跳转到主要内容
此内容尚未提供您的语言版本,正在以英文显示。

TensorRT LLM Inference Serving

技能 已验证 活跃

Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency. Use for production deployment on NVIDIA GPUs (A100/H100), when you need 10-100x faster inference than PyTorch, or for serving models with quantization (FP8/INT4), in-flight batching, and multi-GPU scaling.

目的

To enable users to achieve maximum inference throughput and lowest latency for LLMs on NVIDIA GPUs, particularly for production deployments requiring significant speedups and efficient resource utilization.

功能

  • Optimize LLM inference with NVIDIA TensorRT-LLM
  • Achieve high throughput and low latency
  • Support for production deployment on NVIDIA GPUs
  • Utilize quantization (FP8, INT4)
  • Configure in-flight batching and multi-GPU scaling

使用场景

  • Deploying LLMs on NVIDIA A100/H100 GPUs for maximum performance.
  • Serving LLMs with low latency for real-time applications.
  • Optimizing inference costs by using quantization and efficient batching.
  • Scaling LLM serving across multiple GPUs or nodes.

非目标

  • Model training or fine-tuning
  • Usage on non-NVIDIA hardware (e.g., AMD GPUs, CPUs)
  • General application development beyond LLM inference serving

工作流

  1. Review use case and hardware requirements
  2. Install TensorRT-LLM via Docker or pip
  3. Configure and run basic inference or trtllm-serve
  4. Apply optimizations like quantization and batching
  5. Deploy across multiple GPUs or nodes if needed

先决条件

  • NVIDIA GPUs (A100/H100 recommended)
  • CUDA Toolkit (version compatible with TensorRT-LLM)
  • Python 3.10-3.12
  • Docker (recommended for consistent environment)

安装

npx skills add davila7/claude-code-templates

通过 npx 运行 Vercel skills CLI(skills.sh)— 需要本地安装 Node.js,以及至少一个兼容 skills 的智能体(Claude Code、Cursor、Codex 等)。前提是仓库遵循 agentskills.io 格式。

质量评分

已验证
99 /100
1 day ago 分析

信任信号

最近提交1 day ago
星标27.2k
许可证MIT
状态
查看源代码