All pull requests

Pull requests

• ci(windows-rmt): sync OGA_PR_PATCHES with windows-deps

  #1401·BoarQing opened Nov 1, 2026Contributor·4 tasks done·

  ·
  3

• ci: align the Windows RMT OGA cache key with windows-deps

  #1400·amd-mounikk opened Oct 31, 2026Contributor·2 tasks·

  ·
  4

• feat(perf) Add opt-in KV-Pareto KV-cache compression for Strix Halo (gfx1151)

  large-pr
  #1398·amd-rohrajes opened Oct 31, 2026·

  ·
  4

• perf(laguna): fused key-padding attention, sigmoid-bias qmoe routing, faster qmoe prefill

  large-pr
  #1397·BoarQing opened Oct 31, 2026Contributor·3 of 4 tasks·

  ·
  10

• feat(kernels): add bf16 support to hip_cast

  large-pr
  #1396·noors-amd opened Oct 30, 2026·2 tasks done·

  ·
  5

• fix(memory): prevent invalid host access after buffer reuse

  #1395·fhanuman opened Oct 30, 2026Collaborator·

  ·
  9

• Single-pass softmax for long rows, with typed dispatch and a last-axis guard

  #1394·ayanshul-amd opened Oct 30, 2026Contributor·

  ·
  4

• fix(onnx-to-hip): resolve integer Div when sizing dynamic Slice extents

  #1393·pragyam32 opened Oct 30, 2026Contributor·6 tasks done·

  ·
  3

• Fix/smolvlm vision

  large-pr
  #1392·amd-rmotiram opened Oct 30, 2026·

  ·
  1

• perf(kernels): row-based, vectorized Resize kernels

  #1390·ayanshul-amd opened Oct 30, 2026Contributor·3 of 4 tasks·

  ·
  4

• feat(Scan): add forward axis-0 ONNX Scan support

  large-pr
  #1386·cliu1003 opened Oct 30, 2026Contributor·3 of 4 tasks·

  ·
  5

• perf(resize): faster rank-2 Resize kernels

  #1385·tuhinp-amd opened Oct 30, 2026Contributor·4 tasks done·

  ·
  8

• feat(hip-cast): fold a cast round trip through ops that only move elements

  #1384·ayanshul-amd opened Oct 30, 2026Contributor·4 tasks done·

  ·
  3

• feat(rocmlir): split static multi-head attention into two matmuls

  #1382·nvunnam57128 opened Oct 30, 2026Contributor·5 of 7 tasks·

  ·
  5

• perf(matmul): run bf16 -> fp32 transposed-B projections on a GEMV (decode) and a WMMA GEMM (prefill) on gfx1151

  large-pr
  #1381·masteryi-0018 opened Oct 30, 2026·4 tasks done·

  ·
  4

• feat(matmul): run bf16 matmuls as bf16, with a bf16 or fp32 output

  #1380·masteryi-0018 opened Oct 30, 2026·4 tasks done·

  ·
  3

• feat(fuse-rocmlir): run a conv of dequantized operands as an i8 rocMLIR conv

  #1379·ayanshul-amd opened Oct 29, 2026Contributor·4 tasks done·

  ·
  3

• feat(fuse-rocmlir): run QLinearConv and QLinearMatMul as i8 rocMLIR kernels with a fused quantized epilogue

  large-pr
  #1378·ayanshul-amd opened Oct 29, 2026Contributor·4 tasks done·

  ·
  3

• perf(norm): rewrite the GroupNorm, InstanceNorm and LayerNorm statistics, and lower GroupNormalization

  large-pr
  #1377·ayanshul-amd opened Oct 29, 2026Contributor·4 tasks done·

  ·
  4

• Stack open fixes for If shapes, rocMLIR, and integer Resize

  large-pr
  #1376·amd-mounikk opened Oct 29, 2026Contributor·1 of 3 tasks·

  ·
  13

• fix(matmul): pick the same hipBLASLt kernel in every process, and a fast one

  #1374·masteryi-0018 opened Oct 29, 2026·4 tasks done·

  ·
  5

• feat(resize): support linear Resize on int8 and uint8

  #1373·tuhinp-amd opened Oct 29, 2026Contributor·1 of 2 tasks·

  ·
  6

• test(plugin): let StaticPlugins pass with out-of-tree production plugins linked

  #1371·masteryi-0018 opened Oct 29, 2026·4 tasks done·

  ·
  3

• fix(cast): convert fp32 <-> bf16 in hip_cast

  large-pr
  #1369·masteryi-0018 opened Oct 29, 2026·4 tasks done·

  ·
  4

• fix(hip): keep growing pointwise fusion inside the rocMLIR anchor shape

  #1368·pragyam32 opened Oct 28, 2026Contributor·2 of 3 tasks·

  ·
  3