[Bug] TestWeightStrippedEngine OOM on RTX 3070 (8GB) with CUDA 12.x

[apbose]

Bug Description

TestWeightStrippedEngine tests fail with CUDA out-of-memory errors on GPUs with 8GB VRAM (RTX 3070). The tests attempt to allocate memory that exceeds available GPU capacity.

Environment

• GPU: RTX 3070 (7.67 GiB)
• CUDA: 12.8.1 / 12.6.3
• OS: Ubuntu 22.04 / 24.04
• TensorRT (Myelin): 2.10.26+1103
• cuDNN: 8.9.7.29
• PyTorch: >=2.0
• ONNX: 1.16.0

Failing Tests

FAILED models/test_weight_stripped_engine.py::TestWeightStrippedEngine::test_two_TRTRuntime_in_refitting
FAILED models/test_weight_stripped_engine.py::TestWeightStrippedEngine::test_weight_stripped_engine_results
FAILED models/test_weight_stripped_engine.py::TestWeightStrippedEngine::test_weight_stripped_engine_sizes

Error

torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 50.00 MiB.
GPU 0 has a total capacity of 7.67 GiB ...

Additionally, test_weight_stripped_engine_results also hits an AssertionError (likely after a partial OOM or on retry).

Steps to Reproduce

1. Run on an RTX 3070 (8GB) with the environment listed above
2. Execute: pytest models/test_weight_stripped_engine.py

Expected Behavior

Tests should either:

1. Succeed on GPUs with 8GB VRAM, or
2. Be skipped with an appropriate @unittest.skipIf decorator when insufficient GPU memory is detected

Suggested Fix

Consider adding a memory guard or reducing model/input sizes for these tests so they fit within 8GB GPUs. Alternatively, mark them with a minimum VRAM requirement.