[imD-5]

Effect

• Cold 64K prefill, two GB10 (TP2): 30.8-31.0 s with TF_DS_RS=1, 31.7-32.0 s with it off, in alternating rounds.
• Replies byte-identical with it on and off (26 greedy prompts), and --alone --serial equal in every cell.
• Decode and the other prefill lengths are within noise.

Reproduce

TF_DS_RS=1 python3 tools/prefill_cold.py run http://<head>:8100 <MODEL> prompts.json out.json   # then TF_DS_RS=0, same prompts rebuilt
python3 tools/bench_concurrent.py http://<head>:8100 <MODEL> --alone --serial

• Fresh prompts and a restart per round (TF_DS_SPILL_GIB=0), two rounds each way.

Change

• NCCL.reduce_scatter and Comm.reduced: prompt partials reach the mHC post as bf16(p0 + p1) from an fp32 reduce-scatter and a bf16 all-gather, 3/4 of the bytes.

Why

• The post kernels add the two ranks' rows and round to bf16, which is one fp32 add of two operands: the same bits in either order.

Type

• type: feature: new code (a feature, tool or add-on)

Tests

• tests/test_dsv41_reduced.py: 5 passed in the engine image, 2 skipped without torch or triton.

Checklist

• No personal data, machine names, internal hosts or local paths. No AI attribution lines.

AI disclosure

AI: Claude Opus 5.5 via Hermes Agent
Checked: by author