Skip to main content
nrvnad is the daemon server that loads an LLM model and processes inference jobs from a workspace directory.

​
Usage

nrvnad <model.gguf> <workspace>
nrvnad <model.gguf> <workspace> [workers]
nrvnad [--help | --version]

​
Interactive Mode

Run nrvnad without arguments to launch the interactive dashboard: 
nrvnad
The dashboard displays:
  • Available workspaces with job status (queued, done, failed)
  • Available GGUF models in the models directory
  • Interactive selection prompts

​
Arguments

​
model
string
required
Path to GGUF model file or model name to search for in the models directory.
​
workspace
string
required
Directory path for job storage and processing. Will be created if it doesn’t exist.
​
workers
number
Number of concurrent workers (default: 4, range: 1-64).

​
Options

​
--mmproj
string
Path to vision projection model (mmproj) for multimodal inference. Auto-detected if available.
​
-w, --workers
number
Number of concurrent workers to run (1-64).
​
--workspace
string
Specify workspace path via flag instead of positional argument.
​
-h, --help
boolean
Display help message and interactive dashboard.
​
-v, --version
boolean
Display version number.

​
Environment Variables

​
NRVNA_MODELS_DIR
string
Custom directory for GGUF models. Defaults to ./models or auto-detected relative to binary location.
​
NRVNA_LOG_LEVEL
string
Log level: ERROR, WARN, INFO, DEBUG, TRACE (default: INFO in server mode, ERROR in interactive mode).
​
NRVNA_TEMP
number
Sampling temperature. Auto-set based on model type if not provided:
  • Code models: 0.3
  • Reasoning models (DeepSeek, R1): 0.6

​
Examples

​
Start daemon with specific model

nrvnad qwen2.5-7b.gguf ./workspace

​
Start with 8 workers

nrvnad llama-3.1-8b.gguf ./workspace 8

​
Start with vision model

nrvnad llama-3.2-11b-vision.gguf ./workspace --mmproj llama-3.2-11b-mmproj.gguf

​
Using flags

nrvnad --workers 8 --workspace ./workspace qwen2.5-coder-7b.gguf
You can provide a partial model name and nrvnad will search the models directory: 
nrvnad qwen ./workspace
Matches the first .gguf file containing “qwen” in the models directory.

​
Behavior

​
Model Resolution

  1. If the provided path exists, uses it directly
  2. Otherwise searches NRVNA_MODELS_DIR for .gguf files matching the provided string
  3. Excludes mmproj files from model search
  4. Returns first alphabetically sorted match

​
MMProj Auto-Detection

When a model is loaded, nrvnad automatically searches for matching mmproj files:
  • Looks in the same directory as the model
  • Matches based on model filename stem
  • Prioritizes exact matches

​
Workspace Structure

The daemon expects/creates this structure: 
workspace/
├── input/
│   ├── ready/      # Jobs ready for processing
│   └── writing/    # Jobs being written
├── processing/     # Jobs currently running
├── output/         # Completed jobs
└── failed/         # Failed jobs

​
PID Management

The daemon writes its PID to <workspace>/.nrvnad.pid on startup and removes it on clean shutdown.

​
Signal Handling

The daemon handles SIGINT and SIGTERM for graceful shutdown: 
kill <pid>  # Graceful shutdown

Build docs developers (and LLMs) love

Get started for freeTalk to us