common: migrate the deprecated --mmap/--no-mmap to --load-mode

[fboudra]

Overview

Migrate the deprecated --mmap, --no-mmap, --mlock, and --direct-io flags to the unified --load-mode argument across scripts, examples, and documentation.

Additional information

• Updated SYCL examples, Snapdragon scripts (ADB + Windows), and benchmark scripts
• Updated internal warning message in llama-model-loader.cpp
• Updated env var documentation in server/README.md
• Updated completion/README.md "Model Loading Mode" section

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[ggml-gh-bot]

Hi @fboudra, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• Large PR: Large changes require prior discussion (e.g. an issue or RFC) and maintainers may not be able to review this PR as-is. Consider splitting it into smaller, focused PRs.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[taronaeo]

Thanks! For docs that specified --mmap, by right they should default to --load-mode mmap instead of auto to prevent unwanted behaviour.

The auto mode is relatively new and I don't know if the results will be different from how the original docs wanted it to be.

[arthw]

It's OK to SYCL backend.
I verified it and it's passed.

Thank you!

[fboudra]

* **Large PR**: Large changes require prior discussion (e.g. an issue or RFC) and maintainers may not be able to review this PR as-is. Consider splitting it into smaller, focused PRs.

The PR is large but should be easy to review with usually a one-liner per file modified.

[fboudra]

Thanks! For docs that specified --mmap, by right they should default to --load-mode mmap instead of auto to prevent unwanted behaviour.

The auto mode is relatively new and I don't know if the results will be different from how the original docs wanted it to be.

I had myself the same question. Following #26081 , it seemed safe to map --mmap to --load-mode auto instead of --load-mode mmap. The auto mode default to mmap except on devices that don't support mmap (CUDA/ROCm and Vulkan iGPUs). Looping @0cc4m in case he has additional info on the topic. Nonetheless, I'm fine one way or another.

[0cc4m]

For now auto just means mmap on most devices, none on integrated devices that can't directly use mmapped memory. In the future it may become more complex, ideally it would also pick direct-io where supported. Whatever is best for loading on a specific device.

[taronaeo]

I'm seeing this CI error: https://github.com/ggml-org/llama.cpp/actions/runs/31545705425/job/94473177972?pr=26934#step:9:12948

I don't think it is caused by this PR, but I'm taking a deeper look to confirm before merging.