. Antidoom's own detector + chosen-token sampler then extracted 40 FTPO preference pairs
(rejected = the loop-continuation token, chosen = coherent escapes). FTPO trained a LoRA
(r=32, q/k/v/o + gate/up/down, 12 epochs, lr 3e-5) on the FP16 unpacked model:
chosen-win (prefers the coherent escape over continuing the loop): 0.11 → 0.66 (early-stopped).
LoRA merged (CPU) into FP16, re-quantized to Q1_0. Note the aggressive 1-bit format partially
dilutes the LoRA's fine adjustments, so the anti-repetition effect is strongest at higher precision.
Output remains coherent at 1-bit. This is a demonstration of the antidoom method on a
1-bit hybrid-attention target, packaged with the shipped DSpark drafter.
Provenance
Base weights, DSpark drafter and mmproj are from prism-ml
(re-quantized after the antidoom LoRA merge for the model weights; drafter/mmproj copied unchanged).
Built with antidoom and the PrismML llama.cpp fork.
