v0.10.3

What's Changed

• sync main with dev by @rebel-eunji in #489
• fix: create pr message by @rebel-seinpark in #490
• fix: update uv installation command in README by @rebel-eunji in #495
• other: merge dev into main to update README by @rebel-eunji in #496
• Merge pull request #496 from RBLN-SW/dev by @rebel-eunji in #497
• fix: performance tracker by @rebel-eunji in #503
• fix(NUMA): Cpu split after warmup by @rebel-yskim in #476
• fix(attention): derive num_partition from max_model_len instead of block_length by @rebel-jaehunryu in #513
• fix: change MoE combine by @rebel-ykchoi in #438
• other: bump up v0.18.0 by @rebel-jiwoopark in #514
• fix: num tokens for reorder batch. by @rebel-jiwoopark in #520
• fix: use container-aware thread count to avoid host op performance degradation in pods by @rebel-jonghewk in #512
• other(test): improve worker test coverage by @rebel-jinhwan in #517
• other(tests): tests for bucketing manager by @huijjj in #521
• refactor(scheduler): delay caching instead of undoing by @rebel-jaehwang in #525
• fix(encoder): use RBLNClassifierPooler for classification models by @rebel-jonghewk in #526
• fix: remove unused util func by @rebel-kblee in #530
• fix(config): ensure max_num_batched_tokens >= max_source_positions for enc-dec by @rebel-jonghewk in #527
• fix(core): deduplicate KV cache inputs for torch.export compatibility by @rebel-chanheo in #524
• fix(core): handle meta tensors in KV cache storage key computation by @rebel-chanheo in #533
• fix: skip NUMA cpu affinity on bare metal to prevent 32x latency regression by @rebel-jonghewk in #529
• fix(gemma3): use non-negative sentinel for IMG_PAD_TOKEN_ID by @rebel-jonghewk in #528
• fix(model): fix moe model attribute by @rebel-kblee in #531
• fix: use num_tokens_no_spec in optimum model runner by @rebel-seinpark in #536
• fix(encoder): fix T5EncoderModel scoring mismatch after v0.18.1 bump by @rebel-jonghewk in #532
• feature(pdd): enable P/D disaggregation with NIXL host KV transfer by @rebel-ykchoi in #477
• other: logging argmax for debugging by @rebel-eunji in #540
• feature: add RBLNLMCacheConnectorV1 for LMCache KV cache offloading by @rebel-jinhwan in #523
• other: Revert "other: logging argmax for debugging" by @rebel-eunji in #542
• feature: add sampler benchmark code by @rebel-eunji in #539
• fix(PDD): fixed tensor type for cache copy by @rebel-jindol21 in #550
• fix(whisper): use vLLM block_tables as decoder batch position by @rebel-seinpark in #549
• fix(model): make expert_map with torch export in Moe model by @rebel-kblee in #537
• other: auto-update optimum-rbln to 0.10.3a2 by @rebel-develop in #551
• fix(test): fix torch_compile unit test failures by @rebel-jinhwan in #552
• fix(examples): import get_open_port from vllm.utils.network_utils by @rebel-sjlee in #560
• fix(test): load vllm plugin before test collection by @rebel-jaehwang in #561
• fix(worker): accept profile_prefix arg in profile() by @rebel-sjlee in #559
• fix(model): qwen3 vl with only text input by @rebel-kblee in #557
• fix(model): fix memory estimate logic by @rebel-kblee in #535
• fix: set MKL_NUM_THREADS in OptimumWorker by @rebel-seinpark in #562
• feature(ec_disagg): encoder cache disaggregation for VLMs (NIXL + ZMQ) by @rebel-yskim in #547
• other: auto-update-optimum-rbln-0.10.3a3 by @rebel-seinpark in #567
• fix(worker): use vllm TorchProfilerWrapper in RBLN worker by @rebel-sjlee in #565
• fix(test): assert on n_model_bytes in determine_available_memory tests by @rebel-jinhwan in #568
• core(kernel): rsd support by @rebel-jindol21 in #564
• core(kernel): patch VLLM_RBLN_COMPILE_MODEL for _allocate_kv_cache_tensors by @rebel-jindol21 in #573
• Revert "fix: set MKL_NUM_THREADS in OptimumWorker" by @rebel-seinpark in #574
• other: auto-update optimum-rbln to 0.10.3a4 by @rebel-develop in #576
• core(kernel): change the order of arguments of triton_op along with kernel by @rebel-jindol21 in #577
• other: auto-update optimum-rbln to 0.10.3 by @rebel-develop in #578
• release: v0.13.0 by @rebel-eunji in #579

New Contributors

• @rebel-chanheo made their first contribution in #524
• @rebel-sjlee made their first contribution in #560

Full Changelog: v0.10.2...v0.10.3