nvidia ux & bookworm bpo cleanup

- Removed redundant confirmation dialogs from `_install_nvidia_stack`. Users can now access driver management (including removal) without first confirming installation intent; initial plan display is followed by pause for review instead of blocking prompt.
- Automated CUDA repository enabling on Debian 13/Bookworm paths (590/595). Manual confirmation prompts eliminated; script auto-enables repo via `_enable_cuda_repo` with error handling while maintaining safety checks.
- Split CUDA driver installation into two distinct APT transactions in `nvidia.sh`. Step 1 installs mandatory version pinning (`nvidia-driver-pinning-${ver}`); failure aborts process. Step 2 installs `nvidia-open`, relying on transitive dependencies for firmware rather than explicit package listing.
- Eliminated all Bookworm backports support due to EOL (2026-08-09). Removed `_install_nvidia_bookworm_bpo()`, dead code `install_nvidia_driver()`, and dispatcher logic for Debian 12. Turing now installs standard packages only; Trixie backports remain intact.
- Updated documentation (`docs/gpu.md`) to reflect actual installation flow and removed references to deprecated functions.
This commit is contained in:
stornic56
2026-08-10 17:58:03 -05:00
committed by GitHub
parent 2bcdb511b2
commit dbd2dfca37
4 changed files with 49 additions and 164 deletions
+10 -14
View File
@@ -178,20 +178,16 @@ The NVIDIA driver installation process is inherently complex due to proprietary
- **CUDA Repository Enablement**: On Debian 12 (Bookworm) the `nvidia-cuda` repository is enabled via `extrepo`. On Debian 13 (Trixie) the official `cuda-keyring` package is used instead (`wget` + `dpkg -i`), since extrepo cannot configure the repo correctly on Trixie. After enabling, the script always runs an explicit `apt update`, then installs `nvidia-driver-pinning-<ver>` if available (official NVIDIA pinning, optional), and finally installs the unified metapackage `nvidia-open` plus `firmware-nvidia-gsp`. No manual pinning file is written; if `apt install` fails, the real apt error is shown and the installation aborts.
#### **Installation Flow**
```bash
if [ "$HAS_NVIDIA" = true ]; then
if [ "$(is_backports_kernel)" == "true" ] && \
{ [ "$DEBIAN_CODENAME" != "trixie" ] || ! is_nvidia_blackwell; }; then
# Warn about DKMS compatibility or use enterprise repo
offer_nvidia_enterprise_repo
fi
if _confirm "NVIDIA Driver"; then
install_nvidia_driver # Installs latest stable (535/550) or v590 for Blackwell
fi
fi
```
The NVIDIA flow is driven by `_install_nvidia_stack` (dispatcher in `modules/gpu.sh`):
1. **Plan**`_msg` with detected GPUs, then `_pause`.
2. **Manage menu** (`_nvidia_manage_menu`): "Install / Change NVIDIA Driver Version" or "Remove NVIDIA Driver and Restore Nouveau".
3. **Version menu** (`_show_nvidia_version_menu`): Debian 12 → v535 (Recommended) / v470 (Kepler); Debian 13 → v550 (Recommended) / v590 / v595 (NVIDIA CUDA Repo).
4. **Dispatch by hardware**: Kepler → `nvidia-tesla-470-driver`; Fermi → veto with message; Turing+ → `_install_nvidia_standard` (Debian stable); Blackwell on Trixie → CUDA repo v590.
5. **CUDA repo (590/595)**: two separate APT transactions — `nvidia-driver-pinning-<ver>` (version lock, mandatory) then `nvidia-open` (driver + open kernel modules, firmware via transitive dependency).
6. **Post-install**: `/etc/modprobe.d/nvidia-wayland.conf` + DKMS verification.
> Note: Debian 12 (Bookworm) backports reached EOL (2026-08-09); the NVIDIA flow on Bookworm only uses the stable repository.
---
### 8.**NVIDIA Driver Management & Kernel Compatibility**