[Feature Discussion] Any plans to extend tileRT to support training-side optimization?

[jinbaoziyl]

Great work!!! most of the technical ideas mentioned seem to have potential value for the training phase as well， especially the B card kernel runs very fast..
It would be great to hear the team's thoughts or high-level roadmap regarding the training direction — even if it's not on the short-term plan,
understanding the reasoning behind it would be very helpful.
Thanks for the great work!

[xysmlx]

Hi, @jinbaoziyl , thanks for the kind words and the thoughtful question!

At this stage, TileRT is intentionally scoped to the inference side. The reasoning is twofold:

First, inference and training have meaningfully different optimization objectives and design directions. Inference is largely about pushing latency under tight memory and decoding constraints, while training optimizes for end-to-end throughput across backward pass, optimizer state, distributed parallelism, and numerical stability. A lot of the design choices that pay off on the inference side don't transfer cleanly, and vice versa.

Second, we simply don't have the bandwidth to take on both directions well right now, and we'd rather go deep on one than spread thin across two.

It's not a "never," but a dedicated training path isn't on our near-term roadmap. Really appreciate you raising it, good signal that the direction resonates.