DeepSpeed
3e64f495 - [bug]: fixed comm_dtype in extra_large_param_to_reduce (#7660)

Commit
62 days ago
[bug]: fixed comm_dtype in extra_large_param_to_reduce (#7660) Fixes #7653 The extra-large params were recorded in `param.dtype` but the reducer looks up using `comm_dtype`. https://github.com/deepspeedai/DeepSpeed/blob/d56e847bac2853d5b8819ce176eeafff65a3798e/deepspeed/runtime/zero/stage_1_and_2.py#L1461 cc @sfc-gh-truwase Signed-off-by: Naveenraj Kamalakannan <therealnaveenkamal@gmail.com> Co-authored-by: Masahiro Tanaka <81312776+tohtana@users.noreply.github.com>
Parents
Loading