From 60c48938b3e0fc86c496c800cf8915ee639c6742 Mon Sep 17 00:00:00 2001 From: "M. Chornyi" <99709299+mc-nv@users.noreply.github.com> Date: Fri, 31 Jul 2026 09:28:51 -0700 Subject: [PATCH] test: Skip Torch-TensorRT QA model generation on compute capability 10.7 The TensorRT Myelin autotuner has no maxpool tactic on compute capability 10.7, so building the resnet50 Torch-TensorRT engine fails with 'Autotuner: no tactics to implement operation' followed by 'Could not find any implementation for node {ForeignNode[...]}'. The generated PyTorch script runs under 'set -e', so the abort also skipped ragged, torchvision image and custom-ops model generation and left the whole model repository unpublished. Detect the compute capability with nvidia-smi and skip only the Torch-TensorRT step, so the rest of the repository is still produced. If nvidia-smi is unavailable the step runs as before. --- qa/common/gen_qa_model_repository | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/qa/common/gen_qa_model_repository b/qa/common/gen_qa_model_repository index 3dd8847930..5b9b047c8e 100755 --- a/qa/common/gen_qa_model_repository +++ b/qa/common/gen_qa_model_repository @@ -283,7 +283,11 @@ python3 $TRITON_MDLS_SRC_DIR/gen_qa_implicit_models.py --libtorch --variable --m chmod -R 777 $TRITON_MDLS_QA_VARIABLE_SEQUENCE_IMPLICIT_MODEL python3 $TRITON_MDLS_SRC_DIR/gen_qa_dyna_sequence_models.py --libtorch --models_dir=$TRITON_MDLS_QA_DYNA_SEQUENCE_MODEL chmod -R 777 $TRITON_MDLS_QA_DYNA_SEQUENCE_MODEL -if [ -z "$MODEL_TYPE" ] || [ "$MODEL_TYPE" != "igpu" ]; then +# TRI-1625: exclude Torch-TensorRT model generation on 10.7 compute resources. +COMPUTE_CAP=\$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader | head -1 | tr -d '[:space:]') +if [[ "\$COMPUTE_CAP" == "10.7" ]]; then + echo "Skipping Torch-TensorRT model generation (compute capability '\$COMPUTE_CAP')" +else python3 $TRITON_MDLS_SRC_DIR/gen_qa_torchtrt_models.py --models_dir=$TRITON_MDLS_QA_TORCHTRT_MODEL chmod -R 777 $TRITON_MDLS_QA_TORCHTRT_MODEL fi