[Bizuayeu]

What this changes

Under TF_GLM_KV_SHARE, an idle server whose prompt cache keeps pool pages logs kv pool idle (UNEVEN), because GLM's report also asks for free == total; with this change it reads Pool.idle()'s free + kept == total there, Spark 25's form from 004655c64, and is unchanged without the share. Only the log line changes: no pool, page or token.

Build	TF_GLM_KV_SHARE	Rank 0's line after the last stream closes
main 532589e40	16	kv pool idle (UNEVEN): promised 0, free 6131 + store 16 == total 6147
This head	16	kv pool idle: promised 0, free 6131 + store 16 == total 6147
This head	off	kv pool idle: promised 0, free 6147 + store 0 == total 6147

• Issue: n/a, a log line's word found in a server's log; no issue was filed.
• Stacked on: main (this head is on 532589e40 and merges clean with 071cefe08).
• atc overlap --json: against main 071cefe08, no open pull request found on zig/src/families/glm/cuda_pages.zig (atc who agrees, though ATC could not read a few pulls' file lists). native cuda: TF_KV_SHARE=n keeps a paged family's prompts as its core pool's pages in one prompt_radix store (Kolibri-1 first) #1755 and gemma4 CUDA: the global layers' caches in pages of the core pool, every stream leasing; sliding layers keep their rings #1619, which put other families' prompts in pool pages, do not touch the file.

Checks on this head

• Head commit: 72a78e008
• 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-cuda-host 291/291 steps (4,483 of 4,499 tests, 16 skipped); zig build test 487/487 steps (2,035 of 2,065 tests, 30 skipped); zig build test-kv-pool 33/33 tests; the GLM host steps (test-glm-prompt-settings-host 6/6, test-glm-tp3-host 8/8, test-glm-pages-host, test-glm-fp8-kv-host, test-glm-prompts-host 7/7); zig build compile-cuda-tests 10/10 steps, also with -Dtarget=aarch64-linux-gnu (the CUDA runner, which CI does not build); zig build compile-config-server 7/7; zig build test-roce-multi-host 40/40 tests; zig build test-tp 33/33 steps; zig fmt --check clean. No failures.
• CI: green, all seven checks on this head.
• Not run, and why: Metal, as this file is CUDA's. No test reaches the line: report returns at once under the test build, as Spark 25's does, so the table above is its before and after.

Receipt

• Environment, with the TensorFold commit, Zig and backend toolchain, machine and GPU, checkpoint and revision: main 532589e40 and this head 72a78e008, each built with zig build native (Zig 0.17.0, CUDA 13.3, sm_121); three NVIDIA GB10s at TP=3 over NCCL 2.30.7, driver 580.178.04, --parallel 4 with paged KV; GLM-5.3-Flash XSFP4 (Bizuayeu/GLM-5.3-Flash-XSFP4 at 88736f91). Each launch ran tools/bench_cache.py --require-cold --repeats 1 --streams 1 on an 8K-token prompt (a cold request, a repeated one and one concurrent), and rank 0 printed the same line after each.
• Paths and platforms the change reaches, which I ran and which I could not: GLM's CUDA engine with paged KV, with the share on and off; I ran both on CUDA (the table above). Metal is not reached.
• Exactness: n/a, the change is one log line's word and no token passes through it.
• Decode speed before and after: n/a, as above.
• Prompt speed before and after: n/a, the line prints when the last stream closes, after prompt processing.
• Licence of any new weights or drafter: n/a.

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: report does not run under the test build. The table above is the before and after.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.
• Every token still goes through the lane rounds. No serial path, and nothing that needs drafts off. (A log line.)
• Drafted output equals "draft": false, a resumed prompt equals a fresh one, and concurrent equals solo. (A log line.)
• No precision traded for speed. If the bits change, the description says which sums change and why. (No bits change.)
• Decode and prompt processing are no slower than main at one stream and the widest admitted; with --drafter, no slower than without it, greedy and sampled (the change cannot reach either: the line prints when the last stream closes).

[ashhart]

Thanks @Bizuayeu. An independent review on current main (d38bba3):
Verdict: LAND.

Applying it: cleanly on main d38bba3. The lean check stays at 424, the same as main. The author line is a real
GitHub identity (@Bizuayeu, noreply address).

Nothing was built or run here.

What holds

• One expression changes, in Pages.report (zig/src/families/glm/cuda_pages.zig):
  p.pool.idle() and c.free == c.total becomes p.pool.idle() and (p.e.share_pages or c.free == c.total).
• Pool.idle() already requires nothing promised, nothing leased and free + kept == total. Under
  TF_GLM_KV_SHARE the prompt cache holds pool pages, so kept can be above 0 on a balanced idle pool, and main
  printed (UNEVEN) for it. The new form is Spark 25's (spark25/backend.zig report), word for word in meaning.
• Off is main exactly. share_pages is set only by share() in cuda_native.zig (rank 0) and from rank 0's
  flag in cuda_follow.zig. With the share off it is false and the expression is main's.
• Only rank 0 reaches report (close returns first on other ranks), and only the log word changes. No page,
  pool state or token passes through it.
• No new test: report returns at once under the test build, the same as Spark 25's. The author's table (GB10,
  TP=3, GLM-5.3-Flash, share on and off) is the before and after.

Small point (not blocking): the receipt was taken on an older main (532589e). The changed lines are identical
on d38bba3, so it carries over.

Device run: none needed beyond the author's.

[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.

[ashhart]

Thanks @Bizuayeu! This is in TensorFold 1.2.2, out now: https://github.com/ashhart/TensorFold/releases/tag/v1.2.2