Repository navigation

[Cherry-pick] PRs #2817 #2671 #2892 #2876 #2910 #2911 #2919 #2904 #2937 #2928 #2885 #2908 #2929 #2950 #2972 #2877 #2951 #2816 #3005 #3012 #2971 #3015 #3006 #3026 #2988 #3028 #1860 #3030 #2612 #3033 #2653 #2599 #2655 #2618 #2868 #3032 #3027 #2869 #2984 #2975 #2345 #3020 #2997 - #3072

Merged

kevalmorabia97 merged 43 commits into

release/0.48.0from

cherry-picks/release-0.48.0

Oct 28, 2026

Merged

[Cherry-pick] PRs #2817 #2671 #2892 #2876 #2910 #2911 #2919 #2904 #2937 #2928 #2885 #2908 #2929 #2950 #2972 #2877 #2951 #2816 #3005 #3012 #2971 #3015 #3006 #3026 #2988 #3028 #1860 #3030 #2612 #3033 #2653 #2599 #2655 #2618 #2868 #3032 #3027 #2869 #2984 #2975 #2345 #3020 #2997#3072
kevalmorabia97 merged 43 commits into
release/0.48.0from
cherry-picks/release-0.48.0

Commits on Oct 27, 2026
- Fix MiniMax M2.7 expert selectors and source dtype documentation (#2817)
  
  chadvoegele
  authored and
  rohansjoshi
  committed
  3c44466
- Fix import modelopt.torch without the [hf] extra (#2671)
  
  authored and
  rohansjoshi
  committed
  6060bce
- Pin "leptonai<0.47" to fix tools/launcher CI job (#2892)
  
  kevalmorabia97
  authored and
  rohansjoshi
  committed
  1c7c7b4
- Fix layerwise MoE source-weight accounting (NVBug 6954041) (#2876)
  
  meenchen
  authored and
  rohansjoshi
  committed
  a757235
- [NVBug 6966551] Fix spelling errors and duplicated words in docs and docstrings. (#2910)
  
  authored and
  rohansjoshi
  committed
  2101318
- Leave MTP layers unquantized at any depth in the default PTQ exclusions (#2911)
  
  authored and
  rohansjoshi
  committed
  c603b63
- Fix llm_sparsity finetune saving a sharded state dict under FSDP2 (#2919)
  
  kaix-nv
  authored and
  rohansjoshi
  committed
  d7713ab
- Fix attention quantizer leakage in diffusers HF exports (#2904)
  
  Zhehao-Hu
  authored and
  rohansjoshi
  committed
  bbb07c3
- [NVBug 6972128] Fix broken links (#2937)
  
  authored and
  rohansjoshi
  committed
  f3e9435
- [NVBug 6967406] Fix misspellings in example READMEs and notebooks (#2928)
  
  noeyy-mino
  authored and
  rohansjoshi
  committed
  67aab93
- [https://nvbugs/6820271] Fix skillevaluator schema findings in monitor and evaluation (#2885)
  
  noeyy-mino
  authored and
  rohansjoshi
  committed
  740a77b
- fix(deploy): unblock TensorRT 11 inference (#2908)
  
  joe-nv
  authored and
  rohansjoshi
  committed
  52999c1
- [NVBug 6959002][fix] Supply image inputs for Wan I2V export fusion (#2929)
  
  yingguo-trt
  authored and
  rohansjoshi
  committed
  70ca109
- [OMNIML-6260] Pin the hosted Linux runners to ubuntu-24.04 (#2950)
  
  shengliangxu
  authored and
  rohansjoshi
  committed
  f1b3d41
- [NVBug 6981670][docs] Fix Cache Diffusion prepare calls (#2972)
  
  yingguo-trt
  authored and
  rohansjoshi
  committed
  674f97e
- Launch the fused NVFP4 GPTQ kernel on its tensors' GPU (#2877)
  
  authored and
  rohansjoshi
  committed
  77e22d9
- [NVBug 6934537][fix] Re-tie embeddings in init_quantized_weights (#2951)
  
  authored and
  rohansjoshi
  committed
  3e8d41d
- Fix Windows ONNX external data corruption by upgrading ONNX (#2816)
  
  anmaurya001
  authored and
  rohansjoshi
  committed
  a1ca420
- Quantize FP8 weights in float32 in the functional export (#3005)
  
  authored and
  rohansjoshi
  committed
  74565b2
- [NVBug 6982525] Honor save_modelopt_state in Diffusers export (#3012)
  
  yingguo-trt
  authored and
  rohansjoshi
  committed
  c955816
- [NVBug 6980441] Fix Diffusers subfolder ModelOpt state restoration (#2971)
  
  yingguo-trt
  authored and
  rohansjoshi
  committed
  8319ad8
- [NVBug 6982528] Save LTX full-mode checkpoints as safetensors (#3015)
  
  yingguo-trt
  authored and
  rohansjoshi
  committed
  7484d0c
- Require accelerate>=1.15.0 for llm_qat (#3006)
  
  authored and
  rohansjoshi
  committed
  44950f1
- Fall back to Python INT4 packing when the extension fails to build (#3026)
  
  authored and
  rohansjoshi
  committed
  58e6888
- Fix tensor-parallel gradient scaling in Top-K KL distillation (#2988)
  
  authored and
  rohansjoshi
  committed
  2cce2ac
- Keep calibrated skip-softmax thresholds across save/restore (#3028)
  
  kaix-nv
  authored and
  rohansjoshi
  committed
  5d8e816
- fix: prevent UnboundLocalError masking real errors in fsdp2_aware_weight_update (#1860)
  
  dinhxuanvu
  authored and
  rohansjoshi
  committed
  8d52e20
- [NVBug 6982529] Fix stale ModelOpt state when overwriting checkpoints with plain models (#3030)
  
  Zhehao-Hu
  authored and
  rohansjoshi
  committed
  9ac931b
- docs: fix TensorRT-LLM and autotune links (#2612)
  
  pratikgx
  authored and
  rohansjoshi
  committed
  23f46c3
- [NVBug 6980192] Fix NVFP4 Conv3d activation quantization disable state (#3033)
  
  Zhehao-Hu
  authored and
  rohansjoshi
  committed
  4a5af55
- Fix nested distillation forward contexts (#2653)
  
  Michael-RDev
  authored and
  rohansjoshi
  committed
  92f78b5
- Tile static NVFP4 fake quant and use int64 offsets (#2599)
  
  namgyu-yun
  authored and
  rohansjoshi
  committed
  b4c4f35
- Preserve histogram counts during entropy calibration (#2655)
  
  sylvesterkaczmarek
  authored and
  rohansjoshi
  committed
  8ef5d2f
- Fix q/k/v quantizer wiring in register_attention_for_kv_quant (#2618)
  
  authored and
  rohansjoshi
  committed
  f5e8d0b
- Emit VisualGen-compatible sparse_attention_config for diffusers skip-softmax export (#2868)
  
  authored and
  rohansjoshi
  committed
  bad3cb3
- [NVBug 6980436] Fix per-step averaging in diffusion min-mean calibration (#3032)
  
  Zhehao-Hu
  authored and
  rohansjoshi
  committed
  c342882
- [NVBug 6982527] Fix duplicate final checkpoint save in LTX distillation (#3027)
  
  Zhehao-Hu
  authored and
  rohansjoshi
  committed
  d9069e5
- LTX-2 quantization example: filter calibration extra-params, document supported LTX commit (#2869)
  
  authored and
  rohansjoshi
  committed
  23ec002
- [NVBug 6938136] Use private default locations for TensorRT build artifacts and the AutoTune timing cache (#2984)
  
  shengliangxu
  authored and
  rohansjoshi
  committed
  c38821a
- [NVBug 6901733] fix(hf_ptq): use bundled code for dense Nemotron-H on Transformers 5.5 (#2975)
  
  authored and
  rohansjoshi
  committed
  0174b5d
- Fix selected layers being skipped by post-init sparsity (#2345)
  
  authored and
  rohansjoshi
  committed
  7a6183f
- fix(speculative): pin and hash-verify the Kimi-K2 remote modeling code (#3020)
  
  authored and
  rohansjoshi
  committed
  7f48c9d
- [NVBug 6938144] Stop making /workspace world-writable in the vLLM serve example image (#2997)
  
  shengliangxu
  authored and
  rohansjoshi
  committed
  164e153

Commits

Commits on Oct 27, 2026