Add paged MLA/KDA serving for Ling-3.0

[FENP]

Note

This is a work-in-progress PR for early design and implementation review.
It is blocked on:

• a vLLM release containing Bailing V3 and FP8 support;
• merging MLX-LM PR #1711 and updating the pinned revision.

Please do not merge until these dependencies are available.

Summary

• Add paged serving support for Ling-3.0 Bailing V3 hybrid models.
• Support hybrid MLA and KDA layers with paged latent attention and recurrent-state caching.
• Integrate KDA with the existing hybrid recurrent paged-attention runtime for continuous batching.
• Restrict model detection to BailingMoeV3ForCausalLM and the expected KDA configuration fields.
• Document Ling-3.0 in the supported-model matrix.

Dependencies

This change has two upstream prerequisites:

• vLLM must publish a release containing:

  • Bailing V3 model registration and implementation:
    [Model][Frontend] Add Ling 3.0 Flash BF16, MTP, and parser support vllm#51045
  • Ling-3.0 FP8 loading support:
    [Model][Quantization] Add Ling-3.0-flash-fp8 support vllm#51265

  vLLM 0.27.1 contains neither change. Once a suitable release is available, install.sh must update its pinned vLLM wheel.

• MLX-LM support depends on: Add Bailing V3 support for Ling-3.0 models ml-explore/mlx-lm#1711

  Once merged, pyproject.toml must pin an upstream MLX-LM revision containing that implementation.

Validation

• 54 passed, 1 skipped in the existing focused vLLM-Metal test suites.
• git diff --check passed.
• The full Ling-3.0 Tiny-FP8 checkpoint was previously served through the paged path using local vLLM and MLX-LM source revisions containing the required Bailing V3 support.
• The engine detected native FP8 weights, 6 MLA layers, 18 KDA layers, and patched all 24 hybrid layers.
• Chat API requests for factual answering, multiplication, and translation returned the expected results with thinking disabled.

Scope

• Supports BF16 and native FP8 Ling-3.0 checkpoints through MLX-LM.
• MTP weights may be ignored during loading; this change does not execute MTP decoding.