lints

.github/workflows/rocm_ci.yml

@@ -22,6 +22,9 @@ on:
 jobs:
   build:
     runs-on: self-hosted
+    concurrency:
+      group: ${{ github.workflow }}-${{ github.ref }}
+      cancel-in-progress: true
     if: github.repository == 'rocm/xformers' || github.event_name == 'pull_request'
     container:
       image: 'rocm/pytorch-nightly:latest'

.github/workflows/rocm_wheels.yml

@@ -4,6 +4,13 @@ on:
   push:
     branches:
       - develop
+  pull_request:
+    paths:
+      - "packaging/compute_wheel_version.sh"
+      - ".github/workflows/wheel*"
+      - ".github/actions/setup-windows-runner/action.yml"
+      - "setup.py"
+      - "requirements*.txt"
   workflow_dispatch:
 
 jobs:
@@ -24,8 +31,8 @@ jobs:
 
         include = []
         for os in ['ubuntu-alola']:
-          for python in TORCH_VERSIONS:
-            for torch_version in ['2.4.0']:
+          for python in PY_VERSIONS:
+            for torch_version in TORCH_VERSIONS:
               for toolkit_type, toolkit_short_versions in {'rocm': ["6.0", "6.1"]}.items():
                 for toolkit_short_version in toolkit_short_versions:
                   include.append(dict(
@@ -43,6 +50,9 @@ jobs:
 
   build:
     needs: target_determinator
+    concurrency:
+      group: ${{ github.workflow }}-${{ github.ref }}
+      cancel-in-progress: true
     strategy:
       fail-fast: false
       matrix: ${{ fromJson(needs.target_determinator.outputs.matrix) }}

setup.py

@@ -402,9 +402,8 @@ def get_extensions():
             "--ptxas-options=-O2",
             "--ptxas-options=-allow-expensive-optimizations=true",
         ]
-    elif (
-        (torch.cuda.is_available() and torch.version.hip) or
-        (os.getenv("FORCE_ROCM", "0") == "1")
+    elif (torch.cuda.is_available() and torch.version.hip) or (
+        os.getenv("FORCE_ROCM", "0") == "1"
     ):
         disable_hd256_hip_fmha = os.getenv("DISABLE_HD256_HIP_FMHA", "0")
         if disable_hd256_hip_fmha == "1":

tests/test_mem_eff_attention.py

@@ -37,13 +37,16 @@
 if torch.cuda.is_available():
     compute_capability = torch.cuda.get_device_capability("cuda")
 sm70_or_better_only = pytest.mark.skipif(
-    torch.version.cuda is not None and compute_capability < (7, 0), reason="requires sm70+"
+    torch.version.cuda is not None and compute_capability < (7, 0),
+    reason="requires sm70+",
 )
 sm75_or_better_only = pytest.mark.skipif(
-    torch.version.cuda is not None and compute_capability < (7, 5), reason="requires sm75+"
+    torch.version.cuda is not None and compute_capability < (7, 5),
+    reason="requires sm75+",
 )
 sm80_or_better_only = pytest.mark.skipif(
-    torch.version.cuda is not None and compute_capability < (8, 0), reason="requires sm80+"
+    torch.version.cuda is not None and compute_capability < (8, 0),
+    reason="requires sm80+",
 )
 skip_if_rocm = pytest.mark.skipif(
     torch.version.hip is not None, reason="not supported on ROCm"