ringattention
ringattention copied to clipboard
vmem OOM on TPU
Hi,
I tried to run your script on Cloud TPU v4-64, but failed with following error:
jaxlib.xla_extension.XlaRuntimeError: RESOURCE_EXHAUSTED: XLA:TPU compile permanent error. Ran out of memory in memory space vmem. Used 59.79M of 16.00M vmem. Exceeded vmem capacity by 43.79M.
I tried mesh dim of 1,1,1,32 and 1,1,4,8 all failed.
Any suggestion what caused the error? Thanks.