All pull requests

Pull requests

• [CPU] Fix non-saturating float8 overflow

  #33479·Nathan Chandrasekar (vera-rubin) opened Oct 20, 2026·

  ·
  2

• [CPU] Pow: compute integer powers exactly instead of through std::pow

  #33478·Sathish Kumar Prabaharan (sapr5159) opened Oct 19, 2026·

  ·
  1

  3

• Update stale OrtEnv::m_ references in ort_env.cc comments

  #33477·Mustapha Jaber (mustjab) opened Oct 19, 2026Contributor·

  ·
  1

  

• Allow IR v13 tensor types in custom ops

  #33476·James Ward (jamwar01) opened Oct 19, 2026·

  ·
  3

• Validate runtime tensor shapes and optional inputs in CPU ReverseSequence, DFT, STFT, GatherND, QLinearLookupBase, and Inverse kernels

  #33475·m00dy (professor-moody) opened Oct 19, 2026·

  ·
  9

• Add fused MTP MoE and speculative state fold CUDA operators

  #33474·Akshay Sonawane (apsonawane) opened Oct 19, 2026Contributor·

  ·
  19

• Optimize quantized MoE and block-scaled FP8 CUDA projections

  #33473·Akshay Sonawane (apsonawane) opened Oct 19, 2026Contributor·

  ·
  20

• Extend sparse attention caches and optimize packed indexing

  #33472·Akshay Sonawane (apsonawane) opened Oct 19, 2026Contributor·

  ·
  19

• Extend GDN and causal convolution fused state handling

  #33471·Akshay Sonawane (apsonawane) opened Oct 19, 2026Contributor·

  ·
  8

• Add fused hyper-connection and tensor CUDA operators

  #33470·Akshay Sonawane (apsonawane) opened Oct 19, 2026Contributor·

  ·
  33

• Add exact normalization and rotary CUDA operator modes

  #33469·Akshay Sonawane (apsonawane) opened Oct 19, 2026Contributor·

  ·
  10

• [Draft] Add internal MatMulNBits core operation-count estimator

  #33467·Tommaso Adani (tadani3) opened Oct 19, 2026·

  ·
  

• CPU: fuse transposed INT4 Gemm while preserving activation and output QDQ

  #33458·Eric Sommerlade (es0m) opened Oct 19, 2026·

  ·
  2

• Kleidiops fp32 mlas fallback

  #33457·Damien Dooley (damdoo001-arm) opened Oct 19, 2026Contributor·

  ·
  3

• Fix Debug schema counts with external ONNX static registration

  #33456·miyan (miyanyan) opened Oct 19, 2026Contributor·

  ·
  2

• [CUDA] Add float16 support to GridSample

  #33454·Mohamed BEN HAMDOUNE (mbenhamd) opened Oct 19, 2026·

  ·
  1

  3

• [CUDA] Add QMoEOffload operator with doorbell CUDA-graph decode, cross-layer promotion, and DMA prefill

  #33452·Tianlei Wu (tianleiwu) opened Oct 19, 2026Contributor·3 tasks done·

  ·

• [CUDA] Add session-scoped QMoE offload lifecycle, device memory ledger, and arena shrink hooks

  #33451·Tianlei Wu (tianleiwu) opened Oct 19, 2026Contributor·3 tasks done·

  ·

• [CPU/MLAS] Add QMoE shared 4-bit expert layouts, AVX-512 kernels, and FileExpertStore

  #33450·Tianlei Wu (tianleiwu) opened Oct 19, 2026Contributor·3 tasks done·

  ·
  6

• [WebGPU] Fix FP16 MatMulNBits wide-tile results on Qualcomm D3D12

  #33449·Yang Gu (gyagp) opened Oct 19, 2026Contributor·

  ·
  1

• [WebGPU] Generate component inventories from fetched Dawn dependencies

  #33448·Yang Gu (gyagp) opened Oct 19, 2026Contributor·

  ·
  1

• Add EP device memory capacity query API with CUDA support

  #33447·Selena Yang (selenayang888) opened Oct 19, 2026Contributor·

  ·

• Speed up CPU 1D depthwise ConvTranspose

  #33446· Stern (Hosi121) opened Oct 18, 2026Contributor·

  ·
  7

• Fix: [Bug][CPU] TransposeOptimizer inserts an unsupported INT32 QuantizeLinear after valid dequantization

  #33445·SIDDARTHA REDDY (SIDDARTHAREDDY8) opened Oct 18, 2026Contributor·

  ·
  1

  6

• fix(optimizer): preserve DequantizeLinear before ArgMax when scale is non-positive (#33440)

  #33444·Prashanth (dpr2106) opened Oct 18, 2026·

  ·
  1

  4