moe-cache: pick a GPU device explicitly instead of the first non-meta device

[ALeXssNdR]

Problem

On hosts where the model device list starts with the CPU device, the expert cache silently disabled itself with no GPU device - expert cache disabled, because init_moe_expert_cache() picked the first non-meta device regardless of its type.

While investigating I also hit two cases where the model device list contains no GPU entry at all (early loading stages / mmap path), so a registry fallback is included.

Fix

• select the first GPU-typed device from the model list
• fall back to enumerating every backend registry (including dynamically loaded ones such as libggml-cuda) and then the default device registry

Measured

RTX 5080 Laptop 16 GB, Ornith-1.5-35B-A3B (qwen35moe, 41 blocks) Q4_K_M, built with CUDA arch 120:

-ngl 99 -ncmoe 99 -fa 1 -c 4096
--moe-cache-profile ornith35-merged.csv --moe-cache-slots 176   (9.78 GiB pack)

config	decode
baseline (no cache)	56.8 tok/s
cache, 176 slots	80.0–86.4 tok/s
cache, 176 slots (short bench)	92.0 tok/s

That is +41–62% on the same hardware, bit-identical outputs, with a single slot count that fits VRAM (40 layers × 176 slots ≈ 9.8 GiB).

The routing profile was captured with llama-moe-trace (512-token code + chat prompts, merged) exactly as documented in the README.

Thanks for the great work on this fork — the video walkthrough of the LRU-vs-static tradeoff is excellent.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 427e4924-79a7-49b8-a035-7006a14ba967

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}