[Bizuayeu]

What this changes

The Verified copy drafts section of docs/recipes/glm-5.3-flash-cuda.md said that device exactness and speed were pending (#1910). They are measured now on one checkpoint: copies kept every reply the same token for token, but sped some loads up and slowed others, so TF_GLM_COPIES=1 stays opt-in. The section now says so and gives the rows, in the scope asked for in #1910 (one checkpoint, greedy, two GB10s at TP=2 with --parallel 4, the main commit), so a reader can see where copies pay and where they cost.

Load	1 stream	4 streams, per stream	4 streams, aggregate
A quoted document (~400 tokens back)	+8.1% faster	+0.6% faster	−0.4% slower
Code	+2.3% faster	−1.0% slower	−1.5% slower
One code prompt on every stream	−0.2% inside	+10.1% faster	+11.0% faster
A file edited back	+9.9% faster	+1.5% faster	+2.5% faster
Prose	−3.6% slower	+9.7% faster	+6.5% faster
Prose after a 31K-token prompt	−0.8% inside	+6.7% faster	+4.0% faster
Prose after a 56K-token prompt	−0.6% inside	+6.6% faster	+4.6% faster

• Issue: glm cuda (native): copy drafts are missing, and the 48% acceptance the Python lane gets from them is what keeps it ahead #1910 (the answer to the rows).
• Stacked on: main (3d42a462fb86).
• atc overlap --json (2026-11-14 10:33 UTC, base main 3d42a462fb86; complete: false, with the other pull requests' limit, unavailable pull files, the pull requests' limit and the time limit): 10 open pull requests share a file; atc test-merged them: 7 merge cleanly and 3 (qwen27: serve the 27B pair, refuse what a pair cannot keep, and die with the peer #1434, DeepSeek-V4.1: read the pack's weight spans past the page cache (core/direct_io) #1924, qwen3-asr: host slice, the shared tower and Qwen3 schedules over a backend binding (#2105) #2239) conflict, each in CHANGELOG.md, and each conflicts with main in the same files (git merge-tree). The 9 open pull requests that also change docs/recipes/glm-5.3-flash-cuda.md merge with this branch with the same files as with main: 5 cleanly, kv pool: a named TF_KV_SHARE's slots come where an asked cache goes #3125 and glm cuda: each decode window width keeps its faster expert form #3249 in CHANGELOG.md only, and glm cuda: TF_GLM_GRAPHS=1 replays verify windows as CUDA graphs (#1490) #2279 and cuda: GLM, MiniMax and Nemotron placed from one launch, by flags or aliases (T1b of #2560) #2919 in files this branch does not touch.

Checks on this head

• Head commit: 57c5f9dea166, which only puts this pull request's number in the CHANGELOG line, on 8e567dba1126 (docs only: the recipe's paragraph and table, and one CHANGELOG line). On this head, tools/pr_check.py --base origin/main --pr 3335: 0 errors, 0 warnings (lean 341 vs 341, 0 new).
• Tests run, with the command and pass, fail and skip counts: python -m pytest -q -rs -p no:cacheprovider from the root, on the base and on 8e567dba1126: 459 passed, 86 skipped and 519 subtests passed on each, with the same skip reasons. Main's CI run of the same job on 3d42a462fb86 has 461 passed and 84 skipped: the two more skips here are tools/zig/test_nfkd_tables.py's, since this interpreter has Unicode 15.0.0.
• CI: green on 57c5f9dea166 (runs 43592855621 and 43592855612): benchmark-clients (ubuntu-latest, windows-latest), pytest (461 passed, 84 skipped and 519 subtests passed, as on main), release-paths (ubuntu-24.04), zig, zig-macos-dev-tools, zig-windows and fmt pass; host-nightly-linux and host-nightly-macos were skipped by the workflow.
• Not run, and why: inference, since the change is documentation only; the rows come from the runs in glm cuda (native): copy drafts are missing, and the 48% acceptance the Python lane gets from them is what keeps it ahead #1910.

Receipt

n/a: documentation only, so it cannot reach inference. The rows are my #1910 comment's, measured on main 0fed2e4a5f with Bizuayeu/GLM-5.3-Flash-Jammed-XSFP4 on two GB10s at TP=2, --parallel 4, four fresh launches (off, on, on, off), tools/bench_concurrent.py --alone --serial, greedy, 3 reps, each figure the cell's median. Every item's token_sha256 was one hash over its streams, reps, alone and serial runs and the four launches. The copy-on launches turned copies on through a local default with the same path as TF_GLM_COPIES=1, against TF_GLM_COPIES=0; that comment's "Setup and every row" has each launch's tok/s and range.

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 tests: a recipe paragraph and table, and one changelog line.)
• Comments and docstrings are one line. No measurements or history in the source. (No source changes.)
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.

[ashhart]

Thanks @Bizuayeu. The table keeps the useful gains and the slower cells visible, names the one-checkpoint greedy TP=2 setup, and explains the copy-on default used for those runs. The default remains off, and the text keeps other checkpoints and sampled replies outside that evidence. The figures match the supplied description and the local link checks pass. Ready for this docs update; no device rerun needed.

• Synthena (Ash's Agent)