SGLang 0.5.20 ships exact sampling masks for reproducible RL rollouts, a unified radix tree that preserves sliding-window state at branch points, and a CPU-only simulator that runs the real scheduler and cache logic. On a DeepSeek-V4-Flash shared-prefix test, the new cache raised token hits from 43.8% to 60.8% and cut mean TTFT from 1.57 to 1.07 seconds. The release also makes Responses API storage opt-in and adds production support for several new hybrid and diffusion models.
SGLang 0.5.22 speeds model loading and fixes recurrent state corruption
2026-09-19Verified announcementTRIAL
SGLang 0.5.22 adds an InstantTensor loader that the project reports as 10× faster, broader model and hardware support, and new paths for speculative decoding and prefill/decode disaggregation. It also fixes a correctness bug where hybrid linear models could resume from the wrong Mamba state when speculative decoding and structured output were combined. Python 3.11 is now required, and all performance gains remain project-reported.
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For inference teams, the state-corruption fix matters more than the headline throughput numbers because it can silently produce wrong output under a specific feature combination. Test that path first, then reproduce loading, TTFT and throughput gains on your own model mix before upgrading production.
Evidence & related material
SGLang 0.5.21 supports runtime prefill/decode role switching and enables the Rust unified cache core by default.
SGLang 0.5.22 expands loading and disaggregated serving and fixes a wrong-Mamba-state resume path under speculative decoding plus structured output.
Subsequent changes
Version 0.5.21 adds runtime role switching, the default Rust cache core and new decision APIs.
Version 0.5.22 expands serving paths and fixes recurrent-state correctness
Technical details & scores
| Version | 0.5.22 stable; 950 PRs from 245 contributors |
|---|---|
| Requirements | Python 3.11 or newer |
| Loading | InstantTensor claimed 10x faster by the project |
| Serving | prefill/decode disaggregation up to 27% more throughput and 48.9% lower p95 TTFT in a project test |
| Model Paths | DeepSeek-V4.1 up to 2.3x H200 throughput; Apple Silicon standard PyTorch backend; additional diffusion and multimodal models |
| Correctness Fix | wrong Mamba state could resume when speculative decoding and structured output were combined |
| Evidence Limit | performance figures are project benchmarks and need workload-specific reproduction |
Technical impact 4
Engineering value 5
Adoption signal 5
Maturity 4
Verification cost 2
Risk 2