TensorRT-LLM
NVIDIA's open-source library for optimizing LLM inference with state-of-the-art performance on NVIDIA GPUs.
When to use TensorRT-LLM
Use TensorRT-LLM when:
- Deploying on NVIDIA GPUs (A100, H100, GB200)
- Need maximum throughput (24,000+ tokens/sec on Llama 3)
- Require low latency for real-time applications
- Working with quantized models (FP8, INT4, FP4)
- Scaling across multiple GPUs or nodes
Use vLLM instead when:
- Need simpler setup and Python-first API
- Want PagedAttention without TensorRT compilation
- Working with AMD GPUs or non-NVIDIA hardware
Use llama.cpp instead when:
- Deploying on CPU or Apple Silicon
- Need edge deployment without NVIDIA GPUs
- Want simpler GGUF quantization format
Quick start
Installation
# Docker (recommended)
docker pull nvidia/tensorrt_llm:latest
# pip install
pip install tensorrt_llm==1.2.0rc3
# Requires CUDA 13.0.0, TensorRT 10.13.2, Python 3.10-3.12
Basic inference
from tensorrt_llm import LLM, SamplingParams
# Initialize model
llm = LLM(model="meta-llama/Meta-Llama-3-8B")
# Configure sampling
sampling_params = SamplingParams(
max_tokens=100,
temperature=0.7,
top_p=0.9
)
# Generate
prompts = ["Explain quantum computing"]
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
print(output.text)
Serving with trtllm-serve
# Start server (automatic model download and compilation)
trtllm-serve meta-llama/Meta-Llama-3-8B \
--tp_size 4 \ # Tensor parallelism (4 GPUs)
--max_batch_size 256 \
--max_num_tokens 4096
# Client request
curl -X POST http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "meta-llama/Meta-Llama-3-8B",
"messages": [{"role": "user", "content": "Hello!"}],
"temperature": 0.7,
"max_tokens": 100
}'
Key features
Performance optimizations
- In-flight batching: Dynamic batching during generation
- Paged KV cache: Efficient memory management
- Flash Attention: Optimized attention kernels
- Quantization: FP8, INT4, FP4 for 2-4× faster inference
- CUDA graphs: Reduced kernel launch overhead
Parallelism
- Tensor parallelism (TP): Split model across GPUs
- Pipeline parallelism (PP): Layer-wise distribution
- Expert parallelism: For Mixture-of-Experts models
- Multi-node: Scale beyond single machine
Advanced features
- Speculative decoding: Faster generation with draft models
- LoRA serving: Efficient multi-adapter deployment
- Disaggregated serving: Separate prefill and generation
Common patterns
Quantized model (FP8)
from tensorrt_llm import LLM
# Load FP8 quantized model (2× faster, 50% memory)
llm = LLM(
model="meta-llama/Meta-Llama-3-70B",
dtype="fp8",
max_num_tokens=8192
)
# Inference same as before
outputs = llm.generate(["Summarize this article..."])
Multi-GPU deployment
# Tensor parallelism across 8 GPUs
llm = LLM(
model="meta-llama/Meta-Llama-3-405B",
tensor_parallel_size=8,
dtype="fp8"
)
Batch inference
# Process 100 prompts efficiently
prompts = [f"Question {i}: ..." for i in range(100)]
outputs = llm.generate(
prompts,
sampling_params=SamplingParams(max_tokens=200)
)
# Automatic in-flight batching for maximum throughput
Performance benchmarks
Meta Llama 3-8B (H100 GPU):
- Throughput: 24,000 tokens/sec
- Latency: ~10ms per token
- vs PyTorch: 100× faster
Llama 3-70B (8× A100 80GB):
- FP8 quantization: 2× faster than FP16
- Memory: 50% reduction with FP8
Supported models
- LLaMA family: Llama 2, Llama 3, CodeLlama
- GPT family: GPT-2, GPT-J, GPT-NeoX
- Qwen: Qwen, Qwen2, QwQ
- DeepSeek: DeepSeek-V2, DeepSeek-V3
- Mixtral: Mixtral-8x7B, Mixtral-8x22B
- Vision: LLaVA, Phi-3-vision
- 100+ models on HuggingFace
References
- Optimization Guide - Quantization, batching, KV cache tuning
- Multi-GPU Setup - Tensor/pipeline parallelism, multi-node
- Serving Guide - Production deployment, monitoring, autoscaling
Resources
1---2name: tensorrt-llm3description: 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.4category: ml-inference5version: 1.0.06author: Orchestra Research7license: MIT8tags: [Inference Serving, TensorRT-LLM, NVIDIA, Inference Optimization, High Throughput, Low Latency, Production, FP8, INT4, In-Flight Batching, Multi-GPU]9dependencies: [tensorrt-llm, torch]10metadata:11 source: https://skillmd.com/skills/synthetic-sciences/tensorrt-llm12 upstream: Orchestra-Research/AI-Research-SKILLs13 upstream-url: https://github.com/Orchestra-Research/AI-Research-SKILLs14 upstream-path: 12-inference-serving/tensorrt-llm15 upstream-license: MIT16 upstream-relationship: derived17 skill-author: Orchestra Research18 adapted-by: Synthetic Sciences19---2021# TensorRT-LLM2223NVIDIA's open-source library for optimizing LLM inference with state-of-the-art performance on NVIDIA GPUs.2425## When to use TensorRT-LLM2627**Use TensorRT-LLM when:**28- Deploying on NVIDIA GPUs (A100, H100, GB200)29- Need maximum throughput (24,000+ tokens/sec on Llama 3)30- Require low latency for real-time applications31- Working with quantized models (FP8, INT4, FP4)32- Scaling across multiple GPUs or nodes3334**Use vLLM instead when:**35- Need simpler setup and Python-first API36- Want PagedAttention without TensorRT compilation37- Working with AMD GPUs or non-NVIDIA hardware3839**Use llama.cpp instead when:**40- Deploying on CPU or Apple Silicon41- Need edge deployment without NVIDIA GPUs42- Want simpler GGUF quantization format4344## Quick start4546### Installation4748```bash49# Docker (recommended)50docker pull nvidia/tensorrt_llm:latest5152# pip install53pip install tensorrt_llm==1.2.0rc35455# Requires CUDA 13.0.0, TensorRT 10.13.2, Python 3.10-3.1256```5758### Basic inference5960```python61from tensorrt_llm import LLM, SamplingParams6263# Initialize model64llm = LLM(model="meta-llama/Meta-Llama-3-8B")6566# Configure sampling67sampling_params = SamplingParams(68 max_tokens=100,69 temperature=0.7,70 top_p=0.971)7273# Generate74prompts = ["Explain quantum computing"]75outputs = llm.generate(prompts, sampling_params)7677for output in outputs:78 print(output.text)79```8081### Serving with trtllm-serve8283```bash84# Start server (automatic model download and compilation)85trtllm-serve meta-llama/Meta-Llama-3-8B \86 --tp_size 4 \ # Tensor parallelism (4 GPUs)87 --max_batch_size 256 \88 --max_num_tokens 40968990# Client request91curl -X POST http://localhost:8000/v1/chat/completions \92 -H "Content-Type: application/json" \93 -d '{94 "model": "meta-llama/Meta-Llama-3-8B",95 "messages": [{"role": "user", "content": "Hello!"}],96 "temperature": 0.7,97 "max_tokens": 10098 }'99```100101## Key features102103### Performance optimizations104- **In-flight batching**: Dynamic batching during generation105- **Paged KV cache**: Efficient memory management106- **Flash Attention**: Optimized attention kernels107- **Quantization**: FP8, INT4, FP4 for 2-4× faster inference108- **CUDA graphs**: Reduced kernel launch overhead109110### Parallelism111- **Tensor parallelism (TP)**: Split model across GPUs112- **Pipeline parallelism (PP)**: Layer-wise distribution113- **Expert parallelism**: For Mixture-of-Experts models114- **Multi-node**: Scale beyond single machine115116### Advanced features117- **Speculative decoding**: Faster generation with draft models118- **LoRA serving**: Efficient multi-adapter deployment119- **Disaggregated serving**: Separate prefill and generation120121## Common patterns122123### Quantized model (FP8)124125```python126from tensorrt_llm import LLM127128# Load FP8 quantized model (2× faster, 50% memory)129llm = LLM(130 model="meta-llama/Meta-Llama-3-70B",131 dtype="fp8",132 max_num_tokens=8192133)134135# Inference same as before136outputs = llm.generate(["Summarize this article..."])137```138139### Multi-GPU deployment140141```python142# Tensor parallelism across 8 GPUs143llm = LLM(144 model="meta-llama/Meta-Llama-3-405B",145 tensor_parallel_size=8,146 dtype="fp8"147)148```149150### Batch inference151152```python153# Process 100 prompts efficiently154prompts = [f"Question {i}: ..." for i in range(100)]155156outputs = llm.generate(157 prompts,158 sampling_params=SamplingParams(max_tokens=200)159)160161# Automatic in-flight batching for maximum throughput162```163164## Performance benchmarks165166**Meta Llama 3-8B** (H100 GPU):167- Throughput: 24,000 tokens/sec168- Latency: ~10ms per token169- vs PyTorch: **100× faster**170171**Llama 3-70B** (8× A100 80GB):172- FP8 quantization: 2× faster than FP16173- Memory: 50% reduction with FP8174175## Supported models176177- **LLaMA family**: Llama 2, Llama 3, CodeLlama178- **GPT family**: GPT-2, GPT-J, GPT-NeoX179- **Qwen**: Qwen, Qwen2, QwQ180- **DeepSeek**: DeepSeek-V2, DeepSeek-V3181- **Mixtral**: Mixtral-8x7B, Mixtral-8x22B182- **Vision**: LLaVA, Phi-3-vision183- **100+ models** on HuggingFace184185## References186187- **[Optimization Guide](references/optimization.md)** - Quantization, batching, KV cache tuning188- **[Multi-GPU Setup](references/multi-gpu.md)** - Tensor/pipeline parallelism, multi-node189- **[Serving Guide](references/serving.md)** - Production deployment, monitoring, autoscaling190191## Resources192193- **Docs**: https://nvidia.github.io/TensorRT-LLM/194- **GitHub**: https://github.com/NVIDIA/TensorRT-LLM195- **Models**: https://huggingface.co/models?library=tensorrt_llm196197