Junru Shao
Junru Shao
This is a very good idea! We should eventually make every language binding (swift, Java, C, Rust, etc) a standalone package that people could directly include in their package manager...
> Thank you! Is there a link to the fix? It's patched internally, because the issue originates from internal LLVM/MLIR driver code which invokes TVM-FFI, not the TVM-FFI codebase. @tqchen...
When will our internal patch be reflected to this open source repo? - sorry I haven’t figured out the mechanism yet
@ngimel correct - I misunderstood the process - it will be public as 4.3.1 releases
I tested the patch just now and it prints: ``` >>> python main.py 0 0 cuda:1 cuda:0 0 cuda:1 cuda:0 0 cuda:1 cuda:0 0 cuda:1 cuda:0 0 cuda:1 cuda:0 0...
Also tested exception handling, e.g. throwing an error inside `cute.jit` - the latest patch seems to handle it properly
Perfect - Thanks @ngimel for communicating the expected behaviors and @tqchen for always bringing the most timely fixes! We all want cuteDSL + tvm-ffi to work the best for PyTorch...
@tqchen Do you imagine there’s a way to formalize pytorch convention in all kernel packages tvm-ffi distributes, i.e. includes but not limited to cuteDSL, flashinfer, etc
Oh it was not my intent to propose dramatic change to dlpack without sufficient buy-in from community! On behalf of tvm-ffi, I’d love to see and facilitate dlpack adoption in...