Skip to content

Worker init parity — st_init_multi + COLI_MODEL_DIRS #5

Description

@lineape

What to build

Make the worker's init match a sharded coordinator: honor COLI_MODEL_DIRS via st_init_multi instead of the single-directory st_init, so a worker can resolve expert tensors split across drives. Surface a clear worker error (not a mid-connection fatal) when an expert tensor can't be resolved. Do not port the rest of model_init (tokenizer/KVCache/router/embed/MTP) into the worker — dead weight for an expert-only worker.

Acceptance criteria

  • Worker resolves expert tensors through st_init_multi honoring COLI_MODEL_DIRS.
  • An unresolvable expert tensor surfaces as a clear worker-side error, not a mid-connection fatal/exit.
  • Single-directory models still load identically (no regression).

Environment: codegraph-indexed — run codegraph explore first; refs in /home/eli/dev/_reference-docs-and-repos/colibri.

Metadata

Metadata

Assignees

No one assigned

    Labels

    effort/lowFlash-capable — docs/fixtures/simple changesready-for-agentFully specified, AFK-ready

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions