models icon indicating copy to clipboard operation
models copied to clipboard
verified publisher icon NVIDIA-Merlin

Fix XLA support in Models, benchmark it and document

Open gabrielspmoreira opened this issue 2 years ago • 0 comments

  • [ ] Fix the issue that XLA hits graph execution errors when enabled either a model.compile argument or TF_XLA_FLAGS env var.
  • [ ] Benchmark training time with and without XLA enabled
  • [ ] Document how our users can enable XLA for Merlin Models

Context: @vysarge found these issues trying to train a DLRM model with XLA enabled.

Matrix size-incompatible: In[0]: [65536,857], In[1]: [479,1024] [[{{node model/sequential_block_8/sequential_block_4/private__dense_7/dense_8/MatMul}}]]

gabrielspmoreira avatar Mar 24 '23 15:03 gabrielspmoreira