v0.11.0

What's Changed

• other: sync dev with main by @rebel-eunji in #635
• fix(test): isolate LLM-using tests via patch_and_run to prevent cross-test OOM by @rebel-jaehwang in #629
• fix(test): parametrize sampling method for structured output by @rebel-jinhwan in #636
• fix(test): replace exact match test in spec decode by logprob similarity test by @rebel-jinhwan in #630
• feature(perf): support metrics file output by @rebel-minwook in #638
• fix(test): deterministic engine shutdown by @rebel-jaehwang in #639
• feature(compile): enable torch.compile on NPU-less environment by @rebel-jinhwan in #618
• feature(spec_dec): implement spec decode backfill for fixed length drafts (full_spec only) by @rebel-wonsubkim in #604
• refactor(attn): remove in-graph dtype casts on attn custom op args by @rebel-thkim in #616
• fix(spec-decode): prevent sampler broadcast crash with multi-batch + device tensor by @rebel-jinhwan in #641
• feature(env): default VLLM_RBLN_AUTO_PORT to VLLM_RBLN_USE_DEVICE_TENSOR by @rebel-jonghewk in #645
• other: Revert "fix(optimum): adapt optimum path to vLLM 0.22 API changes" by @rebel-eunji in #650
• fix(attn): normalize seq_lens to int32 so custom kernel compiles by @rebel-jinhwan in #648
• other(test): include e2e lora test to CI by @rebel-jinhwan in #656
• refactor: drop RBLN_FORCE_CCL_ASYNC (async is now the runtime default) by @rebel-jonghewk in #660
• fix(pre-commit): enforce full license header & range-validate copyright year by @rebel-eunji in #655
• feature: support for rejection sampler by @rebel-eunji in #543
• fix(scheduler): re-emit KV block dropped by mixed-batching eviction by @rebel-ykchoi in #662
• fix: update max_num_batched_tokens after max_model_len by @rebel-eunji in #663
• feature: enable all2all communication for MoE model by @rebel-jangys in #605
• other(bump): make MoE models compatible to v0.22.0 by @rebel-jinhwan in #667
• model: assert when executing an unregistered model by @rebel-eunji in #642
• fix(moe): fix unnecessary routing logit transpose by @rebel-jangys in #674
• fix(worker): unquote CompilationTimes return annotation by @rebel-jinhwan in #673
• other(bump): bump vllm to 0.22.0 with optimum by @rebel-eunji in #671
• fix(test): re-enable device-tensor path in sub-block prefix caching test by @rebel-jonghewk in #681
• feature(ci): run e2e in PR CI when 'pytest(e2e)' label is present by @rebel-jinhwan in #679
• model: support gemma4 by @rebel-eunji in #622
• model: fix custom op signature (sync with act_type) by @rebel-kblee in #678
• fix(model): follow the model architecture as mentioned in registry by @rebel-eunji in #686
• fix(tmp): turn off loading gemma4 model by @rebel-eunji in #687
• fix: apply default batch size of 1 across all UsageContext cases by @rebel-eunji in #652
• refactor(model): sync whisper with SupportsTranscription interface by @rebel-eunji in #668
• other(ci): pin fastapi below 0.137 for OpenAI entrypoint tests by @rebel-eunji in #691
• other(ci): pin fastapi below 0.137 for OpenAI entrypoint tests (dev) by @rebel-eunji in #692
• fix(core): accept alignment_tokens kwarg in RBLNSlidingWindowManager.cache_blocks by @rebel-jonghewk in #689
• refactor(multimodal): follow the multimodal interface of upstream vllm by @rebel-eunji in #664
• model: support for EXAONE4.5 by @rebel-eunji in #688
• refactor(paligemma): move dataset/tokenizer access out of async path by @rebel-eunji in #694
• model(whisper): revert the vllm model inheritance and add pytest by @rebel-eunji in #696
• fix(scheduler): clear stranded KV-block stash on preemption by @rebel-ykchoi in #676
• other(test): remove unnecessary rejection sampler test case and skip device tensor scenario by @rebel-jinhwan in #690
• fix: fix minor vLLM 0.22.0 compatibility issues by @pei0033 in #698
• other(bump): make vllm-rbln compatible to vllm 0.22.0 by @rebel-jinhwan in #669
• other: auto-update optimum-rbln to 0.11.0a11 by @rebel-develop in #703
• refactor: refactor encoder models and add pytests by @rebel-eunji in #675
• feature(env): rename env var VLLM_RBLN_TP_SIZE to VLLM_RBLN_NUM_DEVICES_PER_LOCAL_RANK by @rebel-eunji in #697
• other: auto-update optimum-rbln to 0.11.0a12 by @rebel-develop in #704
• fix(deps): pin fastapi below 0.137 in pyproject instead of CI by @rebel-eunji in #699
• other: auto-update optimum-rbln to 0.11.0a13 by @rebel-develop in #707
• other: auto-update optimum-rbln to 0.11.0a14 by @rebel-develop in #708
• model: support deepseek v3 by @rebel-kblee in #624
• fix: compiled model path by @rebel-seinpark in #712
• fix(attn): keep dummy_run block table on CPU in device-tensor mode by @rebel-jonghewk in #709
• feature(warmup): offload host regions during device_tensor warm-up by @rebel-jonghewk in #646
• other: auto-update optimum-rbln to 0.11.0a15 by @rebel-develop in #714
• fix(model): re-enable gemma4 and fix padding logic by @rebel-eunji in #710
• fix: hotfix dtensor condition by @rebel-kblee in #716
• fix(model): support Gemma4 multimodal decode on the vLLM path by @rebel-thkim in #715
• fix(gemma3,4): move the padding logic from model to scheduler by @rebel-eunji in #717
• other: auto-update optimum-rbln to 0.11.0 by @rebel-develop in #724
• release: v0.11.0 by @rebel-eunji in #725
• other(chore): sync main and dev by @rebel-eunji in #727
• fix(gemma3,4): resolve image_prefill_chunk_size (int or list) from optimum-rbln by @rebel-eunji in #726
• release: v0.11.0 2nd by @rebel-eunji in #728

New Contributors

• @rebel-minwook made their first contribution in #638

Full Changelog: v0.10.4...v0.11.0