Manuals / SGLang 101 / Reference · Reference
R.5

Sources

The manual ranks its sources by authority, and this page lists every external source the sections cite, with the sections that use each one.

The ranks match how this manual was made. When two sources disagree, the higher one wins and the section says so. The SGLang source files each section used are recorded in the manual source and checked by the audit. They are not listed here. Every page below was read on 2026-10-16. The studies and reports added for the research update were read on 2026-10-22.

The ranked sources

RankSourceWhat the manual takes from itSource key
1the SGLang source at the tag, baa9befaevery default and every mechanismthe source files of each section, in the manual source
2the captured runs on the Mac and on the GPUsevery listing and every number marked as measuredthe capture file name
3SGLang issues and pull requestsincidents and the design history of each featureissue and its number
4papersthe original design of RadixAttention and the methods SGLang implementspaper and a title
5the SGLang documentation at the tagdocumented behaviour, checked against the codedocs and a heading
6independent field reportsGPUs and scales that this manual could not runfield, the author, and the month
7SGLang team and vendor reportsclaims, always labeled as the claim of the authorteam or the vendor, and the month

Rank 1 and rank 2: the code and the captures

SourceUsed in
SGLang 0.5.22, tag v0.5.22, commit baa9befa381511cd4d56f83980951e5a45fffc6d, github.com/sgl-project/sglangevery section
Miles 0.1.2, tag v0.1.2, commit a22fd824, github.com/radixark/miles7.2
The capture kit and its Mac runs, recorded on 2026-10-16every section with a listing
The GPU kit and its runs, recorded from 2026-10-16 to 2026-10-21 on an NVIDIA L4 node and a DGX Spark (GB10)every section with a GPU measurement
The Miles image radixark/miles:latest at Miles commit 08752a12, with SGLang 0.5.22.dev46+g7b6993a7.2

Rank 3: pull requests

The issues are in the incident index. These pull requests explain a design change.

Pull requestWhat it changedUsed in
#9241warm-up followed by a frozen garbage collector1.2
#11210overlap with two CUDA streams in one thread2.4
#14312renamed sgl-router to sgl-model-gateway6.4
#23906moved the prefill graph field into the per-phase config4.1
#29458breakable prefill graphs on CUDA4.1
#32114deleted Cutlass MLA, Dual Chunk FlashAttention, and the plain GPTQ CUDA kernel4.2, 4.3
#37898raised the decode graph cap on 20 to 35 GB GPUs from 24 to 484.1
#35455loaded the KV scales of compressed-tensors checkpoints, merged 2026-08-204.3
#40795deleted the old benchmark aliases5.1
#41818added --attn-dp-size and gave --dp-size one meaning6.1
#42285, #42286, #42287remove the torch.compile path for prefill graphs4.1

Rank 4: papers

PaperUsed in
Zheng and others, SGLang: Efficient Execution of Structured Language Model Programs, arXiv 2312.071042.2, 3.1, 3.3, 4.5
The tail-optimized LRU policy that the code names, arXiv 2510.151523.5
Lee and others, From Tokens to Layers: Redefining Stall-Free Scheduling for MoE Serving with Layered Prefill, arXiv 2510.080552.2
Liu, Yu, and Yang, When Fancy Eviction Fails: Rethinking Cache Replacement For LLM Prefix Reuse, arXiv 2609.288703.5
Abramovich and others, SPEED-Bench: A Unified and Diverse Benchmark for Speculative Decoding, arXiv 2604.095574.4, 6.2
Kumar, Sanghavi, and Das, Test-Time Speculation, arXiv 2605.093294.4
Agrawal and others, On Evaluating Performance of LLM Inference Serving Systems, arXiv 2507.090194.4
Zhao and others, CoRun: Padding is Simple and Efficient for Deterministic LLM Inference, arXiv 2608.143764.6
Qin and others, Vosti: Specifying, Implementing, and Verifying Deterministic LLM Inference, arXiv 2609.389814.6
Patodiya, Same Request, Different Answer: Quantization Amplifies Cache-Induced Divergence in LLM Serving, arXiv 2609.047484.6
Chandrasekar and Kramberger, Identifying and Mitigating Systemic Measurement Bias in Production LLM Inference Benchmarks, arXiv 2605.242175.1
Li and others, Revisiting Disaggregated Large Language Model Serving for Performance and Energy Implications, arXiv 2601.088336.3

Rank 5: SGLang documentation at the tag

PageUsed in
docs/docs/hardware-platforms/apple_metal.mdxthe front matter, 2.1, 2.4
docs/docs/advanced_features/hyperparameter_tuning.mdx2.3
docs/docs/advanced_features/hicache_design.mdx and radix_eviction_policy.mdx3.5
docs/docs/advanced_features/attention_backend.mdx4.2
docs/docs/advanced_features/quantization.mdx4.3
docs/docs/advanced_features/speculative_decoding.mdx4.4
docs/docs/advanced_features/expert_parallelism.mdx6.2
docs/docs/advanced_features/pd_disaggregation.mdx6.3
docs/docs/get-started/install.mdx7.1
docs/docs/advanced_features/sglang_for_rl.mdx7.2

Rank 6: independent field reports

AuthorReportMonthUsed in
AIMultipleInference engines benchmark2026-01, answer 2026-101.3, 5.3
Alibaba Cloud, RBG and Mooncake teamsA production inference platform with SGLang RBG and Mooncake2026-033.5
Ant GroupDeepSeek-R1 on H20 with SGLang, on the LMSYS blog2025-094.4, 6.1, 6.2, 7.1
basbunarhasan on the NVIDIA developer forumSGLang with NVFP4 and DSpark on a DGX Spark2026-083.2
CERIT-SCRunning LLMs at CERIT-SC and a later post2025-10, 2026-047.1
CloudflareSmaller, faster, safer models2026-083.1, 4.3
CrusoeMLPerf Inference v6.1 results on AMD MI355X2026-095.3, 6.2
David Noel NgGH200 benchmarking, part 42026-084.4
DeepInfravLLM vs SGLang2026-085.3
DoublewordFast SGLang starts2026-047.1
dstackBenchmarking prefill-decode ratios2025-096.3
Hazy ResearchTensor-parallel Llama megakernel2025-094.1
InferenceX by SemiAnalysisGLM-5 FP8 on MI355X, DeepSeek-V4-Pro on MI355X in 26 days, and GLM-5 NVFP4 on B2002026-054.1, 5.3
JarvislabsSGLang vs vLLM on H1002026-054.1, 5.3
JarvislabsManaged endpoints2026-083.4, 6.4
joshua8.aiLLM inference benchmark2026-014.1, 4.3, 5.3
LinkedIn EngineeringScaling LLM-based ranking systems with SGLang at LinkedIn2026-021.2, 2.1, 7.1
Luca LowndesSGLang optimisations2025-116.1
ModalLLM almanac2025-065.3
Novita AISGLang with GLM MoE models, on Hugging Face2026-016.3
Pivot NewsA summary of a mixed Blackwell benchmark2026-054.3, 5.3
RunPodWhen to choose SGLang over vLLM2025-063.3, 5.3
SemiAnalysisInferenceMAX2025-105.3
SpheronvLLM vs TensorRT-LLM vs SGLang20265.3
Steve RawlinsonvLLM vs SGLang on dual H1002025-123.2, 5.3
Thinking MachinesDefeating nondeterminism in LLM inference2025-094.6
Varad MoreInference engine benchmark system2026-043.3, 4.2, 4.5, 5.3
vLLM communityDiscussion #172212025-045.3

Rank 7: SGLang team and vendor reports

AuthorReportMonthUsed in
The vLLM teamThe state of FP8 KV cache and attention quantization in vLLM2026-044.3
LMSYSSGLang with Llama 32024-075.3
LMSYSSGLang v0.4: zero-overhead batch scheduler2024-122.4, 6.1, 6.4
LMSYSDeepSeek with PD disaggregation and large-scale EP on 96 H100 GPUs2025-056.2, 6.3
The Mooncake team on the LMSYS blogKimi K2 with large-scale EP2025-076.3
LMSYSSGLang HiCache2025-093.5
SGLangv0.5.21 release notes2026-103.5
Google CloudEAGLE-3 on Vertex AI2025-114.4
Modal with DecagonDecagon case study2025-114.4
LMSYS and RadixArkMiles2025-117.2
RadixArkMiles v0.12026-087.2
The Miles teamdocs/advanced/p2p-weight-transfer.md at Miles v0.1.22026-107.2

SGLang 0.5.22 · baa9bef · 2026-10-13 · Edition 2026.10