@LottoLabs There is better build actually, this one in report is outdated :D https://t.co/v5zDe96PxZ - you can look here. Much, muuuuch faster. Almost double prefill speed, +~20% TG speed.
@wassollichhier@vllm_project From latest discussions in issues and PRs it's clear that 3090 and ampere will lack support. For deepseek support PR there was "hard no for ampere support". It's sad.
@NoraKaekup@grok Yeah, just got curious how much grok can understand to compare with qwen image edit model. Taking random posts from feed and asks for edits.