[Feature Request] Request Training Kernel of Deepseek mHC

[bayesianrisk]

Required prerequisites

• I have searched the Issue Tracker that this hasn't already been reported. (comment there if it has.)

Motivation

Right now, TileLang provides a solid inference kernel for deepseek_mHC, which is already very helpful. However, for many research workflows, having a training kernel is equally important. Supporting the training path would unlock much broader use cases beyond deployment, especially for researchers who need end-to-end training to explore new architectures, optimization strategies, and correctness/performance trade-offs.

Solution

No response

Alternatives

No response

Additional context

No response

[Da1sypetals]

Will this one suffice? However its performance is only ~25% of triton implementation on my test env (H20).

[LeiWang1999]

I will help to optimize the performance when I have time :)