All pull requests

Pull requests

• Harden CODEOWNERS, action pins, CodeQL wording, and tool argv

  #264·aien-spark opened Nov 2, 2026Collaborator·11 of 12 tasks·

  ·

• Relicense first-party AGPL files to Apache-2.0 WITH LLVM-exception

  #263·aien-spark opened Nov 2, 2026Collaborator·10 of 11 tasks·

  ·

• Refuse a damaged coefficient table in the 27B exact output head

  #262·aien-spark opened Nov 2, 2026Collaborator·

  ·

• perf(gdn): submit scoped prefill through Omega-checked batches

  #261·drakestapleton opened Nov 1, 2026Contributor·

  ·

• GB10 sm_121 latency kit versus NAK

  #254·aien-spark opened Nov 1, 2026Collaborator·5 of 6 tasks·

  ·
  16

• 27B: exact GPU head and reusable command seals

  #251·aien-spark opened Nov 1, 2026Collaborator·

  ·
  11

• fix: l2_stale accessors, tern I2FP fixed latency, fp8 v2 battery path rule

  #247·aien-spark opened Nov 1, 2026Collaborator·

  ·

• DRA-84: chunked tensor-core GDN prefill scan (gdnc) behind AIEN_OMEGA_Q36_PF_GDN_CHUNK=1

  #244·aien-spark opened Oct 31, 2026Collaborator·

  ·

• DRA-84: pfq bf16 in/out for GDN projections (env/field-gated, default bits unchanged)

  #242·aien-spark opened Oct 31, 2026Collaborator·

  ·

• DRA-86: q36 prefix snapshot module (KV + DeltaNet state) + gdn state view/dev copy

  #240·aien-spark opened Oct 31, 2026Collaborator·

  ·

• DRA-84 (pending): IR cap raise 2048 insns / 512 vregs for the chunked GDN scan

  #238·aien-spark opened Oct 31, 2026Collaborator·

  ·
  1

• DRA-89: row-aware dense stroke shape (AIEN_OMEGA_FP8GEMM_ROWAWARE=1, default off) [chip NOT_RUN]

  #237·aien-spark opened Oct 31, 2026Collaborator·

  ·

• DRA-78/79: fp8 GEMM CTA budget knob + KS=1 next-row prefetch (env-gated, default off)

  #235·aien-spark opened Oct 31, 2026Collaborator·

  ·

• DRA-80/81: fp8 GEMM x_resident form + shared-expert one pass (env-gated, default off)

  #234·aien-spark opened Oct 31, 2026Collaborator·

  ·

• DRA-85 [draft]: flash-style tensor-core prefill attention (fattn256) + q36 hook AIEN_OMEGA_Q36_PF_FATTN (default off)

  #232·aien-spark opened Oct 31, 2026Collaborator·

  ·

• DRA-69: FP8 lm_head, one launch for all heads (flag AIEN_OMEGA_Q36_FP8_HEAD, default off)

  #231·aien-spark opened Oct 31, 2026Collaborator·

  ·

• DRA-67 levers 4-5 (6-7 pending): q36 stroke GDN/attn heads, env-gated off (#215)

  #229·aien-spark opened Oct 31, 2026Collaborator·

  ·

• [DRA-82] Tiled mma.sync FP8 GEMM path for prefill (q36, AIEN_OMEGA_Q36_PF_QMMA, default off)

  #228·aien-spark opened Oct 31, 2026Collaborator·

  ·

• [DRA-83] q36 prefill chunk up to 8192 (TC path), per-chunk streamed count, mutant 7 (stacked on DRA-82)

  #227·aien-spark opened Oct 31, 2026Collaborator·

  ·

• DRA-67: gate fp8gemm lever 3 (x_ahead) default off; levers 4-7 pending (#215)

  #225·aien-spark opened Oct 31, 2026Collaborator·

  ·

• [DRA-71] NVFP4 W4A16 grouped GEMV/GEMM generator (host + warp sim)

  #224·aien-spark opened Oct 31, 2026Collaborator·

  ·

• [DRA-76][HOLD] gdn: bf16 state in the multi-token launch + AIEN_OMEGA_GDN_STATE_MUTANT controls

  #222·aien-spark opened Oct 31, 2026Collaborator·

  ·

• [DRA-72] CPU/GPU split feasibility benches (host-ready, chip NOT_RUN)

  #219·aien-spark opened Oct 31, 2026Collaborator·

  ·
  1

• fix(gpu): matmul launch declares peak + 2 registers

  #209·aien-spark opened Oct 31, 2026Collaborator·

  ·

• fix(cubin): tkinfo note + declared registers + covering export grid, so sm_121 images load and run under libcuda and MAX (#68, MAX Gates 2-3)

  #202·aien-spark opened Oct 30, 2026Collaborator·

  ·
  1