All pull requests

Pull requests

• perf(matmul_nbits): add Internal model shapes to the gfx1151 LUT

  large-pr
  #1309·ZhanxingPu opened Oct 23, 2026Contributor·

  ·
  4

• gfx1201 offline LUTs + MatMulNBits launch / autotune fixes

  large-pr
  #1308·pschlan-amd opened Oct 23, 2026Contributor·5 of 6 tasks·

  ·
  6

• fix(gqa,matmul_nbits): fix GQA prefill inf/NaN, remove GQA prefill spills, restore QMoE prefill speed

  large-pr
  #1307·ZhanxingPu opened Oct 23, 2026Contributor·4 tasks done·

  ·
  7

• fix(optimize-memrefs): Keep host-accessed buffers out of slot reuse

  #1305·ximzhang-amd opened Oct 23, 2026·7 of 8 tasks·

  ·
  3

• feat(qmoe): support swiglu_fusion=0 with relu2 activation

  #1304·mingyueliuh opened Oct 23, 2026Collaborator·9 tasks done·

  ·
  4

• fix(mha): Lower Whisper's unshared self-attention and dynamic-batch cross-attention

  #1302·ximzhang-amd opened Oct 23, 2026·9 of 10 tasks·

  ·
  3

• fix(gqa): run a first prompt shorter than its prefill window

  #1301·ximzhang-amd opened Oct 23, 2026·10 tasks done·

  ·
  5

• Add DFlash Support with HIP EP backend

  large-pr
  #1300·devledas opened Oct 22, 2026·3 of 4 tasks·

  ·
  1

  

• feat(onnx-to-hip): lower EmbedLayerNormalization

  #1299·pragyam32 opened Oct 22, 2026Contributor·2 of 3 tasks·

  ·
  3

• [do not merge] Perf/default mempool

  #1298·pschlan-amd opened Oct 22, 2026Contributor·4 tasks·

  ·
  7

• feat(rocmlir): decompose ConvTranspose into one DepthToSpace conv

  #1295·nvunnam57128 opened Oct 22, 2026Contributor·

  ·
  7

• feat(rocmlir): fuse pointwise chains into rocMLIR elementwise kernels

  #1294·nvunnam57128 opened Oct 22, 2026Contributor·

  ·
  10

• Perf/large buffer cache improved

  #1293·pschlan-amd opened Oct 22, 2026Contributor·4 tasks·

  ·
  7

• [do not merge] [experiment] Perf/large buffer cache

  #1292·pschlan-amd opened Oct 22, 2026Contributor·4 tasks·

  ·
  7

• test pr1287+1288

  large-pr
  #1291·cliu1003 opened Oct 22, 2026Contributor·

  ·
  4

• [do not merge] [experiment] Perf/hipmalloc default onnx update

  #1289·pschlan-amd opened Oct 22, 2026Contributor·4 tasks·

  ·
  5

• [DO NOT MERGE][EXPERIMENTAL ONLY] Nemotron 3.5 performance experiments

  large-pr
  #1282·fhanuman opened Oct 21, 2026Collaborator·

  ·
  3

• fix(resize): accept pytorch_half_pixel on main

  #1281·nvunnam57128 opened Oct 21, 2026Contributor·3 tasks·

  ·
  14

• perf(matmul_nbits): tuned int4 WMMA prefill: pipeline options, CU mode, per-weight row pads

  large-pr
  #1274·rnaukark opened Oct 20, 2026Contributor·8 of 12 tasks·

  ·
  1

• experiment: hipMalloc for DEFAULT memory

  #1273·pschlan-amd opened Oct 20, 2026Contributor·4 tasks·

  ·
  17

• fix(compiler): correct memory handling for dynamic shapes

  #1272·cliu1003 opened Oct 20, 2026Contributor·4 tasks done·

  ·
  18

• perf: speculative-decode verify on gfx1151 (QMoE bursts, LinearAttention burst, TopK) + TopK largest=0 fix

  enhancement
  #1264·penglu12-quark opened Oct 20, 2026Contributor·4 tasks done·

  ·
  16

  

• feat(corelib): opt-in RyzenAI CoreLib (AIE4 NPU) backend with GPU fallback

  large-pr
  #1262·bcrane-amd opened Oct 20, 2026Contributor·9 of 11 tasks·

  ·
  1

• Add initial Torch->HIP compiler plugin skeleton.

  #1261·noors-amd opened Oct 19, 2026·

  ·
  3

• Add experimental 4-bit TurboQuant KV with FP16 Ring prefill

  large-pr
  #1254·amdrajeevp1 opened Oct 18, 2026·3 of 5 tasks·

  ·
  1