Add curated-full-sweep matching dashboard chart runs (two active tabs) / 新增匹配仪表盘图表的 curated-full-sweep（两个活跃标签页）

[JordanNanos]

Summary

Adds a curated-full-sweep subcommand to utils/matrix_logic/generate_sweep_configs.py that generates EXACTLY the model×engine×node-type combinations rendered by the ClusterMAX dashboard charts (the source of truth). The dashboard now keeps only two active tabs, so the curated sweep composes the union of exactly two passes:

1. --single-node --model-prefix kimik2.5 --framework vllm
2. --multi-node --model-prefix dsv4 --framework dynamo-vllm llmd-vllm — the chart's dsv4 multi-node is vLLM-only; llmd-vllm emits vllm-prefixed metrics so it renders as vLLM.

The full-sweep filters (--model-prefix/--framework/--single-node/--multi-node) combine with AND within a single invocation, so these distinct chart scenarios cannot be expressed in one full-sweep call, and a bare full-sweep --config-files configs/nvidia-master.yaml sweeps everything (including qwen3.5 and TensorRT).

Excluded: dynamo-sglang, dynamo-trt/trt, and every qwen3.5-* config.

Deprecated / deselected tabs (no longer in the curated sweep):

• dsr1 single-node (SGLang) — deprecated tab, removed from the harness sweep.
• gptoss120b single-node — deprecated, and it has NO active master config (only configs/deprecated/) so it cannot be swept from nvidia-master.yaml anyway.

Only trimming filters are overridable (--precision, --runner-type, --seq-lens, --step-size, --min-conc, --max-conc, --max-tp, --max-ep); model/framework/node-type are fixed by the curated target set.

Data safety — no historical data or config history deleted

This PR only narrows the curated selection. It does NOT delete or rewrite any historical data or config entries:

• perf-changelog.yaml — unchanged (no history entries removed, no whitespace touched).
• configs/nvidia-master.yaml / configs/amd-master.yaml — unchanged; every master-config entry (including dsr1 single-node, qwen3.5, dynamo-sglang, dynamo-trt, and other non-chart configs) is kept intact so other workflows/one-offs can still dispatch them.
• No result artifacts or stored benchmark outputs removed; zero file deletions.

Deselected configs (dsr1 single-node included) simply stop being selected by curated-full-sweep — they no longer run in the curated sweep, but they still exist. The guard test test_master_config_actually_contains_excluded_configs asserts these excluded configs (including dsr1 single-node sglang) are still present in the master config, proving they were deselected, not deleted.

Files changed

• utils/matrix_logic/generate_sweep_configs.py — curated-full-sweep subcommand, generate_curated_full_sweep(), CURATED_SWEEP_PASSES (two passes), _curated_pass_args(), main() routing/validation.
• utils/matrix_logic/test_generate_sweep_configs.py — guard test TestCuratedFullSweepTargetSet running the curated sweep over the real nvidia-master.yaml, asserting the exact set (prefixes ⊆ {kimik2.5, dsv4}; kimik2.5 = SN/vllm; dsv4 = MN/{dynamo-vllm, llmd-vllm}; no dsr1, no qwen, no gptoss, no trt/dynamo-trt/dynamo-sglang) and that excluded configs (incl. dsr1 single-node) still exist in master.
• .github/workflows/README.md, AGENTS.md — documented the two chart-scoped passes, plus the deprecated dsr1/gptoss note.

Testing

• pytest utils/matrix_logic/ -q → 235 passed.
• Manual: curated-full-sweep --config-files configs/nvidia-master.yaml --no-evals → 225 entries = kimik2.5/vllm SN (188) + dsv4/dynamo-vllm MN (32) + dsv4/llmd-vllm MN (5); no dsr1/qwen/gptoss/trt/dynamo-trt/dynamo-sglang.
• perf-changelog.yaml intentionally NOT modified (avoids triggering real runs).

Sweep label

No sweep label applied (intentional — would trigger real GPU runs). A maintainer should add full-sweep-fail-fast to validate.

中文说明

为 utils/matrix_logic/generate_sweep_configs.py 新增 curated-full-sweep 子命令，生成的配置恰好是 ClusterMAX 仪表盘图表（唯一事实来源）所渲染的组合。仪表盘当前仅保留两个活跃标签页，因此 curated 扫描组合恰好两个 pass 的并集：

1. --single-node --model-prefix kimik2.5 --framework vllm
2. --multi-node --model-prefix dsv4 --framework dynamo-vllm llmd-vllm（图表中 dsv4 多节点仅为 vLLM；llmd-vllm 发出 vllm 前缀指标，因此渲染为 vLLM）

排除： dynamo-sglang、dynamo-trt/trt 以及所有 qwen3.5-* 配置。

已弃用/取消选择的标签页（不再纳入 curated 扫描）： dsr1 单节点（SGLang）为已弃用标签页，已从 harness 扫描移除；gptoss120b 单节点已弃用，且无有效 master 配置（仅存在于 configs/deprecated/），本就无法从 nvidia-master.yaml 扫描。

数据安全 —— 未删除任何历史数据或配置历史。 本 PR 仅收窄 curated 选择范围，不删除、不改写任何历史数据或配置条目：perf-changelog.yaml 未改动；configs/nvidia-master.yaml / configs/amd-master.yaml 未改动，所有 master 配置条目（包括 dsr1 单节点、qwen3.5、dynamo-sglang、dynamo-trt 等）均保持完整，其他工作流/一次性任务仍可调度；未删除任何结果产物或存储输出，零文件删除。被取消选择的配置（包括 dsr1 单节点）只是不再被 curated-full-sweep 选中——它们不再参与 curated 扫描，但依然存在。守卫测试 test_master_config_actually_contains_excluded_configs 断言这些被排除的配置（含 dsr1 单节点 sglang）仍存在于 master 配置中，证明它们是被取消选择而非被删除。

测试： pytest utils/matrix_logic/ -q → 235 通过；手动运行得到 225 条配置，符合预期集合，无 dsr1/qwen/gptoss/trt/dynamo-trt/dynamo-sglang。未修改 perf-changelog.yaml。

扫描标签： 未添加扫描标签（有意为之，否则会触发真实 GPU 运行）。维护者应添加 full-sweep-fail-fast 进行验证。

🤖 Generated with Claude Code

[claude]

No bugs found in the implementation, but this PR defines what the canonical production full sweep actually covers (a hardcoded model/framework/engine target set), which is a target-set decision with real GPU-spend implications — worth an explicit maintainer sign-off rather than just a code-correctness check.

What was reviewed: the two-pass composition in generate_curated_full_sweep (single-node kimi/vllm + multi-node dsr1+dsv4/dynamo-vllm+dynamo-sglang+llmd-vllm), the new subcommand's argument wiring (step-size/min-max-conc validation reused from full-sweep), and the guard test against the real nvidia-master.yaml. I also checked whether Kimi's existing multi-node dynamo-vllm/dynamo-trt configs (e.g. kimik2.5-fp4-gb200-dynamo-vllm) are covered — they are not, since neither curated pass's model-prefix/framework filter reaches them; per the PR description this exclusion is intentional (the target set is scoped to kimi single-node vLLM only), not an oversight, but it's worth the author/maintainer confirming that scope is correct.

Extended reasoning...

Overview

This PR adds a new curated-full-sweep subcommand to utils/matrix_logic/generate_sweep_configs.py that composes two generate_full_sweep passes (kimi single-node on vllm; dsr1/dsv4 multi-node on dynamo-vllm/dynamo-sglang/llmd-vllm) into a fixed target set, documents it in AGENTS.md/.github/workflows/README.md as the canonical production full-sweep command, and adds a guard test that runs it against the real configs/nvidia-master.yaml. No existing generate_full_sweep logic is modified.

Security risks

None — this is CLI tooling for generating benchmark matrix configs, no auth/crypto/user-input surface.

Level of scrutiny

Code correctness here is low-risk (additive, isolated, well tested — 233+ passing tests including a new guard test against the real master config). The higher-stakes part is the target-set definition itself: this becomes the documented "canonical production full sweep" that maintainers will dispatch verbatim, so any gap in the target set silently under- or over-covers production GPU benchmarking. I confirmed the generator composition is correct for the stated target set, and that Kimi's existing multi-node dynamo-vllm/dynamo-trt configs are excluded by design per the PR description (kimi scope is single-node vllm only) rather than by a coding mistake — but this is a business/scope call, not something I can validate as "correct" from the code alone.

Other factors

The docs are appropriately exempted from the bilingual-Markdown rule (.github/ internal docs and AGENTS.md are English-only per project convention). Test coverage is strong, including a guard test asserting the exact target set against the real master config. No perf-changelog.yaml change, so this PR itself won't trigger a benchmark run.