Place each engine on a single pool instead of splitting across pools

[negz]

Description of your changes

The scheduler placed an engine's members on one pool when it could, and split them across pools when no single pool fit. Splitting is unsafe, as @dennis-upbound pointed out reviewing #137: the scheduler can't reason about interconnect fabric. Pool identity is the finest grain it has, so when a gang's members land on different pools it can't tell whether those pools share a fabric.

A Leader/Worker gang doing tensor or pipeline parallel talks over its pool's fabric — NVLink within a node, InfiniBand within a pool. Scatter its members across pools that don't share a fabric and the collective never forms: the gang sits NotReady with no signal saying why. Splitting doesn't degrade the interconnect, it removes it — handing back a placement that can't run rather than a slower one.

This change places every member of an engine on a single pool, or rejects the engine on that cluster. Rejection is safe and recoverable — the fill phase tries another cluster, and the deployment surfaces InsufficientCapacity if none fit — whereas a cross-fabric placement is a silent hang. Per-member nodeSelectors stay (a member can still claim different devices from the shared pool, and a claimless member rides along at zero node cost); they just no longer place members on different pools.

Co-scheduling a gang across pools that genuinely share a fabric — an operator exposing low-utilization nodes beside the GPU nodes as a separate pool, say — needs the scheduler to model fabric, which it doesn't yet. That's tracked in #149.

The change is concentrated in _place_engine, now reduced to "place the whole engine on the first single pool that fits, or reject." _place_engine_split and _place_member are gone. That's the part worth reviewer attention.

I have:

• Read and followed Modelplane's contribution process.
• Run nix flake check (or ./nix.sh flake check) and made sure it passes.
• Added or updated tests covering any composition function changes.
• Signed off every commit with git commit -s.

[Copilot]

Pull request overview

Updates the compose-model-deployment scheduler to ensure all members of an engine are placed onto a single GPU pool within a cluster (or the engine is rejected for that cluster), avoiding cross-pool placements that can silently hang multi-node collectives when pools don’t share an interconnect fabric.

Changes:

• Remove cross-pool “split placement” for engine members; place whole engines on the first single pool that satisfies all members and capacity.
• Update scheduler documentation/comments to reflect the new single-pool constraint and its motivation (#149).
• Expand scheduling tests to cover rejection when no single pool fits and successful placement onto an alternate cluster when one cluster can’t fit the whole engine.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
functions/compose-model-deployment/function/scheduling.py	Enforces single-pool-per-engine placement and removes split-member placement logic.
functions/compose-model-deployment/tests/test_scheduling.py	Updates/extends tests to validate rejection and cross-cluster fallback under the new placement rules.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[dennis-upbound]

lgtm