Calibrate and Quantize Passes for Relax

[ksaurabh-cadence]

Hi,

I wanted to know about status of calibrate and quantize passes for relax, which were previously available with relay. I don’t see any equivalent code in the Relax and can’t locate the relevant document as well. Can someone please point me to this?

Is it possible to calibrate and quantize the pre-trained (unquantized models) imported via relax frontend? Is there a sample program to demonstrate this functionality?

We have been using calibrate and quantize passes with relay. When we migrate to Relax, what should we expect with regards to this?

Thanks S.

[geyijun]

I have the same doubt and looking forward to getting an answer.

[ksaurabh-cadence]

Thanks @tqchen for addressing this in a separate thread. Sharing here for @geyijun -

Phasing out Legacy Components

For new flow we don’t have caliberation flow inside framework, as many calberations are now moved to the upper layer. For example, frameworks like MLC-LLM that usually runs quantization first then build up the quantized model through fused dequant mm operators