There is currently no int16 quantization support in torch. This patch
adds a new mlir type to correspond to the missing "torch.qint16" type,
and enables lowering of quantization-related onnx ops using int16 types.
In follow-up patches, custom quantization logic for ops like
aten.matmul/aten.mm/aten.convolution may need to be revisited to allow
support for qint16. The passes in FuseQuantizedOps.cpp may also need
slight modifications.
This addresses 7 of the model failures I'm seeing in the test suite. See
[Shark-Turbine issue
#566](https://github.com/nod-ai/SHARK-Turbine/issues/566).
Need the op ```linalg.conv_2d_ngchw_gfchw_q``` to be added upstream
before merging this. See [llvm-project PR #92136
](https://github.com/llvm/llvm-project/pull/92136).
A small additional expansion to operand quantization is included in this
patch to address a model failure that occurs when unblocking the
quantized group convolutions in one of these onnx models.
Updates:
- some unsupported modes are now going to report a match failure for
unsupported coordinate transformation modes.
- fixes a bug that was introduced in the last patch for resize (my
bad...)
- uses actual x and y coordinates for computing weights in bilinear
interpolation (rather than eps modified values)
- slightly simplifies the bilinear interpolation payload for readability
and performance
- passes coordinate transformation mode information from an onnx.Resize
op to the mode string for the aten._interpolate op. This allows us to
perform custom logic in the torch->linalg lowering to support
onnx.Resize options without losing the default behaviors of the
interpolate op.
The old lowering only had logic for 2d (i.e. images). this patch allows
interpolation for n spatial dims, which is required for some 3d vision
models such as
- onnx/models/pytorch-3dunet_vaiq_int8
which successfully compiles and runs with this patch.
* Enables assume_strict_symbolic_shapes on fx_importer imported
programs, indicating strict shape semantics.
* Reworks the view->reshape lowering to take advantage of strict mode
and do one of:
* Collapse to 0D
* Flatten/Unflatten when there is an inferred dim.
* Fallback to tensor.reshape
* Splits some test cases up and adds an attribute to control the old
pattern (so new corners can be tested in strict mode in isolation).
* Dynamic inferred mode needs upstream work to generalize expand_shape
(so that case is suppressed here).
* Deletes the assert from the existing tensor.reshape lowering if strict
shape mode is enabled (since the condition it is dynamically asserting
cannot happen).
I spent a little while debugging numerics issues with some tests similar
to the ones in quantized_models.py, only to find that pytorch's
quantized conv transpose is catastrophically inaccurate. I'll upstream
the issue and only leave the tests here which are of the form quantize
-> dequantize -> op.
Sparse tensor conversions are represented by special aten operators.
This PR ensures the conversions are recognized (instead of failing the
full torch aten lowering to linalg).
A choice was made to quantize the return type of Relu with a scale and
zero point copied from the input's quantization scheme. With this
choice, the torch-to-linalg conversion of quantized Relu essentially
computes max(input, zeroPoint) in the elementwise payload.
We can map to `tensor.reshape` for handling multiple output dynamic
shapes. Later we can perform a more complex analysis for indentifying
expand/collapse cases from the tensor.reshape.
Initially we planned to handle this identification at the `torch` level
however it will be easier to handle once converted to core
mlir-dialects.
Align corner modes which select what the corners mean.
Either the center of the corner points or the edges of the edge points.
---------
Co-authored-by: Rob Suderman <rob.suderman@gmail.com>
The new cases added for quantized matmuls are:
1. vec-vec
2. vec-mat
3. mat-vec
each of which are now lowered to expand(s), quantized_matmul, and
collapse.
1. onnx.MatMulInteger now converts to aten.matmul instead of aten.mm
2. aten.matmul, for ranks >=2, now allows quantized inputs and will
lower to linalg::quantized_matmul or linalg::quantized_batch_matmul.
3. added AtenMatmulOp to the FuseQuantizeOps rewrite patters
QuantizeOperands, QuantizeTransposedOperands, and QuantizeAccumulator
4. added several tests, including some to test AtenMmOp with varying
quantization signed-ness.
5. a quantized matmul mat-vec test is added to verify the failure to
lower to linalg; cleaned of out-of-date code related to common
torch-mlir lowering xfails.
6. in debugging a real model with quantized matmuls, I found a bug on
the scalarize-shapes pass which resulted from the aten.full op folder
returning an incompatible result type. This is fixed by the small change
here to
[lib/Dialect/Torch/IR/TorchOps.cpp](https://github.com/llvm/torch-mlir/compare/main...zjgarvey:torch-mlir:MatMulIntegerFix?expand=1#diff-dc8ed165c207918e606490eee3984b1ad51d7034e6aac36fc046bf47f6f03f4f).
- Added linalg lowering for `AtenFloorDivideScalarOp`
- Needed `AtenDivScalarModeOp` for the decomp.
- Added linalg lowering for `AtenDivScalarModeOp`
- Moved linalg payload logic to `createDivModePayload()` since the logic
was nearly identical for both `AtenDivScalarModeOp` and
`AtenDivTensorModeOp`. Just a template function
- Added `AtenDivScalarModeOp` lowering for stablehlo
Pytorch's
[`torch.floor_divide()`](https://pytorch.org/docs/stable/generated/torch.floor_divide.html)
in a previous version (for a reason unknown to me) preformed a
truncation instead of "floor". The already implemented op
`AtenFloorDivideTensorOp` was done before this change. However, this
wasn't caught because our testcases only tested positive floor division.
I changed this to floor as well as adding a few test cases.
1. Changes the linalg lowering for dequantization ops to always sign
cast to float to prevent misrepresenting uint32 overflow on subtraction
with zero point.
2. Adds a basic quantized model test which only quantizes and
dequantizes and now passes with these changes in linalg and onnx
configs.
3. Changes the aten.mm lowering to allow mismatched quantized types.
4. If a quantized matmul arg is uint8, we shift by 128 to faithfully
represent the quantization as a signed i8 quantization. This worked fine
in the AtenMmOp lowering, but I'd be happy to move it to a rewrite in
FuseQuantizedOps.cpp instead if that seems more appropriate.
With the changes 3 and 4, the QuantizedMLP_basic and
QuantizedSingleLayer_basic e2e tests now passes with the onnx config.
Squeezes can be ambiguous without the output shape information. For
instance (1, 1, 256) squeezed can be either (1, 256) or (256). We need
to check the resulting shape to know what the shape should look like.
Shapes can be processed as tensors to represent the set of dimensions.
As reshapes take a list of scalars this can result in a single dynamic
dimension blocking the adjacent static dimensions.
This pass attempts to de-couple tensor computations related to shapes
and propagate values to better support lowering scalar tensor
computations.
The `convertTensorToElementType` function expects it's argument to have
a valid tensor type that is not `Torch::NoneType`. This PR checks that
the bias tensor is not of type `Torch::NoneType` before calling
`convertTensorToElementType` on the bias tensor argument in the
`matchAndRewrite` member function of the `ConvertAtenConvolutionOp`
class.
See the related issues here:
[SHARK-Turbine#556](https://github.com/nod-ai/SHARK-Turbine/issues/556)
1. Adds uint8 casting to onnx.Cast op
2. Fixes an issue with onnx.DequantizeLinear when the scale comes with
shape [1].
3. Adds support for unsigned types in an AtenItemOp folder
4. Adds a simpler quantized model for easier debugging
5. Adds a fusion pass to convert [quant -> dequant -> transpose -> mm]
patterns to [transpose -> quant -> mm].
6. Moved some xfails that are still not passing, but for different
reasons than onnx.cast failures.
Reshaping tensors depend on directly matching individual dimensions to
their corresponding dim in the `torch.view` reshape dimensions. This
involves decoupling dynamic dimensions from their static counterparts
and support cleanup / canonicalization.
zjgarvey
Matthias Gehre
Aart Bik
Yuanqiang Liu
aldesilv
penguin_wwy
Vivek Khandelwal
Rob Suderman
Gaurav Shukla
Peiming Liu
Stella Laurenzo
Andreas Falkenberg
NeverRaR
Xinyu Yang
Avinash Sharma
IanWood1
Xinan Jiang(姜曦楠)
Thomas Dietert
ptrifunovic98
schnkmwt