Termux-Vision | Advanced Parameters & Security

Advanced Parameters & Boundary Security

Hyperparameter reference, boundary guards, and device isolation protocols

1. Advanced VLM Inference Hyperparameters

Parameter	Type / Range	Default	Description & Behavior
`prompt`	`str` (Non-empty)	Default Korean prompt	User prompt query. Empty/whitespace raises `ValueError`.
`max_tokens`	`int > 0`	`150`	Maximum token generation budget. Values ≤ 0 are rejected.
`temperature`	`float ≥ 0.0`	`0.2`	Sampling temperature. Lower values produce deterministic outputs.
`top_p`	`0.0 < p ≤ 1.0`	`None`	Nucleus sampling probability threshold.
`top_k`	`int ≥ 1`	`None`	Top-k sampling threshold.
`repeat_penalty`	`float ≥ 0.0`	`1.2`	Penalty applied to repeated token logits.
`seed`	`int`	`None`	Explicit RNG seed for reproducible generation.
`system_prompt`	`str`	`None`	System persona and behavioral steering prompt.
`ngl`	`int ≥ 0`	`99` (Vulkan) / `0` (CPU)	Number of neural layers offloaded to Vulkan GPU.

2. Strict Null / Zero Boundary Validation (Zero Silent Fallbacks)

Termux-Vision adheres to a strict Fail-Closed Validation Protocol. If invalid, null, or empty parameters are provided, the engine will never silently execute dangerous dummy fallbacks; it immediately raises explicit structured exceptions:

ValueError: Parameter 'image' cannot be null/None.
ValueError: Parameter 'prompt' cannot be an empty string.
ValueError: Parameter 'max_tokens' must be a positive integer > 0.
VulkanNotAvailableError: Explicit GPU mode cannot proceed. Switch to --device cpu.

3. Device Backend Execution Policy

# Auto Mode (Recommended for Android):
# Attempts Vulkan GPU -> Gracefully retries CPU on driver errors with warning
termux-vision vlm photo.jpg --device auto

# Strict GPU Mode:
# Enforces Vulkan GPU. Fails immediately without silent CPU fallback if GPU unavailable
termux-vision vlm photo.jpg --device gpu

# Dedicated CPU Mode:
termux-vision vlm photo.jpg --device cpu --threads 4