Centralize runner model path mappings in runners.yaml / 在 runners.yaml 中集中管理 runner 模型路径映射

[cquil11]

Summary

Move every launcher-level model lookup table that maps MODEL_PREFIX / PRECISION / FRAMEWORK to cluster-local model paths and aliases into the typed models section of configs/runners.yaml.

The migration covers all six launchers that contained this pattern:

• B200 DGXC
• B300 NV
• GB200 NV
• GB300 NV
• H100 DGXC
• H200 DGXC

The launchers now share one shell entrypoint and one Python resolver. The resolver supports:

• framework-specific mappings with runner-level defaults;
• exact precision mappings with an optional default precision;
• ordered local path probing and deterministic fallback;
• the existing opt-in MODEL_PATH operator overrides;
• SRT_SLURM_MODEL_PREFIX, SERVED_MODEL_NAME, and MODEL_NAME exports;
• explicit portable-model fallback for launch paths that already allowed it.

Runner-config validation rejects unknown runner labels, relative paths, duplicate paths, and undeclared fields. Model-specific branches that select srt-slurm revisions, recipes, or setup workarounds remain in Bash because they are orchestration behavior rather than model-location metadata.

No perf-changelog.yaml entry is included because this refactor is intended to preserve the existing benchmark matrix and runtime model selection behavior.

Validation

• python -m pytest utils/matrix_logic/ -q — 238 passed
• Resolver coverage for every model entry in configs/runners.yaml
• Resolver coverage for all 56 affected NVIDIA multi-node configurations
• Full fixed-sequence config generation — 1,132 entries
• bash -n for all six migrated launchers and the shared helper
• ShellCheck on the shared helper; no new resolver-source findings in the launchers
• python -m py_compile for the resolver and validation modules
• git diff --check

中文说明

将所有启动器中按 MODEL_PREFIX / PRECISION / FRAMEWORK 映射集群本地模型路径与别名的查找表，统一迁移到 configs/runners.yaml 中带类型校验的 models 配置段。

本次迁移覆盖存在该模式的全部六个启动器：

• B200 DGXC
• B300 NV
• GB200 NV
• GB300 NV
• H100 DGXC
• H200 DGXC

这些启动器现在统一复用一个 Shell 入口和一个 Python 解析工具。解析器支持：

• 框架专属映射及 runner 级默认配置；
• 精确精度映射及可选的默认精度；
• 有序本地路径探测和确定性回退；
• 保留现有的可选 MODEL_PATH 运维覆盖机制；
• 导出 SRT_SLURM_MODEL_PREFIX、SERVED_MODEL_NAME 和 MODEL_NAME；
• 对原本允许使用便携式模型标识的启动路径提供显式回退。

Runner 配置校验会拒绝未知 runner 标签、相对路径、重复路径及未声明字段。用于选择 srt-slurm 版本、配方或环境修复逻辑的模型专属分支仍保留在 Bash 中，因为这些属于作业编排行为，而不是模型存放位置元数据。

本次未修改 perf-changelog.yaml，因为该重构旨在保持现有基准测试矩阵及运行时模型选择行为不变。

验证

• python -m pytest utils/matrix_logic/ -q — 238 项测试通过
• 覆盖 configs/runners.yaml 中的每一条模型映射
• 覆盖全部 56 个受影响的 NVIDIA 多节点配置
• 完整生成固定序列长度配置，共 1,132 项
• 对六个已迁移启动器及共享工具执行 bash -n
• 对共享工具执行 ShellCheck；启动器未引入新的解析器 source 告警
• 对解析器及校验模块执行 python -m py_compile
• git diff --check