[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
Conversation
|
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 |
Archive the deprecated master-config entries into configs/deprecated/ and move their benchmark scripts into the sibling deprecated/ directories, per the Deprecation Notice in MODELS.md. 54 config keys move out of the active NVIDIA and AMD master configs: minimaxm3-8k1k 39 Single-turn 8k1k (agentic coding stays active) kimik2.5-agentic 6 agentic coding (8k1k stays until August 6, 2026) qwen3.5-bf16 9 all bf16 recipes, every scenario The speculative-decoding A/B retirements are NOT enacted here. Removing the non-spec-decode agentic arm would leave MiniMax-M3 and GLM-5.2 with no active config at all, since their replacement EAGLE3/MTP arms have not landed, and would drop all AMD and SGLang agentic coverage for dsv4 and qwen3.5. That round is deferred to a follow-up once the spec-decode arms exist. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Mark the first deprecation table as enacted and update the matrix rows for minimaxm3, kimik2.5 and qwen3.5. The speculative-decoding A/B table gets an explicit "not yet enacted" status note explaining why, and the dsv4 / glm5.2 rows no longer claim MTP-only coverage that is not in force. Mirrored in MODELS_zh.md. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
ecc5f97 to
e607399
Compare
functionstackx
changed the title
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
There was a problem hiding this comment.
Additional findings (outside current diff — PR may have been updated during review):
-
🔴
configs/deprecated/nvidia-agentic-non-spec-master.yaml:92-121— configs/deprecated/nvidia-agentic-non-spec-master.yaml archives four GB300 dynamo-sglang agentic keys (qwen3.5-fp4-gb300-dynamo-sglang-agentic-agg/-disagg, dsv4-fp4-gb300-dynamo-sglang-agentic-agg/-disagg) and removes them from configs/nvidia-master.yaml, but every search-space row in all four uses spec-decoding: mtp — there is no non-MTP sibling for these keys.Extended reasoning...
configs/deprecated/nvidia-agentic-non-spec-master.yamlis meant to hold only the non-speculative-decode arm of the dsv4/qwen3.5/glm5.2 agentic A/B pairs — its own header comment says so explicitly: "Deprecated non-speculative-decoding agentic-coding entries archived from nvidia-master.yaml ... the spec-decode arm is the better Pareto frontier, so only that arm is run and published." That contract is violated by four of the entries this PR moved into the file:qwen3.5-fp4-gb300-dynamo-sglang-agentic-agg— both search-space rows setspec-decoding: "mtp"(lines ~105, 121 of the archive)qwen3.5-fp4-gb300-dynamo-sglang-agentic-disagg— all 7 rows setspec-decoding: "mtp"dsv4-fp4-gb300-dynamo-sglang-agentic-agg— its single row setsspec-decoding: mtpdsv4-fp4-gb300-dynamo-sglang-agentic-disagg— all rows setspec-decoding: mtp
These are the spec-decode (MTP) arm, not the non-spec arm the file is supposed to collect. Contrast with the genuinely non-spec entries correctly archived in the same file (e.g.
dsv4-fp4-gb300-dynamo-vllm-agentic, thedsv4-fp4-gb200-dynamo-vllm-agentic-*entries), which all setspec-decoding: none/omit it.The Deprecation Notice in
MODELS.mdthat this PR itself enacts is explicit: for bothqwen3.5anddsv4, only the non-MTP agentic arm is deprecated; "Agentic coding, MTP" is the arm that "remains"/is published. Grepping the activeconfigs/nvidia-master.yamlafter this PR confirms none of these fourgb300-dynamo-sglang-agentickeys remain active, and there is no like-for-like GB300 dynamo-sglang replacement — the block that stayed active for qwen3.5 GB300-class agentic isqwen3.5-fp4-b300-sglang-agentic-mtp, which is a different recipe (single-node b300, not multi-node gb300-dynamo-sglang), and dsv4 keeps onlygb300-dynamo-vllmagentic MTP, notgb300-dynamo-sglang. So this PR silently drops real, currently-published GB300 dynamo-sglang MTP agentic coverage for both models with no replacement.The likely root cause: the migration script/author probably classified entries into "non-spec" vs "spec" buckets by key-name suffix (looking for a trailing
-mtp) rather than by the actualspec-decodingfield inside the search-space rows. These four keys don't have an-mtpsuffix in their names (unlike e.g.qwen3.5-fp4-b300-sglang-agentic-mtp), even though their contents are MTP recipes — so they fell through the suffix-based filter into the non-spec archive.Proof, step by step:
- Open
configs/deprecated/nvidia-agentic-non-spec-master.yamlin this diff and findqwen3.5-fp4-gb300-dynamo-sglang-agentic-agg:. - Its
search-spacerows read- spec-decoding: "mtp" ... conc-list: [1]and- spec-decoding: "mtp" ... conc-list: [96]— both explicitly MTP, not "none". - Compare to the file's own header: "only the spec-decode arm is run and published" is deprecated-file language for removing the non-spec arm — but these rows ARE the spec-decode arm.
- Check
configs/nvidia-master.yamlpost-PR:grep gb300-dynamo-sglang-agenticreturns nothing — the key is gone from the active config with no replacement of the same platform/engine/multinode shape. - Check
MODELS.md's own updated matrix in this PR: for qwen3.5, "Agentic coding (MTP only from 2026-08-03)" is listed as the active/remains scenario, and for dsv4 the same. Removing the MTP entries directly contradicts what the PR's own documentation says should stay active.
Impact: this is a benchmark-coverage regression, not a crash — no user-facing failure occurs, but real, currently-published GB300 multi-node dynamo-sglang MTP agentic benchmark coverage for two models (qwen3.5 and dsv4) silently disappears from the sweep, with the archive's header and MODELS.md's matrix now factually describing coverage that no longer exists on that platform.
Fix: move these four keys (and their search-space content) back into
configs/nvidia-master.yamlas active configs — they are the arm the Deprecation Notice says to keep — or, if their removal was actually intended for a different reason, updateMODELS.mdand the archive header to correctly describe why the GB300 dynamo-sglang MTP arm no longer runs.
|
|
||
| **Monday, August 3, 2026** is the last day for the scenarios, precisions, and recipe variants listed below; they are deprecated after that date. | ||
|
|
||
| **Enacted on 2026-08-04** in [#2493](https://github.com/SemiAnalysisAI/InferenceX/pull/2493): 80 config keys were removed from the active master configs and archived under [`configs/deprecated/`](configs/deprecated/), and their benchmark scripts moved to the sibling `deprecated/` directories. Kimi-K3 is the one exception — its non-DSpark agentic arm is listed below but was deliberately kept live in that PR, so both arms still run pending a follow-up. | ||
|
|
||
| Scenario and precision retirements: | ||
|
|
||
| | Model | Deprecated | Remains | |
There was a problem hiding this comment.
🟡 The self-referential PR link uses a literal PRNUM placeholder instead of 2493 in both MODELS.md (lines 15, 56, 57, 61) and the mirrored lines in MODELS_zh.md, producing 8 broken /pull/PRNUM links (404) in the docs that are this PR's own deliverable. Fix with a simple find-and-replace of PRNUM → 2493 in both files.
Extended reasoning...
This PR's whole purpose is to record the enactment of the August 3, 2026 deprecation notice and link back to the PR that did it. Four places in MODELS.md (lines 15, 56, 57, 61) and the four mirrored lines in MODELS_zh.md write [#PRNUM](https://github.com/SemiAnalysisAI/InferenceX/pull/PRNUM) where the actual PR number, 2493, should be. This is a leftover template placeholder — the author correctly filled in the real numbers for every other PR reference on those lines (e.g. #2268, #1724, #734), but couldn't have known this PR's own number ahead of time and never went back to substitute it once the PR was opened.
I verified this against the actual committed file content, not the PR diff view: grep -n PRNUM MODELS.md MODELS_zh.md on the current working tree returns all 8 occurrences, and git show HEAD:MODELS.md confirms the same text is present in the commit that shipped this change (5b5606e, "docs(MODELS): record the August 3, 2026 deprecations as enacted"). The diff shown during review substitutes the placeholder with #2493 for display purposes, which is why it can look already-fixed there — but the file actually on disk/in HEAD still has the raw placeholder.
Concretely, anyone who clicks the "Enacted on 2026-08-04 in #PRNUM" link at MODELS.md:15, or either of the "no active config as of 2026-08-04 (#PRNUM)" callouts at lines 56-57, or the "Agentic coding (removed 2026-08-04, #PRNUM)" note at line 61, lands on https://github.com/SemiAnalysisAI/InferenceX/pull/PRNUM, which is not a valid PR number and 404s. Same for the four mirrored Chinese-language lines in MODELS_zh.md. Since the whole point of these notes is to let a future reader jump straight to the PR that performed the archival/removal, a dead link defeats that purpose, even though it doesn't break any code, CI, or benchmark behavior.
Fix is mechanical: replace PRNUM with 2493 in both files (8 substitutions total, matching the #2493 display already shown in the PR's own diff view). Since this is a documentation-only defect with no functional or CI impact, I'm marking it a nit rather than a blocking issue — but it should be fixed before merge since it's trivial and the whole point of the change is to leave working self-referential links.
…-trt / 移除已弃用的 Single-turn 1k1k 场景 Single-turn 1k1k (isl 1024 / osl 1024) is deprecated for all models since 2026-07-17 (#2263) per MODELS.md, and the deprecation was enacted on 2026-08-04 by #2493, which removed 54 config keys from the active master configs. This submission still added a 1k1k scenario, so it is removed here. - Delete the 9 disagg 1k1k STP recipes under benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/1k1k/ - Remove the isl 1024 / osl 1024 scenario (9 search-space points) from the qwen3.5-fp4-gb300-dynamo-trt block in configs/nvidia-master.yaml - Align the perf-changelog entry to describe an 8k1k-only submission The Single-turn 8k1k half is unchanged: 9 recipes and 9 search-space points remain byte-identical. 中文：根据 MODELS.md，Single-turn 1k1k（isl 1024 / osl 1024）自 2026-07-17 起对所有模型弃用（#2263），并已于 2026-08-04 由 #2493 正式执行（从活跃主配置中 移除了 54 个 config key）。本次提交仍新增了 1k1k 场景，故在此移除：删除 gb300-fp4/disagg/1k1k/ 下的 9 个 STP recipe，从 configs/nvidia-master.yaml 的 qwen3.5-fp4-gb300-dynamo-trt 块中删除 isl 1024 / osl 1024 场景及其 9 个搜索空间 配置点，并同步修改 perf-changelog 描述为仅 8k1k。Single-turn 8k1k 部分保持不变， 9 个 recipe 与 9 个搜索空间配置点逐字节保留。
MODELS.md deprecates Single-turn 1k1k for all models (2026-07-17), enacted 2026-08-04 by #2493, which removed 54 config keys and archived them. This PR still added 1k1k, so remove that half: - delete the 9 disagg/1k1k/stp recipes - drop the isl:1024/osl:1024 scenario and its 9 search-space points from qwen3.5-fp4-gb300-dynamo-trt, leaving the 8k1k entry byte-identical - reword the perf-changelog entry to describe an 8k1k-only submission The main merge also conflicted in configs/nvidia-master.yaml: this branch anchored the new block before glm5-fp4-gb300-dynamo-sglang, which main deleted with the rest of the GLM-5 family. Resolved by keeping main's replacement key (qwen3.5-fp4-gb300-dynamo-sglang-agentic-agg) and placing the new dynamo-trt block immediately before it, so the file diff against main is purely the added block.
…GLE3-GQA MTP recipe Port of the validated B300 sibling minimaxm3-fp4-b300-vllm-agentic-mtp to b200-dgxc. Same image, same Inferact/MiniMax-M3-EAGLE3-GQA draft head at three speculative tokens, same golden AL 2.78, same TP2/TP4 topologies and SimpleCPU offload arm, so the two SKU curves stay directly comparable. Spec-decode only, per the AgentX policy in MODELS.md. This is MiniMax-M3's only active B200 configuration: the Single-turn 8k1k scenario that carried the previous B200 curves was deprecated in #2493. Two B200 deltas: the checkpoint-resolution guard (launch_b200-dgxc.sh rewrites MODEL to a cluster-local path, so the HF repo id is kept in HF_MODEL_ID and the checkpoint is verified against its shard index behind a flock) and draft staging (the B300 /data/models path does not exist on b200-dgxc). Concurrency per arm is truncated where B200's 180 GB HBM3e puts the cliff. 中文：新增 MiniMax-M3 NVFP4 在 B200 上的 AgentX（agentic-coding）配方，使用 vLLM 搭配 Inferact/MiniMax-M3-EAGLE3-GQA 草稿头做投机解码（投机 token 数 3，黄金接受 长度 2.78）。镜像、草稿模型、TP2/TP4 拓扑与 SimpleCPU 卸载分支均与 B300 同源配方 一致，保证两个 SKU 的曲线可直接对比。按 MODELS.md 的 AgentX 策略仅发布投机解码 分支；#2493 停用 8k1k 场景后，这是 MiniMax-M3 在 B200 上唯一的有效配置。B200 的 两处差异为检查点解析保护（启动器会把 MODEL 改写为集群本地路径）与草稿模型的暂存 路径；各分支并发上限按 B200 的 180 GB 显存下调。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…GLE3-GQA MTP recipe Port of the validated B300 sibling minimaxm3-fp4-b300-vllm-agentic-mtp to b200-dgxc. Same image, same Inferact/MiniMax-M3-EAGLE3-GQA draft head at three speculative tokens, same golden AL 2.78, same TP2/TP4 topologies and SimpleCPU offload arm, so the two SKU curves stay directly comparable. Spec-decode only, per the AgentX policy in MODELS.md. This is MiniMax-M3's only active B200 configuration: the Single-turn 8k1k scenario that carried the previous B200 curves was deprecated in #2493. Two B200 deltas: the checkpoint-resolution guard (launch_b200-dgxc.sh rewrites MODEL to a cluster-local path, so the HF repo id is kept in HF_MODEL_ID and the checkpoint is verified against its shard index behind a flock) and draft staging (the B300 /data/models path does not exist on b200-dgxc). Concurrency per arm is truncated where B200's 180 GB HBM3e puts the cliff. 中文：新增 MiniMax-M3 NVFP4 在 B200 上的 AgentX（agentic-coding）配方，使用 vLLM 搭配 Inferact/MiniMax-M3-EAGLE3-GQA 草稿头做投机解码（投机 token 数 3，黄金接受 长度 2.78）。镜像、草稿模型、TP2/TP4 拓扑与 SimpleCPU 卸载分支均与 B300 同源配方 一致，保证两个 SKU 的曲线可直接对比。按 MODELS.md 的 AgentX 策略仅发布投机解码 分支；#2493 停用 8k1k 场景后，这是 MiniMax-M3 在 B200 上唯一的有效配置。B200 的 两处差异为检查点解析保护（启动器会把 MODEL 改写为集群本地路径）与草稿模型的暂存 路径；各分支并发上限按 B200 的 180 GB 显存下调。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…GLE3-GQA MTP recipe Port of the validated B300 sibling minimaxm3-fp4-b300-vllm-agentic-mtp to b200-dgxc. Same image, same Inferact/MiniMax-M3-EAGLE3-GQA draft head at three speculative tokens, same golden AL 2.78, same TP2/TP4 topologies and SimpleCPU offload arm, so the two SKU curves stay directly comparable. Spec-decode only, per the AgentX policy in MODELS.md. This is MiniMax-M3's only active B200 configuration: the Single-turn 8k1k scenario that carried the previous B200 curves was deprecated in #2493. Two B200 deltas: the checkpoint-resolution guard (launch_b200-dgxc.sh rewrites MODEL to a cluster-local path, so the HF repo id is kept in HF_MODEL_ID and the checkpoint is verified against its shard index behind a flock) and draft staging (the B300 /data/models path does not exist on b200-dgxc). Concurrency per arm is truncated where B200's 180 GB HBM3e puts the cliff. 中文：新增 MiniMax-M3 NVFP4 在 B200 上的 AgentX（agentic-coding）配方，使用 vLLM 搭配 Inferact/MiniMax-M3-EAGLE3-GQA 草稿头做投机解码（投机 token 数 3，黄金接受 长度 2.78）。镜像、草稿模型、TP2/TP4 拓扑与 SimpleCPU 卸载分支均与 B300 同源配方 一致，保证两个 SKU 的曲线可直接对比。按 MODELS.md 的 AgentX 策略仅发布投机解码 分支；#2493 停用 8k1k 场景后，这是 MiniMax-M3 在 B200 上唯一的有效配置。B200 的 两处差异为检查点解析保护（启动器会把 MODEL 改写为集群本地路径）与草稿模型的暂存 路径；各分支并发上限按 B200 的 180 GB 显存下调。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…GLE3-GQA MTP recipe Port of the validated B300 sibling minimaxm3-fp4-b300-vllm-agentic-mtp to b200-dgxc. Same image, same Inferact/MiniMax-M3-EAGLE3-GQA draft head at three speculative tokens, same golden AL 2.78, same TP2/TP4 topologies and SimpleCPU offload arm, so the two SKU curves stay directly comparable. Spec-decode only, per the AgentX policy in MODELS.md. This is MiniMax-M3's only active B200 configuration: the Single-turn 8k1k scenario that carried the previous B200 curves was deprecated in #2493. Two B200 deltas: the checkpoint-resolution guard (launch_b200-dgxc.sh rewrites MODEL to a cluster-local path, so the HF repo id is kept in HF_MODEL_ID and the checkpoint is verified against its shard index behind a flock) and draft staging (the B300 /data/models path does not exist on b200-dgxc). Concurrency per arm is truncated where B200's 180 GB HBM3e puts the cliff. 中文：新增 MiniMax-M3 NVFP4 在 B200 上的 AgentX（agentic-coding）配方，使用 vLLM 搭配 Inferact/MiniMax-M3-EAGLE3-GQA 草稿头做投机解码（投机 token 数 3，黄金接受 长度 2.78）。镜像、草稿模型、TP2/TP4 拓扑与 SimpleCPU 卸载分支均与 B300 同源配方 一致，保证两个 SKU 的曲线可直接对比。按 MODELS.md 的 AgentX 策略仅发布投机解码 分支；#2493 停用 8k1k 场景后，这是 MiniMax-M3 在 B200 上唯一的有效配置。B200 的 两处差异为检查点解析保护（启动器会把 MODEL 改写为集群本地路径）与草稿模型的暂存 路径；各分支并发上限按 B200 的 180 GB 显存下调。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…nt / 执行 2026 年 8 月 6 日 Kimi-K2.5/2.6/2.7-Code 完全退役 (#2527) * feat: enact the August 6, 2026 Kimi-K2.5/2.6/2.7-Code retirement 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> * chore: fill the PR link for #2527 in MODELS.md / MODELS_zh.md 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> --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
…GLE3-GQA MTP recipe Port of the validated B300 sibling minimaxm3-fp4-b300-vllm-agentic-mtp to b200-dgxc. Same image, same Inferact/MiniMax-M3-EAGLE3-GQA draft head at three speculative tokens, same golden AL 2.78, same TP2/TP4 topologies and SimpleCPU offload arm, so the two SKU curves stay directly comparable. Spec-decode only, per the AgentX policy in MODELS.md. This is MiniMax-M3's only active B200 configuration: the Single-turn 8k1k scenario that carried the previous B200 curves was deprecated in #2493. Two B200 deltas: the checkpoint-resolution guard (launch_b200-dgxc.sh rewrites MODEL to a cluster-local path, so the HF repo id is kept in HF_MODEL_ID and the checkpoint is verified against its shard index behind a flock) and draft staging (the B300 /data/models path does not exist on b200-dgxc). Concurrency per arm is truncated where B200's 180 GB HBM3e puts the cliff. 中文：新增 MiniMax-M3 NVFP4 在 B200 上的 AgentX（agentic-coding）配方，使用 vLLM 搭配 Inferact/MiniMax-M3-EAGLE3-GQA 草稿头做投机解码（投机 token 数 3，黄金接受 长度 2.78）。镜像、草稿模型、TP2/TP4 拓扑与 SimpleCPU 卸载分支均与 B300 同源配方 一致，保证两个 SKU 的曲线可直接对比。按 MODELS.md 的 AgentX 策略仅发布投机解码 分支；#2493 停用 8k1k 场景后，这是 MiniMax-M3 在 B200 上唯一的有效配置。B200 的 两处差异为检查点解析保护（启动器会把 MODEL 改写为集群本地路径）与草稿模型的暂存 路径；各分支并发上限按 B200 的 180 GB 显存下调。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…00 AgentX with EAGLE3-GQA MTP and golden-AL synthetic acceptance / 新增 MiniMax-M3 NVFP4 B200 vLLM EAGLE3-GQA MTP AgentX 配方（黄金 AL 合成接受率） (#2511) * feat(minimaxm3-b200-agentic-mtp): add MiniMax-M3 NVFP4 B200 AgentX EAGLE3-GQA MTP recipe Port of the validated B300 sibling minimaxm3-fp4-b300-vllm-agentic-mtp to b200-dgxc. Same image, same Inferact/MiniMax-M3-EAGLE3-GQA draft head at three speculative tokens, same golden AL 2.78, same TP2/TP4 topologies and SimpleCPU offload arm, so the two SKU curves stay directly comparable. Spec-decode only, per the AgentX policy in MODELS.md. This is MiniMax-M3's only active B200 configuration: the Single-turn 8k1k scenario that carried the previous B200 curves was deprecated in #2493. Two B200 deltas: the checkpoint-resolution guard (launch_b200-dgxc.sh rewrites MODEL to a cluster-local path, so the HF repo id is kept in HF_MODEL_ID and the checkpoint is verified against its shard index behind a flock) and draft staging (the B300 /data/models path does not exist on b200-dgxc). Concurrency per arm is truncated where B200's 180 GB HBM3e puts the cliff. 中文：新增 MiniMax-M3 NVFP4 在 B200 上的 AgentX（agentic-coding）配方，使用 vLLM 搭配 Inferact/MiniMax-M3-EAGLE3-GQA 草稿头做投机解码（投机 token 数 3，黄金接受 长度 2.78）。镜像、草稿模型、TP2/TP4 拓扑与 SimpleCPU 卸载分支均与 B300 同源配方 一致，保证两个 SKU 的曲线可直接对比。按 MODELS.md 的 AgentX 策略仅发布投机解码 分支；#2493 停用 8k1k 场景后，这是 MiniMax-M3 在 B200 上唯一的有效配置。B200 的 两处差异为检查点解析保护（启动器会把 MODEL 改写为集群本地路径）与草稿模型的暂存 路径；各分支并发上限按 B200 的 180 GB 显存下调。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * chore: fill perf-changelog pr-link for #2511 中文：为 #2511 填写 perf-changelog 的 pr-link。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(minimaxm3-b200-agentic-mtp): correct NVFP4 checkpoint size to 250 GB The recipe's memory rationale cited a ~220 GB checkpoint, derived by halving the documented ~444 GB MXFP8 figure. nvidia/MiniMax-M3-NVFP4 is actually 250.1 GB across 88 shards, so the derived KV budgets were ~30 GB optimistic on every arm: TP2 has ~74 GB rather than ~104 GB and TP4 ~398 GB rather than ~428 GB on B200, against ~268 GB and ~787 GB on B300. No search-space change -- the corrected numbers only reinforce the existing truncations (TP2 capped at conc 2, host tier engaged at conc 20, hard stop at 40). 中文：修正配方中 NVFP4 检查点大小的说明。此前按 MXFP8 的 ~444 GB 折半估算为 ~220 GB，实际 nvidia/MiniMax-M3-NVFP4 为 250.1 GB（88 个分片），因此各分支推导 出的 KV 预算均偏乐观约 30 GB。搜索空间不变——修正后的数值只会进一步支持现有的 并发截断。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * fix(minimaxm3-b200-agentic-mtp): drop the TP2 arm, it cannot fit the 1M context Both TP2 cells failed identically at engine init on run 31141913741 while every TP4 cell passed: ValueError: To serve at least one request with the model's max seq len (1048576), 37.62 GiB KV cache is needed, which is larger than the available KV cache memory (31.72 GiB). With the ~250 GB NVFP4 checkpoint resident, two B200s leave a 31.72 GiB KV pool at --gpu-memory-utilization 0.9, and vLLM estimates the achievable max model length at 883,840 -- below the agentic corpus context. This is structural, not a concurrency cliff, so no conc-list would have made the arm run and lowering the grid would only burn engine starts. The B300 sibling keeps its TP2 arm; on 288 GB GPUs the same topology has ~268 GB of KV. Recipe is now TP4-only: 8 GPU-resident points plus 3 SimpleCPU points. 中文：移除 TP2 分支。在 run 31141913741 中两个 TP2 单元均在引擎初始化阶段失败，而 所有 TP4 单元均通过：约 250 GB 的 NVFP4 权重驻留后，两张 B200 在 --gpu-memory-utilization 0.9 下仅剩 31.72 GiB 的 KV 池，而按模型 1,048,576 的 max_model_len 计算，单个请求就需要 37.62 GiB。vLLM 估算 TP2 可支持的最大上下文为 883,840，低于智能体语料的上下文长度，因此这是结构性限制而非并发瓶颈，调整 conc-list 无济于事。B300 同源配方保留 TP2 分支（288 GB 显存下该拓扑约有 268 GB KV）。本配方现为纯 TP4：8 个显存驻留点加 3 个 SimpleCPU 点。 Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> * chore: refresh PR #2511 for sweep reuse [skip-sweep] --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
…arks / [NVIDIA] 添加 GB300 Qwen3.5 NVFP4 Dynamo-TensorRT-LLM 分离式基准测试 (#2206) * add dynamo trtllm gb300 qwen3.5 nvfp4 recipes * update PR link * update PR link * fix no preflight * chore: drop deprecated Single-turn 1k1k from qwen3.5-fp4-gb300-dynamo-trt / 移除已弃用的 Single-turn 1k1k 场景 Single-turn 1k1k (isl 1024 / osl 1024) is deprecated for all models since 2026-07-17 (#2263) per MODELS.md, and the deprecation was enacted on 2026-08-04 by #2493, which removed 54 config keys from the active master configs. This submission still added a 1k1k scenario, so it is removed here. - Delete the 9 disagg 1k1k STP recipes under benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/1k1k/ - Remove the isl 1024 / osl 1024 scenario (9 search-space points) from the qwen3.5-fp4-gb300-dynamo-trt block in configs/nvidia-master.yaml - Align the perf-changelog entry to describe an 8k1k-only submission The Single-turn 8k1k half is unchanged: 9 recipes and 9 search-space points remain byte-identical. 中文：根据 MODELS.md，Single-turn 1k1k（isl 1024 / osl 1024）自 2026-07-17 起对所有模型弃用（#2263），并已于 2026-08-04 由 #2493 正式执行（从活跃主配置中 移除了 54 个 config key）。本次提交仍新增了 1k1k 场景，故在此移除：删除 gb300-fp4/disagg/1k1k/ 下的 9 个 STP recipe，从 configs/nvidia-master.yaml 的 qwen3.5-fp4-gb300-dynamo-trt 块中删除 isl 1024 / osl 1024 场景及其 9 个搜索空间 配置点，并同步修改 perf-changelog 描述为仅 8k1k。Single-turn 8k1k 部分保持不变， 9 个 recipe 与 9 个搜索空间配置点逐字节保留。 * Correct kv-p2p-transfer metadata and the srt-slurm changelog reference Two label defects, no behaviour change: - configs/nvidia-master.yaml declared kv-p2p-transfer: nixl, but all nine retained recipes set cache_transceiver_config.backend: UCX on both prefill and decode. That value is propagated verbatim into every result row as kv_p2p_transfer, so each published point was labelled NIXL while running UCX. Corrected to ucx to describe what actually runs. - The perf-changelog entry said the runner clones srt-slurm@sa-submission-q2-2026 while launch_gb300-nv.sh checks out the immutable tag v1.0.29 on this path. Corrected, and the --no-preflight change this PR also makes is now recorded. --------- Co-authored-by: Ankur-singh <ankusingh@nvidia.com> Co-authored-by: adibarra <93070681+adibarra@users.noreply.github.com>
1 participant
Summary
Today is Tuesday, August 4, 2026, so the August 3, 2026 block of the Deprecation Notice in
MODELS.mdis now past its last run day. This PR enacts the scenario and precision retirements from that block.54 config keys move out of the active NVIDIA and AMD master configs into
configs/deprecated/, and their 45 benchmark scripts move into the siblingdeprecated/directories.minimaxm3-8k1kkimik2.5-agenticqwen3.5-bf16Archive files:
configs/deprecated/{nvidia,amd}-{minimaxm3-8k1k,kimik2.5-agentic,qwen3.5-bf16}-master.yaml.Not in this PR: the speculative-decoding A/B retirements
The second table in the August 3 block retires the non-spec-decode agentic arm of five models. That round is deliberately deferred, because the replacement arms do not exist yet:
Every non-spec-decode agentic config therefore stays active here.
MODELS.mdcarries an explicit "Status: not yet enacted" note under that table with the reasoning, and thedsv4/glm5.2matrix rows no longer advertise MTP-only coverage that is not actually in force.Verification
active ∪ archived == origin/mainat the key level,active ∩ archived == ∅, and every entry's parsed value is byte-identical toorigin/main. Blocks were moved as verbatim text, so comments and formatting are preserved. NVIDIA 157 → 125 active + 32 archived; AMD 78 → 56 active + 22 archived.dsv4-fp4-b200-vllm-agentic,glm5.2-fp4-b300-sglang-agentic,minimaxm3-fp4-mi355x-vllm-agentic,minimaxm3-fp8-h100-vllm-agentic,qwen3.5-fp8-mi355x-sglang-agentic,kimik3-fp4-b300-vllm-agentic.(prefix, precision, runner-family, scenario)fingerprint against all 181 surviving configs — zero collisions.benchmark_lib.shsource,INFERENCEX_ROOT,chat_templates/). Every relative reference resolves to an existing file and all passbash -n.kimik2.5_fp4_b200.sh,models_vllm.yaml,KLAUD_DEBUG.md); all repointed.configs/ci-priority.yamlandconfigs/CONFIGS.mdreference none of the deprecated keys.srt-slurm-recipes/YAMLs are left in place for deprecated models, and noperf-changelog.yamlentry is added.中文说明
今日为 2026 年 8 月 4 日（星期二），
MODELS.md弃用公告中 2026 年 8 月 3 日 一节已过最后运行日。本 PR 执行其中的场景与精度下线：54 个配置键从启用的 NVIDIA 与 AMD master 配置移入configs/deprecated/，其 45 个基准测试脚本移入同级deprecated/目录。投机解码 A/B 下线本轮不执行：其替代分支尚未合入 —— 若此刻移除，MiniMax-M3 与 GLM-5.2 将没有任何启用配置，DeepSeek-V4-Pro 与 Qwen3.5 在 AMD 及 SGLang 上的智能体覆盖亦将全部消失。所有非投机解码智能体配置在本 PR 中保持启用，
MODELS.md已注明该表「尚未执行」及其原因。