All pull requests

Pull requests

• feat: add in-process rocMLIR autotune to hip-rocmlir-compiler

  #1102·gouthamb-amd opened Sep 23, 2026·

  ·
  1

• Perf/gqa flash prefill v6

  #1101·BoarQing opened Sep 23, 2026·

  ·
  5

• perf: keep the GPU pipeline going across Slice, OneHot, TopK, Pad, and CumSum

  large-pr
  #1100·pschlan-amd opened Sep 23, 2026·

  ·
  5

• perf(decode): collapse QDQ chains and the KV append into single kernels

  large-pr
  #1098·AMDmoore opened Sep 23, 2026·

  ·
  3

• feat(skill): decide operator support by compiling, not by scanning source

  large-pr
  #1097·mingyueliuh opened Sep 23, 2026·

  ·
  3

• [EXPERIMENT] perf(allocator): allocate EP tensor memory coarse-grained

  #1095·pschlan-amd opened Sep 22, 2026·

  ·
  5

• perf: integration branch for the QGemm prefill, decode and QElementwise PRs

  #1093·donghanbj opened Sep 22, 2026·

  ·
  3

• test qmoe

  #1091·BoarQing opened Sep 22, 2026·

  ·
  3

• test: Gemma-4 CI for the circulation-gated large pool (#1088 + #599)

  #1090·BoarQing opened Sep 22, 2026·

  ·
  1

• perf(allocator): pool large pinned buffers only for tensors that grow

  #1088·BoarQing opened Sep 22, 2026·

  ·
  6

• test: run #1024 with #599 so CI can exercise Gemma-4 (do not merge)

  #1082·BoarQing opened Sep 21, 2026·

  ·
  3

• [DO NOT MERGE] Feat/reduce cpu gpu syncs

  large-pr
  #1080·pschlan-amd opened Sep 21, 2026·

  ·
  7

• kernels: QMoE kernel optimization for baseline and MTP/EAGLE

  enhancement
  #1067·penglu12-quark opened Sep 20, 2026·

  ·
  7

  

• perf(qelementwise): add a contiguous fast path for unbroadcast operands

  #1063·donghanbj opened Sep 18, 2026·

  ·
  3

• perf(qgemm): optimize W4 prefill kernels

  #1051·donghanbj opened Sep 17, 2026·

  ·
  8

• perf(gqa): optimize D128 sliding-window prefill

  #1050·donghanbj opened Sep 17, 2026·

  ·
  3

• perf(qgemm): vectorize tiled staging loads and double-buffer the tile LDS

  #1047·AMDmoore opened Sep 17, 2026·

  ·
  4

• kernels: int8/dp4a MatMulNBits paths, GQA autotune key bucketing, elementwise flat path

  large-pr
  #1043·penglu12-quark opened Sep 17, 2026·

  ·
  23

  

• feat(skill): classify model compatibility by running convert-onnx-to-hip

  large-pr
  #1039·mingyueliuh opened Sep 17, 2026·

  ·
  3

• build: share LLVM 23 with rocmlirTriton (exp)

  large-pr
  #1038·amd-mounikk opened Sep 16, 2026·

  ·
  3

• [DONT MERGE] Feat/gqa window prefill tile testing

  #1036·pschlan-amd opened Sep 16, 2026·

  ·
  5

• build: move the pinned LLVM to a 23 snapshot

  #1030·amd-mounikk opened Sep 16, 2026·

  ·
  3

• ci(oga): build from AMDmoore fork with device-resident extra inputs

  #1025·AMDmoore opened Sep 16, 2026·

  ·
  3

• EAGLE3 kernel work: matmul_nbits, gqa, elementwise tuning

  enhancementlarge-pr
  #1021·penglu12-quark opened Sep 15, 2026·

  ·
  3

  

• Test: update build oga and ort to test models

  #1016·cliu1003 opened Sep 15, 2026·

  ·
  3