Requirements

The Gradio SDK base image already installs these on every hardware tier — listing them in requirements.txt causes resolution failures or, worse, lets pip silently drift the runtime out of compatibility:

What to list

Everything you actually import, including the often-forgotten:

• torchvision, torchaudio — not preinstalled. Leave unpinned; pip resolves against the installed torch major.minor.
• accelerate — needed whenever you use device_map=. Listing it also silences low_cpu_mem_usage=False warnings.
• sentencepiece — required by most LLM tokenizers; rarely transitive.
• einops — required by flash_attn.layers.rotary and many model repos.
• Domain libs: diffusers, transformers, safetensors, pillow, numpy, etc.

If a research repo ships a Python package directory (models/, pipeline/, …), just upload the directory with the rest of the Space — the whole repo root is importable as /home/user/app. Do not try to reference local paths from requirements.txt.

Pinning torch

ZeroGPU accepts only 2.8.0, 2.9.1, 2.10.0, 2.11.0. Default is unpinned (runtime preinstalls the latest). Pinning is fine — and sometimes warranted — within that set:

• A specific torch is known-good for your model (numerics, attention kernel availability, etc.).
• A direct-URL CUDA wheel encodes a torch2.X tag (see “Prebuilt CUDA wheels” below) — pin torch to match.
• A dep’s setup.py would otherwise downgrade torch outside the supported set.

2.8.0 is the safest fallback for old requirements that refuse modern torch. 2.10.0 / 2.11.0 is the sweet spot for new code. When you pin torch, also pin torchvision / torchaudio to the matching minor — see the side-car drift section.

When a dep would silently downgrade torch (e.g. some forks of demucs, audiocraft pin torchaudio<2.1), install the offender from app.py with --no-deps rather than pinning torch around it:

List the offender’s real runtime deps yourself in requirements.txt.

Torch-family side-car drift

torchvision, torchaudio, torchcodec are built against a specific torch major.minor. Listing them unpinned usually works, but two known drift patterns:

• torchaudio==2.11.0 (and later) dropped its Requires-Dist: torch==X.Y.Z line. With torch pinned to 2.10, pip silently resolves torchaudio to 2.11.0 and the import fails on ABI mismatch.
• torchcodec declares no torch dependency in PyPI metadata at all.

Verification after pip install or uv lock --upgrade:

When PyPI is silent, fall back to the project’s README compatibility table (torchcodec’s lives at https://github.com/pytorch/torchcodec (opens in a new tab)).

Prebuilt CUDA wheels — the Blackwell wheels dataset

This is the first thing to reach for when a CUDA-extension package has no upstream wheel matching the ZeroGPU torch / CUDA / cxx11-abi cell. Prefer it over any runtime workaround (pip install git+… inside @spaces.GPU, committed stub packages, sys.modules injection, monkey-patch shims), all of which are slower, fragile, and eat duration budget. Canonical prebuilt sm_120 wheels:

https://huggingface.co/datasets/multimodalart/zerogpu-blackwell-wheels (opens in a new tab)

Wheels live at wheels/<cell>/<wheel>. A cell encodes torch × CUDA × Python as pt<torch>-cu<cuda>-cp<pyver>. Every cell ships the same seven packages:

flash_attn (two versions: 2.8.3 and 2.7.4.post1), xformers, pytorch3d, nvdiffrast, diff_gaussian_rasterization, torchmcubes.

Current cells (12):

Picking a cell

1. Match cp<pyver> to your python_version:. The wheels are cp-ABI-specific — a cp310 wheel needs Python 3.10, cp312 needs 3.12, etc. (flash_attn now ships cp310 through cp313, so this is a free choice, not a forced pin to 3.10 as in older versions of this doc.)
2. The wheels are torch-minor-tolerant. The flash_attn / xformers filenames encode no torch version, so a pt212-cu130-cp310 wheel runs fine on the live torch-2.11 cp310 runtime. flash_attn ships cp310 through cp313, so the Python choice is free. Pick the highest-torch cell for your Python version unless you’ve pinned an older torch — then match it (torch 2.8 → a pt28-cu128-cp3XX cell).
3. Copy the exact filename from the cell you pick. The xformers build hash differs across cells (0.0.34+3da0fc92… on the cu130 / torch≥2.10 cells, 0.0.34+41531cee… on the cu128 / torch 2.8–2.9 cells). Don’t hardcode one filename across cells — list the cell (hf download --repo-type dataset multimodalart/zerogpu-blackwell-wheels --include "wheels/<cell>/*" or the Hub file browser) and copy what’s there.

Reference by direct URL in requirements.txt:

Per-package status

Each package’s version is constant across cells; only the cp/torch/CUDA tags in the filename change. The "instead of" column is the runtime workaround to avoid — the wheel is the clean path.

Pattern

Resolve the exact filenames from your chosen cell, then:

Do not install these from @spaces.GPU startup. A subprocess.check_call pip-install at first GPU acquire is strictly worse than the wheel URL — slower cold start, eats duration budget, breaks reproducibility, and the build sometimes exceeds the @spaces.GPU(duration=1500) cap.

When you need a wheel that’s not in the dataset

Three options, in preference order:

1. kernels-community — https://huggingface.co/kernels-community (opens in a new tab) handles ABI matching for you. Often the simplest path; no version pinning needed.
2. Upstream wheel matrix — e.g. flash-attention’s releases page ships a fairly complete cu12 / torch / Python matrix at https://github.com/Dao-AILab/flash-attention/releases (opens in a new tab). Pin torch==X.Y.Z in requirements.txt to match the wheel’s torch2.X tag.
3. Build it yourself and host on HF Hub. Last resort — see debugging.md for the in-@spaces.GPU source-build pattern as a stopgap while a wheel is being built.

Reading a CUDA wheel filename

ABI / symbol mismatches at any of these → ImportError on first import. Pin torch to match torch2.X. Set python_version: to match cp3XX.

Don’t pin xformers

Leave bare in requirements.txt (or use the prebuilt URL above). Pip picks the wheel matching your installed torch.

Don’t pin spaces

Even if a uv export produces it, exclude with --no-emit-package spaces. The platform always pins its own version.

Specifically about Python version

Pinning python_version: is effectively required:

• ZeroGPU officially supports 3.10.13 and 3.12.12.
• The runtime default is 3.10.
• Pinning to a cp3XX wheel matrix (e.g. cp310 flash_attn wheel) forces matching Python.

Both "3.12" and "3.12.12" forms are accepted in YAML.