What does beam search in SGLang v0.5.19 not work with?

Beam search in SGLang v0.5.19 works out of the box beside regular requests, but it does not yet mix with speculative decoding, prefill-decode disaggregation, DP attention or HiCache. Teams already running any of those four will need to keep beam requests on a separate server, or wait for a later release to combine them.

What breaks when upgrading to SGLang v0.5.19?

SGLang v0.5.19 lists six breaking changes. The unified radix tree is now mandatory and `SGLANG_ENABLE_UNIFIED_RADIX_TREE` is deprecated, Spark3 is renamed Spark2.5 across configs, DeepSeek-V4 FP4 defaults to the FlashInfer MXFP4 MoE path, FlashInfer 0.6.18 is required, ServerArgs no longer auto-resolves and needs `resolve_once()`, and requests are capped at 32 stop strings of 256 bytes each.

Which models does SGLang v0.5.19 add?

SGLang v0.5.19 adds Qwen3.8 (2.4T-A95B), Qwen3.8-27B, dots3.note, Ling-3.0-flash, Ling-3.0-tiny, Spark2.5, MiniCPM-SALA, Granite 4.2, and the LongCat-Image-Edit and Edit-Turbo diffusion models. The cookbook also gained deployment guides for GLM-5.3, PaddleOCR-VL, Kimi-K3 on Ascend A3, and MiniMax-H3 on a 24 GB GPU.

How much faster is SGLang v0.5.19 on AMD GPUs?

On MI355X, the new Lean attention kernel gives SGLang v0.5.19 up to 1.52x more throughput and up to 3.62x lower inter-token latency, and turns itself on where it helps. The v2 top-k kernel makes DeepSeek-V4 up to three times faster on ROCm, and a shared-experts gate fix gives GLM-5.2 up to 16% better time per output token.

SGLang · 2026-09-05 · major

SGLang v0.5.19 — beam search arrives, plus 786 merged pull requests

SGLang v0.5.19 adds beam search to the inference server: pass beam_width in a request and get the n best sequences back instead of one sample. The release carries 786 pull requests from 214 contributors and nine more models.

SGLang's September release adds beam search, nine more models, and an AMD attention kernel that fills idle compute units.

Quick facts

Version	v0.5.19
Released	September 5, 2026
Scope	786 PRs from 214 contributors
Headline feature	Beam search via beam_width
Required dependency	FlashInfer 0.6.18
License	Apache-2.0
Stars	35,506

What is it?

Beam search is the headline in SGLang v0.5.19. Pass `beam_width` in a request and the server returns the n best sequences instead of a single sample, and beam requests run happily next to normal ones. SGLang is an open-source serving framework for large language and multimodal models; beam search here does not yet combine with speculative decoding, disaggregation, DP attention or HiCache.

How does it work?

The cache and the kernels are where v0.5.19 changes shape. The unified radix tree becomes the prefix cache for every model configuration, not just hybrid ones. A new persistent Lean attention kernel on AMD spreads long or uneven decode batches across all compute units instead of leaving many idle, and `--enable-layernorm-sp` makes each tensor-parallel rank normalize only its own share of prefill tokens.

Why does it matter?

Serving gets cheaper without touching the model. The AMD path moves most in SGLang v0.5.19: disaggregated GLM-5.2 decode on eight MI355X cards drops from 23 ms to 8 ms per output token. On Hopper, teams serving MXFP4 experts can now quantize activations to FP8 as well and gain about 12% output throughput on DeepSeek-V4-Flash with no change in GSM8K accuracy.

Who is it for?

inference and platform teams

Frequently asked questions

What does beam search in SGLang v0.5.19 not work with?: Beam search in SGLang v0.5.19 works out of the box beside regular requests, but it does not yet mix with speculative decoding, prefill-decode disaggregation, DP attention or HiCache. Teams already running any of those four will need to keep beam requests on a separate server, or wait for a later release to combine them.
What breaks when upgrading to SGLang v0.5.19?: SGLang v0.5.19 lists six breaking changes. The unified radix tree is now mandatory and `SGLANG_ENABLE_UNIFIED_RADIX_TREE` is deprecated, Spark3 is renamed Spark2.5 across configs, DeepSeek-V4 FP4 defaults to the FlashInfer MXFP4 MoE path, FlashInfer 0.6.18 is required, ServerArgs no longer auto-resolves and needs `resolve_once()`, and requests are capped at 32 stop strings of 256 bytes each.
Which models does SGLang v0.5.19 add?: SGLang v0.5.19 adds Qwen3.8 (2.4T-A95B), Qwen3.8-27B, dots3.note, Ling-3.0-flash, Ling-3.0-tiny, Spark2.5, MiniCPM-SALA, Granite 4.2, and the LongCat-Image-Edit and Edit-Turbo diffusion models. The cookbook also gained deployment guides for GLM-5.3, PaddleOCR-VL, Kimi-K3 on Ascend A3, and MiniMax-H3 on a 24 GB GPU.
How much faster is SGLang v0.5.19 on AMD GPUs?: On MI355X, the new Lean attention kernel gives SGLang v0.5.19 up to 1.52x more throughput and up to 3.62x lower inter-token latency, and turns itself on where it helps. The v2 top-k kernel makes DeepSeek-V4 up to three times faster on ROCm, and a shared-experts gate fix gives GLM-5.2 up to 16% better time per output token.

Try it

pip install sglang