v0.1.2

What's Changed

• [CI] Calibration precheck: stop CI first and pin exactly one lane bundle by @JiaxinD in #1417
• [docs] Upd After v 0.1.1 by @zhaochenyang20 in #1419
• [Qwen3-Omni Perf] Thinker prefill: batch multimodal embedding merge, remove per-request host syncs by @cr-gao in #1161
• [Feature] Enable automatic language detection for Qwen3-ASR by @SandyLuXY in #1403
• [CI] Cpuset contention: report it, fail the stage, guard calibration rounds by @JiaxinD in #1415
• [CI] Measure cpuset contention from per-CPU counters, not affinity masks by @JiaxinD in #1423
• [Perf] dots.tts: defer batched EOS readback to ModelRunner resolve by @guozhihao-224 in #1402
• [Enhancement] dots.tts: document and gate acoustic-tail pool memory by @guozhihao-224 in #1408
• [dots.tts] Fix non-deterministic reference encoding; add parity-safe batched reference encode by @buffett0323 in #1420
• [Perf] Raise Qwen3-ASR concurrency ceiling to 64 by @SandyLuXY in #1421
• [Perf] Query Qwen3-ASR embedding cache before mel extraction by @SandyLuXY in #1422
• [Perf] Enable pending-build-aware prefill coalescing for Qwen3-ASR by @SandyLuXY in #1404
• [ASR] Add a 90 minute tier and output format gates to the long audio CI by @db-ol in #1036
• feat(qwen3-asr): add RTX 4090 serving profile by @wirybeaver in #1154
• [Fun-ASR] Added bucketed CUDA graphs for encoder forward by @WilsonZheng0327 in #1369
• [Perf] Streamline Qwen3-ASR request construction by @luojiaxuan in #1425
• [Benchmark] Add deterministic mixed-arrival TTS serving workloads by @Ratish1 in #1333
• [Model, Qwen3-TTS] Shim the Transformers mask factories for qwen-tts 0.1.1 by @AkazaAkane in #1407
• [Perf] Keep Qwen3-ASR prefill coalescing during active decode by @luojiaxuan in #1432
• [Docker] Reproduce the validated Omni runtime image by @Hayden727 in #1443
• [ASR][Perf] Enable async lookahead for single-request Qwen3-ASR by @SandyLuXY in #1431
• [CI] Contention report is advisory, never fatal by @JiaxinD in #1456
• [CI] Move TTS serving CI to mixed-arrival router DP2 by @Ratish1 in #1334
• [Perf] dots.tts: stack feedback into the graph buffer by @luojiaxuan in #1439
• [Perf] dots.tts: reuse batched latent denormalization by @luojiaxuan in #1438
• [CPU] Monitor CPU While Calibration by @zhaochenyang20 in #1457
• [Perf] dots.tts: bypass redundant vocoder staging by @luojiaxuan in #1440
• [Perf] dots.tts: reuse the CFG null projection bias by @luojiaxuan in #1441
• [Perf] Speed up Qwen3-ASR low and mid concurrency by @luojiaxuan in #1447
• [Perf] dots.tts: batch compatible streaming AudioVAE steps by @Hayden727 in #1444
• [CI] Confine unit-test and setup work to the runner lane cpuset by @JiaxinD in #1461
• [ASR Refactor][A-PR1] ARK-ASR: move engine setup to AsrEngineBuilder by @jiannan-17 in #1397
• [ASR Refactor] Extract shared speech-to-text serving helpers by @lijrjyan in #1394
• [CI] Recalibrate Qwen3-ASR SeedTTS thresholds by @SandyLuXY in #1465
• [ASR Refactor][F3] Whisper: map prompt to prev-context tokens by @jiannan-17 in #1339
• [ASR][Perf] Add request-event profiling to MOSS-Transcribe-Diarize benchmark by @SandyLuXY in #1430
• [Whisper ASR] Add Bucketed Encoder CUDA Graphs and Atomic Prefill by @Dayuxiaoshui in #1412
• [Fix] realtime VAD dropping audio at adjacent turn boundaries(#1322) by @guozhihao-224 in #1327
• [Comm] Add paged KV cache transfer infrastructure by @JingwenGu0829 in #1315
• [Docs] Refresh the dots.tts H100 benchmark by @luojiaxuan in #1482
• [Feat] Add server-owned barge-in to /v1/realtime audio sessions by @RajPatel17 in #1344
• [ASR] Long-audio chunking for Qwen3-ASR by @0xjeffro in #1347
• [MOSS-TTS] Add consumer-GPU qualification layout by @LG-0927 in #1300
• [Perf][MOSS-TTS] Refactor and optimize vocoder decoding by @CloudRipple in #1387
• [Perf] Streamline MOSS-Transcribe-Diarize request construction by @LG-0927 in #1474
• [Fix] Keep Fun-ASR encoder OOM request-scoped by @LauraGPT in #1309
• [Comm] Add rank-to-rank TP KV cache transfers by @JingwenGu0829 in #1382
• [qwen3-asr] add breakable prefill CUDA graph (BCG) support by @Ratish1 in #1458
• perf(qwen3-asr): vectorize decode-step mrope positions by @0xjeffro in #1472
• [Perf][M-PR2] Enable tuned prefill coalescing for MOSS-TD by @LG-0927 in #1491
• [Perf] Qwen3-TTS: cut per-step host work in the AR decode loop by @JiaxinD in #1462
• [Qwen3-Omni Perf] Graph-safe eager thinker: static hidden capture by @edwingao28 in #1380
• Whisper ASR W-PR2: prefill coalescing to reduce small-batch prefill by @Dayuxiaoshui in #1485
• [ASR Optimization][A-PR3] ARK-ASR: enable shared one-step-lookahead async decode by @jiannan-17 in #1483
• [Qwen3-Omni Perf] Vectorize thinker MRoPE; document talker linearization gate (#1149) by @guozhihao-224 in #1160
• [Perf][M-PR3] Query MOSS-TD embedding cache before mel extraction by @LG-0927 in #1495
• [Perf] Enable MOSS-TD async decode at batch size 1 by @SandyLuXY in #1509
• [Perf] Enable FP8 / CPU offload on TTS pipelines by @xinlij in #1312
• [Qwen3-Omni Perf] Default code2wav onto the thinker GPU so the talker gets its own device by @edwingao28 in #1235
• [NPU] Qwen3-Omni talker support by @Makcum888e in #1476
• [Fix] transcription: skip no-speech chunks instead of hallucinating text by @JiaxinD in #1512
• [Model] Support MiniMax Music 3 Model by @zhaochenyang20 in #1521
• [Perf][ASR] Cache the audio front-end constant tables across requests by @JiaxinD in #1516
• [CI] Register Qwen3-TTS 1.7B in the TTS model rotation by @JiaxinD in #1505
• [Whisper] Enable long audio chunking for Whisper by @0xjeffro in #1520
• [Qwen3-Omni Perf] Add Breakable Prefill CUDA Graph support for Thinker text/audio prefill by @charliechenye in #1381
• [Perf][A-PR4] Batched encoder inference for ARK-ASR by @nagisa-kunhah in #1411
• [Perf] Enable breakable prefill CUDA graph for MOSS-TD by @PrinsYin in #1379
• [Perf] dots.tts: slot-pooled batched streaming AudioVAE by @guozhihao-224 in #1446
• [Intel][XPU] Enable Intel XPU's support for SGLang-Omni by @siju-samuel in #994
• [Comm] Assign unique identities to stream relay transfers by @charliechenye in #1540
• [Perf] Qwen3-TTS: improve playback continuity and add underrun metrics by @AkazaAkane in #1529
• [ASR Feature] Add /v1/audio/translations with per-model capability gating by @lijrjyan in #1453
• feat(cli): integrate Omni with sglang serve by @mickqian in #1531
• [Refactor][MOSS-TTS] Consolidate the audio tokenizer runtime by @CloudRipple in #1551
• [Perf] transcription: soundfile fast path for the audio duration probe by @Ratish1 in #1524
• [CI]: Update Qwen3-ASR CI speed thresholds by @Ratish1 in #1561
• [Perf] Enable decoder compile and tiny prefill graphs for MOSS-TD by @SandyLuXY in #1550
• [CI] Recalibrate MOSS-TD short-audio speed thresholds by @PrinsYin in #1560
• [ASR Refactor][A-PR2] ARK-ASR: pre-lm encoder by @AkazaAkane in #1429
• [Qwen3-Omni Perf] Chunk-aligned Code2Wav dispatch with batched CUDA graph capture by @dyyfk in #1237
• [Bug][Qwen3-Omni] Harden cached eager multimodal cursors by @charliechenye in #1537
• [Bug][Qwen3-Omni] Restore deferred CUDA graph args on bootstrap failure by @charliechenye in #1532
• [Release] Prepare SGLang-Omni 0.1.2 by @Ratish1 in #1571

New Contributors

• @cr-gao made their first contribution in #1161
• @buffett0323 made their first contribution in #1420
• @jiannan-17 made their first contribution in #1397
• @LG-0927 made their first contribution in #1300
• @PrinsYin made their first contribution in #1379
• @siju-samuel made their first contribution in #994
• @mickqian made their first contribution in #1531
• @dyyfk made their first contribution in #1237

Full Changelog: 0.1.1...v0.1.2