GSP: grammar-conditioned sparse LM-head projection - replication and SGLang integration feedback wanted

[StealthEyeLLC]

I am sharing an open research artifact that may be relevant to SGLang's structured-generation path.

GSP (Grammar-Conditioned Sparse Projection) consumes a packed grammar legality mask before/during the target-model LM-head projection instead of computing the full vocabulary projection and masking afterward.

The current artifact includes:

• packed-mask row-sparse Triton projection for latency-oriented decoding
• batched block-GEMM that skips grammar-inactive vocabulary tiles
• density-aware dense fallback for permissive/mixed states
• support/top-1 correctness checks
• raw paired end-to-end Qwen3.5-2B results

On the reference RTX 5060 Laptop GPU, 30 paired AB/BA trials gave median end-to-end ratios of 1.0756x (enum JSON), 1.0642x (tool-call JSON), and 1.0842x (nested-array JSON). The free-string control is not claimed as a reliable mean speedup.

This is not claiming first invention of batch-1 grammar-aware LM-head filtering; the contribution boundary is the packed-mask batched execution, heterogeneous routing, and controlled target-model evaluation.

I would especially value:

• SGLang integration feedback
• datacenter GPU replication
• heterogeneous-batch measurements
• negative results / dense-winning crossover cases

Release, verified preprint, source, raw evidence:
https://github.com/StealthEyeLLC/grammar-sparse-projection/releases/tag/v0.1.1