Skip to content

Rebase #551 onto current dev and restore a compiling tree #2

Description

@lineape

What to build

A fresh branch off current upstream dev carrying PR JustVugg#551's distributed-expert-workers work, with conflicts resolved and a compiling engine. This is the foundation every other ticket builds on.

Rebase, don't re-implement: merge pr551 (head 0af26a3) onto upstream/dev (0f335e5), resolve the one conflicted file (c/coli — two hunks: the openai_server.serve() signature drift and the cluster subparser block), and reconcile the worker's expert-FFN call sequence with the rebased matmul_qt_ex (which gained fp8/CUDA/Metal arms upstream) and the now-merged g_pre_idx pre-routed routing path (JustVugg#432).

Acceptance criteria

  • Branch off current upstream/dev; pr551 merged; zero unresolved conflict markers.
  • Engine compiles clean (make colibri), 0 warnings.
  • Existing test suite still green (make check / make test-c).
  • Worker path (cluster_worker_run) and coordinator routing wire together post-rebase — the g_pre_idx pre-routed path and cluster_moe_batch agree on who consumes routing.
  • Single-machine path unchanged (cluster env-gated OFF; default build byte-identical).

Environment: the repo is codegraph-indexed (.codegraph/ at the clone root) — run codegraph explore (or the codegraph MCP tool) before grep/read. Refs pr550/pr551/upstream/dev are already fetched in /home/eli/dev/_reference-docs-and-repos/colibri; work in a fresh branch there.

Metadata

Metadata

Assignees

No one assigned

    Labels

    effort/mediumNeeds Pro — refactors/cross-module/moderate logicready-for-agentFully specified, AFK-ready

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions