[jschmied]

What

Second step of #1409, on main after #1440. Pure refactor: cuobjdump -sass is identical for every function of every image that includes a touched file.

The last three float warp_sum copies (xor butterfly 16 down to 1, the same loop as common/reduce.cuh) give way to tf_common::warp_sum through #include "common/reduce.cuh" and a using declaration, as in #1440:

• flashnext_batch.cu (inside fn_gdn)
• flashnext_gdn.cu
• qwen4_gdn.cu, the checked-in text Gdn.cuda generates from flashnext_gdn.cu (source_test.zig compares them), so both change together.

qwen4_gdn_cuda_host.zig, which turns that text into the CPU twin, now strips the include and the using declaration instead of the deleted function, the way qwen4_ple_cuda_host.zig handles glue_math.cuh in #1440. flashnext_gdn.cu's gradient-serving pin takes the new bytes. flashnext_batch.cu's two pins already differ on main and are left as they are.

5 files, +9 / −15.

Left alone on purpose

• sample.cu's warp_sum_u: unsigned, a different type.
• deepseek_v41/sampling.cu: double precision.
• glm_kda.cu's warp_sum4: a different signature and reduction.

Verification

• SASS, base main dd0a9cc5e, head aa9184554, nvcc 13.0: 3 images (flashnext_batch, flashnext_gdn, qwen4_gdn), 65 functions, sm_80/86/89/120/121: 0 changed, 0 added, 0 removed, 0 unstable. Before the rebase the same set was also identical with nvcc 13.2. The qwen4_gdn image is built from the checked-in unit, as the tool does for generated units.

  python3 tools/zig/cuda_sass_identity.py --repo . -j 3 --receipt sass.tsv origin/main HEAD
  

• Planted control: a reversed warp_sum butterfly changes 35 functions in 3 of 3 images. The fp4pair control cannot be planted because none of these units includes common/qmm_frag.cuh, so the tool exits 1 on that line alone.
• lean: 425 problems, 425 on the base, 0 new.
• zig build test -Dnvcc=/usr/local/cuda/bin/nvcc: passes (RC 0).

Part of #1409.

[ashhart]

Thanks @jschmied, landed with #1550, #1555 and #1556 together on main (changelog in 5b6c254). We rebuilt every CUDA image from main and from the stack at sm_121 and sm_80: all 1,866 functions in the 104 images compile to identical SASS.