Alexander Korolev
|
f9abbeb509
feat(compose): recursive whole-hierarchy symlink when no glob is set
|
hace 2 meses |
Alexander Korolev
|
0208fc12f1
feat(compose): support subdirectory globs in composed_model sources
|
hace 2 meses |
Alexander Korolev
|
360f50446a
fix(compose): use printf for path assembly to avoid shq-in-quotes bug
|
hace 2 meses |
Alexander Korolev
|
892c28d1a4
feat(serve): add --force flag to ignore occupied nodes
|
hace 2 meses |
Alexander Korolev
|
de18759712
feat(serve): add composed_model for multi-repo model directories
|
hace 2 meses |
Alexander Korolev
|
51ed390d64
refactor(recipe): use relative symlinks instead of rsync for NF4 expert overlay
|
hace 2 meses |
Alexander Korolev
|
8419efeb18
fix(recipe): correct NF4 Lloyd env vars after kernel source inspection
|
hace 2 meses |
Alexander Korolev
|
70ff8fda18
feat(recipe): add GLM-5.2 NF4 Lloyd-Experts-GPTQ recipe, bump glm52 image to 1.4
|
hace 2 meses |
Alexander Korolev
|
67de1f34da
refactor(build): rename docker-images/glm52-int4-int8 to glm52
|
hace 2 meses |
Alexander Korolev
|
febdc3b9d7
feat(build): add Solar Open2 NVFP4 recipe with recipe-pinned versions
|
hace 2 meses |
Alexander Korolev
|
23ae7a6484
feat(serve): add solo mode and fix single-node launch bug
|
hace 3 meses |
Alexander Korolev
|
65198e60fc
feat(recipe): add built-in script aliases and after_stop hook
|
hace 3 meses |
Alexander Korolev
|
1ba1d99e05
fix: resolve review findings across modules
|
hace 3 meses |
Alexander Korolev
|
ac922d02af
fix(ssh)!: harden SSH layer and refuse unpinned host keys
|
hace 3 meses |
Alexander Korolev
|
4767f6fe4d
docs(agents): warn that cluster may be serving live before stop/power off
|
hace 3 meses |
Alexander Korolev
|
7ebb007f08
feat(recipe): GLM-5.2 hf-overrides, full cudagraph compilation, adaptive spec
|
hace 3 meses |
Alexander Korolev
|
b5905ac702
docs(agents): clarify path-ref resolves to container-side HF cache path
|
hace 3 meses |
Alexander Korolev
|
dcb28faf1e
fix(docker): rewrite resolved HF cache path to container mount target
|
hace 3 meses |
Alexander Korolev
|
8c68124cce
feat(recipe): GLM-5.2-MXPF4-Experts-GPTQ recipe
|
hace 3 meses |
Alexander Korolev
|
a7cd759af1
feat(recipe): model-subfolder path refs via {{name}} substitution
|
hace 3 meses |
Alexander Korolev
|
6cd245f9a8
feat(recipe): model-subfolder mounts resolved per-node from HF cache
|
hace 3 meses |
Alexander Korolev
|
ac19fa45b8
fix(serve): dry-run skips SSH probing so it works with cluster off
|
hace 3 meses |
Alexander Korolev
|
2b0deeaffa
feat(recipe): extra_mounts with config alias resolution + tilde + env injection
|
hace 3 meses |
Alexander Korolev
|
7519fcef2a
chore(deepseek-v4-flash): disable temperature/top_p chat-template kwargs
|
hace 3 meses |
Alexander Korolev
|
75a7055af5
feat(glm52): build on production-hybrid-1.3 base with pinned deps
|
hace 3 meses |
Alexander Korolev
|
d102308e93
fix(docker): correct build alive-detection and harden build flow
|
hace 3 meses |
Alexander Korolev
|
d14021070c
feat(recipe): enable DFlash speculative decoding for Laguna S 2.1
|
hace 3 meses |
Alexander Korolev
|
e692ec1813
feat(recipe): update ghcr.io/timothystewart6/vllm-gb10 to v0.25.1-gb10.1
|
hace 3 meses |
Alexander Korolev
|
e844cb481a
feat(recipe): add Laguna S 2.1 NVFP4 (single-node DGX Spark)
|
hace 3 meses |
Alexander Korolev
|
b1a2b5f937
fix(recipe): set XDG_CACHE_HOME to /root/.cache for JIT persistence
|
hace 3 meses |