All pull requests

Pull requests

• AIESW-45531: Disable ASLR during wheel build for deterministic compile…

  #1339·serged-amd opened Sep 30, 2026·3 of 5 tasks·

  ·

• enable dsv41f for gfx1250

  #1336·danichan-mkm opened Sep 30, 2026·

  ·

• [ROCm][gfx1151] RDNA35_HIP_ATTN: HIP decode attention built into _rocm_C, default on gfx1151

  #1330·roberteg16 opened Sep 28, 2026·5 of 6 tasks·

  ·

• Add A4W16 GPT OSS MoE routing for gfx1250

  #1325·JadenMathias opened Sep 23, 2026·

  ·

• [ROCm] Encoder prefill attention on gfx115x: split-D head dims and a narrower KV tile

  #1322·roberteg16 opened Sep 23, 2026·

  ·

• [ROCm][W4A16] Add an opt-in hipBLASLt path for the gfx11 W4A16 GEMM

  #1321·mgehre-amd opened Sep 23, 2026·

  ·

• [Triont/Gluon] [GFX12] DSV4-Pro TP optimization

  #1318·k50112113 opened Sep 21, 2026·

  ·

• initial enablement and gates for dsv4.1f

  #1317·danichan-mkm opened Sep 21, 2026·

  ·

• [ROCm][gfx1151] Default the KV cache layout to HND for decode

  #1315·roberteg16 opened Sep 21, 2026·6 of 8 tasks·

  ·

• [Benchmarks] attention: keep the KV working set out of cache

  #1311·roberteg16 opened Sep 18, 2026·4 tasks done·

  ·
  1

• Remove MOE_BOUND env being set

  #1309·JadenMathias opened Sep 17, 2026·

  ·

• Remove stale upstream-triton override in Dockerfile.rocm_gfx1250 (clobbers ROCm-built wheel)

  #1307·kiran-thumma opened Sep 16, 2026Collaborator·2 tasks·

  ·

• Remove stale upstream-triton override in Dockerfile.rocm_gfx1250 (clobbers ROCm-built wheel)

  #1306·kiran-thumma opened Sep 16, 2026Collaborator·2 tasks·

  ·

• routing through triton attn and padding gemm_afp4wfp4

  #1285·danichan-mkm opened Sep 9, 2026·

  ·
  1

• fixes needed to enable kimi2.6 on gfx1250

  #1282·ahmed-bsod opened Sep 8, 2026·5 tasks·

  ·
  1

• fix(amd-ci): DPX CI fixes — test isolation, memory wait, buildkite config

  #1248·mkunredd opened Sep 3, 2026·2 tasks·

  ·
  1

• [ROCm][gfx1151] Software-pipeline wvSplitK and collapse its gfx11 dispatch

  #1246·mgehre-amd opened Sep 3, 2026·

  ·

• ci: add gfx11 vLLM image build and GHCR publish workflow

  #1242·zohaibm-amd opened Aug 27, 2026·

  ·
  10

• Update Dockerfile.rocm_base

  #1224·jvadakat opened Aug 19, 2026·5 tasks·

  ·

• [ROCm][Attention] Wide MHA decode attention kernel for gfx1151

  #1087·roberteg16 opened Aug 12, 2026·5 of 6 tasks·

  ·

• [ROCm] Stop ROCPROFILER_REGISTER_LIBRARY leaking into engine subprocesses

  #1077·roberteg16 opened Aug 3, 2026·5 tasks done·

  ·
  4

• [bench] vit_attention: model-matching V stride + controllable cache-eviction (--flush-mib)

  #1070·parthash0804 opened Jul 30, 2026·

  ·
  1

• [ROCm][Kernel] W4A16 MoE prefill: M-aware moe_align block size (1.6x TTFT)

  #1069·stoivonen-amd opened Jul 29, 2026·3 of 5 tasks·

  ·
  1

• Release/0.23.0

  #1058·kiran-thumma opened Jul 14, 2026Collaborator·5 tasks·

  ·

• [ROCm][Kernel] W4A16 skinny GEMM: hand-asm kernel for gfx1151

  #1045·mgehre-amd opened Jul 1, 2026·

  ·