DeepSpeed
56e7e0c5 - Consistent fp32 grads flow (#8056)

Commit
79 days ago
Consistent fp32 grads flow (#8056) Avoid bf16 buffers when doing fp32 gradient reduction and accumulation --------- Signed-off-by: tunji-ruwase_snow <tunji.ruwase@snowflake.com>
Parents
Loading