HipKittens MXFP8 support breaks ROCm 7.0 source build on gfx950

[XinyuJiangCMU]

Building current dev from source fails on ROCm 7.0.0 / gfx950 (MI355X).

The same environment builds dev cleanly from a checkout dated 2026-06-21 or earlier, and current dev builds fine on ROCm 7.2 — so this is a ROCm 7.0 regression. It appears to start at the "HipKittens MXFP8 GEMM Support" change (commit 6dd7bcb).

Environment

• gfx950 (MI355X), ROCm 7.0.0
• pip install . --no-build-isolation, USE_ROCM=ON

Two failures

1. Default build (USE_HIPKITTENS_GEMM=ON) — HipKittens fails in ROCm 7.0 backend codegen:

gemm/kittens/mxfp8_gemm.cpp:59:6: error: illegal VGPR to SGPR copy
fatal error: error in backend: Not supported instr: ...

2. With HipKittens off (-DUSE_HIPKITTENS_GEMM=OFF) — the fallback path doesn't compile:

gemm/rocm_gemm.cu: error: use of undeclared identifier 's'

hipStream_t s is declared only inside the HipKittens #ifdef block but is still used in the #else fallback, so disabling HipKittens leaves s undefined.

Repro

git clone --recursive -b dev https://github.com/ROCm/TransformerEngine
cd TransformerEngine
NVTE_FUSED_ATTN=0 pip install . --no-build-isolation -v                                   # failure 1
NVTE_CMAKE_EXTRA_ARGS="-DUSE_HIPKITTENS_GEMM=OFF" NVTE_FUSED_ATTN=0 pip install . --no-build-isolation -v   # failure 2

Ask

Could you please restore a working source build path for ROCm 7.0 / gfx950?

At minimum, -DUSE_HIPKITTENS_GEMM=OFF should leave a buildable fallback path. If HipKittens requires ROCm 7.2, it would also be helpful to avoid enabling it by default on ROCm 7.0.

Workaround

Pin dev to a commit from 2026-06-21 or earlier, or build on ROCm 7.2.

[wenchenvincent]

@XinyuJiangCMU @lucbruni-amd This issue should be fixed in the latest dev now. Could you try it out?