v1.1.0: GPU Architecture, CUDA Extensions & Environment Health

✨ What's New

PyPilot 1.1.0 is built from the most common environment problems reported on Stack Overflow, GitHub and the Hugging Face forums: GPU builds that don't match the card, CUDA packages that won't install, broken venvs and binary mismatches. All of them are caught before they turn into a crash.

🎮 GPU & CUDA

• RTX 50-series (Blackwell) support: PyPilot now reads the GPU's compute capability, not just the driver. Blackwell cards (sm_120) need a CUDA 12.8+ build and torch 2.7+; older builds fail with "no kernel image is available". If no build fits both the driver and the card, PyPilot says exactly what to do (update the driver, or torch==X --index-url cuY)
• Installed torch checked against the GPU: a torch build that's too old for the card, or a CPU-only torch on an NVIDIA machine (e.g. after another install swapped it), gets a warning with the exact reinstall command
• GPU self-test after install: right after PyPilot installs torch, it checks torch.cuda.is_available() and the supported architectures, and explains any mismatch
• uv projects keep the CUDA build: the driver-matched PyTorch index is now recorded in pyproject.toml ([[tool.uv.index]] + [tool.uv.sources], Linux/Windows only), so uv sync and uv run no longer reinstall CPU torch. pypilot fix cuda sticks in uv projects too
• TensorFlow on native Windows: TF 2.11+ has no GPU support on native Windows. PyPilot now says so instead of implying a GPU build, and suggests WSL2 with tensorflow[and-cuda]
• Current CUDA data: torch 2.13–2.14.1 and CUDA 13.0–13.4 driver minimums. New drivers now get the newest matching build (e.g. cu132 instead of cu126)

🧩 CUDA extension packages

• flash-attn, xformers, bitsandbytes, mamba-ssm, causal-conv1d are checked before install: source-only packages that need nvcc and --no-build-isolation, missing Windows/macOS wheels, and xformers releases pinned to a different torch than yours. Each warning suggests an alternative such as PyTorch SDPA or WSL2
• Import quick fixes for these packages explain the problem instead of offering an install that will fail

🩺 Environment health

• Broken .venv detection: if the Python a venv was built on has been moved or removed, you get one clear error and a working Rebuild fix, instead of confusing "not installed" squiggles
• NumPy 1 vs 2 mismatches: packages built for NumPy 1.x next to NumPy 2 are named before they crash with "numpy.dtype size changed". A light pip check covers other installed dependency conflicts
• Source builds explained: PyPilot tells apart packages that definitely need a compiler from ones that may not, gives the fix for your OS (Microsoft C++ Build Tools, xcode-select --install, build-essential, rustup), skips advice for tools you already have, and names a Python version that has ready-made wheels

📓 Notebooks

• ipykernel pinned to >=6.31.0,<7. This avoids both ipykernel 7.0.0 (cannot connect on Python 3.14) and the 7.x notebook hangs reported in VS Code. An ipykernel you installed yourself is left alone

🟦 Zed

• Nested projects: when the opened folder has no Python project but a single subfolder does, PyPilot works on that subfolder
• Optional pyright pin: after setup, Zed users can choose to point pyright/basedpyright at .venv. PyPilot never does this without asking and never overwrites existing config

🛠 Maintenance

• scripts/check_data.py reports when the bundled GPU/framework tables fall behind PyPI and PyTorch, and checks that table changes stay compatible with released engines. A weekly CI job runs it
• AGENTS.md documents the data-first workflow and commit rules for contributors and AI agents

---

📦 Installation

Via Zed (recommended): Install the PyPilot extension from the Zed extension registry

Via VS Code: PyPilot for VS Code

Via pip:

pip install --upgrade pypilot-cli

Full Changelog: v1.0.5...v1.1.0