[jschmied]

What

Part of #1409. Stacked on #1556, with one commit on top, because #1556 edits the same two files (sigm → sigmoidf_).

train_attn.cu (4-bit training, loaded by src/cuda/train_attn.zig) was a copy of the attention part of attn_train.cu (Living Weights, loaded by qwen3_5's cuda_train_ops.zig). It has the same four kernels under tf_gattn_ instead of tf_attn_ names (prep, q, gate_back, prep_back) and the same helpers; only the comments differ. A SASS census of every image at sm_121 found the four functions identical across the two images. That census is how this copy showed up: K0's text comparison missed it because of the names.

The text now lives once, in attn_back.cuh. Each unit names its kernels with #define TF_ATTN_BACK(name) tf_attn_##name (or tf_gattn_##name) before the include. Both images keep their symbols, so the loaders, contract records (gradient_pending.zig), tests and receipts stay as they are. attn_train.cu keeps tf_sum_parts and tf_axpy after the include.

3 files, +167 / −324.

Verification

• SASS: base cuda: sigmoidf_, softplusf_, silu and silu_grad in common/act.cuh, SASS unchanged #1556's head f19336386, head b50786bfb, nvcc 13.0, sm_80/86/89/120/121. Both images (attn_train, train_attn), all 50 functions: 0 changed, 0 added, 0 removed, 0 unstable.

  python3 tools/zig/cuda_sass_identity.py --repo . -j 3 --receipt sass.tsv f19336386 HEAD
  

• Planted control: a reversed warp_sum butterfly changes 30 functions in 2 of 2 images. The fp4pair control cannot be planted, because neither image includes common/qmm_frag.cuh, so the tool exits 1 on that line alone.
• zig build test-gradient-serving: the same set of differing pins as main. Neither file is pinned.
• lean: 425 problems, 425 on the base, 0 new.
• zig build test -Dnvcc=/usr/local/cuda/bin/nvcc: passes (RC 0). It ran before a comment-only fix (two header comment lines folded into one), and the SASS was re-run after it.

Not included

The census also found two pairs of tiny kernels that compile to the same code: ck_lane's tf_ck8_reduce with qmmf's reduce_kernel<false>, and DS4.1's cast_bf16_kernel with GLM vision's glm_vis_to_bf16. Each is 4 to 6 lines, with different parameter types in different families, so a shared body would add more text than it removes. I left them alone.

[ashhart]

Thank you @jschmied. This is in TensorFold 1.2.0, released today: https://github.com/ashhart/TensorFold/releases/tag/v1.2.0

It landed with your commits under your own name, so you're listed as a TensorFold contributor, and it's credited to you in the release notes and the changelog.