Summary :
spec : process images through the draft context — this directly addresses the mmproj + MTP crash. Previously images (mmproj) couldn't be processed through the speculative/draft context at all. This commit adds that capability. That's the actual fix in progress.
server : fix mtmd draft processing — mtmd is the multimodal (mmproj) handler. Explicitly fixing draft processing for multimodal means they know about the crash and are targeting it.
spec : support parallel drafts — this is infrastructure for running multiple draft models simultaneously, which is required for MTP to work properly at scale with parallel slots.
The combination of all three in one build — multimodal draft fix, parallel draft support, and images through draft context — suggests this is a focused push to get MTP + mmproj working together. PR #22673 might not be far behind.
[link] [comments]