Open tiankongdeguiji opened 5 months ago
It what cases would it be necessary to use uint64?
It what cases would it be necessary to use uint64?
In certain scenarios, the uint32 data type is insufficient to contain the indices for user_id and several combo feature.
@tiankongdeguiji Hi, let me bring this to the FBGEMM team's attention and get back to you.
This PR removes uint32 indices protection in quant/embedding_modules.py. Nevertheless, when
indices
andoffsets
inputs of uint64 dtype are provided, int_nbit_split_embedding_codegen_lookup_function triggers aRuntimeError: expected scalar type Int but found Long
on the GPU. Would it be possible to add support for the uint64 dtype on GPU devices?