RCCL issue 2189: first ncclAllReduce aborts with HIP 'invalid device function' in a heterogeneous gfx1200+gfx1201 process
repo
Fields
id
sources/mgpu-rccl-2189-heterogeneous-archPublisher
ROCm/rccl
Kind
repo
Published
2026-08-31
Accessed
2026-10-03
notes
Communicator init succeeds over P2P/direct, then the first collective raises [FATAL ERROR]: HIP failure: 'invalid device function' on the older-arch device; passing when HSA_OVERRIDE_GFX_VERSION makes both report gfx1201.
Notes
Communicator init succeeds over P2P/direct, then the first collective raises [FATAL ERROR]: HIP failure: 'invalid device function' on the older-arch device; passing when HSA_OVERRIDE_GFX_VERSION makes both report gfx1201.
Cross-references
This record references no other records.