add-cuda-kernel

Résultat attendu:

• أضِف نواة CUDA في 10 خطوات:
• 1. Define kernel in include/flashinfer/{name}.cuh using raw pointers
• 2. Create launcher in csrc/{name}.cu with TVM-FFI tensor handling
• 3. Export via TVM_FFI_DLL_EXPORT_TYPED_FUNC in csrc/{name}_jit_binding.cu
• 4. Create JIT generator in flashinfer/jit/{name}.py
• 5. Build Python API in flashinfer/{name}.py with @functools.cache
• 6. Write tests in tests/ directory with pytest.mark.parametrize
• 7. Register in flashinfer/aot.py for pre-compiled packages
• 8. Export from flashinfer/__init__.py
• 9. Add benchmark in benchmarks/ directory
• 10. Run pytest to test; kernel auto-compiles on first use

Résultat attendu:

• الملفات المطلوبة:
• - include/flashinfer/{name}.cuh - تعريف نواة CUDA
• - csrc/{name}.cu - مُشغِّل PyTorch
• - csrc/{name}_jit_binding.cu - ربط TVM-FFI
• - flashinfer/jit/{name}.py - مولّد JIT
• - flashinfer/{name}.py - واجهة Python
• - tests/test_{name}.py - اختبارات وحدات
• - benchmarks/bench_{name}.py - سكربت قياس الأداء