Relace Apply 3
- Source
- relace.ai
- Author
- Relace
- Date
Context
Relace separates deciding how code should change from producing the complete edited file. Apply 3 takes an existing file and an edit description, then merges them. A larger coding model can therefore describe the changes while a smaller specialized model handles their application, including edit descriptions that are awkward for fixed text-replacement rules. Design explanation.
The speed strategy exploits how much of an edited file remains unchanged. Relace describes using : the model guesses a run of upcoming output text, and one pass checks which guesses match, so predictable stretches can be accepted together instead of generated one text unit at a time. For code merging, large sections of the original file and edit are nearly identical to the merged result, which gives long, high-quality guesses. Relace notes that mistakes interrupt a guessed run, so its cleaned training data matters for speed. Inference method.
- speculative decoding — A speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.
- inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
- fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Checking sign-in…
Loading comments…