Motivation

MLX servers fail during startup after #39731 with AttributeError: 'MlxModelRunnerStub' object has no attribute 'kv_cache_configurator'. The new capacity query reaches the Torch configurator, which the MLX stub does not create.

Modifications

Report logical token capacity from the MLX pool, keeping the request pool's capacity adjustment. Add a regression covering worker startup and the request limits reported to the scheduler.

Accuracy Tests

• The new regression fails before the fix; 30 related tests and 3 subtests pass after it on current main.
• On Apple M4, mlx-community/Qwen3-0.6B-4bit now starts and generates text. With default overlap scheduling, both radix-cache modes pass short and chunked-prefill requests, followed by a healthy /health response.
• Changed-file pre-commit checks pass. These are local results; full remote test CI still needs maintainer activation.

Speed Tests and Profiling

No model forward or sampling changes; no throughput claim.

Checklist

• Run pre-commit on the changed files.
• Add a regression that fails before the fix.
• Document the MLX capacity behavior in the property docstring.
• Provide relevant real-model startup and generation checks.
• Follow the SGLang code style guidance.

CI States

Latest PR Test (Base): ⏳ Run #36650163801
Latest PR Test (Extra): ❌ Run #36650163569
Latest PR Test (AMD ROCm 10): ⏳ Run #36650163553