[Bizuayeu]

What this changes

Two sentences at the end of the paragraph #1996 added to the GLM recipe's "Prompt chunk size": what TF_GLM_OVERLAP_PIECES does, and that in #1994's run 4 solo on the wire came 1.7–3.6% sooner at the default 3 pieces and 6.1–7.8% sooner at 8, against the wire without solo, while in run 5 8 pieces without solo stayed within 1% of 3. Documentation only.

• Issue: Refs GLM CUDA at TP=3: under TF_GLM_CHUNK_SOLO=4096, TF_GLM_ROCE_MULTI costs a cold 128K prompt about 7.5%, not 1 to 2% #1994, which has the numbers of runs 4 and 5.
• Base: main 0d43416f9, which has docs: TF_GLM_CHUNK_SOLO gains a little under half at TP=3 on TF_GLM_ROCE_MULTI #1996's paragraph.
• atc overlap --json: it didn't complete (pullFilesLimit, pullFilesUnavailable, timeLimit), so I checked the file by hand: of the 171 open pull requests, only glm cuda: a prompt step takes two chunks while no other stream decodes, each one's exchanges under the other's sites #1423 touches it. glm cuda: a prompt step takes two chunks while no other stream decodes, each one's exchanges under the other's sites #1423 changes the same section about 30 lines above, in another sentence, and the recipe merges without a conflict.

Checks on this head

• Head commit: b4cc532d6
• Lean checker, python3 tools/zig/lean_check.py, on the base and on this head: 424 and 424, no new findings.
• Tests run, with the command and pass, fail and skip counts (Linux x86_64 under WSL2): zig build test 499/499 steps (6,598 of 6,646 tests, 48 skipped). No failures.
• CI: on this head merged onto main b11a8b0a0, five of the seven checks pass (zig, fmt, release-paths and both benchmark-clients). zig-windows hit the job's 60-minute limit inside zig build test, after 35 minutes without output; main's own run at b11a8b0a0 passed it in 23 minutes. zig-macos-dev-tools never started: no hosted macOS runner picked it up (GitHub notes capacity constraints on macOS arm64).
• Not run, and why: the CUDA and Metal steps, as no code changes.

Receipt

n/a: documentation only. The numbers are #1994's runs 4 and 5 (#1994 (comment)): three GB10s at TP=3, Bizuayeu/GLM-5.3-Flash-XSFP4 at 88736f91 with its MTP drafts, --parallel 8, TF_GLM_KV_SHARE=32, TENSORFOLD_PREFILL_SHARE=0.7, TF_GLM_ROCE_MULTI=1, two fresh servers an arm, each cold prompt's first token from tools/bench_cache.py --require-cold, and one token hash for each prompt length across all fourteen servers. The binary is v1.2.1 with one line out that acts only under TF_GLM_PROF=1, which both runs leave unset, so its bits are v1.2.1's. #1996's paragraph above stays run 1's (main 9e8c368b0).

Checklist

• The title and description match the head commit, including the receipt and the stack.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (No new test: documentation only.)
• Comments and docstrings are one line. No measurements or history in the source. (The measurements are in the recipe, as its other rows are.)
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.

[ashhart]

Thanks @Bizuayeu, these two sentences make the recipe's solo paragraph actionable: 8 pieces recovers most of solo's gain on the wire, and they don't move anything without solo. The setting's default and range match the code. This can land as it is.

[ashhart]

Thank you @Bizuayeu. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.