Feature Request: Add Explicit Context Reset for llama-cli or llama-server

[multimediaconverter]

Prerequisites

• I am running the latest code. Mention the version if possible as well.
• I carefully followed the README.md.
• I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• I reviewed the Discussions, and have a new and useful enhancement to share.

Feature Description

Add a built-in way to reset the context without reloading the model, such as:

/reset command in llama-cli interactive mode.

HTTP API endpoint (e.g., POST /reset) for llama-server.

Use Case:
Chat applications needing fresh sessions.
Testing prompts without residual context.
Long-running servers where model reloads are costly.

Motivation

Currently, to start a new conversation with a clean context (no prior chat history), users must restart llama-cli (reloading the model, which is slow).

Possible Implementation

No response

[multimediaconverter]

I noticed the cache_prompt option indirectly manages KV cache state in llama-server. The documentation explains how the KV cache is reused for efficiency, and it may cause slight output variations across runs due to batch-size differences in prompt processing vs. generation. Does llama-cli provide more deterministic/predictable outputs compared to llama-server due to differences in batching or caching?

[ggerganov]

After #14363 is merged, you can disable cache_prompt and the results will be deterministic.