v4 Schritt 2: Cookbook 2.0 — Use-Case-Setups (Stacks)

- recipes.py: 6 kuratierte Use-Case-Stacks (Coding, Vision, Chat, LongDoc, Agents, Fast).
- cookbook.py: /recipes (Stack-Fit pro Modell + Gesamturteil) + /install-recipe
  (download + register aller Modelle mit optimalem, gedeckeltem Kontext).
- cookbook.js: Use-Case-Karten als Haupteinstieg + Setup-Modal + 'Komplettes Setup
  installieren'; Roh-Suche in 'Profi-Modus' (collapsible).
- models.py /register: exists()-Check entfernt (fixt frischen Download-Install).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

routers/cookbook.py

@@ -8,11 +8,19 @@ from fastapi import APIRouter, Depends, HTTPException
 from pydantic import BaseModel
 import psutil
 
+from ruamel.yaml.scalarstring import LiteralScalarString
+
 from auth import auth
 from hw_math import evaluate_fit, max_ctx_for
+from config import MODELS_DIR, CMD_TEMPLATE, DEFAULT_TTL
+from llamaswap import read_config, write_config
+from jobengine import start_job, JOBS
+from recipes import RECIPES
 
 router = APIRouter(prefix="/api/cookbook", dependencies=[Depends(auth)])
 
+_FIT_ORDER = {"perfect": 0, "marginal": 1, "too_tight": 2}
+
 class AnalyzeRequest(BaseModel):
     repo_id: str
     ctx: int = 8192
@@ -22,6 +30,9 @@ class EvaluateRequest(BaseModel):
     quant: str
     ctx: int
 
+class InstallRecipeReq(BaseModel):
+    recipe_id: str
+
 def extract_params_b(repo_id: str) -> float:
     """Extrahiert die Parametergröße (in Milliarden) aus dem Repo-Namen."""
     # z.B. Qwen2.5-Coder-32B -> 32
@@ -100,3 +111,47 @@ def evaluate_single(req: EvaluateRequest):
     fit["optimal_ctx"] = max_ctx_for(req.params_b, req.quant, ram_gb)
     return fit
 
+
+@router.get("/recipes")
+def recipes():
+    """Use-Case-Setups mit Hardware-Fit pro Modell + Stack-Gesamturteil. Da llama-swap nur EIN
+    Modell gleichzeitig lädt, ist das Stack-Urteil der schlechteste (= größte) Einzel-Fit."""
+    ram_gb = psutil.virtual_memory().total / (1024 ** 3)
+    out = []
+    for r in RECIPES:
+        models, worst = [], "perfect"
+        for m in r["models"]:
+            fit = evaluate_fit(m["params_b"], m["quant"], 8192, ram_gb)
+            models.append({**m, "fit": fit, "optimal_ctx": max_ctx_for(m["params_b"], m["quant"], ram_gb)})
+            if _FIT_ORDER[fit["level"]] > _FIT_ORDER[worst]:
+                worst = fit["level"]
+        out.append({**r, "models": models, "fit_level": worst})
+    return {"recipes": out, "sys_ram_gb": round(ram_gb, 1)}
+
+
+@router.post("/install-recipe")
+def install_recipe(req: InstallRecipeReq):
+    """Komplettes Setup installieren: jedes Modell als Download-Job starten UND sofort mit
+    optimalem (gedeckeltem) Kontext in die config.yaml eintragen. llama-swap (-watch-config)
+    übernimmt es, sobald die Datei da ist."""
+    recipe = next((r for r in RECIPES if r["id"] == req.recipe_id), None)
+    if not recipe:
+        raise HTTPException(404, "Setup nicht gefunden.")
+    ram_gb = psutil.virtual_memory().total / (1024 ** 3)
+    cfg = read_config()
+    job_ids = []
+    for m in recipe["models"]:
+        target = MODELS_DIR / m["repo"].split("/")[-1]
+        target.mkdir(parents=True, exist_ok=True)
+        args = ["hf", "download", m["repo"], m["file"], "--local-dir", str(target)]
+        jid = start_job(args, f"download {m['name']}", env={"HF_XET_HIGH_PERFORMANCE": "1"})
+        JOBS[jid]["result_path"] = str(target / m["file"])
+        job_ids.append(jid)
+        # Eintrag jetzt schon schreiben — optimaler Kontext, aber gedeckelt fuer schnellen Erststart.
+        ctx = min(max_ctx_for(m["params_b"], m["quant"], ram_gb), 32768)
+        path = str(target / m["file"])
+        cmd = CMD_TEMPLATE.replace("{model}", path).replace("{ctx}", str(ctx))
+        cfg["models"][m["role"]] = {"cmd": LiteralScalarString(cmd + "\n"), "ttl": DEFAULT_TTL}
+    write_config(cfg)
+    return {"job_ids": job_ids, "count": len(recipe["models"])}
+