V0.9.3 Release

Highlights

• Added ark backend by @Zhenzhong1 in #1075
• reduce vram usage for optimized RTN mode by @wenhuach21 in #1043
• Support alg_ext on windows by @chensuyue in #1082
• adjust 2/3 bits hyperparameters at auto-round-best by @wenhuach21 in #1081
• fix bug of outside block layers do not use best_params by @n1ck-guo in #1128
• Fix nvfp4&fp8 packing ram issue, refine all exporting format ram release for origin layer by @yiliu30 in #1129

What's Changed

• fix accuracy regression by @wenhuach21 in #1041
• add environment.md and remove AutoRoundMLLM usage in readme by @xin3he in #1042
• add memory monitor and import auto-scheme on demand by @wenhuach21 in #1049
• Update get_block_names func by @mengniwang95 in #1047
• set add_bos_token=True for llama model by @n1ck-guo in #1046
• Announce llmc integration by @yiliu30 in #1055
• [High Risk]reduce vram usage for optimized RTN mode by @wenhuach21 in #1043
• Add static FP8 attention support by @yiliu30 in #1045
• Enhance tokenizer saving by @mengniwang95 in #1057
• Revert "Add static FP8 attention support" by @yiliu30 in #1060
• refine readme by @wenhuach21 in #1063
• Fix transformers==4.57.1 in CI by @XuehaoSun in #1066
• Add static FP8 attention support by @yiliu30 in #1061
• Remove tbb by @yiliu30 in #1069
• support for gguf mixed q2_k_s by @n1ck-guo in #1059
• Add compatibility test for ARM by @XuehaoSun in #1073
• Optimize CPU CI pipelines by @XuehaoSun in #1071
• Export KV Scheme in LLMC config by @yiliu30 in #1068
• Add LLMC integration test by @yiliu30 in #1053
• Support transformers loading quantized moe model by @mengniwang95 in #1067
• update alg_ext and add ut by @n1ck-guo in #1064
• simplify what's new and add publication_list by @xin3he in #1070
• Add MXFP8 MOE/Linear and MXFP4 Linear by @yiliu30 in #1034
• improve accuracy for 2bit with auto-round-best by @wenhuach21 in #1078
• Support mxfp nvfp lmhead quant by @WeiweiZhang1 in #1051
• refine sampler by @wenhuach21 in #1077
• fix bf16 option in AutoScheme by @wenhuach21 in #1079
• adjust 3 bits hyperparameters at auto-round-best by @wenhuach21 in #1081
• Support alg_ext on windows by @chensuyue in #1082
• [vLLM Ext]Fix MXFP4 Quant by @yiliu30 in #1088
• remove numpy restriction for gptq kernel by @wenhuach21 in #1084
• Fix MXFP/NVFP + FP8 Attn/KV by @yiliu30 in #1086
• Remove accelerate version limitation by @chensuyue in #1090
• bump version to v0.9.3 by @chensuyue in #1091
• add system checker in backend by @wenhuach21 in #1097
• Refactor input normalization by replaying inputs for consistent preprocessing by @yiliu30 in #1094
• fix gguf acc and oom bug when iters > 0 by @n1ck-guo in #1098
• Add Python 3.14 to compatibility test pipeline by @XuehaoSun in #1096
• Fix typo in README.md by @xin3he in #1102
• refine lmhead ut by @WeiweiZhang1 in #1106
• Move packed res to cpu by @yiliu30 in #1104
• remove non-essential requirements by @n1ck-guo in #1103
• fix auto-scheme/alg-ext multiple devices issue by @wenhuach21 in #1107
• fix release version cuda ut fail by @n1ck-guo in #1110
• fix gguf packing device by @n1ck-guo in #1105
• fix quant fp8 model with iters=0 and scheme=nvfp4 by @n1ck-guo in #1114
• Move quantized block to cpu by @yiliu30 in #1115
• fix gguf extension bug by @wenhuach21 in #1116
• fix bug of triton pow wrong data_type when enable torch compile by @n1ck-guo in #1120
• Use modelscope cache in CPU UT by @XuehaoSun in #1124
• fix bug of outside block layers do not use best_params by @n1ck-guo in #1128
• Fix nvfp4&fp8 packing ram issue, refine all exporting format ram release for origin layer by @yiliu30 in #1129
• fix regression by @xin3he in #1135
• Upgrade llmc to main and add cuda UT by @yiliu30 in #1111
• Enable load MXFP4/MXFP8 + FP8 KV by @yiliu30 in #1095
• Remove duplicate packages by @XuehaoSun in #1139
• fix bug of auto scheme with user layer config by @n1ck-guo in #1133
• Add AutoRound binary build and publish workflow by @chensuyue in #1132
• update readme by @wenhuach21 in #1141
• update document for eval by @xin3he in #1140
• update windows binary for alg_ext by @chensuyue in #1142
• Fix device mismatch of nvfp fuse scale by @WeiweiZhang1 in #1143
• add low_cpu_mem_usage in cli by @n1ck-guo in #1146
• fix cuda ut fail by @n1ck-guo in #1144
• add llmc for cuda ut by @yiliu30 in #1145
• Added ark backend by @Zhenzhong1 in #1075

New Contributors

• @Zhenzhong1 made their first contribution in #1075

Full Changelog: v0.9.2...v0.9.3