llama : wire up ggml_backend_register_host_buffer for mmap'd weights

[bluechiperic]

Summary

ggml_backend_cuda_register_host_buffer() is defined and exposed through
get_proc_address by both the CUDA and SYCL backends — but nothing in the
repository calls it. Page-locking of mmap'd model weights is unreachable today.

A code search across master finds only the two definitions and no consumer.
PR #15615 extended the function to HIP in Aug 2025, so it appears intended to be
used rather than deprecated; the call site simply seems never to have landed.

This adds it.

Implementation

The call goes in llama_mmap's ctor/dtor because their lifetime matches the
mapping exactly — llama_model takes ownership of the loader's mappings, so
registering from the loader would unregister too early.

Resolved through ggml_backend_reg_get_proc_address, so src/ gains no
backend-specific dependency and backends without the hook fall through
unchanged. The CUDA backend gates on GGML_CUDA_REGISTER_HOST, so this is a
no-op unless explicitly requested.

+47/-2 across two files.

Why it helps

Pinned host memory lets host→device copies use DMA instead of the driver's
internal bounce buffer. This matters on MoE offload paths, where expert weights
are uploaded from host memory during batched prefill.

Measurements

OLMoE-1B-7B-0125 Q6_K, RTX A2000 12GB, -ncmoe 16 -ngl 999 -p 8192, mmap on,
alternating rounds to rule out page-cache ordering effects:

round	condition	pp8192 tok/s
1	unpinned	996.55 ± 1.25
1	pinned	1675.77 ± 13.32
2	unpinned	1010.15 ± 18.40
2	pinned	1666.20 ± 12.30

~1.67x on prefill. Round 2's unpinned run executes fourth, fully warm, and
still matches the cold first run — so the effect tracks the condition, not
position in the sequence.

Perplexity is unchanged: PPL = 4.1871 +/- 0.30585 with and without.

This is consistent with the ~+64% reported independently by the
fable5/host-register work on an RTX 3060.

Limitations

Stated plainly, since I can only test one configuration:

• Measured on a single GPU (A2000), a single model, and Windows only. No Linux,
  ROCm, SYCL, or multi-GPU numbers.
• Page-locking makes pages non-pageable. Registering a mapping that is large
  relative to physical RAM fails — I hit this directly: a 17.35GB mapping on a
  32GB host returns false. That path is handled (registration failure is ignored
  and the unpinned path is used unchanged), but it is the reason this is opt-in
  rather than default.
• No benefit to decode paths where experts are computed on the CPU; this only
  helps paths that actually upload weights.

Happy to add Linux and larger-model numbers, or to change the opt-in mechanism
if an env var is not the preferred interface here.

[ggml-gh-bot]

Hi @bluechiperic, thanks for your contribution!

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

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

• AI-generated content: While code is allowed to be generated by AI, please write the PR description and commit messages on your own without the help of AI.

---

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