When the model uses blfloat16 ops, the optimizer fails with the following. We should handle custom types form onnx in _constant_folding
Traceback (most recent call last):
File "/workspace/ONNXConverter/llama.py", line 36, in <module>
onnxscript.optimizer.optimize_ir(program.model)
File "/workspace/onnxscript/onnxscript/optimizer/__init__.py", line 124, in optimize_ir
_constant_folding.fold_constants(model, onnx_shape_inference=onnx_shape_inference)
File "/workspace/onnxscript/onnxscript/optimizer/_constant_folding.py", line 719, in fold_constants
folder.visit_model(model)
File "/workspace/onnxscript/onnxscript/optimizer/_constant_folding.py", line 699, in visit_model
self.visit_graph(model.graph)
File "/workspace/onnxscript/onnxscript/optimizer/_constant_folding.py", line 690, in visit_graph
self.visit_node(node, graph)
File "/workspace/onnxscript/onnxscript/optimizer/_constant_folding.py", line 679, in visit_node
replacement = self.process_node(node)
^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/onnxscript/onnxscript/optimizer/_constant_folding.py", line 650, in process_node
replacement = self.new_constant(node.outputs[0], outputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/onnxscript/onnxscript/optimizer/_constant_folding.py", line 580, in new_constant
irvalue.const_value = _convenience.tensor(value)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/onnxscript/onnxscript/ir/_convenience.py", line 357, in tensor
tensor_ = _core.Tensor(value, dtype=dtype, name=name, doc_string=name)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/onnxscript/onnxscript/ir/_core.py", line 355, in __init__
self._dtype = _enums.DataType.from_numpy(value.dtype)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/onnxscript/onnxscript/ir/_enums.py", line 76, in from_numpy
raise TypeError(f"Unsupported numpy data type: {dtype}")
TypeError: Unsupported numpy data type: (numpy.uint16, [('bfloat16', '<u2')])
justinchuby
commented
1 week ago
When the model uses blfloat16 ops, the optimizer fails with the following. We should handle custom types form onnx in
_constant_folding