commit a54af64872c68913309541f6f30e75da3921ef77
author Shushi Hong <820958424@qq.com> Sun Sep 21 23:36:18 2025 -0400
committer GitHub <noreply@github.com> Sun Sep 21 23:36:18 2025 -0400
tree 0c4344f09f2eb1660c424784e8646abc25609d4a
parent eb45a46d7004c05a0f613b087d7cf82c19ce6196

[Relax][Backend] Implement R.call_py_func operator for calling Python functions from compiled TVM (#18326)

This PR implements the `R.call_py_func` operator that allows compiled
TVM Relax modules to call Python functions at runtime. This enables
integration between TVM's compiled code and Python through a
robust VM backend implementation.

#### Simple Usage with BasePyModule
```python
@I.ir_module
class MyModule(BasePyModule):
    @I.pyfunc
    def torch_relu(self, x):
        return torch.relu(x)
    
    @R.function
    def forward(x: R.Tensor((10,), "float32")) -> R.Tensor((10,), "float32"):
        return R.call_py_func("torch_relu", (x,), out_sinfo=R.Tensor((10,), "float32"))
```

#### Direct VM Backend Usage (Manual)
```python
# Manually register Python function with VM backend
register_func = tvm.get_global_func("vm.builtin.register_py_func")
register_func("my_func", my_python_function)

# Use in Relax function (compiled to VM backend)
@R.function
def test(x: R.Tensor((5,), "float32")) -> R.Tensor((5,), "float32"):
    return R.call_py_func("my_func", (x,), out_sinfo=R.Tensor((5,), "float32"))

# Manual cleanup (required for direct VM backend usage)
clear_func = tvm.get_global_func("vm.builtin.clear_py_func_registry")
clear_func()
```