[ONNX] Add converter for QAttention from Microsoft onnxruntime contrib opset #13654
Conversation
|
Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment. Generated by tvm-bot |
…atmul_tensorcore_cuda
|
Hello @vvchernov, @echuraev, @AndrewZhaoLuo! |
echuraev
left a comment
echuraev
left a comment
There was a problem hiding this comment.
LGTM. But I don't have a lot of knowledge in this codebase. @jwfromm, @AndrewZhaoLuo could you please take a look at this PR?
| # Currently only (batch, past_seq_length + seq_length) shape is supported. | ||
| mask_index = inputs[5] | ||
|
|
||
| # Scalar, which means a per-tensor/layer quantization |
There was a problem hiding this comment.
nit: You have absolutely the same comment for input[3] and input[7]
python/tvm/relay/frontend/onnx.py
Outdated
| result, | ||
| _op.multiply(lhs_scale, rhs_scale), | ||
| zero_point_zero, | ||
| axis=-1, # TODO(agladyshev): what is 'axis' parameter for? |
There was a problem hiding this comment.
Do you still need this todo comment?
|
Apologies, I've been quite sick. I'll try to look at this Thursday. |
AndrewZhaoLuo
left a comment
AndrewZhaoLuo
left a comment
There was a problem hiding this comment.
Thanks! Sorry for late review. LGTM
…b opset (apache#13654) * init QAttention converter * add type and shape checking * add test for QAttention * add tests for optional parameters * change mask_index shape * add support for 'past' input * add support for 'unidirectional' attribute * expand test coverage * fix lint * fix pylint * fix batch dimension for topi/cuda/batch_matmul_tensorcore.py::batch_matmul_tensorcore_cuda * code review fix
4 participants
This PR adds support for QAttention - quantized version of Attention from Microsoft onnxruntime contrib opset.
An explanation and illustration of how this layer works can be found, for example, in @lena-voita NLP course.