Upstream contribution · mohit67890/imajev
imajev · Serving text-only Qwen requests without torchvision
Problem
ImAjev eagerly constructed a multimodal processor during startup. For text-only CPU serving, this activated a video-processor import and prevented the server from starting when torchvision was absent.
Engineering response
The change loads the tokenizer first, lazily constructs the full processor for image inputs, and preserves tokenizer padding through that transition. A text-only install extra makes the supported configuration explicit.
Evidence and scope
The PR reports 28 passing focused tests and three skips, with CPU text-logit and image parity checks on a small generated Qwen checkpoint. The offline full suite retained four baseline failures. Released weights and native Windows ARM64 were outside that validation.