feat(glm5): add GLM5.3-Flash H100 FlagOS support

[CherryLemon]

PR Category

Core

PR Type

New Features | Improvements | Performance

Description

Add GLM5.3-Flash H100 FlagOS support for the vLLM 0.24.0 ABI. This PR is the model-specific layer on top of the shared day0-common branch; the branch is intentionally runnable as-is while the common stack is reviewed separately.

Changes

• Add GLM5.3 configuration, model, multimodal processor, and vLLM 0.24 compatibility patches.
• Add portable sparse MLA/indexer and KPool kernels, MLA prefill, safe KDA, and MHC broadcast paths.
• Make empty-build execution allocator-safe for CUDA graph capture and provide fallbacks when native DeepGEMM or stock indexer paths are unavailable.
• Route empty-build vision attention and MoE alignment through FlagGems.
• Add graph-boundary, KPool, MHC, processor, provider, and portable-path unit coverage.

Testing

• H100 TP16 service path was exercised with effective FULL_AND_PIECEWISE graph mode; mixed batches used PIECEWISE and decode batches used FULL graphs.
• Strict 1k-input/1k-output serving request completed successfully on the empty-build path.
• Added tests under tests/unit_tests/glm5_next and recorded the graph/portable compatibility coverage.

Stacking

The model delta is reviewable with day0-common...day0-glm53. Please land/rebase it after the shared day0-common changes are accepted.

Checklist

• I have run the relevant serving checks
• I have added tests for my changes (if applicable)
• I have updated the documentation (if applicable)

[github-advanced-security]

CodeQL found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.