All pull requests

Pull requests

• Gpuep/stacked all nov9

  large-pr
  #1498·SushmitaThakallapalli1980 opened Nov 9, 2026·4 tasks·

  ·
  1

• ci: slim the Linux LLVM prefix and restore mlir-opt on Windows

  #1496·zz002 opened Nov 9, 2026Collaborator·3 of 8 tasks·

  ·
  1

• feat(conv-transpose): one DepthToSpace conv per ConvTranspose, with its bias and activation in the kernel

  large-pr
  #1495·ayanshul-amd opened Nov 9, 2026Contributor·3 of 4 tasks·

  ·
  1

• feat(rocmlir): run QLinear ops with u8 weights or a weight zero point on rocMLIR

  large-pr
  #1494·ayanshul-amd opened Nov 9, 2026Contributor·3 of 4 tasks·

  ·
  3

• fix(rocmlir): fill rock.prefill kernel arguments before every launch

  #1493·ayanshul-amd opened Nov 9, 2026Contributor·3 of 4 tasks·

  ·
  3

• perf(qsa): fold glue into the sparse attention indexer and attention, pair them, faster kernels

  large-pr
  #1492·zpye opened Nov 8, 2026Collaborator·2 of 4 tasks·

  ·
  10

• fix: i1 cast/cumsum readback, vision op lowerings, and hip-test updates

  large-pr
  #1491·xiaoh-amd opened Nov 8, 2026·2 of 4 tasks·

  ·
  3

• feat(rocmlir): two-pass rocMLIR fusion

  large-pr
  #1490·ayanshul-amd opened Nov 8, 2026Contributor·

  ·
  3

• test(ut): add adaptive HIP kernel timing and native evidence

  large-pr
  #1489·ZhanxingPu opened Nov 8, 2026Contributor·

  ·
  4

• perf(gqa): fold the separate-buffer KV concat into flash decode

  #1488·BoarQing opened Nov 8, 2026Contributor·4 tasks done·

  ·
  6

• feat(gqa): run bf16 GroupQueryAttention on the decomposed path

  large-pr
  #1486·noors-amd opened Nov 7, 2026Contributor·4 tasks·

  ·
  4

• feat(kernels): add bool (uint8) support to GatherND

  #1485·noors-amd opened Nov 7, 2026Contributor·2 tasks·

  ·
  4

• feat(kernels): add bf16 support to Div

  #1484·noors-amd opened Nov 7, 2026Contributor·2 tasks·

  ·
  4

• fix(compiler): size sub-byte (i1) buffers at one byte per element

  #1483·noors-amd opened Nov 7, 2026Contributor·4 tasks done·

  ·
  3

• perf(laguna): router GEMV, cast folds, qmoe 2x2 tiles and Laguna LUT rows on gfx1151

  large-pr
  #1482·BoarQing opened Nov 7, 2026Contributor·4 tasks done·

  ·
  8

• feat(npu): CoreLib model-level, CoreLib per-op and DynamicDispatch NPU backends in one EP

  large-pr
  #1481·bcrane-amd opened Nov 7, 2026Contributor·9 of 10 tasks·

  ·
  1

• build: self-contained Windows EP (static OpenSSL) and wheel packaging for both TheRock hipBLASLt layouts

  #1480·bcrane-amd opened Nov 7, 2026Contributor·7 tasks done·

  ·
  1

• MatMul: pass real dtype to wrap_hipblasLtMatmul for bf16 support

  #1479·noors-amd opened Nov 7, 2026Contributor·2 of 4 tasks·

  ·
  3

• Add DFlash integration for HIP EP with tested Qwen3.6 models with windowed recurrent state enabled

  large-pr
  #1478·devledas opened Nov 6, 2026·5 tasks done·

  ·
  3

  

• Add windowed recurrent-state support and batched slot promotion for speculative decoding

  large-pr
  #1477·devledas opened Nov 6, 2026·10 tasks done·

  ·
  1

  

• feat(corelib): hand a whole decoder graph to a RyzenAI CoreLib model

  large-pr
  #1476·bcrane-amd opened Nov 6, 2026Contributor·8 of 10 tasks·

  ·
  1

• 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·

  ·
  12

• 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