[Klaud Cold] Add H200 Qwen3.8-27B-FP8 vLLM TP1 native MTP 1k1k / 新增 H200 Qwen3.8-27B-FP8 vLLM TP1 原生 MTP 1k1k 配方

[functionstackx]

Add qwen3.827b-fp8-h200-vllm-mtp: Qwen3.8-27B-FP8 (Qwen/Qwen3.8-27B-FP8, fp8 e4m3 weights with dynamic activation scaling) on a single H200 (141 GB), served by vLLM nightly-cd10ed6f9f6b37a8ace9cf380007e66fe12ec0c3 with TP1 and the checkpoint's native MTP head (mtp_num_hidden_layers=1, the mtp.* tensors ship in the repo). No external drafter and no --enforce-eager. Single-turn 1k1k only, concurrency 1-128. Sibling of the bf16 DSpark recipe in the qwen3.8-27b PRs; same qwen3.827b model-prefix, distinguished by precision.

Serve shape: --speculative-config '{"method":"mtp","num_speculative_tokens":3}' (the setting the vLLM recipe page gives for this model; it reports fp8 MTP acceptance 0.771), fp8 KV cache, --max-num-seqs sized to the sweep point (floor 16, 256 for accuracy evals: vLLM's default 1024 exceeded the GDN cache blocks next to the weights on H100 in run 35357364404), --no-enable-prefix-caching so every request prefills its full prompt, --language-model-only (text-only serving of Qwen3_5ForConditionalGeneration), --reasoning-parser qwen3, --tool-call-parser qwen3_xml, and the matrix context as --max-model-len. Prompts go through the chat template so MTP acceptance reflects real text rather than random tokens; acceptance is real. The MODELS.md / MODELS_zh.md rows for Qwen3.8-27B land with the bf16 H100 PR (#3260).

Validation: bash -n, YAML parse, exact-key matrix generation, and changelog append (no deletions) pass. GPU validation pending; full-sweep-fail-fast is set on this PR.

One PR per SKU.

中文

新增 qwen3.827b-fp8-h200-vllm-mtp：Qwen3.8-27B-FP8（Qwen/Qwen3.8-27B-FP8，fp8 e4m3 权重、动态激活缩放）在单张 H200（141 GB）上由 vLLM nightly-cd10ed6f9f6b37a8ace9cf380007e66fe12ec0c3 以 TP1 提供服务，并使用 checkpoint 自带的原生 MTP 头（mtp_num_hidden_layers=1，仓库内含 mtp.* 张量）。无外部草稿模型，也不使用 --enforce-eager。仅单轮 1k1k，并发 1-128。与 qwen3.8-27b 各 PR 中的 bf16 DSpark 配方为姊妹配方；沿用 qwen3.827b model-prefix，以精度区分。

服务参数：--speculative-config '{"method":"mtp","num_speculative_tokens":3}'（vLLM recipe 页面为该模型给出的设置；其报告 fp8 MTP 接受率 0.771），fp8 KV cache，--max-num-seqs 按并发点设定（下限 16，精度评测为 256：运行 35357364404 中 vLLM 默认的 1024 超过 H100 上权重旁的 GDN cache block 数），--no-enable-prefix-caching 使每个请求完整 prefill 其提示词，--language-model-only（仅文本服务 Qwen3_5ForConditionalGeneration），--reasoning-parser qwen3，--tool-call-parser qwen3_xml，--max-model-len 使用矩阵上下文。提示词经 chat template 处理，使 MTP 接受率反映真实文本而非随机 token；接受率为真实值。 Qwen3.8-27B 的 MODELS.md / MODELS_zh.md 行随 bf16 H100 PR（#3260）提交。

验证：bash -n、YAML 解析、精确 key 矩阵生成、changelog 追加（无删除）均通过。GPU 验证待完成；本 PR 已添加 full-sweep-fail-fast。

每个 SKU 一个 PR。

🤖 Generated with Claude Code

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps only run on labeled PRs. Add full-sweep-fail-fast (strongly recommended); use full-sweep-enabled only when matrix jobs should continue after a failure.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**扫描仅在带有标签的 PR 上运行。强烈建议添加 full-sweep-fail-fast；仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。

[claude]

Code review found no issues

No high-confidence issues detected in this change.

This review covers commit 43f9940, which is no longer the latest commit on this pull request; later commits are not covered by it.

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=35465800643

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=35465800643

[functionstackx]

/reuse-sweep-run 35465800643