spectrometerHBH opened a new pull request, #20173: URL: https://github.com/apache/tvm/pull/20173
`force_cu_dtype` admitted only `CU_TENSOR_MAP_DATA_TYPE_TFLOAT32`. The dtype-derived mapping above it can reach `16U4_ALIGN16B` but never `16U4_ALIGN8B`, so an FP4 tile that needs 8-byte alignment is not expressible at all — neither through the derived mapping nor through the override that exists precisely for cases the mapping cannot reach. This admits the three packed sub-byte types alongside `TFLOAT32`, gated on CUDA 12.8 where those enums appear, and checks the storage each requires: a packed sub-byte payload reaches this API either as its logical sub-byte dtype or as the 8-bit byte storage holding it. Anything else now reports which values are accepted instead of asserting that the argument equals 11. Follow-up to #20154, which introduced the version gating this reuses. -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
