Cherry-pick FP4 acts + MXF4 + DG_USE_FP8_COMBINE; wrap mega_moe_pre_dispatch in tvm-ffi#31

Closed

Fridge003 wants to merge 5 commits into

release-0426from

Commits on May 12, 2026

Add FP4 acts + MXF4 kind support and fused mega_moe_pre_dispatch kernel (#27)

authored and
Fridge003
committed
Add DG_USE_FP8_COMBINE: FP8 + per-row UE8M0 SF on the second a2a (combine path) (#28)

authored and
Fridge003
committed
tvm-ffi: wrap mega_moe_pre_dispatch and export from sgl_deep_gemm

Baizhou Zhang
authored and
Fridge003
committed
tests: add sgl_deep_gemm smoke test for tvm-ffi w4a4 wrapping

Baizhou Zhang
authored and
Fridge003
committed
convert_to_torch_tensor: handle zero-numel tensors with empty CUDA tensor

Baizhou Zhang
authored and
Fridge003
committed