Working on v0 and v0-dev branches. - [x] Migrate to VMM and keep IPC as a legacy option - [x] Deprecate `pybind11`-bindings and embrace TVM-FFI for consistency - [x] Introduce support for `multimem`-class instructions - [x] Remove dependency on NVSHMEM/pynvshmem and support GIN - [ ] Native support for kernels on Blackwell and new instructions (e.g. `tma gather4`, `multimem tma`) - [ ] More examples and comprehensive benchmark across baselines - [ ] Clear document and guidance - [ ] Rebase on latest TileLang Also we shall canonicalize memory order in our examples
Working on v0 and v0-dev branches.
pybind11-bindings and embrace TVM-FFI for consistencymultimem-class instructionstma gather4,multimem tma)Also we shall canonicalize memory order in our examples