107 个 · 按相关度排序
Optimize AgentDB performance with quantization (4-32x memory reduction), HNSW indexing (150x faster search), caching, and batch operations. Use when optimizing memory usage, improving search speed, or scaling to millions of vectors.
中文简介正在准备中。当前显示原文说明。
Foundations of digital signal processing — sampling, aliasing, the Nyquist criterion, quantization, convolution, the discrete Fourier transform, FFT, FIR and IIR filter design, windowing, spectral leakage, and practical issues in fixed-point versus floating-point implementation. Use when designing filters, analyzing spectra, choosing sample rates, reasoning about aliasing, or implementing DSP algorithms on a microcontroller, DSP chip, or host CPU.
中文简介正在准备中。当前显示原文说明。
Optimize AgentDB performance with quantization (4-32x memory reduction), HNSW indexing (150x faster search), caching, and batch operations. Use when optimizing memory usage, improving search speed, or scaling to millions of vectors.
中文简介正在准备中。当前显示原文说明。
Optimize AgentDB with quantization, HNSW tuning, caching, batch ops, and pruning.
中文简介正在准备中。当前显示原文说明。
Optimize AgentDB performance with quantization (4-32x memory reduction), HNSW indexing (150x faster search), caching, and batch operations. Use when optimizing memory usage, improving search speed, or scaling to millions of vectors.
中文简介正在准备中。当前显示原文说明。
Use when working on imp's quantization formats, loaders, or dequant paths - GGUF Q4_0...Q8_0/Q*_K/IQ4/MXFP4, NVFP4 two-level scaling (Modelopt vs compressed-tensors layouts), FP8 E4M3, StorageTier, decode cache, KV-cache dtypes (auto/fp16/fp8/int8/int4/nvfp4/mxfp4), "which quant should I use", "which KV dtype", scale-factor layout, dequant kernel wiring, imp-quantize / AWQ, judging quant quality (PPL). Do NOT use for writing/optimizing GEMM/GEMV kernels (sm120-cuda-expert) or measuring quant perf (benchmark-cuda).
中文简介正在准备中。当前显示原文说明。
Use when writing, reviewing, or optimizing CUDA kernels targeting sm_120a (RTX 5090 / Consumer Blackwell, GB202) in the imp inference engine. Triggers on CUDA/PTX kernel code, shared-memory layout, bank conflicts, tensor-core MMA (mxf4nvf4, tf32, 3xTF32/3xFP16), GEMV/GEMM/attention/quantization/GDN-scan kernels, decode tok/s under expected, kernel emitting HMMA instead of mxf4nvf4, occupancy, register pressure, __launch_bounds__, spills, PDL. Pair with `benchmark-cuda` for measurement and `check-degeneration` after hot-path changes.
中文简介正在准备中。当前显示原文说明。
Generalised autonomous optimisation loop — soft RLVR for any artifact a user can measure. Use this skill whenever a user wants to iteratively improve an artifact — code, prompts, documents, configs, designs, content — by running structured experiments, evaluating results against a multi-dimensional rubric, and learning from each attempt. Triggers include: "optimise this", "keep improving until it's good", "run experiments on", "autoresearch", "iterate on this overnight", "try different approaches and pick the best", or any request implying repeated evaluate-and-improve cycles. Also use when the user wants to improve a system prompt, a data pipeline, a writing style, or any artifact where quality can be decomposed into measurable tracks. For inference optimisation tasks (model latency, throughput, quantization, GPU deployment), a* delegates the low-level tuning to AITune while maintaining quality tracking and learning.
中文简介正在准备中。当前显示原文说明。
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。
中文简介正在准备中。当前显示原文说明。
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。
中文简介正在准备中。当前显示原文说明。
Qdrant vector database -- collection management, point operations, payload filtering, named vectors, quantization, recommendations, snapshots
中文简介正在准备中。当前显示原文说明。
Pulp musical/media time primitives, exact beat divisions, tempo and meter maps, transport-range grid projection, inline and order-preserving groove projection, coordinate randomness, streaming cursors, and quantization arithmetic.
中文简介正在准备中。当前显示原文说明。
Tune moflo's memory stack for speed, RAM, and index quality. Covers HNSW parameters (M, efConstruction, ef), vector quantization, batch operations, and common bottlenecks. Use when scaling past ~100k entries or when search latency regresses.
中文简介正在准备中。当前显示原文说明。
Iterate on a GenAI notebook against the self-hosted stack via the genai-stack CLI (config dirs, auth, subdomains, quantization, GPU/VRAM). Arguments: <notebook|service> [--service comfyui|forge|vllm] [--quant int4|fp8] [--validate] [--bg]
中文简介正在准备中。当前显示原文说明。
量化格式契约与位级对齐：从设备字节反推量化器的**编码公式、舍入模式、scale 粒度、退化块规则**， 并据此重实现/融合量化器（MXFP8 e8m0、int8 perblock、fp8 perchannel 等）， 用「同进程同数据 + 独立参考实现 + 中点与退化输入全覆盖」做到**逐字节精确**。 当自研量化 kernel 与框架算子**对不上**、需要逆向某个量化器的数值契约、需要判断 「量化残差算不算可接受」、或要在融合内核里复现 aclnn 量化语义时使用此 skill。 即使用户只说「量化结果和参考不一致」「这个 scale 怎么算出来的」「融合量化器精度对不上」 而未提"契约"，也应触发。**注意**：单纯选量化档位/配置（该不该开量化、开哪一档）走 dit-perf-opt；算子级性能调优与 DSL 选型走 operator-dev； 并行作用域导致的静默失效（DiT 侧「改了并行但不报错也没生效」的判定）走 dit-parallel-opt。
中文简介正在准备中。当前显示原文说明。
DiT 计算模块（L3）：把**已定位的 DiT 计算瓶颈**落成特性级选档与实施—— 量化档（W8A16 / W4A16 / W8A8 系列 / W4A4 / MXFP8 / FA 量化）、稀疏（rf_v2 / ada_bsa）、 缓存（DiTCache / AttentionCache / 时间步优化）、编译启用（MindieSDBackend / Pattern 融合 / ACLGraph） 的**开不开、开哪一档、怎么开、怎么复验**；依据是 `docs/zh/features/*`（特性真源）+ framework-integration/references/framework-support-matrix.md（支持状态）。 即使用户只说"这个模型怎么加速""量化/稀疏/Cache 怎么选怎么开""要不要开量化、开哪一档""这个档位开了有没有效果" 而未提 profiling，也应触发。 **入口条件**：瓶颈点已明确（用户带一句实测锚点，或编排层交付标签）时由域入口 `performance-optimization` 按标签分发到本技能；**瓶颈未明（"怎么加速 / 跑通 / 采 profile"）先走 `model-auto-optimization` 定位**，不在本技能内做占比分析。 near-miss：多卡并行形态 / 通信掩盖 / TP·offload 选型 → `dit-parallel-opt`；VAE 解码段与 host 固定开销 → 各自模块（VAE / host）；单算子实现级实测选型（mindie_bench）→ `benchmark-dev`； 需要新增 pattern / 算子才能落地本档 → `pattern-dev` / `operator-dev`；框架侧开关与使能验证 → `framework-integration`；量化器位级契约与精度对齐（编码公式 / 舍入 / scale 粒度）→ `quantization-dev`；精度验收判据 → `accuracy-gate`；数字入库口径 → `perf-gate`。
中文简介正在准备中。当前显示原文说明。
Optimize AgentDB performance with quantization (4-32x memory reduction), HNSW indexing (150x faster search), caching, and batch operations. Use when optimizing memory usage, improving search speed, or scaling to millions of vectors.
中文简介正在准备中。当前显示原文说明。
vLLM: high-throughput LLM serving, OpenAI API, quantization.
中文简介正在准备中。当前显示原文说明。
Serve LLMs com alta throughput usando PagedAttention do vLLM e continuous batching. Use ao fazer deploy de APIs LLM em produção, otimizar latência/throughput de inferência, ou servir modelos com memória GPU limitada. Suporta endpoints compatíveis com OpenAI, quantização (GPTQ/AWQ/FP8) e tensor parallelism.
中文简介正在准备中。当前显示原文说明。
Executa inferência de LLM em CPU, Apple Silicon e GPUs consumer sem hardware NVIDIA. Use para edge deployment, Macs M1/M2/M3, GPUs AMD/Intel ou quando CUDA não está disponível. Suporta quantização GGUF (1,5-8 bits) para redução de memória e aceleração de 4-10× vs PyTorch em CPU.
中文简介正在准备中。当前显示原文说明。
Quantiza LLMs para 8-bit ou 4-bit com redução de memória de 50-75% e perda mínima de acurácia. Use quando a memória GPU é limitada, precisa ajustar modelos maiores ou quer inferência mais rápida. Suporta formatos INT8, NF4, FP4, treinamento QLoRA e otimizadores 8-bit. Funciona com HuggingFace Transformers.
中文简介正在准备中。当前显示原文说明。
What every Ollama timing field means and the arithmetic that turns it into a claim - tokens per second, time-to-first-token, why total_duration is not latency, why a second run is faster, what changes with context length, quantization, GPU offload and parallel requests - plus the reporting rules that make a benchmark a measurement instead of an anecdote.
中文简介正在准备中。当前显示原文说明。
vLLM: high-throughput LLM serving, OpenAI API, quantization.
中文简介正在准备中。当前显示原文说明。
This work studies the post-training quantization of LLM parameters as a way to improve their runtime GAP: multi-agent orchestration: competitor signal
中文简介正在准备中。当前显示原文说明。