[Klaud Cold] Delete benchmarks/single_node/agentic - #3461
Conversation
|
Thanks for the contribution!
中文感谢你的贡献!
|
There was a problem hiding this comment.
Findings marked 🟡 are optional suggestions and need no follow-up push.
Additional findings (outside the current diff — GitHub can't attach inline comments there):
-
🔴
benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/h100-fp4-mtp/agentic.yaml— Two single-node srt-slurm AgentX recipes already believed migrated off this legacy folder still break after merge: h100-fp4-mtp/agentic.yaml:33 and dsv41flash/sglang/h200-fp4-mtp/agentic.yaml:24 set setup_script: dsv41flash-h100/h200-block32.sh, which runs install_h100/h200_block32_configs.py against kernel_configs/h100_dsv41_block32 -- all deleted by this diff. The PR's own 'Breaks until ported' table names only multi-node dsv41flash-h100/h200-block32.sh configs, omitting these two single-node recipes, so no owner is on the hook to fix them. Fix: restore or repoint the shared configs/dsv41flash-h100/h200-block32.sh setup scripts before deleting the install scripts/kernel_configs, covering both single- and multi-node callers. [also at: benchmarks/single_node/agentic/install_h100_block32_configs.py:48 - Deleting install_h1{0,2}0_block32_configs.py and kernel_configs/h{1,2}00_dsv41_block32 breaks the active multi-node DSv4.1-Flash H100/H200 recipes, which the base branch runs successfully today. benchmarks/multi_node/srt-slurm-recipes/configs/dsv41flash-h100-block32.sh:5 and…; benchmarks/multi_node/srt-slurm-recipes/configs/dsv41flash-h100-block32.sh:5 - Operators running the DSv4.1-Flash H100/H200 multi-node slurm recipes get a hard failure after merge, since this…; +1 more]Why this was flagged
benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/h100-fp4-mtp/agentic.yaml:33 and .../h200-fp4-mtp/agentic.yaml:24 set setup_script: dsv41flash-h100/h200-block32.sh. Grep of every .yaml under benchmarks/ shows only these two single-node files reference either script; no multi_node/.yaml does. Those scripts (benchmarks/multi_node/srt-slurm-recipes/configs/dsv41flash-h100-block32.sh:5) run under set -euo pipefail and call install_h100_block32_configs.py against kernel_configs/h100_dsv41_block32, both deleted here. On base these recipes run; after merge the setup step hard-fails with a…
Verification: normal, acknowledged in diff (the note's "breaks until ported" claim holds; but it mislabels these as multi-node when they are single-node). Two ACTIVE single-node srt-slurm recipes reference the deleted assets via their setup_script and stay unchanged in the tree: - benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/h100-fp4-mtp/agentic.yaml:33 `setup_script:… | normal, acknowledged in…
-
🟡
docs/index.md— Maintainers following docs/index.md, the repo's "mandatory low-context router," hit a dead link after merge: row 46 still points to benchmarks/single_node/agentic/README.md, which this PR deletes entirely. Context rules tell agents to "Open only the focused page" this table names, so any AgentX-benchmark task routed here 404s instead of reaching real docs. Same stale row exists in docs/index_zh.md:46 and is also cited as the authoritative AgentX guide in docs/DOCUMENTATION_PLAN.md:27,44. Fix: remove or repoint all three references (docs/index.md:46, docs/index_zh.md:46, docs/DOCUMENTATION_PLAN.md:27/44) to wherever AgentX documentation now lives (e.g. the srt-slurm-recipes docs), in this same PR.Why this was flagged
docs/index.md:46 is a row in the "Authoritative references" table of the repo's mandatory documentation router, linking to ../benchmarks/single_node/agentic/README.md with description "AgentX trace benchmark implementation". This diff deletes benchmarks/single_node/agentic/README.md (and the whole directory) but does not update this row. Any engineer or coding agent using docs/index.md to find AgentX benchmark docs, per its own "Context rules" ("Open only the focused page ... needed for the task"), gets a missing-file error instead of documentation. On base branch the link resolves; after merge it is dead. The identical stale link also appears in docs/index_zh.md:46, and docs/DOCUMENTATION_PLAN.md:27 and :44 still cite the same deleted README as the authoritative source to consolidate AgentX docs from, so the documentation-migration plan itself now points at nothing.
Verification: normal (documentation regression this change introduces). docs/index.md line 46 is a live row in the "Authoritative references" table of the doc router:
| [benchmarks/single_node/agentic/README.md](../benchmarks/single_node/agentic/README.md) | AgentX trace benchmark implementation |. This PR deletes benchmarks/single_node/agentic/README.md (git diff shows it removed, 65 deletions; Glob…
…/single_node/agentic reached it: - Drop the 17 search-space entries without srt-recipe: six configs that were entirely bash-based (both B300 Qwen3.5 power A/B keys, qwen3.8next H100, Kimi-K3 and MiniMax-M3 MI355X vLLM, DSv4.1 Flash MI355X ATOM) and the ATOM LMCache DRAM-offload arms of Kimi-K3, MiniMax-M3 and GLM-5.2 on MI355X. - Remove the dead dsv41flash bash branches in the GB200/GB300 launchers and the MiniMax-M3 ATOM script special case in the MI355X launcher. - Restore the DSv4.1 Flash H100/H200 block32 installers and kernel configs under benchmarks/multi_node/srt-slurm-recipes/configs/dsv41flash-block32/ and repoint the setup scripts, which the live single-node H100/H200 SGLang AgentX recipes run and #3461 broke. - Point docs and comments at the srt-slurm recipes and srt_agentic.sh. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… delete configs/deprecated and amd_utils/deprecated (#3463) * [Klaud Cold] Remove everything that pointed at the deleted benchmarks/single_node/agentic reached it: - Drop the 17 search-space entries without srt-recipe: six configs that were entirely bash-based (both B300 Qwen3.5 power A/B keys, qwen3.8next H100, Kimi-K3 and MiniMax-M3 MI355X vLLM, DSv4.1 Flash MI355X ATOM) and the ATOM LMCache DRAM-offload arms of Kimi-K3, MiniMax-M3 and GLM-5.2 on MI355X. - Remove the dead dsv41flash bash branches in the GB200/GB300 launchers and the MiniMax-M3 ATOM script special case in the MI355X launcher. - Restore the DSv4.1 Flash H100/H200 block32 installers and kernel configs under benchmarks/multi_node/srt-slurm-recipes/configs/dsv41flash-block32/ and repoint the setup scripts, which the live single-node H100/H200 SGLang AgentX recipes run and #3461 broke. - Point docs and comments at the srt-slurm recipes and srt_agentic.sh. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Drop the DSv4.1 Flash H100/H200 block32 kernel tilings instead of restoring them Per review: do not re-add the block32 installers and kernel configs. Remove the restored dsv41flash-block32 assets, the two block32 setup scripts, the recipes' setup_script entries and the H200 DSV41_BLOCK32_TP env vars; the H100/H200 SGLang AgentX recipes now run with SGLang's default tilings. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Drop the KLAUD_DEBUG script pointer (fixed_seq_len is deleted in #3464) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Delete configs/deprecated and drop remaining fixed_seq_len pointers - Delete configs/deprecated/{amd,nvidia}-master.yaml; nothing reads them. AGENTS.md, docs/configuration-procedures and the Claude review prompt now say retired entries are deleted, with git history and perf-changelog.yaml as the record. - Rewrite the add-model-hardware playbook and configuration procedures for srt-slurm recipes instead of bash scripts under fixed_seq_len/. - Point MODELS links, the KLAUD_DEBUG example and a deprecated multi-node comment away from the deleted folders. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Delete benchmarks/multi_node/amd_utils/deprecated Nothing loads these archived AMD server-registry files or glm5_setup.sh; the live registries are the models*.yaml and node_excludes.yaml beside them. configuration-procedures now says retired AMD registry entries are deleted rather than moved into deprecated/. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * Revert the docs/waiver/2508.md edit Leave the historical waiver record unchanged. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Main migrated single-node AgentX to native srt-slurm (#3428) and deleted benchmarks/single_node/agentic (#3461). Accept the script deletion and move this PR's change into the declarative recipe: bump the engine image to rocm/atom-dev:nightly_202609231248 (ROCm/ATOM#2366) and set ATOM_PA_FLYDSL=1 and ATOM_PA_FLYDSL_PLAN=1. The bash-only GPU-mask fix no longer applies. The changelog entry follows the default eval policy. 中文:main 已将单节点 AgentX 迁移到 srt-slurm 并删除旧 bash 脚本。本 PR 改为在 YAML 配方中更新镜像至 nightly_202609231248 并启用 ATOM_PA_FLYDSL / ATOM_PA_FLYDSL_PLAN;GPU mask 修复随旧脚本一起移除,changelog 采用默认 eval 策略。 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Summary
Delete the rest of
benchmarks/single_node/agentic/(24 files, −2,973 lines). Single-node AgentX runs as declarative srt-slurm recipes since #3428, and #3460 already removed the 53 unused scripts.These still load files from this folder, so they will fail once this merges unless they are ported to srt-slurm recipes first (or their bash-only entries are removed):
Single-node AgentX search-space entries without
srt-recipe:(17 groups across 9 configs)dsv41flash_fp4_mi355x_atom_mtp.shdsv41flash-fp4-mi355x-atom-agentic-dsparkglm5.2_fp4_mi355x_atom_mtp.shglm5.2-fp4-mi355x-atom-agentic-mtpkimik3_fp4_mi355x_atom_mtp.shkimik3-fp4-mi355x-atom-agentic-mtp(DCP8 LMCache arms)kimik3_fp4_mi355x_mtp.shkimik3-fp4-mi355x-vllm-agentic-mtpminimaxm3_fp4_mi355x_atom_mtp.shminimaxm3-fp4-mi355x-atom-agentic-mtpminimaxm3_fp4_mi355x_mtp.shminimaxm3-fp4-mi355x-vllm-agentic-mtpqwen3.5_fp4_b300_sglang_mtp.shqwen3.5-fp4-b300-sglang-agentic-power-abqwen3.5_fp8_b300_sglang_mtp.shqwen3.5-fp8-b300-sglang-agentic-power-abqwen3.8next_fp8_h100_mtp.shqwen3.8next-fp8-h100-sglang-agentic-mtpMulti-node DeepSeek-V4.1-Flash H100/H200 block32 recipes:
benchmarks/multi_node/srt-slurm-recipes/configs/dsv41flash-h{1,2}00-block32.shruninstall_h1{0,2}0_block32_configs.pywithkernel_configs/h{1,2}00_dsv41_block32, all deleted here.Also stale after this: the GB200/GB300 launcher fallbacks that build
benchmarks/single_node/agentic/...paths (unreached today, since every GB single-node AgentX entry has ansrt-recipe:), theminimaxm3_fp4_mi355x_atom_mtp.shspecial case inrunners/launch_mi355x-amds.sh, and doc mentions.Test plan
perf-changelog.yamlchange, so no sweep is triggered.🤖 Generated with Claude Code