tqchen commented on PR #15700: URL: https://github.com/apache/tvm/pull/15700#issuecomment-1711968854
In the case where it is desirable to retain C1 at compile-time, perhaps the simplest way is to embed C1 as compile time constant object (so they can be embedded in VM bytecode) without relying on external format like the ndarray store. The main rationale here is to keep the weight format simple and indeed intended for weight primarily, because they are big and needs some special considerations. We also keep weight format simple and only limit to weight, so loader do not need to worry about other cases. Coming back to the rank in particular(and not other info), indeed it is useful for each of the worker to know the rank. This is somewhat different from both C0 and C1, since it is a runtime information that can be queried. Perhaps the simplest way is to enable a `R.disco.get_rank` builtin that can be queried by the loader, so they loads the respective pre-shaded data -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
