SaiShashank12 opened a new pull request, #36309:
URL: https://github.com/apache/beam/pull/36309

   Perfect — then you should explicitly reference it so the PR auto-closes it. 
Here’s the finalized PR description with your issue number plugged in:
   
   ---
   
   **Description**
   
   This PR introduces a new `TensorRTEngineHandlerNumPy` implementation that 
supports TensorRT **10.x** with the Tensor API (`num_io_tensors`, 
`get_tensor_name`, `execute_async_v3`, etc).
   It resolves runtime errors such as:
   
   * `AttributeError: 'ICudaEngine' object has no attribute 'num_bindings'`
   * `RuntimeError: This handler requires TensorRT 10.x+. Missing attributes on 
ICudaEngine: [...]`
   
   by replacing the legacy TensorRT < 10 API calls with the new Tensor API.
   
   It also fixes shape validation during dynamic batching and ONNX parsing 
(`onnx_path provided but build_on_worker=False`) to allow building engines 
directly on Beam workers if desired.
   
   **Addresses:** fixes #36306
   
   ---
   
   **Checklist**
   
   * [x] Mentioned the relevant issue (`fixes #36306`) so it will auto-close 
once merged.
   * [x] Added code that checks for TensorRT 10.x compatibility 
(`_require_tensorrt_10`) before engine load.
   * [x] Replaced `num_bindings` and `execute_async_v2` with `num_io_tensors` 
and `execute_async_v3`.
   * [x] Validated with dummy ResNet-50 ONNX pipeline on DirectRunner.
   * [x] Validated with dummy ResNet-50 ONNX pipeline on DataRunner.
   * [ ] Update `CHANGES.md` with a new entry under **Python SDK → ML 
Inference**: “Added TensorRT 10.x handler (`TensorRTEngineHandlerNumPy`).”
   * [ ] Write down the test cases.
   
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to