v0.1.1

SGLang-Omni 0.1.1 is the first release available on PyPI, providing a stable version for deployment configurations to pin.

Install it with:

pip install sglang-omni==0.1.1

This release uses SGLang 0.5.16 and Transformers 5.12.1. It also keeps Descript DAC dependencies model-specific and includes updated talker role mappings for Qwen3-TTS and Higgs.

What's Changed

• [Perf] Compile MOSS-TTS Local frame sampler before CUDA graph capture by @Ratish1 in #773
• Fix MOSS local output collection without stream outbox by @Ratish1 in #787
• [TTS] Add OpenAI-compatible speech API by @Ratish1 in #659
• [Perf] Increase MOSS-TTS Local reference audio cache item margin by @SandyLuXY in #788
• fix: resolve tensor device mismatch and CI artifact glob in Video-AMM… by @MelodyyyYin in #767
• feat(qwen3-omni): enable mixed-chunk by default on the thinker path (… by @luojiaxuan in #789
• Run Omni unit CI first and fix profiler route test by @Ratish1 in #790
• [CI] Update Calibration Skills by @zhaochenyang20 in #793
• [CI] CI Host Profiles by @zhaochenyang20 in #794
• [CI] Add Pre Check List by @zhaochenyang20 in #795
• [CI] Calibration 2026/06/15 by @zhaochenyang20 in #796
• [TTS] Add uploaded voice APIs by @Ratish1 in #662
• [Qwen3-Omni] benchmarks: concurrency / rollout stress runner by @luojiaxuan in #770
• feat(ci): add MOSS TTS to TTS CI with random model selection (issue #724) by @yijiangt in #727
• [CI]: Add luojiaxuan to CI permissions by @luojiaxuan in #806
• [MOSS]: Tune MOSS Local colocated AR memory budgeting by @Ratish1 in #810
• [CI] Enable Labels for TTS CI by @zhaochenyang20 in #814
• [Higgs] Fuse top-k/top-p sampling with sgl_kernel renorm by @BBuf in #816
• [RL] distributed weight-sync by @yxs in #784
• [CI] Support targeted TTS CI rerun labels by @Ratish1 in #817
• [Perf] MOSS-TTS-Local: CUDA-graph the streaming vocoder codec decode (default-on, bit-identical) by @JiaxinD in #798
• docs: add MOSS-TTS-Local cookbook by @MelodyyyYin in #792
• [Docs] Fix MOSS-TTS-Local streaming command format by @SandyLuXY in #819
• [RL] add Miles-compatible /generate rollout endpoint by @yxs in #785
• feat(sampling): Framework-level refractor to unify the seed in sampling API->model runner by @JingwenGu0829 in #824
• [CI] Rerun latest failed workflow per event by @Ratish1 in #827
• [CI] Add Ccyest to CI permissions by @Ccyest in #828
• [MOSS]: Use thread_local capture_error_mode for CUDA graph by @AkazaAkane in #826
• [TTS] Add batch and WebSocket speech APIs by @Ratish1 in #665
• [Ming]: Use thread_local capture_error_mode for CUDA graph by @AkazaAkane in #830
• Extract Qwen3-ASR mel at true audio length instead of padding to 3000 mel frames by @0xjeffro in #833
• [Higgs TTS] Raise AR server default to 64 and expose standard batch knobs by @estellaliu233 in #756
• [Docs] Internalize RFC comments + consolidate historical RFCs (#488 part 4.1 + 4.2) by @JiaxinD in #537
• [Qwen3-Omni] TP correctness by @luojiaxuan in #771
• [TTS] Validate generation batch policies (#836 W1) by @luojiaxuan in #843
• [CI] Restart latest Omni CI for targeted reruns by @Ratish1 in #857
• [MOSS]: Add SGLang-backed MOSS vocoder decoder by @Ratish1 in #822
• [CI] Calibration 2026/06/22 by @zhaochenyang20 in #863
• [MOSS] Raise colocated codec_mem_reserve 0.05->0.15 for CI VRAM headroom by @JiaxinD in #864
• [MOSS] Decouple MOSS-TTS Local codec from processor by @Ratish1 in #874
• [CI] Expand Stage 1 to cover full set of Seed TTS on ASR by @zhaochenyang20 in #880
• [TTS] Standardize generation batch conventions (#836) by @luojiaxuan in #850
• [MOSS] Capture all streaming vocoder CUDA graph step lengths by @Ratish1 in #886
• [Qwen3-ASR] Parallelize request building by @codestrongestx in #885
• [CI] Update Threshold For Qwen3 ASR and Moss Local by @zhaochenyang20 in #887
• Optimize Qwen3-TTS predictor sampling hot path by @luojiaxuan in #871
• [MOSS] Decouple streaming vocoder first-chunk default from the coalesce join floor by @JiaxinD in #888
• [CI] Migrate Qwen3-Omni CI from H20 to 2×H100 by @yxs in #883
• [CI] Remove decommissioned H20 host profile and stale repro-host doc ref by @yxs in #901
• [CI] Clean Up Calibration Docs by @zhaochenyang20 in #902
• [Qwen3-Omni] Raise talker_ar default max_running_requests 16->32 by @edwingao28 in #847
• [bigfix] Ming-Omni Streaming TTS all-segment overlap bugfix by @AkazaAkane in #866
• [Realtime]: define audio buffer overflow error by @YzXiao101 in #706
• feat(higgs/rl): add rollout logprob + delay-pattern action-mask kernels by @JingwenGu0829 in #823
• [Qwen3]: Remove redundant Qwen3 Omni deepstack clone by @Ratish1 in #909
• perf(serve): gate thinker custom all-reduce on P2P/NVLink topology by @luojiaxuan in #783
• [Perf] Moss streaming chunk coalescing by @edwingao28 in #892
• T1: shared pipeline-state base + all 6 TTS migrations (+ quick-wins) by @Hayden727 in #807
• [Qwen3-Omni] Lazy TensorRef path for large inter-stage tensors (#797) by @luojiaxuan in #808
• [Perf] Defer Omni stage factory imports to worker startup by @stephenkgli in #900
• [CI] Warm the TTS-stage ASR the same way as the standalone gate by @db-ol in #911
• [Logging]: Use f-strings for simple logging calls by @Ratish1 in #922
• [Misc]: Use direct access for strict fields by @Ratish1 in #928
• Support MOSS-Transcribe-Diarize model by @CloudRipple in #914
• Refactor audio utilities for ASR models by @CloudRipple in #947
• Add MOSS diarization unit test package marker by @Ratish1 in #950
• [CI] Reuse ASR SeedTTS benchmark in CI by @zhaochenyang20 in #949
• [CI] Rootless Runner For reclaiming by @zhaochenyang20 in #952
• [CI] Add MOSS-Transcribe-Diarize ASR CI stages by @zhaochenyang20 in #955
• [TTS Refactor][M2] TTS engine factory: Extract shared TTS engine builder by @SandyLuXY in #923
• [TTS Refactor][M8] Model capabilities: Add static TTS model capability metadata by @SandyLuXY in #957
• [CI] Unify ASR Eval Script with CI by @zhaochenyang20 in #963
• [CI] Clean up GPU state after ASR stage startup failure by @SandyLuXY in #958
• refactor(moss-td): fold per-item encoder loop into a single forward in get_audio_feature by @draftbk in #971
• [ASR] Default transcription requests to greedy decoding by @Ratish1 in #969
• [router] Stream proxied responses instead of buffering (Phase 0 of #907 router optimization) by @JiaxinD in #920
• docs: whisper_asr temperature default reflects #969 by @Ccyest in #974
• [ASR] Eval Update by @zhaochenyang20 in #980
• [router] Fix in-flight counter leak on mid-stream upstream failure by @JiaxinD in #931
• [router] Auto-size --max-connections to the worker pool by @JiaxinD in #938
• [Qwen3-Omni] Async decode (one-step lookahead) for the thinker text decode by @JiaxinD in #908
• [TTS Refactor][M4a] ReferenceEncodeService: base + FishAudio S2-Pro migration by @luojiaxuan in #926
• [CI] Cover Long Sequence of ASR for MTD by @zhaochenyang20 in #982
• [TTS Refactor][T5a]Batch Vocoder Base (non-streaming) by @AkazaAkane in #940
• Encoder LRU cache for MOSS-Transcribe-Diarize: +17% throughput under load, +28.6% ceiling (M-PR3) by @xinlij in #972
• [TTS Refactor][M4a] ReferenceEncodeService: Qwen3-TTS taxonomy migration by @luojiaxuan in #927
• [ASR] Update MTD sampling defaults by @zhaochenyang20 in #989
• [Perf] Async decode (one-step lookahead) for MOSS-TD; fix lookahead overrun KV slot leak by @Ccyest in #966
• [CI] Let Omni CI inherit runner GPU assignment by @yxs in #987
• [ASR] Default Sampling with Greedy Search by @zhaochenyang20 in #993
• [CI] Consolidate AS Shell Threshold by @zhaochenyang20 in #995
• [Docs]: Revamp SGLang-Omni README by @Ratish1 in #997
• [Feat] Add MMAU and MMAR audio benchmarks under audio understanding by @JingwenGu0829 in #897
• [Docs] Add cookbook for MTD by @zhaochenyang20 in #1001
• [Cookbook] ACK for MTD Model by @zhaochenyang20 in #1002
• Fix typo in SGLang Omni Team member name by @xinlij in #1004
• [Qwen3-Omni] Capture talker predictor decode with CUDA graph by @Hayden727 in #1007
• [Refactor] T7: extract shared PreparedRequestQueue (#661) by @MelodyyyYin in #903
• [higgs] Enable --mem-fraction-static for the tts_engine stage by @JiaxinD in #977
• [Perf] Add stream output builder for MOSS-Transcribe-Diarize model by @guozhihao-224 in #965
• [Async]: Fix async decode stale batch KV slot ownership by @Ratish1 in #1010
• [Docs] Add architecture and optimization principles to MOSS-TD cookbook by @Ccyest in #1003
• encoder cuda graph for Moss TD by @yijiangt in #973
• perf(qwen3-omni): drop unused deepstack embeds from the talker projection by @MelodyyyYin in #953
• [TTS Refactor][M6] StreamingVocoderBase: base + MOSS-TTS-Local migration by @GaokaiZhang in #936
• [TTS Refactor][M9] Fish -> OmniScheduler: migrate the AR stage, delete FishScheduler by @GaokaiZhang in #937
• [CI] Trim Calibration Skills with New Grok Model by @zhaochenyang20 in #1021
• fix(tts): MOSS-TTS Local non-streaming FlashAttention and RoPE caching by @CloudRipple in #1017
• [CI] Extend MTD CI for MOSS_TD_STREAM_N_ABOVE_50_CER_MAX by @zhaochenyang20 in #1028
• [Docs]: Add MOSS-TD to README by @Ratish1 in #1029
• Add fast path for WAV audio decoding by @AkazaAkane in #1031
• [Serving] Add same-GPU DP with CUDA MPS guide and KV token cap by @JiaxinD in #986
• [TTS Refactor][T5b] StreamingVocoderBase: Higgs migration by @GaokaiZhang in #939
• feat(moss-td): opt-in torch.compile for the Whisper encoder by @draftbk in #979
• [Quant]: Add unified quantization support for AutoRound W4A16 and FP8 by @lvliang-intel in #722
• [MPS-DP]: Enforce equal KV capacity for same-GPU DP by @Ratish1 in #1037
• [Docs] Add same-GPU DP concept figures and common questions by @JiaxinD in #1042
• [TTS Refactor] Prune migrated-surface duplicate unit tests (moss / moss-local / qwen3-tts) by @luojiaxuan in #1019
• [Bench] Pre-encode MMMU images before request dispatch by @Ratish1 in #1059
• fix(moss-td): drop reduce-overhead from encoder torch.compile (illegal memory access) by @Ccyest in #1046
• [Benchmark] Fix TTS serving stress scenario contracts by @Ratish1 in #889
• [Perf] Isolate Fish S2-Pro reference encoding from GPU stages by @Hayden727 in #1047
• [TTS Refactor] Use upstream Transformers Higgs tokenizer by @SandyLuXY in #1057
• [CI] Align Qwen3-Omni gates with H100 calibration report by @Hayden727 in #1072
• [Perf] Bound Qwen3-Omni colocated stage CPU thread pools by @Hayden727 in #1060
• [TTS Refactor][M4a] Higgs: route ad-hoc reference codes through ReferenceEncodeService by @luojiaxuan in #1051
• [TTS Refactor] Simplify Fish S2-Pro inference prompt encoding by @SandyLuXY in #1058
• [TTS Refactor] Remove unreachable implementation code by @luojiaxuan in #1069
• [Perf] Reuse idle vocoder lanes for MOSS-TTS Local streaming by @Hayden727 in #1055
• [MPS]: Add H100 and H200 MPS DP hardware profiles by @Ratish1 in #1076
• [TTS Refactor] Declarative wire serialization for TTS pipeline states by @luojiaxuan in #1050
• [TTS Refactor] Load Fish Fast Audio Decoder Directly by @SandyLuXY in #1083
• [TTS Refactor] Remove TTS engine builder forwarding by @luojiaxuan in #1068
• [Feat] Add support for Fun-ASR (https://arxiv.org/abs/2509.12508)[new] by @PoTaTo-Mika in #1078
• feat(relay): intra-node NVLink CUDA-IPC data plane + centralized transport by @JingwenGu0829 in #869
• Optimize Higgs TTS streaming code collection by @Hayden727 in #1008
• [CI] Recalibrate Qwen3-Omni speed gates for #869 (NVLink CUDA-IPC transport) by @yxs in #1087
• [TTS Refactor] Unify Omni example launchers by @luojiaxuan in #1070
• [Bugfix] Avoid unretrieved completion-future errors for streaming requests by @AkazaAkane in #1082
• [Model] Add Ming-Omni-TTS 16.8B standalone TTS support by @YzXiao101 in #858
• [Perf] Overlap MOSS-TD audio encoding with LM compute (pre-LM encoder service) by @Ccyest in #1045
• [TTS Refactor] Reduce reference encode hook boilerplate by @luojiaxuan in #1096
• [CI] Replace Qwen3-ASR with Fun-ASR in ASR stage 2 and refresh ASR perf numbers on H100 by @db-ol in #1093
• [TTS] Land Audar-TTS V1 Turbo on main by @luojiaxuan in #1104
• [Perf][Higgs TTS] Prefill admission coalescing + out-of-process vocoder + compiled codec decode by @MayDomine in #1071
• [CI] Use direct Hugging Face Xet downloads by @Ratish1 in #1107
• [router] Overload protection: admission control + unified --max-connections semantics by @JiaxinD in #1014
• [TTS Refactor] Adopt shared state serialization and checkpoint/config reuse for Ming-Omni-TTS by @luojiaxuan in #1103
• [Ming-Omni] Offload mel preprocessing from event loop by @AkazaAkane in #1013
• [CI] Register Google Time Dataset to ASR CI by @zhaochenyang20 in #1127
• Perf(fun-asr): enable async decoding in Fun-ASR executor by @guozhihao-224 in #1086
• [CI] Tighten Higgs TTS quality gates from post-#1071 recalibration by @JiaxinD in #1133
• [Ming-Omni] talker duration guard fix by @AkazaAkane in #1016
• fix(client): correct GenerateRequest prompt type annotation by @cuzmi in #1064
• [CI]: Restart current Omni CI for tag-and-rerun by @Ratish1 in #1123
• [Fun-ASR] Add pre-LM audio encoder service with caching (FPR-3) by @AkazaAkane in #1095
• [RL] Add processed multimodal input contract by @JingwenGu0829 in #1129
• [CI] Add TTS serving benchmark stage by @Ratish1 in #1085
• [Fun-ASR] SANM encoder torch.compile with dynamic=True (F-PR4) by @WilsonZheng0327 in #1092
• [Qwen3-Omni Perf] Talker AR loop: sampling-state reuse by @AkazaAkane in #1043
• [Perf][MOSS-TTS] Restrict MOSS-TTS Delay audio-state text sampling by @CloudRipple in #1158
• [Qwen3-Omni Perf] Fix async drop-stale KV slot reslice for extend/mixed batches by @edwingao28 in #1027
• [Model] Add Zyphra ZONOS2 (MoE TTS) support on shared TTS surfaces by @luojiaxuan in #1112
• [Benchmark] Extract WER outlier threshold into a named constant by @Peterren in #1108
• [MPS-DP] Same-GPU DP weight sharing over CUDA IPC + autodp auto-sizing launcher by @yl3469 in #1124
• [Fix][TTS][MOSS] Support all MOSS-TTS Local v1.5 languages by @CloudRipple in #1181
• [Qwen3-Omni Perf] exact-shape CUDA Graph replay for Code2Wav by @edwingao28 in #1101
• [Qwen3-Omni Perf] Raise thinker max_running_requests default to 64 by @edwingao28 in #1135
• [ZONOS2] Release decode state at request termination by @Ratish1 in #1190
• feat: add model-free GPU diagnostics by @Dayuxiaoshui in #1141
• [Qwen3-Omni Perf] Reduce talker per-frame host overhead by @edwingao28 in #1167
• [Qwen3-Omni Perf] Bounded wait/floor batching for Code2Wav vocoder by @edwingao28 in #1126
• [Fun-ASR] Add pad+mask batched audio encoding in get_audio_feature by @guozhihao-224 in #1109
• [Perf] CUDA-graph the Qwen3-TTS code-predictor chain: ~1.9x at c8, ~3x at c32 by @JiaxinD in #1134
• [TTS] Improve Fish S2-Pro streaming continuity by @luojiaxuan in #1207
• [Perf] Add stream output builder for Fun-ASR model by @keke0315 in #1094
• [Benchmark][ASR] Report RTFx and repeat coverage by @Ratish1 in #1203
• [ASR] Add ARK-ASR-3B support by @Peterren in #1143
• docs(router): add user-facing troubleshooting guide by @estellaliu233 in #442
• [Bugfix] Abort backend inference when a streaming response ends early by @liuhuijiayou in #1080
• [Ming-Omni] Add thinker payload projection functions by @AkazaAkane in #1000
• [TTS] Use model-owned streaming vocoder defaults by @luojiaxuan in #1225
• [CI] Recalibrate Fun-ASR SeedTTS thresholds by @SandyLuXY in #1220
• [Perf] Add process isolation override for pipeline stages by @AkazaAkane in #1125
• [TTS][MOSS] Decouple MOSS-TTS Delay codec and run reference encoding on GPU by @CloudRipple in #1222
• [Fix] Add MOSS preprocessing GPU budget by @MayDomine in #1242
• Bump SGLang to 0.5.16 by @junliu-mde in #1183
• [TTS][Qwen3] Fix seeded sampling distribution by @luojiaxuan in #1239
• [TTS][Fish] Fix streaming final decode and cadence by @luojiaxuan in #1241
• ci: use Docker Hub omni image cache by @Hayden727 in #1254
• [CI] ASR stage 2: random Qwen3-ASR/Fun-ASR model selection (#1214) by @0xjeffro in #1215
• Calibration 07 31 by @zhaochenyang20 in #1260
• [CI]: Add ASR model selectors to slash commands by @Ratish1 in #1289
• fix(qwen3-asr): avoid silently truncating audio beyond 30 seconds by @lijrjyan in #1176
• [CI] Add Stage 5 TTS MPS validation by @JiaxinD in #1186
• [Consumer GPU] Add architecture-aware Fish backend and runtime diagnostics by @Hayden727 in #1199
• [Fix] Keep MOSS-TD encoder OOM request-scoped by @Hayden727 in #1290
• [Docs] Announce TTS architecture refactor by @luojiaxuan in #1316
• [TTS] Add MOSS-TTS Delay streaming by @luojiaxuan in #1311
• [ASR] Duration-aware output budget, context clamp, and 400 mapping for long audio transcription by @db-ol in #1034
• [Feature] ：support FishAudio S2-Pro on consumer GPUs by @Dayuxiaoshui in #1243
• [ASR Refactor][M3] Extract shared pre-LM encoder service mechanics by @AkazaAkane in #1273
• [ASR Refactor][M1] Extract shared audio preparation by @WilsonZheng0327 in #1274
• [ASR][Perf] Enable async decode for Qwen3-ASR by @SandyLuXY in #1326
• feat(qwen3-tts): add true incremental streaming by @Hayden727 in #1286
• [Perf][MOSS-TTS] Add MOSS-TTS sampling CUDA graphs by @CloudRipple in #1303
• [Fix] Restore MOSS-TD request-builder test import by @SandyLuXY in #1337
• [Feat] Add Qwen3-Omni audio streaming to /v1/realtime by @RajPatel17 in #1314
• [CI] Recalibrate Qwen3-ASR SeedTTS speed thresholds by @SandyLuXY in #1338
• [CI] Raise talker_ar mem fraction by @db-ol in #1298
• [TTS][Qwen3] Reduce first audio latency by @luojiaxuan in #1336
• [Perf][MOSS-TTS] Batch and cache reference encoding by @CloudRipple in #1348
• [ASR Refactor][M2] Extract shared SGLang generation engine builder by @Dayuxiaoshui in #1277
• [ASR][Perf] Current-main concurrency benchmark and bottleneck profile for Qwen3-ASR by @luojiaxuan in #1340
• [ASR][Perf] Pre-LM audio encoder service for Qwen3-ASR by @luojiaxuan in #1341
• [ASR][Perf] Tune Qwen3-ASR admission defaults from the measured knob sweep by @luojiaxuan in #1345
• [Qwen3-Omni Perf] Predictor attention: SDPA enable_gqa instead of materialized KV expansion by @EdwardZhang1108 in #1164
• fix(serve): stop pipeline workers after handled signals by @wirybeaver in #1153
• [Benchmark] Add reproducible SeedTTS ASR benchmark foundation by @wirybeaver in #1155
• [Qwen3-TTS] Preserve termination reason on speech responses by @lijrjyan in #1185
• [router] Add multiprocess CP/DP router with shared admission and crash-safe control plane by @JiaxinD in #1049
• [ASR Refactor][M4] Extract shared token text streaming mechanics by @guozhihao-224 in #1275
• [Qwen3-ASR][Perf] Route MRoPE positions to fused kernel by @SandyLuXY in #1358
• [Model] Add framework-native dots.tts support by @Hayden727 in #1349
• [ASR] Multilingual support for Qwen3-ASR by @SandyLuXY in #1325
• [Qwen3-Omni] Fix 555-sample audio gap at every streamed code2wav chunk boundary by @edwingao28 in #1355
• [Qwen3-Omni Perf] Cache talker thinker-embed shard resolution and rows by @edwingao28 in #1187
• [dots.tts Perf] Skip the full vocabulary LM head on AR steps by @db-ol in #1370
• fix(qwen3-asr): validate context capacity before preprocessing by @wirybeaver in #1193
• [CI] Recalibrate Qwen3-ASR SeedTTS speed thresholds by @SandyLuXY in #1373
• [Perf] dots.tts single-request path: async EOS readback, effective-backend logging, benchmark --server-config by @luojiaxuan in #1372
• Align Omni CI with image Python 3.12 by @Hayden727 in #1343
• [Perf] dots.tts: SGLang backbone decode CUDA graph via a model-owned feedback buffer by @luojiaxuan in #1377
• [Qwen3-ASR] Default to BF16 greedy decoding by @SandyLuXY in #1365
• [NPU] Add hardware abstraction layer and Ascend NPU support for Qwen3-Omni text only mode by @Makcum888e in #1306
• [Perf] dots.tts: CUDA graph batched acoustic tail by @Hayden727 in #1385
• [Perf] dots.tts: default optimize=true for the latent engine and vocoder by @luojiaxuan in #1376
• [CI] Pin perf-gate test sessions to OMNI_CI_CPUSET when set by @JiaxinD in #1321
• [ASR Refactor][M5] Extract transcription serving by @wirybeaver in #1285
• [Model] dots.tts: serve the SOAR checkpoint on the single-request solver by @luojiaxuan in #1391
• [dots.tts] Fix streaming crash under backbone decode CUDA graphs by @db-ol in #1395
• [CI] Pass OMNI_CI_CPUSET through to perf stage containers by @JiaxinD in #1388
• [Docs] Add the dots.tts cookbook by @luojiaxuan in #1393
• [Perf] dots.tts: batch AudioVAE decoding by length bucket by @AkazaAkane in #1374
• [Perf] dots.tts: skip full-batch DiT KV gathers by @Hayden727 in #1389
• [CI] Recalibrate Fun-ASR and Qwen3 TTS gates on the pinned cpuset condition by @JiaxinD in #1405
• [Perf] Generalize prefill admission coalescing across supported AR stages by @MayDomine in #1073
• [higgs-tts] add breakable prefill CUDA graph (BCG) support by @Ratish1 in #1364
• Prepare SGLang-Omni 0.1.1 for PyPI by @Ratish1 in #1414
• [Router] Add complete TTS routing support to Omni Router by @Ratish1 in #1332
• build: make Descript codecs model-specific by @Ratish1 in #1416
• [TTS] Fix talker SGLang role mappings for Qwen3-TTS and Higgs by @charliechenye in #1413

New Contributors

• @AkazaAkane made their first contribution in #826
• @codestrongestx made their first contribution in #885
• @YzXiao101 made their first contribution in #706
• @stephenkgli made their first contribution in #900
• @db-ol made their first contribution in #911
• @CloudRipple made their first contribution in #914
• @guozhihao-224 made their first contribution in #965
• @lvliang-intel made their first contribution in #722
• @PoTaTo-Mika made their first contribution in #1078
• @MayDomine made their first contribution in #1071
• @cuzmi made their first contribution in #1064
• @WilsonZheng0327 made their first contribution in #1092
• @Peterren made their first contribution in #1108
• @yl3469 made their first contribution in #1124
• @Dayuxiaoshui made their first contribution in #1141
• @keke0315 made their first contribution in #1094
• @liuhuijiayou made their first contribution in #1080
• @lijrjyan made their first contribution in #1176
• @RajPatel17 made their first contribution in #1314
• @EdwardZhang1108 made their first contribution in #1164
• @wirybeaver made their first contribution in #1153
• @Makcum888e made their first contribution in #1306
• @charliechenye made their first contribution in #1413

Full Changelog: 0.1.0...0.1.1