kali20gakki/msAgent / skills/quantizer/quant-tuning-quantize-dit
quant-tuning-quantize-dit
Verified| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 15.7 KB |
| assets/calib_dataset.md | markdown | 2.7 KB |
| assets/index.example.jsonl | other | 395 B |
| assets/w8a8_dynamic.example.yaml | config | 5.5 KB |
| assets/w8a8_mxfp8.example.yaml | config | 5.0 KB |
| references/error_handling.md | markdown | 3.5 KB |
assets/w8a8_mxfp8.example.yaml
View on GitHub# W8A8 MXFP8(data-free)量化 YAML 示例 — DiT(扩散/多模态生成)
#
# 本示例以 HunyuanVideo(重构路径)为参考,结构与 msmodelslim 主仓
# `lab_practice/hunyuan_video/hunyuan_video_w8a8f8_mxfp.yaml` 对齐。
# HunyuanVideo 是当前 **唯一**已迁移到 `MultimodalPipelineInterface` 且
# lab_practice YAML 已 validated 的 DiT;其它 DiT(Wan2.2-T2V/I2V/TI2V、
# FLUX.1-dev(已迁移)、SD3、Sana、HunyuanDiT、CogViewX 等)的 YAML
# 在 `inference_config` 字段集合上需按各自推理仓的 `parse_args` 调整。
#
# 重要:
# - apiversion 固定为 multimodal_sd_modelslim_v1;
# - 走 MultimodalSDModelslimV1QuantService,统一 MultimodalPipelineInterface 重构路径;
# 不再保留 LegacyMultimodalPipelineInterface;
# - "data-free" 指 **不引入外部激活校准数据**;MXFP8 path 通常显式
# 设 `dump_config.enable_dump: False`(per-block 浮点量化不走
# inference_dump_calib_data 路径),但 spec.dataset 仍建议显式写,
# 保持与 int8 路径一致的语义;
# - 本模板仅展示 `linear_quant` 一种 process step;不展示 smooth / quarot /
# iter_smooth / flex_smooth_quant / awq 等抑制/校准策略的 YAML 形态;
# 若引入,由 orchestrator 调 quantization-expert-experience-tuning-rules/
# scripts/apply_rollback.py(整层回退)在 `spec.process` 中追加对应 step
# 后再回到 quantize-dit 执行。
# - 双专家模型(如 Wan2.2-T2V-A14B)的 expert key 须与 init_model 返回 dict 一一对应
# ({"low_noise_model": ..., "high_noise_model": ...})。
#
# 注意:YAML 顶层除 apiversion / metadata / spec 外,**禁止**额外字段
# (model_family / inference_repo 不要写在顶层,
# 会被 MultimodalSDModelslimV1QuantConfig 的 Pydantic schema 拒绝)。
# orchestrator 路由所需的 model_family / inference_repo
# 通过 analysis_result.yaml 与 subagent_io 协议字段传递。
apiversion: multimodal_sd_modelslim_v1
metadata:
config_id: hunyuan_video_w8a8_mxfp8 # 配置标识;不是 label,是 config_id
score: 0.95
verified_model_types: # 与主仓 hunyuan_video_w8a8f8_mxfp.yaml 对齐
- HunyuanVideo
label:
w_bit: 8
a_bit: 8
is_sparse: false
kv_cache: false # DiT 不依赖 KV cache
fa_quant: false # 本阶段不启用 FA3(仅 linear_quant)
spec:
# 本模板仅演示 `linear_quant` 一种 process step;如需叠加其它
# 策略(smooth / quarot / fa3_quant / …),在下方追加对应 step 即可,
# 各 step 的 `exclude` 互不影响,由 yaml_utils.merge_exclude 选择首个
# 含 exclude 的 step 写入。
process:
- type: "linear_quant"
qconfig:
act:
scope: "per_block" # 匹配规则: dtype=int8 → scope=per_token, dtype=mxfp8 → scope=per_block
dtype: "mxfp8"
symmetric: true
method: "minmax"
weight:
scope: "per_block" # 匹配规则: dtype=int8 → scope=per_channel, dtype=mxfp8 → scope=per_block
dtype: "mxfp8"
symmetric: true
method: "mse_round" # MXFP8 weight 经验值用 mse_round(act 仍 minmax)
include:
- "*"
exclude: # 按目标模型实际需要排除;HunyuanVideo 主仓排除 *mod*
- "*mod*"
# 校准 prompt 集:短名(在 msmodelslim/lab_calib/ 下解析)/ 绝对路径 / 相对路径。
# 建议显式写 —— 保持与 int8 路径一致的语义;MXFP8 path 多数场景下不消费
# inference_dump_calib_data,但仍建议给一个正确的短名以便回退/复现。
# 现成短名:hunyuanvideo | wan2_2_t2v | wan2_2_i2v | wan2_2_ti2v
# 自建:目录内放一个 index.jsonl,每行至少含非空 text,I2V 追加 image(相对路径);
# 模板见 assets/index.example.jsonl,说明见 assets/calib_dataset.md。
dataset: hunyuanvideo
# 默认 MindIE-SD 格式
save:
- type: "mindie_format_saver"
part_file_size: 0
multimodal_sd_config:
# MXFP8 是 per-block 浮点量化,不走 inference_dump_calib_data 路径,
# 显式关闭 dump 以避免无谓浮点推理;data-free 语义不变。
dump_config:
enable_dump: false
capture_mode: "args"
dump_data_dir: "" # 留空 → 默认 save_path
# 重构路径统一使用 inference_config(Pydantic 校验,与原推理仓 parse_args 对齐)
# 字段名 / 默认值 / 入口脚本 详见:
# ../../msmodelslim-model-adapt/references/dit/inference_config_field_map.md
# 定位方法见 inference_repo 的 README.md。
inference_config:
model_resolution: "720p" # HunyuanVideo hyvideo.config parse_args
video_size: [720, 1280] # (height, width)
video_length: 129
infer_steps: 50
cfg_scale: 1.0
embedded_cfg_scale: 6.0
flow_shift: 7.0
num_videos: 1
batch_size: 1Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |