BBuf/vllm-glm5-glm51-optimization

vLLM GLM-5 / 5.1 Optimization

Overview

GLM-5 support in vLLM is not a standalone glm5.py runtime. The current landed path adapts GLM-5 into the existing DeepSeek-V2/V3 MLA/MoE implementation and then fixes the GLM-5 MTP draft-model correctness bug separately.

Evidence snapshot:

• vLLM mainline checked around 0f7be0f2f76814f80f9091220a5fbbb53912ad00
• Support status: partially supported on current mainline
• Canonical PR notes: references/pr-history.md
• History mirrors: model-pr-optimization-history/vllm/glm5-glm51/README.zh.md and README.en.md

Non-Negotiable Evidence Rule

Use skills/model-optimization/model-pr-diff-dossier/SKILL.md as the bar. For GLM-5, do not pretend support comes from the older glm4* modules; the actual landed path is the DeepSeek-based alias plus MTP follow-up.

Runtime Surfaces

• vllm/vllm/model_executor/models/deepseek_v2.py
• vllm/vllm/model_executor/models/registry.py
• vllm/vllm/config/speculative.py
• vllm/vllm/transformers_utils/model_arch_config_convertor.py
• vllm/vllm/v1/spec_decode/eagle.py
• vllm/tests/models/registry.py
• vllm/tests/models/test_initialization.py

Current Main Summary

• #34124 adds GlmMoeDsaForCausalLM as a DeepSeek-V2-derived runtime rather than introducing a dedicated GLM-5 file.
• The same PR also flips rope style through indexer_rope_interleave and teaches speculative config conversion to treat glm_moe_dsa like a DeepSeek MTP family.
• #34385 fixes GLM-5 MTP accuracy by explicitly sharing the target lm_head into every MTP layer shared_head.head; without that, logits could become zero or NaN.

Key Landed PRs

• #34124 GLM adaptation
• #34385 Fix MTP accuracy for GLM-5

Open Optimization Items

• Re-run PR search before claiming broader GLM-5 router, parser, or multimodal support beyond the checked mainline commit.

Validation Lanes

• Plain GLM-5 launch through the GlmMoeDsaForCausalLM alias.
• MTP correctness on held-out prompts after #34385.
• Rope/interleaving checks when touching the DeepSeek indexer path, because the GLM adaptation changed is_neox_style handling in deepseek_v2.py.

References

• references/pr-history.md: diff-reviewed GLM-5 / 5.1 cards.

Non-Negotiable Evidence Rule

Use skills/model-optimization/model-pr-diff-dossier/SKILL.md as the production bar. Every PR cited for this family must be based on diff reading, not only PR titles.