You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
My current strategy involves calling Python tpu-info via popen and sscanf parsing the output.
Would you be open to upstreaming the TPU support in some version (perhaps a Python-less one eventually)?
For Python-less access to TPU data, the TPU metrics are available through a GRPC server, but a C client dependency might be heavy. Another possibility is a C++ or Go library loaded via dlopen .
The text was updated successfully, but these errors were encountered:
I'd rather if we stick to loading a library dynamically (the way it's done for other devices, e.g. à la libnvidia-ml). But I'm not against having conditional build for TPUs using a sub-module (or external lib) that we would link with.
I'm a big fan of your tool and I was hoping we could add TPU support (I work for Google).
I wrote a prototype here: https://github.com/rdyro/nvtop-with-tpu
My current strategy involves calling Python tpu-info via
popen
andsscanf
parsing the output.Would you be open to upstreaming the TPU support in some version (perhaps a Python-less one eventually)?
For Python-less access to TPU data, the TPU metrics are available through a GRPC server, but a C client dependency might be heavy. Another possibility is a C++ or Go library loaded via
dlopen
.The text was updated successfully, but these errors were encountered: