Put torchao (0.13.0) back to benchmark workflow (#162227)

0.13.0 was released on Sep 3rd https://pypi.org/project/torchao/#history, which should have fixed the crashing issue on transformers now
Pull Request resolved: https://github.com/pytorch/pytorch/pull/162227
Approved by: https://github.com/malfet

.ci/docker/common/install_inductor_benchmark_deps.sh

@@ -43,4 +43,4 @@ install_huggingface
 install_timm
 
 # Clean up
-conda_run pip uninstall -y torch torchvision torchaudio triton torchao
+conda_run pip uninstall -y torch torchvision torchaudio triton

.ci/pytorch/macos-test.sh

@@ -181,9 +181,6 @@ checkout_install_torchbench() {
   popd
 
   pip install -r .ci/docker/ci_commit_pins/huggingface-requirements.txt
-  # https://github.com/pytorch/pytorch/issues/160689 to remove torchao because
-  # its current version 0.12.0 doesn't work with transformers 4.54.0
-  pip uninstall -y torchao
 
   echo "Print all dependencies after TorchBench is installed"
   python -mpip freeze

.github/ci_commit_pins/torchao.txt

@@ -1 +1 @@
-51c87b6ead6b7e098ada95d6a7609ee873b854cf
+f32431e593d0e9db86c502d3872dd67ee40a005f

.github/workflows/inductor-perf-test-nightly-h100.yml

@@ -137,7 +137,6 @@ jobs:
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
       timeout-minutes: 720
-      # disable monitor in perf tests, next step is to enable it
       disable-monitor: false
       monitor-log-interval: 15
       monitor-data-collect-interval: 4
@@ -154,7 +153,6 @@ jobs:
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
       timeout-minutes: 1440
-      # disable monitor in perf tests, next step is to enable it
       disable-monitor: false
       monitor-log-interval: 15
       monitor-data-collect-interval: 4
@@ -173,7 +171,6 @@ jobs:
       docker-image: ${{ needs.build.outputs.docker-image }}
       test-matrix: ${{ needs.build.outputs.test-matrix }}
       timeout-minutes: 720
-      # disable monitor in perf tests for more investigation
       disable-monitor: false
       monitor-log-interval: 15
       monitor-data-collect-interval: 4

.github/workflows/inductor-periodic.yml

@@ -36,6 +36,8 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-default-label-prefix
     with:
+      # More memory is needed to build torchao
+      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
@@ -128,6 +130,8 @@ jobs:
     needs:
       - get-default-label-prefix
     with:
+      # More memory is needed to build torchao
+      runner: linux.2xlarge.memory
       runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks

.github/workflows/inductor.yml

@@ -49,6 +49,8 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
+      # More memory is needed to build torchao
+      runner: linux.2xlarge.memory
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm86
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks
       cuda-arch-list: '8.6'