feat(glm52): add checkpoint-native MTP speculative decoding

[xiaguan]

Description

Add checkpoint-native GLM5.2 MTP speculative decoding with a separate layer-78 KV lane, proposal/recycle execution, scheduler-wide collectives, and an official-vLLM oracle harness.

The integration also:

• feeds final-normalized target hidden states into MTP, matching the official vLLM model boundary;
• represents drafter selection and MTP rounds as enums so invalid combinations are unrepresentable;
• bumps the rank-host wire protocol and covers the drafter payload with a round-trip test;
• drains every rank response before returning collective errors, preserving the root cause;
• corrects the shared EP8 routed-expert order used by every GLM5.2 target MoE layer: quantize the unweighted SwiGLU activation before W2, apply route weights after W2, then apply routed scaling before adding the unscaled shared expert;
• includes official-vLLM front and layer-78 fixtures plus a full production-path slot-reuse gate;
• integrates the native-MTP lane with the latest TP4 prefill-only build and scheduler contracts.

Accuracy

• Official-vLLM MTP fixture: MLP RMS 8.62e-3, chained hidden RMS 1.87e-2, exact top-1, top-8 overlap 8/8, and top-32 overlap 30/32.
• Matched c8 mean accepted length: OpenInfer 1.753 before the hidden-boundary fix, 3.725 after it, versus official vLLM 3.786.
• The selected 251-token target trajectory improves from 1.000 to 5.795 mean accepted length; the first draft matches token 98825.
• Post-change GSM8K: plain target 195/200, native MTP 196/200. There is no identical pre-change plain-task A/B, so this is a regression check rather than a task-accuracy improvement claim.
• The latest-main 8×H200 production gate passes in 234.13 s, including a released-slot reuse request, at least 32 reuse verification rounds, and eight mixed-length rank-pinned requests.
• The layer-6 EP8 oracle passes all g64/g32/g16/g8 groupings with 62/64 probes inside tolerance per grouping and two permitted near-tie outliers.

Performance

All comparisons below use the same 8×H200 node, checkpoint, greedy sampling, fixed output length, disabled prefix cache, and matched request corpus.

Engine	Concurrency	TPOT	Mean accepted length	TPOT × accepted length
OpenInfer native MTP	1	7.749 ms	3.8762	30.04 ms
official vLLM native MTP	1	8.814 ms	3.4858	30.72 ms
OpenInfer before boundary fix	8	23.44 ms	1.753	41.10 ms
OpenInfer after boundary fix	8	11.31 ms	3.725	42.11 ms
official vLLM native MTP	8	16.90 ms	3.786	63.98 ms

The repeated c1 result shows no remaining decode TPOT deficit: OpenInfer is 1.065 ms lower, while proxy round cost differs by 0.69 ms. OpenInfer TTFT is higher (432.50 ms versus 232.57 ms), so this is a decode result, not a broad end-to-end latency win. The c8 engines use different target topologies and generation trajectories; accepted length is reported as serving evidence, not isolated MTP-forward parity.

Validation

• cargo fmt --all -- --check
• cargo check --release -p openinfer-server --features glm52
• cargo test --release -p openinfer-glm52 --lib — 86 passed, 0 failed, 20 ignored
• official-vLLM MTP front golden gate
• official-vLLM layer-78 EP8 golden gate
• latest-main 8×H200 production-path hidden-boundary, slot-reuse, and mixed-rank gate — 1 passed in 234.13 s
• latest-main 8×H200 layer-6 EP8 oracle gate — 1 passed

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update

Checklist

• My code follows the style guidelines of this project (see docs/conventions/coding-style.md).
• I have performed a self-review of my own code.
• I have formatted my commits according to Commitizen conventions.
• I have run the local test suite and all tests pass (see CLAUDE.md).

[xiaguan]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Swish!

Reviewed commit: 3b2c74353d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".