Potential fix for pull request finding

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>

Author: Aishwarya-Tonpe

superbench/benchmarks/micro_benchmarks/_export_torch_to_onnx.py

@@ -18,7 +18,6 @@
 
 class torch2onnxExporter():
     """PyTorch model to ONNX exporter."""
-
     def __init__(self):
         """Constructor."""
         from transformers import BertConfig, GPT2Config, LlamaConfig
@@ -360,7 +359,6 @@ def _build_vision_export_inputs(self, model, batch_size, model_dtype, device):
         dynamic_axes = {'pixel_values': {0: 'batch_size'}, 'output': {0: 'batch_size'}}
 
         class VisionModelWrapper(torch.nn.Module):
-
             def __init__(self, model):
                 super().__init__()
                 self.model = model
@@ -398,7 +396,6 @@ def _build_nlp_export_inputs(self, model, batch_size, seq_length, device):
         }
 
         class NLPModelWrapper(torch.nn.Module):
-
             def __init__(self, model):
                 super().__init__()
                 self.model = model

superbench/benchmarks/micro_benchmarks/huggingface_model_loader.py

@@ -154,6 +154,11 @@ def load_model(
         # Reject malformed / path-like identifiers before any network or disk activity.
         validate_model_identifier(model_identifier)
 
+        # Fall back to CPU on hosts without CUDA so default device='cuda' callers don't fail.
+        if device == 'cuda' and not torch.cuda.is_available():
+            logger.warning('CUDA not available; falling back to CPU.')
+            device = 'cpu'
+
         try:
             load_kwargs = self._build_load_kwargs(torch_dtype, revision, kwargs)
 

superbench/benchmarks/micro_benchmarks/tensorrt_inference_performance.py

@@ -23,6 +23,7 @@
 
 class TensorRTInferenceBenchmark(MicroBenchmarkWithInvoke):
     """TensorRT inference micro-benchmark class."""
+
     def __init__(self, name, parameters=''):
         """Constructor.
 
@@ -238,9 +239,10 @@ def _preprocess_huggingface_models(self):
             hf_config = AutoConfig.from_pretrained(
                 self._args.model_identifier, trust_remote_code=allow_remote_code, **load_kwargs
             )
-            precision_str = self._args.precision    # already a string: 'fp16', 'fp32', 'int8'
+            # ONNX export is always done in float32 (see _build_trtexec_command_for_hf), so gate
+            # the pre-download check on fp32 memory regardless of the requested runtime precision.
             fits, _, _, _ = HuggingFaceModelLoader.check_memory_fits(
-                self._args.model_identifier, hf_config, precision_str, mode='inference', token=hf_token
+                self._args.model_identifier, hf_config, 'fp32', mode='inference', token=hf_token
             )
             if not fits:
                 self._result.set_return_code(ReturnCode.MICROBENCHMARK_EXECUTION_FAILURE)

tests/benchmarks/micro_benchmarks/test_huggingface_e2e.py

@@ -28,7 +28,6 @@
 )
 class TestHuggingFaceE2E:
     """End-to-end tests for HuggingFace model loading."""
-
     @pytest.fixture
     def loader(self, tmp_path):
         """Create a loader instance with an isolated per-test cache dir."""