We do have to optimize inference code paths for fast training. We are currently adding optimizations that target inference-only use cases (some PRs are already

We do have to optimize inference code paths for fast training. We are currently adding optimizations that target inference-only use cases (some PRs are already up). Will post data as soon as that is done. Stay tuned!