From 84a6c649c0ea49cce3feebb3cbd78f2972edc204 Mon Sep 17 00:00:00 2001
From: jialin <jialinkuang@126.com>
Date: Thu, 24 Sep 2026 16:41:49 +0800
Subject: [PATCH] fix(llmodels): keep the scaling baseline in sync with list
 replica actions

The inline replica edit and start/stop sent only `replicas`, leaving
`baseline_replicas` stale, so the scheduler reconciled the count straight
back and the action silently did nothing. Start also fell back to a
hardcoded 1 instead of the declared baseline.

Refs #6257
---
 src/pages/llmodels/components/table-list.tsx | 44 ++++++++++++++++++--
 src/pages/llmodels/config/types.ts           |  4 ++
 2 files changed, 45 insertions(+), 3 deletions(-)

diff --git a/src/pages/llmodels/components/table-list.tsx b/src/pages/llmodels/components/table-list.tsx
index 826d04f8c..39894629c 100644
--- a/src/pages/llmodels/components/table-list.tsx
+++ b/src/pages/llmodels/components/table-list.tsx
@@ -111,6 +111,40 @@ const getFormattedData = (record: any, extraData = {}) => ({
   }
 });
 
+// While scheduled scaling is on, `replicas` is the live count the scheduler
+// writes and `baseline_replicas` is the idle count the user actually declared;
+// the deploy form keeps the two equal (see `forms/index.tsx`). A list action
+// that moved only `replicas` left the baseline behind, so the next reconcile
+// drove the count straight back to it and the action looked like it had done
+// nothing at all.
+const getReplicasUpdate = (record: ListItem, replicas: number) => {
+  const schedule = record.scaling_schedule;
+  if (!schedule?.enabled) {
+    return { replicas };
+  }
+  return {
+    replicas,
+    scaling_schedule: { ...schedule, baseline_replicas: replicas }
+  };
+};
+
+// Start is only offered on a row already at `replicas: 0`, so the fallback is
+// what actually decides the count — the row's own value can never be anything
+// but the hardcoded 1. Under a schedule the baseline the user declared is the
+// better answer. A baseline of 0 is deliberately not one of those: stopping
+// from this list writes exactly that, and honouring it would make the next
+// start a no-op. Whether a zero baseline should mean something else is the
+// open question in gpustack/gpustack#6257.
+const getStartReplicas = (record: ListItem) => {
+  if (record.replicas) {
+    return record.replicas;
+  }
+  const baseline = record.scaling_schedule?.enabled
+    ? record.scaling_schedule.baseline_replicas
+    : null;
+  return baseline || 1;
+};
+
 const Models: React.FC<ModelsProps> = ({
   handleNameChange,
   handleSearch,
@@ -208,7 +242,9 @@ const Models: React.FC<ModelsProps> = ({
 
   const handleOnCell = useMemoizedFn(async (record: any, extra: any) => {
     try {
-      await updateModel(getFormattedData(record, { replicas: extra.newValue }));
+      await updateModel(
+        getFormattedData(record, getReplicasUpdate(record, extra.newValue))
+      );
       message.success(intl.formatMessage({ id: 'common.message.success' }));
       if (extra.newValue > extra.oldValue) {
         updateExpandedRowKeys([record.id, ...expandedRowKeys]);
@@ -229,11 +265,13 @@ const Models: React.FC<ModelsProps> = ({
   );
 
   const handleStartModel = async (row: ListItem) => {
-    await updateModel(getFormattedData(row, { replicas: row.replicas || 1 }));
+    await updateModel(
+      getFormattedData(row, getReplicasUpdate(row, getStartReplicas(row)))
+    );
   };
 
   const handleStopModel = async (row: ListItem) => {
-    await updateModel(getFormattedData(row, { replicas: 0 }));
+    await updateModel(getFormattedData(row, getReplicasUpdate(row, 0)));
     removeExpandedRowKey([row.id]);
   };
 
diff --git a/src/pages/llmodels/config/types.ts b/src/pages/llmodels/config/types.ts
index c02d8232e..d79bd0d7d 100644
--- a/src/pages/llmodels/config/types.ts
+++ b/src/pages/llmodels/config/types.ts
@@ -51,6 +51,10 @@ export interface ListItem {
   };
   gpu_type_selector?: GPUTypeSelector | null;
   worker_selector?: object;
+  // Present on the row so a list action can carry the schedule back untouched
+  // apart from its baseline. While it is enabled, `replicas` above is the
+  // scheduler-driven live count, not the user's declared one.
+  scaling_schedule?: ScalingSchedule | null;
 
   // --- PD: user intent ---
   roles?: RoleSpec[] | null;