apache/beam · error · ValueError

Failed to load ONNX file

Error message

Failed to load ONNX file: {onnx_path}

What it means

Raised in tensorrt_inference._load_onnx when the TensorRT OnnxParser fails to parse the ONNX file; the individual parser errors are logged first, then a ValueError naming the ONNX path is raised so engine building aborts early.

Solutions

1. Check the worker logs for the per-error output of parser.get_error to identify the failing op/opset
2. Re-export the model with an opset version supported by your TensorRT version (e.g. opset 13-17 for TRT 8.x)
3. Upgrade the TensorRT (and onnx/onnxruntime) versions in the worker image, or simplify/replace unsupported ops in the exported graph
4. Validate the file locally with onnx.checker.check_model before submitting the pipeline

Example fix

// before
# exported with opset 18, worker has TensorRT 8.2 (max opset 17)
// after
# re-export: torch.onnx.export(model, x, path, opset_version=17)

import onnx
m = onnx.load_model(onnx_path)
onnx.checker.check_model(m)  # raises on invalid/unsupported models
assert m.ir_version <= 8, 'ONNX IR version may exceed TensorRT support'

try:
    handler = TensorRTEngineHandlerNumpy(...)
except ValueError as e:
    if e.args and str(e).startswith('Failed to load ONNX file'):
        raise RuntimeError('Re-export ONNX with an opset supported by the worker TensorRT version; see logs for parser errors') from e
    raise

When it happens

Trigger: Loading an ONNX file whose format/IR version is unsupported by the installed TensorRT version, a corrupted/truncated file, or a model with unsupported operators — parser.parse returns False.

Common situations: ONNX exported by a newer opset than the TensorRT version supports; model exported with ops (e.g. custom or latest transformers ops) TRT cannot parse; downloading/serializing the file incorrectly (partial upload to GCS); mismatch between onnx and tensorrt package versions.

Understand the failure class

Background: "failed to read file", EACCES, ENOENT and "could not read <path>" errors: when a program can't read a file from disk — this error's family across 49 libraries.

Related errors

• Cuda Error
• inference_args were provided, but should be None because…
• Unable to load the TensorFlow model
• Callable create_model_fn must be passedwith…
• Cannot make make an unkeyed model handler with pre or…

AI-assisted analysis of apache/beam@12126d8942 (2026-09-13). Data as JSON: /api/errors/66b575f56acb866e. Report an issue: GitHub.

Appendix: source

Thrown at sdks/python/apache_beam/ml/inference/tensorrt_inference.py:72

  file = FileSystems.open(engine_path, 'rb')
  runtime = trt.Runtime(TRT_LOGGER)
  engine = runtime.deserialize_cuda_engine(file.read())
  assert engine
  return engine


def _load_onnx(onnx_path):
  import tensorrt as trt
  builder = trt.Builder(TRT_LOGGER)
  network = builder.create_network(
      flags=1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH))
  parser = trt.OnnxParser(network, TRT_LOGGER)
  with FileSystems.open(onnx_path) as f:
    if not parser.parse(f.read()):
      LOGGER.error("Failed to load ONNX file: %s", onnx_path)
      for error in range(parser.num_errors):
        LOGGER.error(parser.get_error(error))
      raise ValueError(f"Failed to load ONNX file: {onnx_path}")
  return network, builder


def _build_engine(network, builder):
  import tensorrt as trt
  config = builder.create_builder_config()
  runtime = trt.Runtime(TRT_LOGGER)
  plan = builder.build_serialized_network(network, config)
  engine = runtime.deserialize_cuda_engine(plan)
  builder.reset()
  return engine


def _assign_or_fail(args):
  """CUDA error checking."""
  from cuda import cuda
  err, ret = args[0], args[1:]
  if isinstance(err, cuda.CUresult):

View on GitHub (pinned to 12126d8942)