sglang

batch_invariant_ops

[Auto Sync] Update batch_invariant_ops.py (20260221) (#19098 )

2026-02-20 18:27:31 -08:00

batch_overlap

[AMD][TBO] Fix mori ep dual stream accuracy (#19888 )

2026-03-18 02:00:55 -07:00

checkpoint_engine

update pre-commit config (#18860 )

2026-02-16 00:18:31 +08:00

compilation

fix: piecewise_cuda_graph get correct qo_indptr (#21452 )

2026-03-28 15:57:29 -07:00

configs

Fix kimi-linear launch server error (#21752 )

2026-03-31 21:07:08 +08:00

connector

Remove sync points in mamba cache + prefill cudagraph plumbing for DP (#19639 )

2026-03-02 15:03:42 +08:00

constrained

[MUSA] apply_vocab_mask support musa device (#21296 )

2026-03-25 21:00:58 -07:00

debug_utils

Add e2e demo test in dump comparator (#21031 )

2026-03-20 22:41:01 +08:00

disaggregation

[PD] Tiny cleanup after KVReceiver refactor (#21760 )

2026-03-31 21:07:57 +08:00

distributed

Support Qwen3 MoE context parallel (#18233 )

2026-03-22 01:27:20 -07:00

dllm

Decouple scheduler log printing from metrics collection (#20107 )

2026-03-07 22:09:10 -08:00

elastic_ep

[5/N] (Elastic EP) Use GPU P2P to exchange expert weights during EPLB as much as possible (#12068 )

2026-03-16 18:40:58 +08:00

entrypoints

chore: bump flashinfer version to 0.6.7 (#21422 )

2026-03-31 21:18:16 -07:00

eplb

[5/N] (Elastic EP) Use GPU P2P to exchange expert weights during EPLB as much as possible (#12068 )

2026-03-16 18:40:58 +08:00

function_call

[feat] Enhance Kimi-K2/K2.5 function call and reasoning detection (#19552 )

2026-03-19 12:57:57 -07:00

grpc

[gRPC] Extract gRPC servicer into standalone package (#20478 )

2026-03-13 09:13:29 -07:00

hardware_backend

[Bugfix][NPU] Skip FRACTAL_NZ format for MoE weights with unaligned dimensions (#21209 )

2026-03-30 23:22:17 +03:00

layers

VLM: change default mm-attention backend from triton_attn to fa4 (on blackwell) (#21595 )

2026-04-01 14:29:59 +08:00

lora

Fix added tokens config with sensible filter (#17905 )

2026-03-31 23:32:21 -07:00

managers

Fix remote weight info nnode>1 and dp>1 (#17389 )

2026-03-31 21:17:18 +08:00

mem_cache

[HiMambaTree]: Optimize mamba host lock mechanism (#21750 )

2026-03-31 21:52:24 +08:00

model_executor

Fix remote weight info nnode>1 and dp>1 (#17389 )

2026-03-31 21:17:18 +08:00

model_loader

Fix remote weight info nnode>1 and dp>1 (#17389 )

2026-03-31 21:17:18 +08:00

models

[3/n] lora moe - Support Qwen3-VL-30B-A3B-Instruct (#21469 )

2026-03-31 23:15:16 -07:00

multimodal

[bugfix]GLM-4V model (#17122 )

2026-04-01 10:37:40 +08:00

multiplex

Revert "[Scheduler] Decouple maybe_send_health_check_signal from process_batch_result" (#20259 )

2026-03-10 01:58:48 -07:00

observability

MFU metrics in Prometheus (#19395 )

2026-03-29 23:40:06 -07:00

parser

Add Mistral Small 4 (Pixtral) support (#20708 )

2026-03-18 14:15:32 -07:00

ray

Add subprocess liveness monitor to detect scheduler crashes (#18582 )

2026-03-29 00:09:13 -07:00

sampling

[Feature Restoration] repetition_penalty is essential for GLM-V models (#21258 )

2026-03-31 23:29:49 -07:00

speculative

[Feature Restoration] repetition_penalty is essential for GLM-V models (#21258 )

2026-03-31 23:29:49 -07:00

tokenizer

[Piecewise Cuda Graph] rename, refactor and add more logging (#13675 )

2025-11-21 13:28:38 +08:00

utils

chore: bump flashinfer version to 0.6.7 (#21422 )

2026-03-31 21:18:16 -07:00

weight_sync

[RL] Allow passing tensors of different dtypes for FlattenedTensorBucket (#13413 )

2025-11-19 17:24:48 +08:00

constants.py

Clean up detokenizer and remove dead multimodal_gen code (#21588 )

2026-03-28 21:44:40 -07:00

environ.py

[AMD] Add SGLANG_DISAGGREGATION_NUM_PRE_ALLOCATE_REQS env var for configurable KV transfer overlap (#20410 )

2026-03-30 14:37:16 -07:00

server_args_config_parser.py

[ConfigArgumentMerger] Improve ConfigArgumentMerger compatibility with external callers (#17051 )

2026-01-16 15:32:42 +08:00

server_args.py

CUTLASS NVFP4 GEMM improvement of SM120 (#21314 )

2026-04-01 09:04:34 +08:00