All pull requests

Pull requests

• fix: use calibrated NVFP4 v2 tensor maxima

  #2586·0z5a opened Nov 4, 2026·

  ·
  1

  8

• fix: honor max_scale in NVFP4 v2 global quantizer

  #2585·0z5a opened Nov 4, 2026·

  ·
  1

  5

• fix: restore torch.compile for NVFP4 quantization

  #2582·WeiweiZhang1 opened Nov 4, 2026Contributor·2 of 4 tasks·

  ·
  1

• fix: support Diffusers weight filenames in low-memory mode

  #2580·changwangss opened Nov 4, 2026Contributor·4 tasks·

  ·

  · 0.17.0
  1

  10

• perf: vectorize Qwen4-Exp QSA indexer to speed up full-attention blocks

  #2579·lvliang-intel opened Nov 4, 2026Contributor·1 of 4 tasks·

  ·
  11

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

  #2578·luoyu-intel opened Nov 4, 2026Contributor·2 of 4 tasks·

  ·
  1

  13

• Add fuzz tests for parsing and path safety in auto_round

  #2577·chensuyue opened Nov 4, 2026Contributor·4 tasks·

  ·
  5

• fix: drop redundant --enable_minmax_tuning CLI flag

  #2576·Cross2pro opened Nov 4, 2026Contributor·2 of 4 tasks·

  ·
  1

  4

• feat: shard the AWQ pipeline under --data_parallel

  #2575·avtc opened Nov 3, 2026Collaborator·3 of 4 tasks·

  ·
  4

• fix: default AWQ to optimized RTN (disable_opt_rtn=False)

  #2572·WeiweiZhang1 opened Nov 3, 2026Contributor·2 tasks·

  ·
  12

• [High Risk] Support Tensor Parallelism and LoRA, Optimize FP4 MoE Tuning Speed, and Refactor the Code

  ready-for-review
  #2567·wenhuach21 opened Nov 2, 2026Contributor·4 tasks·

  ·
  51

• Enhance model-free quantization with improved error handling and shard management

  #2566·xin3he opened Nov 2, 2026Contributor·1 of 4 tasks·

  ·
  18

• fix: support nested Qwen2.5-Omni thinker layout

  #2564·XuehaoSun opened Nov 2, 2026Contributor·3 of 4 tasks·

  ·
  12

• Investigate transient CodeQL C/C++ upload failure

  #2563·chensuyue with CopilotAI opened Oct 30, 2026·4 tasks·

  ·
  1

  

• Refactor quantization functions and enhance backend compatibility checks

  #2562·chensuyue opened Oct 30, 2026Contributor·4 tasks·

  ·
  6

• add Qwen27B data and WnA16 INT exp results to awq_detail doc

  #2561·WeiweiZhang1 opened Oct 30, 2026Contributor·4 tasks·

  ·
  1

• Improve CI UT coverage

  #2560·chensuyue opened Oct 30, 2026Contributor·4 tasks·

  ·
  5

• Fix FP8 override aliases and validation in mixed INT export

  #2555·changwangss opened Oct 29, 2026Contributor·

  ·
  4

• feat: Enhance MoE routing by introducing batch processing and improv

  #2551·xin3he opened Oct 29, 2026Contributor·2 of 4 tasks·

  ·
  17

• fix: register copied-through MTP modules in compressed-tensors ignore list

  #2546·lkk12014402 opened Oct 27, 2026Contributor·

  ·

  · 0.17.0
  9

• Refactor special_model_handler into a package by model/arch family

  #2545·Cross2pro opened Oct 27, 2026Contributor·1 of 3 tasks·

  ·
  1

  19

• feat: share one scheme across vLLM-fused projections in AutoScheme by default

  #2542·lvliang-intel opened Oct 27, 2026Contributor·1 of 4 tasks·

  ·
  12

• refine export and format code and provide unify api

  #2540·n1ck-guo opened Oct 27, 2026Contributor·4 tasks·

  ·
  21

• fix: support Nemotron-H Omni without mamba_ssm

  #2537·urakozz opened Oct 26, 2026·

  ·
  9

• feat: add MXFP4 export to SVDQuant Omni format

  #2535·yiliu30 opened Oct 26, 2026Contributor·3 of 4 tasks·

  ·
  13