Installation

Use Linux with an NVIDIA GPU supported by FlashAttention 2, such as an A100, H100, RTX 3090, or RTX 4090. Python 3.12 is recommended. All four model workflows require CUDA-enabled PyTorch and FlashAttention.

Set up GPU dependencies

In your Python environment, install PyTorch before FlashAttention. This setup uses PyTorch 2.10.0 with CUDA 12.8 and FlashAttention 2.8.3. Skip these steps if you already have these versions working.

python -m pip install --upgrade pip setuptools wheel packaging ninja psutil
python -m pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cu128
python -m pip install flash-attn==2.8.3 --no-build-isolation

The CUDA wheel must be compatible with your NVIDIA driver; see the official PyTorch installation options. Install FlashAttention after selecting PyTorch, because its compiled extension must match that environment. --no-build-isolation lets its installer use the PyTorch you just installed.

FlashAttention attempts to use a matching prebuilt wheel when available. Plan for a source build with the PyTorch 2.10 / CUDA 12.8 combination above, unless a compatible cached or separately supplied wheel is available. The upstream release includes PyTorch 2.10 wheels labelled cu13; those target CUDA 13 and do not match this CUDA 12.8 setup. Compilation can take a long time. See the official wheel list and the build guidance below.

Install PoreFM

git clone https://github.com/WGLab/PoreFM.git
cd PoreFM
python -m pip install .

Pip installs the remaining Python dependencies, including those used to read, summarize, and plot outputs. For an editable installation, use python -m pip install -e . instead.

Input preparation also uses the external samtools command. Install it through your system or cluster package manager and check that samtools --version works. Model files are supplied separately.

Check the installation

Run this on a machine with an allocated GPU:

porefm --help
python - <<'PY'
import torch
import flash_attn

print("PyTorch:", torch.__version__, "CUDA:", torch.version.cuda)
print("FlashAttention:", flash_attn.__version__)
assert torch.cuda.is_available(), "No CUDA GPU is visible to PyTorch"
print("GPU:", torch.cuda.get_device_name(0))
PY

If FlashAttention needs to compile

Source builds require a CUDA toolkit with nvcc and a C++ compiler. Use a toolkit matching the CUDA version of PyTorch (12.8 for the command above). The CUDA runtime in a PyTorch wheel does not provide the nvcc compiler. On a cluster, load the appropriate CUDA module for your system before building.

To limit build memory use, cap the number of compiler jobs:

MAX_JOBS=4 python -m pip install flash-attn==2.8.3 --no-build-isolation

Leave FLASH_ATTENTION_FORCE_BUILD unset for normal installation: setting it to TRUE forces compilation even if a matching wheel exists. Keep pip's cache enabled so subsequent installs can reuse a built wheel. A prebuilt wheel must match Python, PyTorch, CUDA, the platform, and the C++ ABI; a wheel for another PyTorch version is not interchangeable. See the FlashAttention installation guide.

Continue with the HeLa RNA example for a first result using downloadable data, or prepare your own inputs.