Closed radudiaconu0 closed 5 months ago
@GiggleLiu can you take a look at tests? the majority of tests that are failing are failing because somehow in
this function
function _batched_gemm!(C1::Char, C2::Char, alpha, A::ROCArrayTypes{T1,3}, B::ROCArrayTypes{T2,3}, beta, C::ROCArrayTypes{T3,3}) where {T1<:ROCBlasFloat,T2<:ROCBlasFloat,T3<:ROCBlasFloat} AMDGPU.rocBLAS.gemm_strided_batched!(C1, C2, alpha, T1 == T3 ? A : T3.(A), T2 == T3 ? B : T3.(B), beta, C) end
alpha beacomes boolean and it is not allowed in AMDGPU.rocBLAS.gemm_strided_batched!
https://github.com/JuliaGPU/AMDGPU.jl/pull/616 this needs to be approved and the only test that fails is the one with gradients
Attention: Patch coverage is 0%
with 83 lines
in your changes are missing coverage. Please review.
Project coverage is 82.54%. Comparing base (
8c07022
) to head (1634b1b
).:exclamation: Current head 1634b1b differs from pull request most recent head e74d7e9. Consider uploading reports for the commit e74d7e9 to get more accurate results
Files | Patch % | Lines |
---|---|---|
ext/AMDGPUExt.jl | 0.00% | 83 Missing :warning: |
:exclamation: Your organization needs to install the Codecov GitHub app to enable full functionality.
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
@radudiaconu0 Thank you very much for the amazing work!
The alpha
and beta
are default to true
and false
in Julia. You need to cast the types manually to make it consistent with the output type. i.e. you can use the following code instead
AMDGPU.rocBLAS.gemm_strided_batched!(C1, C2, T3(alpha), T1 == T3 ? A : T3.(A), T2 == T3 ? B : T3.(B), T3(beta), C)
I was trying to trying to find a AMD cloud machine to inspect the problem better, however, AWS rejected my request to increase the quota. :(
ok i pushed the code @GiggleLiu
@GiggleLiu and yeah i found what the problem was yesterday. i forgot to mention. i had to go to CUDAjl source code to find the problem to be honest, There alpha and beta are Number (meaning they can be boolean). so yeah now what needs to be done is that pull request to be approved and solve that gradient test that fails.
I merged the PR in AMDGPU.jl.
Looks great! I just merged the PR, thanks again.
No problem
Add support for AMDGPU