 Alexander Korolev
|
4e24cafb19
fix: revert kernel.py overlay (uses private b12x symbols)
|
3 ヶ月 前 |
|
Alexander Korolev
|
ff8340a20a
fix: add real CosmicRaisins DCP patches + sm120-return-lse.patch
|
3 ヶ月 前 |
|
Alexander Korolev
|
d8048a2b1d
fix: sync overlay with madeby561/vllm-glm52-nvfp4-nf3-hybrid:v3
|
3 ヶ月 前 |
|
Alexander Korolev
|
126fe81115
fix: drop_caches after image sync (clear docker load page cache)
|
3 ヶ月 前 |
|
Alexander Korolev
|
810d952c72
fix: b12x non-editable install (files land in site-packages)
|
3 ヶ月 前 |
|
Alexander Korolev
|
b37592678f
feat: drop_caches (echo 3 | sudo tee /proc/sys/vm/drop_caches) before start
|
3 ヶ月 前 |
|
Alexander Korolev
|
673acbe1ff
fix: switch to voipmonitor/b12x@f3686b55 (eldritch stack)
|
3 ヶ月 前 |
|
Alexander Korolev
|
e89fe5d869
fix: sync_image_to_nodes selects most recent image (by creation timestamp)
|
3 ヶ月 前 |
|
Alexander Korolev
|
c04454ce6d
feat: remote build via tmux (disconnect-safe, polling)
|
3 ヶ月 前 |
|
Alexander Korolev
|
731937a625
feat: --build-node, docker_image_dir recipe field
|
3 ヶ月 前 |
|
Alexander Korolev
|
9391712331
feat: --build-node flag to auto-build image on a cluster node
|
3 ヶ月 前 |
|
Alexander Korolev
|
837eb17d52
feat: self-contained build system in docker-images/glm52-nf3-hybrid (hash-pinned locks, versions.env, patches)
|
3 ヶ月 前 |
|
Alexander Korolev
|
2d5c87150b
chore: add nf3-hybrid branch ref + patches + build README
|
3 ヶ月 前 |
|
Alexander Korolev
|
b0c3daeca5
fix: install b12x with --no-deps to preserve base torch 2.11.0 compatibility
|
3 ヶ月 前 |
|
Alexander Korolev
|
7d4253fe26
fix: use run_ssh_long_running for docker save|load pipe (no inactivity timeout)
|
3 ヶ月 前 |
|
Alexander Korolev
|
acadda1259
feat: auto-sync Docker image (ANY source node) + --sync-image / --sync-image-source flags
|
3 ヶ月 前 |
|
Alexander Korolev
|
cae048890c
feat: auto-sync Docker image from head node to workers before serve
|
3 ヶ月 前 |
|
Alexander Korolev
|
3ae04cd229
fix: remove explicit port from NF3 hybrid recipes (use default)
|
3 ヶ月 前 |
|
Alexander Korolev
|
1b4f49aaa6
feat: bump to vllm-gb10 v0.24.0-gb10.5 base (vLLM 0.24.1 + b12x)
|
3 ヶ月 前 |
|
Alexander Korolev
|
421d123a31
feat: GLM-5.2 NF3-Hybrid image builds successfully (22GB, vLLM 0.23.1 + b12x)
|
3 ヶ月 前 |
|
Alexander Korolev
|
f5e65e3338
feat: vllm-gb10 based Dockerfile for GLM-5.2 NF3-Hybrid
|
3 ヶ月 前 |
|
Alexander Korolev
|
fc2593bb0b
feat: rewrite GLM-5.2 NF3-Hybrid Dockerfile (single-stage, local repos, real overlay kernels)
|
3 ヶ月 前 |
|
Alexander Korolev
|
15698ba6b7
feat: add GLM-5.2 NF3-Hybrid recipes (context/fast lanes) + build system
|
3 ヶ月 前 |
|
Alexander Korolev
|
31aaf141e5
refactor: split ends (HA/SSH) from control logic with traits and orchestrator
|
3 ヶ月 前 |
|
Alexander Korolev
|
56bcdf4bf5
feat(recipe): update deepseek-v4-flash-dspark from NVIDIA reference
|
3 ヶ月 前 |
|
Alexander Korolev
|
66b9879240
fix: auto-power-on now powers on per-node entities, not just cluster extras
|
3 ヶ月 前 |
|
Alexander Korolev
|
05f923dfaf
fix: add served-model-name with two names to dspark recipe
|
3 ヶ月 前 |
|
Alexander Korolev
|
732cef4a04
fix: remove served-model-name and custom port from dspark recipe
|
3 ヶ月 前 |
|
Alexander Korolev
|
d6cd9196d4
feat: add deepseek-v4-flash-dspark recipe
|
3 ヶ月 前 |
|
Alexander Korolev
|
7158e63649
refactor: separate cluster state from SSH probing for testability
|
3 ヶ月 前 |
