[Bug] 17 dynamo test failures across all GPUs on CUDA 13.x — SDP use_fp32_acc, BERT accuracy, AutomaticPlugin, refit

Bug Description

17 tests fail across the dynamo test suites on all tested GPU architectures and CUDA 13.x versions. The failures group into four distinct categories, suggesting multiple independent regressions.

Environment

• GPUs: RTX 3070, B100-TS2, H100, A100
• Arch: x86_64
• CUDA: 13.2.0 / 13.1.1 / 13.0.2
• OS: Ubuntu 24.04
• cuDNN: 8.9.7.29
• TensorRT: 10.16.0.59
• Myelin: 2.17.78+7
• CASK: 5.16.17+1
• Python: 3.12
• Package: qa_tar_py3.12

Failure Categories

1. scaled_dot_product_attention — unexpected keyword argument use_fp32_acc (8 tests)

FAILED lowering/test_decompositions.py::TestLowering::test_lowering_scaled_dot_product_attention_0
FAILED lowering/test_decompositions.py::TestLowering::test_lowering_scaled_dot_product_attention_1
FAILED lowering/test_decompositions.py::TestLowering::test_lowering_scaled_dot_product_cudnn_attention_0
FAILED lowering/test_decompositions.py::TestLowering::test_lowering_scaled_dot_product_cudnn_attention_1
FAILED lowering/test_decompositions.py::TestLowering::test_lowering_scaled_dot_product_efficient_attention_0
FAILED lowering/test_decompositions.py::TestLowering::test_lowering_scaled_dot_product_efficient_attention_1
FAILED lowering/test_decompositions.py::TestLowering::test_lowering_scaled_dot_product_flash_attention_0
FAILED lowering/test_decompositions.py::TestLowering::test_lowering_scaled_dot_product_flash_attention_1

Error: TypeError: scaled_dot_product_attention() got an unexpected keyword argument 'use_fp32_acc'

The use_fp32_acc kwarg may have been removed or renamed in the current PyTorch/CUDA version.

2. scaled_dot_product_attention with dynamic shape — output mismatch (1 test)

FAILED lowering/test_decompositions.py::TestLowering::test_lowering_scaled_dot_product_attention_with_dynamic_shape_0

Error: AssertionError: Scaled_dot_product_attention_with_dynamic_shape TRT outputs don't match with the original model.

3. BERT base-uncased — accuracy regression and dtype issue (5 tests)

FAILED models/test_models.py::test_bert_base_uncased[dtype0] - Cosine sim: 0.5738 (threshold: 0.99)
FAILED models/test_models.py::test_bert_base_uncased[dtype1] - TypeError: Unsupported numpy dtype
FAILED models/test_models.py::test_bert_base_uncased[dtype2] - Cosine sim: 0.5090 (threshold: 0.99)
FAILED models/test_models.py::test_bert_base_uncased_cpu_offload - Cosine sim: 0.4028 (threshold: 0.99)
FAILED models/test_models_export.py::test_bert_base_uncased - Cosine sim: 0.4203 (threshold: 0.99)

Cosine similarity scores are far below the 0.99 threshold, indicating significant numerical divergence in the compiled BERT model.

4. AutomaticPlugin — array conversion error (2 tests)

FAILED automatic_plugin/test_automatic_plugin_with_attrs.py::TestAutomaticPlugin::test_scale_mul_plugin_float_0
FAILED automatic_plugin/test_automatic_plugin_with_attrs.py::TestAutomaticPlugin::test_scale_mul_plugin_float_1

Error: TypeError: only 0-dimensional arrays can be converted to Python scalars

5. Refit cumsum fallback — missing PyTorch segment (1 test)

FAILED models/test_model_refit.py::test_refit_cumsum_fallback

Error: AssertionError: False is not true : test_refit_cumsum_fallback test found 0 pytorch segments but expected 1

Reproducible Configurations

GPU	CUDA	Test Suites Affected
RTX 3070/x86_64	r13.0.2, r13.1.1, r13.2.0	backend, conversion, models, partitioning, runtime
A100/x86_64	r13.0.2, r13.1.1, r13.2.0	backend, conversion, models, partitioning, runtime
H100/x86_64	r13.1.1, r13.2.0	runtime
B100-TS2/x86_64	r13.0.2, r13.1.1, r13.2.0	runtime

Overall Test Results

17 failed, 2317 passed, 22 skipped, 2 xpassed, 4088 warnings in 3644.85s (1:00:44)

Steps to Reproduce

1. Run on any of the listed GPUs with CUDA 13.x and the environment above
2. Execute the dynamo test suites:

   pytest lowering/test_decompositions.py
   pytest models/test_models.py
   pytest models/test_models_export.py
   pytest models/test_model_refit.py
   pytest automatic_plugin/test_automatic_plugin_with_attrs.py
   

Start by running the five listed pytest commands in the Python 3.12/CUDA 13.x environment and group failures by category. Read the named tests in lowering/test_decompositions.py, models/test_models.py, models/test_models_export.py, models/test_model_refit.py, and automatic_plugin/test_automatic_plugin_with_attrs.py. Done means the 17 failures are resolved and the BERT similarity and refit expectations pass.