V0.15.0 release

Highlights

• Experimentally support algorithm combination
• Support SVDQuant by @changwangss in #2095
• Feature: speedup autoscheme with multi-process and cache files by @xin3he in #2083
• Stream AutoScheme's per-layer sensitivity scoring block-by-block by @aquilarubra in #2063
• Add OpenCodeInstruct calibration dataset by @changwangss in #2096
• Support fused MoE quantization and export for GGUF format by @n1ck-guo in #2072
• Enable torch.compile by default and update related documentation by @xin3he in #2070
• Ark: add SpargeAttn support by @yiliu30 in #2027

Bug fixes and Improvements

• Ark: Support Dense Gemm using sycl-tla by @Zhenzhong1 in #1968
• Add moe prefill/ decode with int2/int4/int8 sym/asym and fp8 e4m3 e5m2 by @a32543254 with @Copilot in #1813
• Fix checkpoint remap by @yiliu30 in #2001
• Fix: implement dedicated cache for streaming source shards by @xin3he in #2021
• Update: torch version to 2.13.0 by @XuehaoSun in #2045
• Fix: invalid group reference in revert_checkpoint_conversion_mapping by @A-S-Manoj in #2035
• Fix: handle router_logits tuple for compatibility with transformers >= 5.13.0 by @xin3he in #2046
• Enhance: logging and data loading for AutoScheme and model-free compatibility by @xin3he in #2050
• Fix: add triton to requirements for LLMC CPU tests by @XuehaoSun in #2054
• Stabilize MoE prefill accuracy test by @a32543254 in #2060
• Fix AutoScheme embedding budget and gguf mixed-bits config for multimodal models by @n1ck-guo in #2055
• Add CUDA CI coverage by @XuehaoSun in #2074
• Refactor compressor and quantizer by @wenhuach21 in #2039
• Fix Flux diffusion state propagation during block replay by @changwangss in #2086
• Fix: gguf format gemma4 High RAM and glm-4.7 error by @n1ck-guo in #2084
• Feature: speedup autoscheme with multi-process and cache files by @xin3he in #2083
• Decoupling oneDNN by @Zhenzhong1 in #2049
• Ark: add SpargeAttn support by @yiliu30 in #2027
• Refine alg and support lfq by @wenhuach21 in #2082
• Fix dataset and batch_size not taking effect in CLI by @changwangss in #2102
• Stream AutoScheme's per-layer sensitivity scoring block-by-block by @aquilarubra in #2063
• Fix: MXFP model free auto-round format bug by @xin3he in #2106
• Fix: skip exporting fp8 attention q_max by @yiliu30 in #2103
• Disable parallel for non low_gpu_mem_usage by @xin3he in #2116
• Automatically select calibration datasets for code models by @changwangss in #2107
• Refactor scheme,layer_config and formats by @n1ck-guo in #2013
• Support fused MoE quantization and export for GGUF format by @n1ck-guo in #2072
• Enable torch.compile for selected tests and update related configurations by @xin3he in #2124
• Fix: XPU crash due to circle call of nonzero() and index_select() by @xin3he in #2117
• Refator UT and improve code coverage by @lvliang-intel in #2057
• Fix: sparge on torch2.13 by @yiliu30 in #2130
• Add llmc ut back by @yiliu30 in #2131
• Support one more mxfp4 variant by @wenhuach21 in #2132
• Add python 3.14t compatibility tests by @XuehaoSun in #2110
• Enhance unit tests with timeout handling and stability improvements by @XuehaoSun in #2135
• Fix performance regression bug by @n1ck-guo in #2136
• Feat: add per-head fp8 kv attention calibration by @zeel2104 in #2044
• Fix: prevent quantized layers from being reinitialized by Transformers by @n1ck-guo in #2147
• Refactor unit test scripts by @XuehaoSun in #2148
• Migrate cpu sdpa from NS flash attn by @jijiaz in #2092
• Refine eval, add more option by @WeiweiZhang1 in #2114
• Add timeout decorators to improve test case stability by @XuehaoSun in #2155
• Feat: enable search scale for mxfp in model-free by @xin3he in #2141
• Refine awq, bugfix by @WeiweiZhang1 in #2115
• Set the quant_method to auto-round by default by @Zhenzhong1 in #2138
• Ark: Optimize SYCL-TLA woqgemm_s8 on B70 by @Zhenzhong1 in #2133
• Add cpu & xpu nightly test by @XuehaoSun in #2149
• Add support for kimi_k25 INT4 packed source models and related tests by @xin3he in #2156
• Refactor: optimize expert selection and output handling in moe forward by @xin3he in #2209
• Fix: awq torch_compile issue and set disable_opt_rtn to true by default for awq by @WeiweiZhang1 in #2208
• Add enable_torch_compile mark to SGLang and vLLM throughput tests by @xin3he in #2226
• Fix: FP8 quantization config for native fp8 export by @Zhenzhong1 in #2228
• Fix: AutoScheme parallel scoring OOMs (MoE memory, VLM gate) and disk-stream tuning crash by @avtc in #2220
• Fix: remove stale quantizer access in dual-transformer diffusion by @changwangss in #2230
• Fix: disable torch compile in rtn/opt-rtn/iters<10 and fix glm-5.3 tuning exception by @wenhuach21 in #2246

New Contributors

• @A-S-Manoj made their first contribution in #2035
• @a32543254 made their first contribution in #2060
• @aquilarubra made their first contribution in #2063
• @zeel2104 made their first contribution in #2044
• @jijiaz made their first contribution in #2092

Full Changelog: v0.14.0...v0.15.0