-
-
Notifications
You must be signed in to change notification settings - Fork 122
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
use vmap for all activations #221
base: master
Are you sure you want to change the base?
Conversation
could you update the OP with the benchmark result for your system? |
we are missing |
@@ -118,7 +119,7 @@ elu(x::RealOrFloatType, α = one(x)) = vifelse(x ≥ 0, x / one(x), α * (exp(x) | |||
activation function. | |||
""" | |||
function gelu(x::RealOrFloatType) | |||
p = oftype(x / 1, π) | |||
p = oftype(x / 1, Float64(π)) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why is this hardcoding the type here?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Zygote willl fail for oftype(..., ::Irrational)
Needs testing against Zygote and CUDA to make sure we don't break any dispatch that we are relying on. |
Zygote is already tested against here. But testing against CUDA may be out of the scope of NNlib? |
Do we need to update the lower bound for the LoopVectorization version in |
given #224 and the discussion in FluxML/Flux.jl#1272, I no longer think this is the correct way forward. The whole vectorization logic should live in Flux's layers definitions, and we should revert NNlib to its pre-LoopVectorization state |
We should revert the vectorisation stuff and release a patch that drops the packages from dependencies. |
Can you also add the same benchmarks using a simple
|
fix #220