rocm-systems

Author SHA1 Message Date

Author	SHA1	Message	Date
David Addison	b8dcb4dd83	Make verifiable a DSO and add NAME_SUFFIX support Build option DSO=1 generates libverifiable.so which can be used to reduce the combined binary size. Build option NAME_SUFFIX can be used to a add suffix to all generated binaries. e.g. NAME_SUFFIX=_mpi Added new make target: clean_intermediates [ROCm/rccl-tests commit: `1021260ca9`]	2025-04-23 17:07:24 -07:00
David Addison	8d71063e05	Add support for FP8 datatypes Added new datatypes: f8e4m3, f8e5m2 Only supported on H100+ architectures and NCCL versions >= 2.24.0 [ROCm/rccl-tests commit: `501a149d57`]	2025-04-18 19:20:59 -07:00
John Bachan	b5d746b58e	Resync with NCCL 2.13 * Added "verifiable", a suite of kernels for generating and verifying reduction input and output arrays in a bit-precise way. * Data corruption errors now reported in number of wrong elements instead of max deviation. * Use ncclGetLastError. * Don't run hypercube on non-powers of 2 ranks. * Fix to hypercube data verification. * Use "thread local" as the defaut CUDA capture mode. * Replaced pthread_yield -> sched_yield() * Bugfix to the cpu-side barrier/allreduce implementations. [ROCm/rccl-tests commit: `51af5572bf`]	2022-08-22 17:51:06 -07:00

David Addison

b8dcb4dd83

Make verifiable a DSO and add NAME_SUFFIX support

Build option DSO=1 generates libverifiable.so which can be
used to reduce the combined binary size.

Build option NAME_SUFFIX can be used to a add suffix to all
generated binaries. e.g. NAME_SUFFIX=_mpi

Added new make target: clean_intermediates


[ROCm/rccl-tests commit: 1021260ca9]

2025-04-23 17:07:24 -07:00

David Addison

8d71063e05

Add support for FP8 datatypes

Added new datatypes: f8e4m3, f8e5m2

Only supported on H100+ architectures and NCCL versions >= 2.24.0


[ROCm/rccl-tests commit: 501a149d57]

2025-04-18 19:20:59 -07:00

John Bachan

b5d746b58e

Resync with NCCL 2.13

* Added "verifiable", a suite of kernels for generating and verifying reduction
  input and output arrays in a bit-precise way.
* Data corruption errors now reported in number of wrong elements instead of max
  deviation.
* Use ncclGetLastError.
* Don't run hypercube on non-powers of 2 ranks.
* Fix to hypercube data verification.
* Use "thread local" as the defaut CUDA capture mode.
* Replaced pthread_yield -> sched_yield()
* Bugfix to the cpu-side barrier/allreduce implementations.


[ROCm/rccl-tests commit: 51af5572bf]

2022-08-22 17:51:06 -07:00

3 Commits