Add new INT4 quantization features to model builder #940
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
This PR adds new INT4 quantization features to the model builder.
MatMul
can now be quantized toMatMulNBits
via RTN.int4_op_type_to_quantize
has been added to allow more flexibility with INT4 quantization.Motivation and Context
With these PR changes, the size of the ONNX models can be reduced by quantizing the embedding layer and/or the language modeling head.
For the ONNX models built from already-quantized PyTorch models, one example is with using AutoAWQ. AutoAWQ does not quantize the language modeling head. The resulting ONNX model typically contains a
MatMul
op for the language modeling head. Now, thatMatMul
op will be quantized via RTN toMatMulNBits
to reduce memory.