Fix PR test

.github/workflows/linter.yaml

@@ -19,7 +19,7 @@ jobs:
           path: tritonbench
       - name: Install deps
         run: |
-          pip install ruff-api
+          pip install ruff-api==0.1.0
       - name: Check Formatting
         uses: omnilib/ufmt@action-v1
         with:

.github/workflows/pr.yaml

@@ -27,9 +27,9 @@ jobs:
           sudo nvidia-smi -pm 1
           sudo ldconfig
           nvidia-smi
-      - name: Test Tritonbench operators
+      - name: Test Tritonbench operators on H100 GPU
         run: |
-          bash ./.ci/tritonbench/test-operators.sh
+          bash ./.ci/tritonbench/test-gpu.sh
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}

pyproject.toml

@@ -1,3 +1,7 @@
 [tool.ufmt]
 formatter = "ruff-api"
 excludes = ["submodules/"]
+
+[tool.black]
+line-length = 88
+target-version = ["py312"]

test/test_gpu/skip_tests_h100_pytorch.yaml

@@ -2,10 +2,23 @@
 # This file is regarding to the Triton version bundled with pytorch
 # Use <op-name> to skip an entire operator
 # Use <op-name/impl-name> to skip an impl
-- test_op
 - bf16xint16_gemm/bf16xint16
 - fp8_attention/colfax_fmha
 - fp8_fused_quant_gemm_rowwise
 - fp8_gemm/triton_persistent_fp8_gemm
 - fp8_gemm/triton_tma_persistent_fp8_gemm
 - fp8_gemm_rowwise
+- gemm
+- grouped_gemm
+- int4_gemm
+- jagged_layer_norm
+- jagged_mean
+- jagged_softmax
+- jagged_sum
+- layer_norm
+- low_mem_dropout
+- rms_norm
+- rope
+- swiglu
+- template_attention
+- test_op