From 0a344578d6001b337020263a612910516e1eb71f Mon Sep 17 00:00:00 2001 From: Bruno Volpato Date: Mon, 2 Nov 2026 19:40:14 -0500 Subject: [PATCH] [CI] Disable TF32 for the L0 PyTorch numerics tests Transformer Engine selects TF32 for pure FP32 GEMM, and the PyTorch references in the numerics tests do not. Where cuBLAS uses TF32, the FP32 cases fail on correct code. These launchers already set NVIDIA_TF32_OVERRIDE=0 for test_mhc.py for the same reason. Set it for the numerics tests too, so that the L0 launchers give the same result outside the CI image. Signed-off-by: Bruno Volpato --- qa/L0_pytorch_debug_unittest/test.sh | 7 +++++-- qa/L0_pytorch_unittest/test.sh | 3 ++- 2 files changed, 7 insertions(+), 3 deletions(-) diff --git a/qa/L0_pytorch_debug_unittest/test.sh b/qa/L0_pytorch_debug_unittest/test.sh index 9e024e22e32..fddd86c9363 100644 --- a/qa/L0_pytorch_debug_unittest/test.sh +++ b/qa/L0_pytorch_debug_unittest/test.sh @@ -33,16 +33,19 @@ export NVTE_FLASH_ATTN_V4=0 pip install pytest==8.2.1 || error_exit "Failed to install pytest" +# The numerics tests compare FP32 results with PyTorch references that do not use TF32, while TE +# selects TF32 for pure FP32 GEMM, so these tests set NVIDIA_TF32_OVERRIDE=0. + pytest -v -s --junitxml=$XML_LOG_DIR/test_sanity.xml $TE_PATH/tests/pytorch/debug/test_sanity.py --feature_dirs=$NVTE_TEST_NVINSPECT_FEATURE_DIRS || test_fail "test_sanity.py" pytest -v -s --junitxml=$XML_LOG_DIR/test_config.xml $TE_PATH/tests/pytorch/debug/test_config.py --feature_dirs=$NVTE_TEST_NVINSPECT_FEATURE_DIRS || test_fail "test_config.py" -pytest -v -s --junitxml=$XML_LOG_DIR/test_numerics.xml $TE_PATH/tests/pytorch/debug/test_numerics.py --feature_dirs=$NVTE_TEST_NVINSPECT_FEATURE_DIRS || test_fail "test_numerics.py" +NVIDIA_TF32_OVERRIDE=0 pytest -v -s --junitxml=$XML_LOG_DIR/test_numerics.xml $TE_PATH/tests/pytorch/debug/test_numerics.py --feature_dirs=$NVTE_TEST_NVINSPECT_FEATURE_DIRS || test_fail "test_numerics.py" pytest -v -s --junitxml=$XML_LOG_DIR/test_log.xml $TE_PATH/tests/pytorch/debug/test_log.py --feature_dirs=$NVTE_TEST_NVINSPECT_FEATURE_DIRS --configs_dir=$NVTE_TEST_NVINSPECT_CONFIGS_DIR || test_fail "test_log.py" NVTE_TORCH_COMPILE=0 pytest -v -s --junitxml=$XML_LOG_DIR/test_api_features.xml $TE_PATH/tests/pytorch/debug/test_api_features.py --feature_dirs=$NVTE_TEST_NVINSPECT_FEATURE_DIRS --configs_dir=$NVTE_TEST_NVINSPECT_CONFIGS_DIR || test_fail "test_api_features.py" pytest -v -s --junitxml=$XML_LOG_DIR/test_perf.xml $TE_PATH/tests/pytorch/debug/test_perf.py --feature_dirs=$NVTE_TEST_NVINSPECT_FEATURE_DIRS --configs_dir=$NVTE_TEST_NVINSPECT_CONFIGS_DIR || test_fail "test_perf.py" # standard sanity and numerics tests with initialized debug NVTE_TEST_NVINSPECT_ENABLED=1 NVTE_TEST_NVINSPECT_CONFIG_FILE=$NVTE_TEST_NVINSPECT_DUMMY_CONFIG_FILE NVTE_TEST_NVINSPECT_FEATURE_DIRS=$NVTE_TEST_NVINSPECT_FEATURE_DIRS PYTORCH_JIT=0 NVTE_TORCH_COMPILE=0 NVTE_ALLOW_NONDETERMINISTIC_ALGO=0 pytest -v -s --junitxml=$XML_LOG_DIR/test_sanity_2.xml $TE_PATH/tests/pytorch/test_sanity.py || test_fail "debug test_sanity.py" -NVTE_TEST_NVINSPECT_ENABLED=1 NVTE_TEST_NVINSPECT_CONFIG_FILE=$NVTE_TEST_NVINSPECT_DUMMY_CONFIG_FILE NVTE_TEST_NVINSPECT_FEATURE_DIRS=$NVTE_TEST_NVINSPECT_FEATURE_DIRS PYTORCH_JIT=0 NVTE_TORCH_COMPILE=0 NVTE_ALLOW_NONDETERMINISTIC_ALGO=0 NVTE_FUSED_ATTN=0 pytest -v -s --junitxml=$XML_LOG_DIR/test_numerics_2.xml $TE_PATH/tests/pytorch/test_numerics.py || test_fail "debug test_numerics.py" +NVIDIA_TF32_OVERRIDE=0 NVTE_TEST_NVINSPECT_ENABLED=1 NVTE_TEST_NVINSPECT_CONFIG_FILE=$NVTE_TEST_NVINSPECT_DUMMY_CONFIG_FILE NVTE_TEST_NVINSPECT_FEATURE_DIRS=$NVTE_TEST_NVINSPECT_FEATURE_DIRS PYTORCH_JIT=0 NVTE_TORCH_COMPILE=0 NVTE_ALLOW_NONDETERMINISTIC_ALGO=0 NVTE_FUSED_ATTN=0 pytest -v -s --junitxml=$XML_LOG_DIR/test_numerics_2.xml $TE_PATH/tests/pytorch/test_numerics.py || test_fail "debug test_numerics.py" if [ "$RET" -ne 0 ]; then echo "Error in the following test cases:$FAILED_CASES" diff --git a/qa/L0_pytorch_unittest/test.sh b/qa/L0_pytorch_unittest/test.sh index bb459dad3b8..86e1a662461 100644 --- a/qa/L0_pytorch_unittest/test.sh +++ b/qa/L0_pytorch_unittest/test.sh @@ -34,7 +34,8 @@ python3 -m pytest --tb=auto --junitxml=$XML_LOG_DIR/pytest_test_sanity.xml $TE_P python3 -m pytest --tb=auto --junitxml=$XML_LOG_DIR/pytest_test_recipe.xml $TE_PATH/tests/pytorch/test_recipe.py || test_fail "test_recipe.py" python3 -m pytest --tb=auto --junitxml=$XML_LOG_DIR/pytest_test_custom_recipe.xml $TE_PATH/tests/pytorch/test_custom_recipe.py || test_fail "test_custom_recipe.py" python3 -m pytest --tb=auto --junitxml=$XML_LOG_DIR/pytest_test_deferred_init.xml $TE_PATH/tests/pytorch/test_deferred_init.py || test_fail "test_deferred_init.py" -PYTORCH_JIT=0 NVTE_TORCH_COMPILE=0 NVTE_ALLOW_NONDETERMINISTIC_ALGO=0 NVTE_FUSED_ATTN=0 python3 -m pytest --tb=auto --junitxml=$XML_LOG_DIR/pytest_test_numerics.xml $TE_PATH/tests/pytorch/test_numerics.py || test_fail "test_numerics.py" +# TE selects TF32 for pure FP32 GEMM, while the PyTorch reference in this test does not. +NVIDIA_TF32_OVERRIDE=0 PYTORCH_JIT=0 NVTE_TORCH_COMPILE=0 NVTE_ALLOW_NONDETERMINISTIC_ALGO=0 NVTE_FUSED_ATTN=0 python3 -m pytest --tb=auto --junitxml=$XML_LOG_DIR/pytest_test_numerics.xml $TE_PATH/tests/pytorch/test_numerics.py || test_fail "test_numerics.py" PYTORCH_JIT=0 NVTE_TORCH_COMPILE=0 NVTE_ALLOW_NONDETERMINISTIC_ALGO=0 NVTE_FUSED_ATTN=0 python3 -m pytest --tb=auto --junitxml=$XML_LOG_DIR/pytest_test_tied_delayed_wgrad.xml $TE_PATH/tests/pytorch/test_tied_delayed_wgrad.py || test_fail "test_tied_delayed_wgrad.py" PYTORCH_JIT=0 NVTE_TORCH_COMPILE=0 NVTE_ALLOW_NONDETERMINISTIC_ALGO=0 NVTE_FUSED_ATTN=0 python3 -m pytest --tb=auto --junitxml=$XML_LOG_DIR/pytest_test_cuda_graphs.xml $TE_PATH/tests/pytorch/test_cuda_graphs.py || test_fail "test_cuda_graphs.py" python3 -m pytest --tb=auto --junitxml=$XML_LOG_DIR/pytest_test_jit.xml $TE_PATH/tests/pytorch/test_jit.py || test_fail "test_jit.py"