[Klaud Cold] Fix benchmark_lib source paths broken by #1860 script move / 修复 #1860 移动脚本导致的 benchmark_lib 引用路径错误

[functionstackx]

Summary

• PR fix: move dsv4-fp4-mi355x MTP scripts into fixed_seq_len/ #1860 moved dsv4_fp4_mi355x_{atom,sglang,vllm}_mtp.sh from benchmarks/single_node/ into fixed_seq_len/ as a pure git mv, leaving their source "$(dirname "$0")/../benchmark_lib.sh" one level short — all three configs have failed at launch with No such file or directory ever since (surfaced by this #1981 sweep failure).
• Fixes the atom_mtp and sglang_mtp scripts to ../../benchmark_lib.sh (matching the other 151 fixed_seq_len scripts); the vllm_mtp copy is fixed on PR [Klaud Cold] [AMD] DeepSeek-V4 FP4 MI355X vLLM MTP: bump image to latest nightly / DeepSeek-V4 FP4 MI355X vLLM MTP：升级镜像至最新 nightly #1981's branch.
• Appends a perf-changelog.yaml entry for dsv4-fp4-mi355x-atom-mtp and dsv4-fp4-mi355x-sglang-mtp so both configs re-run and restore their data.

中文说明

• fix: move dsv4-fp4-mi355x MTP scripts into fixed_seq_len/ #1860 将 dsv4_fp4_mi355x_{atom,sglang,vllm}_mtp.sh 以纯 git mv 移入 fixed_seq_len/，其 source "$(dirname "$0")/../benchmark_lib.sh" 相对路径少了一级 - 自那以后三个配置全部在启动时报 No such file or directory（由 [Klaud Cold] [AMD] DeepSeek-V4 FP4 MI355X vLLM MTP: bump image to latest nightly / DeepSeek-V4 FP4 MI355X vLLM MTP：升级镜像至最新 nightly #1981 的 sweep 失败暴露）。
• 将 atom_mtp 与 sglang_mtp 脚本修正为 ../../benchmark_lib.sh（与其余 151 个 fixed_seq_len 脚本一致）；vllm_mtp 已在 PR [Klaud Cold] [AMD] DeepSeek-V4 FP4 MI355X vLLM MTP: bump image to latest nightly / DeepSeek-V4 FP4 MI355X vLLM MTP：升级镜像至最新 nightly #1981 分支上修复。
• 为 dsv4-fp4-mi355x-atom-mtp 与 dsv4-fp4-mi355x-sglang-mtp 追加 perf-changelog.yaml 条目，重新运行这两个配置以恢复数据。

🤖 Generated with Claude Code

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-enabled or full-sweep-fail-fast label to this PR — the benchmark sweep only runs on labeled PRs.

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-enabled 或 full-sweep-fail-fast 标签 — 基准测试 sweep 仅在带有标签的 PR 上运行。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档

[github-actions]

see unofficial run visualizer at https://inferencex.semianalysis.com/inference?unofficialRun=28719812805
see unofficial run visualizer at https://inferencex.semianalysis.com/evaluation?unofficialRun=28719812805

[claude]

🔴 The --chat-template argument at benchmarks/single_node/fixed_seq_len/dsv4_fp4_mi355x_sglang_mtp.sh:205 is missing ../ — it still points at $(dirname "$0")/chat_templates/deepseek_v4_thinking.jinja, which resolves to the nonexistent benchmarks/single_node/fixed_seq_len/chat_templates/… after the #1860 move (the file only exists at benchmarks/single_node/chat_templates/…). This is the exact same off-by-one relative-path defect this PR fixes on the source line above it, and the sibling non-MTP script dsv4_fp4_mi355x_sglang.sh:95 already uses the correct ../chat_templates/… form — please fix line 205 to "$(dirname "$0")/../chat_templates/deepseek_v4_thinking.jinja" in the same commit, otherwise the dsv4-fp4-mi355x-sglang-mtp re-run appended in perf-changelog.yaml will still fail at launch (chat-template file not found).

---

中文翻译: benchmarks/single_node/fixed_seq_len/dsv4_fp4_mi355x_sglang_mtp.sh:205 的 --chat-template 参数少了一级 ../，仍然指向 $(dirname "$0")/chat_templates/deepseek_v4_thinking.jinja；在 #1860 移动脚本后，这会解析到不存在的 benchmarks/single_node/fixed_seq_len/chat_templates/…（该模板文件只存在于 benchmarks/single_node/chat_templates/…）。这与本 PR 修复的 source 行属于同一类相对路径错位问题，同目录的非 MTP 脚本 dsv4_fp4_mi355x_sglang.sh:95 已经使用正确的 ../chat_templates/… 形式 — 请在同一个 commit 中将第 205 行改为 "$(dirname "$0")/../chat_templates/deepseek_v4_thinking.jinja"，否则 perf-changelog.yaml 中追加的 dsv4-fp4-mi355x-sglang-mtp 重跑仍会在启动时报错（找不到 chat-template 文件）。

Extended reasoning...

What the bug is

PR #1860 git mvd dsv4_fp4_mi355x_sglang_mtp.sh from benchmarks/single_node/ into benchmarks/single_node/fixed_seq_len/. This PR fixes one of the resulting relative-path breaks — the source "$(dirname "$0")/../benchmark_lib.sh" line — but misses a second, identical break at line 205:

--chat-template "$(dirname "$0")/chat_templates/deepseek_v4_thinking.jinja" \

Since $(dirname "$0") now resolves to benchmarks/single_node/fixed_seq_len/, this concatenates to benchmarks/single_node/fixed_seq_len/chat_templates/deepseek_v4_thinking.jinja — a path that does not exist. The template file lives only at benchmarks/single_node/chat_templates/deepseek_v4_thinking.jinja (verified: find benchmarks -name deepseek_v4_thinking.jinja yields exactly one match, and ls benchmarks/single_node/fixed_seq_len/chat_templates errors with No such file or directory).

The expected form is already in the same directory

The sibling non-MTP script benchmarks/single_node/fixed_seq_len/dsv4_fp4_mi355x_sglang.sh:95 was correctly updated for the #1860 move:

--chat-template "$(dirname "$0")/../chat_templates/deepseek_v4_thinking.jinja" \

Every other sglang script referencing this template (agentic/dsv4_fp4_mi355x_sglang.sh:147, agentic/dsv4_fp4_b200_sglang.sh:182, agentic/dsv4_fp4_b300_sglang.sh:180) uses ../chat_templates/…. Only dsv4_fp4_mi355x_sglang_mtp.sh:205 is missing the extra ../ — the exact same defect class as the source line this PR fixes.

Step-by-step proof this breaks after merge

1. Sweep launches dsv4-fp4-mi355x-sglang-mtp (forced to re-run by the perf-changelog.yaml entry appended in this PR).
2. dsv4_fp4_mi355x_sglang_mtp.sh now sources ../../benchmark_lib.sh — succeeds (that is what this PR fixes).
3. Script proceeds to python3 -m sglang.launch_server … --chat-template "$(dirname "$0")/chat_templates/deepseek_v4_thinking.jinja".
4. $(dirname "$0") → benchmarks/single_node/fixed_seq_len (script-local, CWD-independent).
5. Resolved path: benchmarks/single_node/fixed_seq_len/chat_templates/deepseek_v4_thinking.jinja.
6. That path does not exist. SGLang’s --chat-template loader opens the file at server startup, so the server crashes before becoming ready and wait_for_server_ready times out. The config fails again — just in a new way. (Or worse, if SGLang silently falls back to no template: MTP acceptance regresses because the script’s own comments call out that MTP benchmarks require chat-formatted inputs.)

This directly defeats the PR’s stated goal (“both configs failed at launch … this PR restores them”) for the sglang_mtp half. The atom_mtp half is genuinely restored — dsv4_fp4_mi355x_atom_mtp.sh does not use --chat-template, so it is unaffected.

Why nothing catches this upstream

The shell expands nonexistent paths without complaint, and SGLang’s --chat-template argument is a caller-supplied path — no upstream guard validates it. This has to be fixed at the call site.

Fix

In the same commit as the source fix, change benchmarks/single_node/fixed_seq_len/dsv4_fp4_mi355x_sglang_mtp.sh:205 from:

--chat-template "$(dirname "$0")/chat_templates/deepseek_v4_thinking.jinja" \

to:

--chat-template "$(dirname "$0")/../chat_templates/deepseek_v4_thinking.jinja" \

One character added, and the sglang_mtp config the PR is chartered to restore actually reaches the benchmark instead of dying at chat-template load on the very next sweep.