is there a cost to cublaslt `retcode != 0`?

[ExpandingMan]

With this PR to OneHotArrays, basic examples with Lux seem to be working with them, but one gets the warning

┌ Warning: cuBLASLt failed for the given inputs relu, CuArray{Float32, 2, CUDA.DeviceMemory} [(8, 4)], OneHotMatrix{UInt32, CuArray{UInt32, 1, CUDA.DeviceMemory}} [(4, 32)], CuArray{Float32, 1, CUDA.DeviceMemory} [8]. Falling back to generic implementation.
└ @ LuxLibCUDAExt ~/dev/Lux/lib/LuxLib/ext/LuxLibCUDAExt/cublaslt.jl:311

This occurs if the return code of cublaslt_matmul_fused! is non-zero. Note that there is another branch in which hasmethod returns false and only a @debug is returned (as I recall I had asked you and you said it's fine if we made it a debug in that case).

This case seems different, because it is not fully eliding the call to cuBLAS, so I really have no idea if this is a bigger issue or again this should just be ignored.

I suggest one of the following:

• If there's no real issue here, this should just be @debug.
• If there's a real problem, we need some way for stuff to opt out of this so that it doesn't try to cal cuBLAS on stuff it doesn't work on.

[avik-pal]

We recently changed the type signature for

Lux.jl/lib/LuxLib/ext/LuxLibCUDAExt/cublaslt.jl

Lines 61 to 71 in 37ef45f

	function cublaslt_matmul_fused!(
	transy::Bool,
	@nospecialize(y::CuMatrix{yT}),
	σ::F,
	transw::Bool,
	@nospecialize(w::CuMatrix{wxT}),
	transx::Bool,
	@nospecialize(x::CuMatrix{wxT}),
	b::Optional{<:CuVector},
	aux::Optional{<:CuMatrix},
	) where {F,yT,wxT}

. Can you check if this is still an issue?

[ExpandingMan]

I still get the same warning. Looks to me like it's coming from the same place, so I don't think the situation has changed.