Skill: DiT 调优历史追加

SkillAI & models

Append a DiT-tuning round to history.yaml. DiT-side counterpart of the LLM-path accuracy_append.py; records practice.md5, inference_outputs, fp_baseline_outputs, and the scoring fields (scores / overall_score / loss_vs_baseline / is_satisfied) populated by quant-tuning-score-dit. Idempotent on practice_id.

Available today. Use it from your connected AI after setup.

Add ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Then ask your AI: use the Skill: DiT 调优历史追加 skill

What this skill tells your AI

The instructions your AI receives, as published by kali20gakki/msagent in skills/quantizer/quant-tuning-history-append-dit/SKILL.md and read by ahel’s review.

1. 概述

每轮 DiT 调优结束后,orchestrator 调用本 Skill 把本轮记录追加到 {workdir}/history/history.yaml。

2. 适用与不适用

  • 适用:model_family=dit 调优回路;每轮 inference 跑完即追加
  • 不适用:
    • LLM/VLM 路径(用既有 accuracy_append.py)
    • 评分字段本身由 quant-tuning-score-dit 填充;本 skill 仅负责把这些字段透传到 history.yaml(若 quant-tuning-score-dit 未触发则保持 null)

3. 协作关系

quant-tuning-evaluate DiT 扩展节 (产出 infer_outputs/round_N/...)
   │
   ▼
quant-tuning-history-append-dit (本 skill)
   │  scripts/append.py
   ▼
{workdir}/history/history.yaml  ←  dit_records 段

4. 输入参数

参数类型必填说明
history_pathstr✅{workdir}/history/history.yaml
practice_idstr✅唯一 ID(如 dit-round-2)
practice_pathstr✅{workdir}/round_{N}/practice.yaml(自动算 md5)
inference_outputslist✅本轮推理产物路径列表
fp_baseline_outputslist⛏️FP baseline 产物路径(quant-tuning-evaluate DiT 扩展节跑 FP baseline 模式时填)
scoresobject⛏️每维度 score dict;来自 quant-tuning-score-dit 的 scores 字段
overall_scorefloat⛏️加权总分;来自 quant-tuning-score-dit 的 overall_score 字段
loss_vs_baselinefloat⛏️量化 vs FP 的 overall 差;启用 --baseline-outputs 时存在
is_satisfiedbool⛏️loss_vs_baseline >= -tolerance;orchestrator 据此决定是否退出回路
append_asstrYAML 段名,默认 dit_records

5. 工作流

┌──────────────────────────────────────┐
│ 1. 入参校验                            │
│ - history_path / practice_path 存在   │
│ - inference_outputs 非空             │
└──────────────┬───────────────────────┘
               ▼
┌──────────────────────────────────────�
│ 2. 计算 practice.yaml md5            │
└──────────────┬───────────────────────┘
               ▼
┌──────────────────────────────────────┐
│ 3. 读现有 history.yaml(若有)         │
│ - 找 dit_records 段                   │
│ - 按 practice_id upsert(替换/追加)  │
└──────────────┬───────────────────────┘
               ▼
┌──────────────────────────────────────┐
│ 4. 写回 history.yaml(保留 LLM 记录) │
└──────────────────────────────────────┘

6. CLI 调用

python msagent/skills/quantizer/quant-tuning-history-append-dit/scripts/append.py \
    --history-path output/wan22-t2v-a14b-w8a8/history/history.yaml \
    --practice-id dit-round-2 \
    --practice-path output/wan22-t2v-a14b-w8a8/round_2/practice.yaml \
    --inference-outputs \
        "output/wan22-t2v-a14b-w8a8/infer_outputs/round_2/overall_consistency/0000.mp4,output/wan22-t2v-a14b-w8a8/infer_outputs/round_2/subject_consistency/0001.mp4"

7. 输出结果

{
  "protocol": "msagent.subagent_io",
  "subagent_type": "quant-tuning-history-append-dit",
  "status": "ok",
  "output": {
    "ok": true,
    "record": {
      "practice_id": "dit-round-2",
      "quant_config_md5": "44c42e68...",
      "time": "2026-08-08 12:34:56",
      "practice_path": "/abs/path/to/round_2/practice.yaml",
      "inference_outputs": [
        "/abs/path/to/infer_outputs/round_2/overall_consistency/0000.mp4",
        "/abs/path/to/infer_outputs/round_2/subject_consistency/0001.mp4"
      ],
      "fp_baseline_outputs": null,
      "scores": null,
      "overall_score": null,
      "loss_vs_baseline": null,
      "is_satisfied": null
    }
  }
}

8. 错误处理

错误处理
history_path parent not writable立即中止
practice_path not found立即中止
inference_outputs 为空立即中止(防止无效 history 记录)
YAML 解析失败报 stderr 摘要,立即中止

9. 约束

  • 幂等:同一 practice_id 重复调用覆盖旧记录而非重复追加
  • 不破坏 LLM 记录:仅在 dit_records 段追加,与 LLM records 段平行
  • 不修改既有字段名:practice_id / quant_config_md5 / time 与既有 LLM schema 对齐
  • 错误即停

10. 参考

Signals

GitHub stars
31
Forks
8
Last commit
Sep 2026
Advanced
Item type
skill
Key
quant-tuning-history-append-dit
Source
github.com/kali20gakki/msagent