V0.14.2 patch release

What's Changed

• Fix the mismatch between weight and scale in gguf format (#2014)
• Fix CLI ignore --scheme when bits/act_bits are unset (#2019)
• Fix gradient_accumulate_steps cli issue (#2041)

Full Changelog: v0.14.1...v0.14.2

V0.14.1 patch release

What's Changed

• Remove transformers version limit
• [ARK] feat: expose smooth_k (use_mean_bias) parameter from Python to C++ in SAGE attention kernels (#1995)

Full Changelog: v0.14.0...v0.14.1

V0.14.0 release

Highlights

• [Model Free] Add MXFP mix-precision quantization and support RoutedExperts by @xin3he in #1921
• [Model Free] Implement AutoScheme support for model-free packing by @xin3he in #1990
• Support Wan2.2 S2V Quantization by @changwangss in #1924
• Add other non-integer scheme support for AWQ by @WeiweiZhang1 in #1892
• DDP for multiple cards by @yiliu30 in #1897

Bug fixes and Improvements

• Refine code to support all devices in torch and hot fix for gemma4-unified by @wenhuach21 in #1879
• Feat: add MXFP4/MXFP8 quantization support (llmc_compressor format) and related tests by @xin3he in #1865
• Fix slow startup time of pytest coverage for unit tests by @XuehaoSun in #1899
• Feat: improve review-pr skill score from 76% to 90% by @yogesh-tessl in #1901
• Fix rotation issue by @wenhuach21 in #1903
• Enable MXFP clamp for model free by @xin3he in #1914
• Refactor and support for multi algorithms fusion by @n1ck-guo in #1920
• Enhance b70 check by @yiliu30 in #1937
• Refine AutoScheme logic for gguf, reduce memory consumption and improve speed by @wenhuach21 in #1916
• Fix SIGSEGV in pack-weight unpack/GEMM on malformed blob by @luoyu-intel in #1928
• Update PyTorch and torchvision versions to 2.12.1 and 0.27.1 by @XuehaoSun in #1943
• Fix few gguf bugs of hf-name and checkpoint name by @n1ck-guo in #1939
• Fix saving hadamard rotation model without rotation config by @lkk12014402 in #1933
• Implement shard parallelism configuration and update documentation by @xin3he in #1942
• Use shared tokenizer save helper in exporters by @changwangss in #1951
• Fix auto_scheme gguf format wrong bits if use tiny model by @n1ck-guo in #1953
• Adapt the latest sglang integration test Engine kwargs. by @changwangss in #1964
• [ARK] Add varlen support for sdpa and update tests by @luoyu-intel in #1957
• Add the wrapper for woqgemm by @Zhenzhong1 in #1965
• Skip fused_moe replacement when llm-compressor oneshot already linearizes experts by @changwangss in #1966
• Fix layer_config bug and CI unstable issue by @xin3he in #1949
• Fix gguf cuda ut by @n1ck-guo in #1975
• Rename the base rotation (deterministic/random hadamard) directory name. by @lkk12014402 in #1956
• Fix: avoid silent weight loss when OffloadManager fails to save a block by @xin3he in #1982
• Fix: promote None-initialized kwarg to list in forward_capture by @ramkrishs in #1979
• Fix few bugs cause less accuracy of gguf format by @n1ck-guo in #1960
• Fix: use spawn instead of fork for dataset subprocess on macOS by @ramkrishs in #1980
• [AutoScheme] Fix lm_head no_grad issue and support MOE model with shared mix_score by @xin3he in #1971
• Update CUDA UT to cu130 by @XuehaoSun in #1992
• Enable auto-round-bot for CI comments by @chensuyue in #1993
• Feat: add return_lse support for SDPA and SAGEV1 kernel — enable sequence-parallel (Ring Attention) merge by @luoyu-intel in #1991
• Fix cuda ut fail caused by used instructions for base model by @n1ck-guo in #1994
• Update rotation doc. by @lkk12014402 in #1999

New Contributors

• @yogesh-tessl made their first contribution in #1901
• @Entrpi made their first contribution in #1909
• @ramkrishs made their first contribution in #1979

Full Changelog: v0.13.1...v0.14.0

V0.13.1 patch release

What's Changed

• GGUF format add support for mtp quantization by @n1ck-guo in #1866
• Add bf16 + NHD layout support, refactor sage_dynamic_quant by @luoyu-intel in #1882
• Suppress misleading warning when detecting model type for GGUF export directories by @lvliang-intel in #1887
• Fix performance regression by @wenhuach21 in #1886
• Fix random rotation and update rotation doc. by @lkk12014402 in #1884
• Update auto-round-lib release package build by @chensuyue in #1895
• Fix CI coverage & bug grep issue by @chensuyue in #1893
• Fix gguf opt-rtn regression by @wenhuach21 in #1905
• Fix: guard zero-division in GGUF quant kernels to avoid NaN block scales by @Entrpi in #1909
• Fallback compute type on b70 if needed by @yiliu30 in #1904

Full Changelog: v0.13.0...v0.13.1

V0.13.0 release

Highlights

• Support MTP quantization by @xin3he in #1758

• Support one cpu/xpu backend by @Zhenzhong1 in #1723, @luoyu-intel in #1806, @Copilot in #1819

• Support model_free WOQ quantization by @xin3he in #1699, @xin3he in #1809

• Expanded support for new model architectures across LLMs, VLMs, diffusion, audio, and MoE models, including Gemma4 (#1655), MiMo-V2-Flash (#1718), BAGEL-7B-MoT (#1633), WAN2.2 (#1678), Ovis-Image-7B (#1616), Qwen-TTS and MiMo-Audio (#1810), and Qwen3.6-35B-A3B (#1705).

• Add compressed-tensors format export support for W4A16 and W8A16 by @thuang6 in #1669

• New architecture for auto_round by @n1ck-guo in #1542, #1761, #1781, #1796, in #1807, #1765, #1808, #1817, #1832

• Support AWQ based smoothing algorithm by @WeiweiZhang1 in #1749

Bug fixes and Improvements

• Reduce XPU memory usage in CI by @xin3he in #1812
• Reduce flux CUDA CI tuning memory from 30G to 5GB by @xin3he in #1694
• Support MXINT4 scheme by @mengniwang95 in #1666
• Remove IPEX related code, doc, and test by @xin3he in #1787
• Enhance dataset preprocessing memory management and fix hash failure by @xin3he in #1621
• Enhance quantization configuration support for mixed precision and schemes in utils and tests by @xin3he in #1643
• Force auto-scheme low_gpu to True in CLI by @wenhuach21 in #1653
• Enhance performance test by @XuehaoSun in #1610
• Fix auto-scheme accuracy drop bug w/o low_gpu, add CI test by @WeiweiZhang1 in #1658
• Add gptqmodel 6.0 compatible change and MTP quantization skip warning. by @xin3he in #1663
• Import gptqmodel in confest to avoid error in transformers by @xin3he in #1668
• Fix missing extra_config export for unsupported ignore_layers like mlp.gate by @lvliang-intel in #1660
• Update unit test requirements for compressed-tensors and transformers… by @XuehaoSun in #1670
• Fix omni model test CI issue by @lvliang-intel in #1667
• Refactor module access to use PyTorch get_submodule / set_submodule by @scopophobic in #1590
• Support longcat_next by @xin3he in #1637
• Stabilize Qwen3-Omni MoE weight fidelity test for matching NaNs by @lvliang-intel in #1681
• Fix hadamard transform weight dtype, using float32 as default and in-place transformed weight . by @lkk12014402 in #1665
• fp8_block bug fix by @mengniwang95 in #1693
• Support diffusion model AIDC-AI/Ovis-Image-7B quantization by @lvliang-intel in #1616
• Gate large FP4 packing test by GPU memory by @yiliu30 in #1696
• Add Claude skills for AutoRound by @lvliang-intel in #1686
• Add missing run_mllm entry point alias by @JGSphaela in #1695
• Rename scheme INT8_W8A8 to INT8 by @thuang6 in #1687
• Update mtp quant for special cases by @xin3he in #1691
• Update gaudi-docker to v1.24.0 & fix CUDA UT by @XuehaoSun in #1708
• Add support for gemma4 model by @n1ck-guo in #1655
• Ignore mtp.fc for qwen3_5 due to vllm failure by @xin3he in #1710
• Introduce INT4 support at the algorithm level by @wenhuach21 in #1641
• Refine int4 doc by @wenhuach21 in #1720
• Revert "ignore mtp.fc for qwen3_5 due to vllm failure (#1710)" by @xin3he in #1730
• Skip quantizing mtp.fc since vLLM doesn't support by @xin3he in #1731
• Create model_support_request.yml by @xin3he in #1738
• Remove threaded packing from exporters by @yiliu30 in #1719
• Reduce XPU memory usage with patch_xpu_sdpa_drop_causal_mask by @xin3he in #1716
• Add MLX format export support and AutoScheme for vlm support by @wenhuach21 in #1732
• Add warnings for lm_head activation scale fallback by @n1ck-guo in #1728
• Add support for MiMo-V2-Flash by @n1ck-guo in #1718
• Fix vllm CUDA CI by @XuehaoSun in #1750
• Fix hpu error by @n1ck-guo in #1766
• MTP split gate_up_proj and fix accu gap in rtn quantization by @xin3he in #1758
• Support gptqmodel 7.0.0 and fix bug in CI by @xin3he in #1772
• Optimize CUDA CI and Code Scan workflows by @XuehaoSun in #1770
• Fix accuracy regression and check it in CUDA CI by @xin3he in #1785
• Fix amp by @wenhuach21 in #1768, @wenhuach21 in #1767
• Fix incompatible weight names by @mengniwang95 in #1759
• Fix CT export metadata for KV cache and attention by @yiliu30 in #1752, @yiliu30 in #1861
• Enhance AutoRound Lib test workflow by @chensuyue in #1805, @chensuyue in #1801
• Fix mixed-precision accuracy regression when AutoScheme runs with CPU offloading and Hadamard rotation enabled by @lvliang-intel in #1753
• Reduce XPU memory usage in CI by @xin3he in #1812
• Add shared agent config layout by @yiliu30 in #1700
• Support ByteDance-Seed/BAGEL-7B-MoT quantization in w4a16 format by @lvliang-intel in #1633
• Adjust gguf tuning algorithm by @wenhuach21 in #1649, @wenhuach21 in #1824
• Feats: Quantize/save/evaluate the Wan-AI/WAN2.2 models in w4a16 format by @lvliang-intel in #1678
• Fix layer name mismatch of VLM(qwen3.5-2B) in hf loading by @xin3he in #1823
• Update dependencies in CI and installation scripts for cuda compatibility by @XuehaoSun in #1825
• Reduce RAM usage of quantizing VLM models and fix some issues of quantizing gemma4 by @lvliang-intel in #1791
• Add mimo-audio, Qwen-TTS model backbone quantization by @WeiweiZhang1 in #1810
• Reduce VRAM usage of quantizing VLM models by @lvliang-intel in #1777
• Support quarot/spinquant rotation before quantization by @lkk12014402 in #1797
• Support Exporting Block-Wise FP8 AR Format by @Zhenzhong1 in #1798
• Fix Gemma4 KeyError sliding_attention issue by @lvliang-intel in #1839
• Fix gpt-j-6b RTN RuntimeError by @lvliang-intel in #1848
• Sage fast sfm by @luoyu-intel in #1843
• Security: HTTP requests are performed without timeout safeguards by @tomaioo in #1683
• Dynamic map checkpoint naming based on model objective. by @xin3he in #1840
• Refine/fix gptq format by @wenhuach21 in #1853
• Fix save_quantized log conflict by @WeiweiZhang1 in #1845
• Fix Qwen Omni quantization model issue for long form audio generation by @lvliang-intel in #1698
• Fix bug of qwen and gguf export by @n1ck-guo in #1846
• Refactor quarot/spinquant rotation with simplying code. by @lkk12014402 in #1849
• Fix gemma4 crash issue during quantizing by @lvliang-intel in #1860
• Fix SDPA bug by @luoyu-intel in #1862
• Fix special-model predef...

V0.12.3 patch release

What's Changed

• fp8_block bug fix by @chensuyue in #1736
• Remove threaded packing from exporters by @yiliu30 in #1740

Full Changelog: v0.12.2...v0.12.3

v0.12.2 patch release

fix autoscheme accuracy drop bug w/o low_gpu #1658
add gptqmodel 6.0 compatible change #1663
force auto-scheme low_gpu to beTrue in CLI #1653

V0.12.0 release

Highlights

• Support hadamard transform for mxfp4 with RTN or AutoRound method by @lkk12014402 in #1515, @lkk12014402 in #1607
• Support transformers 5.0 by @n1ck-guo in #1474, @lvliang-intel in #1475, @lvliang-intel in #1475, @lvliang-intel in #1535, @XuehaoSun in #1604, @xin3he in #1627, @xin3he in #1626
• Support Qwen3 and Qwen2.5 Omni model quantization by @lvliang-intel in #1404, @lvliang-intel in #1568, @xin3he in #1598
• Support GLM-Image model quantizaiton by @lvliang-intel in #1512
• Support block-wise fp8 quant by @mengniwang95 in #1487
• Support MTP module WOQ quantization and step3p5 model quantization by @xin3he in #1526, @xin3he in #1554
• Enable CUDA CI Test @XuehaoSun in #1473
• Enable Performance CI Test by @XuehaoSun in #1561

Bug fixes and Improvements

• Refactor evaluation in tests to use evaluate_accuracy function by @xin3he in #1402
• Remove require_intel_extension_for_pytorch and fix TypeError: unhashable type: 'set' by @xin3he in #1425
• Fix cuda model ut [glm4][Molmo] by @Kaihui-intel in #1428
• Update BackendInfos for AutoGPTQ based on transformer version and save g_idx for gptqmodel backend by @xin3he in #1429
• Refine global scale calculation to blockwise by @WeiweiZhang1 in #1421
• Support GPTQ_FORMAT for "gptqmodel:exllamav2" backend by @xin3he in #1434
• Update diffusion README by @mengniwang95 in #1439
• Enable new cpu pool for CI test by @chensuyue in #1435
• Update auto-round-lib README.md by @chensuyue in #1437
• Update logic for compatibility with gptqmodel:exllamav2 backend by @xin3he in #1438
• Update readme for Intel xpu usage by @chensuyue in #1441
• Fix bug of evaluate_accuracy by @xin3he in #1430
• Attach act_max_hook for FP8 model by @yiliu30 in #1447
• [Regression] fix FP8_STATIC loading by @xin3he in #1452
• Fix requirements packaging for source distribution by @timkpaine in #1455
• Fix: preserve classmethod descriptor in from_pretrained monkey patch by @yiliu30 in #1460
• Support load FP8 model on HPU by @yiliu30 in #1449
• Update compatibility test by @XuehaoSun in #1463
• Support multiple device evaluation for activation quantized model by @wenhuach21 in #1394
• Fix KeyError when GPU is missing from accelerate max_memory by @lvliang-intel in #1457
• Support glm5 by @wenhuach21 in #1466
• Fix cuda ut workflow by @XuehaoSun in #1472
• Fix: qwen3-next NVFP4 quantization returns act_max incorrectly by @xin3he in #1470
• Reduce AutoGPTQ priority and add transformer requirement by @xin3he in #1467
• Add envs.py for environment checks and update imports in test files by @xin3he in #1432
• Fix HPU CI: apply temporary version restriction on compressed_tensors by @XuehaoSun in #1479
• Fix typo: remove extra backtick in comment by @04cb in #1481
• Support qwen3_5 moe by @wenhuach21 in #1476
• Add qwen35 ut by @wenhuach21 in #1482
• Grep core dumped issue in UT test by @chensuyue in #1484
• Update torch to 2.10.0 in CPU CI by @XuehaoSun in #1492
• Update MOE linear_loop implementation for speedup and matching name in vLLM by @xin3he in #1478
• Add handle_generation_config function to manage model generation_config saving failure by @xin3he in #1448
• Force ignore mlp.gate by @xin3he in #1501
• Update README.md: Add nightly installation instructions for auto-round by @chensuyue in #1505
• Fix FP8 quantizer for Transformers v4 by @yiliu30 in #1504
• Support minimax_m2 ignore layer: block_sparse_moe.gate by @xin3he in #1508
• Reduce ram&vram usage for vlm calib stage by @WeiweiZhang1 in #1488
• [BUG] update moe check logic and make .gate ignore general by @xin3he in #1517
• Fix low_gpu default value, refine doc by @WeiweiZhang1 in #1520
• Release CUDA memory in WeightConverter and avoid meaningless print by @xin3he in #1498
• Fix CUDA UT and HPU UT by @XuehaoSun in #1514
• Fix bug of quantizing Z-image by @xin3he in #1516
• Enable more xpu instance for xpu CI by @chensuyue in #1521
• Fix Inference tensors do not track version counter error by @mengniwang95 in #1502
• Refactor build process to use 'uv build' instead of 'python setup.py' by @XuehaoSun in #1495
• Enable --eval for diffusion model by @xin3he in #1522
• Support gpt-oss mxfp4 directly loading by @xin3he in #1401
• Support diffusion model saving by @mengniwang95 in #1519
• Optimize CPU RAM peak memory during quantization by @lvliang-intel in #1386
• Fix dynamic int8 w8a8 export issue with tuning by @thuang6 in #1525
• Fix missing trust_remote_code in unfused_moe by @xin3he in #1532
• Fix sglang ut by @mengniwang95 in #1541
• Fix alg_ext torch compile issue by @wenhuach21 in #1540
• Add markdown-link-check by @XuehaoSun in #1531
• Fix google/gemma-3-4b-it by @xin3he in #1547
• Fix wrong match of ignore_layers and use warning instead of error for mismatch by @xin3he in #1553
• Use AR_WORK_SPACE/offload/ as base dir for offload temp files by @yiliu30 in #1552
• Fix shard finalization issue to prevent skipping non-LLM tail layers by @lvliang-intel in #1548
• Fix missing change for compress_layer_names and add UTs by @xin3he in #1556
• Fix MoE for llama/gpt-oss by @yiliu30 in #1557
• Fix reload issue for lm_head quantization and update CI to cover it by @xin3he in #1563
• Refine model save folder name by @Kaihui-intel in #1549
• Add gptqmodel:awq backend support and keep autoawq WQLinear_GEMM for compatibility by @WeiweiZhang1 in #1545
• Refine CPU RAM offload strategy by @lvliang-intel in #1544
• Fix deadlock issue when do quantization on hpu. by @lkk12014402 in #1571
• Fix compatibility with latest compressed_tensors compressor refactor by @yiliu30 in #1576
• [Regression] fix copy missing tensors for llama4 by @xin3he in #1579
• Disable torch compile for gguf with alg_ext by @wenhuach21 in #1582
• Update awq backend infos, refine UT by @WeiweiZhang1 in #1581
• Fix CI test FP8 static backend matching failure by @lvliang-intel in #1589
• Fix packing nvfp/mxfp max_wokers & extend xpu ut by @Kaihui-intel in #1555
• Fix contiguous issue by @xin3he in #1594
• Update ipex backend to be deprecated. by @xin3he in #1599
• Fix CUDA UT for 0.12.0 by @xin3he in #1600
• Fix gguf bug for qwen3.5 moe and some small models by @n1ck-guo in #1575
• Sup...

v0.10.2 patch release

#1455

V0.10.1 patch release

What's Changed

• [Regression] fix FP8_STATIC loading (#1452)

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