Historie revizí

Autor SHA1 Zpráva Datum
  Alexander Korolev 876f54ad27 aiden-4x-labdrafter: point at the dequantized -bf16 repo (aiden registry arch DFlash2DraftModel; MXFP8 scales unsupported) před 1 měsícem
  Alexander Korolev e0eeffe3b1 lab-dflash2: use dequantized bf16 drafter variant (arch DFlashDraftModel for lab image; MXFP8 scales unsupported by its loader) před 1 měsícem
  Alexander Korolev 509654fbed ledger: glm53-flash-aiden-4x row (89/156-176, 37.6 tg c1 flat to d8192, prefill 2,185 = his ref) + TC-43 note před 1 měsícem
  Alexander Korolev 6922cb8584 ledger: glm53-flash-aiden-2x row (aiden production-1.1 + incoai DFlash2 k=7, 89/157-176) před 1 měsícem
  Alexander Korolev 0672182b34 vision 4x: parked note — same o_proj einsum assert in his exact production shape; image-level DeepGEMM layout issue on our GB10s před 1 měsícem
  Alexander Korolev 8e49f12e5a vision 2x: parked note — B12X sparse o_proj DeepGEMM fp8_einsum assert at TP2 (his verified values); 4x DCP2 is the profile lane před 1 měsícem
  Alexander Korolev 33cde9c3ef vision 2x: his verified TP2 values (GMU 0.85, KV pin off) — 0.80+pin hit a DeepGEMM layout assert před 1 měsícem
  Alexander Korolev 7bb4c8a7b1 lab-dflash2: spec swap applied (mtp-3 -> dflash + lab drafter), name updated před 1 měsícem
  Alexander Korolev 200aae23f0 new lanes: aiden 4x + lab DFlash2 drafter (incoai swap); lab image + lab DFlash2 (MTP swap) — competitive A/Bs on the lab's speculative model před 1 měsícem
  Alexander Korolev 923e6b2bfc vision recipes: args block was absorbed into [mods] table (TOML section scoping) — moved top-level před 1 měsícem
  Alexander Korolev 2b52f97ac3 aiden lanes: drop manual HF mounts (launcher auto-mounts ~/.cache/huggingface = shared NAS on all 4 nodes); GLM 4x -> NVFP4 (in shared cache, +23% vs FP8) před 1 měsícem
  Alexander Korolev 01c4ccad57 vision aiden lanes: flag route with --entrypoint sh override (baked-in env-script doubles the tail); 4x gets same fix + shared NAS HF mount před 1 měsícem
  Alexander Korolev 7497729957 vision 2x: MODEL_PATH -> cached snapshot (his script's unconditional --revision '' breaks offline resolution) před 1 měsícem
  Alexander Korolev f8ab43bcd5 vision 2x: MODEL_PATH -> cached snapshot (his script's unconditional --revision '' breaks offline resolution) před 1 měsícem
  Alexander Korolev 8c292e8a5d vision 2x: point SPARK_HF_HOME at the shared NAS HF cache (DFlash2/DSpark drafts live there; the NVMe copy lacked the DSpark draft) před 1 měsícem
  Alexander Korolev b2ec9376cc vision 2x: drop the revision pin (offline refs/main resolves; pinned hash broke the draft lookup) před 1 měsícem
  Alexander Korolev dc13bd0afe vision 2x: aiden image via HIS env-script entrypoint — mods derive NODE_RANK/MASTER_ADDR from VLLM_HOST_IP + sed off the $@ tail (env-driven contract, launcher-tail dropped) před 1 měsícem
  Alexander Korolev f2ba69ee5d vision 2x: use his /opt/vllm-src/.venv python (dist-packages wheel lacks the multimodal wrap) + his env block před 1 měsícem
  Alexander Korolev 41f2808d7a vision 2x: cudagraph cap from sizes list (pydantic consistency); MAX env kept před 1 měsícem
  Alexander Korolev a0c57b9e2f aiden lanes: entrypoint=[] for built-in-entrypoint images (launcher doubles CMD otherwise); vision 2x uses proven overlay img + aiden DSpark-greedy additions (his env-script needs a dedicated pass) před 1 měsícem
  Alexander Korolev ce4727d2a3 aiden lanes: 2x+4x recipes from aidenle.com guides (glm prod-1.1 DFlash2-7 + dsv4 vision prod-3.72 DSpark-6); 2x to verify, 4x deferred (FP8 ckpt downloading) před 1 měsícem
  Alexander Korolev fd4c74dcc1 checkpoint A/B on ds4: NVFP4 lab ckpt = same 88 quality on r17 stack but +4-7 tok/s; lab day-0 stack retains 93; r17abi recipe added před 1 měsícem
  Alexander Korolev 9079638ee4 lab recipes: instanttensor baked into lab image + enabled (verified: 631s vs 700s, two-pass loader, all boot markers + smoke 3901) před 1 měsícem
  Alexander Korolev 04984a169d ledger: one-backend bisect solved (linear b12x = the toxin); ship config locked před 1 měsícem
  Alexander Korolev 2d379f9752 final bisect: moe-b12x kept (88, d8192+2.8), linear-b12x is the toxin (44), align perf-negative — recipe locked před 1 měsícem
  Alexander Korolev 582343c9c6 r17 composition ship anchor: backend-flags are the quality toxin (bisect 46/47/74 vs 87); knobs env clean; graphs+MTP fixed on public r17 refs; ledger před 1 měsícem
  kilork 08a930e477 r12 image: adopt eugr lab build patches (12x patch_vllm_*.py) + pre-staged build sub-deps (SRC_DIR escapes) for flaky node github; ledger: lil render reference + perf push state před 1 měsícem
  kilork cc41c48e03 glm53 spark recipe: full lil policy alignment (b12x linear/moe backends, k=5 MTP + moe_backend, instanttensor load, mamba-cache-mode align, INSTANTTENSOR env) + mounted pre-baked JIT caches (~/sm121-cache/glm53-jit-121f, dsv4 pattern); boot 35m -> 7m, exact-391 verified před 1 měsícem
  kilork 5283d11ce1 clean r12 image shipped: 90/100 c1 (was 88 day-0-base), entrypoint+MAX_JOBS boot fixes, ledger před 1 měsícem
  kilork 236ce5a7f3 spark recipe: entrypoint (clean image is ENTRYPOINT []) + MAX_JOBS=2 (JIT nvcc OOM at boot); clean r12 image serving, exact-391 verified před 1 měsícem