Examples of training integration are needed.

[tongxin]

LigerKernel has done a great job in democratizing the use of Triton in real world common training use cases. In contrast, our prior focus was almost entirely on building kernels only. Until now we've had a couple of in-house verification for training but we're not yet able to release the code.

Fuller integrated training examples are immensely helpful to users. Shall we prioritize this feature in next steps?