biondizzle f19932d8db fix: compile CuTeDSL kernel once per process, not per MoE layer ...

The warmup was running for every MoE layer (61 layers × 8 ranks = 488
compile attempts). The kernel is cached after the first compile —
subsequent calls are instant. But the print spam was insane.

Now uses a class-level flag to compile exactly once per process.
All 61 layers on a rank share the same compiled kernel.

2026-05-16 05:16:53 +00:00

..

patches

fix: compile CuTeDSL kernel once per process, not per MoE layer

2026-05-16 05:16:53 +00:00

nvfp4_cutedsl.py

fix: L1 gate/up split — intermediate_size is per-projection, not fused

2026-05-16 04:04:40 +00:00