Yixin Dong
Yixin Dong
There has been increased interest from the community in using TVM for training. Relax, the next generation graph level IR of TVM, also faces the demand of training model. We...
This PR supports enabling constrained decoding and Speculative decoding v2 at the same time, resolving #13019. Signed-off-by: Ubospica cc @merrymercy @hnyls2002 @jiapingW ## Motivation ## Modifications ## Accuracy Tests ##...
This PR removes huggingface as dependency. It also adds a py.typed file to enable type checking for callers. Signed-off-by: Ubospica