Quick Start

This quick start provides examples of tuning and deployment paths available in NVIDIA AITune.

NVIDIA AITune enables seamless tuning of models for deployment (for example, converting them to TensorRT) without requiring changes to your original Python pipelines.

NVIDIA AITune supports two modes:

• Ahead-of-time tuning — provide a model or a pipeline, and a dataset/dataloader. You can either rely on inspect to detect promising modules to tune or manually select them.
• Just-in-time tuning — set a special environment variable, run your script without changes, and AITune will, on the fly, detect modules and tune them one by one.

Ahead-of-time mode is more powerful and allows you to tweak more settings, whereas just-in-time works out of the box but offers less control over the tuning process. For a more detailed comparison, see the Comparison between AOT and JIT tuning section.

Enabling logging

The tuning process guides the user through decisions and steps that are performed to tune every selected module.

We recommend to enable the INFO logging level for better verbosity in the quick start steps:

1
import logging
2
3
logging.basicConfig(level=logging.INFO, force=True)

Learn about more options in observability.

Ahead-of-time tuning

The code below demonstrates Stable Diffusion pipeline tuning.

You can annotate torch.nn.Modules manually or use the inspect functionality to have modules picked automatically; you can then verify them and schedule them for tuning.

First, install the required third-party dependencies:

$
pip install transformers diffusers torch

Then initialize the pipeline:

1
import torch
2
from diffusers import DiffusionPipeline
3
4
import aitune.torch as ait
5
6
# Initialize pipeline
7
pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16)
8
pipe.to("cuda")

Next, inspect the pipeline components and display the summary:

1
# Prepare input data
2
input_data = [{"prompt": "A beautiful landscape with mountains and a lake"}]
3
4
# Inspect pipeline to get modules
5
modules_info = ait.inspect(pipe, input_data)
6
7
8
# Optional: inference function, if you need more control over execution
9
def infer(prompt):
10
    return pipe(prompt, width=1024, height=1024, num_inference_steps=10)
11
12
# modules_info = ait.inspect(pipe, input_data, inference_function=infer)
13
14
# Display modules info
15
modules_info.describe()

Finally, wrap the selected modules and tune within the pipeline:

1
# Wrap modules for tuning
2
modules = modules_info.get_modules()
3
pipe = ait.wrap(pipe, modules)
4
5
# Tune pipeline
6
ait.tune(pipe, input_data)

At this point, you can use the pipeline to generate predictions with the tuned models directly in Python:

1
# Run inference on tuned pipeline
2
images = pipe(["A beautiful landscape with mountains and a lake"])
3
image = images[0][0]
4
5
# Save image for preview
6
image.save("landscape.png")

Once the pipeline has been tuned, you can save the best-performing version of the modules for later deployment:

1
ait.save(pipe, "tuned_pipe.ait")

And load the tuned pipeline directly:

1
pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16)
2
pipe.to("cuda")
3
ait.load(pipe, "tuned_pipe.ait")

Just-in-time tuning

In this mode, there is no need to modify the user’s code. AITune records inference calls until jit_config.min_samples are collected, then tries to tune modules one by one starting from the top. If there is one of the following conditions:

• a graph break is detected, i.e., torch.nn.Module contains conditional logic on inputs, meaning there is no guarantee of a static, correct graph of computations, or
• there is an error during tuning

that module is left unchanged and AITune tries to tune its children. This process continues until the module depth reaches a configured limit.

First, install the required third-party dependencies:

$
pip install transformers diffusers torch

Prepare the script with the model for tuning my_script.py:

1
# Enable JIT tuning - single import
2
import aitune.torch.jit.enable
3
4
from diffusers import DiffusionPipeline
5
6
# Initialize pipeline
7
pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16)
8
pipe.to("cuda")
9
10
# First call - tuning the model
11
pipe("A beautiful landscape with mountains and a lake")
12
13
# Second call - using tuned model
14
pipe("A beautiful landscape with mountains and a lake")

You can then run your script:

$
python my_script.py

Note: The import aitune.torch.jit.enable must be a first import in your code. The alternative option is to use export AUTOWRAPT_BOOTSTRAP=aitune_enable_jit_tuning to avoid any source code modification.

Configuring just-in-time tuning

If there is a need to adjust just-in-time options, you can do it but currently this requires modifying code to import the JIT config:

1
from aitune.torch import jit_config
2
from aitune.torch.backend import TensorRTBackend
3
from aitune.torch.tune_strategy import FirstWinsStrategy
4
5
jit_config.max_depth_level = 2 # change the default maximum depth level for nested modules to be tuned
6
jit_config.detect_graph_breaks = False # turn off graph break detection
7
jit_config.strategy = FirstWinsStrategy(backends=[TensorRTBackend()]) # change the tune strategy

Deferred mode

By default JIT tuning uses eager mode — each module is tuned automatically as soon as enough samples have been collected. For pipelines where different modules are called a variable number of times per step (e.g. text-to-image or text-to-video diffusion models), use deferred mode instead. In deferred mode, AITune records samples until you mark a safe synchronization point; tuning then starts on the next normal forward pass.

1
import aitune.torch.jit.enable
2
3
from aitune.torch import jit_config
4
from aitune.torch.jit.config import JITMode
5
from aitune.torch.jit.tune import deferred as jit_deferred
6
from diffusers import DiffusionPipeline
7
8
jit_config.mode = JITMode.TUNE_DEFERRED
9
10
pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-3-medium-diffusers")
11
pipe.to("cuda")
12
13
# First full step — records samples for all modules
14
pipe("A beautiful landscape")
15
16
# Mark that deferred tuning may start
17
jit_deferred()
18
19
# This call triggers tuning from the normal pipeline flow
20
pipe("A snowy mountain at sunset")
21
22
# Subsequent calls run on the tuned pipeline
23
pipe("A snowy mountain at sunset")

See JIT Tuning Guide for a detailed comparison of the two modes.

Comparison between ahead-of-time and just-in-time tuning

The ahead-of-time tuning gives you the most control over the tuning process:

• it detects the batch axis and dynamic axes (axes that change shape independently of batch size, e.g., sequence length in LLMs)
• allows picking modules to tune
• you can pick a tuning strategy (e.g., best throughput) for the whole process or per-module
• you can pick tuning backends (e.g., TensorRT, TorchInductor, TorchAO, ONNXRuntime) which will be used by the strategy
• you can mix different backends in the same model/pipeline
• you can manually verify the tuning process (note: AITune performs basic checks for NaNs and errors)
• you can save the resulting artifact and later read it from disk

The big advantage of just-in-time tuning is that you don’t need to modify the user’s script to tune a model. However, it has some disadvantages - since it cannot access data directly (you don’t provide a dataloader):

• it cannot deduce batch size nor do benchmarking
• input/output shapes depend on the data seen, so for example, TRT backend will build a profile only for that data
• it needs at least one inference call to record inputs before tuning; later calls use tuned modules where tuning succeeded
• if you need dynamic axes (e.g., TRT backend), you need to provide two different batch sizes
• benchmarking-based strategies are limited because JIT cannot extrapolate to controlled batch sizes
• you can specify a global tune strategy for the whole model

The following table summarizes the difference between modes:

Note: JIT mode writes build artifacts and logs under jit_config.cache_dir / AITUNE_JIT_CACHE_DIR, but it does not reuse them as tuned checkpoints across Python interpreter runs. Every new process starts tuning from scratch.