Issues

• [DRA-89] Dense MIX/O decode projections: row-aware mt for N<8 strokes (≥190 GB/s at N=1)

  aien-dev/CUDApp#236 · Filter by author aien-spark opened on Oct 31, 2026

• DRA-84: chunked tensor-core gated delta rule scan for prefill (30 GDN layers)

  aien-dev/CUDApp#233 · Filter by author aien-spark opened on Oct 31, 2026
  1

• DRA-85: flash-style tensor-core prefill attention for q36 full-attention layers

  aien-dev/CUDApp#230 · Filter by author aien-spark opened on Oct 31, 2026

• [DRA-83] q36 prefill: chunk 128 -> 4k-8k on the tensor-core path, weights once per chunk

  aien-dev/CUDApp#226 · Filter by author aien-spark opened on Oct 31, 2026

• [DRA-71] 4-bit experts and head (NVFP4, W4A16 decode) and bf16 DeltaNet state

  aien-dev/CUDApp#223 · Filter by author aien-spark opened on Oct 31, 2026

• [DRA-76] Qualify bf16 DeltaNet state (AIEN_OMEGA_GDN_STATE_BF16=1) on the 35B

  aien-dev/CUDApp#221 · Filter by author aien-spark opened on Oct 31, 2026

• [DRA-82] Tiled mma.sync FP8 GEMM path for prefill

  aien-dev/CUDApp#220 · Filter by author aien-spark opened on Oct 31, 2026
  2

• [DRA-72] CPU/GPU feasibility: CPU bandwidth under GPU load, GPU slowdown, flag hand-off cost

  aien-dev/CUDApp#218 · Filter by author aien-spark opened on Oct 31, 2026

• [DRA-69] FP8 lm_head, one launch for all heads

  needs-GB10
  aien-dev/CUDApp#217 · Filter by author aien-spark opened on Oct 31, 2026
  1

• [DRA-68] Bring the expert down GEMM and the M=8 dense GEMMs to ~215 GB/s

  needs-GB10
  aien-dev/CUDApp#216 · Filter by author aien-spark opened on Oct 31, 2026
  1

• [DRA-67] Run the queued 35B gates for levers 3–6 (stroke8l3–l6)

  needs-GB10
  aien-dev/CUDApp#215 · Filter by author aien-spark opened on Oct 31, 2026
  1

• Track omega imports: persistent device, hazard checker, FP8 slot pool

  aien-dev/CUDApp#206 · Filter by author drakestapleton opened on Oct 30, 2026

• Pin the LDG load modifier in omega_blackwell_codegen.c

  needs-GB10
  aien-dev/CUDApp#205 · Filter by author drakestapleton opened on Oct 30, 2026

• Drift follow-up: 5 deferred omega b5ab3fe94304 mirrors recorded as divergences in #203

  aien-dev/CUDApp#204 · Filter by author drakestapleton opened on Oct 30, 2026

• CI: drift gate red on main, 49 of 763 GPU paths changed in omega b5ab3fe94304 and not mirrored

  aien-dev/CUDApp#201 · Filter by author drakestapleton opened on Oct 30, 2026

• CI: add an x86-64 host-tests job (portable suite + freestanding link)

  aien-dev/CUDApp#196 · Filter by author drakestapleton opened on Oct 30, 2026

• sm80-family cubin writer: alu.s2 schedule race (WRONG / error 700 on sm_89), root cause open

  aien-dev/CUDApp#191 · Filter by author drakestapleton opened on Oct 30, 2026
  1

• CI: run make test-runtime-api on PRs and main (ADR 0043 follow-up)

  aien-dev/CUDApp#189 · Filter by author drakestapleton opened on Oct 30, 2026

• [CUDA-MISSION] Qualify unchanged binaries on GB10, AMD and Intel and package evaluation release

  aien-dev/CUDApp#186 · Filter by author drakestapleton opened on Oct 29, 2026
  1

• [CUDA-MISSION C2] Implement bounded PTX and fatbinary translation into real backend code

  aien-dev/CUDApp#185 · Filter by author drakestapleton opened on Oct 29, 2026

• [CUDA-MISSION C0] Freeze binary corpus, compatibility profile and physical runners

  aien-dev/CUDApp#184 · Filter by author drakestapleton opened on Oct 29, 2026
  1

• [CUDA-MISSION] Hive execution plan: unchanged CUDA binaries on AMD and Intel

  aien-dev/CUDApp#183 · Filter by author drakestapleton opened on Oct 29, 2026
  1

• [MOE-INDEX-1] Native indexed/grouped MoE work scheduler for Q36 decode, batch and prefill

  aien-dev/CUDApp#159 · Filter by author drakestapleton opened on Oct 28, 2026
  2

• [KSTITCH-1] Verified IR composition of load, dequant, MMA and epilogue components

  aien-dev/CUDApp#157 · Filter by author drakestapleton opened on Oct 28, 2026
  7

• [QSCHEMA-1] Canonical quantization descriptor and target-specific native repacking

  aien-dev/CUDApp#156 · Filter by author drakestapleton opened on Oct 28, 2026
  2