V4 Flash 0731 (Q4K chat-v2 imatrix): greedy / low-temp decoding enters a verbatim repetition loop on long analytical answers

Summary

On M3 Ultra 256G, macOS, ds4 at bd66c40, the V4 Flash 0731 chat-v2 Q4 checkpoint falls into a verbatim paragraph-level repetition loop when answering a long analytical prompt in Chinese (find-the-bug task, ~4 KB script + symptom description). The loop persists regardless of thinking mode or DSpark, and disappears at higher temperature.

Reproduction matrix (same prompt, same model, ds4 -p one-shot, ctx 32768)

Run	Config	Result
1	DSpark + --temp 0	loops after stating the correct root cause; repeats the same paragraph dozens of times (113 KB output)
2	no DSpark + --temp 0.2 --top-p 0.9 --min-p 0.01	loops (numbered analysis points cycle 18→19→20→…)
3	no DSpark + --temp 0.2 + --think	loops (reaches item 55 and keeps cycling)
4	no DSpark + --temp 0.8 --top-p 0.95	clean completion, 4.5 KB, zero repetition

Notes

• Run 1 shows the model actually derives the correct root cause before degenerating, so this is a decode-sampling pathology, not a capability limit.
• DSpark is not the trigger (runs 2-3 loop without it), but since DSpark requires greedy sampling, the combination is effectively unusable for analytical prompts on this checkpoint.
• No repetition of this kind was observed on DeepSeek V4.1 Flash (Q2 resident / Q4 SSD streaming) or GLM 5.3 Flash Q4 with temp 0.2 on the identical prompt.

Happy to provide the raw outputs if useful.