All pull requests

Pull requests

• Lower runtime shape expressions directly to GPU fill kernels

  #5449·shivadbhavsar opened Oct 27, 2026Contributor·

  ·
  2

• Check runtime dependent ops for symbol simplification

  #5448·shivadbhavsar opened Oct 27, 2026Contributor·

  ·

• Parse LinearAttention and fuse causal gemm-elementwise-gemm chains with rocMLIR

  #5447·pfultz2 opened Oct 27, 2026Collaborator·

  ·
  12

• prepare_mlir: push a reshape of an int4 unpack under the unpack

  #5446·pfultz2 opened Oct 27, 2026Collaborator·

  ·
  3

• Block output handling for data-dependent operators in split_sym_dim

  #5445·CharlieL7 opened Oct 26, 2026Collaborator·

  ·
  62

• Scatter elements analyzer for split_sym_dim

  #5444·CharlieL7 opened Oct 26, 2026Collaborator·

  ·
  4

• Handle empty tensors in select module by short-circuiting

  #5443·CharlieL7 opened Oct 26, 2026Collaborator·

  ·
  4

• Dynamic concat kernel for handling concating over symbolic dimensions

  #5442·CharlieL7 opened Oct 26, 2026Collaborator·

  ·
  6

• KV cache attention flash decoding

  #5439·bdevorem opened Oct 26, 2026Member·

  ·
  1

• Data dependent handling in split_sym_dim

  #5438·CharlieL7 opened Oct 26, 2026Collaborator·

  ·
  35

• simplify_reshapes: require an unambiguous entry chain in find_entry_shape_transform_op

  #5435·VishalX opened Oct 26, 2026Contributor·2/2·

  ·
  11

• gpu: make gpu::slice_at skip out-of-range writes instead of throwing

  #5434·VishalX opened Oct 26, 2026Contributor·1/2·

  ·
  16

• Bump gitpython from 3.1.62 to 3.2.1 in /docs/sphinx

  documentation
  #5433·dependabot[bot] opened Oct 26, 2026Bot·

  ·

• Bump certifi from 2024.7.4 to 2026.10.19 in /docs/sphinx

  documentation
  #5432·dependabot[bot] opened Oct 24, 2026Bot·

  ·

• Improve select_module perf

  #5431·shivadbhavsar opened Oct 23, 2026Contributor·

  ·
  3

  

• Forward slices spanning several concat segments through reshaped concats

  #5430·pfultz2 opened Oct 23, 2026Collaborator·

  ·
  4

• Onnxruntime Weekly Sync 2026-10-23

  automateddependenciesonnxruntimeskip bot checks
  #5428·github-actions[bot] opened Oct 23, 2026Bot·

  ·
  

• Resolve channels_auto convolution layout to channels_first on gfx1100 and gfx1151

  Windows
  #5426·ppetrovi-amd opened Oct 23, 2026·1 of 6 tasks·

  ·
  7

• Prevent FP16 GroupNorm overflow with float intermediates

  #5425·royzhang-amd opened Oct 23, 2026·

  ·
  2

• Add ONNX parser for the CausalConvWithState contrib op

  #5423·pfultz2 opened Oct 22, 2026Collaborator·

  ·
  6

• Sink the view chain onto pointwise inputs so flat-layout kv-cache appends fuse

  #5419·pfultz2 opened Oct 21, 2026Collaborator·2/2·

  ·
  3

• Fuse the attention-sink gate into the attention epilogue and the o_proj reduce

  #5417·pfultz2 opened Oct 21, 2026Collaborator·2/2·

  ·
  5

• Share loaded modules between identical code objects

  #5414·pfultz2 opened Oct 20, 2026Collaborator·

  ·
  6

• Preserve arg-reduction axes during reshape fusion

  #5413·royzhang-amd opened Oct 20, 2026·

  ·
  14

• Apply MatMulInteger zero-point corrections in int32

  #5412·royzhang-amd opened Oct 20, 2026·

  ·
  3