Update pytorch.yml

This reverts commit 4b66b6d7be, reversing
changes made to ebb6f29b58.

.azuredevops/nightly/pytorch.yml

@@ -99,7 +99,7 @@ parameters:
   default:
     - rocminfo
     - MIOpen
-    - clr
+  #  - clr
     - hipBLAS
     - hipFFT
     - hipRAND
@@ -120,10 +120,11 @@ parameters:
     - rocm-core
     - rocPRIM
     # below are additional dependencies not called out by build script, but throw errors during cmake
-    - hipCUB
-    - rocThrust
-    - hipBLAS-common
     - composable_kernel
+    - hipBLAS-common
+    - hipCUB
+    - rocminfo
+    - rocThrust
 - name: rocmTestDependencies
   type: object
   default:
@@ -166,11 +167,11 @@ jobs:
   - template: /.azuredevops/variables-global.yml
 # various flags/parameters expected by bash scripts in pytorch repo's .ci directory
   - name: ROCM_VERSION
-    value: 6.3.0
+    value: 6.4.0
   - name: ROCM_PATH
     value: /opt/rocm
   - name: DESIRED_CUDA
-    value: 6.3.0
+    value: 6.4.0
   - name: MKLROOT
     value: /opt/intel
   - name: AOTRITON_INSTALLED_PREFIX
@@ -211,11 +212,36 @@ jobs:
       script: |
         sudo mkdir -p /opt/python/cp310-cp310/lib/python3.10
         sudo ln -s /usr/local/lib/python3.10/dist-packages /opt/python/cp310-cp310/lib/python3.10/site-packages
+  - task: DownloadPipelineArtifact@2
+    displayName: Download Specific HIP
+    inputs:
+      buildType: 'specific'
+      project: ROCm-CI
+      definition: 145
+      specificBuildWithTriggering: true
+      itemPattern: '**/*'
+      buildVersionToDownload: specific
+      targetPath: '$(Pipeline.Workspace)/d'
+      pipelineId: 16515
+  - task: ExtractFiles@1
+    displayName: Extract clr
+    inputs:
+      archiveFilePatterns: '$(Pipeline.Workspace)/d/**/*.tar.gz'
+      destinationFolder: '$(Agent.BuildDirectory)/rocm'
+      cleanDestinationFolder: false
+      overwriteExistingFiles: true
+  - task: DeleteFiles@1
+    displayName: Cleanup Compressed clr
+    inputs:
+      SourceFolder: '$(Pipeline.Workspace)/d'
+      Contents: '**/*.tar.gz'
+      RemoveDotFiles: true
   - template: ${{ variables.CI_TEMPLATE_PATH }}/steps/dependencies-rocm.yml
     parameters:
       dependencyList: ${{ parameters.rocmDependencies }}
       dependencySource: staging
       gpuTarget: $(JOB_GPU_TARGET)
+      setupHIPLibrarySymlinks: true
   - task: Bash@3
     displayName: ROCm symbolic link
     inputs:
@@ -226,8 +252,14 @@ jobs:
     displayName: git clone upstream pytorch
     inputs:
       targetType: inline
-      script: git clone https://github.com/pytorch/pytorch.git --depth=1 --recurse-submodules
+      script: git clone https://github.com/pytorch/pytorch.git --recurse-submodules
       workingDirectory: $(Build.SourcesDirectory)
+  - task: Bash@3
+    displayName: checkout pytorch 2.5
+    inputs:
+      targetType: inline
+      script: git checkout release/2.5
+      workingDirectory: $(Build.SourcesDirectory)/pytorch
 # builder clone still needed due to run_tests.sh at end of build_common.sh call
   - task: Bash@3
     displayName: git clone pytorch builder
@@ -271,6 +303,18 @@ jobs:
       targetType: inline
       script: sudo bash ./common/install_aotriton.sh /opt/rocm
       workingDirectory: $(Build.SourcesDirectory)/pytorch/.ci/docker
+ # - task: Bash@3
+ #   displayName: Temporarily Patch HIP
+ #   inputs:
+ #     targetType: inline
+ #     script: git apply $(Build.SourcesDirectory)/.azuredevops/patches/pytorch_hip_fp16.diff
+ #     workingDirectory: $(Agent.BuildDirectory)/rocm
+ # - task: Bash@3
+ #   displayName: Temporarily Patch CK Submodule
+ #   inputs:
+ #     targetType: inline
+ #     script: git pull origin develop
+ #     workingDirectory: $(Build.SourcesDirectory)/pytorch/third_party/composable_kernel
   - task: Bash@3
     displayName: Run ROCm Build Script
     inputs:
@@ -318,8 +362,14 @@ jobs:
     displayName: git clone pytorch vision
     inputs:
       targetType: inline
-      script: git clone https://github.com/pytorch/vision.git --depth=1 --recurse-submodules
+      script: git clone https://github.com/pytorch/vision.git --recurse-submodules
       workingDirectory: $(Build.SourcesDirectory)
+  - task: Bash@3
+    displayName: checkout release vision
+    inputs:
+      targetType: inline
+      script: git checkout release/0.20
+      workingDirectory: $(Build.SourcesDirectory)/vision
   - task: Bash@3
     displayName: Build vision
     inputs:

.azuredevops/patches/pytorch_hip_fp16.diff

@@ -0,0 +1,27 @@
+From 342133a5cb404beae4d7e1994338120ff99a76d2 Mon Sep 17 00:00:00 2001
+From: Jatin Chaudhary <JatinJaikishan.Chaudhary@amd.com>
+Date: Mon, 09 Dec 2024 11:24:29 +0000
+Subject: [PATCH] SWDEV-503299 - Do not use operator to check for nan
+
+Some libs use __HIP_NO_HALF_OPERATORS__ and __HIP_NO_HALF_CONVERSIONS__
+which results in operators being hidden and can cause errors.
+
+Change-Id: I83c194d7d727cba30b46d7c296f7d396549f5fca
+---
+
+diff --git a/include/hip/amd_detail/amd_hip_fp16.h b/include/hip/amd_detail/amd_hip_fp16.h
+index c8117b1..1a08bb8 100644
+--- a/include/hip/amd_detail/amd_hip_fp16.h
++++ b/include/hip/amd_detail/amd_hip_fp16.h
+@@ -1679,8 +1679,9 @@
+             __HOST_DEVICE__
+             bool __hisinf(__half x)
+             {
+-                // +Inf/-Inf
+-                return x == HIPRT_INF_FP16 || x == __ushort_as_half((unsigned short)0xFC00U);
++                __half_raw hr = x;
++                // +/-Inf
++                return hr.x == 0x7C00U || hr.x == 0xFC00U;
+             }
+             inline
+             __HOST_DEVICE__