[Question] Could we consider utilizing the index_mapping feature of IntNBitTableBatchedEmbeddingBagsCodegen to enhance the inference performance of QuantManagedCollisionEmbeddingCollection

[tiankongdeguiji]

In the current version of _mch_remap within the MCHManagedCollisionModule, the searchsorted method is used for remapping, which may not be the most efficient approach for inference. To enhance the inference performance of the QuantManagedCollisionEmbeddingCollection, could we consider using the index_mapping feature of IntNBitTableBatchedEmbeddingBagsCodegen?

[tiankongdeguiji]

hi, @kausv @dstaay-fb @dracifer @sarckk @TroyGarden @PaulZhang12 can you see this problem?

[tiankongdeguiji]

Hi, @henrylhtsang @joshuadeng @YLGH can you see this problem?

[iamzainhuda]

Hey, I've relayed this information to a member on our team who wrote this - they should get back to you shortly

[tiankongdeguiji]

Hey, I've relayed this information to a member on our team who wrote this - they should get back to you shortly

thx!

[kausv]

@tiankongdeguiji I think that could work.
May be you could write QuantMCHManagedCollisionEmbeddingCollection to try it out?

[tiankongdeguiji]

@tiankongdeguiji I think that could work. May be you could write QuantMCHManagedCollisionEmbeddingCollection to try it out?

thx!. The default value of the index_mapping is -1, whereas the default value of the _mch_remap is zch_size. Could you provide some suggestions on how to deal with the difference?