[bugfix]Indexer init skip and MTP TopK share for iteration

[JaredforReal]

Purpose

fix GLM-5.2 BF16 init Indexer in skip_topk layer
fix GLM-5.2 MTP Port-Norm cycle

Test Plan

# setup with
VLLM_DEEP_GEMM_WARMUP=skip  vllm serve zai-org/GLM-5.2-FP8 \
  --trust-remote-code \
  --tensor-parallel-size 8 \
  --tool-call-parser glm47 \
  --enable-auto-tool-choice \
  --reasoning-parser glm45 \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 5 \
  --max-num-seqs 32 \
  --cudagraph-capture-sizes 1 2 4 8 16 32 \
  --kv-cache-dtype fp8_e4m3

Test Result

MTP Mean Acceptance Length raise form 3~ to 4~, while Mean Acceptance Rate raise to 60%
IFBench remain 74.62

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

This PR adds support for GLM-DSA MTP “post-final-norm hidden recycling” in speculative decoding and refactors sparse MLA attention to consume a shared top‑k indices buffer explicitly (enabling “skip” layers without an indexer).

Changes:

• Add mtp_recycle_post_norm plumbing across config override, draft model (MTP) forward output shape, and proposer tuple-handling logic.
• Pass topk_indices_buffer explicitly through MLA layers/attention into sparse MLA backends, decoupling them from indexer.
• Refine DeepSeek v3.2 IndexCache “skip topk” logic to always build indexers for MTP/nextn layers while optionally skipping some backbone layers.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
vllm/v1/spec_decode/llm_base_proposer.py	Detects GLM-DSA recycling and adjusts whether MTP returns tuples.
vllm/v1/attention/backends/mla/xpu_mla_sparse.py	Switches sparse MLA backend to take a shared topk_indices_buffer directly.
vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py	Same as above for ROCm aiter sparse MLA backend.
vllm/v1/attention/backends/mla/flashmla_sparse.py	Same buffer wiring change for FlashMLA sparse backend.
vllm/v1/attention/backends/mla/flashinfer_mla_sparse.py	Same buffer wiring change, removing the “indexer required” assertion.
vllm/model_executor/models/deepseek_v2.py	Updates IndexCache skip/topk pattern logic and MTP-layer exceptions.
vllm/model_executor/models/deepseek_mtp.py	Adds optional tuple return for GLM-DSA recycling behavior.
vllm/model_executor/layers/mla.py	Plumbs topk_indices_buffer into the MLA attention layer.
vllm/model_executor/layers/attention/mla_attention.py	For sparse MLA, passes topk_indices_buffer down to backend impl args.
vllm/config/speculative.py	Sets mtp_recycle_post_norm based on original GLM model_type.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.