v0.10.2a1

What's Changed

• fix: hotfix for dependencies by @rebel-seinpark in #455

Full Changelog: v0.10.2a0...v0.10.2a1

v0.10.2a0

What's Changed

• other: sync dev with main by @rebel-eunji in #421
• fix(kernel): added sinks as parameters for attention and causal attention by @rebel-jindol21 in #422
• release: v0.10.1post1 by @rebel-eunji in #424
• other: Merge pull request #424 from RBLN-SW/dev by @rebel-eunji in #425
• feature(spec-dec): initial works on ngram and suffix by @huijjj in #408
• fix: add sink argument for swa prefill by @rebel-jaehunryu in #427
• fix: add sink argument for swa prefill by @rebel-eunji in #428
• other: sync with dev by @rebel-eunji in #429
• other: sync main with dev by @rebel-eunji in #432
• other: sync dev with main by @rebel-eunji in #433
• fix: update format of auto-created PR title by @rebel-seinpark in #426
• model: support minimax by @rebel-kblee in #435
• fix: guide what to do on oom by @rebel-jaehwang in #437
• fix: update batch attention logic to handle padding and size conditions by @rebel-jaehunryu in #436
• fix(moe): order of token mask by @rebel-kblee in #441
• fix(context): Quickfix for global context by @rebel-yskim in #444
• fix: enable to turn off RBLN_FORCE_CCL_ASYNC in debugging by @rebel-eunji in #446
• fix: input ref for pr ci by @rebel-seinpark in #447
• other: auto-update optimum-rbln to 0.10.2a0 by @rebel-develop in #452
• feature(context): always use global_ctx by @rebel-yskim in #413
• fix(specdec): remove block size limitation by @junstar92 in #440
• other: replace vLLM GPU with CPU dependency by @rebel-eunji in #439
• fix: set num_threads in TP=1 & DP=1, and change env variable by @rebel-yskim in #453

New Contributors

• @rebel-kblee made their first contribution in #435

Full Changelog: v0.10.1...v0.10.2a0

v0.10.1.post2

What's Changed

• fix: add sink argument for swa prefill by @rebel-jaehunryu in #427

Full Changelog: v0.10.1post1...v0.10.1post2

v0.10.1.post1

What's Changed

• other: sync dev with main by @rebel-eunji in #421
• fix(kernel): added sinks as parameters for attention and causal attention by @rebel-jindol21 in #422

Full Changelog: v0.10.1...v0.10.1.post1

v0.10.1

What's Changed

• chore: pin the version of optimum-rbln by @rebel-eunji in #328
• fix: resolve timeout issue in OpenAI tests by @rebel-eunji in #329
• bump: initial vllm v0.13 bump up by @huijjj in #336
• core: bump vllm to 0.13.0 by @rebel-eunji in #298
• add: pytorch native rejection sampler by @huijjj in #341
• chore: update pre-commit on dev-0.13 by @rebel-eunji in #345
• chore: add .git-blame-ignore-revs by @rebel-eunji in #355
• fix: update imports to reflect package path/name changes by @junstar92 in #356
• fix(test): common unit test by @rebel-jiwoopark in #357
• update: enable kv cache meta tensor only when it is not a user custom kernel by @rebel-jiwoopark in #358
• fix: eagle/eagle3 for v0.13.0 by @junstar92 in #346
• chore: bump dev branch to dev-0.13 by @rebel-eunji in #361
• fix(model): skip computing encoder budget by @rebel-eunji in #368
• feat(kernel): support normal and causal normal kernel in mlir by @rebel-jindol21 in #360
• test(ci): add retry wrapper by @rebel-jaebin in #359
• fix(kernel): added missing kernel decl by @rebel-jindol21 in #374
• fix: adapt to 0.13 FusedMoEMethod signature by @rebel-jaehwang in #369
• fix: multimodal processing vllm 0.13.0 and sampler by @rebel-eunji in #378
• fix(platform): not support disabling chunked prefill by @rebel-jiwoopark in #366
• changed an order of arguments by @rebel-jindol21 in #372
• refactor: Remove useless workflow by @rebel-jaebin in #385
• feature(kernel): modify user kernel mode option by @rebel-jiwoopark in #383
• feature(attn): remove sinks attn param for sliding window attn. by @rebel-jiwoopark in #390
• other(ci): add PR title checker by @rebel-seinpark in #387
• fix(ci): fix pr title checker by @rebel-jaehwang in #392
• fix: handle 0-sized logits by @rebel-jaehwang in #381
• fix: change the pooling type of encoder model to CLS by @rebel-eunji in #379
• fix(model): fix Qwen3 embedding support in vLLM 0.13.0 by @rebel-eunji in #380
• fix: reshape logits to avoid stride-triggerd recompilation by @rebel-eunji in #364
• fix: correct sampler output under CPU eager fallback of torch.compile by @rebel-eunji in #382
• feature: merge dev-0.12-rebase into dev by @rebel-eunji in #395
• feature(ccl): Enable autoport in rbln-ccl by @rebel-yskim in #397
• fix(core): Sampler with RBLN_CTX_STANDALONE by @rebel-jonghewk in #401
• core(platform): check prerequisite for parallelism by @rebel-jiwoopark in #393
• fix(attn): attn backend related to sinks by @rebel-jiwoopark in #403
• other: Revert "fix(core): Sampler with RBLN_CTX_STANDALONE" by @rebel-eunji in #405
• fix: set RBLN_CTX_STANDALONE to False by @rebel-eunji in #406
• refactor: fix available memory estimation by @rebel-ykchoi in #404
• fix: Re-enable standalone ctx by @rebel-jonghewk in #407
• other: Auto-update optimum-rbln to 0.10.1a2 by @rebel-develop in #412
• fix: reshape logits only in prefill phase by @rebel-eunji in #409
• fix: set num_blocks outside of model_runner by @rebel-eunji in #410
• fix: add sink argument for contrib.custom_ops by @rebel-jaehunryu in #416
• other: Auto-update optimum-rbln to 0.10.1 by @rebel-develop in #419
• release: v0.10.1 by @rebel-seinpark in #420

Full Changelog: v0.10.0...v0.10.1

v0.10.1a0

What's Changed

• chore: pin the version of optimum-rbln by @rebel-eunji in #328
• fix: resolve timeout issue in OpenAI tests by @rebel-eunji in #329
• bump: initial vllm v0.13 bump up by @huijjj in #336
• core: bump vllm to 0.13.0 by @rebel-eunji in #298
• add: pytorch native rejection sampler by @huijjj in #341
• chore: update pre-commit on dev-0.13 by @rebel-eunji in #345
• chore: add .git-blame-ignore-revs by @rebel-eunji in #355
• fix: update imports to reflect package path/name changes by @junstar92 in #356
• fix(test): common unit test by @rebel-jiwoopark in #357
• update: enable kv cache meta tensor only when it is not a user custom kernel by @rebel-jiwoopark in #358
• fix: eagle/eagle3 for v0.13.0 by @junstar92 in #346
• chore: bump dev branch to dev-0.13 by @rebel-eunji in #361
• fix(model): skip computing encoder budget by @rebel-eunji in #368

Full Changelog: v0.10.0...v0.10.1a0

v0.10.0

What's Changed

• fix: add guard_filter to rbln_sampler by @rebel-eunji in #241
• fix environment variables for rbln ccl by @rebel-ykchoi in #242
• other(ci): change the runner of PR CI by @rebel-eunji in #245
• Revert "fix: add guard_filter to rbln_sampler" by @rebel-eunji in #251
• update: Change log level in CI by @rebel-jonghewk in #250
• other(ci): log the length of generated tokens when the request is preempted by @rebel-eunji in #248
• fix(model): fix whisper model by @rebel-eunji in #247
• other(CI): clean up workflow by @rebel-seinpark in #233
• update(core): customize blockpool to increase prefix cache hit rate by @rebel-eunji in #205
• fix(ci): pkg auto update by @rebel-seinpark in #259
• fix(ci): pkg update by @rebel-seinpark in #260
• other: add enable_expert_parallel in basic example by @rebel-jiwoopark in #255
• Auto-update optimum-rbln to 0.9.5a0 by @rebel-shshin in #261
• update(worker): Set NUMA aware CPU affinity and OMP_NUM_THREADS by @rebel-yskim in #232
• other(ci): fix bug in sampler and update pytest for sampler by @rebel-eunji in #239
• Add dev0.12 event type by @rebel-jaebin in #267
• Auto-update optimum-rbln to 0.9.5a1 by @rebel-shshin in #268
• update: enable profiler for optimum-rbln based vllm by @rebel-jonghewk in #256
• fix: warm up for swa hybrid model by @rebel-jaehwang in #231
• other(ci): remove explicit secrets by @rebel-seinpark in #276
• Auto-update optimum-rbln to 0.9.5a2 by @rebel-shshin in #275
• Update runner name by @rebel-jaebin in #281
• other: repo migration (sw) by @rebel-seinpark in #271
• Auto-update optimum-rbln to 0.9.5a4 by @rebel-develop in #282
• other(test): add unit test for attention backend. by @rebel-jiwoopark in #273
• Revert "other(test): add unit test for attention backend." by @rebel-jiwoopark in #285
• Auto-update optimum-rbln to 0.9.5a5 by @rebel-develop in #286
• update(triton): replace triton with torch_triton by @rebel-jindol21 in #274
• fix moe data parallel for v1 engine by @rebel-ykchoi in #252
• Auto-update optimum-rbln to 0.9.5a6 by @rebel-develop in #288
• Revert "update(triton): replace triton with torch_triton" by @rebel-jindol21 in #289
• feat: enable topk_topp_sampler by @rebel-eunji in #284
• feat: prefill performance by request_id and exclude warmup requests from performance tracking by @rebel-eunji in #279
• feat(triton): support torch_triton by @rebel-jindol21 in #292
• fix(model): pad image tokens using an out-of-vocab index by @rebel-eunji in #291
• Auto-update optimum-rbln to 0.9.5a7 by @rebel-develop in #296
• fix: use rbln_sampler when both top_k and top_p are None by @rebel-eunji in #295
• update(triton): fixed along with triton kernels in rebel_compiler by @rebel-jindol21 in #299
• fix: apply argmax to greedy by @rebel-eunji in #300
• fix(kernel): fix kvcache by @rebel-jindol21 in #301
• Auto-update optimum-rbln to 0.9.5a8 by @rebel-develop in #302
• fix(triton): alignment enforced by @rebel-jindol21 in #303
• fix(attn): remove invalid attention param by @rebel-jiwoopark in #305
• sync main-dev by @rebel-eunji in #314
• Auto-update optimum-rbln to 0.10.0.post1 by @rebel-develop in #315
• Release v0.10.0 by @rebel-eunji in #318

New Contributors

• @rebel-yskim made their first contribution in #232
• @rebel-develop made their first contribution in #282

Full Changelog: v0.9.4...v0.10.0

v0.9.4

What's Changed

• fix(core): async engine bug by @rebel-jiwoopark in #140
• feat(core): add structured output support and add benchmark scripts by @pei0033 in #121
• fix(bug): Fix the bug of recompile_limit when compiling sampler by @rebel-eunji in #153
• Batched rope by @rebel-jaehwang in #150
• feat(model): Support for Structured Ouput by @rebel-eunji in #151
• fix(core): customize the KV Cache Manager for prefix caching by @rebel-eunji in #146
• fix(core): Fix the logic of dummy block selection for padding in case of prefix caching by @rebel-eunji in #154
• fix(core): exclude cache-hit blocks when checking free block availability by @rebel-eunji in #158
• trivial: remove hardcoded strict compilation by @huijjj in #161
• fix(core): skip touch function of cached blocks by @rebel-eunji in #164
• ci: add GitHub Actions workflow for ARC CI testing by @rebel-jaebin in #120
• refactor: clean up input padding logic by @rebel-jaehwang in #168
• dev revert 251124 by @rebel-sunwook in #170
• feat: mixed precision quantization by @rebel-jaehwang in #159
• fix(core): fix debug logging in case of abortion by @rebel-eunji in #173
• ci: Add internal ci trigger workflow by @rebel-jaebin in #172
• ci: add basic model test. by @rebel-jiwoopark in #165
• chore(deps): bump optimum-rbln package from 0.9.2.a7 to 0.9.3(stable) by @rebel-eunji in #149
• update MoE PoC features & bfloat16 model load by @rebel-wonsubkim in #145
• add(CI): model coverage on single device by @huijjj in #178
• feat: sliding window attention by @rebel-jaehwang in #167
• feat(model): prevent excessive recompilation of sampler by @rebel-eunji in #171
• fix(core): fix handling logprobs when use rbln_sampler by @rebel-eunji in #184
• [feat] LoRA support on V1 (with limitations) by @junstar92 in #169
• feat(model): fix type casting and enable bfloat16 optimum-rbln compiled models by @rebel-eunji in #183
• fix(core): v1 scheduler by @huijjj in #181
• update(ci): add openai server test in PR CI by @rebel-eunji in #182
• fix(core): fix sampler buckets to include max_num_seqs by @rebel-eunji in #186
• chore: bump version of optimum-rbln to 0.9.4a0 by @rebel-eunji in #189
• other: cache rbln_sampler by @rebel-eunji in #187
• fix(model): fix gemma3 to use left-padding by @rebel-eunji in #190
• other: use standard cache dir by @rebel-jaehwang in #188
• other(v0): log model loading & compilation time by @rebel-jaebin in #193
• other(ci): add common unit tests. by @rebel-jiwoopark in #195
• Moe PoC v1 migration by @rebel-wonsubkim in #180
• fix(other): disable autograd in logits by wrapping with torch.inference_mode by @rebel-eunji in #194
• Restore revert 251124 by @rebel-jaebin in #174
• fix: cleanup installation by @rebel-sunwook in #198
• fix: python 3.9 compat by @rebel-jaehwang in #200
• feat(other): unify the contexts in sampler by @rebel-eunji in #197
• feat(model): add tokens mask for MoE custom kernel by @rebel-ykchoi in #192
• feat(core): add pooling model initial support for V1 engine by @pei0033 in #152
• other(ci): add pytest-cov by @rebel-jiwoopark in #203
• update vllm moe kernel by @rebel-jaehunryu in #202
• test: add comprehensive LoRA tests by @junstar92 in #199
• fix(model): fix custom_moe_glu custom operation signature by @rebel-jaehunryu in #204
• feat(model): add paligemma, paligemma2 by @rebel-eunji in #162
• other: log format by @rebel-jaehwang in #207
• feat(core): allow dynamic block_size for prefix caching by @rebel-eunji in #185
• Revive requirements txt by @rebel-sunwook in #210
• feat(model): support hybrid attention text-only models in optimum (gpt-oss, gemma2) by @rebel-eunji in #208
• fix(script): DP placeholder condition by @rebel-myeongbo in #201
• feat(ci): add paligemma to optimum ci by @rebel-seinpark in #211
• other(ci): add core pytest by @rebel-jiwoopark in #212
• fix: missing sinks attention param by @rebel-jiwoopark in #213
• chore: bump version of optimum-rbln up to 0.9.4rc0 by @rebel-seinpark in #218
• fix: get dtype from optimum model by @rebel-seinpark in #217
• fix(env): change default value of VLLM_RBLN_MOE_USE_OPT_KERNEL to True by @rebel-jaehunryu in #220
• fix: modify physical device ids for v1 dp rsd by @rebel-wonsubkim in #214
• feat(ci): auto update for optimum-rbln by @rebel-seinpark in #222
• fix(core): Apply keep_tensor_guards_unsafe for torch.compile by @rebel-jonghewk in #229
• fix: block caching logics to enable prefix caching by @huijjj in #228
• fix(sampler): Disable sampler cache by @rebel-jonghewk in #234
• fix: Resolve dev main merge conflict by @rebel-jonghewk in #236
• Release v0.9.4 by @rebel-shshin in #235
• fix(packaging): revive setup.py by @rebel-sunwook in #237
• fix(CI): revive setup.py by @rebel-shshin in #238

New Contributors

• @rebel-jaebin made their first contribution in #120
• @rebel-ykchoi made their first contribution in #192
• @rebel-jaehunryu made their first contribution in #202
• @rebel-myeongbo made their first contribution in #201

Full Changelog: v0.9.3...v0.9.4

v0.9.3.post2

Full Changelog: v0.9.3.post1...v0.9.3.post2

v0.9.3.post1

Full Changelog: v0.9.3...v0.9.3.post1