All pull requests

Pull requests

• [OMNIML-6079] [3/3] Make Kimi-K3 exports from the general PTQ recipes load in vLLM

  #3370·shengliangxu opened Nov 12, 2026Collaborator·

  ·

• [OMNIML-6533] [2/2] Move the vLLM quantization plugins' named-model code into modelopt/torch/models

  #3368·shengliangxu opened Nov 12, 2026Collaborator·2/2·

  ·
  5

• [OMNIML-6533] [1/2] Move the HF quantization plugin's remaining named-model code into modelopt/torch/models

  #3367·shengliangxu opened Nov 12, 2026Collaborator·1/2·

  ·
  11

• [OMNIML-6104] Prepare QDQ weights during PTQ and preserve GGML initialization

  #3366·hychiang-git opened Nov 12, 2026Contributor·

  ·
  8

• Avoid repeated calibration inference for shared ONNX AWQ activations

  #3365·Tracy-123-dev opened Nov 12, 2026Contributor·

  ·
  3

• Add Kimi-K3.1-Pro NVFP4 recipe: recipe-driven per-module formats in the K3 converter

  #3363·juhi10071998 opened Nov 12, 2026Contributor·

  ·
  32

• [OMNIML-6531] [5/6] Decompress packed experts with fewer kernels

  #3361·shengliangxu opened Nov 12, 2026Collaborator·5/5·

  ·
  4

• [OMNIML-6531] [4/6] Keep compressed-tensors experts packed under expert parallelism

  #3360·shengliangxu opened Nov 12, 2026Collaborator·4/5·

  ·
  4

• [OMNIML-6531] [3/6] Keep a compressed-tensors checkpoint's experts packed as stored

  #3359·shengliangxu opened Nov 12, 2026Collaborator·3/5·

  ·
  3

• [OMNIML-6531] [2/6] Quantize DeepSeek-V4's packed FP4 experts as transformers stores them

  #3358·shengliangxu opened Nov 12, 2026Collaborator·2/5·

  ·
  3

• [OMNIML-6531] [1/6] Keep an MXFP4 checkpoint's experts packed through PTQ

  #3357·shengliangxu opened Nov 12, 2026Collaborator·1/5·

  ·
  4

• [OMNIML-6493] [3/3] Add --calib_pack to hf_ptq to pack the calibration text into full rows

  #3356·shengliangxu opened Nov 12, 2026Collaborator·

  ·
  7

• [OMNIML-6493] [2/3] Dump hf_ptq's Python stacks on SIGTERM, on a crash and, on request, periodically

  #3355·shengliangxu opened Nov 12, 2026Collaborator·

  ·
  11

• [OMNIML-6493] [1/3] Read hf_ptq's calibration text once, on rank 0, for every rank

  #3354·shengliangxu opened Nov 12, 2026Collaborator·

  ·
  23

• docs: Training standalone speculative decoding drafts for Fast Inference on Groq 3 LPX

  #3353·phicksy opened Nov 11, 2026Contributor·

  ·
  32

• docs: refresh the external draft guide

  #3352·phicksy opened Nov 11, 2026Contributor·

  ·
  3

  

• Derive the GGML IQ super-block scale and match each block's output energy for max and imatrix

  #3351·cjluo-nv opened Nov 11, 2026Collaborator·

  ·
  7

• [OMNIML-6281] Deploy-test DeepSeek-V4.1-Flash-1017 and V4-Pro-0813 NVFP4 on vLLM and SGLang

  #3295·yingguo-trt opened Nov 11, 2026Contributor·

  ·
  25

• Fix Qwen3.6 hybrid ONNX calibration and INT4 export (NVBug 7035423)

  #3293·Tracy-123-dev opened Nov 11, 2026Contributor·

  ·
  2

• Load position ids on every PP stage when distilling VL models that use mROPE

  #3292·younes-hourri opened Nov 11, 2026Contributor·

  ·
  11

• fix(quantization): pad MXFP4QTensor blocks like its sibling QTensor types

  #3291·SID-6921 opened Nov 11, 2026Contributor·

  ·
  55

• [3/3] [OMNIML-6340] Add the WikiText-2 IQ GSQ example

  #3289·ajrasane opened Nov 10, 2026Contributor·

  ·
  2

• [2/3] [OMNIML-6340] Add quantize_iq_gsq and GPTQ capture

  #3288·ajrasane opened Nov 10, 2026Contributor·

  ·
  2

• [1/3] [OMNIML-6340] Add native IQ GSQ candidates

  #3287·ajrasane opened Nov 10, 2026Contributor·

  ·
  7

• [5/5] Add draft skill cards and catalog eval cases

  #3286·chadvoegele opened Nov 10, 2026Contributor·

  ·
  2