[onnx regression] Investigate onnx_ops errors on top of main #23189
Description
What happened?
Starting #22813 seven ONNX ops are failing.
| test | reason |
|---|---|
| test_gru_seq_length | hangs |
| test_rnn_seq_length | hangs |
| test_scan9_sum | hangs |
| test_stft_with_window | hangs |
| test_tfidfvectorizer_tf_batch_onlybigrams_skip0 | Does not generate any dispatches & hangs when run |
| test_tfidfvectorizer_tf_batch_onlybigrams_skip5 | Does not generate any dispatches & hangs when run |
| test_tfidfvectorizer_tf_batch_uniandbigrams_skip5 | Does not generate any dispatches & hangs when run |
While the last three do not generate any dispatches, there is a valid vmfb file and iree-run-module will attempt to run it.
Steps to reproduce:
Go to the appropriate directory and execute the following commands
iree-compile model.mlir --iree-hal-target-device=hip --iree-hip-target=${GPU} --iree-input-demote-f64-to-f32=false --iree-opt-level=O3 -o model_gpu_hip_rdna3.vmfb
iree-run-module --device=hip --module=model_gpu_hip_rdna3.vmfb --flagfile=run_module_io_flags.txt
Below are the --mlir-print-ir-after-all output of all the tests above.
test_gru_seq_length.txt
test_rnn_seq_length.txt
test_scan9_sum.txt
test_stft_with_window.txt
test_tfidfvectorizer_tf_batch_onlybigrams_skip0.txt
test_tfidfvectorizer_tf_batch_onlybigrams_skip5.txt
test_tfidfvectorizer_tf_batch_uniandbigrams_skip5.txt
See here for each of the model's MLIR after AutoInputConversionPipelinePass https://gist.github.com/amd-eochoalo/4dec6eea1a449ea577fd6e9f357c7b56