https://stondo.github.io/tags/vllm-omni/Recent content in Vllm-Omni on Bits-EntangledHugo -- 0.162.1en-usThu, 27 Aug 2026 17:00:00 +0000https://stondo.github.io/posts/minimax-h3-all-modes-video-editing-two-rtx-pro-6000/Thu, 27 Aug 2026 17:00:00 +0000https://stondo.github.io/posts/minimax-h3-all-modes-video-editing-two-rtx-pro-6000/My MiniMax H3 deployment could only start videos from still images. Getting the Ref2VA partition online (video-in generation and text-instructed video editing) turned into a lesson in how vLLM-Omni actually resolves model layouts: a 196 GB download of the wrong layout, a crash loop caused by a mount point basename, and an API field whose plural I needed.