执行模型量化。通过 msmodelslim quant 依据 Practice YAML 对模型进行量化。
The English summary is being prepared. Showing the original description.
58 skills (repositories by kali20gakki) · sorted by popularity
执行模型量化。通过 msmodelslim quant 依据 Practice YAML 对模型进行量化。
The English summary is being prepared. Showing the original description.
Aggregate DiT inference outputs into per-dimension VBench metrics via AISBench. Run after the quant-tuning-evaluate DiT workflow produces a directory of mp4s and the user supplies a VBench-1.0-mini JSON + pre-populated cache dir. Single mode (Standard).
The English summary is being prepared. Showing the original description.
在 W8A8 动态量化后,用 MindIE-SD 冒烟测试 DiT 量化权重是否可加载(可选 skill)。与 `quant-tuning-evaluate`(DiT 扩展节)的区别:本 skill 仅做"能跑通 + 权重可用"的轻量验证(4~16 条 prompt,几分钟出结果),不算分;后者产出视频集交 `quant-tuning-score-dit`(AISBench-VBench 16 维度)评分,是调优回路一环。不引入 vLLM-Ascend、不做精度指标评估(FID/CLIP-score 等)。
The English summary is being prepared. Showing the original description.
End-to-end automated model quantization and accuracy tuning workflow. Use when user asks for automated model quantization and accuracy tuning, e.g. "自动量化", "量化调优", "一键量化", "精度调优", etc. 当用户给出多卡(≥2 卡)卡号时,量化阶段通过 EP 并行适配(msmodelslim-ep-parallel-adaptation)保证多卡 EP 并行,调优主流程仍在本 Skill。
The English summary is being prepared. Showing the original description.
对 DiT(扩散/多模态生成)模型执行 W8A8 动态(data-free)或 W8A8 MXFP8(data-free)量化。覆盖 HunyuanVideo、Wan2.2-T2V/I2V/TI2V、FLUX.1-dev(已迁移)、SD3、Sana、HunyuanDiT、CogViewX 等。YAML apiversion 固定 multimodal_sd_modelslim_v1,统一 MultimodalPipelineInterface 重构路径(不再保留 LegacyMultimodalPipelineInterface)。模板与字段以 msmodelslim/model/hunyuan_video/model_adapter.py(唯一已迁移到重构路径且 lab_practice YAML 已 validated 的 DiT)为参考;其它 DiT 需按各自推理仓的 `parse_args` 调整字段集合。data-free 指不引入外部激活校准数据,spec.dataset(校准集短名,短名如 wan2_2_t2v;enable_dump: false 时不参与 dump)仍必填;精度调优通过下游经验库 L2 §7 + 直接调 `quantization-expert-experience-tuning-rules/scripts/apply_rollback.py`(整层回退)扩展 exclude 列表,闭环由 `quantization-accuracy-tuning-orchestrator` 调度。
The English summary is being prepared. Showing the original description.
Use when 量化调优闭环中需要生成或修改一轮调优所需的 Practice YAML,包括敏感层分析、策略决策、写出 YAML 文件和校验。
The English summary is being prepared. Showing the original description.
量化专家经验调优库:按「L1 通用专家调优意见 → L2 结构化专家调优意见 → L3 模型专属专家调优意见」三级递进, 输出量化精度调优的完整手段(离群值抑制 / 量化方法·粒度·对称性选型 / 校准集调整 / 敏感层回退 / 模型专属策略), 每条结论附专家原因分析与专家意见可信度等级。 本 Skill 是专家经验库,负责说明可选离群值抑制算法并询问用户选择哪些算法,以及回答「怎么调、哪些层需要回退及为什么」;不运行离群值抑制 processor,不修改 YAML、不执行量化、不做 EP 检查 / 服务化 / 任务评测。
The English summary is being prepared. Showing the original description.
重采样:把已有多轮对话数据的每条 assistant 回答丢弃,用 verifier 模型/vLLM 服务逐轮重生成,产出可直接进训练的预分词样本(每行含 input_ids/loss_mask)。依赖支持 return_token_ids 的 OpenAI 兼容 chat 服务(服务需由用户提前就绪,本技能不负责拉起服务)。触发:用某模型对某个数据集做重采样/重生成(如“对这个数据集用这个模型做重采样”)、把对话回答用模型重写、生成预分词样本/on-policy/MTP 训练数据。
The English summary is being prepared. Showing the original description.
Interactively gather parameters for `python -m cli.inference.text_generate`, generate a concrete text generation simulation command, explain assumptions, ask for execution confirmation, then run the command and summarize performance metrics. Use when the user wants to validate a specific model, hardware profile, batch/query size, prefill or decode scenario, TP/DP/EP/MOE parallel strategy, profiling database, chrome trace, empirical metrics export, or a best row produced by throughput_optimizer.
The English summary is being prepared. Showing the original description.
为 msModelSlim 适配流程注入反量化能力。先识别模型权重是否可反量化,再实现反量化脚本并接入 model_adapter。当前仅覆盖 FP8 的 per-block 与 per-channel 两类;若格式不确定或无公开反量化规则,要求用户提供反量化脚本或浮点权重。
The English summary is being prepared. Showing the original description.
当用户需要部署 msmodeling optix 服务化自动寻优工具时使用。负责安装与验证。
The English summary is being prepared. Showing the original description.
提供 msModelSlim 的通用快速量化指引,包含安装、最简 YAML 配置与基础执行校验。适用于用户询问 msmodelslim 安装、快速量化、配置 yaml、linear_quant 或 minmax 基础参数时。
The English summary is being prepared. Showing the original description.
在实现适配器前对候选模型做分析。确定模型实现来源(transformers / 模型目录 / diffusers)、结构特征、内存约束下的逐层量化建议(可选)及 MoE 融合权重风险。适用于用户询问模型适配可行性或做适配前分析时使用。识别为 DiT 扩散模型时,索取推理仓路径后回传 `next_step: model-adapt`,由 orchestrator 委派 `msmodelslim-model-adapt`(DiT 扩展节承接)。
The English summary is being prepared. Showing the original description.
运行程序采集 Profiling 数据,再用 msprof-analyze 分析TFLOPS 与 MFU。
The English summary is being prepared. Showing the original description.
当首次使用 msmodeling optix 的用户需要根据硬件、模型、负载和优化目标推荐 MindIE/vLLM 寻优参数、搜索范围、benchmark 侧字段或 config.toml 片段时使用。
The English summary is being prepared. Showing the original description.
在需要根据自然语言规格为未支持硬件新增或更新 DeviceProfile 设备画像条目时使用
The English summary is being prepared. Showing the original description.
当用户提供 MindStudio profiler 采集的性能数据(框架 profiler、msprof 命令行)时,对数据完整性、采集状态及关键配置进行校验,确保后续分析工具能正常运行。
The English summary is being prepared. Showing the original description.
Use when diagnosing NPU + PyTorch or LLM Serving Host CPU affinity, NUMA locality, cgroup/cpuset constraints, CPU range conflicts, PyTorch/runtime threading, DataLoader, tokenizer, scheduler, vLLM-Ascend, SGLang, TTFT, TPOT, tokens/s, QPS, or multi-rank/multi-worker CPU binding issues.
The English summary is being prepared. Showing the original description.
为 msModelSlim 适配器实现逐层量化(按层加载/懒加载)能力。仅在用户明确要求逐层量化或基础适配因 CPU 内存不足无法全量加载权重时使用。该特性为高阶可选项,不是基础适配必需项。
The English summary is being prepared. Showing the original description.
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.
The English summary is being prepared. Showing the original description.
分析 msProbe compare 精度比对结果文件(CSV/XLSX),定位整网精度问题与数值差异。支持统计量模式,输出分析报告。支持辅助信息协助定位:辅助信息(历史结论/已知根因/已排除算子/嫌疑区域等)建议在请求中附带(不依赖交互工具)。若请求中未附带且运行环境无 AskUserQuestion,则跳过辅助信息收集、照常完成分析,报告中省略辅助信息章节并注明原因。
The English summary is being prepared. Showing the original description.
多卡分布式训练中的 loss/gnorm 精度溢出检测与根因追溯。基于 MSProbe dump 数据,先跨 rank 定位首次出现 NaN 的源卡,再在源卡上追溯具体的溢出根因算子。 当用户需要:(1) 多卡分布式训练场景下的 NaN/Inf 溢出检测 (2) 找出首先出现 NaN 的源卡 (3) 追溯根因计算算子 (4) loss/gnorm NaN 问题定位 时使用此 skill。
The English summary is being prepared. Showing the original description.
训练与推理数据不一致的端到端根因分析。当模块映射/值比较不足以定位问题时使用,可追踪首个可信的分歧边界,过滤融合或结构性误报,遵循生产者-消费者链,并生成包含具体假设和证据的根因报告。
The English summary is being prepared. Showing the original description.
Analyze Ascend NPU collective communication profiling data with a DB-first workflow. Use when the user provides `cluster_analysis_output/cluster_analysis.db`, rank-level `analysis.db`, rank-level `ascend_pytorch_profiler_{rank_id}.db`, together with `profiler_info.json`, and asks about HCCL or hcom communication cost, collective communication TOP ops, wait time, slow rank/straggler, Notify Wait, bandwidth, retry, relay, SDMA/RDMA/HCCS links, communication matrix, or Ascend communication fault patterns.
The English summary is being prepared. Showing the original description.