- Added models table to database
- Created admin routes for model CRUD operations
- Added models.html template
- Added models link to admin dropdown
- Implemented Hugging Face model download functionality
- Added download_huggingface_model function in utils.py

Fix timezone issue by explicitly treating SQLite timestamps as UTC before converting to Unix timestamp

Fix last seen timestamp parsing to handle both ISO and SQLite timestamp formats for proper timezone display

Fix cluster status GPU detection to use available_backends instead of gpu_info for more reliable GPU client identification

Show connected cluster nodes in analyze page right panel, displaying hostname, IP, GPU status, weight, and last seen time

Add documentations dropdown menu in navbar, moving API documentation and admin API documentation links

Add admin dropdown menu in navbar for admin users, moving train, cluster tokens, cluster nodes, and configurations links

- Store connected_at as UTC ISO string with Z suffix
- Handle datetime objects from database as UTC timestamps
- Ensure uptime calculation uses correct UTC times regardless of server timezone

- Store connected_at as proper UTC timestamp in database using FROM_UNIXTIME/datetime
- Update web interface to handle datetime objects and timestamps correctly
- Ensure uptime starts from actual connection time, not offset by timezone

- Fix cluster master to properly detect GPU backends from client capabilities
- Extract available_backends from capabilities list instead of gpu_info
- Ensure clients with GPU workers are correctly identified as GPU-enabled

- Fix cluster client uptime calculation to start from 0 by using explicit UTC timestamps
- Fix cluster client workers not showing by populating cluster_master.clients dictionary
- Ensure connected_at uses current UTC time instead of database CURRENT_TIMESTAMP

- Get real worker process information from cluster master instead of placeholder data
- Display correct number of workers and their actual backends
- Improve accuracy of cluster node statistics

- Update connected_at timestamp on each successful connection
- Uptime now resets to 00:00:00 on reconnections as expected

- Extract real client IP address from websocket connection
- Preserve connected_at timestamp for accurate uptime calculation
- Send full GPU device info from client to master for proper VRAM reporting

- Handle timestamp strings in sorting function
- Parse ISO timestamp strings to numeric values for proper sorting
- Prevent TypeError when sorting by last_seen timestamp

- Add ALTER TABLE to create connected_at column in existing databases
- Handle migration gracefully for databases created before schema update

- Move subprocess import to function scope to avoid UnboundLocalError
- Ensure subprocess is available for fallback GPU detection

- Add connected_at timestamp to track when client first connected
- Calculate uptime from connection time instead of last seen time
- Update database schema and API to use proper uptime tracking

- Update detect_gpu_backends to collect actual VRAM for each GPU device
- Store device info including VRAM in gpu_info sent to master
- Use real VRAM data in cluster nodes API instead of hardcoded values
- Ensure consistent VRAM reporting between master and clients

- Use green background color to indicate connected nodes
- Remove status text column for cleaner interface
- Update colspan values for table messages

- Store cluster client info in database for persistence
- Update API to read connected clients from database
- Maintain compatibility with existing web interface