babusid opened a new pull request, #18422: URL: https://github.com/apache/tvm/pull/18422
This updated implementation of longrope allows for the consideration of `long_factors` and `short_factors`, which are scaling dictionaries provided via HF configs for MSFT's Phi3+ models. In the HF canonical implementation of longrope, once the sequence length exceeds a certain pre-configured dimension, you must use a different set of `ext_factors` than you were previously. This patch enables this by packing both sets of scaling factors into one argument, and selecting which to use dynamically within the returned `prim_func`. -- 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]
