Hybrid Gated DeltaNet models keep fast prompt processing but collapse ~25x in decode past ~80K context

With the model fully resident on GPU (zero 'assigned to device CPU' lines in the log), decode falls from ~33 t/s at 68K KV position to 1.4 t/s (714 ms/token) at 91K position, or times out, while prompt processing stays ~1300 t/s at the same positions. Zero-context decode is normal (~38-39 t/s). Reproduced on native Windows 11 (CUDA 13.3) and WSL2 Ubuntu 22.04, on master 8144f31 and release b10488.

Not fully root-caused in the issue; the reproduction matrix localizes it to the KV-cache path at high positions rather than to GPU residency or quantization. The cliff survives with the model fully offloaded (-ngl 999), so it is not CPU offload. It persists across KV quant types (q4_0 and q5_0 both stall or time out) and with FA on or off; only -c 73728 with q8_0 KV and no FA stayed at 33.1 t/s @ 68K. Affected files: unsloth/Qwen3.8-27B-GGUF UD-Q3_K_XL, UD-IQ4_XS, IQ4_XS on RTX 4080 SUPER (sm_89, driver 591.86).

Cap -c below roughly 80K KV position for these models on this stack, or use q8_0 KV with FA off at the smaller context. Treat as an open upper bound on usable context for hybrid GDN GGUF models on sm_89 rather than a solved configuration.

Confidence is medium: the symptom is well reproduced with a clean position sweep, but the mechanism is not identified, so the workaround is a bound rather than a fix. The pp/tg split is the key diagnostic - fast pp with collapsed tg at long context rules out memory-capacity and offload explanations.