All pull requests

Pull requests

• perf(matmul_nbits): add bounded K8-XOR routing and refresh gfx1151 LUT

  large-pr
  #1475·ZhanxingPu opened Nov 6, 2026Contributor·

  ·
  6

• perf(sai): fold QSA rotary tables and counts into the indexer, faster kernels

  large-pr
  #1474·zpye opened Nov 6, 2026Collaborator·2 of 4 tasks·

  ·
  4

• ci(windows): add Strix GPU test with public HF models

  #1473·zz002 opened Nov 6, 2026Collaborator·3 of 7 tasks·

  ·
  3

• fix(onnx-to-hip): preserve ONNX dtype enum through Cast lowering

  #1472·kirankumarmatam-amd opened Nov 6, 2026Collaborator·3 of 4 tasks·

  ·
  3

• fix(cmake): guard nested find_package(onnxruntime|GTest) behind TARGET checks

  #1471·kirankumarmatam-amd opened Nov 6, 2026Collaborator·3 of 4 tasks·

  ·
  3

• refactor(hip): align dialect and C API source layout with MLIR

  large-pr
  #1470·fhanuman opened Nov 6, 2026Collaborator·

  ·
  5

• DO NOT MERGE: test #1468's cold-cache path

  #1469·xingy16-amd opened Nov 6, 2026Collaborator·1 of 4 tasks·

  ·
  3

• ci: write the LLVM caches only from push runs

  #1468·xingy16-amd opened Nov 6, 2026Collaborator·5 of 7 tasks·

  ·
  3

• Feat/elementwise bf16

  #1467·noors-amd opened Nov 5, 2026Contributor·

  ·
  4

• perf(nemotron): fix QMoE top-k > 16, speed up gated LinearAttention prefill, fold Mamba2 q/k broadcast

  #1466·zz002 opened Nov 5, 2026Collaborator·7 of 8 tasks·

  ·
  7

• ci(windows): install only the LLVM components hip-ep consumes

  #1465·zz002 opened Nov 5, 2026Collaborator·7 of 8 tasks·

  ·
  3

• docs(readme): fix hip-mlir-opt startup crash on Windows and stale build/test docs

  #1464·cliu1003 opened Nov 5, 2026Contributor·4 tasks done·

  ·
  3

• test(e2e): name native artifacts with the platform's library suffix

  #1463·xingy16-amd opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  3

• feat(compiler): HIP IR as an output and an input (hip-compiler --mode HIP_IR)

  #1462·xingy16-amd opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  3

• ci(oga): bump OGA to v0.19.0 and drop merged PR patches

  #1461·zz002 opened Nov 5, 2026Collaborator·2 of 7 tasks·

  ·
  5

• perf: fuse decomposed attention and speed up fp16 convolutions for SD1.5 (~1.6x)

  large-pr
  #1458·donghanbj opened Nov 5, 2026Contributor·

  ·
  4

• perf: add offline autotune LUTs for matmul_nbits and GQA

  large-pr
  #1456·ZhanxingPu opened Nov 5, 2026Contributor·

  ·
  3

• feat(resize): ONNX Resize's antialias and cubic as hip.resize attributes, covering GGML's interpolate

  #1455·xingy16-amd opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  4

• feat(swiglu): a clamped form with alpha, beta and a limit

  #1454·xingy16-amd opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  4

• fix(mha): broadcast self-attention's seqlens_k on the device

  #1453·xingy16-amd opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  3

• fix(scatter_nd): do not copy data onto itself when the output is data

  #1452·xingy16-amd opened Nov 5, 2026Collaborator·3 of 4 tasks·

  ·
  4

• fix(output-allocator): report a static output's returned shape through any reshape chain

  #1451·xingy16-amd opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  3

• fix(hip): name the optional operands of causal_conv_with_state and rope

  #1450·xingy16-amd opened Nov 5, 2026Collaborator·4 tasks done·

  ·
  3

• feat(rocmlir): name fused kernels mlir_<ops>_<n>

  #1449·nvunnam57128 opened Nov 4, 2026Contributor·3 of 4 tasks·

  ·
  1

• perf(qmoe): fold sigmoid-bias routing, speed up prefill, keep SwiGLU beyond fp16

  large-pr
  #1448·BoarQing opened Nov 4, 2026Contributor·4 tasks done·

  ·
  4