 Alexander Korolev
|
bf593fd4fd
test: add extra_mounts_manifest: None to Recipe test literals (fix test-profile build)
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
08f53fc875
exl3: per-file bind mounts via extra_mounts_manifest — the runtime opens shards with O_NOFOLLOW, so a symlink final component => ELOOP (deterministic, nothing to do with NFS); manifest generated per node by before-serve hook, launcher reads it over ssh after the hooks. EXL3 2x now boots and serves
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
a1472225ad
refactor: move prebuilt-image check to sync::ensure_prebuilt_image (mod.rs back under 500)
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
89c21cc236
fix: stop --cluster narrows multi-deployment (no more 'use --node' when --cluster given); --node without --cluster resolves via node containment. prebuilt images (no docker_image_dir) are presence-checked + node-to-node synced before docker run — no more silent Docker Hub pull + false 'started'
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
42e78480b8
qwen: VLLM_MARLIN_USE_ATOMIC_ADD off (live garble caught co-batched with 70K prefill); day-shift notes
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
7250e2ad11
qwen: VLLM_MARLIN_USE_ATOMIC_ADD off (experimental atomic-add path, suspected intermittent garble)
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
faca4a98da
exl3: two-mount staging (repo at /hub + relative-symlink dir at /model), kv_cap 0; qwen served name qwen3.8
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
ad7069edcd
docker: --entrypoint takes tail[0], rest is argv (no repeat); port convention 30000 in recipes + docs
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
3bcc7760a0
docker: --entrypoint override for entrypoint-driven images; trim cluster to launch set before hooks/sync; recipe.cluster on auto
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
7f2f15d656
09-26 shift: second half (aiden/qwen fixes, exl3 2x recipe + launcher extensions)
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
0edc52495e
display unlock scripts: sudo reads for /sys module parameters
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
3270c2b967
serve: honor recipe.cluster when --cluster is auto (docker path parity with tmux)
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
07919e5fc1
exl3 2x: recipe (rc2 image, /model mount, 9g/6cpu cguard, display unlock) + launcher memory_limit/cpus flags
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
2dcf9a3e9e
exl3: temporary 2 GB display-reserve unlock (unlock/restore builtins) + per-recipe device passthrough
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
45462379ae
bench: re-attach the perf table the --json-file path drops; fix aiden recipe toml
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
548fcdfc75
dsv41 4x sglang aiden-1.6 (162/176) + qwen38 int4 1x (165/176) lanes, night shift 09-26
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
5bd465adb0
MiMo-V2.6-Flash-RL 2x on SGLang: recipe + first run 161/176 (91/100), needle 132K, code 71-79 t/s
|
3 nedēļas atpakaļ |
|
Alexander Korolev
|
59c99002ea
mimo26 recipe: context 512K (pool 2.19M tokens, model natively 1M) — takes effect at next boot
|
4 nedēļas atpakaļ |
|
Alexander Korolev
|
2aa9023c2f
bench: MiMo-V2.6-Pro-RL 8x first run — 165/176 (94/100), perf pp ~1.7K t/s
|
4 nedēļas atpakaļ |
|
Alexander Korolev
|
759910ef68
history: 09-25 shift — MiMo-V2.6-Pro-RL 8x bring-up (build, sync, serve from NAS, smoke pass)
|
4 nedēļas atpakaļ |
|
Alexander Korolev
|
d5663ec4d9
feat: MiMo-V2.6-Pro-RL 8x Spark recipe + build tree (SGLang a14, DFlash, RoCEnante); dry-run path-ref placeholder
|
4 nedēļas atpakaļ |
|
Alexander Korolev
|
51b7404aa8
feat: serve --hf-cache <path> overrides the host-side HF cache dir per serve (config hf_cache_dir stays the global knob; mutual exclusion with --mirror)
|
4 nedēļas atpakaļ |
|
Alexander Korolev
|
3f52b0072a
config: merge [mikrotik] from cluster_sources (with the HA endpoint it merges from) — a RouterOS-named extra_entity 404s as raw HA when the section is dropped
|
1 mēnesi atpakaļ |
|
Alexander Korolev
|
36872f238c
web: recipes iframe loads /recipes with the token (the page document itself is token-gated)
|
1 mēnesi atpakaļ |
|
Alexander Korolev
|
d1e9c05c35
web: dashboard token in the browser (?token= seed -> localStorage, auth() on all data URLs, 401 prompt; recipes iframe shares it)
|
1 mēnesi atpakaļ |
|
Alexander Korolev
|
c2e6c33d79
4x r39 perf: warm re-baseline via perf-only reruns (cold d0 c1 1772 -> 2018-2103), spec numbers corrected
|
1 mēnesi atpakaļ |
|
Alexander Korolev
|
314812116c
feat: sm121 bench --perf-only (skip scenarios, re-baseline the perf table after a cold first row)
|
1 mēnesi atpakaļ |
|
Alexander Korolev
|
5dc8f91220
4x r39 verified: boot clean (comm.roce 81/81), bench 164/176 v2.7.0, perf +53% prefill / 2x long-decode vs 2x
|
1 mēnesi atpakaļ |
|
Alexander Korolev
|
bb1b4c83fa
recipe: glm53-flash-nvfp4-lab-4x-r39 (r39 tip on the 4x 512K config)
|
1 mēnesi atpakaļ |
|
Alexander Korolev
|
f1e87ec828
history: night 09-21/22 shift log (r39 verified, r38 4x repro, control baseline)
|
1 mēnesi atpakaļ |
