tinytorchprofiler 0.1.0

A lightweight PyTorch model profiler for latency, parameters, memory, and layer-wise bottlenecks.

TinyTorchProfiler

TinyTorchProfiler is a lightweight Python library for profiling PyTorch models.

It helps users understand model performance beyond accuracy, including latency, parameter count, model size, activation memory, layer-wise runtime, and likely bottlenecks.

The goal of v0.1.0 is to stay small, readable, and practical for ML systems workflows where model performance matters as much as model quality.

Why This Library Matters

A model with good accuracy can still be too slow, too large, or too memory-heavy for production use.

TinyTorchProfiler helps answer practical deployment questions:

• How many parameters does this model have?
• How many parameters are trainable?
• How large is the model in memory?
• What is the average forward-pass latency?
• Which layers are the slowest?
• Which layers create the largest activations?
• Does this model look suitable for deployment constraints?

Installation

From the project root, install the package in editable mode:

pip install -e .

For development, install the optional test dependencies:

pip install -e ".[dev]"

You can also install dependencies directly:

pip install -r requirements.txt

Quickstart

from torch import nn

from tinytorchprofiler import profile_model


model = nn.Sequential(
    nn.Linear(128, 64),
    nn.ReLU(),
    nn.Linear(64, 10),
)

report = profile_model(
    model,
    input_shape=(1, 128),
    device="cpu",
    warmup=5,
    runs=20,
)

report.summary()
report.to_csv("profile.csv")

Core API

from tinytorchprofiler import profile_model

report = profile_model(
    model,
    input_shape=(1, 3, 224, 224),
    device="cpu",
    warmup=5,
    runs=20,
)

report.summary()
report.to_dict()
report.to_csv("profile.csv")

Supported Metrics

TinyTorchProfiler v0.1.0 supports:

• Total parameter count
• Trainable parameter count
• Estimated model size in MB
• Average forward-pass latency in milliseconds
• CPU profiling
• CUDA profiling when available
• Layer-wise names
• Layer-wise module types
• Layer-wise input and output shapes
• Layer-wise parameter counts
• Approximate activation size in MB
• Approximate layer-wise forward latency

Device Support

CPU profiling is supported by default:

report = profile_model(model, input_shape=(1, 3, 224, 224), device="cpu")

CUDA profiling is supported when PyTorch detects an available CUDA device:

report = profile_model(model, input_shape=(1, 3, 224, 224), device="cuda")

For CUDA timing, TinyTorchProfiler synchronizes the device around measured regions with torch.cuda.synchronize().

Examples

Profile a small CNN:

python examples/01_profile_simple_cnn.py

This prints a summary and writes:

simple_cnn_profile.csv

Profile a torchvision model if torchvision is installed:

python examples/02_profile_torchvision_model.py

If torchvision is missing, the example exits gracefully with a short message.

Testing

Run the test suite with:

pytest

The v0.1.0 tests cover:

• Parameter counting
• Trainable parameter counting
• Model size estimation
• Tensor size estimation
• Report serialization with to_dict()
• Basic end-to-end profiling on a tiny model

Notes on Profiling

TinyTorchProfiler uses PyTorch forward passes and hooks.

Layer-wise latency is approximate. It is useful for identifying likely bottlenecks, but exact timing can vary depending on hardware, backend libraries, CPU load, CUDA synchronization, and model structure.

Activation memory is estimated from forward outputs. It should be treated as a helpful approximation, not as a replacement for full memory tracing.

Roadmap

Planned future improvements:

• Batch-size scaling analysis
• Memory peak tracking
• Visualization
• ONNX export profiling
• ViT and DINOv2 examples

Version

Current version: 0.1.0