gpkg

GPU package manager. Stop compiling CUDA extensions.

pip install gpkg

gpkg add flash-attn causal-conv1d mamba-ssm
# done. resolved, locked, installed.

The problem

These packages take 20-60 minutes to compile and regularly fail:

Package	Typical compile time	Common failure
flash-attn	25 min	OOM during build, CUDA mismatch
flash-attn-3	20 min	SM90+ only, rare wheel coverage
causal-conv1d	10 min	Torch ABI mismatch
mamba-ssm	15 min	Cascading causal-conv1d failure
natten	30 min	CUTLASS dependency, arch-specific
sageattention	10 min	Windows build nightmare
grouped-gemm	10 min	MoE stack dependency

Prebuilt wheels exist across dozens of GitHub repos and pip indexes. Finding the right wheel for your exact python + torch + cuda + platform combo is a scavenger hunt nobody should repeat.

Install

pip install gpkg

# or with uv
uv tool install gpkg

# or from source
git clone https://github.com/Mapika/gpkg && cd gpkg
uv tool install .

Usage

Just add packages

# Auto-detect torch + cuda, resolve wheels, lock, install
gpkg add flash-attn causal-conv1d mamba-ssm

# If no prebuilt wheel exists, it builds from source (optimized)
gpkg add flash-attn causal-conv1d mamba-ssm  # --build-missing is automatic

Resolve without installing

# Explicit versions
gpkg --torch 2.11.0 --cuda 130 flash-attn flash-attn-3 -o pyproject.toml

# Auto-detect torch and CUDA from your environment
gpkg flash-attn causal-conv1d mamba-ssm

Lockfile for reproducible installs

# Write a lockfile with exact versions and URLs
gpkg --torch 2.11.0 --cuda 130 flash-attn causal-conv1d --lock

# Install from lockfile (no network needed)
gpkg install -o pyproject.toml

# Install + sync in one step
gpkg install --sync

Verify your environment

gpkg test flash-attn causal-conv1d mamba-ssm

  ok  torch 2.11.0+cu128  GPU: NVIDIA GeForce RTX 5070 Ti
  ok  flash-attn 2.8.3
  ok  causal-conv1d 1.6.1
  ok  mamba-ssm 2.3.1

See what's available

gpkg --available causal-conv1d natten

causal-conv1d -- available wheels (linux_x86_64)
┏━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ CUDA ┃ PyTorch versions        ┃
┡━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━┩
│ 12   │ 2.10, 2.9, 2.8, 2.7    │
│ 13   │ 2.10, 2.9              │
└──────┴────────────────────────┘

Diagnostics

gpkg --explain --torch 2.11 --cuda 130 flash-attn   # why a wheel was/wasn't selected
gpkg --doctor --torch 2.11 --cuda 130 flash-attn     # verify URLs are accessible

Cache management

gpkg --cache-info                    # show cache statistics
gpkg --cache-clean                   # clean all cached data
gpkg --cache-clean --older-than 1h   # clean entries older than 1 hour

When there's no wheel: compile fast

When gpkg add can't find a prebuilt wheel, it automatically builds from source with optimized settings:

• Detects your GPU arch via nvidia-smi → builds for only that arch
• Uses all CPU cores (capped by RAM to prevent OOM)
• Enables ninja for parallel builds
• Caches the built wheel so you never compile the same package twice

You can also use build-env.sh manually:

source build-env.sh
uv add causal-conv1d    # 5-10x faster than default

How it works

1. Checks the hosted registry at wheels.mapika.dev for cached wheels (fast)
2. Falls back to GitHub releases API and pip find-links indexes
3. Matches wheels against your torch + cuda + python + platform
4. Picks the best match per package (latest version, prefers non-manylinux)
5. If no wheel exists and --build-missing is set, compiles from source
6. Emits a valid pyproject.toml with [tool.uv.sources] pointing at direct URLs

Registry

The registry tracks 9 sources across 7 packages:

Package	Sources
flash-attn	mjun0812/flash-attention-prebuild-wheels, Dao-AILab/flash-attention
flash-attn-3	mjun0812/flash-attention-prebuild-wheels
causal-conv1d	Dao-AILab/causal-conv1d
mamba-ssm	state-spaces/mamba
natten	SHI-Labs/NATTEN
grouped-gemm	fanshiqing/grouped_gemm
sageattention	woct0rdho/SageAttention, mobcat40/sageattention-blackwell

Adding a source

Edit src/gpkg/registry.toml and open a PR:

[[sources]]
package      = "causal-conv1d"
description  = "causal-conv1d -- your torch 2.11 builds"
type         = "github"
repo         = "yourname/causal-conv1d-wheels"
wheel_name   = "causal_conv1d-{version}+cu{cuda}torch{torch}-{pytag}-{platform}.whl"
cuda_style   = "full"
scan_tags    = 5

Configuration

Env var	Purpose
GITHUB_TOKEN	Raise API rate limit 60 to 5000 req/hr
UVFORGE_TOKEN	Bearer token for private registries
UVFORGE_TOKEN_<HOST>	Host-specific token (e.g. UVFORGE_TOKEN_WHEELS_MYCO_COM)
UVFORGE_REGISTRY	Override default registry path/URL

Private registries also support ~/.netrc for credential storage.

CI Usage

- name: Install GPU packages
  env:
    GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
  run: |
    pip install gpkg
    gpkg add flash-attn causal-conv1d mamba-ssm

For machine-readable output:

gpkg --list --json                           # all registered sources
gpkg --available flash-attn --json           # cuda/torch combos
gpkg --torch 2.10 --cuda 128 flash-attn --json  # resolved wheel URLs

Development

git clone https://github.com/Mapika/gpkg && cd gpkg
uv sync
uv run pytest -v
uv run ruff check src/

License

MIT