when tracing into an operator, I get the following error. When code is inlined rather than called as operator, this passes
(this is not a workaround, though, because it interferes with down-stream IR manipulation of an easy-to-recognize operator"
File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/torch/_subclasses/fake_tensor.py", line 918, in _cached_dispatch_impl
output = self._dispatch_impl(func, types, args, kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/torch/_subclasses/fake_tensor.py", line 1402, in _dispatch_impl
r = func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/torch/_ops.py", line 610, in __call__
return self_._op(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/Users/runner/work/torchat/torchat/quantized_ops.py", line 52, in embedding_int8
rw_dtype = result_weights.to(dtype=scales.dtype)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Internal error: pybind11::error_already_set called while Python error indicator not set.
While executing %embedding_int8 : [num_users=3] = call_function[target=torch.ops.torchat.embedding_int8.default](args = (%arg72_1, %l__self___model_tok_embeddings_weight, %l__self___model_tok_embeddings_scales), kwargs = {})
Original traceback:
File "/Users/runner/work/torchat/torchat/export.py", line 60, in forward
logits = self.model(idx, input_pos)
File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
return forward_call(*args, **kwargs)
File "/Users/runner/work/torchat/torchat/model.py", line 208, in forward
x = self.tok_embeddings(idx)
File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1541, in _call_impl
return forward_call(*args, **kwargs)
File "/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
return func(*args, **kwargs)
File "/Users/runner/work/torchat/torchat/quantize.py", line 599, in forward
return torch.ops.torchat.embedding_int8(
Versions
This happens in the pytorch CI environment, with nightlies
cc @ezyang @msaroufim @bdhirsh @anijain2305 @chauhang
🐛 Describe the bug
https://github.com/pytorch/torchat/actions/runs/8643410590/job/23696389630?pr=136
when tracing into an operator, I get the following error. When code is inlined rather than called as operator, this passes (this is not a workaround, though, because it interferes with down-stream IR manipulation of an easy-to-recognize operator"
Versions
This happens in the pytorch CI environment, with nightlies
cc @ezyang @msaroufim @bdhirsh @anijain2305 @chauhang