Commit History
8 days ago
-
Fix Mammoth’s CLI selecting the wrong engine
Verified signatureI noticed `qev serve` selects the multimodal engine because Mammoth’s config includes `vision_config`. That engine misses the block-bidirectional attention handling in `serve.patch`. This change routes block-bidirectional checkpoints through the patched text engine, matching the model card’s Python example. Explicit `--multimodal` requests for these checkpoints fail with a clear error. **Checks:** patch applies to the pinned runtime; routing regression fails before and passes after. No new GPU inference run.3881ab5
