All pull requests

Pull requests

• Add Qwen FSDP2 loading and export tests

  #2699·SloptimistPrime opened Oct 8, 2026·

  ·
  2

• Add FSDP2 AutoQuant search and resume tests

  #2698·SloptimistPrime opened Oct 8, 2026·

  ·
  2

• [https://nvbugspro.nvidia.com/bug/6886263] Fix FlashInfer sparse attention packed KV cache support

  #2697·yingguo-trt opened Oct 8, 2026Contributor·

  ·
  9

• Fix ONNX MXFP8 get_amax for 2D weights blocked on axis 0

  #2696·raashish1601 opened Oct 7, 2026·

  ·
  2

• Start NumPy histogram calibration bins at 0

  #2695·raashish1601 opened Oct 7, 2026·

  ·
  2

• Keep all-zero blocks finite in FP8QTensor.quantize

  #2694·raashish1601 opened Oct 7, 2026·

  ·
  9

• Round MXFP4 export cast half to even

  #2693·raashish1601 opened Oct 7, 2026·

  ·
  2

• Honour dflash_freeze_base_model=False in the DFlash base forward

  #2692·yeyu-nvidia opened Oct 7, 2026Contributor·

  ·
  2

• Add the missing datasets dependency to the speculative-decoding example

  #2690·yeyu-nvidia opened Oct 7, 2026Contributor·

  ·
  2

• Make the multimodal OpenAI generation path usable with vLLM

  #2689·yeyu-nvidia opened Oct 7, 2026Contributor·

  ·
  4

• [8/8] Add an external-draft training runbook

  #2688·phicksy opened Oct 7, 2026·

  ·
  5

• [7/8] Add an optional vocabulary swap for cross-tokenizer external drafts

  #2687·phicksy opened Oct 7, 2026·

  ·
  4

• Weight the IQ2 encoders by the square root of GPTQ's Hessian diagonal

  #2686·cjluo-nv opened Oct 7, 2026Collaborator·

  ·
  5

• [4/4] [6466805] Integrate helper-owned Dynamo ONNX export

  #2685·ajrasane opened Oct 7, 2026Contributor·

  ·
  2

• [3/4] [6466805] Normalize canonical Dynamo quantized weight paths

  #2684·ajrasane opened Oct 7, 2026Contributor·

  ·
  9

• [1/4] [6466805] Make quantization carriers safe for CPU export capture

  #2682·ajrasane opened Oct 7, 2026Contributor·

  ·
  5

• Deprecate the Diffusers ONNX/TensorRT deployment workflow

  #2680·Zhehao-Hu opened Oct 6, 2026Contributor·

  ·
  20

• [6860073] Fix ONNX quantization and separate DLA IQ/EQ workflows

  #2678·ajrasane opened Oct 6, 2026Contributor·

  ·
  12

• Restrict final RC cherry-pick audits to QA-validated committed NVBugs

  #2677·chadvoegele opened Oct 6, 2026Contributor·

  ·
  6

• [2/2] Support Qwen text-only FSDP2 AutoQuant loading and export

  #2676·meenchen opened Oct 6, 2026Contributor·

  ·
  2

• Add activation-only scale search option for AWQ calibration

  #2675·thuiaaa opened Oct 6, 2026·

  ·
  1

  2

• Register the FP8 per-tensor real-quant GEMM again

  #2673·LinCanNerd opened Oct 6, 2026·

  ·
  1

  2

• Fix import modelopt.torch without the [hf] extra

  #2671·LinCanNerd opened Oct 6, 2026·

  ·
  2

• Fix MGDLoss teacher gradient leak and enforce abstract loss balancer

  #2668·LinCanNerd opened Oct 6, 2026·

  ·
  4

• [1/2] Support generic FSDP2 weight and KV AutoQuant

  #2664·meenchen opened Oct 5, 2026Contributor·

  ·
  17