How llvm/torch-mlir works
Torch-MLIR bridges PyTorch-family inputs into MLIR: documented entry routes include FX and ONNX, with a Torch dialect and conversion passes forming the compiler core. The tree also shows a JIT importer, a lazy-tensor integration, C/Python APIs, and several lowering targets. The principal path is input import → Torch IR → target lowering; the reference backend provides a separate execution-oriented component. Some lowerings and subsystem links are inferred from the tree rather than sampled call sites, so those links are dashed and unverified.
Main entry points: TorchOps.cpp Torch dialect · DefaultDomainQtoZ.cpp ONNX conversion · InitAll.cpp Pass registration · function_importer.cpp JIT importer
Components 19
Input frontends
- FX importer
python/torch_mlir/extras/fx_importer.py - FX API
python/torch_mlir/fx.py - ONNX importer
projects/onnx_c_importer/OnnxImporter.cpp - ONNX conversion
lib/Conversion/TorchOnnxToTorch/DefaultDomainQtoZ.cpp - JIT importer
projects/jit_ir_common/csrc/jit_ir_importer/function_importer.cpp
Torch IR
- Torch dialect
lib/Dialect/Torch/IR/TorchOps.cpp - Torch types
lib/Dialect/Torch/IR/TorchTypes.cpp
Compiler passes
- Torch transforms
lib/Dialect/Torch/Transforms/ScalarizeShapes.cpp - Pass registration
lib/InitAll.cpp
Targets and integrations
- Linalg lowering
lib/Conversion/TorchToLinalg/Uncategorized.cpp - StableHLO lowering
lib/Conversion/TorchToStablehlo/TorchToStablehlo.cpp - TOSA lowering
lib/Conversion/TorchToTosa/TorchToTosa.cpp - TMTensor lowering
lib/Conversion/TorchToTMTensor/TorchToTMTensor.cpp - Reference backend
lib/RefBackend/RefBackend.cpp - Lazy backend
projects/ltc/csrc/base_lazy_backend/mlir_lowering_context.cpp - C API
lib/CAPI/TorchOps.cpp - MLIR optimizer
tools/torch-mlir-opt/torch-mlir-opt.cpp
Other components
- PyTorch
- ONNX models
Connections 22
- PyTorch → FX API: provides FX
README.md - FX API → FX importer: uses importer
python/torch_mlir/fx.py - ONNX models → ONNX importer: provides model
README.md - ONNX importer → ONNX conversion: imports for conversion
- ONNX models → ONNX conversion: ONNX entry
README.md - PyTorch → JIT importer: provides JIT input
- FX importer → Torch dialect: emits Torch IR
README.md - ONNX conversion → Torch dialect: rewrites to Torch ops
lib/Conversion/TorchOnnxToTorch/DefaultDomainQtoZ.cpp - JIT importer → Torch dialect: imports Torch IR
- Torch dialect → Torch types: uses types
lib/Dialect/Torch/IR/TorchOps.cpp - Torch transforms → Torch dialect: rewrites ops
lib/Dialect/Torch/Transforms/ScalarizeShapes.cpp - Pass registration → Torch transforms: registers passes
lib/InitAll.cpp - Pass registration → ONNX conversion: registers passes
lib/InitAll.cpp - Torch dialect → Linalg lowering: lowering input
- Torch dialect → StableHLO lowering: lowering input
- Torch dialect → TOSA lowering: lowering input
- Torch dialect → TMTensor lowering: lowering input
lib/Conversion/TorchToTMTensor/TorchToTMTensor.cpp - Lazy backend → JIT importer: imports functions
projects/ltc/csrc/base_lazy_backend/mlir_lowering_context.cpp - MLIR optimizer → Pass registration: registers dialects
tools/torch-mlir-opt/torch-mlir-opt.cpp - C API → Torch dialect: exposes ops
lib/CAPI/TorchOps.cpp - Reference backend → Torch dialect: uses conversion ops
lib/RefBackend/RefBackend.cpp - Linalg lowering → Reference backend: backend path