• missing files by @vince62s in #395
• Add MetricX scoring, scorer-based early stopping, and external scorers by @dameikle in #393
• Replace torchaudio audio preprocessing with torchcodec and pure torch mel by @dameikle in #397

Full Changelog: 0.5.3...0.6.0

What's Changed

• Support MoE Marlin fused cuda kernel by @vince62s in #348
• update docusaurus by @vince62s in #350
• Self contained Marlin kernel (do not depend on gptqmodel) by @vince62s in #352
• split fused_moe_int4 into triton and marlin by @vince62s in #353
• fix compile bug with marlin kernel by @vince62s in #356
• Fix AudioPredictor missing streamer parameter by @dameikle in #355
• fix bug default sliding window should be 0 not 4096 at conversion by @vince62s in #357
• Support anthropic like api + Claude Code by @vince62s in #354
• fix sliding attention - add chunked prefill + chunk cache by @vince62s in #340
• Add autocast support for MPS by @dameikle in #365
• Fix id_tokenization check to scan all transforms by @yjymosheng in #364
• "-gpu" flag doesn't exist; it should be "-gpu_ranks" by @Geremia in #363
• Move LoRa embedding validation to config layer by @yjymosheng in #367
• convert_HF.py: add --check-tensors and --config-only flags by @vince62s in #361
• Enable runnning inference directly on supported HF models by @vince62s in #374
• fix pydantic and doc stuff by @vince62s in #383
• Add trackio experiment tracking as an optional logging backend by @dameikle in #385
• Add Gemma4 support by @vince62s in #382
• Add native COMET/KIWI scoring with encoder scorer models by @dameikle in #390

New Contributors

• @Geremia made their first contribution in #363

Full Changelog: 0.5.2...0.5.3

What's Changed

• Support Qwen3.5-27B/9B/4B/2B/0.8B by @vince62s in #329
• Add Whisper speech-to-text model support with HF conversion, timestamp-seeking, and word-level timestamps by @dameikle in #328
• Make triton fused_moe import optional so MoE module loads on platforms without triton (e.g. macOS) by @dameikle in #332
• Support Intel/autoround - gptqmodel by @vince62s in #331
• Add streamer by @vince62s in #333
• Support gguf conversion and inference by @vince62s in #334
• fix #327 (large model saving) by @vince62s in #335
• fix lora tensor_parallel training by @vince62s in #336
• fix #223 by @vince62s in #337
• Support Comet scorer by @vince62s in #338
• fix #107 by @vince62s in #339
• fix #312 by @vince62s in #341
• add scorer chrF/chrF++ by @yjymosheng in #297
• LM scoring / Perplexity by @vince62s in #342
• few fixes by @vince62s in #343
• docs update by @vince62s in #344
• yarn.lock renewal by @vince62s in #345

New Contributors

• @yjymosheng made their first contribution in #297

Full Changelog: 0.5.1...0.5.2

• stupid bug by @vince62s in #321
• Add a static batch size decoding to optimize torch compile usage by @vince62s in #322
• for some reason did not compile with pytorch 2.10 by @vince62s in #323
• Fixtrain by @vince62s in #324
• Dynamicshapes by @vince62s in #325

What's Changed

• fix space tok by @vince62s in #301
• fix gemma3 text model by @vince62s in #302
• add cuda ops by @vince62s in #303
• add backward pass to eole_ops rmsnorm + benchmark script by @vince62s in #305
• replace vLLM fused_moe by adhoc triton fused_moe with similar signature by @vince62s in #306
• add rope benchmark for custom ops vs pytorch implementation by @vince62s in #307
• refactor Optimizer class by @vince62s in #308
• compare inference speed Eole vs HF vs CT2 vs vLLM by @vince62s in #309
• add script to show diff between SDPBackend.EFFICIENT_ATTENTION and Fl… by @vince62s in #310
• add pytorch fused act by @vince62s in #311
• bug fixes and optimizations by @vince62s in #313
• torch.compile compliant transformer decoder + MHA - speed goes vrrrrrrmmmm by @vince62s in #316
• Cudagraphs implementation along torch compile by @vince62s in #318
• update benchmarks with cudagraphs by @vince62s in #319

Full Changelog: 0.4.4...0.5.0

What's Changed

• refactor use of vLLM ops - optimize tensor shapes for Rope by @vince62s in #289
• Refactor model class by @vince62s in #290
• Refactor dynamic iterator and tokenization workflow by @vince62s in #291
• Refactor train / train_single / trainer workflow for better readability - small fixes to distributed ops by @vince62s in #292
• Cleanrecipe by @vince62s in #293
• refactor inference_engine class and adjust distributed classes by @vince62s in #294
• Refactor encoders (more pythonic) by @vince62s in #296
• refactor and clean adapters classes by @vince62s in #298
• Refactor decoder classes - cross attention / other implications by @vince62s in #299

Full Changelog: 0.4.3...0.4.4

What's Changed

• Dynamic batching server / OpenAI like API (on top of the proprietary one) / load test tool by @vince62s in #283
• few fixes by @vince62s in #284
• fix vllm regression in dispatch forward by @vince62s in #286

Full Changelog: 0.4.2...0.4.3

What's Changed

• Add Deepseek OCR demo scripts by @vince62s in #277
• various fixes + Ministral3 (3B/8B/14B) support by @vince62s in #279
• Support batches of images with HunyuanOCR by @vince62s in #280
• Add xdrope for HunyuanOCR (even though does not seem to improve results) by @vince62s in #281

Full Changelog: 0.4.1...0.4.2

What's Changed

• FusedMoE (vllm) + PDF to markdown with DeepSpeed-OCR by @vince62s in #270
• support HunyuanOCR model + fixes by @vince62s in #273
• fix #272 by @vince62s in #274
• Demo HunyuanOCR by @vince62s in #275

What's Changed

• feat: Add attention entropy monitoring during training by @chillum-codeX in #250
• replace awq_ext rmsnorm by vllm rmsnorm by @vince62s in #260
• Better timing log and performance boost with fused layers by @vince62s in #261
• Allow estimation only (no step decoding) by @vince62s in #263
• hunyuan estimator by @vince62s in #264
• partial fix to lora_embedding training by @vince62s in #265
• Deepseekocr by @vince62s in #266
• Takeover PR #238 by @vince62s in #267
• update README by @vince62s in #268

New Contributors

• @chillum-codeX made their first contribution in #250

Full Changelog: 0.3.0...0.4.0