Open mo-tenstorrent opened 1 week ago
Because CCL ops basically grab a hold of eth cores, UMD R/W are unusable on all remote chips, if any chip is running a CCL op.
This make dump device calls hang on async runs that call dump before syncing all devices.
Syncing all devices is not a requirement for profiler and if we move to an FD based read, we don't require device sync.
we have a workaround, right? can I drop this to P2?
Yes Thanks
Because CCL ops basically grab a hold of eth cores, UMD R/W are unusable on all remote chips, if any chip is running a CCL op.
This make dump device calls hang on async runs that call dump before syncing all devices.
Syncing all devices is not a requirement for profiler and if we move to an FD based read, we don't require device sync.