E2E HuggingFace Bert using LTC Backend #912
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
- Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization
- Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor
- Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message
- Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models
- Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy
|
Note that for this PR, we are executing on the JIT graph (which we confirmed matches exec on CPU) and not the emitted MLIR. We're operating under the assumption that torch-mlir is doing its job in emitting MLIR that accurately represents the JIT graph. In a future PR, we will lower the MLIR to linalg and execute on a reference backend to validate that the MLIR itself is able to produce the correct numerics. |
|
Awesome!!! 🥳 Is there anything in particular you would like me to look at in this big PR? Otherwise I trust you folks and LGTM. |
Nothing in particular. The changes made are quite isolated to the LTC backend and won't affect any other part of Torch-MLIR. Thanks for the vote of confidence! |
antoniojkim
added a commit
that referenced
this pull request
Jun 30, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
antoniojkim
added a commit
that referenced
this pull request
Jun 30, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
antoniojkim
added a commit
that referenced
this pull request
Jul 5, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
antoniojkim
added a commit
that referenced
this pull request
Jul 7, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
henrytwo
added a commit
that referenced
this pull request
Jul 8, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
henrytwo
added a commit
that referenced
this pull request
Jul 8, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
henrytwo
added a commit
that referenced
this pull request
Jul 12, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
antoniojkim
added a commit
that referenced
this pull request
Jul 15, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
antoniojkim
added a commit
that referenced
this pull request
Jul 19, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
antoniojkim
added a commit
that referenced
this pull request
Jul 22, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
henrytwo
added a commit
that referenced
this pull request
Jul 29, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
henrytwo
added a commit
that referenced
this pull request
Jul 29, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
henrytwo
added a commit
that referenced
this pull request
Jul 30, 2022
* Update native function definitions * Add ops to support bert lowering - Add empty_strided and as_strided - Restore zeros_like to op blacklist (Without this, tensors will be unintentionally created with a CPU device rather than lazy) - Check for composite implicit ops and add device data IR - Also fix codegen for functionalization * Add autogen to CMakeList * Remove PyTorch submodule * Reduced BERT model size * Print Mark Step status in Torch MLIR LTC debug string * Apply fixes to work with latest upstream/main - Pass importOptions into getMlirTypeFromTorchType during NodeImporter::importNode Without this, the tensor type created may have a mismatched type as ImportOptions may cause vtensor to be used instead of tensor * Update shape inference functions - Fixed compute_shape_native_batch_norm when mean and var are uninitialized Previously, the number of shapes returned would be <3 if either mean or val was didn't exist. Instead, we now initialize them with a vector matching the number of channels. - Implemented compute_shape_mul - Fixed bug in reshape shape inference error message * Get MLIR backend more consistent with TS backend - Remove LazyNativeFunctions::_unsafe_view from autogen - Blacklist ops to make JIT graph more like output of TS backend - Print graph when SSA value has mismatch of types and results - Remove normalize_index from LazyShapeInference - Fix seeds for LTC example models * Update and clean up shape inference functions - Prune shape inference functions - Add shape inference function for GenerateSlice - Add shape inference function for GenerateCopy Co-authored-by: Henry Tu <henry.tu@cerebras.net>
qedawkins
pushed a commit
to nod-ai/torch-mlir
that referenced
this pull request
Oct 3, 2022
* Add links to onnxai Signed-off-by: Tung D. Le <tung@jp.ibm.com> * Reduce spaces on the left and righ sides Signed-off-by: Tung D. Le <tung@jp.ibm.com> Co-authored-by: Alexandre Eichenberger <alexe@us.ibm.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
We finally have the full training graph for HuggingFace's implementation of Bert working using the LTC Backend!
The
torch_mlir_ltc_backendstaging branch still won't be ready to merge intomainafter this PR though. We are still waiting on a couple things in upstream PyTorch to be merged in:torchgenrelies on non-packagednative_functions.yamlpytorch/pytorch#77445Plus, there are a few more minor changes I'd like to make before we merge it in that will help make it easier to maintain.
CC: @ke1337 @henrytwo