cutedsl_megamoe

Replace the five kernel packages under src/ verbatim with the new drop, then audit the shim imports and kernel construct/launch signatures against the new sources. Finish by running the multi-rank cutedsl test suite with torchrun on 4 Blackwell GPUs.

src/ is a verbatim copy of the NVIDIA kernel-team drop that must never be edited, while shim/ contains all FlashInfer adaptation code such as path bootstrap, symmetric-buffer wrappers, tuners, and autotuners. Backends import only the package __init__, never src/ directly.

nvidia-cutlass-dsl 4.5.2 is the performance floor: 4.5.0 fails at cute.compile, and 4.5.2 needs the MR!27 mainloop workaround while 4.5.3 and later run at full speed natively. The shim warns when an older runtime is detected.

The highest-churn surface is the kernel constructor and launch kwargs, where arguments change without names changing, so a symbol-existence grep misses it. The shim's _ensure_mega_compiled and _build_mega_runtime_kwargs must be re-audited against the kernel __init__ and __call__ signatures on every drop.

No, both trees expose the same top-level module names, so bootstrap_paths raises if the sibling tree's modules are already imported. Use a separate process per architecture, which matches hardware reality since a process runs on either Hopper or Blackwell.