Ziel

/root/.ollama

Standard

/mnt/user/appdata/ollama

Wert

/mnt/user/appdata/ollama

Port number where ollama listens on.

Ziel

11434

Standard

11434

Wert

11434

IP and Port the server binds to. Set to 127.0.0.1:11434 for internal only access.

Standard

0.0.0.0:11434

Wert

0.0.0.0:11434

Comma-separated list of allowed CORS origins.

Standard

*

Wert

*

How long a model stays in VRAM, e.g. 60m or 24h (Set to -1 for infinite, 0 for none).

Standard

5m

Wert

5m

Timeout for stall detection during model loads.

Standard

5m

Wert

5m

Max number of parallel requests a single model can handle.

Standard

1

Wert

1

Default context window (tokens) if not specified by the model.

Standard

4096

Wert

4096

Quantization type for the K/V cache, e.g. f16, q8_0, q4_0.

Standard

f16

Wert

f16

The path where model weights and blobs are stored.

Standard

/root/.ollama/models

Wert

/root/.ollama/models

Maximum number of models loaded per GPU at once (Set to 0 for infinite).

Standard

0

Wert

0

Max requests that can wait in line when the server is busy.

Standard

512

Wert

512

Log detail level: 0 for INFO, 1 for DEBUG, 2 for TRACE.

Standard

0|1|2

Reserved VRAM (in bytes) to leave empty on each GPU.

Standard

0

Wert

0

Enables experimental Flash Attention optimizations.

Standard

false|true

If true, always spreads model layers across all visible GPUs.

Standard

false|true

Optimizes prompt caching when multiple users share a model.

Standard

false|true

If true, does not delete unused model blobs on startup.

Standard

false|true

Disables the readline history in the interactive CLI.

Standard

false|true

Enables the experimental new Ollama engine.

Standard

false|true

Enables experimental Vulkan hardware acceleration.

Standard

false|true

Proxy for downloading models over HTTP.

Proxy for downloading models over HTTPS.

Comma-separate list of hosts/IPs that bypass the proxy.