[Klaud Cold] Enact the August 6, 2026 Kimi-K2.5/2.6/2.7-Code retirement / 执行 2026 年 8 月 6 日 Kimi-K2.5/2.6/2.7-Code 完全退役 - #2527
Conversation
MODELS.md set 2026-08-06 as the last day for the Single-turn 8k1k scenario on kimik2.5, and — combined with the agentic-coding deprecation enacted on 2026-08-04 in #2493 — declared the model fully retired after that date. Today is 2026-08-07, so this carries it out. Removes 17 kimik2.5 config keys from the active master configs and archives them under configs/deprecated/: - nvidia-kimik2.5-8k1k-master.yaml (10): int4 b200/b300/h200, fp4 b200/b300, fp4 b200/gb200/gb300 dynamo-trt, fp4 gb200/gb300 dynamo-vllm - amd-kimik2.5-8k1k-master.yaml (7): int4 mi300x/mi325x/mi355x, fp4 mi355x vllm/atom, fp4 mi355x atom-disagg/vllm-disagg Their 12 benchmark scripts move to the sibling deprecated/ directories (10 under single_node/fixed_seq_len/, 2 under multi_node/). kimik2.5 now has no active configuration in any master config. Also archives kimik2.5-int4-h100-vllm, an agentic-coding key #2493 left behind in nvidia-master.yaml after moving its script to single_node/agentic/deprecated/ -- the config had been pointing at a path that no longer existed. It joins its siblings in nvidia-kimik2.5-agentic-master.yaml. benchmarks/single_node/speedbench/kimik2.5_fp4_b300_vllm.sh is deliberately kept: speedbench is driven by speedbench-al.yml rather than the master configs, which is how #2493 treated MiniMax-M3. perf-changelog.yaml is untouched -- removals do not trigger benchmarks, also per #2493. 中文：执行 MODELS.md 中 2026 年 8 月 6 日的 Kimi-K2.5/2.6/2.7-Code 退役计划。单轮 8k1k 场景的最后运行日为 2026-08-06，叠加 #2493 已于 2026-08-04 执行的智能体编码 弃用，该模型此后完全退役；今日为 2026-08-07，故予以执行。 从启用的主配置中移除 17 个 kimik2.5 配置项并归档至 configs/deprecated/（NVIDIA 10 个、AMD 7 个），对应的 12 个基准测试脚本移入同级 deprecated/ 目录。此后 kimik2.5 在所有主配置中均无启用配置。 同时归档 kimik2.5-int4-h100-vllm：#2493 将其脚本移入 agentic/deprecated/ 时，该 配置项被遗留在 nvidia-master.yaml 中，导致配置指向已不存在的路径。 SPEED-Bench 脚本予以保留（由 speedbench-al.yml 驱动，不经过主配置，与 #2493 处理 MiniMax-M3 的方式一致）；perf-changelog.yaml 未改动（移除操作不触发基准测试）。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
中文：在 MODELS.md 与 MODELS_zh.md 中填入 #2527 的 PR 链接。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
1 similar comment
|
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 For PR verification, add the 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 如需进行 PR 验证，请为此 PR 添加 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 |
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| # Removed from the active master config so sweep generation no longer selects them. | ||
|
|
||
| kimik2.5-int4-b200-vllm: | ||
| image: vllm/vllm-openai:v0.25.0 | ||
| model: moonshotai/Kimi-K2.5 | ||
| model-prefix: kimik2.5 | ||
| runner: b200 | ||
| precision: int4 | ||
| framework: vllm | ||
| multinode: false | ||
| scenarios: | ||
| fixed-seq-len: | ||
| - isl: 8192 | ||
| osl: 1024 | ||
| search-space: |
There was a problem hiding this comment.
🟡 The 10 kimik2.5 fixed_seq_len scripts moved into benchmarks/single_node/fixed_seq_len/deprecated/ (e.g. kimik2.5_int4_b200.sh, kimik2.5_fp4_b200.sh, etc.) still source $(dirname "$0")/../../benchmark_lib.sh, which is one ../ short from the new location and resolves to a nonexistent path. Fix by changing to ../../../benchmark_lib.sh, matching every other already-archived script in the same directory (e.g. glm5_fp4_b200.sh).
Extended reasoning...
When this PR archived the 10 kimik2.5 fixed_seq_len scripts via git mv into a new deprecated/ subdirectory (benchmarks/single_node/fixed_seq_len/deprecated/), it moved each script one directory level deeper than its old location. Since git mv is a pure rename, the source "$(dirname "$0")/../../benchmark_lib.sh" line at the top of every one of these scripts was left untouched.
That relative path was correct at the old location (fixed_seq_len/ → single_node/ → benchmarks/ = 2 levels up), but from the new deprecated/ subdirectory it now resolves to benchmarks/single_node/benchmark_lib.sh — a file that does not exist. The real file lives at benchmarks/benchmark_lib.sh, three levels up from deprecated/.
Proof:
$ ls benchmarks/benchmark_lib.sh
benchmarks/benchmark_lib.sh # exists
$ ls benchmarks/single_node/benchmark_lib.sh
ls: cannot access 'benchmarks/single_node/benchmark_lib.sh': No such file or directory
$ grep benchmark_lib.sh benchmarks/single_node/fixed_seq_len/deprecated/kimik2.5_int4_b200.sh
source "$(dirname "$0")/../../benchmark_lib.sh" # -> benchmarks/single_node/benchmark_lib.sh (broken)
$ grep benchmark_lib.sh benchmarks/single_node/fixed_seq_len/deprecated/glm5_fp4_b200.sh
source "$(dirname "$0")/../../../benchmark_lib.sh" # -> benchmarks/benchmark_lib.sh (correct)
glm5_fp4_b200.sh is a script that was already archived into the same deprecated/ directory in an earlier PR, and it correctly uses three ../s — confirming both that this is the established convention for scripts at this depth and that the convention was simply missed for these 10 scripts in this PR.
Nothing in CI or sweep generation catches this: these scripts are deprecated and no longer selected by generate_sweep_configs.py, so the broken source line is never exercised automatically. The only way it manifests is a human manually re-running or sourcing one of these 10 scripts for reference/debugging, at which point it fails immediately with a "No such file or directory" error on the source line before anything else in the script executes.
The fix is mechanical and consistent with the rest of the directory: change ../../benchmark_lib.sh to ../../../benchmark_lib.sh in all 10 moved scripts (kimik2.5_fp4_b200.sh, kimik2.5_fp4_b300.sh, kimik2.5_fp4_mi355x.sh, kimik2.5_fp4_mi355x_atom.sh, kimik2.5_int4_b200.sh, kimik2.5_int4_b300.sh, kimik2.5_int4_h200.sh, kimik2.5_int4_mi300x.sh, kimik2.5_int4_mi325x.sh, kimik2.5_int4_mi355x.sh).
Severity-wise, this is a real, concrete breakage with a guaranteed failure mode and an unambiguous fix pattern already established in the same directory, but the blast radius is limited: these scripts are archived/deprecated and not part of any active sweep or CI path, so nothing breaks for existing automated workflows if this ships unfixed. It only affects a manual reference/rerun scenario. Given the rubric's guidance to reserve "normal" for things that break automated behavior or user-facing correctness, this is best treated as a nit — worth fixing before merge since the fix is trivial and the pattern is already established, but not blocking.
| # does not have a B300-specific recipe, so this config reuses the existing | ||
| # Kimi-K2.5 FP4 B200 vLLM recipe as-is until B300-specific tuning is available. | ||
|
|
||
| kimik2.5-fp4-b200-vllm: | ||
| image: vllm/vllm-openai:nightly-e2fa28594f7baad142a426b0b6a2cfe2c79201c7 | ||
| model: nvidia/Kimi-K2.5-NVFP4 | ||
| model-prefix: kimik2.5 | ||
| runner: b200 | ||
| precision: fp4 | ||
| framework: vllm | ||
| multinode: false | ||
| scenarios: | ||
| fixed-seq-len: | ||
| - isl: 8192 | ||
| osl: 1024 | ||
| search-space: | ||
| - { tp: 8, ep: 1, conc-start: 1, conc-end: 512 } | ||
| - { tp: 4, ep: 1, conc-start: 1, conc-end: 512 } | ||
| - { tp: 8, ep: 8, dp-attn: false, conc-start: 1, conc-end: 512 } | ||
| - { tp: 4, ep: 4, dp-attn: false, conc-start: 1, conc-end: 512 } | ||
| - { tp: 8, ep: 8, dp-attn: true, conc-list: [512, 1024] } | ||
|
|
||
| # NOTE: At the time of submission, https://docs.vllm.ai/projects/recipes/en/latest/moonshotai/Kimi-K2.5.html | ||
| # does not have a B300-specific recipe, so this config reuses the existing | ||
| # Kimi-K2.5 FP4 B200 vLLM recipe as-is until B300-specific tuning is available. | ||
| kimik2.5-fp4-b300-vllm: | ||
| image: vllm/vllm-openai:nightly-e2fa28594f7baad142a426b0b6a2cfe2c79201c7 | ||
| model: nvidia/Kimi-K2.5-NVFP4 | ||
| model-prefix: kimik2.5 | ||
| runner: b300 | ||
| precision: fp4 | ||
| framework: vllm | ||
| multinode: false | ||
| scenarios: | ||
| fixed-seq-len: | ||
| - isl: 8192 | ||
| osl: 1024 | ||
| search-space: | ||
| - { tp: 8, ep: 1, conc-list: [1] } | ||
| - { tp: 4, ep: 1, conc-start: 1, conc-end: 512 } | ||
| - { tp: 8, ep: 8, dp-attn: false, conc-list: [1] } | ||
| - { tp: 4, ep: 4, dp-attn: false, conc-start: 1, conc-end: 512 } | ||
| - { tp: 4, ep: 4, dp-attn: true, conc-start: 128, conc-end: 512 } | ||
| dsr1-fp8-b200-sglang-mtp: | ||
| image: lmsysorg/sglang:v0.5.12-cu130 | ||
| model: deepseek-ai/DeepSeek-R1-0528 |
There was a problem hiding this comment.
🟡 This 3-line NOTE comment about reusing the Kimi-K2.5 FP4 B200 vLLM recipe was left behind above the unrelated dsr1-fp8-b200-sglang-mtp entry after both kimik2.5-fp4-b200-vllm and kimik2.5-fp4-b300-vllm were removed in this PR. It's a stale editing artifact (the identical duplicate comment above the b300 entry was correctly deleted) that now falsely implies the DeepSeek-R1 entry reuses a Kimi-K2.5 recipe — delete lines 1459-1462 in configs/nvidia-master.yaml.
Extended reasoning...
configs/nvidia-master.yaml previously carried two identical copies of a 3-line # NOTE: comment explaining that kimik2.5-fp4-b300-vllm reuses the kimik2.5-fp4-b200-vllm recipe as-is, pending B300-specific tuning: one copy directly above kimik2.5-fp4-b200-vllm, and one directly above kimik2.5-fp4-b300-vllm.
This PR correctly removes both kimik2.5-fp4-b200-vllm and kimik2.5-fp4-b300-vllm (archiving them, comment included, into configs/deprecated/nvidia-kimik2.5-8k1k-master.yaml lines 72-76). The diff shows the second copy of the comment — the one immediately preceding kimik2.5-fp4-b300-vllm — was removed as part of that hunk. However, the first copy, which sat above kimik2.5-fp4-b200-vllm, was an unchanged context line and was never deleted.
With both Kimi-K2.5 entries gone, that orphaned comment now sits at lines 1459-1461 of configs/nvidia-master.yaml, directly above the unrelated dsr1-fp8-b200-sglang-mtp: entry (line 1463), separated only by a blank line. Read in context, it appears to explain why the DeepSeek-R1 SGLang MTP entry reuses a Kimi-K2.5 FP4 B200 vLLM recipe — which is false. The dsr1-fp8-b200-sglang-mtp/dsr1-fp8-b300-sglang-mtp pair already has its own correct NOTE comment a few lines further down (1478-1480), explaining the actual DSR1 B300→B200 SGLang reuse, so the misplaced Kimi-K2.5 comment is pure noise with no correct referent nearby.
Step-by-step proof:
- In the parent commit (
8c241a3^),nvidia-master.yamlline ~1505 has NOTE-copy-1 abovekimik2.5-fp4-b200-vllm, and line ~1528 has NOTE-copy-2 abovekimik2.5-fp4-b300-vllm. - This PR's diff removes
kimik2.5-fp4-b200-vllm(and NOTE-copy-1's neighbor context is not removed — the comment lines themselves are left as unchanged context), and separately removeskimik2.5-fp4-b300-vllmalong with NOTE-copy-2 (that hunk's-lines include the comment text). - Reading current HEAD at
configs/nvidia-master.yaml:1459-1465confirms NOTE-copy-1 is still present, immediately followed by a blank line and thendsr1-fp8-b200-sglang-mtp:. - There is no longer any Kimi-K2.5 entry anywhere near this comment for it to legitimately describe.
Since YAML comments are inert to generate_sweep_configs.py and the sweep/CI pipeline, this has zero effect on functionality, config generation, or the 1912-config full-sweep validation the PR cites. The only harm is that a future reader (human or LLM) skimming this file could be misled into thinking the DeepSeek-R1 entry has some dependency on a retired Kimi-K2.5 recipe. Fix is a trivial one-line-block deletion: remove lines 1459-1462 (the 3 comment lines plus the now-redundant blank line) in configs/nvidia-master.yaml.
|
|
||
| **Thursday, August 6, 2026** is the last day for the **Single-turn 8k1k** scenario on **Kimi-K2.5/2.6/2.7-Code** (`kimik2.5`); the scenario is deprecated for these models after that date. Rationale: Kimi-K3 launched on July 27, 2026, so GPU cluster time shifts to the newer frontier model. Combined with the Agentic coding deprecation above, this leaves `kimik2.5` with no active scenario — the model is **fully retired after August 6, 2026**. | ||
|
|
||
| **Enacted on 2026-08-07** in [#2527](https://github.com/SemiAnalysisAI/InferenceX/pull/2527): 17 `kimik2.5` config keys were removed from the active master configs and archived under [`configs/deprecated/`](configs/deprecated/) as `nvidia-kimik2.5-8k1k-master.yaml` (10) and `amd-kimik2.5-8k1k-master.yaml` (7), and their 12 benchmark scripts were moved to the sibling `deprecated/` directories. `kimik2.5` now has **no active configuration in any master config** and is fully retired. The same PR archived `kimik2.5-int4-h100-vllm`, an agentic-coding key that #2493 left behind in `nvidia-master.yaml` after moving its script to `benchmarks/single_node/agentic/deprecated/`; it is now in `nvidia-kimik2.5-agentic-master.yaml` with its siblings. The SPEED-Bench acceptance-length script `benchmarks/single_node/speedbench/kimik2.5_fp4_b300_vllm.sh` is intentionally kept — speedbench is driven by `speedbench-al.yml`, not the master configs, matching how #2493 treated MiniMax-M3. |
There was a problem hiding this comment.
🟡 The self-referential PR links in the Deprecation Notice enactment note and the model-support-matrix row use the literal unfilled placeholder #XXXX (linking to the nonexistent /pull/XXXX) instead of #2527. This occurs 4 times total: MODELS.md lines 43 and 147, and the equivalent lines in MODELS_zh.md — all four should read #2527 / /pull/2527.
Extended reasoning...
What the bug is
The PR adds two new self-referential links to this very PR (#2527) — one in the "Enacted on 2026-08-07" paragraph under the Deprecation Notice, and one in the kimik2.5 row of the Model support matrix. In both MODELS.md and MODELS_zh.md, the author left the template placeholder #XXXX unfilled instead of substituting the real PR number. The committed text reads:
**Enacted on 2026-08-07** in [#XXXX](https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX): ...
and in the matrix row:
| Kimi-K2.5/2.6/2.7-Code | `kimik2.5` | ... | — (fully retired 2026-08-07, [#XXXX](https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX)) | ... Single-turn 8k1k (removed 2026-08-07, [#XXXX](https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX)) |
The same two instances are duplicated in MODELS_zh.md (Chinese doc), for 4 broken links total.
Why this is real and not a display artifact
The PR description text (the human-written summary) correctly uses #2527 throughout — that is what makes this easy to miss on a first read. But the actual file content added by this commit (verified directly against the repo, not the description) contains the literal string XXXX. git show on this merge commit confirms these lines were added with XXXX already in them — this is not a pre-existing issue inherited from a prior PR, it is a template placeholder introduced and never filled in by this PR.
Impact
Each [#XXXX](.../pull/XXXX) renders as a clickable markdown link that resolves to a nonexistent PR (/pull/XXXX is not a valid GitHub PR number). This defeats the entire purpose of the enactment note, which exists specifically to give readers a traceable link from the deprecation record to the PR that carried it out — the same pattern the #2493 precedent (referenced elsewhere in this same PR) implements correctly with a real PR number.
Proof / how to see it
- Open
MODELS.mdat line 43 in the merged file (not the PR description prose). - Read:
**Enacted on 2026-08-07** in [#XXXX](https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX): ... - Click the link (or navigate to
https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX) — it 404s, since XXXX is not a PR number. - Compare to the intended target,
https://github.com/SemiAnalysisAI/InferenceX/pull/2527, which is this very PR and works correctly. - Repeat for line 147 (matrix row, 2 occurrences) and the equivalent lines 43/147 in
MODELS_zh.md.
Fix
Replace all 4 occurrences of #XXXX / pull/XXXX with #2527 / pull/2527 in MODELS.md and MODELS_zh.md. This is a trivial find-and-replace with no other code impact — generate_sweep_configs.py, the config YAMLs, and the archived scripts are all unaffected, since this is purely a documentation link.
Severity
Docs-only defect with no effect on sweep generation, config parsing, or benchmark execution — nothing crashes and no data is lost. However it is an objective, unambiguous error (not a style nit or PR-description mismatch) that ships 4 dead links in the canonical model-tracking document. Marking as nit: worth fixing before merge, but not blocking.
1 participant
Enacts the Thursday, August 6, 2026 deprecation recorded in
MODELS.md. That entry set 2026-08-06 as the last day for the Single-turn 8k1k scenario on Kimi-K2.5/2.6/2.7-Code (kimik2.5) and — combined with the agentic-coding deprecation already enacted on 2026-08-04 in #2493 — declared the model fully retired after that date. Today is 2026-08-07, so this carries it out.Follows the #2493 precedent exactly: archive the config keys, move the scripts, update both
MODELSdocs, and leaveperf-changelog.yamlalone.Config keys removed (17)
Archived under
configs/deprecated/, removed from the active masters so sweep generation no longer selects them.nvidia-kimik2.5-8k1k-master.yaml(10) —kimik2.5-int4-b200-vllm,-int4-b300-vllm,-int4-h200-vllm,-fp4-b200-vllm,-fp4-b300-vllm,-fp4-b200-dynamo-trt,-fp4-gb200-dynamo-trt,-fp4-gb300-dynamo-trt,-fp4-gb200-dynamo-vllm,-fp4-gb300-dynamo-vllmamd-kimik2.5-8k1k-master.yaml(7) —kimik2.5-int4-mi355x-vllm,-int4-mi325x-vllm,-int4-mi300x-vllm,-fp4-mi355x-vllm,-fp4-mi355x-atom,-fp4-mi355x-atom-disagg,-fp4-mi355x-vllm-disaggnvidia-master.yaml134 → 123 keys,amd-master.yaml56 → 49.kimik2.5now has no active configuration in any master config.Scripts moved (12)
10 from
benchmarks/single_node/fixed_seq_len/and 2 frombenchmarks/multi_node/into their siblingdeprecated/directories, asgit mvrenames.Bug fixed along the way
kimik2.5-int4-h100-vllmis an agentic-coding key that #2493 left behind innvidia-master.yamlwhen it moved that recipe's script tobenchmarks/single_node/agentic/deprecated/. The key has since been pointing atbenchmarks/single_node/agentic/kimik2.5_int4_h100.sh, which does not exist — so the config was dangling. It is archived here intonvidia-kimik2.5-agentic-master.yamlalongside its siblings, where #2493 intended it to go.Deliberately not changed
benchmarks/single_node/speedbench/kimik2.5_fp4_b300_vllm.sh— speedbench is driven byspeedbench-al.yml, not the master configs. [Klaud Cold] Enact the August 3, 2026 scenario and precision deprecations / 执行 2026 年 8 月 3 日场景与精度下线 #2493 left MiniMax-M3's speedbench script in place for the same reason.perf-changelog.yaml— removals don't trigger benchmarks, and [Klaud Cold] Enact the August 3, 2026 scenario and precision deprecations / 执行 2026 年 8 月 3 日场景与精度下线 #2493 touched 58 files without adding a changelog entry.configs/ci-priority.yaml,utils/evals/thresholds.yaml,runners/,srt-slurm-recipes/— these are keyed by model prefix or are inert reference data; [Klaud Cold] Enact the August 3, 2026 scenario and precision deprecations / 执行 2026 年 8 月 3 日场景与精度下线 #2493 touched none of them.Validation
pytest utils/matrix_logic/→ 224 passedgenerate_sweep_configs.py full-sweepover both masters → 1912 configs, 0 withmodel-prefix: kimik2.5origin/main: every removed key lands in exactly one archive, no key lost, and no non-kimik2.5key removed (an earlier attempt silently droppeddsr1-fp8-b200-sglang-mtp; caught and redone)mainhad中文说明
执行
MODELS.md中 2026 年 8 月 6 日（星期四） 的弃用计划：该日为 Kimi-K2.5/2.6/2.7-Code（kimik2.5）单轮 8k1k 场景的最后运行日；叠加 #2493 已于 2026-08-04 执行的智能体编码弃用，该模型此后完全退役。今日为 2026-08-07，故予以执行。完全遵循 #2493 的先例：归档配置项、移动脚本、同步更新中英文MODELS文档、不改动perf-changelog.yaml。移除 17 个配置项并归档至
configs/deprecated/：nvidia-kimik2.5-8k1k-master.yaml（10 个）与amd-kimik2.5-8k1k-master.yaml（7 个）。nvidia-master.yaml由 134 项减至 123 项，amd-master.yaml由 56 项减至 49 项。此后kimik2.5在所有主配置中均无启用配置。移动 12 个脚本：单节点
fixed_seq_len/10 个、多节点 2 个，均以git mv移入同级deprecated/目录。顺带修复一处缺陷：
kimik2.5-int4-h100-vllm是一个智能体编码配置项，#2493 将其脚本移入agentic/deprecated/时，该配置项被遗留在nvidia-master.yaml中，导致其指向的benchmarks/single_node/agentic/kimik2.5_int4_h100.sh已不存在。现将其归入nvidia-kimik2.5-agentic-master.yaml。刻意保留：SPEED-Bench 脚本（由
speedbench-al.yml驱动，不经过主配置，与 #2493 处理 MiniMax-M3 一致）；perf-changelog.yaml（移除操作不触发基准测试）；ci-priority.yaml、thresholds.yaml、runners/、srt-slurm-recipes/（按模型前缀索引或为惰性参考数据，#2493 亦未改动）。验证：
pytest utils/matrix_logic/224 项全部通过；两个主配置的全量扫描生成 1912 个配置，其中model-prefix: kimik2.5为 0；对照origin/main的配置项守恒检查确认无遗失、且未误删任何非kimik2.5配置项（首次尝试曾误删dsr1-fp8-b200-sglang-mtp，已发现并重做）；归档条目内容与main完全一致；所有改动的 YAML 均无重复键。