[WIP]Add initial support for transpiler

[chandrasekaranpradeep]

Summary

This PR introduces a new ONNX Transpiler Framework that converts ONNX ModelProto inputs into a fully executable Transpiler Intermediate Representation (TIR) and generates Forge Module. The transpiler provides a modular flow covering ONNX parsing, intermediate graph construction, PyTorch-based computation, attribute transformation, and detailed per-node debugging support.

This establishes the foundation for replacing the existing TVM-based workflow with a complete in-house transpilation pipeline.

Key Features

1. ONNX Frontend & Input Handling

• Directly accepts ONNX ModelProto as input.
• Extracts ONNX computational graph, operators, weights, and model metadata.

2. TIR Graph Construction

• Builds a framework-agnostic TIRGraph.
• Creates a 1:1 mapping between ONNX operations and TIRNode objects.
• Each TIRNode stores operator name, inputs/outputs, attributes, and execution details.

3. TIRNode Operation Pipeline

Each ONNX operation goes through the following pipeline:

• Framework Attrs → Torch Attrs Conversion
  Converts ONNX operator attributes (e.g., kernel_shape, pads, strides) into PyTorch-compatible formats.

• Torch Attrs → Forge Attrs Conversion
  Normalizes PyTorch attributes into Forge-specific parameters used for final lowering.

• PyTorch-Based Computation
  Utilizes torch.* operators for numerical computation and validation.

• Forge CodeGen Emission
  Generates Forge operation metadata including op name, function, inputs, outputs, and attributes.

4. Executable TIR Graph

• The entire TIR graph is runnable end-to-end using PyTorch.
• Executes nodes in topological order, ensuring correct data dependency handling.
• Includes activation garbage collection to reduce memory usage during execution.

5. Debug Mode & Validation

• Supports per-node output comparison against ONNXRuntime.

Workflow Overview

ONNX ModelProto
    ↓
[ONNX Frontend Engine]
    ↓
TIRGraph Construction
    ├── TIRNode (Operation 1)
    │   ├── Framework Attrs → Torch Attrs
    │   ├── Torch Attrs → Forge Attrs
    │   ├── PyTorch Computation
    │   └── Emit Forge Operation Info
    ├── TIRNode (Operation 2)
    │   └── ...
    └── ...
    ↓
Executable TIRGraph
    ├── Runnable (accept inputs, produce outputs)
    └── Debug Mode (compare with ONNXRuntime)
    ↓
Forge Module Construction

[github-advanced-security]

flake8 found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.

To fix the issue, remove the unused imports from the import statement on line 4 in forge/forge/transpiler/core/registry.py. Specifically, delete Dict and Type from the from typing import ... statement. Since neither Dict nor Type are used in this file, no further action is required. Ensure that only the line with the unused imports is removed or amended, leaving the remaining code unchanged.

To fix this issue, we should remove the unused import of TIRNode from the import statement on line 5 of forge/forge/transpiler/core/registry.py. This can be accomplished by editing the import line to only import get_op_registry as _get_op_registry from ..ir.nodes, thereby eliminating the unnecessary dependency and reducing clutter in the code. No other changes need to be made, as there are no references to TIRNode elsewhere in the provided code.

---

The best way to fix the problem is to simply remove the unused import onnx statement from line 4 in forge/forge/transpiler/frontends/onnx/converters/utils.py. This cleans up the code by removing unnecessary dependencies, as all necessary symbols from onnx are already imported via from onnx import ModelProto. No other changes to code or dependencies are required.

---

To fix the problem, simply delete the line from .graph import TIRGraph from forge/forge/transpiler/test_verify.py, as this import is not used anywhere in the provided code. No additional changes, imports, or definitions are required elsewhere in the file. Double-check that no other usage or definition of TIRGraph exists in the given snippet before removal.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 63.55%. Comparing base (6f9586b) to head (8d274e1).
⚠️ Report is 4 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3063   +/-   ##
=======================================
  Coverage   63.55%   63.55%           
=======================================
  Files         156      156           
  Lines       11908    11908           
=======================================
  Hits         7568     7568           
  Misses       4340     4340           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

To fix this issue, simply remove Optional from the from typing import ... import statement on line 8 in forge/forge/transpiler/frontends/onnx/engine.py. Leave the rest of the imports unchanged, and do not remove the entire line if other types are still used elsewhere in the file. This change ensures that the code does not include unnecessary dependencies, reduces potential confusion, and follows best practices for code hygiene. No new functionality or external dependencies are required for this fix.

To fix this problem, we should remove the unused import statement, specifically the line from typing import List as ListType. This will clean up the unnecessary dependency, improving code readability and slightly reducing module load time. No other sections of the presented code are dependent on ListType being present, so removal is safe and cannot affect existing functionality.

Locate the from typing import List as ListType line (line 9) in forge/forge/transpiler/frontends/onnx/engine.py and delete it. No further changes are necessary.

---

The best way to fix this problem is to simply delete the unused import statement: from ...ir.nodes import TIRNode from line 12 of forge/forge/transpiler/frontends/onnx/engine.py. This removes unnecessary code and avoids confusion about where dependencies exist. No further changes are required since no code in the shown snippet references TIRNode.

The best way to fix this problem is to delete the assignment to the unused variable kernel_size, taking care not to remove any right-hand-side expressions that might have side effects (in this case, there are none; the assignment is straightforward). Specifically, lines 227 and 230 handle assignment to kernel_size, but since it is never used, both assignments should be removed, and we should restructure the conditional so that all references to kernel_size are omitted. The surrounding logic relies only on kernel_dims, which is used, so the fix amounts to removing the assignment lines for kernel_size (lines 227 and 230), and not introducing any variable in its place.

No imports, definitions, or method changes are needed—simply remove the assignments for the unused variable.

To fix the issue, simply remove the assignment to kernel_size since it is not being used anywhere in the provided code. Deleting this assignment avoids confusion, clarifies intent, and eliminates the warning flagged by the static analysis tool. Since kernel_size does not appear to have any side effects on the right-hand side (it's just a value extracted from kernel_shape), this change is safe and has no effect on functionality.

Edit only line 230 in forge/forge/transpiler/frontends/onnx/engine.py to remove the assignment to kernel_size, leaving assignments to kernel_dims untouched. No other code or imports need to be altered.

To fix the issue, we should delete the import statement: from ...frontends.onnx.converters.autopad import AutoPad on line 10 of forge/forge/transpiler/ir/operations/pooling.py. There is no need to add any replacement or substitute code, as no references to AutoPad exist in the shown code. This change preserves the current functionality while cleaning up an unnecessary import.

To fix this issue, simply remove the statement assigning ndim = 1 on line 34 in the MaxPoolNode.create static method. This change does not affect functionality because the value is not used in any subsequent code or logic for this branch. Care should be taken not to remove any assignment with side effects, but here the right hand side is a constant and has no side effects. Only this line needs to be deleted.

---

To fix the problem, remove the unused assignment to the variable ndim on line 109 within the create method of the AveragePoolNode class. Since its assignment has no side effects and is not referenced elsewhere, delete just the line containing ndim = 1 and leave the rest of the method unchanged. No additional methods, imports, or definitions are required. Only edit forge/forge/transpiler/ir/operations/pooling.py at line 109.

The best way to fix the problem is to remove the unused import of Optional from the import statement on line 5 in forge/forge/transpiler/ir/operations/reduction.py. We should not remove the entire typing import, but only delete Optional to prevent unnecessary dependencies, while leaving the other imports (Dict, List, Union, Tuple) intact because they are used in type annotations throughout the file. No further changes are required.

To fix the problem, simply remove the unused import onnx statement on line 4 of forge/forge/transpiler/ir/types.py. This eliminates the redundant dependency and conforms to Python best practices, making the code cleaner and easier to maintain. No further changes are required, since all necessary references to ONNX are via the explicit import from onnx import TensorProto.