Warn when --gpu-layers puts MoE layers on the CPU; mark #176 fixed in README

[solderzzc]

Summary

• Print a startup warning when --gpu-layers puts MoE layers on the CPU. The CPU runs the quantized expert matmuls on a single core, so it is very slow: ~0.4 tok/s prefill and ~3 s per token on Gemma 4 26B-A4B, on both M5 and M6. The warning recommends --stream-experts instead.
• README: the --gpu-layers timeout (--gpu-layers N crashes the server with a Metal GPU timeout on the first request #176) is fixed in chore: bump mlx-swift for compile scoped-device fix (#176) #177. The MoE speed caveat stays, and the note points 32 GB Macs to --stream-experts first.

Test plan

• Gemma 4 26B-A4B with --gpu-layers 23 prints 7 MoE layers will run on the CPU ...
• Full-GPU runs print no warning

🤖 Generated with Claude Code

[solderzzc]

Cross-review (M6 session): no blocking issues. OK to merge once CI is green.

Reviewed the full diff (Server.swift +8, README.md):

• The warning is gated on modelIsMoE && cpuCount > 0, so a FULL GPU load (no split) prints nothing. modelIsMoE comes from the same ModelProfile the partition plan uses. When the profile isn't available (first-run download, modelDirectory == nil), partitionPlan is also nil, total == actual and cpuCount == 0, so no warning is lost relative to the split message itself.
• README: the --gpu-layers note now says "fixed in chore: bump mlx-swift for compile scoped-device fix (#176) #177" and keeps the MoE CPU speed caveat and the --stream-experts recommendation. Merge note: docs: correct M6 TurboKV data — #175 affects every chip, not just M5 #180 rewrites the TurboKV line just above, with one unchanged line in between. Git should merge the two cleanly, but whichever lands second may want a quick look.

Verified on a Mac mini M6 (32 GB), branch 60c0bf4:

• gemma-4-26b-a4b-it-4bit --gpu-layers 23 prints ⚠️ 7 MoE layers will run on the CPU, which is very slow (expect well under 1 tok/s). and For MoE models that don't fit in GPU memory, prefer --stream-experts over --gpu-layers.
• The same model without --gpu-layers (FULL GPU): no warning.
• Clean build, no new compiler warnings.