[torch-bisect] fix repro cmdline bug and enhance with more cases (#172)

Author: xuzhao9

.github/scripts/bisect/run.sh

@@ -3,7 +3,7 @@
 set -euo pipefail
 
 readonly SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-readonly DETECTOR="${WORKSPACE_DIR}/.ci/bisect/regression_detector.py"
+readonly DETECTOR="${WORKSPACE_DIR}/.github/scripts/bisect/regression_detector.py"
 
 required_envs=(
   WORKSPACE_DIR
@@ -15,6 +15,7 @@ required_envs=(
   CUDA_HOME
   FUNCTIONAL
   REPRO_CMDLINE
+  USE_UV
 )
 
 for env_name in "${required_envs[@]}"; do
@@ -24,6 +25,11 @@ for env_name in "${required_envs[@]}"; do
   fi
 done
 
+if [ ! -e ${DETECTOR} ]; then
+    echo "Missing detector script: ${DETECTOR}."
+    exit 1
+fi
+
 checkout_pytorch_commit() {
   local repo_dir="$1"
   local commit="$2"
@@ -47,14 +53,15 @@ bash ${tritonparse_dir}/bisect/scripts/prepare_build_pytorch.sh
 BASELINE_LOG="${LOG_DIR}/baseline.log"
 checkout_pytorch_commit "${PYTORCH_SRC_DIR}" "${GOOD_COMMIT}"
 bash ${tritonparse_dir}/bisect/scripts/build_pytorch.sh
+cd ${PYTORCH_SRC_DIR}
 eval ${REPRO_CMDLINE} 2>&1 | tee "${BASELINE_LOG}"
 
 # step 3: build and run the bad commit
 checkout_pytorch_commit "${PYTORCH_SRC_DIR}" "${BAD_COMMIT}"
 bash ${tritonparse_dir}/bisect/scripts/build_pytorch.sh
 # allow the regression detector to exit with error code
 set +e
-BASELINE_LOG="${BASELINE_LOG}" python ./.ci/bisect/regression_detector.py
+BASELINE_LOG="${BASELINE_LOG}" python "${DETECTOR}"
 PREFLIGHT_RC=$?
 set -e
 
@@ -70,7 +77,7 @@ elif [ ${PREFLIGHT_RC} -ne 1 ] && [ ${FUNCTIONAL} -ne 1 ]; then
 fi
 
 # kick off the bisect!
-BASELINE_LOG="${BASELINE_LOG}" USE_UV=0 \
+BASELINE_LOG="${BASELINE_LOG}" USE_UV="${USE_UV}" \
 tritonparseoss bisect \
   --no-tui \
   --target torch \

.github/workflows/pytorch-bisect.yaml

@@ -49,14 +49,16 @@ jobs:
     name: Run PyTorch Bisect
     runs-on: ${{ inputs.runner }}
     env:
+      USE_UV: '1'
       WORKSPACE_DIR: ${{ github.workspace }}
       GOOD_COMMIT: ${{ inputs.good_commit }}
       BAD_COMMIT: ${{ inputs.bad_commit }}
+      REPRO_CMDLINE: ${{ inputs.repro_cmdline }}
       REGRESSION_THRESHOLD: ${{ inputs.regression_threshold }}
       PYTORCH_REPO: pytorch/pytorch
       FUNCTIONAL: ${{ inputs.type == 'functional' && '1' || '0' }}
-      REPRO_CMDLINE: ${{ inputs.command_arguments }}
       CUDA_HOME: ${{ github.workspace }}/torch-benchmarks/cuda
+      CUDA_VERSION: "13.0"
     permissions:
       id-token: write
       contents: read
@@ -69,7 +71,7 @@ jobs:
         with:
           repository: pytorch-labs/tritonparse
           path: torch-benchmarks/tritonparse
-          ref: xz9/add-torch-bisect
+          ref: xz9/fix-bisect
           fetch-depth: 0
 
       - name: Checkout PyTorch
@@ -80,6 +82,13 @@ jobs:
           submodules: recursive
           fetch-depth: 0
 
+      - name: Checkout torchbench
+        uses: actions/checkout@v4
+        with:
+          repository: pytorch/benchmark
+          path: torch-benchmarks/benchmark
+          fetch-depth: 0
+
       - name: Checkout torchvision
         uses: actions/checkout@v4
         with: