Recreate python_frontend test_basic for nvfuser_direct

[rdspring1]

This PR recreates test_basic, which is the first test from test/python/test_python_frontend.py.

• It checks correctness against pytorch eager mode and the validity of the python repro via check_captured_python_definition.

New functions to nvfuser_direct FusionDefinition:

• Added __repr__, define_scalar, and from_pytorch

Updated python/python_next/ir.cpp for define_scalar:

1. Created define_scalar binding
2. Added is_symbolic to Val binding

Created pytorch_utils.py for from_pytorch:

• Added torch_dtype_to_nvfuser_dtype

PR Stack:
#4513 Add bindings for FusionExecutorCache
#4516 Add the remaining binary operations
#4517 Add the bindings for unary operations
#4518 Add the bindings for reduction operations
#4519 Move helper functions from python_frontend to python_common
#4520 Create python reproducer from Fusion IR for nvfuser_direct
#4521 Recreate python_frontend test_basic for nvfuser_direct <<< This PR.

[github-actions]

Review updated until commit c1b0ee3

Description

• Added __repr__, define_scalar, and from_pytorch to FusionDefinition

• Created define_scalar binding in ir.cpp

• Added is_symbolic to Val binding in ir.cpp

• Created pytorch_utils.py for torch_dtype_to_nvfuser_dtype

---

Changes walkthrough 📝

	Relevant files
Bindings	ir.cpp Add define_scalar and is_symbolic bindings                              python/python_direct/ir.cpp Added define_scalar binding Added is_symbolic to Val binding +39/-2
Enhancement	__init__.py Enhance FusionDefinition with new methods                                python/nvfuser_direct/init.py Added __repr__ method Added define_scalar method Added from_pytorch method +67/-0
Utility	pytorch_utils.py Add torch_dtype_to_nvfuser_dtype function                                python/nvfuser_direct/pytorch_utils.py Added torch_dtype_to_nvfuser_dtype function +36/-0    utils.py Add utility functions and NVFuserTest class                            tests/python_direct/utils.py Added is_pre_volta, is_pre_ampere, is_pre_hopper functions Added check_captured_python_definition function Added NVFuserTest class +91/-0
Tests	test_python_frontend.py Add test_basic for nvfuser_direct                                                tests/python_direct/test_python_frontend.py Added test_basic test case +42/-0

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review Overloading The define_scalar function is overloaded with two definitions. Ensure that this overloading does not cause ambiguity or unexpected behavior. void bindScalar(py::module& nvfuser) { nvfuser.def( "define_scalar", [](PolymorphicValue::VariantType value, std::optional<PrimDataType> dtype) { PolymorphicValue cast_value( dtype.has_value() ? castToDtype(std::move(value), dtype.value()) : std::move(value)); PrimDataType value_dtype( dtype.has_value() ? dtype.value() : std::get<PrimDataType>(getDataType(value).type)); return IrBuilder::create<Val>(value, value_dtype); }, py::arg("value"), py::arg("dtype") = std::nullopt, py::return_value_policy::reference); nvfuser.def( "define_scalar", [](PrimDataType dtype) { return IrBuilder::create<Val>(std::monostate{}, dtype); }, py::arg("dtype") = DataType::Double, py::return_value_policy::reference); } Type Handling The torch_dtype_to_nvfuser_dtype function maps Python scalars directly to DataType. Verify that this mapping is correct and does not lead to unintended type conversions. complex: DataType.ComplexDouble, float: DataType.Double, int: DataType.Int, bool: DataType.Bool, Test Coverage The test_basic function checks correctness against PyTorch eager mode. Ensure that additional tests cover more complex scenarios and edge cases. @pytest.mark.skipif(is_pre_volta(), reason="Only supported on Volta and newer devices.") class TestNvFuserFrontend(NVFuserTest): def test_basic(self): inputs = [ torch.ones(2, 4, 8, device="cuda"), torch.ones(2, 4, 8, device="cuda"), ] def fusion_func(fd: FusionDefinition): t0 = fd.from_pytorch(inputs[0]) t1 = fd.from_pytorch(inputs[1]) c0 = fd.define_scalar(3.0) t2 = fd.ops.add(t0, t1) t3 = fd.ops.mul(t2, c0) t4 = fd.ops.sum(t3, [-1], False, DataType.Float) fd.add_output(t4) # Expected Output is a tensor of 48's nvf_out, _ = self.exec_nvfuser(fusion_func, inputs) eager_out = torch.sum((inputs[0] + inputs[1]) * 3.0, dim=-1)