# SGLang v0.5.19

- Product: SGLang (https://whatsnew.fyi/product/sglang)
- Vendor: LMSYS
- Date: 2026-09-05
- Version: v0.5.19
- Original notes: https://github.com/sgl-project/sglang/releases/tag/v0.5.19
- Permalink: https://whatsnew.fyi/product/sglang/releases/v0.5.19

What's New is an index, not a publisher: every entry below links to the vendor's own release notes, which are the authoritative source. Entries are labelled where they are hand-curated sample data, pre-releases, or drawn from a secondary source such as a developer blog.

Reuse: the summaries, labels and curation here are © What's New. Quote freely with attribution and a link back; wholesale republication of the corpus is not permitted — terms: https://whatsnew.fyi/terms. The vendors' own release notes remain their publishers'.

---

- **added** — Support for beam search with configurable beam_width parameter to return the n best sequences
- **added** — DeepEP v2 ElasticBuffer engine available as --moe-a2a-backend deepep_v2 for DeepSeek-V3/V4 and Qwen3-MoE in FP8
- **added** — LayerNorm sequence parallelism with --enable-layernorm-sp flag to reduce prefill latency for dense Qwen3 models
- **added** — W4A8 MoE quantization on Hopper with --flashinfer-mxfp4-moe-precision fp8 option
- **added** — Decode context parallelism support on trtllm_mla default Blackwell MLA backend
- **added** — Opt-in fused accept path for KDA models with SGLANG_OPT_KDA_FUSED_ACCEPT_STATE=1 environment variable
- **added** — Ability to attach or detach L3 storage on a running server with unified radix tree cache
- **added** — Persistent Lean attention kernel on AMD MI300X and MI355X that automatically optimizes decode batch processing
- **added** — Support for Qwen3.8 (2.4T-A95B) autoregressive model
- **added** — Support for Qwen3.8-27B autoregressive model
- **added** — Support for dots3.note autoregressive model
- **added** — Support for Ling-3.0-flash and Ling-3.0-tiny autoregressive models
- **added** — Support for Spark2.5 autoregressive model
- **added** — Support for MiniCPM-SALA autoregressive model
- **added** — Support for Granite 4.2 autoregressive model
- **added** — Support for LongCat-Image-Edit and LongCat-Image-Edit-Turbo diffusion models
- **changed** — Unified radix tree is now the default cache for all models instead of only hybrid models
- **changed** — DSA prefill top-k kernel moved to v2 version for improved performance on B200
- **fixed** — PD decode workers can now reuse cached prefixes for SWA hybrid models like gpt-oss
- **fixed** — Pipeline parallelism with HiCache L3 now stays consistent across ranks

#### Highlights

*786 PRs from 214 contributors.*

**New models in this release** (see the [cookbook](https://docs.sglang.io/cookbook) for all supported models):

| Model | Type | PRs | Cookbook |
|---|---|---|---|
| Qwen3.8 (2.4T-A95B) | Autoregressive | [#35758](https://github.com/sgl-project/sglang/pull/35758) ⭐ | [link](https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8) |
| Qwen3.8-27B | Autoregressive | [#34859](https://github.com/sgl-project/sglang/pull/34859) | [link](https://docs.sglang.io/cookbook/autoregressive/Qwen/Qwen3.8-27B) |
| dots3.note | Autoregressive | [#33829](https://github.com/sgl-project/sglang/pull/33829) ⭐ | [link](https://docs.sglang.io/cookbook/autoregressive/RedNote/Dots3-Note) |
| Ling-3.0-flash | Autoregressive | [#33561](https://github.com/sgl-project/sglang/pull/33561) | [link](https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.0-flash) |
| Ling-3.0-tiny | Autoregressive | [#33561](https://github.com/sgl-project/sglang/pull/33561) | [link](https://docs.sglang.io/cookbook/autoregressive/InclusionAI/Ling-3.0-tiny) |
| Spark2.5 | Autoregressive | [#35963](https://github.com/sgl-project/sglang/pull/35963) ⭐ | |
| MiniCPM-SALA | Autoregressive | [#30360](https://github.com/sgl-project/sglang/pull/30360) | |
| Granite 4.2 | Autoregressive | [#36286](https://github.com/sgl-project/sglang/pull/36286) | [link](https://docs.sglang.io/cookbook/autoregressive/IBM/Granite-4.2) |
| LongCat-Image-Edit & Edit-Turbo | Diffusion | [#35829](https://github.com/sgl-project/sglang/pull/35829) | |

**Cookbook updates:**

* GLM-5.3 deployment guide: [link](https://docs.sglang.io/cookbook/autoregressive/GLM/GLM-5.3)
* PaddleOCR-VL deployment guide: [link](https://docs.sglang.io/cookbook/autoregressive/Baidu/PaddleOCR-VL)
* Kimi-K3 on Ascend A3: [#35508](https://github.com/sgl-project/sglang/pull/35508)
* Kimi-K2.7-Code-MXFP4 on MI355X: [#36246](https://github.com/sgl-project/sglang/pull/36246)
* Qwen3.5 MXFP4 on MI355X with an FP8 KV cache or a HiCache host-memory tier: [#35445](https://github.com/sgl-project/sglang/pull/35445), [#36245](https://github.com/sgl-project/sglang/pull/36245)
* MiniMax-H3 on a 24 GB GPU or DGX Spark, with a consumer-GPU tuning guide: [#36169](https://github.com/sgl-project/sglang/pull/36169), [#35816](https://github.com/sgl-project/sglang/pull/35816)
* Ling-3.0-flash on DGX Spark: [#36364](https://github.com/sgl-project/sglang/pull/36364)
* Qwen3.8-27B on RTX 5090, RTX PRO 6000, and DGX Spark, re-measured: [#35825](https://github.com/sgl-project/sglang/pull/35825)

**Beam search.** SGLang can now do beam search. Pass `beam_width` in your request and you get back the `n` best sequences instead of a single sample. It works out of the box next to regular requests, though it does not yet mix with speculative decoding, disaggregation, DP attention, or HiCache ([#31626](https://github.com/sgl-project/sglang/pull/31626)).

**DeepEP v2.** DeepEP's new ElasticBuffer engine is available as `--moe-a2a-backend deepep_v2` for DeepSeek-V3/V4 and Qwen3-MoE in FP8. Its buffers have a fixed size, so decode can run under CUDA graphs even across nodes. Performance is on par with the classic backend ([#35634](https://github.com/sgl-project/sglang/pull/35634), [#34923](https://github.com/sgl-project/sglang/pull/34923)).

**LayerNorm sequence parallelism.** With `--enable-layernorm-sp`, each tensor-parallel rank normalizes only its own share of the prefill tokens instead of all of them. That takes 3.5% off Qwen3-8B prefill on H100 and 5.6% on B200, and the saving grows with the TP degree. Dense Qwen3 models only for now ([#30915](https://github.com/sgl-project/sglang/pull/30915)).

**W4A8 MoE on Hopper.** If you serve MXFP4 experts on Hopper, you can now quantize the activations to FP8 as well with `--flashinfer-mxfp4-moe-precision fp8`. DeepSeek-V4-Flash gains about 12% output throughput with no change in GSM8K accuracy. Needs FlashInfer 0.6.18 ([#34967

_[Truncated at 4000 characters — full notes: https://github.com/sgl-project/sglang/releases/tag/v0.5.19]_