[Question][Feature Request] Provides dispatch primitives based on dtypes.
Open
- Dominant language
- C++
- Stars
- 462
- Forks
- 99
- Avg merge
- 8h 37m
- Merged PRs (30d)
- 73
Description
PyTorch's `AT_DISPATCH_FLOATING_TYPES_AND_HALF` (and similar ops) are very helpful when implementing kernel for multiple dtypes, flash-infer also implement [`DISPATCH_DLPACK_*`](https://github.com/flashinfer-ai/flashinfer/blob/main/csrc/tvm_ffi_utils.h).
It would be great if tvm-ffi could provide those utils.
Contributor guide
Research direction
Start by comparing PyTorch's AT_DISPATCH_FLOATING_TYPES_AND_HALF and related macros with the DISPATCH_DLPACK_* utilities in csrc/tvm_ffi_utils.h. Clarify which dtype dispatch primitives tvm-ffi should provide and what API shape fits its FFI; done means the requested utilities and their supported dtype coverage are defined and implemented.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp, pytorch
- Domain
- backend-api-design, machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100