Use linux.g4dn.4xlarge.nvidia.gpu for cuda 12.4 legacy driver tests (pytorch#164172)

Use linux.g4dn.4xlarge.nvidia.gpu for cuda 12.4 legacy driver tests (pytorch#163956)

Workaround for pytorch#163658

Looks like the workflow passes on 12.8 build that use inux.g4dn.4xlarge.nvidia.gpu but its failing on 12.6 builds that use linux.4xlarge.nvidia.gpu: https://github.com/pytorch/pytorch/actions/runs/17953843505/job/51080623612#step:13:470

Pull Request resolved: pytorch#163956
Approved by: https://github.com/malfet


(cherry picked from commit 349c960)

Co-authored-by: atalman <[email protected]>
Co-authored-by: Mark Saroufim <[email protected]>

Author: 3 people

.github/workflows/periodic.yml

@@ -59,13 +59,14 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-cuda12.4-py3.10-gcc11
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11
+      cuda-arch-list: 7.5
       test-matrix: |
         { include: [
-          { config: "legacy_nvidia_driver", shard: 1, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
-          { config: "legacy_nvidia_driver", shard: 2, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
-          { config: "legacy_nvidia_driver", shard: 3, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
-          { config: "legacy_nvidia_driver", shard: 4, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
-          { config: "legacy_nvidia_driver", shard: 5, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu" },
+          { config: "legacy_nvidia_driver", shard: 1, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "legacy_nvidia_driver", shard: 2, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "legacy_nvidia_driver", shard: 3, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "legacy_nvidia_driver", shard: 4, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
+          { config: "legacy_nvidia_driver", shard: 5, num_shards: 5, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
         ]}
     secrets: inherit
 

test/test_cuda.py

@@ -7045,9 +7045,7 @@ def test_graph_external_wait_and_record(self):
         """
         from torch.cuda import _compile_kernel
 
-        spin_wait_kernel = _compile_kernel(
-            kernel_source, "wait_for_cpu", compute_capability="70"
-        )
+        spin_wait_kernel = _compile_kernel(kernel_source, "wait_for_cpu")
 
         x = torch.ones(4, device="cuda")
         x_cpu = torch.zeros(x.shape, device="cpu").pin_memory()