About the known bugs fix plan of amp

[666DZY666]

I have asked a few questions before and there has been some discussion(#3996). Now I would like to ask about the estimated fix plan for these issues, Tansks.

1，Please remove this constraint. The operator i_node_pred==2 is definitely not limited to mul:
https://github.com/quic/aimet/blob/develop/TrainingExtensions/common/src/python/aimet_common/amp/convert_ops_reduction.py#L579

2，Please remove this constraint. The candidate items I need to support are: W8A8, W8A16, W16A16:
https://github.com/quic/aimet/blob/develop/TrainingExtensions/common/src/python/aimet_common/amp/convert_ops_reduction.py#L681

3，Phase2, setting allowed_accuracy_drop to None will only generate one point and will not generate the complete curve as described in the document:
https://quic.github.io/aimet-pages/releases/latest/featureguide/mixed%20precision/amp.html#use-cases

[quic-ykota]

Hello @666DZY666 AMP works in multiple phases. Currently we have Phase-1 (sensitivity list generation), Phase-2 (pareto curve generation), Phase-3 (reducing convert ops). At the end of phase-2, you will be able to see a mixed precision profile. Phase-3 is optional and also experimental right now. Would you be able to try your pipeline without setting the flag GreedyMixedPrecisionAlgo.ENABLE_CONVERT_OP_REDUCTION to True (leave it at the default value of False). With this you'd be able to avoid your errors. Could you let me know if its works for you?

[666DZY666]

Hello @666DZY666 AMP works in multiple phases. Currently we have Phase-1 (sensitivity list generation), Phase-2 (pareto curve generation), Phase-3 (reducing convert ops). At the end of phase-2, you will be able to see a mixed precision profile. Phase-3 is optional and also experimental right now. Would you be able to try your pipeline without setting the flag GreedyMixedPrecisionAlgo.ENABLE_CONVERT_OP_REDUCTION to True (leave it at the default value of False). With this you'd be able to avoid your errors. Could you let me know if its works for you?Hello AMP 分多个阶段进行。目前我们有 Phase-1 （灵敏度列表生成）、Phase-2 （帕累托曲线生成）、Phase-3 （减少转换作）。在阶段 2 结束时，您将能够看到混合精度配置文件。第 3 阶段是可选的，目前也是实验性的。您是否能够在不将标志 GreedyMixedPrecisionAlgo.ENABLE_CONVERT_OP_REDUCTION 设置为 True （将其保留为默认值 False） 的情况下试用管道。这样，您就可以避免错误。您能告诉我它是否适合您吗？

hello，I think the third stage is more important, because if the conversion operations are not reduced, it is very likely that the mixed precision benefits will be greatly reduced or even negative due to the existence of a large number of conversion operations. So I hope that the third stage can be officially released as soon as possible, which should have a great promoting effect on AMP. Thans!

[666DZY666]

New question: I found that the results of amp are not reproducible when using aimet's built-in sqnr as the first and second stage evaluation? Specifically: running amp twice with exactly the same settings (allow accuracy drop, clean_start=False, etc.), the fp32 precision and the final mixed precision configuration are different? Is this as expected?

[666DZY666]

New question: AIMET v2 quantsim cannot use AMP phase 3 and will error ?