C++ CuTe / CUTLASS vs CuTeDSL (Python) in 2026 — what should new GPU kernel / LLM inference engineers actually learn?

For people just starting out in GPU kernel engineering or LLM inference (FlashAttention / FlashInfer / SGLang / vLLM style work), most job postings still list “C++17, CuTe, CUTLASS” as hard requirements. At the same time NVIDIA has been pushing CuTeDSL (the Python DSL in CUTLASS 4.x) hard since late 2025 as the new recommended path for new kernels — same performance, no template metaprogramming, JIT, much faster iteration, and direct TorchInductor integration. The shift feels real in FlashAtte