Unsolved problems:
1. One test failure was caused by a bug in Cudnn rnn kernels, when they can allocate a buffer and partially initialize it, the garbage data near tail of the buffer caused problem in some of the hardware. To attack this problem in a broader sense, should we add code in our allocators, and during a memory fuzzing test, fill an allocated buffer with garbage before returning to the caller?
2. Prepacking is used more widely than we know. For instance, Cudnn rnn kernels also cache their weights. They mix several weight tensors together into a single buffer, and never touch the original weight tensor anymore. This is the same idea with pre-pack, but they didn't override the virtual function, and they never tried to release those weight tensors, leading to memory waste. It also seems to me that there are some other kernels have similar behavior. Wonder how much memory we can save if we try to cleanup those too.
3. Turning off memory pattern planning does increase memory fragmentation, leading to out of memory error in some training test cases. Perhaps we can revisit the idea of pushing kernels-creation stage earlier, and then during initializer deserialization, we only avoid tracing those that will be prepacked.
* remove tests to speed up CI
* add back _into_data_parallelism tests to see how long the CI test takes
* remove unnecessary save calls
* add back data_parallelism_full_precision_bart_path
* add data_parallelism_full_precision_path
* remove data parallelism tests
Co-authored-by: Jingyan Wang <jingywa@OrtTrainingDev3.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
* Partial updating of ROCM reduction code.
* Update reduction_all.cu
* Add reduce template parameters.
* miopen common
* Reuse CUDA's reduction_functions.cc
* Reduction ops.
* Update remaining reduction ops to use MIOpen. double datatype is not supported, so disable those typed kernels.
* Disable a couple more unsupported tests.
* Code formatting.
* Delete ROCM-specific reduction code that is identical to CUDA reduction code.
* Fix scratch buffer early free.
* Fix merge conflict.
* first attempt nightly amd ci pipeline
* try fix bad yaml file
* try again with corrected model directory
* add convergence test as well
* update reference loss for amd mi100
* include mi100 test results csv
* update the mi100 convergence test reference values
* update batch sizes for mi100 32g
* fix gpu sku for run_convergence_test.py
* undo unrelated changes to master
* pr comments
* pr comment
Co-authored-by: Jesse Benson <jesseb@microsoft.com>
* Add warning when GetGradientForOp() silently fails.
In some cases, `GetGradientForOp()` can return without creating any nodes, which may lead to an invalid graph being created.
Remove condition from ORT_RETURN_IF[_NOT] macro output as repeating the condition doesn't add much value compared to the explicit error message, and the error message includes the file and line anyway so it's easy enough to find the condition if needed.
Update the few places where the macros were used without an explicit error message to provide an explicit error message.
Saves 12.5KB in a minimal MinSizeRel build with all DNN ops, 16KB in full release build.
* Support to allow user to specify compute stream per session
Create computation cuda stream explicitly rather than use default legacy stream or per-thread default stream.
remove some redudant cudaStreamSynchronize
fix gpt2 model test failures
don't use default stream in nccl either.
add stream schronization in OnRunEnd()
using cub::DeviceScan::InclusiveSum which can be called with stream specified.
fix topK failure due to latest rebase
fix tensorrt
support user specified stream
add user_stream support in tensorrt EP
use same stream for both tensort and CUDA EP.
fix ScatterND
specify stream for adasum and p2p kernels.
fix loop
fix CApiTest.custom_op_handler
fix CApiTest.varied_input_custom_op_handler
change for cudaMemcpyFromSymbol
improve provider options for user specified compute stream
* add changes for ROCM EP
* fix GatherGrad UT for ROCM EP
* clean code and fix NonMaxSuppression
* use default stream for ROCM now
* fix CApiTest.custom_op_handler:OrtFormatCustomOpTests.ConvertOnnxModelToOrt
* fix tensorrt ut: CApiTest.io_binding_cuda
Co-authored-by: Weixing Zhang <wezhan@microsoft.com>
1. Merge Nuget CPU pipeline, Java CPU pipeline, C-API pipeline into a single one.
2. Enable compile warnings for cuda files(*.cu) on Windows.
3. Enable static code analyze for the Windows builds in these jobs. For example, this is our first time scanning the JNI code.
4. Fix some warnings in the training code.
5. Enable code sign for Java. Previously we forgot it.
6. Update TPN.txt to remove Jemalloc.
* minimal_build with training ops
* Removing redundant comment from an earlier attempt at a fix
* Fixing a bad merge conflict resolution
* Responding to PR feedback
* tweaking the makefiles based on feedback
* combining two enable_training blocks in CMakeLists.txt
* Add ReluGrad and ConvGrad ops for the dnnl provider
* the mnist sample is updated to add the --use_dnnl option that
will cause the sample to use the dnnl execution provider for
nodes that exist in dnnl provider.
* Added the ability to find forward ops. Dnnl backward gradient
ops require the forward primitive description and workspace
from the forward operation.
* Enable specifying the execution provider for Gradient Checker Tests
* Prevent memory leak when running dnnl_provider in training mode
Prevent creating a SubgraphPrimitivePool when the code is built with the
ENABLE_TRAINING build flag. Instead create a SubgraphPrimitive directly.
The SubgraphPrimitivePool was causing a pool of SubgraphPrimitives to be
stashed in a map for reuse. Due to the way the Training Loop uses threads
the pool of SubgraphPrimitives were not being reuse instead a new pool of
SubgraphPrimitives being created each run. The old pool was not instantly
freed. This behavior could be a language error when using thread_local
memory.
Signed-off-by: George Nash <george.nash@intel.com>
* Added fixes to maxpoolgrad and memory leak.
Maxpoolgrad will now pass all unit tests.
With the conv and convgrad disabled for dnnl, mnist is able to train till 95%
Signed-off-by: Chethan Palangotu Keshava <chethan.palangotu.keshava@intel.com>
* Fixed misc issues when testing training code with dnnl provider
* fix conv_grad dnnl tests with dilation to run dnnl execution provider
* update mnist training sample to accept convolution type models
convolution models require the input shape to be {1, 28, 28}
instead of the flat {728} image that is used for the gemm models
this will enable models that require the different shape by adding
`--model_type conv` to the command line when running the mnist sample.
(while testing a workaround was used see #4762)
* Disable weight caching in dnnl conv operator when using training
When training we can not use cached weights because the weight
will be updated each run. This re-enables dnnl Conv and ConvGrad Ops.
The weight caching was the source of the error from Conv when training.
* Fix issues found when building grad ops on Linux
* The dnnl_convgrad code was over using the scope operator
causing a compilation problem.
* The dnnl_maxpoolgrad code had a logic error that is was
comparing with the source description when it should have
been comparing with the destination despription.
* Update BUILD.md so it shows DNNL for training
* Updated the table of contents. Since the same providers
are listed twice. Once for Infrance and again for Training
an HTML anchor was added to distinguish the second header
from the first for the TOC.
* Fix build failure when not using --enable-training build option
* reorganize the gradient operators so they are grouped together
* Fix issues found when running onnx_backend_test_series.py
* Pooling code only supports 2 outputs when built with --enable-training
* Address code review feedback
* class member variables end in underscore_
* use dst instead of dist to match pattern use elsewhere in DNNL code.
* Remove workaround that was introduced to handle problems running
convolution based training models. See issue #4762
Signed-off-by: George Nash <george.nash@intel.com>
* Isolate training code and code cleanup
* Do not build if dnnl_gpu_runtime if enable_training is set training code
does not support dnnl_gpu_runtime yet.
* Isolated Training code inside ifdefs so that they wont affect
project if built without training enabled
* Inadvertant changes in whitespace were removed to make code review simpler
* Undid some code reordering that was not needed
* comments added to closing #endif statments to simplify reading complex ifdefs
* Modified the GetPrimitiveDesc functions to return shared_ptr instead of raw
pointer. This matches what was done in Pool code and is safer memory code.
Signed-off-by: George Nash <george.nash@intel.com>
* Address code review issues
- whitespace changes caused by running clang-format on the code
- Several spelling errors fixed
- Removed/changed some ifdefs to improve readability
- other misc. changes in responce to code review.
Signed-off-by: George Nash <george.nash@intel.com>
* Code changes to address code review
- Simplify iteration code using `auto` keyword
- remove C style cast that was not needed
- remove instance variable that was not needed [relugrad.h]
- added the execution providers to `ComputeGradientErrorInternal()`
and `ComputeTheoreticalJacobianTranspose()` instead of using
a pointer to an instance varaible [gradient_checker.h/.cc]
Signed-off-by: George Nash <george.nash@intel.com>
* Combined the default gradient ops test and dnnl gradient ops test for ConvGrad and MaxPoolGrad into one function with the help of a helper function.
This will reduce repeated code.
Signed-off-by: Palangotu Keshava, Chethan's avatarChethan Palangotu Keshava <chethan.palangotu.keshava@intel.com>
* Replaced the stack used by convgrad to vector so that the vector(used as stack) can be easily cleared everytime the graph is created.
This will prevent memory leak from convolution kernels being pushed constantly onto the stack.
Signed-off-by: chethan.palangotu.keshava@intel.com
* Code clean up and formating updates
- Removed empty else statment
- updated indentation of code that was causing double curly brackets to look unususal
- Changed check for NumDimensions to Size in Relu and ReluGrad error checking code.
- isolated training code
Signed-off-by: George Nash <george.nash@intel.com>
* Restore inadvertantly removed ConvGrad tests
When combining the DNNL and CPU version of the ConvGrad
tests two test were inadvertantly excluded. This adds
back the Conv3d and Conv3d with strides test cases.
Signed-off-by: George Nash <george.nash@intel.com>
* Add validation to ConvGrad
This validates the dimensions of the ConvGrad match the
passed in Convolution forward primitive description.
The current code for DNNL ConvGrad makes the assumption that the ConvGrad
nodes will be visited in the reverse order from the corresponding Conv nodes
The added validation will return an error if this assumption is not true.
Signed-off-by: George Nash <george.nash@intel.com>
* Do not create new execution providers in provider_test_utils
This removes the code that generated new execution providers in the
OpTester::Run function. This was added because the std::move was
leaving the `entry` value empty so subsequent calls would cause a
segfault.
Problem is this potentially changed the execution_provider because it
would create the default provider dropping any custom arguments.
When the now removed code was originally added the std::move was causing
crashes when the GradientChecker unit tests were run. However, it is no
longer causing problems even with the code removed.
Signed-off-by: George Nash <george.nash@intel.com>
* Change the forward conv stack to a forward conv map
This changes how the forward conv kernel is mapped to the bwd ConvGrad
kernel the problematic stack is no longer used.
The convolution stack made the assumption that the corresponding
ConvGrad operator would be visited in reverse order of the forward
Conv operators. This was always problematic and was unlikely to
work for inception models.
Important changes:
- The weight_name is added to the ConvGrad dnnl_node making it
possible to use the weight_name as a lookup key to find the
Conv forward Kernel
- the `std::vector fwd_conv_stack_` has been replaced with a
`std::map fwd_conv_kernel_map_`
- Although it is not needed lock_guards were added when writing
to and reading from the fwd_conv_kernel_map_ as well as the
fwd_kernel_map_. These should always be accessed by a single
thread when preparing the dnnl subgraphs so the guard should not
be needed but its added just in case.
- Updated the comments ConvGrad.h code to no longer mention the
stack. The error check is not removed. It will be good to verify
there are no errors as we continue to test against more models.
Signed-off-by: George Nash <george.nash@intel.com>
Co-authored-by: Chethan Palangotu Keshava <chethan.palangotu.keshava@intel.com>
Co-authored-by: unknown <63478620+jeyblu@users.noreply.github.com>
* Permit dense seq optimization on BERT-L pytorch export by enabling ReduceSumTraining, Equal, and NonZero on AMD
* enable Equal tests
* enable fast_matrix_reduction test case
* expose optimized_model_filepath in SessionOptions as `debug.graph_save_paths.model_with_training_graph_after_optimization_path` in `ORTTrainerOptions`
* merged alloc_plan
* pass compilation
* Start running, incorrect allocation memory info
* add in comments
* fix a bug of recording pattern too early.
* debugging lifetime
* fix lifetime
* passed mnist
* in process of visualization
* Add code to generate chrome trace for allocations.
* in process of collecting fragmentation
* before rebuild
* passed mnist
* passed bert tiny
* fix the inplace reuse
* fix the exception of weight in pinned memory
* add guards to ensure the tensor is in AllocPlan
* add customized profiling
* debugging
* debugging
* fix the reuse of differnt location type
* add rank
* add the rank
* add fragmentation
* add time_step_trace
* Add summary for each execution step (total bytes, used/free bytes).
* add top k
* change type of top k parameter
* remove prints
* change heap to set{
* add the name pattern
* add the useage for pattern
* add partition
* change to static class
* add custom group
* remove const
* update memory_info
* in process of adding it as runtime config
* change the memory profiling to be an argument
* add some comments
* add checks to recored meomry_info in traaining session
* set the "local rank setting" to correct argument.
* addressing comments
* format adjustment
* formatting
* remove alloc_interval
* update memory_info.cc to skip session when there is no tensor for a particular memory type
* fix memory_info multiple iteration seg-fault
* consolidate mainz changes
* fixed some minor errors
* guard by ORT_MINIMAL_BUILD
* add ORT_MEMORY_PROFILE flag
* added compiler flag to turn on/off memory profiling related code
* clean up the code regarding comments
* add comments
* revoke the onnx version
* clean up the code to match master
* clean up the code to match master
* clean up the code to match master
Co-authored-by: Jesse Benson <benson.jesse@gmail.com>
Co-authored-by: Wei Zuo <wezuo@OrtTrainingDev3.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
Co-authored-by: wezuo <wezuo@az-eus-v100-32gb-5-worker-mgtbby.eastus.cloudapp.azure.com>
Co-authored-by: wezuo <wezuo@az-eus-v100-32gb-5-worker-yclzsf.eastus.cloudapp.azure.com>
* traing bert using bf16
* Adam support bf16
* bugfix
* add fusedmatmul support
* fix after merge from master.
* bugfix
* bugfix after merge from master
* fast reduction for bf16.
* resolve comments
* fix win build
* bugfix
* change header file.
Co-authored-by: Vincent Wang <weicwang@microsoft.com>
Update Python API to allow more flexibility for setting providers and provider options.
The providers argument (InferenceSession/TrainingSession constructors, InferenceSession.set_providers()) now also accepts a tuple of (name, options dict).
Fix get_available_providers() API (and the corresponding function in the C API) to return the providers in default priority order. Now it can be used as a starting point for the providers argument and maintain the default priority order.
Convert some usages of the deprecated global configuration functions to use EP-specific options instead.
Update some EP-specific option parsing to fail on unknown options.
Other clean up.
