All pull requests

Pull requests

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

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

  ·

• Admit GB10 SM_VERSION 10_04 (0xa04) with the Blackwell-B class set

  #246·aien-spark opened Nov 1, 2026Contributor·

  ·
  1

• Move omega-bridge pin to 916d9f6d (seeded loop guards match omc 1,000,000-pass trap)

  #245·aien-spark opened Oct 31, 2026Contributor·

  ·

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

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

  ·

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

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

  ·

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

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

  ·

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

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

  ·
  1

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

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

  ·

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

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

  ·

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

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

  ·

• 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, 2026Contributor·

  ·

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

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

  ·

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

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

  ·

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

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

  ·

• [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, 2026Contributor·

  ·

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

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

  ·

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

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

  ·

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

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

  ·

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

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

  ·
  1

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

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

  ·

• 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, 2026Contributor·

  ·
  1

• mathlibs: GB10 xoshiro256** RNG and batched complex FFT from a seeded .om unit (audit rows 29, 30)

  #147·aien-spark opened Oct 27, 2026Contributor·

  ·
  2

• aien-gpu: second platform implementation over /dev/aien-gpu (hardware NOT_RUN)

  #97·aien-spark opened Oct 26, 2026Contributor·

  ·
  3