Merge pull request #18 from LuxDL/ap/amdgpu

Test AMDGPU

.buildkite/pipeline.yml

@@ -21,8 +21,37 @@ steps:
       setup:
         julia:
           - "1"
-          - "1.6"
-          - "1.9-nightly"
+          - "nightly"
+      adjustments:
+        - with:
+            julia: "nightly"
+          soft_fail: true
+  - label: ":julia: Julia: {{matrix.julia}} + AMD GPU"
+    plugins:
+      - JuliaCI/julia#v1:
+          version: "{{matrix.julia}}"
+      - JuliaCI/julia-test#v1:
+          test_args: "--quickfail"
+      - JuliaCI/julia-coverage#v1:
+          codecov: true
+          dirs:
+            - src
+            - ext
+    env:
+      JULIA_AMDGPU_CORE_MUST_LOAD: "1"
+      JULIA_AMDGPU_HIP_MUST_LOAD: "1"
+      JULIA_AMDGPU_DISABLE_ARTIFACTS: "1"
+      GROUP: "AMDGPU"
+    agents:
+      queue: "juliagpu"
+      rocm: "*"
+      rocmgpu: "*"
+    if: build.message !~ /\[skip tests\]/
+    timeout_in_minutes: 60
+    matrix:
+      setup:
+        julia:
+          - "1"
           - "nightly"
       adjustments:
         - with:

.github/workflows/CI.yml

@@ -19,8 +19,6 @@ jobs:
       matrix:
         version:
           - "1"
-          - "1.6"
-          - "~1.9.0-0"
     steps:
       - uses: actions/checkout@v3
       - uses: julia-actions/setup-julia@v1
@@ -46,4 +44,3 @@ jobs:
       - uses: codecov/codecov-action@v3
         with:
           files: lcov.info
-          flags: ${{ matrix.group }}

Project.toml

@@ -1,7 +1,7 @@
 name = "LuxLib"
 uuid = "82251201-b29d-42c6-8e01-566dec8acb11"
 authors = ["Avik Pal <avikpal@mit.edu> and contributors"]
-version = "0.2.1"
+version = "0.2.2"
 
 [deps]
 ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"

ext/LuxLibTrackerExt.jl

@@ -102,10 +102,12 @@ end
     epsilon::Real) where {T <: FP_32_64}
     LuxLib._assert_same_backend(data(x), data(scale), data(bias))
     if length(scale) != length(bias) != size(x, 3)
-        throw(ArgumentError("Length of `scale` and `bias` must be equal to the number of channels (N - 1 dim of the input array)."))
+        throw(ArgumentError("Length of `scale` and `bias` must be equal to the number of \
+                             channels (N - 1 dim of the input array)."))
     end
     if size(x, 3) % groups != 0
-        throw(ArgumentError("Number of channels $(size(x, 3)) must be divisible by the number of groups $groups."))
+        throw(ArgumentError("Number of channels $(size(x, 3)) must be divisible by the \
+                             number of groups $groups."))
     end
 
     y, mu, rsig = LuxLib._groupnorm(data(x), groups, data(scale), data(bias), epsilon)

src/api/dropout.jl

@@ -66,9 +66,7 @@ function dropout(rng::AbstractRNG,
     ::Val{false};
     dims,
     invp::T=inv(p)) where {T, T1, T2, N}
-    if size(x) != size(mask)
-        return dropout(rng, x, p, Val(true); dims, invp)
-    end
+    size(x) != size(mask) && return dropout(rng, x, p, Val(true); dims, invp)
     return x .* ignore_derivatives(mask), mask, rng
 end
 

test/Project.toml

@@ -1,6 +1,7 @@
 [deps]
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 JET = "c3a54625-cd67-489e-a8e7-0a5a0ff4e31b"
+LuxAMDGPU = "83120cb1-ca15-4f04-bf3b-6967d2e6b60b"
 LuxCUDA = "d0bbae9a-e099-4d5b-a835-1c6931763bda"
 LuxTestUtils = "ac9de150-d08f-4546-94fb-7472b5760531"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"

test/api/dropout.jl

@@ -1,5 +1,4 @@
-using LuxCUDA, Statistics, Test
-using LuxLib
+using Statistics, Test, LuxLib
 
 include("../test_utils.jl")
 
@@ -145,6 +144,7 @@ end
         @test y isa aType{T, length(x_shape)}
         @test size(y) == x_shape
         @test rng != rng_
+
         @test_broken isapprox(std(y), std(x); atol=1.0f-2, rtol=1.0f-2)
 
         __f = x -> sum(first(alpha_dropout(rng, x, T(0.5), Val(true))))

test/test_utils.jl

@@ -1,21 +1,23 @@
 using LuxLib, LuxTestUtils, StableRNGs, Test, Zygote
-using LuxCUDA  # CUDA Support
+using LuxCUDA, LuxAMDGPU
 using LuxTestUtils: @jet, @test_gradients, check_approx
 
 const GROUP = get(ENV, "GROUP", "All")
 
 cpu_testing() = GROUP == "All" || GROUP == "CPU"
 cuda_testing() = (GROUP == "All" || GROUP == "CUDA") && LuxCUDA.functional()
-amdgpu_testing() = (GROUP == "All" || GROUP == "AMDGPU") # && LuxAMDGPU.functional()
+amdgpu_testing() = (GROUP == "All" || GROUP == "AMDGPU") && LuxAMDGPU.functional()
 
 const MODES = begin
     # Mode, Array Type, GPU?
     cpu_mode = ("CPU", Array, false)
     cuda_mode = ("CUDA", CuArray, true)
+    amdgpu_mode = ("AMDGPU", ROCArray, true)
 
     modes = []
     cpu_testing() && push!(modes, cpu_mode)
     cuda_testing() && push!(modes, cuda_mode)
+    amdgpu_testing() && push!(modes, amdgpu_mode)
     modes
 end