per channel QAT is really slow

[shuyuan-wang]

I'm currently using 1.29 AIMET. when I try to do QAT using per_channel config, the time spent is almost 50x longer than using default config. Is there a solution?

[quic-mtuttle]

Hi @shuyuan-wang, could you provide a bit more information on how you are instantiating your QuantizationSimModel and running QAT (e.g., which QuantScheme you are using)?

[quic-klhsieh]

Hi @shuyuan-wang , as of AIMET 2.0 onwards, QAT with per channel should be on par with per tensor. Please reopen this ticket if you continue to face any issues with newer AIMET version.