intel/auto-round

Pull Requests Commits

refactor: decouple tuning data handling from SignRound (#2524)

changwangss committed 11 hours ago

Verified 42db5d3a

fix: register copied-through MTP modules in compressed-tensors ignore list (#2546)

lkk12014402 committed 15 hours ago

Verified 54907c95

feat: add support for enable_torch_compile flag in NVFP4 E5M3 quantization functions and tests (#2619)

xin3he committed 20 hours ago

Verified 09530317

feat: chunked tuning for lm_head-class layers (#2383)

avtc committed 20 hours ago

Verified 7c47e73f

Recurrent Residual Quantization (RRQ) for LLMs (#2308)

luoyu-intel committed 1 day ago

Verified b58069ea

fix: support nested Qwen2.5-Omni thinker layout (#2564)

XuehaoSun committed 1 day ago

Verified 8680a3c4

fix: preserve GLM nextn weights during checkpoint copying (#2430)

Shubham-Padkonde committed 1 day ago

Verified cd261ae9

add Qwen27B data and WnA16 INT exp results to awq_detail doc (#2561)

WeiweiZhang1 committed 1 day ago

Verified 8a3cb2ac

support teq algo (#2301)

WeiweiZhang1 committed 1 day ago

Verified 17e7cb6d

fix: disable trust_remote_code by default for security (#2457)

chensuyue committed 1 day ago

Verified 298894c2

feat: implement optimized routing for sequential MoE forward (#2615)

xin3he committed 1 day ago

Verified 55d1d45f

Refine device operations into a unified DeviceManager and ARDevice abstraction (#2344)

lvliang-intel committed 1 day ago

Verified 1c447696

Clean up RunPod GPU targets for CUDA unit tests (#2618)

XuehaoSun committed 1 day ago

Verified 933da9bd

perf: vectorize Qwen4-Exp QSA indexer to speed up full-attention blocks (#2579)

lvliang-intel committed 1 day ago

Verified b1f816bb

fix: only export algorithm-owned fields in quantization_config (#2605)

n1ck-guo committed 1 day ago

Verified ec55f2e0

refactor: separate schemes from alg configs (#2594)

n1ck-guo committed 1 day ago

Verified bca92fbb

fix: honor max_scale in NVFP4 v2 global quantizer (#2585)

0z5a committed 2 days ago

Verified a85997e1

fix: use calibrated NVFP4 v2 tensor maxima (#2586)

0z5a committed 2 days ago

Verified a2641e77

fix(autoscheme): do not warn on BF16+int mixed data_types (#2600)

ompatil1906 committed 2 days ago

Verified 4a7d1ff6

Fix quant block list (#2610)

WeiweiZhang1 committed 2 days ago

Verified 6f1d6652

Refactor special_model_handler into a package by model/arch family (#2545)

Cross2pro committed 2 days ago

Verified a80cb096

fix: always save key AutoRound hyper-parameters to config.json (#2597)

n1ck-guo committed 2 days ago

Verified c09906f5

Improve CI UT coverage (#2560)

chensuyue committed 2 days ago

Verified db3a3ee1

fix: use one-dimensional masks for concatenated calibration data (#2607)

changwangss committed 2 days ago

Verified 520f5d4c

Fix: ARK `return_lse` returns identical LSE for every 16-query-row block (and wrong LSE under `smooth_k=True`) (#2578)

luoyu-intel committed 2 days ago

Verified ee6e601d

feat: add MXFP4 export to SVDQuant Omni format (#2535)

yiliu30 committed 2 days ago

Verified edf0d278

fix: reset BOS/EOS bookkeeping per sample when packing calibration data (#2591)

urakozz committed 3 days ago

Verified b3d11504

fix: update test_quant_lm_head to use tiny untied Qwen model path and assert quantization config (#2604)

chensuyue committed 3 days ago

Verified c9c3ce88

Enhance model-free quantization with improved error handling and shard management (#2566)

xin3he committed 3 days ago

Verified 2679ca9a

[fix] use BF16 KV for reference output; use QDQ NVFP4 KV for tuning inference. (#2592)

xin3he committed 3 days ago

Verified a6efc3e4

Older