[Klaud Cold] feat(golden-al): add GLM-5.2 MTP golden AL distribution / 新增 GLM-5.2 MTP 黄金接受长度(AL)分布

[functionstackx]

Thanks to @qiching @xinli-sw !

Summary

Adds the GLM-5.2 MTP golden acceptance-length (AL) distribution to golden_al_distribution/ — the committed reference curve for synthetic-acceptance speculative decoding on GLM-5.2, per the AgentX fairness guidelines.

• Measurement: SPEED-Bench coding category, temperature 1.0, top_p 0.95, output_len 4096, thinking_on (chat_template_kwargs: {"enable_thinking": true}), on B300 with vLLM MTP, per num_speculative_tokens 1–8.
• Provenance: values are verbatim from the speedbench-al artifact of the SpeedBench AL Collection run 28058352479 (source line embedded in the YAML header, matching the convention of the existing golden files).
• Checkpoint: measured on the GLM-5.2 FP8 checkpoint; one curve is committed per model, and the NVFP4/MXFP4 checkpoints ship the same nextn head.

Why a standalone PR

This curve is split out of #2447 so it can be reviewed against the golden-curve checklist on its own, and so the recipe PRs that consume it can land independently. Current consumers:

PR	SKU	Precision	SGLANG_SIMULATE_ACC_LEN
#2447	B300	NVFP4	2.99
#2448	B200	NVFP4	2.99
#2459	MI355X	MXFP4	2.99

All three pin SGLANG_SIMULATE_ACC_LEN=2.99 — the K=3, thinking_on value from this curve. PRs #2447 and #2448 still carry copies of these same golden-curve files and must drop them once this PR lands.

Data-only change (no benchmark configs touched); [skip-sweep] on the head commit.

中文说明

新增 GLM-5.2 MTP 黄金接受长度(AL)分布至 golden_al_distribution/——按照 AgentX 公平性准则，作为 GLM-5.2 合成接受率投机解码的统一参考曲线。

• 测量条件：SPEED-Bench coding 类别、temperature 1.0、top_p 0.95、output_len 4096、thinking_on（chat_template_kwargs: {"enable_thinking": true}），B300 + vLLM MTP，覆盖 num_speculative_tokens 1–8。
• 数据来源：数值与 SpeedBench AL Collection 运行 28058352479 的 speedbench-al 产物逐字节一致（YAML 头部已按现有黄金文件约定嵌入 Source 链接）。
• 模型检查点：基于 GLM-5.2 FP8 模型测量；每个模型仅提交一条曲线，NVFP4/MXFP4 量化使用同一 nextn 头。

纯数据变更（不涉及基准测试配置）；头部提交带 [skip-sweep] 标记。

Reviewer checklist

Before accepting this curve, please verify (per golden_al_distribution/README.md):

• every requested draft length and thinking mode completed;
• detailed outputs are coherent and use the intended thinking mode;
• server logs contain no fallback, draft-disable, or chat-template errors;
• the YAML metadata matches the dispatched image, sampling settings, model, and speculative method;
• the source Actions run is linked at the first line of the YAML; and
• the committed values exactly match the workflow artifact.

🤖 Generated with Claude Code

[claude]

LGTM — data-only addition of the GLM-5.2 MTP golden AL curve, following the established file format.

What was reviewed: compared glm5.2_mtp.yaml against the other golden-curve YAMLs (qwen3.5_mtp.yaml, kimik2.5_eagle3.yaml, minimaxm3_eagle3.yaml) for structural and formatting consistency; checked the README/README_zh table additions link to the same source Actions run and file as the new YAML; noted the file only ships a thinking_on data block even though the header comments document both thinking_on/thinking_off chat_template_kwargs — the PR description only claims a thinking_on measurement, so this appears to be an intentionally partial collection (all three downstream consumer PRs also only pin the thinking_on, K=3 value) rather than a missing block.

Extended reasoning...

Overview

This PR adds a new golden acceptance-length reference file (golden_al_distribution/glm5.2_mtp.yaml) plus one-line table entries in README.md/README_zh.md pointing to it. No benchmark configs, workflow code, or CI logic are touched — it is a pure data/documentation change.

Security risks

None. The change adds static YAML data and Markdown table rows; there is no executable code path, no secrets, and no user input.

Level of scrutiny

Low. This is exactly the kind of small, mechanical, well-precedented change (new file following an existing template, one README row added in two language files) that doesn't need deep design review. I compared the new YAML's structure, comment header conventions, and key naming against the three existing golden curve files and found it consistent (source-run link on line 1, sampling metadata in comments, model: thinking_mode: {level: AL} nesting).

Other factors

The one point worth noting for the record: unlike the three existing golden files, which each carry both thinking_on and thinking_off data blocks, this file only includes thinking_on despite documenting both sets of chat_template_kwargs in the header comments. However, the PR description explicitly scopes the measurement to thinking_on only, and all three listed downstream consumers (#2447, #2448, #2459) pin only the thinking_on, K=3 value (2.99), so this looks like an intentionally partial collection rather than an omission — a bug-hunting pass already raised and ruled out this same concern. No outstanding review comments are visible in the PR timeline, and the reviewer checklist items around provenance and source-run linkage all check out from the diff.