All pull requests

Pull requests

• Add vLLM NVFP4 MLA KV-cache fake quant

  #2620·sychen52 opened Oct 1, 2026Contributor·

  ·
  2

• Fix q/k/v quantizer wiring in register_attention_for_kv_quant

  #2618·javierdejesusda opened Oct 1, 2026Contributor·

  ·
  1

  6

• fix(speculative): keep the DFlash draft's RoPE frequencies in fp32 on a fresh convert

  #2617·h-guo18 opened Oct 1, 2026Contributor·

  ·
  4

• Share one CUDA encoder per IQ family

  #2615·cjluo-nv opened Oct 1, 2026Collaborator·

  ·
  6

• fix(onnx): support per-tensor scales in qdq_to_dq

  #2613·Yi-111-a opened Oct 1, 2026·

  ·
  1

  3

• docs: fix TensorRT-LLM and autotune links

  #2612·pratikgx opened Sep 30, 2026·

  ·
  2

• [5/5] llm_sparsity: export from a local checkpoint

  #2611·shengliangxu opened Sep 30, 2026Collaborator·

  ·
  2

• [4/5] Megatron export: copy non-model files from local disk only

  #2610·shengliangxu opened Sep 30, 2026Collaborator·

  ·
  2

• [3/5] hf_ptq: load from a local checkpoint, carry its files verbatim

  #2609·shengliangxu opened Sep 30, 2026Collaborator·

  ·
  2

• [2/5] HF exporters write off-index safetensors and non-model files

  #2608·shengliangxu opened Sep 30, 2026Collaborator·

  ·
  2

• [1/5] Add ensure_local_checkpoint and copy_non_model_files

  #2607·shengliangxu opened Sep 30, 2026Collaborator·

  ·
  4

• docs(qwen3.6): add the decode-time fix for SciCode non-termination

  #2606·kevalmorabia97 opened Sep 30, 2026Collaborator·

  ·
  37

• docs(ptq): explicitly select PyTorch base image for HF PTQ

  #2605·chadvoegele opened Sep 30, 2026Contributor·

  ·
  5

• Pack each IQ weight once and decode the IQ formats on CUDA

  #2604·cjluo-nv opened Sep 30, 2026Collaborator·

  ·
  3

• Keep an unset ONNX dimension in the reported shape

  #2603·SashaMIT opened Sep 30, 2026·

  ·
  2

• skills(specdec): add DFlash2 and LiLiCorr algorithm sheets

  #2602·yeyu-nvidia opened Sep 30, 2026Contributor·

  ·
  11

• Fix LiLiCorr exports for vLLM

  #2601·askliar opened Sep 30, 2026·

  ·
  2

• Fuse the LSQ NVFP4 fake quant forward and backward

  #2600·namgyu-youn opened Sep 30, 2026·

  ·
  6

• Tile static NVFP4 fake quant and use int64 offsets

  #2599·namgyu-youn opened Sep 30, 2026·

  ·
  6

• Default GPTQ to the fused NVFP4 update

  #2598·namgyu-youn opened Sep 30, 2026·

  ·
  2

• [OMNIML-5935] Fix hf_ptq overwriting valid pad_token with eos_token

  #2597·Edwardf0t1 opened Sep 30, 2026Contributor·

  ·
  4

• Trim redundant hf_ptq example test cases

  #2596·cjluo-nv opened Sep 29, 2026Collaborator·

  ·
  2

• Reduce diffusers test compute and TensorRT build time

  #2594·ajrasane opened Sep 29, 2026Contributor·

  ·
  6

• Make the QAT/QAD guide the central place for concepts, background, and framework selection

  #2590·danielkorzekwa opened Sep 29, 2026Contributor·

  ·
  16

  

• [Speculative Decoding] LiLiCorr: offline/streaming training and an optional calibration loss

  #2589·mrusanovsky opened Sep 29, 2026Contributor·

  ·
  16