torch-continuum

Accelerate any PyTorch workload in one line.

Quickstart • Benchmarks • Usage • API • Citation

---

Why torch-continuum?

Most PyTorch users leave 40-60% performance on the table because they don't know the right combination of hardware-specific settings for their GPU. torch-continuum detects your hardware and applies the optimal configuration — automatically.

import torch_continuum
torch_continuum.optimize()
# That's it. Your code runs faster.

No code changes. No config files. No guesswork.

---

Quickstart

Install

pip install torch-continuum

Use

import torch
import torch.nn as nn
import torch_continuum

# One line — before your training loop
torch_continuum.optimize("fast")

# Your existing code — unchanged
model = nn.TransformerEncoder(...)
optimizer = torch.optim.AdamW(model.parameters())

for batch in dataloader:
    loss = model(batch).sum()
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

---

Benchmarks

Measured on NVIDIA H100 80GB with PyTorch 2.10. Real training loop with AdamW optimizer, CrossEntropyLoss, 5 independent trials, 200 iterations each.

GPT-style Decoder 6 layers · d=768 · 8 heads · vocab 32K Time Speedup PyTorch 9.622s — torch-continuum 3.912s +59.3%

CNN (ResNet-style) 5-layer ConvNet · 224×224 · batch 64 Time Speedup PyTorch 3.173s — torch-continuum 1.539s +51.5%

Dense Linear Stack 67M params · 4096→4096 ×4 · batch 256 Time Speedup PyTorch 0.900s — torch-continuum 0.554s +38.4%

Standard deviations: 0.001–0.004s across 5 trials. Fully reproducible.

---

Three Optimization Levels

Choose your tradeoff:

import torch_continuum

torch_continuum.optimize("safe")    # No precision change — pure speed
torch_continuum.optimize("fast")    # ~2x matmul throughput (recommended)
torch_continuum.optimize("max")     # Maximum — fused kernels + compilation

Level	Precision Impact	Speedup	Best For
safe	None	Moderate	Any workload — zero risk
fast	Minor	Up to 60%	Training & inference
max	Mixed precision	Maximum	LLM training, large transformers

---

Usage

Training a Model

import torch
import torch.nn as nn
import torch_continuum

torch_continuum.optimize("fast")

model = nn.Sequential(
    nn.Linear(4096, 4096),
    nn.GELU(),
    nn.Linear(4096, 4096),
).cuda()

x = torch.randn(256, 4096, device="cuda")
optimizer = torch.optim.AdamW(model.parameters())

for step in range(1000):
    loss = model(x).sum()
    loss.backward()
    optimizer.step()
    optimizer.zero_grad()

Smart Model Compilation

import torch_continuum

# Auto-selects the best compilation strategy for your model size
model = torch_continuum.smart_compile(model)

Benchmark Your Own Model

import torch_continuum

# See exactly how much faster your model runs
torch_continuum.benchmark(model, example_input, level="fast")

Output:

[torch-continuum benchmark]
  model params : 67,125,248
  input shape  : (256, 4096)
  device       : cuda:0
  iters        : 100  (warmup 10)

  Config                     Time    Speedup   Peak Mem
  ────────────────────── ──────── ────────── ──────────
  PyTorch baseline         0.261s        ---       983 MB
  torch-continuum          0.100s +   61.6%      1520 MB

LLM Training with Fused Kernels

pip install torch-continuum[liger]

import torch_continuum

# Patches HuggingFace models with fused GPU kernels
# +20% throughput, -60% memory on LLM training
torch_continuum.optimize("max")
torch_continuum.apply_liger_kernels()

Supports: LLaMA, Mistral, Gemma, Qwen2, Phi-3

Hardware Detection

import torch_continuum

info = torch_continuum.detect_device()
print(info.summary())

  device : NVIDIA H100 80GB HBM3 (cuda)
  memory : 85.0 GB  |  GPUs: 1

---

Hardware Support

Platform	Status
NVIDIA Ampere (A100, RTX 30xx)	Full acceleration
NVIDIA Hopper (H100, H200)	Full acceleration
NVIDIA Ada (RTX 40xx, L40)	Full acceleration
Apple Silicon (M1/M2/M3/M4)	Supported
CPU	Supported

---

API Reference

Function	Description
optimize(level)	Apply hardware-tuned optimizations ("safe", "fast", "max")
status()	Print current optimization state
smart_compile(model)	Compile with auto-tuned settings
benchmark(model, input)	Measure speedup on your model
detect_device()	Get hardware capability profile
apply_liger_kernels()	Enable fused kernels for LLM training
liger_available()	Check if fused kernel backend is installed

---

Citation

If you use torch-continuum in your research, please cite:

@software{torch_continuum,
  title     = {torch-continuum: Hardware-Aware PyTorch Acceleration},
  author    = {Badaramoni, Avinash},
  year      = {2026},
  url       = {https://github.com/badaramoni/torch-continuum},
  version   = {0.2.0}
}

---

Contributing

We welcome contributions! Please open an issue or pull request on GitHub.

License

MIT — see LICENSE for details.

---

Made with speed in mind.
PyPI • GitHub