biondizzle 2f68c7ba77 fix: cache E2M1 step_to_idx LUT per device (no CPU->CUDA copy in forward) ...

torch.tensor() and new_tensor() both trigger CPU->CUDA copies during
cudagraph capture. Pre-cache the LUT on first use per device.

2026-05-16 18:48:31 +00:00

..

kernel

refactor: copy CuTeDSL kernel into repo with local imports

2026-05-16 02:57:54 +00:00

__init__.py

refactor: copy CuTeDSL kernel into repo with local imports

2026-05-16 02:57:54 +00:00

bridge.py

fix: cache E2M1 step_to_idx LUT per device (no CPU->CUDA copy in forward)

2026-05-16 18:48:31 +00:00

moe_pipeline.py

fix: same gate/up split fix in moe_pipeline.py

2026-05-16 04:04:53 +00:00