Conversation
Addresses 27 user-reported issues from live testing on an RTX 3080 system booting from USB. All changes deployed and verified on hardware. Crash handling: - Override App._handle_exception() to capture Textual runtime crashes - Write crash dumps to persistent disk (/var/lib/neuraldrive/logs/) - Screenshots routed to persistent disk via TEXTUAL_SCREENSHOT_LOCATION - Outer try/except in __main__ catches startup crashes Chat screen: - Fix TypeError from RichLog.write(end='') — removed invalid param - Move streaming response to @work(exclusive=True) to unblock UI - Add on_screen_resume to refresh model list on every screen visit - Add model selector (Select widget) on dedicated row with amber border Models screen: - Rewrite catalog with two-zone keyboard navigation (list + buttons) - Arrow keys navigate, Enter/Space toggle, PgUp/PgDn page jump - Add download cancel button with worker cancellation - Handle asyncio.CancelledError in _start_pull - Add model load/unload via Ollama generate API (keep_alive) - Show both Load and Unload buttons per model (disable irrelevant one) - Fix ModelItem._size/_name collision with Textual Widget internals Services screen: - Fix DuplicateIds crash: await remove_children() before mounting - Use sudo systemctl for service start/stop/restart - Arrow-key service selection with yellow highlight - Use Binding() objects for show/priority params (not 4-element tuples) Dashboard: - Expand GPU StatsBox to show Device, VRAM, Temp, Utilization - Rename 'Loaded Models' to 'Active Models (VRAM)' Wizard: - Rewrite _create_persistence_partition(): fix parted start position, detect actual free space, immediate mount, correct Ollama dirs, proper ownership, restart Ollama after partition creation - Add YAML config persistence (persistent disk with overlay fallback) Navigation: - Replace single-letter hotkeys with F2-F6 function keys (priority=True) - Remove old silent hotkeys entirely - Disable command palette via ENABLE_COMMAND_PALETTE=False (COMMAND_PALETTE_BINDING=None crashes Textual 8.2.4) Security: - Add scoped NOPASSWD sudoers (/etc/sudoers.d/neuraldrive-tui) that survives wizard _finalize() stripping NOPASSWD from neuraldrive-admin - Covers systemctl, parted, mkfs, mount, chpasswd, and file ops New files: - utils/config.py: YAML config read/write with persistent/overlay fallback - utils/hardware.py: Boot device detection, partition enumeration - etc/sudoers.d/neuraldrive-tui: Scoped NOPASSWD rules for TUI ops - dev-reset.sh: Development reset script (password, NOPASSWD, sentinel) Build: - Add pyyaml to TUI venv dependencies - Set neuraldrive-tui sudoers permissions in build hook
Display parameter count, quantization level, disk size, and VRAM usage for each installed model. VRAM is cached to persistent config on first load so it remains visible after unloading. Fix model-item height (3->5) so Load/Unload buttons render inside the bordered container instead of being clipped. Show both buttons per model with the irrelevant one disabled. Add disabled button styles.
Write api.key and credentials.conf to both /etc/neuraldrive/ (overlay) and /var/lib/neuraldrive/config/ (persistent disk) when available. Update wizard completion text to show where the key is stored instead of telling the user to save it manually.
Updates every 2 seconds alongside the system stats refresh. Shows HH:MM:SS so the user can tell at a glance the dashboard is live.
- Compact model selector into horizontal row with inline label - Remove clipping on Select widget (border removed, height auto) - Enable text wrapping in chat log (wrap=True on RichLog) - Remove dock:bottom on input row to prevent footer collision - Center Send button label vertically
Save Select value before refreshing options list, restore it if the model is still available. Falls back to first model only when previous selection is no longer present.
- Add red Delete button to each installed model item - Auto-unload from VRAM before deleting if model is loaded - Fix httpx DELETE with json body (use client.request instead) - Preserve selected chat model when returning to chat screen
- Gate sentinel write behind errors check: sentinel is only written after config.save() and all prior writes succeed, preventing the wizard from being silently skipped after partial failures - Guard partition detection: reject if lsblk returns base device instead of new partition, preventing accidental whole-disk format - Add --wizard CLI flag to force wizard rerun on demand - Add on_input_submitted to ModelsScreen so Enter in the pull-input field triggers model download
…ror checking - Launcher now forwards "$@" so neuraldrive-tui --wizard works - Partition detection uses before/after diff instead of fragile last-line - Wizard completion uses sentinel file as single source of truth - config.save() and wizard._sudo_write() check all subprocess return codes
lsblk before-snapshot was taken after mkpart, which could show the new partition if the kernel auto-detected the table change. Snapshot now taken before mkpart so the diff is always reliable.
- Abort before mkpart if pre-lsblk snapshot fails (no disk mutation without a valid baseline) - Check partprobe return code; poll lsblk with bounded retry loop instead of fixed sleep(2) - Replace fragile regex in get_boot_device() with lsblk PKNAME (supports NVMe, MMC, and sd devices) - Guard Enter-to-pull against re-submission during active download
Set _pulling=True immediately in both user-facing entry points before scheduling the @work worker, closing the race window. Pull button handler now mirrors the Enter-to-pull guard.
…ve-media path, guard _pull_next - Remove wizard_complete config key write from wizard finalize; sentinel file is now the single source of truth for wizard completion - Remove unused wizard_complete() function from config.py - Check return codes for all subprocess calls in partition creation: mkdir, chown, umount, systemctl (warning-only for restart) - Normalize live-media= cmdline path through lsblk PKNAME for NVMe/MMC - Set _pulling=True in _pull_next() before _start_pull() to prevent concurrent pull submissions from all entry points
Instead of returning the raw live-media= partition path when lsblk PKNAME resolution fails, fall through to the findmnt detection path. This prevents handing an unvalidated partition/symlink path to the storage wizard for partition creation.
Replace Textual's Header with SafeHeader subclass that catches NoMatches during title watcher updates. Textual 8.2.4 only catches NoScreen in the set_title watcher but not NoMatches, causing crashes when screens are pushed/popped and HeaderTitle hasn't recomposed yet. This is a known upstream bug (Textualize/textual#4258, PR #4817). Simplify --wizard: instead of a separate force_wizard constructor flag, --wizard now removes the sentinel file before launch so the existing on_mount check triggers the wizard naturally.
…e filter - Add ExecStartPre to load nvidia-current-uvm module and create /dev/nvidia-uvm device nodes before Ollama starts (with - prefix for non-fatal failure on non-NVIDIA systems) - Remove DeviceAllow lines that blocked CUDA access under cgroup v2 - Add nvidia-modprobe to NVIDIA package list for device node creation - Add /etc/modules-load.d/nvidia-uvm.conf for early boot module load - Show [GPU]/[CPU] tags with VRAM usage per model on dashboard
Rich interprets [GPU] and [CPU] as style tags and silently drops them. Escape with backslash-bracket on dashboard. Also change model_item status from 'VRAM' to 'GPU' for consistency.
Up/Down/PgUp/PgDn navigate between model items with a yellow highlight border. The scroll container follows the highlighted item via scroll_visible(), matching the catalog popup behavior.
Tab cycles between zones: model list, Browse button, Pull input, Pull button. Within the model list zone, Up/Down navigates models with scroll-follow, Left/Right selects Load/Unload/Delete per model, Enter activates the selected button. All ModelItem buttons are non-focusable — navigation is fully managed by the screen.
Left/Right nav now skips disabled buttons (Unload when not loaded, Load when already loaded). Load button shows 'Loading...' and disables during VRAM load. Added column header row (Params, Quant, Disk, VRAM, Status) aligned with model item columns.
Poll /api/ps after unload until model is actually evicted (Ollama returns 200 before eviction completes). Await remove_children() to prevent stale widgets. Use keep_alive=-1 for manual loads so models stay loaded until explicitly unloaded.
Ollama rejects "-1" with 'missing unit in duration', but accepts the integer -1 for infinite keep-alive.
Each service gets its own row with inline Start/Stop/Restart buttons. Arrow keys navigate services (Up/Down) and buttons (Left/Right). Disabled buttons are skipped. Enter activates the highlighted button. Service status auto-polls every 5 seconds and updates in place.
Poll fires every 5s but _load_services clears and remounts items. Skip poll while _loading flag is set to avoid NoMatches on .svc-state.
Set OLLAMA_MAX_LOADED_MODELS=0 (auto) so Ollama manages concurrency based on available VRAM. Add persistent EnvironmentFile override so config on /var/lib/neuraldrive/config/ollama.conf survives reboots, falling back to baked-in defaults when persistent disk is unavailable.
Wizard was missing /var/lib/neuraldrive/webui from the directory list, causing systemd NAMESPACE failure (status=226) when ReadWritePaths referenced the missing path.
…hanges Rewrite 17 docs files across user-guide and dev-guide to match the current implementation after the TUI UX overhaul, GPU/VRAM fixes, and Ollama configuration changes. Key updates: - Replace old single-letter hotkeys with F1-F5 function key nav - Rewrite models and services screen docs for zone-based navigation - Correct first-boot wizard steps, sentinel file path, and --wizard flag - Update OLLAMA_MAX_LOADED_MODELS from 1 to 0 (auto/LRU eviction) - Document DeviceAllow removal (cgroup v2 eBPF incompatibility) - Document nvidia-current-uvm module naming and boot-time loading - Add nvidia-uvm and cgroup v2 GPU troubleshooting sections - Add persistent config override (EnvironmentFile) documentation - Document crash dump logging, VRAM cache, and chat model selector
Chat model dropdown now prefixes loaded models with * so users can see which models are ready without loading delay. A 10-second poll timer keeps the indicators current as models load/unload. Input focus is restored after each response completes so users can type follow-up messages without re-clicking the input box.
This was referenced Apr 24, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Addresses 27 user-reported bugs and UX issues discovered during live testing on an RTX 3080 system booting NeuralDrive from USB, plus a critical GPU acceleration bug where Ollama was silently falling back to CPU-only inference. Includes comprehensive documentation updates across 17 files to reflect all implementation changes. All code changes have been deployed to the live system and verified on hardware.
Commits
6ecd510334ef93927df1aa1e82c4bdfd694d0f1ca8ff34cabfbac7b66efe83a5534f54c0e802cb0d8a8864a9514b1003b1493fe4e5e1d3765f8908eb555d1f3c12f7d8064d81fe0a28f4bc2f3278fbc0dc8e3a71ea9fcfc2a704c637d03302160fa85dec79cc621827Issues Addressed
Closes #6, #8, #9, #10, #11, #13. References #4, #5, #7, #12, #14.
TUI Changes
Navigation Overhaul
Models Screen (complete redesign)
/var/lib/neuraldrive/config//api/psuntil confirmed)keep_alive: -1(integer) for infinite retention on manual loadsServices Screen (complete redesign)
ServiceItemwidget with inline Start/Stop/Restart buttons (colored: green/red/amber)_loadingguard against widget rebuild raceChat Screen
*prefix in selector, refreshed every 10 seconds@work(exclusive=True)Dashboard
[GPU]/[CPU]) rendered correctly (escaped Rich markup)Wizard & First Boot
/var/lib/neuraldrive/webui/--wizardCLI flag to force re-run/etc/neuraldrive/first-boot-completeReliability
SafeHeaderwidget catches Textual Header NoMatches bug (#4258)/var/lib/neuraldrive/logs/tui-crash-*.log/var/lib/neuraldrive/screenshots/GPU / System Changes
Critical: GPU Acceleration Fix
modprobe nvidia-current-uvm+nvidia-modprobe -uasExecStartPrein Ollama service and/etc/modules-load.d/nvidia-uvm.conffor boot-time loadingDeviceAllowfrom Ollama service, keptPrivateDevices=noOllama Configuration
OLLAMA_MAX_LOADED_MODELS=0(auto, was 1) — concurrent model loading with LRU evictionEnvironmentFile=-/var/lib/neuraldrive/config/ollama.confDocumentation Updates (17 files)
User Guide
--wizardflagOLLAMA_MAX_LOADED_MODELSto 0 (auto), added persistent config override docs, expanded config inventoryDeveloper Guide
Testing
All TUI changes deployed and verified on live hardware (RTX 3080, 500GB USB, Debian 12, kernel 6.1, Ollama 0.21.1). GPU acceleration confirmed:
inference compute: CUDA compute=8.6, NVIDIA GeForce RTX 3080, 10.0 GiB available.