[Klaud Cold] MI325X Qwen3.8-27B BF16 MTP, synthetic AL 2.51 / MI325X Qwen3.8-27B BF16 MTP，合成 AL 2.51

[Oseltamivir]

Goal: Rerun MI325X Qwen3.8-27B BF16 with thinking on, synthetic AL 2.51 and 3 native MTP draft tokens.
Baseline recipe: #3295 · vllm/vllm-openai-rocm:nightly-a8d1aa9c99b8698a2a78b611b7a10c30e6b3995b
1k/1k · TP1 · concurrency 1, 2, 4, 8, 16, 32, 64, 128.

qwen3.827b-bf16-mi325x-vllm-mtp uses Qwen/Qwen3.8-27B, BF16 target/native MTP weights and BF16 KV cache. The image, runner, graph mode, prefix-cache setting and concurrency-sized scheduler follow the source recipe.

Throughput sets rejection_sample_method: synthetic and synthetic_acceptance_length: 2.51, using the BF16 thinking_on[3] measurement in #3304. Require THINKING_MODE=thinking_on; explicitly enable thinking in server chat-template defaults. The fixed-sequence client retains --use-chat-template, whose checkpoint default renders a thinking-on prompt. Runs requesting accuracy through EVAL_ONLY or RUN_EVAL retain real verification.

Validation: Bash syntax, whitespace, append-only changelog and full matrix checks pass (8 throughput points; the current 1k/1k policy selects no default evals). All four actual recipes reject thinking-off before GPU startup. Rendering the checkpoint template confirms its default equals explicit thinking-on. The previous real-verification sweep passed. Synthetic sweep dispatched: run 35495280070 on 452108aed28ef506babacb15ff92f893fa38a178.

AI model disclosure

GPT-6 prepared the BF16 and synthetic-acceptance changes and validation. The exact runtime model/version identifier was not exposed and could not be verified. The source FP8 recipe was credited to Claude Code; its underlying model/version was not disclosed and could not be verified.

中文

**目标：**重跑 MI325X Qwen3.8-27B BF16，使用 thinking 开启、合成 AL 2.51 和 3 个原生 MTP 草稿 token。
基线配方：#3295 · vllm/vllm-openai-rocm:nightly-a8d1aa9c99b8698a2a78b611b7a10c30e6b3995b
1k/1k · TP1 · 并发 1、2、4、8、16、32、64、128。

qwen3.827b-bf16-mi325x-vllm-mtp 使用 Qwen/Qwen3.8-27B，目标权重、原生 MTP 权重及 KV cache 均为 BF16。沿用原配方的镜像、runner、graph 模式、prefix-cache 设置和按并发配置的 scheduler。

吞吐测试设置 rejection_sample_method: synthetic 和 synthetic_acceptance_length: 2.51，采用 #3304 中 BF16 的 thinking_on[3] 测量值。要求 THINKING_MODE=thinking_on，服务端默认 chat-template 参数显式启用 thinking。固定序列长度客户端保留 --use-chat-template，checkpoint 默认模板会生成 thinking 开启的提示词。通过 EVAL_ONLY 或 RUN_EVAL 请求准确率评测时保留真实验证。

验证：Bash 语法、空白、changelog 只追加约束及完整矩阵检查均通过（8 个吞吐测试点；当前 1k/1k 策略默认不选取 eval）。实际执行四个配方，均在 GPU 启动前拒绝 thinking 关闭。实际渲染 checkpoint 模板，确认默认行为与显式开启 thinking 相同。此前的真实验证 sweep 已通过。合成 AL sweep 已启动：run 35495280070，提交为 452108aed28ef506babacb15ff92f893fa38a178。

**AI 模型披露：**GPT-6 完成 BF16、合成接受长度修改与验证；运行时未提供精确模型/版本标识，无法核实。原 FP8 配方署名 Claude Code，未披露底层模型/版本，无法核实。

[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。

[github-actions]

View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=35495280070

View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=35495280070

[functionstackx]

InferenceX has switched away from unmaintainable bash scripts to YAML files that don't repeat the same stuff over and over again. Please merge the latest main into this PR: we have migrated single-node AgentX onto native srt-slurm (#3428), so AgentX configs are now declarative YAML recipes, not per-config 1000+ line bash slop scripts. Please also delete the old benchmarks/single_node/** scripts (see this recipe for the new format).

[functionstackx]

Sorry, over the weekend, there was 2 major refactors to clean up the technical debt accumalated over the past 11 months of moving at the speed of light. We don't see any major refactors in the forthseeable future besides cleaning up AMD multinode AgentX pile of bash. As much, due to the refactors, u would need to ask your agent to rebase from remote main@latest. Thank you in advance for ur understanding