LoginSign Up
← Back to all sparks
S

SGLang

AI-ASSISTANTS
Velocity2.5

Fast serving framework for large language and vision models

sgl-project.github.io ↗

Only patch tags reach this feed, and every one of them is frontier-model firefighting

llm-servinginferencedeepseekglmquantizationpatch-releases
◆Current state
SGLang is a serving engine for large language models, and the three entries captured here are all .post patch releases rather than feature versions. Their content is narrow and specific: GLM 5.2 failing under prefill/decode disaggregation and context parallelism, DeepSeek V4 emitting garbled text during single-token decode on B200/B300 hardware, NaN outputs from FlashInfer TRT-LLM FP4 MoE kernels on long inputs, and a FlashInfer version bump to fix its JIT cubin downloader.
◆Where it's heading
What these patches describe is the real cost of supporting frontier architectures early: each new model family brings its own interaction with speculative decoding, sliding-window KV allocation, quantised MoE kernels and disaggregated serving, and the failures surface as wrong output rather than crashes. The recurring FlashInfer dependency issues point to a kernel layer moving as fast as the models above it. Because only .post tags are captured, none of the actual feature releases appear, so this feed shows the stabilisation work and none of the shipping.
◆Prediction
Expect further .post patches tracking whichever model family lands next; a read on SGLang's feature direction isn't possible until the minor releases themselves appear in this feed.

◆Recent moves

  1. 17d ago

    Patch fixes GLM 5.2 under disaggregation and FP4 MoE NaNs

    Repairs GLM 5.2 IndexShare under both prefill/decode disaggregation and context parallelism, plus NaN outputs from FlashInfer FP4 MoE kernels on long inputs. Model-specific stabilisation across the serving paths that are hardest to get right.

    View source ↗
  2. 2mo ago

    Patch cherry-picks twelve DeepSeek V4 stability fixes

    Twelve fixes backported onto the release branch, mostly DeepSeek V4: garbled single-token decode on B200/B300 from a scale-packing path, and a sliding-window allocator assertion that crashed EAGLE/MTP disaggregated decode around 2000 requests. Failures that appear only under sustained load.

    View source ↗
  3. 3mo ago

    Patch bumps FlashInfer to fix its JIT cubin downloader

    A single dependency bump resolving a FlashInfer JIT cubin download failure. The smallest possible release, and another instance of the kernel layer driving patch traffic.

    View source ↗