Move separated Q, K and V (without input projection) from Attention to a
new operator CrossAttention.
The Attention operator is hard to maintain when we need support with and
without input projection in one class. Add a new operator according to
feedback.
Some change might need in the future, but not in this PR:
(1) bias could be optional (We will not proceed that route unless
experiments show that fusing Add bias with MatMul instead of this op
could improve performance).
(2) support packed KV. There are two ways to support it: when key and
value are same Tensor, they are packed; or we can make value as
optional, and use packed mode when value is empty and the key has packed
K/V.
(3) support cached key and value, and other (like relative position
bias), or more attention mask format. They can be added easily without
breaking backward compatible.
(4) ROCm/CPU implementation of this op.
### Description
1. The graph pattern search introduced in
https://github.com/microsoft/onnxruntime/pull/13914/ needs to be
enhanced so that SkipLayerNormalization is supported
2. Fix fp32 parity for GPT-2 while using `SkipLayerNormalization`
fusion. The optional output of SLN needs to also include the bias (if
present) and the added output should be a sum of `input + skip + (bias)`
### Motivation and Context
Fix some breaking tests
### Description
T5 uses a layer_norm which only scales and doesn't shift, which is also
known as Root Mean Square Layer Normalization.
ORT already have the simplified_layer_norm which is the RMS layer_norm.
This PR extends this T5 layer_norm with support of skip/bias and the
residual output.
This new op is named SkipSimplifiedLayerNorm and has similar interface
as SkipLayerNorm but removes the beta as input
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Co-authored-by: Ubuntu <wy@v100-2.0cdb2e52twzevn1i4fi45bylyg.jx.internal.cloudapp.net>
### Description
<!-- Describe your changes. -->
Sampling op for cpu and cuda
support huggingface case and custom case
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Co-authored-by: Ubuntu <wy@v100-2.0cdb2e52twzevn1i4fi45bylyg.jx.internal.cloudapp.net>
### Description
Update CUDA ArgMin/ArgMax op kernels to have end version 11 since opset
12+ is not supported yet.
With the way these kernels are currently registered, the documentation
shows support for opset 11+. This is not accurate.
### Motivation and Context
Fix#13781
This PR Registers the following operators for opset 16 to the DML EP:
- LeakyRelu-16
- PRelu-16
- Where-16
- GreaterOrEqual-16
- LessOrEqual-16
Identity-16 was not added in this PR due to pipeline failures
Co-authored-by: Numfor Mbiziwo-Tiapo <numform@microsoft.com>
Implement reuse kv_cache past and present tensor in Attention Ops.
Unit test for abover feature.
Utilize the reuse kv_cache for past and present tensor in Greedy Search.
Correctness test for it.
Co-authored-by: Zhang Lei <phill.zhang@gmail.com>
### Description
This PR adds support for `float64` kernels in the latest versions of
operators: Floor, Ceil and IsNaN.
### Motivation and Context
The lack of these kernels is non-trivial to work around and easily lead
to performance losses when it is attempted. When equivalence with an
existing implementation is required, precision is easily lost when
casting to `float32` instead.
IsNaN is common when cleaning up data in an ML pipeline. Floor and Ceil
have uses for discretising values and single-precision floats are
insufficient to round well when values get larger than a few million.
According to my measurement this only increases the binary size by a few
kilobytes (on the Python wheel of RelWithDebInfo).
Closes#13673 (Round already has float64 support)
Partially solves #8791 (Looks like there's parallel issues/PR open for
Split, but it is also hard to work around and hence useful)
Signed-off-by: jbachurski <kbachurski@gmail.com>
Split copies data - we can add support for all data types without too much binary size impact by using data type size-based implementations. The DispatchStridedCopy() function used here does this.
### Description
Add the NonZero op for DML
### Motivation and Context
NonZero is used in a few transformer models, so having a DML
implementation will stop large tensors from being transferred to the CPU
and back to the GPU
### Description
Add mixed datatype support for DML's LayerNorm contrib op.
### Motivation and Context
The fusion logic removes casts around LayerNorm in the graph because the
contrib version of the op supports mixed datatypes. Scale, Bias and
Output's datatypes must match, but input's datatype can be different.
### Description
Add a DML registration for Shape to avoid copying back to the CPU just
to get the shape of a GPU tensor.
### Motivation and Context
When using free dimensions, many Transformers models extensively use the
`Shape` operator. This causes hundreds of GPU->CPU copy that should be
completely avoidable. Note that this change also uses the same
heuristics as other providers (e.g. CUDA) to force some tensors on the
CPU in certain situations.
Co-authored-by: Patrice Vignola <pavignol@microsoft.com>
Some models have QuickGelu(x)=x*sigmoid(1.702x), which has 3 Ops for
forward and 5 Ops for backward. The PR is to fuse this to a single Op
named QuickGelu and its gradient QuickGeluGrad.
For CUDA, tested in V100 using input tensor with shape [64,128,2048] and
float16 type:
Before, FW takes 335us, BW takes 614us
After, FW takes 115us, BW takes 139us, which is much faster.
For CPU kernel, using same shape and float type:
Before, FW takes 10us, BW takes 49us
Mul: 3480[µs]
Sigmoid: 1996[µs]
Mul: 4789[µs]
Mul: 4642[µs]
Mul: 4195[µs]
SigmoidGrad: 18328[µs]
Mul: 2988[µs]
Sum: 18576[µs]
After, FW takes 4us, BW takes 5us, which is also much faster.
QuickGelu: 3939[µs]
QuickGeluGrad: 5089[µs]
Co-authored-by: Vincent Wang <weicwang@microsoft.com>
### Description
<!-- Describe your changes. -->
Fix document generation CI. It's not currently updating the docs as
we're skipping the tests, which is the invocation of build.py that would
have generated the documentation.
Setup specific task to generate documentation for greater clarity.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Operator kernel documentation is not getting updated and is now out of
date.
### Description
Allow separated Q, K and V inputs to support cross attention:
* Q: [batch_size, sequence_length, hidden_size]
* K: [batch_size, kv_sequence_length, hidden_size]
* V: [batch_size, kv_sequence_length, v_hidden_size]
* Output: [batch_size, sequence_length, v_hidden_size]
To use separated Q/K/V inputs, the input tensor is for query, and two
optional inputs are added for key and value. Weights for input
projection is not included for now, so the MatMul of input projection
shall be done out of Attention operator, but Add bias is included for
performance consideration.
### Description
<!-- Describe your changes. -->
Fix a bug in GreedySearch Op when batch > 1
Support custom attention mask in GreedySearch and BeamSearch with GPT2
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
# Motivation
Currently, ORT minimal builds use kernel def hashes to map from nodes to
kernels to execute when loading the model. As the kernel def hashes must
be known ahead of time, this works for statically registered kernels.
This works well for the CPU EP.
For this approach to work, the kernel def hashes must also be known at
ORT format model conversion time, which means the EP with statically
registered kernels must also be enabled then. This is not an issue for
the always-available CPU EP. However, we do not want to require that any
EP which statically registers kernels is always available too.
Consequently, we explore another approach to match nodes to kernels that
does not rely on kernel def hashes. An added benefit of this is the
possibility of moving away from kernel def hashes completely, which
would eliminate the maintenance burden of keeping the hashes stable.
# Approach
In a full build, ORT uses some information from the ONNX op schema to
match a node to a kernel. We want to avoid including the ONNX op schema
in a minimal build to reduce binary size. Essentially, we take the
necessary information from the ONNX op schema and make it available in a
minimal build.
We decouple the ONNX op schema from the kernel matching logic. The
kernel matching logic instead relies on per-op information which can
either be obtained from the ONNX op schema or another source.
This per-op information must be available in a minimal build when there
are no ONNX op schemas. We put it in the ORT format model.
Existing uses of kernel def hashes to look up kernels are replaced
with the updated kernel matching logic. We no longer store
kernel def hashes in the ORT format model’s session state and runtime
optimization representations. We no longer keep the logic to
generate and ensure stability of kernel def hashes.
* Fix bug in pybind get_all_operator_schema due to premature reference dropping
* Add updated operator kernels markdown table
* Update build.py to include documentation generation for DML operators too
* Update GPU pipeline to include DML in the build to so operators can be generated.
* Use a separate pipeline stage, feedback from Changming and Scott
* Appease annoying Python linter
* Add onnxruntime_BUILD_UNIT_TESTS=OFF and remove stale --use_dml in cuda stage
* Register signal ops for op set 17
Note code is mostly being moved, not added. These ops were previously
only registered as Microsoft contrib ops and only built if
`BUILD_MS_EXPERIMENTAL_OPS=1`. They've been added to the ai.onnx
standard op set in version 17.
Main components of this change:
* Move the kernels from the conrib_ops directory to the
core directory.
* Add function bodies for ms experimental ops. This will allow
old models that use the contrib ops to continue to function.
All the function bodies consist of a single op (the
new standard op), so performance overhead should be minimal.
Minor clean-up also in this change:
* De-duplicate get_scalar_value_from_tensor: put it in a new utils.h.
* Fix some bugs that caused compilation errors with the experimental
ops. Tested with `build.sh --ms_experimental`
* Fix some spelling errors and lint violations.
* Replace a couple of switch statements with `MLTypeCallDispatcher`.
* Use `InlineVector` instead of `std::vector`.
Unblocks https://github.com/microsoft/onnxruntime/issues/11640
* Implement BitmaskDropout and associated unit tests.
* Implement BitmaskDropoutGrad and associated unit tests.
* Implement Dropout -> BitmaskDropout rewrite rule and associated unit tests.
* Implement (Dropout,DropoutGrad) -> (BitmaskDropout,BitmaskDropoutGrad) rewrite rule.
This commit does not yet include unit tests for this rewrite rule.
This commit also introduces improved documentation for all changes which will be grouped
into this PR.
* bitmask dropout
* fix win build
* bugfix for rocm
* bugfix
* fix code format
* fix ut
* fix build break
* fix ut in win
* resolve comments
* fix ut in trt
* resolve comments
* fix rocm build error
* fix typo
Co-authored-by: Aidan Beggs <aidanbeggs@microsoft.com>
* Implement TreeEnsemble for opset(ai.onnx.ml)==3
* use of InlineVector
* refactoring
* improve attributes retrieval
* avoid creating a temporary buffer
* modifies onnx.ml.cpu.json
* use unordered_map
* update docs/OperatorKernels.md
* address PR comments (TH -> ThresholdType, ORT_RETURN...)
* add a python unit test to load a TreeEnsembleRegressor following ai.onnx.ml==3 specifications
* improve NonZero
* fix megatron_fp16 optimzier, fix the doc
* multi_tensor_applier
* resolve comment
* fix building warning
* fix build error when enabling training and use tensorrt
* Fix incorrect type constraint registration for RoiAlign. This led to the input type not actually being checked when matching a kernel as the invalid constraint name is treated as a missing optional input.
* fix missing dependency for the unit test exe. Whilst it doesn't link against the CUDA providers lib, without the dependency VS doesn't know it needs to rebuild the library if there are changes.
* Add check for invalid type constraints.
* Fix invalid registrations for other kernels.
* Add hash replacement logic to provide backwards compatibility in ORT format models when the registration is fixed.
* Add tests
* Changes to fuse embed layer for gpt2, kernal changes pending
* verified add output and regular add match
* Test added for additional output embedlayernorm, working on CUDA
* Test passing on CPU
* updated convert_to_onnx toll to check parity correctly
* removed some debugs
* couple of TODO left as in optimizer.py
* removed changes to optimizer.py
* fixing build
* fixing build
* updated order of initilization
* added a test case for float16
* updating the docs
* updating tests failing due to embed layer fusion
* update unit tests
* updating CUDA documentation in operatorkernels.md
* addressing comments
* OperatorKernels.md updated with CUDA
* adding TODO to qembed_layer
* minor edit
* updated docs
* addressing comments
* adding position ids to embed layer gpt2
* updating fused gpt2 model
* added extra test
* remove comments
* addressing comments
* contrib_defs.cc updated
* all tests passing
* fixing a typo
* minor edit
* trigger build
* qembedlayernorm checkinputs updated
* fixing build error
* fixing build error
* fixing build error
```
Component for aggressive decoding. Find the bifurcation index of predicted tokens, between source tokens,
starting from previous suffix match index, and predicted tokens.
Concat predicted tokens, starting from bifurcation index, to the back
of current tokens. This forms the output tokens.
Detect suffix match index in source tokens, between source tokens and output tokens.
Detection is based on finding the appearances of last n-gram in output tokens
in source tokens.
A match is considered found if source tokens contain a single matching n-gram.
Return the index of the start of the n-gram in source tokens.
No matching if found if src tokens contain multiple or zero matching n-grams. Return -1.
```
* GridSample OP implementation for CPU and CUDA
**Description**: This change contains implementation for torch grid_sample OP.
Cuda implementation contains contribution from Muscle Wu.
* Use interpolation for out-of-bound points in zero padding mode
Out-of-bound points in zeros padding mode changed from constant 0 to
interpolation of surrounding pixels. This aligns with Pytorch implementation.
A bug in CUDA batch offset calculation is fixed.
Custom op exporter type is added.
* Fix nearest bug in CPU
* Update per CI build finding and review comments
* Force float to avoid potential integer T issue
* Style update
* PR update
* Remove c++17 feature from cuda code
* changes
* tile grad unsqueeze fix for opset 13
* clean up
* remove bool support for opset 2 to 12 for Pad as it is not supported.
* Copy OperatorKernels.md from artifacts of Windows CI build.
* updates for picking pnnx commit
* add tests filter to c# tests
* plus test fixes
* fix versioning for contrib ops
* fix tests
* test filter for optional ops
* more versioning related updates
* fix test
* fix layernorm spec
* more updates
* update docs
* add more test filters
* more filters
* update binary size threshold
* update docs
* plus more fixes
* updates per review
* update to release commit
* add filters for optional type tests
* plus updates
QGemm takes in quantized A, B, C, and quantization parameters of output Y, in which C and quantization parameters of Y are optional. Its output can be quantized or full precision, which depends on whether quantization parameters of Y exists or not. If quant params of Y are provided, the output will be requantized or is full precision.
Comparing with QLinearMatMul and MatMulInteger, QGemm supports transpose, apha and beta attribute.
The formula for quantized GEMM is:
Y = alpha * scale_a * scale_b * ((A_int8 - zp_a) * (B_int8 - zp_b) + C_int32), in which,
C_int32 is quantized with formula: C_int32 = (beta * C) / (alpha * scale_a * scale_b)
SparseTensor support
Implement Builder pattern
Fix support for 1-D and 2-D COO indices
Implement and test CSR support.
Handle shape inference for SparseTensors
Implement conversion for COO, CSR and tests.
Address the case where constant sparse initializer is the output.
Implement test infra for SparseTensors
Implement SparseDenseMatMul for Csr and COO and tested it.
Add hash for SparseToDenseMatMul
Finish shared provider refactor
Refactor GetOrCreate to Create
Working on py interface
Expose OrtDevice and use it in allocate_numpy
Adjust Sparse interfaces, add support for string SparseTensor. Add tests.
Add and test to_cuda()
Add accessors to format specific indices
Test values and indices views, read-only flag, after GC access
Add sparse related methods to OrtValue
Re-work SparseTensor wrapper, add OrtValue methods
Rework numpy_array_to_cuda/to_cpu
Add run_with_ort_values
Add models and test sparse_mat_mul with run_with_ort_values
Refactor sparse tensor to use a single buffer
Ifdef x86 Eigen CSR sparse matmul implementation
Exclude broken test, check for string type when copying cross device
Split pybind schema, regenerate docs, add exclusion
Conditionally exclude schema module
Update docs fix cuda build
Add test to a filter and renerate JS docs
Add conversion and test string support for sparse tensors
Exclude conversion utils from minimal build
Add CUDA Memcpy and adjust provider interfaces
* changes working to convert akv nodes
* changes to replace nodes
* changes to accomodate qkv hidden sizes as attributes
* kernel to accept qkv_hidden_size attributes
* Working till compute for varied dimension, todo applyattention()
* changes to make all regression tests work
* inference running successfully without prepack
* success inference with pre-pack weights
* add test for diff sizes
* bias shape need not be a mul of 3
* get the output_hidden_size from input
* infer output shape from input
* merge with master
* cleaning up files that got merged wrong
* accurancy at accepted level
* added unit test case for different dimensions
* all unit tests passing
* packed weights working for attention
* prepacked weights working
* added test case for newly added extra qk input
* updated unit test to test only extra add qk
* fixing build error
* removing few debugs
* reverting test changes
* all python test passing
* cleaning up
* new unit test added, major clean up of code
* removed extra code
* minor
* minor fix to tests
* prepack weights code cleaned up
* compacted compute() in attention.cc
* reformat compute()
* making a parameter T
* adding 3 q,k,v buffers in all cases
* fixing build
* running tests only on cpu
* Updating docs
* trigger ci builds
* Addressing comments in PR
* addressing some more comments
* get add_qk_str from add_qk node directly
* updating docs, added extra check to verify attn inputs
* Optimized the extra add by parallelizing
* added attention_shape to symbolic_shape_infer.py
* minor refactoring to address comments
**Description**:
Enforce no repetition of n-grams. Scores are set to `-inf` for tokens that form a repeated n-gram if added to the back of the input_ids.
**Motivation and Context**
Needed by transformer models in sequence generation algorithms (greedy search and beam search). This module has heavy impact on performance, and can be highly parallelized.
* Update the operator documentation generation
- Make layout a little nicer
- Update to latest supported operators including training
- Fix some links that are broken when the docs content is copied to github-pages
- Fix incorrect usage of 'onnx.ai.ml' as the default domain
- ML ops are now separated from the real default domain of 'onnx.ai'
- Include CPU, CUDA and training kernels
- exclude DNNL as it's not an EP we own
* There are separate paths for CUDA and CUDNN as they are not guaranteed to be in the same location on a Windows machine. Use the CUDNN path when looking for the CUDNN library.
* Enable validation of both contrib ops and operator kernels in build
Filter generation so it's deterministic
Add ability for CI to publish the md files as build artifacts if they differ so a developer can download and add to their PR to resolve any diffs.
Remove workarounds for github-pages as that will now link to the github docs which display correctly
* checkin
* add 4dmask support in attention cuda op
* trim
* add comments
* fix build/test error
* review comments and add tests
* sync doc
* review comments
* minor change
* Implement qlinear concat and unit test.
Add quantization tools for QLinearConcat and it quantization tests.
* Add kernel def hash for QLinearConcat.
* Change according to PR. Add qdq transformer support for QLinearConcat.
* Add QDQ Transformer unittest. Fix typo on domain.
* remove dup logic of no use.
* fix x86 build error.
* Update operator docs.
- Added python script for generating markdown doc from the registered opkernels.
- Made some conditional changes in the pybind to expose necessary python API
- Added some missing type-constraints in the op kernel registrations
