[Klaud Cold] Enact the September 8, 2026 DeepSeek-V4-Pro Single-turn 8k1k deprecation / 执行 2026 年 9 月 8 日 DeepSeek-V4-Pro 单轮 8k1k 场景下线

[functionstackx]

Enacts the Tuesday, September 8, 2026 deprecation recorded in MODELS.md (#2728). That entry set 2026-09-08 as the last day for the Single-turn 8k1k scenario on DeepSeek-V4-Pro 1.6T (dsv4). Today is 2026-09-09, so this carries it out. Agentic coding is unaffected and stays active for dsv4, including its MTP and DSpark arms. The model is not retired: agentic coding becomes its only scenario.

Follows the #2493 / #2527 precedent: archive the config keys, move the scripts, update both MODELS docs, leave perf-changelog.yaml alone.

Config keys removed (46)

Archived under configs/deprecated/, removed from the active masters so sweep generation no longer selects them. Archived bodies are byte-identical to main (a patience diff of each master shows zero inserted lines; every comment line is conserved).

nvidia-dsv4-8k1k-master.yaml (33) — single-node: dsv4-fp4-b200-sglang, -fp4-b200-vllm, -fp4-b200-vllm-mtp, -fp4-b200-trt, -fp4-b200-trt-mtp, -fp4-b300-sglang, -fp4-b300-sglang-mtp, -fp4-b300-vllm, -fp4-b300-vllm-mtp, -fp4-b300-trt, -fp4-b300-trt-mtp, -fp8-h200-vllm, -fp8-h200-vllm-mtp, -fp8-h200-sglang, -fp8-h200-sglang-mtp; multi-node: -fp4-b200-dynamo-vllm, -fp4-b200-dynamo-vllm-mtp, -fp4-b200-dynamo-sglang, -fp4-b200-dynamo-sglang-mtp, -fp4-b300-dynamo-vllm, -fp4-b300-dynamo-sglang, -fp4-b300-dynamo-sglang-mtp, -fp4-gb200-dynamo-vllm, -fp4-gb200-dynamo-vllm-mtp2, -fp4-gb200-dynamo-vllm-mtp2-nosynthetic, -fp4-gb200-dynamo-sglang, -fp4-gb200-dynamo-sglang-mtp, -fp4-gb200-llmd-vllm, -fp4-gb300-dynamo-vllm, -fp4-gb300-dynamo-sglang, -fp4-gb300-dynamo-sglang-mtp, -fp4-gb300-dynamo-trt, -fp4-gb300-dynamo-trt-mtp

amd-dsv4-8k1k-master.yaml (13) — dsv4-fp8-mi300x-vllm, -fp8-mi300x-vllm-mtp, -fp8-mi325x-vllm, -fp8-mi325x-vllm-mtp, -fp4-mi355x-vllm, -fp4-mi355x-vllm-mtp, -fp4-mi355x-sglang, -fp4-mi355x-sglang-mtp, -fp4-mi355x-atom, -fp4-mi355x-atom-mtp, -fp4-mi355x-sglang-disagg, -fp4-mi355x-sglang-disagg-mtp, -fp4-mi355x-atom-disagg

nvidia-master.yaml 156 → 123 keys, amd-master.yaml 56 → 43. The notice said 45; it is 46 because dsv4-fp4-b200-dynamo-sglang (#2560) landed after the notice was written. All 19 dsv4 agentic-coding keys are untouched.

Scripts moved (28)

25 from benchmarks/single_node/fixed_seq_len/ and 3 from benchmarks/multi_node/ (dsv4_fp4_gb200_llmd-vllm-disagg.sh, dsv4_fp4_mi355x_sglang-disagg.sh, dsv4_fp4_mi355x_atom-disagg.sh) into their sibling deprecated/ directories, as git mv renames. The agentic multi_node/agentic/dsv4_fp4_mi355x_sglang-disagg.sh and every single_node/agentic/dsv4_* script stay. The comment in single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh that cites the fixed_seq_len recipe now points at its deprecated/ path, as #2493 did for amd_utils/models_vllm.yaml.

Test removed

utils/test_gb300_power_official_contract.py::test_gb300_dsv4_recipe_images_match_their_master_configs and its two private helpers. It asserted that the runner: gb300 / dynamo-sglang / dsv4 keys pin their srt-slurm recipe model.container to the master image, and required at least one such key to exist. Every key it matched is archived here. The remaining GB300 Dynamo SGLang dsv4 keys are the agentic agg/disagg pair, whose recipes carry the dynamo-sglang placeholder container rather than an image, so the test cannot be retargeted without changing what it checks. The three launcher-routing tests in that file are unchanged.

Deliberately not changed

• benchmarks/single_node/speedbench/dsv4* — speedbench is driven by speedbench-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, [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 precedent).
• benchmarks/multi_node/srt-slurm-recipes/{vllm,sglang}/deepseek-v4/8k1k/, trtllm/…/ISL8K_OSL1K/, llm-d-recipes/dsv4-* — inert reference data once no key references them; [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 and [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 left the MiniMax-M3 and Kimi-K2.5 recipe YAMLs in place the same way.
• perf-changelog.yaml — removals do not trigger benchmarks ([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, [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).
• configs/ci-priority.yaml, utils/evals/thresholds.yaml, runners/ — keyed by model prefix; dsv4 still runs.
• The historical September 8 notice text and its "45 keys" rationale paragraph stay as written; the enactment paragraph below it explains the 46.

Validation

• pytest utils/matrix_logic/ utils/test_gb300_power_official_contract.py utils/test_ci_priority.py utils/test_process_changelog.py → 309 passed
• generate_sweep_configs.py full-sweep over both masters: 1691 → 1308 generated configs. The 383 dropped are all dsv4 fixed-sequence-length rows; 0 added; dsv4 agentic rows unchanged at 141. (Both main and this branch were run with the keys pinned to the retired cluster:b300-nv runner filtered out, because that pre-existing gap in runners.yaml makes the unfiltered full-sweep fail identically on main.)
• Key-conservation check against origin/main: every removed key lands in exactly one archive, no key lost, no non-dsv4 key and no dsv4 agentic key removed; archived bodies parse identical to main; no duplicate keys in any touched YAML.

---

中文说明

执行 MODELS.md（#2728）中 2026 年 9 月 8 日（星期二） 的弃用计划：该日为 DeepSeek-V4-Pro 1.6T（dsv4）单轮 8k1k 场景的最后运行日。今日为 2026-09-09，故予以执行。智能体编码不受影响，dsv4 的该场景继续启用，含 MTP 与 DSpark 分支；该模型不退役，智能体编码成为其唯一场景。完全遵循 #2493 / #2527 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 46 个配置项并归档至 configs/deprecated/：nvidia-dsv4-8k1k-master.yaml（33 个）与 amd-dsv4-8k1k-master.yaml（13 个），归档内容与 main 逐字节一致。nvidia-master.yaml 由 156 项减至 123 项，amd-master.yaml 由 56 项减至 43 项。公告所述为 45 个，实际为 46 个，是因为 dsv4-fp4-b200-dynamo-sglang（#2560）在公告发布后合入。19 个 dsv4 智能体编码配置项全部未做改动。

移动 28 个脚本：单节点 fixed_seq_len/ 25 个、多节点 3 个，均以 git mv 移入同级 deprecated/ 目录；所有智能体编码脚本原地保留。single_node/agentic/dsv4_fp4_mi355x_vllm_mtp.sh 中引用 fixed_seq_len 配方的注释已改为其 deprecated/ 路径。

移除一个测试：test_gb300_dsv4_recipe_images_match_their_master_configs 及其两个私有辅助函数。该测试断言 runner: gb300 / dynamo-sglang / dsv4 配置项的 srt-slurm 配方容器与主配置镜像一致，且要求至少存在一个此类配置项；其匹配的全部配置项已在此归档，余下的 GB300 Dynamo SGLang dsv4 智能体配置项的配方使用 dynamo-sglang 占位容器而非镜像，无法在不改变检查内容的前提下重新定向。该文件中三个启动器路由测试未改动。

刻意保留：dsv4 SPEED-Bench 脚本（由 speedbench-al.yml 驱动）；已归档多节点配置项引用的 srt-slurm 与 llm-d 配方 YAML（惰性参考数据，与 #2493/#2527 一致）；perf-changelog.yaml（移除操作不触发基准测试）；ci-priority.yaml、thresholds.yaml、runners/（按模型前缀索引，dsv4 仍在运行）；9 月 8 日公告原文及其「45 个」的说明段落保留，其下新增的执行段落解释了 46 的来由。

验证：pytest 309 项全部通过；两个主配置的全量扫描生成由 1691 个配置减至 1308 个，减少的 383 个全部为 dsv4 固定序列长度配置，无新增，dsv4 智能体配置保持 141 个不变（main 与本分支均在过滤掉固定到已退役 cluster:b300-nv 运行器的配置项后运行，因该既有问题会使未过滤的全量扫描在 main 上同样失败）；对照 origin/main 的配置项守恒检查确认无遗失、未误删任何非 dsv4 配置项或 dsv4 智能体配置项；归档条目内容与 main 完全一致；所有改动的 YAML 均无重复键。

🤖 Generated with Claude Code

---

Note

Low Risk
Operational/config retirement only: shrinks sweep surface for dsv4 fixed-seq-len while agentic publishing paths stay active; no auth or runtime logic changes in the diff shown.

Overview
Enacts the planned DeepSeek-V4-Pro (dsv4) Single-turn 8k1k deprecation (effective after 2026-09-08): dsv4 keeps only agentic coding in active sweeps; all fixed-seq-len 8k1k work for this model is retired from the masters.

46 fixed-seq-len config keys are removed from nvidia-master.yaml and amd-master.yaml and archived as configs/deprecated/nvidia-dsv4-8k1k-master.yaml (33) and amd-dsv4-8k1k-master.yaml (13), so sweep generation no longer schedules those runs. 28 matching benchmark scripts move under benchmarks/.../deprecated/ (25 single-node fixed_seq_len, 3 multi-node). 19 agentic dsv4 keys, SPEED-Bench dsv4* scripts, and referenced srt-slurm / llm-d recipe YAMLs stay in place.

MODELS.md and MODELS_zh.md mark the notice as enacted (#2921), explain the 46-vs-45 key count, and update the support matrix and draft-model table (MTP no longer tied to 8k1k for dsv4). One agentic script comment now points at the deprecated fixed-seq-len recipe path.

Per the PR description, test_gb300_dsv4_recipe_images_match_their_master_configs is dropped because every GB300 Dynamo SGLang 8k1k key it targeted was archived.

^{Reviewed by Cursor Bugbot for commit 0b3e4d4. Bugbot is set up for automated code reviews on this repo. Configure here.}

[claude]

🔴 Both MODELS.md and MODELS_zh.md merge with the literal placeholder '#PRNUM' / 'pull/PRNUM' instead of the real PR number, so the new deprecation-notice links (4 occurrences total) point to a nonexistent https://github.com/SemiAnalysisAI/InferenceX/pull/PRNUM and render as broken markdown links in the published docs. Fix: replace all four '#PRNUM' / 'pull/PRNUM' occurrences (MODELS.md lines 55 and 157, MODELS_zh.md lines 55 and 157) with the actual merged PR number before/at merge time.

Extended reasoning...

MODELS.md line 55 and 157, and MODELS_zh.md line 55 and 157, all introduce the text '#PRNUM' as the citation for this enactment PR, following the exact pattern used elsewhere in the same files for real PRs (e.g. #2493, #2527, #2728). Every other citation in these docs is a resolved number; this one was left as a template placeholder. Once merged, any reader clicking the link gets a GitHub 404, and the historical record (used elsewhere in the file to reconstruct the deprecation timeline) is broken for this entry specifically.

Verification: nit. The candidate is factually correct: this PR introduces the literal template placeholder PRNUM in four spots, all absent on the base branch. Confirmed via git diff and grep: - MODELS.md:55 — added line: **Enacted on 2026-09-09** in [#PRNUM](https://github.com/SemiAnalysisAI/InferenceX/pull/PRNUM): ... - MODELS.md:157 — `... Single-turn 8k1k (removed 2026-09-09,… | nit. The diff…

[claude]

🔴 Same unfilled #PRNUM placeholder as in MODELS.md is duplicated in the Chinese doc (enactment note and dsv4 history table row), so both language versions ship a broken PR link after merge. Fix: substitute the actual PR number in all four occurrences across MODELS.md and MODELS_zh.md.

Extended reasoning...

MODELS_zh.md's enacted-notice sentence and the dsv4 row in the deprecation-history table both cite [#PRNUM](https://github.com/SemiAnalysisAI/InferenceX/pull/PRNUM), which was never replaced with this PR's real number, leaving a permanently dead link in the merged documentation.

Verification: nit. The diff introduces literal #PRNUM placeholders that were never replaced with this PR's real number, producing dead links in the merged docs. In MODELS_zh.md the enactment note (line 55) reads **已于 2026-09-09 执行**（[#PRNUM](https://github.com/SemiAnalysisAI/InferenceX/pull/PRNUM)） and the dsv4 deprecation-history row reads `单轮 8k1k（已于 2026-09-09… | nit. The candidate is real. The diff for…

[claude]

Additional findings (outside the current diff — GitHub can't attach inline comments there):

• 🟡 docs/configuration-procedures.md — This PR moves benchmarks/multi_node/dsv4_fp4_gb200_llmd-vllm-disagg.sh to benchmarks/multi_node/deprecated/dsv4_fp4_gb200_llmd-vllm-disagg.sh but doesn't update the 'Register an llm-d recipe' doc, whose only concrete wrapper example link still points at the old (now nonexistent) path, breaking the doc for anyone following it after this merge. Fix: update the link (and its MODELS_zh.md-analog docs/configuration-procedures_zh.md:131) to the new deprecated/ path.

  Extended reasoning...

  docs/configuration-procedures.md:131 (and the identical Chinese docs/configuration-procedures_zh.md:131) cite [llmd-vllm benchmark wrapper](../benchmarks/multi_node/dsv4_fp4_gb200_llmd-vllm-disagg.sh) as the current example script for registering a new llm-d recipe. This diff's own git-mv moves that file into benchmarks/multi_node/deprecated/, so the relative link now 404s. A contributor following the doc to see a live example clicks through to a dead file, unlike on main where the link resolved.

  Verification: nit. The failure is real: docs/configuration-procedures.md:131 (and the identical zh line) reads and the current [\llmd-vllm` benchmark wrapper](../benchmarks/multi_node/dsv4_fp4_gb200_llmd-vllm-disagg.sh), and this PR git-mv's that file. git diff --statshows the renamebenchmarks/multi_node/{ => deprecated}/dsv4_fp4_gb200_llmd-vllm-disagg.sh; ls` confirms the old path no longer…

• 🟡 configs/nvidia-master.yaml:1056 — Deleting dsv4-fp4-b300-sglang-mtp left its 10-line preceding documentation comment (describing benchmarks/single_node/dsv4_fp4_b300_sglang_mtp.sh's DP_ATTENTION recipe selection) orphaned in the active master, now dangling under dsv4-fp4-b300-sglang-agentic-hicache-mtp right before the unrelated qwen3.5-fp8-b200-sglang key, and the comment was never copied into configs/deprecated/nvidia-dsv4-8k1k-master.yaml. This contradicts the PR's own claim that every comment line is conserved, and leaves stale/misattributed documentation pointing at a script path that no longer exists at that location. Fix: when archiving a key, also remove/relocate any 2-space-indented comment block that immediately precedes it (even without a separating blank line), and add it to the archived body.

  Extended reasoning...

  In the base branch, this same comment (2-space indented, matching the previous key's body indent, with no blank line separating it from the key line) sat directly above dsv4-fp4-b300-sglang-mtp:. The removal in this PR deleted only the key stanza starting at dsv4-fp4-b300-sglang-mtp:, not the comment above it, because the comment's indentation made it look like it belonged to the prior surviving key (dsv4-fp4-b300-sglang-agentic-hicache-mtp). Result: nvidia-master.yaml:1047-1056 now contains a comment about a moved/archived script sitting before qwen3.5-fp8-b200-sglang, and configs/deprecated/nvidia-dsv4-8k1k-master.yaml's dsv4-fp4-b300-sglang-mtp entry (line ~132) has no preceding documentation at all — verified by grep showing the comment text exists only in nvidia-master.yaml, not in the archive.

  Verification: nit. The defect is real but purely cosmetic (an inert YAML comment), so it does not break parsing, sweep generation, or any behavior — it makes documentation stale, not the base branch functionally worse. Facts verified: - In the base branch (git show 4caeb31:configs/nvidia-master.yaml), a 2-space-indented comment block "# DeepSeek-V4-Pro on B300 with EAGLE/MTP speculative decoding. Recipe is /…