[imD-5]

What this changes

tools/dsv41/REPORT.md: section 6's open item on prefill overlap becomes its result, and the launch line's chunk size
goes from 2048 to 4096 rows.

Effect

• Two GB10 (TP2), DeepSeek-V4.1-Flash-1017 EXL3 2.9 bpw, tools/prefill_cold.py, cold 64K prompt, spill off, fresh prompts a round.
• Chunk 2048 (today's line): 30.87 s.
• Chunk 4096: 28.14 s, and 16K is 7.35 s against 7.84 s, 32K is 14.40 s against 15.49 s.
• Decode is unchanged: 60.4 / 59.9 / 64.5 / 69.2 tok/s.
• Chunk 8192: does not run, because its MoE step needs about 1.3 GiB more than engine._memory_ceiling leaves.

Why

• A chunk is one prompt pass of TF_DS_PREFILL_CHUNK rows, one row a prompt token, so a bigger chunk gives each expert launch more rows and reads the weights once for more of them.
• The gather overlap was built as the open item described (two halves, one stream for every collective, stage-major) and measured 0%: half 1's attention waits for half 0's ring writes, and splitting the rows halves each launch.

Type

• type: results: run results or a close-out

Receipt

• Exactness: the chunk size is not bit-neutral, so 4096 gives other tokens than 2048, reproducible within a configuration (26 greedy prompts, same file on two runs).
• Not run: other checkpoints or families; a single GB10, where a 4096-row chunk has less headroom.

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: (pending the author's review)