Give your PyTorch models superpowers ⚡

 

Source-to-source compiler for PyTorch. Understandable. Inspectable. Extensible.

✅ Run PyTorch 40% faster   ✅ Quantization                ✅ Kernel fusion        
✅ Training recipes         ✅ FP4/FP6/FP8 precision       ✅ Distributed TP/PP/DP 
✅ Inference recipes        ✅ Ready for NVIDIA Blackwell  ✅ CUDA Graphs          
✅ LLMs, non LLMs and more  ✅ Custom Triton kernels       ✅ Compose all the above

Thunder is a source-to-source deep learning compiler for PyTorch that focuses on making it simple to optimize models for training and inference.

It provides:

• a simple, Pythonic IR capturing the entire computation
• a rich system of transforms that simultaneously operate on the computation IR, the model, and the weights
• an extensible dispatch mechanism to fusers and optimized kernel libraries

With Thunder you can:

• profile deep learning programs easily, map individual ops to kernels and inspect programs interactively
• programmatically replace sequences of operations with optimized ones and see the effect on performance
• acquire full computation graphs without graph breaks by flexibly extending the interpreter
• modify programs to fully utilize bleeding edge kernel libraries on specific hardware
• write models for single GPU and transform them to run distributed
• quickly iterate on mixed precision and quantization strategies to search for combinations that minimally affect quality
• bundle all optimizations in composable recipes, so they can be ported across model families

Ultimately, you should think about Thunder as a highly efficient tool to go from “unoptimized” to “optimized”.

If that is of interest for you, read on to Install Thunder and get started quickly.

Quick start • Examples • Performance • Docs

 

 

Quick start

Install Thunder via pip (more options):

pip install lightning-thunder

pip install -U torch torchvision
pip install nvfuser-cu128-torch28 nvidia-cudnn-frontend  # if NVIDIA GPU is present

For older versions of torch

torch==2.7 + CUDA 12.8

pip install lightning-thunder

pip install torch==2.7.0 torchvision==0.22
pip install nvfuser-cu128-torch27 nvidia-cudnn-frontend  # if NVIDIA GPU is present

torch==2.6 + CUDA 12.6

pip install lightning-thunder

pip install torch==2.6.0 torchvision==0.21
pip install nvfuser-cu126-torch26 nvidia-cudnn-frontend  # if NVIDIA GPU is present

torch==2.5 + CUDA 12.4

pip install lightning-thunder

pip install torch==2.5.0 torchvision==0.20
pip install nvfuser-cu124-torch25 nvidia-cudnn-frontend  # if NVIDIA GPU is present

Advanced install options

Install optional executors

# Float8 support (this will compile from source, be patient)
pip install "transformer_engine[pytorch]"

Install Thunder bleeding edge

pip install git+https://github.com/Lightning-AI/lightning-thunder.git@main

Install Thunder for development

git clone https://github.com/Lightning-AI/lightning-thunder.git
cd lightning-thunder
pip install -e .

Hello world

Define a function or a torch module:

import torch.nn as nn

model = nn.Sequential(nn.Linear(2048, 4096), nn.ReLU(), nn.Linear(4096, 64))

Optimize it with Thunder:

import thunder
import torch

thunder_model = thunder.compile(model)

x = torch.randn(64, 2048)

y = thunder_model(x)

torch.testing.assert_close(y, model(x))

Examples

LLM training

Install LitGPT (without updating other dependencies)

pip install --no-deps 'litgpt[all]'

and run

import thunder
import torch
import litgpt

with torch.device("cuda"):
    model = litgpt.GPT.from_name("Llama-3.2-1B").to(torch.bfloat16)

thunder_model = thunder.compile(model)

inp = torch.ones((1, 2048), device="cuda", dtype=torch.int64)

out = thunder_model(inp)
out.sum().backward()

HuggingFace BERT inference

Install Hugging Face Transformers (recommended version is 4.50.2 and above)

pip install -U transformers

and run

import thunder
import torch
import transformers

model_name = "bert-large-uncased"

tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)

with torch.device("cuda"):
    model = transformers.AutoModelForCausalLM.from_pretrained(
        model_name, torch_dtype=torch.bfloat16
    )
    model.requires_grad_(False)
    model.eval()

    inp = tokenizer(["Hello world!"], return_tensors="pt")

thunder_model = thunder.compile(model)

out = thunder_model(**inp)
print(out)

HuggingFace DeepSeek R1 distill inference

Install Hugging Face Transformers (recommended version is 4.50.2 and above)

pip install -U transformers

and run

import torch
import transformers
import thunder

model_name = "deepseek-ai/DeepSeek-R1-Distill-Llama-8B"

tokenizer = transformers.AutoTokenizer.from_pretrained(model_name)

with torch.device("cuda"):
    model = transformers.AutoModelForCausalLM.from_pretrained(
        model_name, torch_dtype=torch.bfloat16
    )
    model.requires_grad_(False)
    model.eval()

    inp = tokenizer(["Hello world! Here's a long story"], return_tensors="pt")

thunder_model = thunder.compile(model)

out = thunder_model.generate(
    **inp, do_sample=False, cache_implementation="static", max_new_tokens=100
)
print(out)

Vision Transformer inference

import thunder
import torch
import torchvision as tv

with torch.device("cuda"):
    model = tv.models.vit_b_16()
    model.requires_grad_(False)
    model.eval()

    inp = torch.randn(128, 3, 224, 224)

out = model(inp)

thunder_model = thunder.compile(model)

out = thunder_model(inp)

Benchmarks

Although is Thunder a tool for optimizing models, rather than an opaque compiler that gets you speedups out of the box, here is a set of benchmarks.

Perf-wise, out of the box Thunder is in the ballpark of torch compile, especially when using CUDAGraphs. Note however that Thunder is not a competitor to torch compile! It can actually use torch compile as one of its fusion executors.

The script examples/quickstart/hf_llm.py demonstrates how to benchmark a model for text generation, forward pass, forward pass with loss, and a full forward + backward computation.

On an H100 with torch=2.8.0 and nvfuser-cu128-torch28 and Transformers 4.55.4 running Llama 3.2 1B we see the following timings:

Transformers with torch.compile and CUDAGraphs (reduce-overhead mode):  521ms
Transformers with torch.compile but no CUDAGraphs (default mode):       814ms
Transformers without torch.compile:                                    1493ms
Thunder with CUDAGraphs:                                                542ms

Plugins

Plugins are a way to apply optimizations to a model, such as parallelism and quantization.

Thunder comes with a few plugins included of the box, but it's easy to write new ones.

• scale up with distributed strategies with DDP, FSDP, TP ()
• optimize numerical precision with FP8, MXFP8
• save memory with quantization
• reduce latency with CUDAGraphs
• debugging and profiling

For example, in order to reduce CPU overheads via CUDAGraphs you can add "reduce-overhead" to the plugins= argument of thunder.compile:

thunder_model = thunder.compile(model, plugins="reduce-overhead")

This may or may not make a big difference. The point of Thunder is that you can easily swap optimizations in and out and explore the best combination for your setup.

How it works

Thunder works in three stages:

1. ⚡️ It acquires your model by interpreting Python bytecode and producing a straight-line Python program

2. ️⚡️ It transforms the model and computation trace to make it distributed, change precision

3. ⚡️ It routes parts of the trace for execution

   • fusion (NVFuser, torch.compile)
   • specialized libraries (e.g. cuDNN SDPA, TransformerEngine)
   • custom Triton and CUDA kernels
   • PyTorch eager operations

 

 

This is how the trace looks like for a simple MLP:

import thunder
import torch
import torch.nn as nn

model = nn.Sequential(nn.Linear(1024, 2048), nn.ReLU(), nn.Linear(2048, 256))

thunder_model = thunder.compile(model)
y = thunder_model(torch.randn(4, 1024))

print(thunder.last_traces(thunder_model)[-1])

This is the acquired trace, ready to be transformed and executed:

def computation(input, t_0_bias, t_0_weight, t_2_bias, t_2_weight):
# input: "cuda:0 f32[4, 1024]"
# t_0_bias: "cuda:0 f32[2048]"
# t_0_weight: "cuda:0 f32[2048, 1024]"
# t_2_bias: "cuda:0 f32[256]"
# t_2_weight: "cuda:0 f32[256, 2048]"
t3 = ltorch.linear(input, t_0_weight, t_0_bias) # t3: "cuda:0 f32[4, 2048]"
t6 = ltorch.relu(t3, False) # t6: "cuda:0 f32[4, 2048]"
t10 = ltorch.linear(t6, t_2_weight, t_2_bias) # t10: "cuda:0 f32[4, 256]"
return (t10,)

Note how Thunder's intermediate representation is just (a subset of) Python!

Performance

Thunder is fast. Here are the speed-ups obtained on a pre-training task using LitGPT on H100 and B200 hardware, relative to PyTorch eager.

Community

Thunder is an open source project, developed in collaboration with the community with significant contributions from NVIDIA.

💬 Get help on Discord 📋 License: Apache 2.0