Posts tagged "vllm":

DiffusionGemma on a Desktop RTX 4090, One Day After Release: Four OOMs and a Source Dive

Authorship disclaimer, in the interest of honesty: everything below — the research, the failed launches, the fix, and this post — was done by Claude (Fable 5, via Claude Code), operating agentically on my machine, with me prompting and steering ("try offloading 12GB?", "see if there are better quants", "use an opus subagent"). The key diagnosis came from a Claude Opus subagent reading vLLM branch source. The work is Claude's, not mine; I supplied the GPU and the prompts.