Skip to main content
Ctrl+K
TensorRT Edge-LLM - Home TensorRT Edge-LLM - Home

TensorRT Edge-LLM

TensorRT Edge-LLM - Home TensorRT Edge-LLM - Home

TensorRT Edge-LLM

Table of Contents

Getting Started

  • Overview
  • Official Support Matrix
  • Supported Models
  • Installation
  • Quick Start
  • Experimental ONNX-less Engine Build
  • Limitations and Known Issues

Examples

  • Examples
  • Speculative Decoding
  • Phi-4-Multimodal
  • ASR (Automatic Speech Recognition)
  • TTS (Text-to-Speech)
  • Vision-Language-Action
    • Alpamayo-R1-10B (Vision-Language-Action)
    • Cosmos3-Edge
  • Omni (Audio + Vision + Speech I/O)
  • Experimental Python API and Server
  • Evaluate with NeMo Evaluator

Features

  • LoRA (Low-Rank Adaptation)
  • Quantization
  • Multi-Device Inference
  • Vocabulary Reduction
  • FP8 KV Cache
  • FP8 Embedding
  • Streaming Output
  • Visual-Token Pruning (DART)
  • KV Cache Reuse

Input & Chat Format

  • Input JSON Format
  • Chat Template Format

Performance

  • Performance Benchmarks

Software Design

  • Checkpoint Exporter Design
  • Experimental ONNX-less Builder Design
  • Quantization Package Design
  • Engine Builder
  • C++ Runtime Overview
  • LLM Inference Runtime
  • LLM Streaming — Design
  • Memory monitoring
  • HTTP Inference Server Architecture

Models

  • Cosmos3-Edge Design
  • Nemotron-3.5-ASR

Customization

  • Customization Guide
  • Calibration Dataset Customization
  • TensorRT Plugins Guide

Testing

  • Unit-Testing the Runtime with a Substitute Engine
  • Code Coverage with SonarQube
  • Few-Layer Numeric Validation

APIs

  • Python API Reference
  • C++ API Reference
    • Action Module
      • Action KV Batch
      • Action Model Types
      • Action Utils
      • Alpamayo1 Action Runner
      • Alpamayo1 Runner Utils
    • Builder Module
      • Action Builder
      • Audio Builder
      • Builder Utils
      • LLM Builder
      • Visual Builder
    • Common Module
      • All Reduce Path
      • Binding Names
      • Check Macros
      • CUDA Macros
      • CUDA Utils
      • File Utils
      • Hash Utils
      • Input Limits
      • Logger
      • Math Utils
      • MMAP Reader
      • Paged KV Types
      • Parallel Artifact Names
      • Rope Utils
      • Safetensors Utils
      • String Utils
      • Tensor
      • TRT Utils
      • Utf8
      • Version
    • Kernels Module
      • Nv Fp4 Mo E Utils
      • Apply Rope Write KV
      • Argmax Kernel
      • Attention Scale Utils
      • Audio Fbank Kernels
      • Batch Evict Kernels
      • Build Layout
      • Causal Conv1d
      • Checkpoint Source Batch
      • Common
      • Conversion
      • Cute Dsl F16 Moe Runner
      • Cute Dsl FMHA Runner
      • Cute Dsl Fmhav2 Runner
      • Cute Dsl FMHA Params
      • Cute Dsl Gdn Runner
      • Cute Dsl GEMM Nv Fp4 Runner
      • Cute Dsl GEMM Runner
      • Cute Dsl Module Loader
      • Cute Dsl Nvfp4 Moe Runner
      • Cute Dsl Nvfp4 Moe Sm110 Runner
      • Cute Dsl Rms Norm Runner
      • Cute Dsl Ssd Runner
      • Cute Dsl Tensor Descriptors
      • Dart Gather Kernels
      • Dart Selector Kernels
      • Ddtree Kernels
      • Decoder XQA Jit Compiler
      • Decoder XQA Runner
      • Dequant
      • Dequantize
      • Dflash Runtime Kernels
      • Dspark Kernels
      • EAGLE Accept Kernels
      • EAGLE Util Kernels
      • Embedding Kernels
      • F16 Moe Support Kernels
      • Fp16 Layout Convert
      • Fp16 Moe Assemble
      • Fp4 Quantize
      • Gdn Kernel Utils
      • Gdn Tree Chunk Kernels
      • Gemma4 Audio Attention
      • Gemma4 Mtp Runtime Kernels
      • Gptq Activation Permutation
      • Gptq Marlin Repack
      • Image Util Kernels
      • Initialize Cos Sin Cache
      • Int4 Cute Dsl Repack
      • Int4 Groupwise GEMM
      • Int4 Moe Scale Repack
      • Int4 Plugin V1 Repack
      • Int4 Qkv Scale Concat
      • Int4 Source Accessors
      • Kernel
      • Kernel Selector
      • KV Cache Utils Kernels
      • Marlin
      • Marlin Repack Index
      • Marlin Dtypes
      • Marlin Mma
      • Marlin Template
      • Moe Activation Kernels
      • Moe Align Sum Kernels
      • Moe Marlin
      • Moe Marlin Indices Kernels
      • Moe Per Expert Scale Kernels
      • Moe Sigmoid Group Topk Kernels
      • Moe Topk Softmax Kernels
      • Mtp State Scatter Kernels
      • Nvfp4 A16 Blackwell GEMM Runner
      • Nvfp4 A16 Blackwell Gemv Jit Compiler
      • Nvfp4 A16 Blackwell Gemv Jit Runner
      • Nvfp4 A16 Blackwell Support
      • Nvfp4 Moe Layout
      • Nvfp4 Moe Normalize
      • Nvfp4 Moe Plugin Geforce Transform
      • Nvfp4 Moe Plugin Transform
      • Nvfp4 Moe Scale Transform
      • Nvfp4 Moe Types
      • Nvfp4 Moe Weight Transform
      • Nvfp4 Dequant
      • Nvfp4 Scale
      • Nvfp4 Tensor
      • Plugin Jit Compile Cache
      • Plugin Jit Compiler
      • Plugin Jit Embedded Sources
      • Scale Convert
      • Selective State Update
      • Speculative Kernels Utils
      • Ssd Varlen Metadata
      • Talker Mlp Kernels
      • Util Kernels
      • Vectorized Types
      • XQA Kernel Types
    • Multimodal Module
      • Audio Runner
      • Audio Utils
      • Clone Encoder Runner
      • Code2 Wav Runner
      • Cosmos3 Edge ViT Runner
      • Gemma4 Audio Runner
      • Gemma4 Unified Audio Runner
      • Gemma4 Unified Vision Runner
      • Gemma4 ViT Runner
      • Image Utils
      • Intern ViT Runner
      • Model Types
      • Multimodal Runner
      • Nemotron Omni Audio Runner
      • Nemotron Omni ViT Runner
      • Phi4mm ViT Runner
      • Qwen25vl ViT Runner
      • Qwen3omni ViT Runner
      • Qwen3vl ViT Runner
      • Qwen ViT Runner
    • Plugins Module
      • All Reduce Plugin
      • Attention Plugin
      • Causal Conv1d Plugin
      • Cute Dsl Int4 GEMM
      • Cute Dsl Int4 Gemv
      • Dflash Target KV Cache Update Plugin
      • Fp16 Moe Plugin
      • Fused Nvfp4 GEMM All Reduce Plugin
      • Gated Delta Net Plugin
      • Gemma4 Audio Attention Plugin
      • Int4 Groupwise GEMM Plugin
      • Int4 Groupwise GEMM Plugin V2
      • Int4 Moe Plugin
      • Mamba Plugin
      • Nvfp4 A16 Blackwell Dispatch Policy
      • Nvfp4 A16 GEMM Plugin
      • Nvfp4 A16 GEMM Plugin V2
      • Nvfp4 A16 Moe Plugin
      • Nvfp4 Moe Plugin
      • Nvfp4 Moe Plugin Geforce
      • Plugin Utils
      • Qkv Concat Plugin
      • Rms Norm Plugin
      • VIT Attention Plugin
    • Profiling Module
      • Layer Profiler
      • Metrics
      • Nvtx Wrapper
      • Timer
    • Runtime Module
      • Audio Loader
      • Audio Utils
      • Block Diffusion Decoder
      • Block Hash
      • Block Index
      • Cache Record
      • Checkpoint Reader
      • Checkpoint Weight Assemble
      • Collective Group
      • Context Cache Config
      • Context Cache Coordinator
      • Context Cache Deployment
      • Context Cache Manager
      • Context Cache Metrics
      • Context Cache Request
      • Context Cache Types
      • Dart Pruner
      • Decoder Registry
      • Decoder Utils
      • Decoding Inference Context
      • Decoding Strategy
      • Deepstack Binding
      • Deployment Config
      • Dflash Decode Utils
      • Dflash Decoder
      • Dspark Decoder
      • EAGLE Decoder
      • Embedding Preprocessor
      • Encoder Embedding Cache
      • Engine Executor
      • Eviction Planner
      • External Weight Manager
      • Gemma4 Embedding Preprocessor
      • Gemma4 Mtp Decoder
      • Hybrid Cache Manager
      • Hybrid Snapshot Storage
      • Image Utils
      • Inference Dims
      • Inference Phase
      • KV Cache Manager
      • KV Page Table
      • Layer Debugger
      • LLM Engine Config
      • LLM Inference Runtime
      • LLM Rank Runtime
      • LLM Runtime Utils
      • Logit Bias
      • Lora Manager
      • Mamba Cache Manager
      • Mel Spectrogram
      • Model Artifacts
      • Mtp Decoder
      • Multi Device Plugin Comm Registry
      • Multi Device Plugin Resource Factory
      • Multi Device Plugin Resources
      • Nccl Collective Backend
      • Parallel Config
      • Pipeline Io
      • Qwen3 Omni Tts Runtime
      • Registry Builder
      • Resource Pools
      • Reuse Plan
      • Rope Cache
      • Runtime Coordinator
      • Shared Resources
      • Spec Common State Tracker
      • Spec Reuse Contract
      • Spec State Plan
      • Step Preparer
      • Streaming
      • System Prompt KV Cache
      • Tensor Map
      • Tensor Parallel Nccl Resources
      • Tensor Parallel Plugin Resources
      • Tensor Registry
      • Vanilla Decoder
      • Visual Token Pruner
    • Sampler Module
      • Diffusion Gemma Sampling
      • Sampling
    • Tokenizer Module
      • Pre Tokenizer
      • Token Encoder
      • Tokenizer
      • Tokenizer Utils
      • Unicode Data

Quick Links

  • Releases
  • GitHub
  • Roadmap
  • Overview: module code

All modules for which code is available

  • experimental.server.config
  • experimental.server.runtime.engine
  • experimental.server.runtime.engine_build
  • tensorrt_edgellm.checkpoint.checkpoint_utils
  • tensorrt_edgellm.checkpoint.loader
  • tensorrt_edgellm.config
  • tensorrt_edgellm.model
  • tensorrt_edgellm.onnx.export
NVIDIA NVIDIA
Privacy Policy | Your Privacy Choices | Terms of Service | Accessibility | Corporate Policies | Product Security | Contact

Copyright © 2025, Nvidia.

Last updated on September 03, 2026.

This page is generated by TensorRT-Edge-LLM commit e8b2952.