Add support for automatic scheduler in direct bindings

[rdspring1]

• Define HeuristicParams struct
• Update manual_execute to take launch_params, compile_params, and scheduler_type from HeuristicParams
• Add can_schedule, find_compatible_schedulers, and schedule operations to schedule submodule.
• Create tests for pointwise, reduction, and inner persistent schedulers

[github-actions]

Review updated until commit 8710824

Description

• Add automatic scheduler support in direct bindings

• Introduce HeuristicParams and derived scheduler parameter types

• Implement schedule, can_schedule, and find_compatible_schedulers operations

• Add comprehensive tests for pointwise, reduction, and persistent schedulers

---

Changes walkthrough 📝

	Relevant files
Enhancement	heuristic_params.cpp Bind HeuristicParams and scheduler-specific parameter types to Python python/python_direct/heuristic_params.cpp Define HeuristicParams class with lparams, cparams, and scheduler_type Add __repr__ and __eq__ methods for string representation and comparison Expose hash function and readonly scheduler_type field to Python Define PointwiseParams and ReductionParams as subclasses of HeuristicParams +33/-0    schedule.cpp Implement Python bindings for scheduler compatibility and execution python/python_direct/schedule.cpp Add can_schedule to check scheduler compatibility with debug messages Implement find_compatible_schedulers to detect valid schedulers Add schedule function to apply a scheduler and return HeuristicParams Include necessary headers for scheduler runtime and type support +121/-0  __init__.py Extend manual_execute to support heuristic-based execution python/nvfuser_direct/init.py Update manual_execute to accept optional heuristic_params Pass heuristic parameters to compile and run when provided Maintain backward compatibility when heuristic_params is None +17/-2    heuristic.h Export HeuristicParams via NVF_API for external use            csrc/scheduler/heuristic.h Add NVF_API macro to HeuristicParams class for symbol export Enable external linkage of HeuristicParams in shared libraries +1/-1      registry.h Export canSchedule function via NVF_API                                    csrc/scheduler/registry.h Add NVF_API to canSchedule function for external access Allow Python bindings to call canSchedule directly +1/-1      runtime_info.h Export SchedulerRuntimeInfo constructor via NVF_API            csrc/scheduler/runtime_info.h Add NVF_API to SchedulerRuntimeInfo constructor Enable external instantiation of runtime info objects +1/-1
Tests	test_tutorial.py Add tests for automatic scheduler functionality                    tests/python/direct/test_tutorial.py Add check_auto_schedule helper to validate scheduler selection Test automatic scheduling for pointwise, reduction, and inner persistent cases Verify correct scheduler rejection with error message checks Compare NVFuser output with eager PyTorch results +152/-1
Bug fix	executor.h Fix scheduler parameter name typo in KernelExecutor            csrc/runtime/executor.h Fix typo in SchedulerType parameter name (sceduler -> scheduler) Maintain API consistency in KernelExecutor::compile method +1/-1

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review Debug Settings Persistence The can_schedule function modifies global debug dump options via DebugDumpOptionsGuard and sets FusionSegmenterLog, but it does not explicitly restore previous settings after use. This could lead to unintended side effects in other parts of the system if the guard does not properly reset the state. DebugDumpOptionsGuard debug_dump_options_guard; DebugDumpOptionsGuard::getCurOptions().set( DebugDumpOption::FusionSegmenterLog); // Send debug messages to stringstream std::stringstream ss; DebugStreamGuard dsg(ss); Default HeuristicParams Handling The manual_execute method accepts an optional heuristic_params argument, but when it is not provided, the code falls back to a default compile path without scheduler type. This may lead to inconsistent behavior between automatic and manual execution paths if the scheduler type is not properly inferred or handled in the fallback case. if heuristic_params is not None: self.ke.compile( self.fusion, inputs, heuristic_params.lparams, heuristic_params.cparams, heuristic_params.scheduler_type, ) else: self.ke.compile(self.fusion, inputs) Incomplete Scheduler Enumeration The check_auto_schedule helper function filters out SchedulerType.none, but uses if not SchedulerType.none which is always true and does not correctly filter the enum value. This may result in incorrect filtering and unintended inclusion or exclusion of scheduler types during testing. if not SchedulerType.none ]

[rdspring1]

!test

[rdspring1]

I will added fields to PointwiseParams and ReductionParams later but the class definition was necessary to pass tests.

[rdspring1]

!test --pybench