Alexander Korolev
|
ae1285dfdf
fix: reduce gpu-memory-utilization to 0.75 for deepseek-v4-flash
|
hai 3 meses |
Alexander Korolev
|
04e04b4e8c
fix: add initial boot delay and retry logic to auto-power-on
|
hai 3 meses |
Alexander Korolev
|
9123df86cd
fix: use original cluster name for solo instances instead of 'solo/<node>'
|
hai 4 meses |
Alexander Korolev
|
27c02331d8
refactor: rename --power-on-if-off to --auto-power-on
|
hai 4 meses |
Alexander Korolev
|
811b8b0848
feat: add --power-on-if-off and --power-off flags
|
hai 4 meses |
Alexander Korolev
|
176946c99f
feat: power off command stops containers first with timeout
|
hai 4 meses |
Alexander Korolev
|
98abe64703
fix: address soundness issues from audit
|
hai 4 meses |
Alexander Korolev
|
0ae454c259
feat: add --node parameter to logs command
|
hai 4 meses |
Alexander Korolev
|
4b08863c18
chore: update qwen-agentworld-35b recipe with tool call args and model name
|
hai 4 meses |
Alexander Korolev
|
81f43858be
feat: health checks NCCL socket interface status per node
|
hai 4 meses |
Alexander Korolev
|
0c19edf62d
feat: health checks HA cluster entities (switch) to detect switched-off clusters
|
hai 4 meses |
Alexander Korolev
|
604f61491b
feat: improved health command with cluster-scoped container view
|
hai 4 meses |
Alexander Korolev
|
cefdfa2368
feat: deployment discovery + stop ambiguity resolution
|
hai 4 meses |
Alexander Korolev
|
3774e765e5
feat: cluster selection with occupancy awareness
|
hai 4 meses |
Alexander Korolev
|
ea8fb4e287
fix: select_cluster should use >= not == for required_nodes matching
|
hai 4 meses |
Alexander Korolev
|
2513ac6f85
feat: add Qwen-AgentWorld-35B-A3B recipe (2× DGX Spark, TP=2)
|
hai 4 meses |
Alexander Korolev
|
3d986923c7
fix: drop incompatible -q flag from container label query (docker -q + --format conflict)
|
hai 4 meses |
Alexander Korolev
|
68087045df
fix: resolve_cluster_by_container_label now actually reads the sm121.cluster label from container metadata instead of inferring cluster from node ownership
|
hai 4 meses |
Alexander Korolev
|
2946e07f79
fix: stop_cluster filters by cluster label to avoid cross-cluster interference
|
hai 4 meses |
Alexander Korolev
|
d478170365
fix: switch MiMo recipe to mp backend, remove Ray-specific env vars
|
hai 4 meses |
Alexander Korolev
|
5837205e68
feat: add MiMo-V2.5 Omni NVFP4 recipe (2× DGX Spark, Ray, 1M ctx)
|
hai 4 meses |
Alexander Korolev
|
4936530710
feat: RecipeArgs enum — inline string or list for recipe args
|
hai 4 meses |
Alexander Korolev
|
5931869d69
refactor: DistributedBackend as Option — remove None_ variant
|
hai 4 meses |
Alexander Korolev
|
275e5f47ba
fix: keep serde(rename = "none") so existing recipes with 'distributed_backend = "none"' still parse
|
hai 4 meses |
Alexander Korolev
|
09841103d8
fix: default DistributedBackend to None_ (absent means single-node)
|
hai 4 meses |
Alexander Korolev
|
dbdbc7ed62
refactor: replace String distributed_backend with DistributedBackend enum
|
hai 4 meses |
Alexander Korolev
|
6081ccf443
feat: mods system — entrypoint wrapping + global/recipe mods with volume mount
|
hai 4 meses |
Alexander Korolev
|
09ac2dc3c4
fix(deepseek-v4-flash): revert wrong tp pp changes
|
hai 4 meses |
Alexander Korolev
|
746049fc18
fix(cli): switch clap help texts to English
|
hai 4 meses |
Alexander Korolev
|
aa34a575e4
fix(i18n): wire up unused Fluent keys in before_serve, host_keys, health, exec
|
hai 4 meses |