Error when the context window is full, since v 1.94

[ProcrastinatingUnit]

Since version 1.94, when the context window overflows with SWA and FastForwarding enabled, instead of prompt processing, as in older versions(v1.93.2), I get an error as in the screenshot.

Model gemma-3-12b-it-q4_0_s, all context memory and layers in video memory, all ffn_up and ffn_gate tensors override in the CPU RAM. Other settings are standard

[LostRuins]

Confirmed, related to ggml-org#14217, I will investigate

[LostRuins]

Please try the newest version 1.94.2

[ProcrastinatingUnit]

Oh, thank you! Now everything works like clockwork 🤝