Quantized _out functions don't follow same conventions as other out functions in the codebase

[ezyang]

Currently they're defined like:

.op("quantized::add_out(Tensor qa, Tensor qb, Tensor(a!) out)"
     "-> Tensor(a!) out",
    c10::RegisterOperators::options()
      .aliasAnalysis(at::AliasAnalysisKind::FROM_SCHEMA)
      .kernel<QAddOut</*ReLUFused=*/false>>(DispatchKey::QuantizedCPU))

However, a standard out function looks like this:

- func: angle(Tensor self) -> Tensor
  use_c10_dispatcher: full
  variants: function, method
  supports_named_tensor: True

- func: angle.out(Tensor self, *, Tensor(a!) out) -> Tensor(a!)
  supports_named_tensor: True

cc @jerryzh168 @jianyuh @raghuramank100 @jamesr66a @vkuzo @jgong5 @Xia-Weiwen @leslie-fang-intel @dzhulgakov @kevinbchen (who added the alias analysis annotations to these functions) and @z-a-f (who appears to have originally added these out variants at #23971 )

[supriyar]

cc @vkuzo

[z-a-f]

The quantized ops live under torch.quantized.ops.XXX and are not expected to be used by the end-user in Python. However the native_functions exposes the functions in python (mostly). Do we really need to have identical apis for the two? Also, I am not sure if dispatcher schema supports *?

[dzhulgakov]

Only kwarg arguments should work for custom op registration api too. Switching to it is probably not BC-compatible in practice but if jit serializes explicit kwargs properly it might be fine

[ezyang]

Do we really need to have identical apis for the two?

That is a great question! This isn't really related to this issue, but I don't know why quantization also registers their operators under the quantized namespace. Does anyone else know? (@jerryzh168 ?)

[andrewor14]

@jerryzh168 Any context on this? Is it too late to change it now?

[jerryzh168]

yeah this is no longer relevant I feel, since these ops are for server/edge CPU use case and used in original eager mode quantization. we have moved on in both backend (more focusing on server GPU + edge) and flow (more focusing on pt2e flow, and the current GPU + triton flow)