Clean up configuration ownership and redundant defaults after PR #2127 / 清理 PR #2127 后的配置归属与冗余默认值

[cquil11]

Summary

Follow up after #2127 merges to clean up configuration ownership and redundant shell defaults across the multi-node DSV4 Agentic path.

Do not modify #2127 for this cleanup; preserve its validated sweep and /reuse-sweep-run merge path. Start the implementation from main after #2127 is merged.

Cleanup rules

Apply these rules consistently across every file introduced or modified by #2127:

• Caller/workflow-owned inputs must be required and consumed directly. Do not add ${VAR:-default} after the caller has already supplied or validated the variable.
• Recipe-owned policy must be assigned once as an explicit value in the recipe.
• Retain an environment override only when it is an intentional, documented interface with a real caller or operational use.
• Remove pass-through self-assignments such as export MODEL_PATH=$MODEL_PATH when they do not change scope or behavior.
• Restore unrelated global behavior to main or move it to a separate infrastructure PR.
• Keep the DSV4 recipe specific to dsv4-fp4-mi355x-sglang-disagg-agentic-hicache.

Initial findings

benchmarks/multi_node/agentic/dsv4_fp4_mi355x_sglang-disagg.sh

Audit every assignment. In particular:

• DURATION and KV_OFFLOADING are already checked by check_env_vars; consume them directly without defaults.
• MODEL_PREFIX, PRECISION, RESULT_FILENAME, MAX_MODEL_LEN, topology, concurrency, and framework values are supplied by the reusable workflow or launcher; do not silently replace them.
• Decide explicitly which of TIME_LIMIT, HiCache settings, MoRI queue tuning, router policy, metrics, and MTP settings are recipe-owned constants versus real supported caller inputs.
• Remove redundant self-exports and fallback chains.
• Derive EP/DP flags from already-required topology inputs without adding fallback values.

Shared runtime files

Review these against the same ownership model:

• benchmarks/multi_node/amd_utils/env.sh
• benchmarks/multi_node/amd_utils/job.slurm
• benchmarks/multi_node/amd_utils/models.yaml
• benchmarks/multi_node/amd_utils/server_sglang.sh
• benchmarks/multi_node/amd_utils/setup_deps.sh
• benchmarks/multi_node/amd_utils/submit.sh
• benchmarks/multi_node/amd_utils/trace_replay.sh
• runners/launch_mi355x-amds.sh

For each added default or override, identify its owner and caller before retaining it.

Generic infrastructure scope

Reassess generic changes separately from recipe cleanup:

• .github/workflows/run-sweep.yml array serialization for multi-node Agentic conc-list/conc.
• utils/matrix_logic/generate_sweep_configs.py changing from up to four concurrencies per allocation to one allocation per concurrency.

The workflow serialization fix may be valid generic infrastructure, but should not be hidden inside a recipe cleanup. The global one-concurrency scheduling policy requires an explicit independent decision.

Validation

After cleanup:

• rebase onto post-[AMD][AgentX] DeepSeek V4 multi-node Agentic framework and recipe / DeepSeek V4 多节点 Agentic 框架与配方 #2127 main;
• run bash -n on every changed shell/Slurm script;
• parse all changed YAML;
• run matrix logic tests;
• run Agentic aggregation tests;
• run relevant dry-run/server command generation checks;
• verify the generated DSV4 matrix and workflow inputs;
• confirm no unrelated existing config changes;
• use a new sweep only if runtime behavior changes.

中文说明

在 #2127 合并后，针对多节点 DSV4 Agentic 路径进行一次独立的配置归属和冗余默认值清理。不要为了本次清理继续修改 #2127，以免影响已经验证通过的扫描结果和 /reuse-sweep-run 合并流程；实现应从 #2127 合并后的 main 开始。

统一遵循以下规则：

• 由 workflow 或调用方提供的输入必须直接使用；如果变量已经由调用方设置或通过校验，不要再次使用 ${VAR:-default}。
• 属于配方自身策略的参数应在配方中明确赋值一次。
• 只有确实存在调用方或运维场景、并作为正式接口记录的参数，才保留环境变量覆盖能力。
• 删除不改变作用域或行为的自赋值，例如 export MODEL_PATH=$MODEL_PATH。
• 与配方无关的全局行为应恢复到 main，或拆分为独立的基础设施 PR。
• DSV4 配方应仅服务于 dsv4-fp4-mi355x-sglang-disagg-agentic-hicache。

需要逐文件审查 #2127 涉及的 recipe、共享 runtime、launcher、workflow 和 matrix 逻辑，明确每个默认值和覆盖项的所有者及真实调用链。清理完成后，执行 Shell、YAML、Python、矩阵逻辑、Agentic 聚合以及相关 dry-run 验证，并确认没有修改无关的既有配置。