v0.9.0

Highlights

• support automatic mixed bits assignment by @wenhuach21 in #851
• optimize rtn for int woq by @wenhuach21 in #924
• support for model scope by @n1ck-guo in #957
• enhance auto device map and support XPU by @xin3he in #961
• support for immediate saving to reduce ram usage by @Kaihui-intel in #965
• update gguf alg ext by @n1ck-guo in #1026

What's Changed

• Fix rtn tuning_device issue by @Kaihui-intel in #893
• fix vlm gguf ut by @n1ck-guo in #895
• update alg_ext.abi3.so with python compatible version by @chensuyue in #894
• move ste from quant to round for nvfp4 by @xin3he in #889
• Add GPT-OSS quant support by @yiliu30 in #887
• better help printing information by @n1ck-guo in #883
• speedup quant and evaluation, fix recompile issue by @xin3he in #897
• fix nvfp act quantization bug by @WeiweiZhang1 in #891
• support automatic mixed bits assignment by @wenhuach21 in #851
• try to fix gguf vram issue on windows by @wenhuach21 in #886
• remove numba from requirments by @yiliu30 in #905
• Extend mxfp loading dtypes by @yiliu30 in #907
• block dataset logger info by @n1ck-guo in #908
• fix torch compile issue in AutoScheme by @wenhuach21 in #909
• Revert "Extend mxfp loading dtypes" by @wenhuach21 in #915
• support disable_opt_rtn in auto-scheme by @wenhuach21 in #913
• fix llama 4 ut by @n1ck-guo in #896
• Add numba for cpu lib by @yiliu30 in #919
• Loosen the packing restrictions for mxfp&nvfp by @WeiweiZhang1 in #911
• Extend mxfp loading dtypes by @yiliu30 in #916
• Fix act config exporting for mixed schemes by @WeiweiZhang1 in #903
• optimize rtn for int woq by @wenhuach21 in #924
• fix bug of gguf and support for LiquidAI/LFM2-1.2B by @n1ck-guo in #927
• remove numpy<2.0 limitation by @xin3he in #921
• enable regex quantization config saving for mixed bits by @WeiweiZhang1 in #825
• Fix Flux tuning issue by @mengniwang95 in #936
• gguf support for inclusionAI/Ling-flash-2.0 by @n1ck-guo in #940
• remove low_cpu_mem by @n1ck-guo in #934
• Add compatibility test by @XuehaoSun in #918
• Add commit hash to version by @XuehaoSun in #941
• gguf weight type align with original, output.weight, token_embed by @n1ck-guo in #900
• support attention mask in user's dataset by @wenhuach21 in #930
• Add diffusion README by @mengniwang95 in #923
• update readme by @wenhuach21 in #949
• refactor utils file by @n1ck-guo in #943
• update readme for sglang support by @WeiweiZhang1 in #953
• update gguf and support for CompressedLinear by @n1ck-guo in #950
• Reduce AutoSchem VRAM usage by up to 10X by @wenhuach21 in #944
• add self attribution and fix avg_bits error by @xin3he in #956
• add logo by @wenhuach21 in #960
• refine AutoScheme readme/code by @wenhuach21 in #958
• update readme by @wenhuach21 in #962
• fix critic disable_opt_rtn regression by @wenhuach21 in #963
• [1/N] Initial vllm-ext evaluation support (MXFP4 MOE) by @yiliu30 in #935
• fix bug of imatrix contains 0 by @n1ck-guo in #955
• fix rtn bug by @mengniwang95 in #966
• enhance flux doc by @mengniwang95 in #967
• clean code by @wenhuach21 in #968
• support for model scope by @n1ck-guo in #957
• merge main branch to alg_ext by @wenhuach21 in #970
• fix cuda CI backend issue, fixtypo by @WeiweiZhang1 in #974
• disable compile packing by default by @yiliu30 in #975
• enhance auto device map and support XPU by @xin3he in #961
• refine readme by @wenhuach21 in #978
• cli support for positional arguments model by @n1ck-guo in #979
• update bits in UT by @xin3he in #986
• fix guff scheme and device_map bug by @n1ck-guo in #969
• add support for Magistral-Small by @n1ck-guo in #980
• support model_dtype and fix bug of scheme contains quotes, mllm eval by @n1ck-guo in #985
• fix bug of cannot create adam compressor by @n1ck-guo in #992
• [CI] Update python to 3.12 and torch to 2.8.0 by @XuehaoSun in #741
• fix lm head bug and rm clear_mem_reach_threhold by @wenhuach21 in #997
• Reduce peak gpu memory usage and support moe estimation by @xin3he in #981
• fix cuda ut bug by @n1ck-guo in #999
• fix mllm device_map ut by @Kaihui-intel in #1000
• refine md tables by @WeiweiZhang1 in #994
• Refine exllamav2 ut by @WeiweiZhang1 in #1001
• Support for immediate saving to reduce ram usage by @Kaihui-intel in #965
• Fix diffusion multi-device ut issue by @mengniwang95 in #1002
• fix multiple devices map issue in calibration by @wenhuach21 in #1003
• Fix non auto device map by @WeiweiZhang1 in #1005
• fix multiple devices issue in Compressor and AutoScheme by @wenhuach21 in #1007
• fix cuda low_cpu_mem_usage ut by @Kaihui-intel in #1010
• Fix param missing bug by @mengniwang95 in #1008
• add device list to clear memory by @wenhuach21 in #1009
• Minor refactor for LLMC by @yiliu30 in #993
• fix one clear memory issue by @wenhuach21 in #1011
• add ut for gguf alg_ext and update so file by @n1ck-guo in #1012
• fix multi cuda ut bug by @n1ck-guo in #1014
• Including auto_scheme.default_alg into pypi by @chensuyue in #1018
• add num_device check for set_auto_device_map_for_block_with_tuning by @xin3he in #1021
• dispatch model with real max memory by @xin3he in #1022
• fix cuda ut by @n1ck-guo in #1020
• disable itrex format first by @WeiweiZhang1 in #998
• fix bug of lm_head and dispatch model,gguf eval by @n1ck-guo in #1025
• Fix the missing temporary name by @yiliu30 in #1029
• Reduce mem usage of GPT-OSS by @yiliu30 in #1013
• update gguf alg ext by @n1ck-guo in #1026
• optimize vram for gguf and add momentum by @wenhuach21 in #1031
• fix incorrect model name in readme by @wenhuach21 in #1035
• Bump into v0.9.0 by @XuehaoSun in #1024

Full Changelog: v0.8.0...v0.9.0