Quantized op implementation: quantized_decomposed::dequantize_per_channel_group

[sheetalarkadam]

🚀 The feature, motivation and pitch

The op quantized_decomposed::dequantize_per_channel_group implementation would help in using executorch's quantized CPU ops in models like llama3

Came across this issue when trying to build llama3 without XNNPack. More details here #6975

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

cc @larryliu0820 @manuelcandales @kimishpatel @jerryzh168

[mcr229]

@kimishpatel @manuelcandales is this something we are planning on supporting eventually?

[kimishpatel]

Can you say more on the use case? I agree in general that this is useful to add. But it is likely gonna be slower without xnnpack so just trying to get more context to figure out how to prioritize this

[sheetalarkadam]

I can't think of production usecases as of now. Currently, we are trying measure how much of a boost xnnpack gives. If the performance boost significant, it makes sense to migrate workflows not using xnnpack to use them. But I understand if it will be a lower priority task. My usecase mainly is for getting stats around performance.

[mcr229]

@sheetalarkadam generally performance without xnnpack will be bad. It's generally not a fair comparison between no-xnnpack --> xnnpack, because XNNPACK is the primary cpu accelerator.

[digantdesai]

This was asked before as well. We should just add at least to make it "runnable" without XNNPACK.