[ROCm] Guard triton backend call around cuda.is_available (pytorch#14…

…3570)

To resolve: pytorch/test-infra#6082

Calling into Triton's get_backend_options will initialise CUDA and break CPU-only environments that may have hip installed.

Pull Request resolved: pytorch#143570
Approved by: https://github.com/atalman, https://github.com/jeffdaily

torch/_inductor/kernel/conv.py

@@ -79,7 +79,7 @@ def conv3d_grid(n, c, d, h, w, meta):
 )
 
 # On ROCm convert num_stages to 1 as pipelining provides no benefit
-if torch.version.hip:
+if torch.version.hip and torch.cuda.is_available():
     platform_configs = build_rocm_gemm_configs(platform_configs)
 
 

torch/_inductor/kernel/mm_common.py

@@ -388,7 +388,7 @@ def filtered_configs(
 )
 
 # On ROCm convert num_stages to improve performance
-if torch.version.hip:
+if torch.version.hip and torch.cuda.is_available():
     mm_platform_configs = build_rocm_gemm_configs(mm_platform_configs)
     extra_mm_platform_configs = build_rocm_gemm_configs(extra_mm_platform_configs)
     int8_platform_configs = build_rocm_gemm_configs(int8_platform_configs)