`NCCL_TESTS_SPLIT` serves as new way of computing the color for splitting communicators.
Will be overrided by `NCCL_TESTS_SPLIT_MASK`.
Examples:
NCCL_TESTS_SPLIT_MASK="0x7" # color = rank & 0x7. What we do today to run on a DGX with one GPU per node.
NCCL_TESTS_SPLIT="AND 0x7" # color = rank & 0x7. New way to run on one GPU per node on a DGX, equivalent to NCCL_TESTS_SPLIT_MASK=0x7
NCCL_TESTS_SPLIT="MOD 72" # color = rank % 72. One GPU per NVLink domain on an NVL72 system.
NCCL_TESTS_SPLIT="DIV 72" # color = rank / 72. Intra NVLink domain on NVL72.
You can also use: "%" "&" "|" "/" for short.
Extra spaces in the middle will be automatically ignored.
Not case sensitive.
The followings are all equivalent:
NCCL_TESTS_SPLIT="%0x7"
NCCL_TESTS_SPLIT="%0b111"
NCCL_TESTS_SPLIT="AND 7"
NCCL_TESTS_SPLIT="and 0x7"
* 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.
Add error checking for minbytes and maxbytes arguments
Also accept lowercase literals when parsing size arguments and print errors and usage on stderr.
Major rework to merge most of the changes from the NCCL internal
tests into the public ones
Added "-m <agg_iters>" operation aggregation option.
Data integrity checking is now much more performant at scale.
Startup times at scale are improved.
Test latency units are now displayed in usec.
David Addison
Junyu Ma
John Bachan
David Addison
Stefano Salsano
Kaiming Ouyang
Giuseppe Congiu
Sylvain Jeaugey
Greg Inozemtsev
Luke Yeager