Repository navigation
[Klaud Cold] Remove pointers to the deleted single-node bash folders; delete configs/deprecated and amd_utils/deprecated - #3463
Merged
functionstackx merged 6 commits intoSep 26, 2026
Conversation
functionstackx
requested review from
1am9trash,
Ankur-singh,
billishyahao,
chunfangamd,
kedarpotdar-nv,
seungrokj and
yctseng0211
as code owners
September 26, 2026 06:23
Contributor
|
Thanks for the contribution!
中文感谢你的贡献！
|
3 tasks done
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.
Looks good — mostly mechanical cleanup/restoration (removing dead references to the deleted directory, restoring the 14 kernel-config files, trimming launcher branches, updating docs). One inline nit flags a stale comment block that no longer matches the trimmed config; it doesn't affect behavior.
Extended reasoning...
The change touches benchmark/config YAML, shell launcher scripts, restored JSON kernel-config data files, and docs — no auth/crypto/permission code and no security-sensitive surface. The single confirmed finding is a stale explanatory comment in configs/amd-master.yaml (describing bands and a companion key this same diff removed) which is cosmetic/documentation-only, not a functional bug, so it doesn't block approval.
Comment on lines
655
to
657
| # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark | ||
| # speculative decoding. Acceptance is pinned to the committed golden curve in | ||
| # golden_al_distribution/kimik3_dspark_probabilistic_sample_method_block_rejection_sample_method.yaml |
Contributor
There was a problem hiding this comment.
🟡 Nit (optional): maintainers reading amd-master.yaml get explanatory comment blocks that still describe search-space arms and a sibling config key this PR deleted, so the docs no longer match the config. The kimik3-fp4-mi355x-atom-agentic-mtp comment here says 'Companion to kimik3-fp4-mi355x-vllm-agentic-mtp' (that key is deleted by this same diff) and still describes the mid (14,16) and throughput (48,56,72) LMCache/DCP8 bands and two dram-utilization blocks, but the search-space below now has only the interactive band and one dram-utilization: 0.343. …
Why this was flagged
…Fix: update or remove these comments to match the trimmed search-space at all three sites: configs/amd-master.yaml:655-673 (kimik3), :1339-1345 (glm5.2, still describes the removed DCP4 C16-C48 band), and :700-701 (minimaxm3, '0.687 targets ~256 GB/rank on the dram-offload arms' though none remain).
The kimik3-fp4-mi355x-atom-agentic-mtp block comment at configs/amd-master.yaml:655-673 references 'Companion to kimik3-fp4-mi355x-vllm-agentic-mtp', a config key this same diff removes from this file entirely, and describes a 'mid (14,16)' and 'throughput (48,56,72)' LMCache/dcp-size 8 band plus a second dram-utilization: 0.513 block, none of which exist in the search-space at line ~684-686 anymore (only { tp: 8, kv-offloading: none, conc-list: [1, 4] } remains). Same staleness at :1339-1345 (glm5.2-fp4-mi355x-atom-agentic-mtp, still says large concurrency C16-C48 adds DCP4/LMCache though that arm was deleted) and :700-701 (minimaxm3-fp4-mi355x-atom-agentic-mtp, '0.687 targets ~256 GB/rank on the dram-offload arms' though the two LMCache arms were deleted, leaving only kv-offloading:…
Verification: nit. This is a comment-staleness issue introduced by this diff (comments are not executed, so nothing functional breaks; base branch had matching comments+arms). Site 1 (kimik3), configs/amd-master.yaml:655-675: line 660 says "# Companion to kimik3-fp4-mi355x-vllm-agentic-mtp: same checkpoint, same runner." but the diff deletes the whole kimik3-fp4-mi355x-vllm-agentic-mtp: block from this…
…/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>
…toring 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>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- 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>
functionstackx
force-pushed
the
klaud/remove-single-node-agentic-references
branch
from
September 26, 2026 06:38
5f8bd8c to
43cc8af
Compare
functionstackx
changed the title
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>
functionstackx
changed the title
Leave the historical waiver record unchanged. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
functionstackx
deleted the
klaud/remove-single-node-agentic-references
branch
September 26, 2026 06:47
1 task done
Oseltamivir
added a commit
that referenced
this pull request
Sep 28, 2026
…ides / 修正文档、操作手册与实验性指南中的过时内容 Correctness only; no trimming or restructuring. Each English docs/ change is mirrored in its _zh twin. - AGENTS.md, configuration-procedures: *_mtp.sh scripts no longer exist; describe the recipe USE_CHAT_TEMPLATE mechanism. Drop the removed atom-agentic-dspark section, block32 configs and missing test file; sync image pins with the master configs. - MODELS: configs/deprecated and script archives were deleted (#3463, #3461, #3464); fix PR attribution. - ci-procedures: remove the obsolete source-snapshot bullets; add the GB200 deadline. klaud-reporting: add --exclude-newer PT12H. - KLAUD_DEBUG: full-sweep-fail-fast, run-sweep concurrency cancellation, prepare_perf_changelog_merge for changelog conflicts. - architecture: tests live in infx/tests; add multinode_agentic_evals. - workflows README: sync matrix generator usage with argparse. - eval-agentx: re-point all #L anchors. - RECIPES: srt-slurm is v2.23.2 (8dace5f). RUNNER_SETUP, recovery, CONFIGS, golden_al, KNOWN_LIMITATION: runner names, paths, fields. - .claude/commands: Klaud PR filters match klaud/, klaud-cold/, klaude/ (none use claude/); nuke also bumps recipe model.container and no longer truncates perf-changelog.yaml; REST label call in recover-failed-ingest; derive runner pool counts from runners.yaml. - debug-runs skill: single-node reproduction goes through srt recipes. - experimental: CollectiveX MoRI LL coverage and graph replay, swap-blocks --ref main, operatorx QoS/cluster list/workflow status, configs/ paths, InferenceX naming.
yhl-amd
added a commit
that referenced
this pull request
Sep 28, 2026
Move the MiniMax-M3 MI355X ATOM AgentX recipe to rocm/atom-dev:nightly_202609281543, which contains ROCm/ATOM#2414, and restore the TP2 C20/C25/C30 and TP4 C40/C48 LMCache DRAM-offload arms that #3463 dropped. srtctl reserves ATOM's kv-transfer-config, so each DRAM variant enables the in-process lmcache_offload connector with ATOM_KV_OFFLOAD=lmcache and carries the legacy script's LMCache settings. 中文：将 MiniMax-M3 MI355X ATOM AgentX 配方镜像更新为包含 ROCm/ATOM#2414 的 rocm/atom-dev:nightly_202609281543，并恢复 #3463 删除的 TP2 C20/C25/C30 与 TP4 C40/C48 LMCache DRAM 卸载臂。srtctl 占用 ATOM 的 kv-transfer-config， 因此各 DRAM 变体通过 ATOM_KV_OFFLOAD=lmcache 启用进程内 lmcache_offload 连接器，并沿用旧脚本的 LMCache 设置。 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This was referenced Sep 29, 2026
seungrokj
added a commit
that referenced
this pull request
Oct 8, 2026
…ghtly dsv41flash-fp4-mi355x-atom-agentic-dspark: - Restore DeepSeek-V4.1-Flash FP4 ATOM AgentX on MI355X (#3387) on the native srt-slurm recipe srt-slurm-recipes/dsv41flash/atom/mi355x-fp4-mtp/ agentic.yaml (legacy script and master entry were removed in #3461/#3463). TP2 conc [1,2,4,5,6,8,16,32,64,128] (adds 4,5,6,128); TP4 conc [1,4,5,6,8,32] (adds 1,4,5,6; drops 2,16,64). No EP or KV offload. Conc 1-8 covers the batch sizes V4.1 mono decode serves (<= 8 requests). - Image rocm/atom-dev:nightly_202609250902 -> nightly_202610070333 (sha256:2f605a3e8b459a0385f784edb0335c11892277fee650fe42044e4af39e36dd01; ATOM 0.1.7.dev60+g7fc0926c6, ROCm 7.2.4): - ROCm/ATOM#2419: group32 FP8 weight preshuffle and BF16 RoPE tables for V4.1. - ROCm/ATOM#2479: V4.1 mono decode (on by default, <= 8 requests a step for this TP2/TP4, BF16-KV, five-token DSpark deployment) and the FP4 index plane used by upstream recipes/DeepSeek-V4.1-Flash-Agentic.md. - ROCm/aiter#6174: gfx950 tuned rows for six V4.1-Flash TP4 bf16 GEMM shapes that fell back to torch (DSpark verify batches, TP4 lm_head). - ROCm/ATOM#2483: replaces the Triton mono step fence that could spin forever on gfx950; on the 1006 nightlies the TP2 c1 canary hung in warmup on the DSpark draft's first mono step (runs 37486079757, 37493871230). - Server settings follow the upstream recipe (BF16 KV, index cache FP8 -> FP4 per upstream at 7fc0926c, max-num-seqs 128, 16K batching/prefill chunks, prefix caching block 16, 8K state checkpoints, level 3, FULL graphs, dsml_v41 tool parser); every point now captures [1..32, 48, 64, 128]. Five-token DSpark uses golden AL 3.51 (thinking_on) for throughput, real acceptance for eval; 3600 s weka_062126 replay with five warmup requests per lane. - DSpark draft (layers 37-39, inline): no --online_quant_config; attention and projection linears stay native FP8 32x32. Via the image bump, the draft's routed FP4 experts take MXFP8 instead of MXFP4 activations (ROCm/ATOM#2479) and the V4.1 RoPE tables are BF16 instead of FP32 (ROCm/ATOM#2419). The draft has no indexer, so index-cache-dtype fp4 does not apply to it. - TP4 conc 32: ATOM_PREFILL_DECODE_INTERVAL=16 (local 3600 s replay: 108.2 -> 150.9 tok/s/user, ITL p90 9.24 -> 6.63 ms, TTFT p50 0.51 -> 1.14 s, throughput unchanged). - TP2 conc 64: ATOM_PREFILL_DECODE_INTERVAL=4 (30.3 -> 38.8 tok/s/user, 105.4k -> 115.2k tok/s/GPU, 6487 -> 6941 requests, TTFT p90 2.3 -> 6.4 s). All other points unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
seungrokj
added a commit
that referenced
this pull request
Oct 26, 2026
- Image rocm/atom-dev:nightly_202609211553 (ATOM 0.1.7.dev517+g52a6a96fd) -> nightly_202610181506 (ATOM 0.1.8.dev77+gdf86c4829); the atomesh router container_image digest moves from f00a9dc5 (nightly_202609161445) to 1bb1909a, the same nightly as the workers. - Settings follow the GLM-5.2 recipe in ROCm/ATOM PR 2632 (open; recipe document only, measured on df86c4829). - Add TP4 C1 (TP4 now C1-C10) and restore the TP4+DCP4 arm at C16-C48 with the LMCache DRAM tier (0.5.6.dev139+gf22dec28), which #3463 dropped with the other search-space entries without an srt-recipe; TP8 C1-C4 unchanged. - dram-utilization 0.171 -> 0.74, so TOTAL_CPU_DRAM_GB is 1109 GB per TP4 server, covering the 4 x 256 GiB LMCache tier. - TP4 C1 and C2 run GLM-5.2 mono decode (ATOM_MONO_ENABLE=1) at 3 draft tokens (golden AL 2.99), max-num-seqs equal to the concurrency, graphs [1] and [1,2]; C2 was 5 draft tokens, max-num-seqs 4, graphs [1,2,4]. - Every other point sets ATOM_MONO_ENABLE=0, which the new ATOM defaults on; TP4 C4/C8 (5 drafts, AL 3.61), C10 (4, AL 3.33) and TP8 keep their settings. - DCP4 points: decode-context-parallel-size 4, ATOM_DCP_PREFILL_QREP=1, LMCache 256 GiB per rank via extra-kv-connectors lmcache_offload, 2x admission, 4 draft tokens (AL 3.33), 3 at C48 (AL 2.99). - Relevant ATOM changes: GLM-5.2 mono decode (#2479, #2483, #2535); DCP prefill QREP and QREP with MTP (#2437, #2290); DCP prefill fixes with MTP (#2430, #2496, #2506, #2348); MTP verify/draft-graph fixes (#2451, #2367, #2566, #2495); lmcache_offload fixes (#2305, #2339, #2547) and the LMCache pin (#2438). - Draft data types unchanged: the BF16 MTP layer 78 stays BF16 (excluded from ptpc_fp8 by model.layers.78.*), with the target's FP8 KV and FP4 indexer cache; mono decode covers only target layers 0-77. - 镜像 rocm/atom-dev:nightly_202609211553（ATOM 0.1.7.dev517+g52a6a96fd）-> nightly_202610181506（ATOM 0.1.8.dev77+gdf86c4829）；atomesh router 的 container_image digest 从 f00a9dc5（nightly_202609161445）改为 1bb1909a，与 worker 使用同一 nightly。 - 设置与 ROCm/ATOM PR 2632 中的 GLM-5.2 recipe 一致（该 PR 尚未合并，只改 recipe 文档，测量基于 df86c4829）。 - 新增 TP4 C1（TP4 覆盖 C1-C10），并恢复带 LMCache DRAM 层（0.5.6.dev139+gf22dec28）的 TP4+DCP4 C16-C48 分支；该分支曾在 #3463 中随其他没有 srt-recipe 的搜索空间条目一起被删除。TP8 C1-C4 不变。 - dram-utilization 0.171 -> 0.74，每个 TP4 服务的 TOTAL_CPU_DRAM_GB 为 1109 GB，可容纳 4 x 256 GiB 的 LMCache 层。 - TP4 C1 和 C2 以 3 个 draft token（golden AL 2.99）运行 GLM-5.2 mono decode（ATOM_MONO_ENABLE=1），max-num-seqs 等于并发数，graph 为 [1] 和 [1,2]；C2 原为 5 个 draft token、max-num-seqs 4、graph [1,2,4]。 - 其他点位都设置 ATOM_MONO_ENABLE=0（新 ATOM 默认开启）；TP4 C4/C8（5 个 draft，AL 3.61）、C10（4 个，AL 3.33）和 TP8 设置不变。 - DCP4 点位：decode-context-parallel-size 4，ATOM_DCP_PREFILL_QREP=1，通过 extra-kv-connectors lmcache_offload 每个 rank 使用 256 GiB LMCache，准入为 2 倍并发，4 个 draft token（AL 3.33），C48 为 3 个（AL 2.99）。 - 相关 ATOM 改动：GLM-5.2 mono decode（#2479、#2483、#2535）；DCP prefill QREP 及 QREP 与 MTP 同时使用（#2437、#2290）；DCP prefill 与 MTP 的修复（#2430、#2496、#2506、#2348）；MTP verify/draft graph 修复（#2451、#2367、#2566、#2495）；lmcache_offload 修复（#2305、#2339、#2547）和 LMCache 版本固定（#2438）。 - Draft 数据类型不变：BF16 的 MTP 第 78 层保持 BF16（被 model.layers.78.* 排除在 ptpc_fp8 之外），与 target 共用 FP8 KV cache 和 FP4 indexer cache；mono decode 只覆盖 target 第 0-77 层。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Follow-up to #3461 and #3464, which deleted
benchmarks/single_node/agentic/andbenchmarks/single_node/fixed_seq_len/: remove everything that still pointed at them, and deleteconfigs/deprecated/andbenchmarks/multi_node/amd_utils/deprecated/. Rebased onmainafter #3464.1. DSv4.1 Flash H100/H200 SGLang AgentX: drop the block32 tilings
The single-node recipes
dsv41flash/sglang/h{1,2}00-fp4-mtp/agentic.yamlransetup_script: dsv41flash-h{1,2}00-block32.sh, which loaded the block32 installers andkernel_configs/from the deleted folder (broken onmainsince #3461). The tilings are not re-added: this deletes both setup scripts, drops the recipes'setup_scriptentries and the H200 recipe's 16DSV41_BLOCK32_TPenv vars. Both recipes now run with SGLang's default kernel tilings (tiling only; checkpoint data, scales and dtypes were never changed by the setup script).2. Remove the 17 bash-only AgentX search-space entries
Every single-node
agentic-codingentry withoutsrt-recipe:ran a deleted script. Removed:qwen3.5-fp8-b300-sglang-agentic-power-abqwen3.5-fp4-b300-sglang-agentic-power-abqwen3.8next-fp8-h100-sglang-agentic-mtpkimik3-fp4-mi355x-vllm-agentic-mtpminimaxm3-fp4-mi355x-vllm-agentic-mtpdsv41flash-fp4-mi355x-atom-agentic-dsparkkimik3-fp4-mi355x-atom-agentic-mtpminimaxm3-fp4-mi355x-atom-agentic-mtpglm5.2-fp4-mi355x-atom-agentic-mtpThe ATOM LMCache arms can't move to srt-slurm yet (srtctl reserves ATOM's
kv-transfer-configfor disaggregated workers); the recipe header comments now say they were removed.srt-slurm-recipes/kimik3/vllm/mi355x-fp4-mtp/agentic.yamlis kept (it points nowhere legacy) and could revive the Kimi-K3 vLLM config.3. Launchers
launch_gb200-nv.sh/launch_gb300-nv.sh: remove the deaddsv41flashbash branch (theSRT_RECIPEbranch above it already exits for every single-node AgentX point).launch_mi355x-amds.sh: remove theminimaxm3_fp4_mi355x_atom_mtp.shspecial case.4. Docs and comments
MODELS(_zh).mdscenario table,docs/index(_zh).md,docs/DOCUMENTATION_PLAN(_zh).md, the codeowner verify prompt's applicability rule,KLAUD_DEBUG.md, and two code comments now point at the srt-slurm recipes /benchmarks/srt_agentic.shor note the removal.Intentionally left:
docs/waiver/2508.md(historical waiver record), dated history (MODELS deprecation note for #2493,perf-changelog.yaml), the SHA-pinned PR #3205 example in the codeowner prompt, and thefixture_*.shstrings inutils/test_srt_single_node.pythat exercise the generic launcher fallback.5. Delete
configs/deprecated/and move the docs to the srt-slurm flow#3464 merged with only its first commit (the
fixed_seq_len/deletion); its follow-up is carried here:configs/deprecated/{amd,nvidia}-master.yaml(nothing reads them).AGENTS.md"Deprecating benchmark configs",docs/configuration-procedures(_zh).mdand the Claude review prompt in.github/workflows/claude.ymlnow say retired entries are deleted, with git history andperf-changelog.yamlas the record..claude/commands/add-model-hardware.md(the add-model-hardware skill) and the configuration-procedures new-config steps for srt-slurm recipes: no bash script underfixed_seq_len/, no launcher routing,USE_CHAT_TEMPLATE: 'true'for MTP.MODELS(_zh).md: links intoconfigs/deprecated/become plain text; present-tense "archived in" statements now say the archive was removed.benchmarks/multi_node/amd_utils/deprecated/(archivedmodels*.yaml,node_excludes.yaml,glm5_setup.sh); nothing loads them, the live registries sit beside it inamd_utils/. configuration-procedures now says retired AMD registry entries are deleted too.Test plan
generate_sweep_configs.py full-sweepover both master configs passes (it fails onmainbecause of thecluster:b300-nvpower-ab keys this PR deletes); the three trimmed ATOM configs generate 23 GPU-resident srt points; no single-node agentic entry withoutsrt-recipe:remains.bash -non the three edited launchers andserver_sglang.sh.utils/test_srt_single_node.py,utils/test_srt_cluster_config.py,utils/matrix_logic,utils/changelog_gate_tests: 439 passed, 14 failed, identical failure set onmain(pool-launcher tests hitmapfile: command not foundon macOS bash 3.2).perf-changelog.yamlchange, so no sweep is triggered.🤖 Generated with Claude Code