All pull requests

Pull requests

• Fix megatron_preprocess_data multi-worker segfault: spawn instead of fork

  #2726·shengliangxu opened Oct 9, 2026Collaborator·

  ·
  3

• Add GLM-5.3-Flash indexer scorer keyword arguments to vLLM fake-quant serving

  #2725·sychen52 opened Oct 9, 2026Contributor·

  ·
  4

• Speed up the diffusers example tests

  #2724·cjluo-nv opened Oct 9, 2026Collaborator·

  ·
  6

• fix(onnx): require onnxslim 0.1.78 for ONNX 1.21

  #2721·joe-nv opened Oct 9, 2026·

  ·
  4

• Add a NemotronH omni VL W4A4 PTQ + QAD tutorial for Megatron-Bridge

  #2720·yueshen2016 opened Oct 9, 2026Contributor·

  ·
  15

• fix(quantization): isolate optional Megatron imports

  #2719·joe-nv opened Oct 9, 2026·

  ·
  4

• Add a prebuilt DeepGEMM option to vLLM fake-quant serving

  #2717·sychen52 opened Oct 8, 2026Contributor·

  ·
  36

• Fix automatic device mapping for Nemotron VL in HF PTQ

  #2715·meenchen opened Oct 8, 2026Contributor·

  ·
  3

• Fix unsafe Megatron preprocessing worker startup

  #2713·kinjalpatel27 opened Oct 8, 2026Contributor·

  ·
  3

• docs: add AutoQuant KV mixed-precision quantization blog

  #2712·meenchen opened Oct 8, 2026Contributor·

  ·
  13

• Fix the GLM-5.1 alias: it is not the input_scale1 recipe

  #2711·shengliangxu opened Oct 8, 2026Collaborator·

  ·
  3

• Add upcast_ggml="bf16" export for evaluating GGML checkpoints in vLLM

  #2710·cjluo-nv opened Oct 8, 2026Collaborator·

  ·
  37

• Check IQ CUDA encoder ties per block instead of by count

  #2708·LinCanNerd opened Oct 8, 2026Contributor·

  ·
  1

  2

• Measure qtensor GPU memory tests with allocator stats

  #2707·LinCanNerd opened Oct 8, 2026Contributor·

  ·
  1

  2

• Support a new model (NemotronH_Omni_Reasoning_V3) in Megatron-Bridge PTQ/QAD and export

  #2704·yueshen2016 opened Oct 8, 2026Contributor·

  ·
  16

• Add a NemotronH aggressive-mse NVFP4 PTQ recipe

  #2703·juhi10071998 opened Oct 8, 2026Contributor·

  ·
  25

  

• Add rank-local tensor-parallel reload tests

  #2702·SloptimistPrime opened Oct 8, 2026·

  ·
  2

• 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

  cherry-pick-0.48.0
  #2697·yingguo-trt opened Oct 8, 2026Contributor·

  ·
  10

• 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