🔄 Авто-синхронизация: из Discussion #1090 каждые 6 часов.
Qwen3-235B-A22B-Instruct-2507-FP8 with FP8 KVcache as the first multi model PoC launch model¶
Автор: @beatifull · Категория: 
Proposals · Создано: 2026-04-18 06:46 UTC · Обновлено: 2026-04-18 06:49 UTC
📝 Описание¶
Hello, I'm a gonka host with almost half a year of experience.
Motivation - I would like to suggest a first model to test multi model PoC, which will fit my setup and prepare the network for the smooth multi model PoC launch. I know the main goal: provide models to consumers that are in demand, but this is a nice option to test everything and make life for hosts easier. This will boost PCI-E cards, especially: 4x H100 PCI-E with two NVLINK islands with TP=2.
High-Level Solution - Enable Qwen3-235B-A22B-Instruct-2507-FP8 with FP8 kvcache as the first multi model PoC launch model.
Implementation: Just enable it as a first option in multi-model PoC
Roadmap - I think this is a nice option to test multi model PoC and make life of hosts easier. The same model, but with a different configuration.
Open Questions - This will make life easier for users with PCI-E setup and nvlink islands, like me. Because BF16 kvcache can't fit with TP=2 setup, and I get a lot of "MISS" penalties if I enable FP8 KVcache.
But I'm not sure if this is in demand by users or gonka developers.