feat: retry evicted models in one load action

README.md

@@ -47,7 +47,10 @@ The Available downloads controls support:
 
 Popularity and date ordering use upstream metadata only; the plugin does not invent popularity, dates, RAM requirements, or token metrics.
 
-## Per-model placement
+## Staged multi-model loading
+
+A single **Load selected permanently** action now loads models in verified stages. RAM-only models are attempted first, followed by GPU + RAM models. The backend checks Ollama `/api/ps` after each runner starts and automatically performs one recovery pass for any selected model Ollama evicted. Already resident models are not reloaded. The result includes a retry count and reports when automatic eviction recovery completed, so users do not need to click the load action again manually.
+
 
 Each installed model in the Model pool has a persistent placement selector:
 

dashboard/dist/index.js

@@ -336,7 +336,7 @@
         if (endpoint === "/models/load") {
           var resident = Array.isArray(result.resident) ? result.resident : ((result.runtime && result.runtime.model_memory) || []).map(function (item) { return item.name; });
           var missing = Array.isArray(result.not_resident) ? result.not_resident : requested.filter(function (name) { return resident.indexOf(name) < 0; });
-          setNotice(missing.length ? { warning: "Ollama kept resident: " + resident.join(", ") + ". Not resident: " + missing.join(", ") + ". This is an Ollama scheduler/capacity warning, not a plugin error." } : { ok: "Loaded and resident: " + resident.join(", ") });
+          setNotice(missing.length ? { warning: "Ollama kept resident: " + resident.join(", ") + ". Not resident: " + missing.join(", ") + ". This is an Ollama scheduler/capacity warning, not a plugin error." } : { ok: result.message || ("Loaded and resident: " + resident.join(", ")) });
         } else {
           setNotice({ ok: label + ": " + requested.join(", ") });
         }

dashboard/manifest.json

@@ -3,7 +3,7 @@
   "label": "Ollama Models",
   "description": "Inspect, manage, and chat with local Ollama models, including shared persistent conversations, performance metrics, images, PDFs, URLs, and live memory telemetry.",
   "icon": "Cpu",
-  "version": "1.5.15",
+  "version": "1.5.16",
   "tab": {"path": "/ollama-manager", "position": "after:models"},
   "entry": "dist/index.js",
   "css": "dist/style.css",

dashboard/plugin_api.py

@@ -1446,21 +1446,29 @@ def models_load(body: ModelsRequest) -> dict[str, Any]:
     if not names:
         raise HTTPException(400, "Select at least one model to load")
     placements = {name: _model_placement(body.placements.get(name)) for name in names}
-    results = []
+    # Load RAM-only models first, then GPU+RAM models. This avoids asking the
-    for name in names:
+    # GPU scheduler to rearrange an already GPU-resident runner unnecessarily.
-        _model_load_update(name, active=True, state="queued", stage="Waiting for Ollama", started_at=time.time(), error="", placement=placements[name])
+    ordered_names = sorted(names, key=lambda name: 0 if placements[name] == "ram_only" else 1)
-    for name in names:
+    results_by_name: dict[str, dict[str, Any]] = {}
-        placement = placements[name]
+    for pass_index in range(2):
-        stage = "Loading into GPU + RAM" if placement == "gpu_ram" else "Loading into system RAM only"
+        resident_now = {str(row.get("name") or row.get("model")) for row in _local_ps()}
-        _model_load_update(name, state="loading", stage=stage)
+        missing_now = [name for name in ordered_names if name not in resident_now]
-        try:
+        if not missing_now:
-            results.append({"name": name, "placement": placement, "ok": True, "result": _load_model(name, placement)})
+            break
-            _model_load_update(name, state="checking", stage="Checking Ollama resident state")
+        for name in missing_now:
-        except Exception as exc:
+            placement = placements[name]
-            results.append({"name": name, "placement": placement, "ok": False, "error": str(exc)})
+            stage = "Loading into GPU + RAM" if placement == "gpu_ram" else "Loading into system RAM only"
-            _model_load_update(name, active=False, state="failed", stage="Ollama load failed", finished_at=time.time(), error=str(exc))
+            _model_load_update(name, active=True, state="loading", stage=stage + " · pass " + str(pass_index + 1), attempt=pass_index + 1)
+            try:
+                result = _load_model(name, placement)
+                results_by_name[name] = {"name": name, "placement": placement, "ok": True, "result": result, "attempts": pass_index + 1}
+                _model_load_update(name, state="checking", stage="Checking Ollama resident state", attempt=pass_index + 1)
+            except Exception as exc:
+                results_by_name[name] = {"name": name, "placement": placement, "ok": False, "error": str(exc), "attempts": pass_index + 1}
+                _model_load_update(name, active=False, state="failed", stage="Ollama load failed", finished_at=time.time(), error=str(exc), attempt=pass_index + 1)
     resident_rows = _local_ps()
     resident_names = {str(row.get("name") or row.get("model")) for row in resident_rows}
+    results = [results_by_name.get(name, {"name": name, "placement": placements[name], "ok": name in resident_names, "attempts": 0}) for name in names]
     for item in results:
         item["resident"] = item["name"] in resident_names
     not_resident = [name for name in names if name not in resident_names]
@@ -1470,16 +1478,20 @@ def models_load(body: ModelsRequest) -> dict[str, Any]:
         elif item["ok"]:
             _model_load_update(item["name"], active=False, state="evicted", stage="Ollama did not retain this model", finished_at=time.time())
     runtime = _runtime_snapshot()
+    retry_count = sum(max(0, int(item.get("attempts") or 0) - 1) for item in results)
     return {
         "ok": bool(results) and not not_resident and all(item["ok"] for item in results),
         "requested": names,
         "resident": sorted(resident_names),
         "not_resident": not_resident,
         "results": results,
+        "retries": retry_count,
         "runtime": runtime,
         "keep_alive": "permanent",
         "message": (
-            "All selected models are resident."
+            "All selected models are resident. Automatic Ollama eviction recovery completed."
+            if not not_resident and retry_count
+            else "All selected models are resident."
             if not not_resident
             else "Ollama did not keep every selected model resident; see not_resident."
         ),

plugin.yaml

@@ -1,5 +1,5 @@
 name: ollama-manager
-version: 1.5.15
+version: 1.5.16
 description: Native dashboard manager and chat interface for local Ollama models, attachments, URLs, shared persistent conversations, performance metrics, and live runtime telemetry.
 auto_install_dependencies: true
 python_dependencies: