Baizhou Zhang 6ea491e439 Overlap shared experts with deepep dispatch for single batch overlap on Blackwell (#17289)

2026-01-21 02:56:55 +08:00

..

batch_invariant_ops

…

batch_overlap

Overlap shared experts with deepep dispatch for single batch overlap on Blackwell (#17289)

2026-01-21 02:56:55 +08:00

checkpoint_engine

…

compilation

tiny refactor pcg split op registration (#16863)

2026-01-10 23:45:28 +08:00

configs

Disable swa memory for trtllm-mha backend (#17429)

2026-01-20 23:55:00 +08:00

connector

…

constrained

Fix grammar sync across TP ranks (#17100)

2026-01-15 18:38:01 -08:00

debug_utils

Support multiple engines for router simulation in schedule simulator (#16368)

2026-01-04 11:46:05 +08:00

disaggregation

[PD] Optimize MHA models pp util calculation logic (#17306)

2026-01-19 11:23:15 +08:00

distributed

[3/N] Achieve fault tolerance at the DP level (#11657)

2026-01-20 18:47:08 +08:00

dllm

[DLLM] Implement initial dynamic batching for diffusion LLM (#14883)

2026-01-17 16:48:15 +08:00

elastic_ep

…

entrypoints

[New Model] GLM4.7-Flash (#17247)

2026-01-20 23:44:16 +08:00

eplb

Re-introduce the unit test of test_mooncake_ep_small (#16019)

2026-01-12 17:01:24 +08:00

function_call

fix function calling for Trinity (#17364)

2026-01-19 18:33:17 -08:00

grpc

[gRPC] Add GetLoads RPC for comprehensive load metrics (#17087)

2026-01-14 12:05:10 -08:00

hardware_backend/npu

[NPU]Support GPT-OSS for NPU (#14197)

2026-01-19 04:13:41 +08:00

layers

Use attn_tp_group for all reduce in token embedding (#17403)

2026-01-20 23:56:21 +08:00

lora

[Feature] overlap LoRA weight loading with compute (#15512)

2026-01-19 10:43:17 +08:00

managers

Add flag that enables NCCL mlp sync batch for overlap scheduler (#17288)

2026-01-20 23:06:55 +08:00

mem_cache

[RadixTree][1/N Refactor]: Support unified match_prefix params (#17142)

2026-01-19 22:39:40 +08:00

metrics

Code clean up for fp8 quantization (#16982)

2026-01-13 12:38:39 -08:00

model_executor

[FIX] fix mambaish model pp kv cache compute (#17334)

2026-01-20 23:59:05 +08:00

model_loader

ci: enable offline mode when local cache is complete to avoid HF Hub … (#16121)

2026-01-15 20:15:33 -08:00

models

Overlap shared experts with deepep dispatch for single batch overlap on Blackwell (#17289)

2026-01-21 02:56:55 +08:00

multimodal

[VLM][Reland] Refactor load_mm_data to improve performance (#16152)

2026-01-18 14:11:17 +08:00

multiplex

Fix deprecation warning of sre_parse for python 3.13 (#16247)

2025-12-31 21:25:09 -08:00

parser

VLM: enhance VL embedding model with video input support and revise warm-up strategy (#16635)

2026-01-08 22:12:01 +08:00

sampling

Revert "feat: reduce constrained-decoding overhead in TP" (#16845)

2026-01-10 11:39:38 +08:00

speculative

Evict swa kv cache during decoding (#17220)

2026-01-19 22:36:52 +08:00

tokenizer

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

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

tracing

SGLang Tracing: fix attribute errors (header extraction & bootstrap span closing) (#15693)

2025-12-27 22:44:31 -06:00

utils

Update default attn backend to fa3 for mimo model (#17419)

2026-01-20 21:22:27 +08:00

weight_sync

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

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

constants.py

…

environ.py

Overlap shared experts with deepep dispatch for single batch overlap on Blackwell (#17289)

2026-01-21 02:56:55 +08:00

server_args_config_parser.py

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

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

server_args.py

Disable swa memory for trtllm-mha backend (#17429)

2026-01-20 23:55:00 +08:00