Results for “inference-latency”

Accelerate LLM inference using speculative decoding, Medusa multiple heads, and lookahead decoding techniques. Use when optimizing inference speed (1.5-3.6× speedup), reducing latency for real-time applications, or deploying models with limited compute. Covers draft models, tree-based attention, Jacobi iteration, parallel token generation, and production deployment strategies.

Accelerate LLM inference using speculative decoding, Medusa multiple heads, and lookahead decoding techniques. Use when optimizing inference speed (1.5-3.6× speedup), reducing latency for real-time applications, or deploying models with limited compute. Covers draft models, tree-based attention, Jacobi iteration, parallel token generation, and production deployment strategies.

Reduce per-token latency on Jetson vLLM servers by appending speculative decoding configuration, with guidance on when to enable and how to benchmark the improvement.

Slow down interpretation from observation to action. Use when students or adults need to examine assumptions in conflict, dialogue, or inquiry.

Trains, evaluates, exports, and runs inference for FastFoundationStereo (FFS) stereo depth estimation models, a distilled variant of FoundationStereo with lower latency.

Accelerate LLM inference using speculative decoding, Medusa multiple heads, and lookahead decoding techniques for 1.5-3.6× speedup without quality loss.

The layers trade timeliness against definitiveness.

Patent prior-art and landscape intelligence skill — not generic patent help. Commits to one of five sub-use-cases via forcing intake (novelty search / freedom-to-operate / competitive landscape / acquisition diligence / litigation prior-art) before any search runs. Searches Google Patents, Espacenet, USPTO, and optionally Lens.org for citation-graph signals. Output is an editable Word document (.docx) with verdict, ranked closest art (claim-text extracted), CPC-class-aware landscape, family-resolved hits, geographic coverage, FTO flags where applicable, strategy recommendations, and full audit log. Use when the user asks for patent searching or analysis (e.g., 'prior art search for [invention]', 'freedom to operate analysis for [product]'). Produces search signal, not legal advice — always recommends consulting a patent attorney before filing or licensing decisions. Trademark, copyright, and trade-secret questions are out of scope.

Authenticates and connects to Google Cloud Agent Platform for inference with Gemini and third-party OpenMaaS models (Llama, DeepSeek, Qwen). Generates code for multiple SDKs, configures endpoints, and troubleshoots common errors.

Orchestrate token-heavy research, coding, and testing by delegating bounded tasks to cheaper subagents while reserving Claude Fable for architecture, synthesis, and final review.

Diagnoses why someone knows what to do but doesn't do it, using Adlerian psychology to identify root causes like procrastination, perfectionism, or avoidance.

Design privacy-aware observability for AI agents using traces, spans, structured events, metrics, cost attribution, dashboards, alerts, and investigation workflows. Use when instrumenting an agent, debugging intermittent tool or model failures, defining service-level objectives, analyzing latency or spend, auditing agent decisions, or preparing production monitoring.

Expert en inference llama.cpp avancé (GGUF, quantization, local models, HTTP server, hardware)

Strategic patterns for solving intractable problems through cascading approximation, self-improvement, and heterogeneous evaluation from DeepMind's AlphaGo system

research-synthesis

Econometrics skill for panel data models. Activates when the user asks about: "panel data", "fixed effects", "random effects", "Hausman test", "within estimator", "between estimator", "two-way fixed effects", "clustered standard errors panel", "FE model", "RE model", "pooled OLS", "unobserved heterogeneity", "panel regression", "first difference estimator", "entity fixed effects", "time fixed effects", "面板数据", "固定效应", "随机效应", "豪斯曼检验", "双向固定效应", "面板回归", "个体效应", "时间效应", "一阶差分"

Temporal interval algebra for reasoning about time relationships in planning and knowledge representation

Proactively hunt for threats by analyzing IOCs, detecting behavioral anomalies in telemetry, and prioritizing signals mapped to MITRE ATT&CK.

`analysis-agent`/`task-agent`/`review-agent`: use when latency, throughput, bundle, memory, CPU, query, rendering, or resource cost needs a budget; skip without performance risk.

Diagnoses whether a user's current approach is trading short-term speed for long-term pain, and recommends slower methods that build compoundable assets through friction.

Think through the consequences of consequences — not just what happens immediately, but what happens next, and next after that, across time. Load when a decision looks obviously good or obviously bad on initial read, when the user is optimising for a short-term outcome that might create a long-term problem, when unintended consequences are a concern, or when deep-thinking diagnoses a second-order frame. Triggers on "what are the downstream effects", "what happens after that", "unintended consequences", "think ahead on this", "long-term vs short-term", or "what comes after that". Based on Howard Marks second-level thinking and Farnam Street mental models. Most powerful for decisions with delayed consequences or systemic effects.

Build AI applications with the inference.sh Python SDK: run apps, build agents, and integrate with 250+ models using sync/async, streaming, file uploads, and a tool builder API.

After any AI-generated explanation, require the learner to identify one place it could be wrong, one thing to check, and one source to consult. Builds epistemic vigilance — treats AI output as a claim to evaluate, not truth to absorb.

Econometrics skill for panel data models. Activates when the user asks about: "panel data", "fixed effects", "random effects", "Hausman test", "within estimator", "between estimator", "two-way fixed effects", "clustered standard errors panel", "FE model", "RE model", "pooled OLS", "unobserved heterogeneity", "panel regression", "first difference estimator", "entity fixed effects", "time fixed effects", "面板数据", "固定效应", "随机效应", "豪斯曼检验", "双向固定效应", "面板回归", "个体效应", "时间效应", "一阶差分"

Investigates a question systematically before answering — verifying claims through sources, distinguishing known facts from inferences, and calibrating expressed confidence to actual certainty. Use when asked to research a topic, verify a claim, evaluate a technology, or answer a question where accuracy matters more than speed.

通用时间抽象接口。当定义时间维度、时序关系、时间演化时调用此技能。

Resolves legacy references to the cyber-intel capability by routing to the current runtime agent or skill.

根据研究者提供的**研究计划书（Research Proposal）**执行基于**外国（美国/欧盟/英国/日本/跨国）制度环境**的公司金融类实证研究全流程。**启动后第一件事：根据计划书的主题、识别策略、贡献边际与样本范围，从外国 CF 顶刊池（AER/QJE/JPE/JF/JFE/RFS/JFQA/JAR/JAE/MS/JCF/JBF 等 25+ 期刊）中推荐 5 本最匹配的目标期刊（[J1]–[J5]），等待研究者明确选定一本；该期刊决定 main.tex 的 bibliographystyle、Section 骨架、Introduction 风格与表注规范**。然后用 Python 完成数据清洗、描述性统计、基准回归、内生性检验（IV/2SLS、DML）、平行趋势、异质性、机制、稳健性检验与图表绘制。LaTeX 表格和图像严格遵循 template/ 示例格式，研究逻辑与排版严格遵循 rule/ 下的《通用实证研究逻辑与规范总结》与《回归表写作规范总结》。数据集与政策集从 asset/ 中按计划书中的关键词检索（WRDS / NBER/Fed releases / FRED / 全球宏观库）。**当计划书预期的实证结果无法实现时（系数不显著、平行趋势不通过、IV 弱工具、机制不成立等），skill 自动切换备选方案直至完成研究项目**。最终交付物：Python 代码 + LaTeX 表格 + 图像（.pdf/.png）。触发条件：研究者提交研究计划书（含 X→Y 假设、识别策略、样本、政策冲击等）。

選択肢(A/B/C)の深掘り比較→淘汰→推奨で判断負担を下げ判断の質を上げるスキル。5段階（L1固定3点/L1.5案拡張Diverge・自動/L2評価軸マトリクス/L3複数LLM弁証論/L4過去判断照合）。 「比較して」「深掘りして」「メリデメ教えて」「お勧めは？」「徹底的に」「過去の判断と照合」「前にどう決めたっけ」「/sentaku」等で発火。teian(浅)の深掘り要求を受け取り、brainstorming(深:設計全体)と棲み分け。

Creates digital banners for social, display, hero, and email formats, applying behavioral science principles to maximize attention and conversion.