Commit Verlauf

Autor SHA1 Nachricht Datum
  Alexander Korolev c957beb710 eager A/B done (159 vs 158, rumor not confirmed); aiden 4x 163 = campaign record; night closed vor 1 Monat
  Alexander Korolev 6f39548ba3 aiden 4x benched: 163/176 campaign record; r18 2x eager A/B recipe vor 1 Monat
  Alexander Korolev d806bd2c2f aiden 4x: use the extended image (production-1.2 + instanttensor), restore instanttensor load vor 1 Monat
  Alexander Korolev 9e55382f40 aiden 4x: drop --load-format instanttensor (not in production-1.2 image) vor 1 Monat
  Alexander Korolev bbe82a903e aiden 4x un-deferred: production-1.2 image, zai FP8 weights now cached vor 1 Monat
  Alexander Korolev 9c648a2ee4 state doc: night-3 close — final campaign table, DFlash2 blocker, queue vor 1 Monat
  Alexander Korolev 5b47e80fd6 r27 dflash2 BLOCKED: C4 parent-page stride assert (tree-level, both r18+r27), root cause documented vor 1 Monat
  Alexander Korolev b3babad4a2 r27 DFlash2 lane: r18-dflash2 clone retargeted (method dflash, k=7, tree-picked backends) vor 1 Monat
  Alexander Korolev 191c035248 r27 4x DCP4 benched: 155 (-6 vs r18), KV 7.7M @ 16GiB; r27 still trailing r18 quality vor 1 Monat
  Alexander Korolev 1e1df7eb06 r27 4x DCP4 recipe (r18-4x-dcp4 clone retargeted at the r27 engine) vor 1 Monat
  Alexander Korolev 9cfed14f41 vision-373 lane benched: 161/176 new 2x record; launcher entrypoint-image support (model='', NODE_RANK/MASTER_ADDR env) vor 1 Monat
  Alexander Korolev 9cb56de106 state doc: night 2 results (aiden 159 first canonical, r27 156), next queue vor 1 Monat
  Alexander Korolev 2d74fdaca9 aiden 2x benched (159, first canonical aiden), r27 lane 156 (-2 vs r18), reports logged vor 1 Monat
  Alexander Korolev 8ad35cba78 night part 2: aiden 2x benched (159, first canonical), r27 lane 156 (-2 vs r18), reports saved vor 1 Monat
  Alexander Korolev e9928e3ac7 night list: prod-3.73-vision = DSV4F-Vision aiden lane (replaces blocked eugr row) vor 1 Monat
  Alexander Korolev 03edc92027 night revisit: aiden lanes + prod-3.73-vision tag queued, statuses ledgered vor 1 Monat
  Alexander Korolev d736662719 r27: explicit TORCH_CUDA_ARCH_LIST=12.1a in vllm-builder (stage-env loss -> sm120-only kernels) vor 1 Monat
  Alexander Korolev afd95f742d r27 lane: glm53-flash-r27-spark (best 2x config on the r27 engine) vor 1 Monat
  Alexander Korolev f599efe741 skill: jovian-images (clone-at-pin builds, rev adoption, GB10 deltas, pitfall ledger) vor 1 Monat
  Alexander Korolev 35d8f090b2 r27 fetch-deps: recursive gitlink population (FlashMLA/qutlass/FlashKDA cutlass vendors; -d stub trap fixed) vor 1 Monat
  Alexander Korolev 3e63d19947 state doc: image build v2 (clone-at-pin) contract vor 1 Monat
  Alexander Korolev 9118a4648b r27: WORKDIR in builder stages (stage split dropped inherited cwd) vor 1 Monat
  Alexander Korolev 97329c6b10 r27 build dir: clone-at-pin (no tarballs), gated vllm/b12x/lmcache stages, in-build deps fetch w/ pin gates vor 1 Monat
  Alexander Korolev 0a1a5dbcc1 typo vor 1 Monat
  Alexander Korolev b564bd561d ops layout: 2nodes-dac = production (operator agent), ds4 = dev lane vor 1 Monat
  Alexander Korolev ca8bb6f448 campaign extended: 2x nvfp4 KV (154), 2x Spark (158, best 2x), 4x Spark blocked (mm_mxfp8 TP4); 4x 16G pins; JIT warm-start 9min vor 1 Monat
  Alexander Korolev fe78fec205 add nvfp4kv spark variants (2x + 4x 16G, -Spark checkpoint on r18 img) vor 1 Monat
  Alexander Korolev c296f735a3 4x cells: KV pin 10 -> 16 GiB (NVFP4 4x has the headroom; 2x stays 10) vor 1 Monat
  Alexander Korolev 79c40c9482 add r18 2x nvfp4_ds_mla KV lane (parity with the 4x cell) vor 1 Monat
  Alexander Korolev cc19b5f581 persist JIT caches per node (flashinfer/torchinductor/vllm mounts in all glm53 lanes) vor 1 Monat