[TRACKER] Bloom-PyTorch Model #1340
Assignees
Labels
model support
Hub issue for progress on adding support for a specific model
Comments
|
this seems like a local build issue. |
it's might because george-cumsum-op-support and prashant-max-other-op-support both did change on cumsum op, might lead to conflict |
vivekkhandelwal1
added
the
model support
|
#1348 aten::_reshape_alias op need to be fixed first. |
qedawkins
pushed a commit
to nod-ai/torch-mlir
that referenced
this issue
Oct 3, 2022
…#1340) Leverage the template function 'shapeHelperInferShapes' to reduce code "duplication" in ShapeInference.cpp for several ONNX operators. As an example, the following implementation of the Transpose operator inferShapes member function: ONNXTransposeOp::inferShapes( std::function<void(mlir::Region &)> doShapeInference) { // Cannot infer shape if no shape exists. if (!data().getType().isa<RankedTensorType>()) return success(); auto elementType = data().getType().cast<ShapedType>().getElementType(); ONNXTransposeOpAdaptor operandAdaptor(*this); ONNXTransposeOpShapeHelper shapeHelper(this); if (failed(shapeHelper.computeShape(operandAdaptor))) return emitError("Failed to scan Transpose parameters successfully"); SmallVector<int64_t, 4> outputDims; IndexExpr::getShape(shapeHelper.dimsForOutput(), outputDims); getResult().setType(RankedTensorType::get(outputDims, elementType)); return success(); } Becomes: ONNXTransposeOp::inferShapes( std::function<void(mlir::Region &)> doShapeInference) { return shapeHelperInferShapes<ONNXTransposeOpShapeHelper, ONNXTransposeOp, ONNXTransposeOpAdaptor>(this, elementType); } Signed-off-by: Ettore Tiotto [email protected]
qedawkins
pushed a commit
to nod-ai/torch-mlir
that referenced
this issue
Oct 3, 2022
* Use shapeHelperInferShapes template to reduce boilerplate code. (llvm#1340) Leverage the template function 'shapeHelperInferShapes' to reduce code "duplication" in ShapeInference.cpp for several ONNX operators. As an example, the following implementation of the Transpose operator inferShapes member function: ONNXTransposeOp::inferShapes( std::function<void(mlir::Region &)> doShapeInference) { // Cannot infer shape if no shape exists. if (!data().getType().isa<RankedTensorType>()) return success(); auto elementType = data().getType().cast<ShapedType>().getElementType(); ONNXTransposeOpAdaptor operandAdaptor(*this); ONNXTransposeOpShapeHelper shapeHelper(this); if (failed(shapeHelper.computeShape(operandAdaptor))) return emitError("Failed to scan Transpose parameters successfully"); SmallVector<int64_t, 4> outputDims; IndexExpr::getShape(shapeHelper.dimsForOutput(), outputDims); getResult().setType(RankedTensorType::get(outputDims, elementType)); return success(); } Becomes: ONNXTransposeOp::inferShapes( std::function<void(mlir::Region &)> doShapeInference) { return shapeHelperInferShapes<ONNXTransposeOpShapeHelper, ONNXTransposeOp, ONNXTransposeOpAdaptor>(this, elementType); } Signed-off-by: Ettore Tiotto [email protected] Signed-off-by: Philip Lassen <[email protected]> * Allow lowering ONNXGemm with dynamic dims to ZHigh and fix zDNN Conv condition (llvm#1332) * Allow lowering ONNXGemm with dynamic dims to ZHigh Signed-off-by: Tung D. Le <[email protected]> * Update lit tests Signed-off-by: Tung D. Le <[email protected]> * Fix zDNN Conv condition Signed-off-by: Tung D. Le <[email protected]> Signed-off-by: Philip Lassen <[email protected]> * Fix builders with boolean output types Signed-off-by: Philip Lassen <[email protected]> * Fix format issue Signed-off-by: Philip Lassen <[email protected]> * Fix legality check of ONNXToZHigh for MaxPool. (llvm#1343) * Fix legality check of NNPA for 1d maxpool Signed-off-by: Haruki Imai <[email protected]> * Apply the same fix to conv Signed-off-by: Haruki Imai <[email protected]> * Add lit test for 1d maxpool and averagepool Signed-off-by: Haruki Imai <[email protected]> * Insert diation check after checking shape Signed-off-by: Haruki Imai <[email protected]> * Simplify lit test for pooling and update func name Signed-off-by: Haruki Imai <[email protected]> * Change func name to test_pool_not_lowered_pool1d and test_pool_not_lowered_pool3d Signed-off-by: Haruki Imai <[email protected]> Signed-off-by: Philip Lassen <[email protected]> * embed libzdnn in model.so (llvm#1324) * - Build libzdnn.a with -fPIC and embed in model.so when -maccel=NNPA specified - Add CompilerConfigMap to store states associated with certain options - Move options in main() to CompilerOptions.cpp - Fix compiler warning in Stickify.cpp Signed-off-by: Gong Su <[email protected]> * - fix NNPA_ENABLED for lit test - install zdnn.h so we no longer need third_party/zdnn-lib - move options back to onnx-mlir.cpp::main (consolidation requires much more effort, deferred) - use DEPENDS in add_onnx_mlir_library for libzdnn dependency Signed-off-by: Gong Su <[email protected]> * Remove zdnn-lib from .gitmodules Signed-off-by: Gong Su <[email protected]> * Make libzdnn ALL target so it gets built before other NNPA components Signed-off-by: Gong Su <[email protected]> * Fix libzdnn dependency for NNPA components Signed-off-by: Gong Su <[email protected]> * Build NNPA in dev image as well Signed-off-by: Gong Su <[email protected]> * Comment out BYPRODUCTS in target libzdnn since generator support requires cmake 3.20+ which is not yet available on official Ubuntu Focal Signed-off-by: Gong Su <[email protected]> * - Force setting cached MLIR_DIR to honor command line argument - unset cached LLVM_DIR so it changes along with MLIR_DIR - surround third_party with set(CMAKE_MESSAGE_LOG_LEVEL NOTICE) and set(CMAKE_MESSAGE_LOG_LEVEL STATUS) to mask out their cmake output so we can see more clearly output by onnx-mlir only. third_party cmake output can still be turned on by the --log-level command line option Signed-off-by: Gong Su <[email protected]> * revert set(CMAKE_MESSAGE_LOG_LEVEL NOTICE) and set(CMAKE_MESSAGE_LOG_LEVEL STATUS) around third_party to make another PR Signed-off-by: Gong Su <[email protected]> * revert force setting cached MLIR_DIR and unsetting cached LLVM_DIR to make another PR Signed-off-by: Gong Su <[email protected]> Co-authored-by: Charles Volzka <[email protected]> Co-authored-by: Tung D. Le <[email protected]> Signed-off-by: Philip Lassen <[email protected]> * ScatterElements operator code gen. (llvm#1352) Implement support for the ONNX ScatterElement operator: - verification (verify diagnostic completeness) - shape inference (should be trivial, but verify) - initial codegen support - codegen for negative indices - add lit test to check code generation - enable end-to-end tests (backend tests) Signed-off-by: Ettore Tiotto [email protected] Signed-off-by: Philip Lassen <[email protected]> * Improve variable naming of builder lists Signed-off-by: Philip Lassen <[email protected]> Co-authored-by: Ettore Tiotto <[email protected]> Co-authored-by: Tung D. Le <[email protected]> Co-authored-by: Haruki Imai <[email protected]> Co-authored-by: gongsu832 <[email protected]> Co-authored-by: Charles Volzka <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hi
I am working on providing the support for bloom-pytorch model via torch-mlir. As of now, I used 2 unmerged patch to solver the ops that not supported. Those 2 patch pass the check independently, but when I try to merge them together. It results in the following segmentation error:
Here is the patch to run:
https://github.com/AmosLewis/SHARK/tree/bloom
To test it,
The torch_mlir used is https://github.com/AmosLewis/torch-mlir/tree/bloom3.
It is the combination of george-cumsum-op-support and prashant-max-other-op-support.
I tried to run the python test
cmake --build build --target check-torch-mlir-pythonThe george-cumsum-op-support and prashant-max-other-op-support op support patches pass the check independently. But when combining them together, it leads to the following error:
The text was updated successfully, but these errors were encountered: