Align CI business priority across node counts and model families

configs/ci-priority.yaml

@@ -6,9 +6,6 @@ version: 1
 base-score: 1.0
 
 adjustments:
-  # Preserve business priority as the primary key, then prefer smaller Slurm
-  # allocations within an otherwise identical class.
-  additional-node: -0.001
   event:
     push: 2.0
   multi-node: 1.25
@@ -26,7 +23,7 @@ adjustments:
     dynamo-sglang: 0.5
     dynamo-vllm: 0.5
   model-prefix:
-    kimik3: 1.5
+    kimik3: 0.75
     glm5: 0.75
     glm5.1: 0.75
     kimik2.5: 0.75

utils/ci_priority.py

@@ -149,7 +149,6 @@ def calculate_priority(
         or node_count < 1
     ):
         raise ValueError(f"node-count must be a positive integer, got {node_count!r}")
-    score += _decimal(adjustments.get("additional-node", 0)) * (node_count - 1)
 
     if entry.get("prefill") is not None:
         score += _decimal(adjustments.get("multi-node", 0))

utils/runner_setup/RUNNER_SETUP.md

@@ -209,9 +209,10 @@ hardware label bypasses admission accounting.
 Set `NODE_SLOT_SCHEDULER_ENABLED=true` only after the deployed priority
 controller supports `nodes:N` and `ci-lease-*`. `PRIORITY_SCHEDULER_ENABLED`
 must also remain enabled. If either variable is disabled, workflows omit
-`nodes:N` and retain the existing unweighted behavior. The priority score also
-subtracts `0.001` per additional node so otherwise equal work prefers smaller
-allocations without overriding the existing business-priority signals.
+`nodes:N` and retain the existing unweighted behavior. Node count is capacity
+demand, not a business-priority signal. The dashboard controller packs jobs
+within each business-priority band, maximizing admitted jobs and then occupied
+nodes without allowing lower-priority work to displace a higher-priority band.
 
 Aggregated multi-node search-space entries must declare one aggregate `worker`
 role and `num-nodes`; that value becomes the generated `node-count` directly.

utils/test_ci_priority.py

@@ -54,22 +54,30 @@ def test_main_branch_jobs_receive_an_automatic_boost():
     ) == Decimal("3.000")
 
 
-def test_smaller_node_allocations_win_otherwise_equal_priority_ties():
+def test_kimi_k3_and_dsv4_receive_the_same_model_priority():
+    entry = {"runner": "h100", "framework": "trt"}
+
+    assert calculate_priority(
+        {**entry, "model-prefix": "kimik3"}, POLICY
+    ) == calculate_priority({**entry, "model-prefix": "dsv4"}, POLICY)
+
+
+def test_node_count_does_not_change_business_priority():
     entry = {"runner": "h100", "framework": "trt"}
 
     assert calculate_priority({**entry, "node-count": 1}, POLICY) == Decimal("1.000")
-    assert calculate_priority({**entry, "node-count": 2}, POLICY) == Decimal("0.999")
-    assert calculate_priority({**entry, "node-count": 3}, POLICY) == Decimal("0.998")
+    assert calculate_priority({**entry, "node-count": 2}, POLICY) == Decimal("1.000")
+    assert calculate_priority({**entry, "node-count": 3}, POLICY) == Decimal("1.000")
 
 
-def test_node_count_tiebreaker_survives_classifier_projection():
+def test_node_count_validation_survives_classifier_projection():
     entry = {"runner": "h100", "framework": "trt", "node-count": 3}
 
     assert calculate_priority(
         entry,
         POLICY,
         PriorityContext(criteria=frozenset()),
-    ) == Decimal("0.998")
+    ) == Decimal("1.000")
 
 
 @pytest.mark.parametrize("node_count", [0, -1, 1.5, True])