server : support reading arguments from environment variables

[ngxson]

Motivation

When deploying to HF inference endpoint, we only have control over the environment variables that can be passed to docker. That's why currently we need to build a custom container and specify these variables via LLAMACPP_ARGS (ref: #9041)

This PR add some server-related arguments to environment variables (see a full list in server/README.md)

Variables are being prefixed LLAMA_ARG_ to distinguish them from compile-time variables like LLAMA_CURL.

Example

LLAMA_ARG_MODEL=../models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf LLAMA_ARG_CTX_SIZE=1024 LLAMA_ARG_N_PARALLEL=2 LLAMA_ARG_ENDPOINT_METRICS=1 ./llama-server

In case the same variable is specified in both env and arg, we prioritize env variable:

LLAMA_ARG_MODEL=my_model.gguf ./llama-server -m another_model.gguf
# Expected behavior: we load my_model.gguf
# (in other words, "-m another_model.gguf" is ignored)

On HF infrefence endpoint, these variables can be set from "Settings" tab. (In near future, these variable will be exposed as pre-defined input fields in the UI)

---

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High