Yes, constrained decoding and speculative decoding work together, although they need to be carefully integrated and incorrect integrations give result in bugs (see point 3 under Section "Differences in Constrained-Decoding Behavior" in the blog post). Performance overhead does not generally seem to be a large problem. Of the speculative tokens, those that are not allowed by the grammar get discarded.