kali20gakki/msAgent / skills/quantizer/msmodelslim-model-adapt
msmodelslim-model-adapt
Verified为 msModelSlim 创建基础 Transformers 模型适配器(Model Adapter)。 包含创建适配器、实现必需接口与注册安装流程。 适用:Decoder-only LLM、理解类 VLM(仅 LLM/text 部分)。 多模态生成(DiT/扩散,含双/多专家)见下方「多模态生成扩展」节,按需加载。 不适用:Encoder-only、非 Transformers 架构。
references/dit/diagram_fallback.yaml
View on GitHub# 通用多模态生成(DiT/扩散)量化 YAML — 步 2 fallback(adapter 骨架验证用)
#
# 适用:msmodelslim-adapter-verification 步 2 全回退量化。
# 不针对具体模型(不写死 Wan2.2 / HunyuanVideo 字段名)。
# 真实模型量化时由 subagent 按 inference_repo 的 parse_args 调整 inference_config 字段。
#
# 关键约定(与具体模型无关):
# - apiversion 固定 multimodal_sd_modelslim_v1(与 LLM/VLM 的 modelslim_v1 区分);
# - 走 MultimodalSDModelslimV1QuantService + MultimodalPipelineInterface 重构路径;
# - "data-free" W8A8 全动态量化 = 显式写 spec.dataset(推荐正确短名)+
# dump_config.enable_dump: false(短路 prepare_calib_data,不走 inference_dump);
# fallback 阶段不消费 inference_dump_calib_data;
# - inference_config 字段名 = 目标推理仓 parse_args 的 --key(任何具体模型按需改);
# - 顶层除 apiversion / metadata / spec 外禁止额外字段;
# - 双专家 / 多专家 expert key(如 low_noise_model / high_noise_model)须与
# init_model() 返回 dict key 完全一致。
apiversion: multimodal_sd_modelslim_v1
metadata:
config_id: <model_id>_w8a8_fallback # 例:wan2_2_t2v_a14b_w8a8_fallback
score: 0.5 # fallback 阶段无精度评估
verified_model_types:
- <registered_model_type> # 例:Wan2.2-T2V-A14B
label:
w_bit: 8
a_bit: 8
is_sparse: false
kv_cache: false # DiT 默认 false
fa_quant: false # fallback 不启用 FA3
spec:
process:
- type: "linear_quant"
qconfig:
act:
scope: "per_token" # int8 → per_token;mxfp8 → per_block
dtype: "int8"
symmetric: true
method: "minmax"
weight:
scope: "per_channel" # int8 → per_channel;mxfp8 → per_block
dtype: "int8"
symmetric: true
method: "minmax"
include:
- "*"
exclude:
- "*mod*" # 与主仓 hunyuan_video_w8a8f8_mxfp.yaml 对齐;新模型首次 exclude 空
# 双专家 DiT(如 Wan2.2):include 可加 "*.low_noise_model.*" / "*.high_noise_model.*"
# dataset 字段不写 = data-free(不要写 dataset: <name>)
# save_path 由 CLI --save_path 给出
save:
- type: "mindie_format_saver"
part_file_size: 0
multimodal_sd_config:
dump_config:
capture_mode: "args"
dump_data_dir: "" # 留空 = 默认 save_path
# W8A8 全动态量化显式 enable_dump: false(短路 prepare_calib_data,
# 不走 inference_dump_calib_data;spec.dataset 仍建议显式写)
inference_config: # 按目标 DiT parse_args 字段命名
# === 必填字段(按目标 DiT 调整) ===
# Wan2.2-T2V-A14B 字段:
# size: "1280*720"
# frame_num: 81
# sample_steps: 40
# sample_guide_scale: 7.0
# base_seed: 1234
# HunyuanVideo 字段:
# model_resolution: "720p"
# video_size: [720, 1280]
# infer_steps: 50
# cfg_scale: 1.0
# embedded_cfg_scale: 6.0
# flow_shift: 7.0
# FLUX.1-dev 字段:
# num_inference_steps: 28
# guidance_scale: 3.5
# height: 1024
# width: 1024
# === 字段白名单 ===
# 任何字段必须命中 inference_repo parse_args 的 --key;否则 Pydantic 校验 fail-fastRelated 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 |