 Alexander Korolev
|
02b94db96f
redhat lane: add MTP k=5 — model_mtp.safetensors (7.6 GB) now in the main NVFP4 repo; card 'Enable Speculative Decoding' section (2026-09-05)
|
hace 1 mes |
|
Alexander Korolev
|
15c930fa14
r18: bump jovian-judgement pins to branch tip (RoCEnante #597 + NVFP4 KV + DFlash split-KV + SM121 A16 precision); fix stale r12 pins; Dockerfile header -> versions.env is source of truth
|
hace 1 mes |
|
Alexander Korolev
|
1a6df4d3d8
AGENTS.md: use rg, not grep — recursive grep aborts/hangs in this tree
|
hace 1 mes |
|
Alexander Korolev
|
5b24806bc4
bump remaining GLM53 aiden recipes to production-1.2 (2x + full-nvfp4-spark; 1.2 = 1.1 + fused mHC prefill kernel)
|
hace 1 mes |
|
Alexander Korolev
|
4ac6815579
glm53-flash-aiden image: bump base to production-1.2 (fused mHC prefill kernel, 1.37x mHC section @4K tokens per aiden; VLLM_MHC_FUSED_LARGE=0 to revert)
|
hace 1 mes |
|
Alexander Korolev
|
bd72e30934
AGENTS.md: tool-eval-bench runs in a separate dir, never in this repo
|
hace 1 mes |
|
Alexander Korolev
|
d31e3b0b0c
glm53 aiden: local image (prod-1.1 + instanttensor), MTP spec, 512K ctx
|
hace 1 mes |
|
Alexander Korolev
|
cda2dfa272
full GLM-5.3 4x lanes (aiden + lab images): PARKED with findings
|
hace 1 mes |
|
Alexander Korolev
|
7cbebdee5a
aiden-4x-labdrafter: parked note — warmup fp8 fused-KV dtype mismatch with the dequantized lab drafter; belongs to the lab fork
|
hace 1 mes |
|
Alexander Korolev
|
9225c3be97
lab-dflash2: parked note — lab image GLM5Next lacks SupportsEagle3 (dflash requires aux hidden states); the A/B moves to the aiden image
|
hace 1 mes |
|
Alexander Korolev
|
2ce25b0081
lab-dflash2: refs-main 40-hex note; add 4x aiden bench report (89/156-176)
|
hace 1 mes |
|
Alexander Korolev
|
876f54ad27
aiden-4x-labdrafter: point at the dequantized -bf16 repo (aiden registry arch DFlash2DraftModel; MXFP8 scales unsupported)
|
hace 1 mes |
|
Alexander Korolev
|
e0eeffe3b1
lab-dflash2: use dequantized bf16 drafter variant (arch DFlashDraftModel for lab image; MXFP8 scales unsupported by its loader)
|
hace 1 mes |
|
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
|
hace 1 mes |
|
Alexander Korolev
|
6922cb8584
ledger: glm53-flash-aiden-2x row (aiden production-1.1 + incoai DFlash2 k=7, 89/157-176)
|
hace 1 mes |
|
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
|
hace 1 mes |
|
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
|
hace 1 mes |
|
Alexander Korolev
|
33cde9c3ef
vision 2x: his verified TP2 values (GMU 0.85, KV pin off) — 0.80+pin hit a DeepGEMM layout assert
|
hace 1 mes |
|
Alexander Korolev
|
7bb4c8a7b1
lab-dflash2: spec swap applied (mtp-3 -> dflash + lab drafter), name updated
|
hace 1 mes |
|
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
|
hace 1 mes |
|
Alexander Korolev
|
923e6b2bfc
vision recipes: args block was absorbed into [mods] table (TOML section scoping) — moved top-level
|
hace 1 mes |
|
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)
|
hace 1 mes |
|
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
|
hace 1 mes |
|
Alexander Korolev
|
7497729957
vision 2x: MODEL_PATH -> cached snapshot (his script's unconditional --revision '' breaks offline resolution)
|
hace 1 mes |
|
Alexander Korolev
|
f8ab43bcd5
vision 2x: MODEL_PATH -> cached snapshot (his script's unconditional --revision '' breaks offline resolution)
|
hace 1 mes |
|
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)
|
hace 1 mes |
|
Alexander Korolev
|
b2ec9376cc
vision 2x: drop the revision pin (offline refs/main resolves; pinned hash broke the draft lookup)
|
hace 1 mes |
|
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)
|
hace 1 mes |
|
Alexander Korolev
|
f2ba69ee5d
vision 2x: use his /opt/vllm-src/.venv python (dist-packages wheel lacks the multimodal wrap) + his env block
|
hace 1 mes |
|
Alexander Korolev
|
41f2808d7a
vision 2x: cudagraph cap from sizes list (pydantic consistency); MAX env kept
|
hace 1 mes |
