Change the context length or server arguments

Each recipe runs llama-server with exactly the settings that were measured, including a context of 65,536 tokens. You can change the context length, or add llama-server arguments of your own. The launcher then calls the run untuned: it is no longer the configuration the receipt measured, so the receipt’s numbers don’t apply to it.

You need the launcher installed and a host that passes tensorless doctor. See Prepare a Linux host for a P100. Words such as recipe and receipt are in the glossary.

See the measured settings first

tensorless info pn3

Under Run, context is the context length and args lists every other argument the recipe passes, in order. See exactly what will run explains the rest of the output.

Change the context length

Pass --context to run or serve:

tensorless run iamMrDuncan/Qwen3.8-27B-PN3-GGUF --context 32768

Before it starts, the launcher says what differs from the recipe:

UNTUNED: --context 32768 instead of 65536. This is not the configuration the receipt measured; its numbers do not apply.

The value goes to llama-server as -c. With -np 1 in the recipes, the server has one slot, and that slot gets the whole context.

Caution

The recipes’ 65,536-token context already takes nearly all of the P100’s 16 GB: with Qwen3.8-27B-PN3, the published image used 13,997 MiB once loaded and 14,577 MiB while generating (the receipt). A larger context needs more memory than the card has, and the server fails while it loads. A smaller one frees memory.

Pass extra llama-server arguments

Everything after -- goes to llama-server, after the recipe’s own arguments:

tensorless run iamMrDuncan/Qwen3.8-27B-PN3-GGUF -- --temp 0.7

UNTUNED: extra arguments --temp 0.7. This is not the configuration the receipt measured; its numbers do not apply.

Because they come last, your arguments can also override one of the recipe’s. The launcher sets the model, address, context, alias and threads itself, so it refuses -m, --host, --port, -c, --alias, -t and -tb (and their long forms) after --:

tensorless: do not pass -c after --: the launcher sets it. Use --host, --port or --context instead; the model, alias and threads come from the recipe
Run 'tensorless help' for usage.

Only run, serve and info take arguments after --. The others refuse them.

Tip

Sampling settings such as temperature can also be sent with each request, which leaves the run tuned. See Connect an OpenAI-compatible client.

Combine both

tensorless run iamMrDuncan/Qwen3.8-27B-PN3-GGUF --context 32768 -- --temp 0.7

UNTUNED: --context 32768 instead of 65536 and extra arguments --temp 0.7. This is not the configuration the receipt measured; its numbers do not apply.

Check that it worked

Preview the command before you run it, with the same flags and --dry-run. The plan’s header includes the UNTUNED line, and the podman run (or docker run) command shows your -c value and, at the end, your extra arguments:

tensorless run iamMrDuncan/Qwen3.8-27B-PN3-GGUF --context 32768 --dry-run -- --temp 0.7

To measure an untuned run, start the server, then run tensorless bench with the same --context. The result notes that the run is untuned, and its comparison with the receipt is only a rough guide. Tuned and untuned runs explains why.