v0.2.3 - Per-model Advanced settings + catalog-add for downloaded models

Backend:
- overrides.py: read/write /data/models-overrides.yaml (knobs + custom entries)
- apply_knobs_to_args(): strip matching flags from bundled vllm_args and append knob values, so knob changes properly override bundled defaults
- extract_knobs_from_args(): seed UI knob values from bundled args so the Advanced dialog has correct starting state
- models.py: load_catalog merges overrides on top of bundled yaml
- GET /api/models returns effective_knobs per model
- PUT /api/models/{key}/knobs persists knob changes
- POST /api/models adds a custom catalog entry
- DELETE /api/models/{key} removes a custom entry (bundled models cannot be deleted)
- swap_manager.reload_catalog() called after each mutation so swaps see latest

Frontend:
- New 'Advanced' button on every card opens a modal dialog: max-model-len input, gpu-memory-utilization slider, three optimization checkboxes (fastsafetensors, prefix caching, FP8 KV cache). Save persists; Cancel discards. Custom models also have a Delete button.
- After a successful download, automatically open the 'Add to catalog' dialog pre-filled with the repo, with the same knob defaults — user just enters key, display name, and clicks Save.
- Custom catalog entries are tagged with a blue 'custom' pill on the card.

Package: bump 0.2.3:0; main.ts sets MODELS_OVERRIDES=/data/models-overrides.yaml so overrides persist on the StartOS volume.

image/app/models.py

@@ -1,8 +1,10 @@
 from __future__ import annotations
-from typing import Literal
+from typing import Literal, Optional
 import yaml
 from pydantic import BaseModel, Field
 
+from .overrides import apply_knobs_to_args, load_overrides
+
 
 class ModelDef(BaseModel):
     display_name: str
@@ -13,6 +15,8 @@ class ModelDef(BaseModel):
     expected_ready_seconds: int = 300
     vllm_args: list[str] = Field(default_factory=list)
     description: str | None = None
+    knobs: dict | None = None       # user-customized; merged at launch time
+    custom: bool = False             # True if this came from /data overrides
 
 
 class Defaults(BaseModel):
@@ -25,17 +29,52 @@ class Catalog(BaseModel):
     models: dict[str, ModelDef]
 
 
+def _merge_overrides(catalog: Catalog) -> Catalog:
+    """Apply user overrides + custom entries from /data/models-overrides.yaml."""
+    ov = load_overrides()
+    knobs_by_key = ov.get("knobs") or {}
+    custom_entries = ov.get("custom") or []
+
+    new_models: dict[str, ModelDef] = {}
+    for key, m in catalog.models.items():
+        k = knobs_by_key.get(key)
+        new_models[key] = m.model_copy(update={"knobs": k}) if k else m
+
+    for entry in custom_entries:
+        key = entry.get("key")
+        if not key:
+            continue
+        defaults_dump = {
+            "display_name": entry.get("display_name", key),
+            "repo": entry["repo"],
+            "size_gb": float(entry.get("size_gb", 0)),
+            "mode": entry.get("mode", "solo"),
+            "capabilities": entry.get("capabilities") or [],
+            "expected_ready_seconds": int(entry.get("expected_ready_seconds", 300)),
+            "vllm_args": entry.get("vllm_args") or [],
+            "description": entry.get("description"),
+            "knobs": entry.get("knobs"),
+            "custom": True,
+        }
+        new_models[key] = ModelDef.model_validate(defaults_dump)
+
+    return Catalog(defaults=catalog.defaults, models=new_models)
+
+
 def load_catalog(path: str) -> Catalog:
     with open(path) as f:
         data = yaml.safe_load(f)
-    return Catalog.model_validate(data)
+    bundled = Catalog.model_validate(data)
+    return _merge_overrides(bundled)
 
 
 def build_launch_command(key: str, model: ModelDef, defaults: Defaults) -> str:
     """Return the shell command to launch `model` on Spark 1.
 
+    User knobs (if any) override matching flags in the bundled vllm_args.
     Assumes cwd will be `~/spark-vllm-docker` (we cd in the SSH wrapper).
     """
     solo = "--solo " if model.mode == "solo" else ""
-    args = [f"--port={defaults.port}", f"--host={defaults.host}", *model.vllm_args]
+    base_args = apply_knobs_to_args(list(model.vllm_args), model.knobs)
+    args = [f"--port={defaults.port}", f"--host={defaults.host}", *base_args]
     return f"./launch-cluster.sh {solo}-d exec vllm serve {model.repo} {' '.join(args)}"

image/app/overrides.py

@@ -0,0 +1,145 @@
+"""User overrides + custom catalog entries, persisted on the package's main volume.
+
+File: /data/models-overrides.yaml (only when something has been customized).
+Survives package updates because /data is the StartOS volume.
+
+Shape:
+    knobs:
+      qwen36:
+        max_model_len: 65536
+        gpu_memory_utilization: 0.85
+        fastsafetensors: true
+        prefix_caching: true
+        kv_cache_dtype: fp8     # one of: 'fp8' | 'auto' | null
+    custom:
+      - key: my-new-model
+        display_name: My New Model (from download)
+        repo: my-org/my-model
+        size_gb: 20
+        mode: solo
+        description: null
+        capabilities: []
+        knobs:
+          max_model_len: 32768
+          gpu_memory_utilization: 0.85
+          fastsafetensors: true
+          prefix_caching: true
+          kv_cache_dtype: fp8
+"""
+from __future__ import annotations
+import os
+from pathlib import Path
+from typing import Any, Optional
+import yaml
+
+
+KNOWN_FLAG_MAP = {
+    "max_model_len": "--max-model-len",
+    "gpu_memory_utilization": "--gpu-memory-utilization",
+}
+
+
+def _path() -> str:
+    # Container: /data/models-overrides.yaml. Local dev: cwd or env override.
+    return os.environ.get("MODELS_OVERRIDES", "/data/models-overrides.yaml")
+
+
+def load_overrides() -> dict:
+    p = _path()
+    try:
+        with open(p) as f:
+            data = yaml.safe_load(f) or {}
+    except FileNotFoundError:
+        return {"knobs": {}, "custom": []}
+    data.setdefault("knobs", {})
+    data.setdefault("custom", [])
+    return data
+
+
+def save_overrides(data: dict) -> None:
+    p = _path()
+    parent = Path(p).parent
+    parent.mkdir(parents=True, exist_ok=True)
+    with open(p, "w") as f:
+        yaml.safe_dump(data, f, sort_keys=False)
+
+
+def set_knobs(key: str, knobs: dict) -> dict:
+    data = load_overrides()
+    if not knobs:
+        data["knobs"].pop(key, None)
+    else:
+        data["knobs"][key] = knobs
+    save_overrides(data)
+    return data
+
+
+def add_custom(entry: dict) -> dict:
+    data = load_overrides()
+    # Replace if same key already exists
+    data["custom"] = [c for c in data["custom"] if c.get("key") != entry["key"]]
+    data["custom"].append(entry)
+    save_overrides(data)
+    return data
+
+
+def delete_custom(key: str) -> dict:
+    data = load_overrides()
+    data["custom"] = [c for c in data["custom"] if c.get("key") != key]
+    data["knobs"].pop(key, None)
+    save_overrides(data)
+    return data
+
+
+def apply_knobs_to_args(vllm_args: list[str], knobs: Optional[dict]) -> list[str]:
+    """Return a new vllm_args list with knob values overriding matching flags."""
+    if not knobs:
+        return list(vllm_args)
+    skip_prefixes: list[str] = []
+    for k, flag in KNOWN_FLAG_MAP.items():
+        if k in knobs:
+            skip_prefixes.append(f"{flag}=")
+            skip_prefixes.append(flag + " ")  # rare style
+    # opt-in flag toggles
+    if "fastsafetensors" in knobs:
+        skip_prefixes.append("--load-format=")
+    if "prefix_caching" in knobs:
+        skip_prefixes.append("--enable-prefix-caching")
+    if "kv_cache_dtype" in knobs:
+        skip_prefixes.append("--kv-cache-dtype=")
+
+    filtered = [a for a in vllm_args if not any(a.startswith(p) or a == p.rstrip("= ") for p in skip_prefixes)]
+
+    for k, flag in KNOWN_FLAG_MAP.items():
+        if k in knobs and knobs[k] is not None:
+            filtered.append(f"{flag}={knobs[k]}")
+
+    if knobs.get("fastsafetensors"):
+        filtered.append("--load-format=fastsafetensors")
+    if knobs.get("prefix_caching"):
+        filtered.append("--enable-prefix-caching")
+    kvd = knobs.get("kv_cache_dtype")
+    if kvd and kvd != "auto":
+        filtered.append(f"--kv-cache-dtype={kvd}")
+
+    return filtered
+
+
+def extract_knobs_from_args(vllm_args: list[str]) -> dict:
+    """Reverse: read default knob values from a model's bundled vllm_args, so the UI
+    has a starting point even before the user has saved any overrides."""
+    knobs: dict[str, Any] = {}
+    for a in vllm_args:
+        if a.startswith("--max-model-len="):
+            try: knobs["max_model_len"] = int(a.split("=", 1)[1])
+            except ValueError: pass
+        elif a.startswith("--gpu-memory-utilization="):
+            try: knobs["gpu_memory_utilization"] = float(a.split("=", 1)[1])
+            except ValueError: pass
+        elif a == "--load-format=fastsafetensors":
+            knobs["fastsafetensors"] = True
+        elif a == "--enable-prefix-caching":
+            knobs["prefix_caching"] = True
+        elif a.startswith("--kv-cache-dtype="):
+            knobs["kv_cache_dtype"] = a.split("=", 1)[1]
+    return knobs

image/app/server.py

@@ -13,6 +13,7 @@ from .config import Settings
 from .download import DownloadManager
 from .health import check_magpie, check_parakeet, check_vllm
 from .models import load_catalog
+from .overrides import add_custom, delete_custom, extract_knobs_from_args, load_overrides, set_knobs
 from .services import docker_state, run_action, services_from_settings
 from .ssh import ssh_run
 from .swap import SwapManager
@@ -46,14 +47,75 @@ async def get_config() -> dict:
     }
 
 
+def _reload_catalog() -> None:
+    global catalog
+    catalog = load_catalog(settings.models_yaml)
+    swap_manager.reload_catalog(catalog)
+
+
 @app.get("/api/models")
 async def get_models() -> dict:
+    out_models: dict[str, dict] = {}
+    for key, m in catalog.models.items():
+        d = m.model_dump()
+        # Always include effective knobs for the UI (defaults from base args + any overrides)
+        d["effective_knobs"] = {**extract_knobs_from_args(m.vllm_args), **(m.knobs or {})}
+        out_models[key] = d
     return {
         "defaults": catalog.defaults.model_dump(),
-        "models": {k: v.model_dump() for k, v in catalog.models.items()},
+        "models": out_models,
     }
 
 
+class KnobsBody(BaseModel):
+    knobs: dict
+
+
+@app.put("/api/models/{key}/knobs")
+async def put_model_knobs(key: str, body: KnobsBody) -> dict:
+    if key not in catalog.models:
+        raise HTTPException(404, f"unknown model: {key}")
+    # Strip empty/None values
+    clean = {k: v for k, v in body.knobs.items() if v not in (None, "")}
+    set_knobs(key, clean)
+    _reload_catalog()
+    return {"ok": True, "key": key, "knobs": clean}
+
+
+class CustomModelBody(BaseModel):
+    key: str
+    display_name: str
+    repo: str
+    size_gb: float = 0
+    mode: Literal["solo", "cluster"] = "solo"
+    description: str | None = None
+    capabilities: list[str] = []
+    vllm_args: list[str] = []
+    knobs: dict | None = None
+
+
+@app.post("/api/models")
+async def post_model(body: CustomModelBody) -> dict:
+    if not body.key or not body.key.replace("-", "").replace("_", "").isalnum():
+        raise HTTPException(400, "key must be alphanumeric/-/_ only")
+    if body.key in catalog.models and not catalog.models[body.key].custom:
+        raise HTTPException(409, f"'{body.key}' is a bundled model — pick a different key")
+    add_custom(body.model_dump())
+    _reload_catalog()
+    return {"ok": True, "key": body.key}
+
+
+@app.delete("/api/models/{key}")
+async def del_model(key: str) -> dict:
+    if key not in catalog.models:
+        raise HTTPException(404, f"unknown model: {key}")
+    if not catalog.models[key].custom:
+        raise HTTPException(400, "cannot delete a bundled model; you may override its knobs instead")
+    delete_custom(key)
+    _reload_catalog()
+    return {"ok": True, "key": key}
+
+
 @app.get("/api/services")
 async def get_services() -> dict:
     """Lifecycle state of always-on support services (Parakeet, Magpie, …).

image/app/static/app.js

@@ -53,24 +53,32 @@ function renderCards() {
     const desc = m.description
       ? `<div class="desc">${escapeHtml(m.description)}</div>`
       : '';
+    const customPill = m.custom ? `<span class="tag custom-pill">custom</span>` : '';
     card.innerHTML = `
       <div class="name">${escapeHtml(m.display_name)}</div>
       <div class="meta">
         <span class="tag mode-${m.mode}">${m.mode}</span>
         <span class="tag">${m.size_gb} GB</span>
+        ${customPill}
         ${(m.capabilities || []).map(c => `<span class="tag cap">${escapeHtml(c)}</span>`).join('')}
       </div>
       ${desc}
       <div class="muted small repo">${escapeHtml(m.repo)}</div>
       <div class="spacer"></div>
-      <button class="btn ${isActive ? '' : 'primary'}" data-key="${key}" ${isActive || isSwapping ? 'disabled' : ''}>
+      <div class="card-actions">
+        <button class="btn ${isActive ? '' : 'primary'}" data-swap-key="${key}" ${isActive || isSwapping ? 'disabled' : ''}>
           ${isActive ? 'Current' : 'Switch to this'}
         </button>
+        <button class="btn adv-btn" data-adv-key="${key}" title="Advanced settings">Advanced</button>
+      </div>
     `;
     root.appendChild(card);
   }
-  for (const btn of $$('.card .btn')) {
+  for (const btn of root.querySelectorAll('[data-swap-key]')) {
-    btn.addEventListener('click', () => triggerSwap(btn.dataset.key));
+    btn.addEventListener('click', () => triggerSwap(btn.dataset.swapKey));
+  }
+  for (const btn of root.querySelectorAll('[data-adv-key]')) {
+    btn.addEventListener('click', () => openAdvanced(btn.dataset.advKey));
   }
 }
 
@@ -544,6 +552,8 @@ async function startDownload() {
     alert('Enter a HuggingFace repo in the form "org/name", e.g. RedHatAI/Qwen3.6-35B-A3B-NVFP4');
     return;
   }
+  dlState.last_repo = repo;
+  dlState.last_mode = mode;
   try {
     const r = await fetchJSON('/api/download', {
       method: 'POST',
@@ -623,12 +633,126 @@ function handleDownloadDone(d) {
     el('#dl-phase').textContent = 'Failed';
   } else {
     el('#dl-title').textContent = 'Done';
-    el('#dl-phase').textContent = 'Done ✓ — you can now add this model to the catalog and swap to it.';
+    el('#dl-phase').textContent = 'Done ✓';
     el('#dl-progress-fill').style.width = '100%';
+    // Offer to add to catalog
+    const repo = dlState.last_repo;
+    const mode = dlState.last_mode;
+    if (repo) {
+      setTimeout(() => openCatalogDialog(repo, mode), 600);
+    }
   }
   dlState.job_id = null;
 }
 
+// ===================== Advanced / Add to catalog =====================
+
+function openAdvanced(key) {
+  const m = state.models[key];
+  if (!m) return;
+  const dlg = el('#advanced-dialog');
+  el('#adv-title').textContent = `Advanced — ${m.display_name}`;
+  const k = m.effective_knobs || {};
+  el('#adv-mml').value = k.max_model_len ?? '';
+  el('#adv-gmu').value = k.gpu_memory_utilization ?? 0.85;
+  el('#adv-gmu-out').value = parseFloat(el('#adv-gmu').value).toFixed(2);
+  el('#adv-fst').checked = !!k.fastsafetensors;
+  el('#adv-pcache').checked = !!k.prefix_caching;
+  el('#adv-fp8').checked = k.kv_cache_dtype === 'fp8';
+  const del = el('#adv-delete');
+  del.classList.toggle('hidden', !m.custom);
+  del.onclick = async () => {
+    if (!confirm(`Delete "${m.display_name}" from the catalog? The model weights on disk are NOT deleted.`)) return;
+    try {
+      await fetchJSON(`/api/models/${encodeURIComponent(key)}`, { method: 'DELETE' });
+      dlg.close();
+      await loadModels();
+      pollStatus();
+    } catch (e) { alert('Delete failed: ' + e.message); }
+  };
+  const form = el('#advanced-form');
+  form.onsubmit = async (e) => {
+    e.preventDefault();
+    const knobs = {};
+    const mml = parseInt(el('#adv-mml').value, 10);
+    if (Number.isFinite(mml) && mml > 0) knobs.max_model_len = mml;
+    const gmu = parseFloat(el('#adv-gmu').value);
+    if (Number.isFinite(gmu)) knobs.gpu_memory_utilization = gmu;
+    if (el('#adv-fst').checked) knobs.fastsafetensors = true; else knobs.fastsafetensors = false;
+    if (el('#adv-pcache').checked) knobs.prefix_caching = true; else knobs.prefix_caching = false;
+    knobs.kv_cache_dtype = el('#adv-fp8').checked ? 'fp8' : 'auto';
+    try {
+      await fetchJSON(`/api/models/${encodeURIComponent(key)}/knobs`, {
+        method: 'PUT',
+        headers: { 'content-type': 'application/json' },
+        body: JSON.stringify({ knobs }),
+      });
+      dlg.close();
+      await loadModels();
+      pollStatus();
+    } catch (e) { alert('Save failed: ' + e.message); }
+  };
+  dlg.showModal();
+}
+
+function openCatalogDialog(repo, mode) {
+  const dlg = el('#catalog-dialog');
+  const key = repo.split('/').pop().toLowerCase().replace(/[^a-z0-9_-]/g, '-');
+  el('#cd-key').value = key;
+  el('#cd-name').value = repo.split('/').pop();
+  el('#cd-repo').value = repo;
+  el('#cd-size').value = '';
+  el('#cd-mode').value = mode || 'solo';
+  el('#cd-desc').value = '';
+  el('#cd-mml').value = 32768;
+  el('#cd-gmu').value = 0.85;
+  el('#cd-gmu-out').value = '0.85';
+  el('#cd-fst').checked = true;
+  el('#cd-pcache').checked = true;
+  el('#cd-fp8').checked = true;
+  dlg.showModal();
+}
+
+function setupCatalogDialog() {
+  el('#cd-cancel').addEventListener('click', () => el('#catalog-dialog').close());
+  el('#cd-gmu').addEventListener('input', (e) => { el('#cd-gmu-out').value = parseFloat(e.target.value).toFixed(2); });
+  el('#catalog-form').addEventListener('submit', async (e) => {
+    e.preventDefault();
+    const body = {
+      key: el('#cd-key').value.trim(),
+      display_name: el('#cd-name').value.trim(),
+      repo: el('#cd-repo').value.trim(),
+      size_gb: parseFloat(el('#cd-size').value) || 0,
+      mode: el('#cd-mode').value,
+      description: el('#cd-desc').value.trim() || null,
+      vllm_args: [],
+      knobs: {
+        max_model_len: parseInt(el('#cd-mml').value, 10) || 32768,
+        gpu_memory_utilization: parseFloat(el('#cd-gmu').value),
+        fastsafetensors: el('#cd-fst').checked,
+        prefix_caching: el('#cd-pcache').checked,
+        kv_cache_dtype: el('#cd-fp8').checked ? 'fp8' : 'auto',
+      },
+    };
+    try {
+      await fetchJSON('/api/models', {
+        method: 'POST',
+        headers: { 'content-type': 'application/json' },
+        body: JSON.stringify(body),
+      });
+      el('#catalog-dialog').close();
+      closeDownloadPanel();
+      await loadModels();
+      pollStatus();
+    } catch (e) { alert('Add to catalog failed: ' + e.message); }
+  });
+}
+
+function setupAdvancedDialog() {
+  el('#adv-cancel').addEventListener('click', () => el('#advanced-dialog').close());
+  el('#adv-gmu').addEventListener('input', (e) => { el('#adv-gmu-out').value = parseFloat(e.target.value).toFixed(2); });
+}
+
 // ===================== updates (spark-vllm-docker) =====================
 
 const updState = {
@@ -769,6 +893,8 @@ async function init() {
     list.open = !list.open;
   });
   el('#ub-apply').addEventListener('click', applyUpdate);
+  setupCatalogDialog();
+  setupAdvancedDialog();
   await loadModels();
   await pollStatus();
   await renderServices();

image/app/static/index.html

@@ -74,6 +74,54 @@
         <button id="open-download" class="btn small-btn">+ Download a new model</button>
       </div>
 
+      <dialog id="catalog-dialog" class="modal">
+        <form method="dialog" class="modal-form" id="catalog-form">
+          <h3>Add downloaded model to catalog</h3>
+          <p class="muted small">It will appear as a new card you can swap to. Knob values become its default launch flags — you can tweak later via the model's "Advanced" panel.</p>
+          <label class="modal-row"><span>Key (URL-safe id)</span><input type="text" id="cd-key" required pattern="[a-zA-Z0-9_-]+"></label>
+          <label class="modal-row"><span>Display name</span><input type="text" id="cd-name" required></label>
+          <label class="modal-row"><span>Repo (read-only)</span><input type="text" id="cd-repo" readonly></label>
+          <label class="modal-row"><span>Size (GB)</span><input type="number" id="cd-size" step="0.1" min="0"></label>
+          <label class="modal-row"><span>Mode</span>
+            <select id="cd-mode">
+              <option value="solo">solo (Spark 1 only)</option>
+              <option value="cluster">cluster (both Sparks via Ray)</option>
+            </select>
+          </label>
+          <label class="modal-row"><span>Description (optional)</span><textarea id="cd-desc" rows="3"></textarea></label>
+          <fieldset class="modal-fieldset">
+            <legend>Default launch knobs</legend>
+            <label class="modal-row"><span>Max context (tokens)</span><input type="number" id="cd-mml" step="1024" min="1024" value="32768"></label>
+            <label class="modal-row"><span>GPU memory %</span><input type="range" id="cd-gmu" min="0.5" max="0.95" step="0.01" value="0.85"> <output id="cd-gmu-out">0.85</output></label>
+            <label class="modal-row inline"><input type="checkbox" id="cd-fst" checked> Fast safetensors loading</label>
+            <label class="modal-row inline"><input type="checkbox" id="cd-pcache" checked> Prefix caching</label>
+            <label class="modal-row inline"><input type="checkbox" id="cd-fp8" checked> FP8 KV cache</label>
+          </fieldset>
+          <div class="modal-actions">
+            <button type="button" id="cd-cancel" class="btn">Cancel</button>
+            <button type="submit" class="btn primary">Add to catalog</button>
+          </div>
+        </form>
+      </dialog>
+
+      <dialog id="advanced-dialog" class="modal">
+        <form method="dialog" class="modal-form" id="advanced-form">
+          <h3 id="adv-title">Advanced settings</h3>
+          <p class="muted small">Custom values are stored in the package volume and survive package updates. Empty fields fall back to defaults.</p>
+          <label class="modal-row"><span>Max context (tokens)</span><input type="number" id="adv-mml" step="1024" min="1024"></label>
+          <label class="modal-row"><span>GPU memory %</span><input type="range" id="adv-gmu" min="0.5" max="0.95" step="0.01"> <output id="adv-gmu-out"></output></label>
+          <label class="modal-row inline"><input type="checkbox" id="adv-fst"> Fast safetensors loading <span class="muted small">(faster cold start)</span></label>
+          <label class="modal-row inline"><input type="checkbox" id="adv-pcache"> Prefix caching <span class="muted small">(speeds up repeated prefixes)</span></label>
+          <label class="modal-row inline"><input type="checkbox" id="adv-fp8"> FP8 KV cache <span class="muted small">(halves context memory)</span></label>
+          <div class="modal-actions">
+            <button type="button" id="adv-delete" class="btn danger hidden">Delete model</button>
+            <span class="spacer"></span>
+            <button type="button" id="adv-cancel" class="btn">Cancel</button>
+            <button type="submit" class="btn primary">Save</button>
+          </div>
+        </form>
+      </dialog>
+
       <section id="download-panel" class="download-panel hidden">
         <div class="download-form" id="download-form">
           <label class="dl-row">

image/app/static/style.css

@@ -217,6 +217,57 @@ main {
   word-break: break-word;
 }
 
+/* ===== Modal dialogs (Advanced / Add to catalog) ===== */
+
+.modal {
+  background: var(--surface);
+  color: var(--text);
+  border: 1px solid var(--border);
+  border-radius: var(--radius);
+  padding: 0;
+  max-width: 520px;
+  width: 92vw;
+}
+.modal::backdrop {
+  background: rgba(0, 0, 0, 0.6);
+  backdrop-filter: blur(2px);
+}
+.modal-form { padding: 22px 24px; display: flex; flex-direction: column; gap: 12px; }
+.modal-form h3 { margin: 0; font-size: 17px; }
+.modal-row {
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+  font-size: 13px;
+  color: var(--muted);
+}
+.modal-row.inline { flex-direction: row; align-items: center; gap: 8px; color: var(--text); font-size: 14px; }
+.modal-row > span { color: var(--muted); font-size: 12px; text-transform: uppercase; letter-spacing: 0.05em; }
+.modal-row input[type='text'],
+.modal-row input[type='number'],
+.modal-row textarea,
+.modal-row select {
+  background: var(--surface-2);
+  border: 1px solid var(--border);
+  color: var(--text);
+  padding: 7px 10px;
+  border-radius: 6px;
+  font: 13px ui-monospace, SFMono-Regular, "SF Mono", Menlo, monospace;
+}
+.modal-row textarea { font-family: inherit; resize: vertical; }
+.modal-row input:focus, .modal-row textarea:focus, .modal-row select:focus { outline: 1px solid var(--info); border-color: var(--info); }
+.modal-row input[type='range'] { padding: 0; flex: 1; }
+.modal-fieldset {
+  border: 1px solid var(--border);
+  border-radius: 6px;
+  padding: 12px 14px 4px;
+  display: flex;
+  flex-direction: column;
+  gap: 10px;
+}
+.modal-fieldset legend { color: var(--muted); font-size: 11px; text-transform: uppercase; letter-spacing: 0.05em; padding: 0 6px; }
+.modal-actions { display: flex; gap: 8px; justify-content: flex-end; margin-top: 8px; align-items: center; }
+
 /* ===== Update banner ===== */
 
 .update-banner {
@@ -436,7 +487,13 @@ main {
 .btn.primary { background: var(--accent); color: #052e16; border-color: var(--accent); }
 .btn.primary:hover:not(:disabled) { background: #6ee19a; }
 .btn:disabled { opacity: 0.45; cursor: not-allowed; }
+.btn.danger { color: var(--error); border-color: rgba(239, 68, 68, 0.3); }
+.btn.danger:hover:not(:disabled) { background: rgba(239, 68, 68, 0.08); border-color: var(--error); }
 .card.active .btn { background: rgba(74, 222, 128, 0.12); color: var(--accent); border-color: rgba(74, 222, 128, 0.4); }
+.card-actions { display: flex; gap: 6px; }
+.card-actions .btn.primary { flex: 1; }
+.card .adv-btn { padding: 8px 12px; font-size: 12px; }
+.card .custom-pill { color: var(--info); border-color: rgba(96, 165, 250, 0.4); }
 
 .footer {
   margin-top: 28px;

package/startos/main.ts

@@ -46,6 +46,7 @@ export const main = sdk.setupMain(async ({ effects }) => {
         MAGPIE_HOST: cfg.magpie_host,
         MAGPIE_USER: cfg.magpie_user,
         MAGPIE_CONTAINER: cfg.magpie_container,
+        MODELS_OVERRIDES: '/data/models-overrides.yaml',
         BIND_PORT: String(uiPort),
       },
     },

package/startos/versions/v0_1_0.ts

@@ -1,10 +1,10 @@
 import { VersionInfo, IMPOSSIBLE } from '@start9labs/start-sdk'
 
 export const v0_1_0 = VersionInfo.of({
-  version: '0.2.2:0',
+  version: '0.2.3:0',
   releaseNotes: {
     en_US:
-      'Update checking for spark-vllm-docker. Dashboard footer shows "N commits behind upstream" when applicable; click for the commit log, then "Apply update" runs git pull + ./build-and-copy.sh -c on Spark 1 with a streamed log and elapsed timer. No auto-apply — you confirm each update.',
+      'Per-model Advanced settings + downloaded-model catalog flow. Each card now has an Advanced button: max context tokens, GPU memory %, and optimization toggles (fastsafetensors, prefix caching, FP8 KV cache). After a download finishes, a dialog appears to add the model to the catalog with those same knobs as launch defaults. Custom models can be deleted. Overrides persist in /data/models-overrides.yaml and survive package updates.',
   },
   migrations: {
     up: async ({ effects }) => {},