feat: echter Download-Fortschritt via Datei-Polling

`hf` gibt im Nicht-TTY-Modus keinen Fortschritt aus (am Bosgame verifiziert:
0 CR-Frames). Stattdessen pollt jobengine.attach_download_progress die
wachsende <local-dir>/.cache/huggingface/download/*.incomplete-Datei gegen die
Gesamtgroesse aus der HF-Tree-API (cookbook.hf_file_size) -> exaktes %.

- attach_download_progress an /api/download, install-recipe, install-model
- Frontend (Aktivitaet + Server-Karte): nutzt job.progress bevorzugt,
  Log-%-Parsing bleibt Fallback fuer Tools, die selbst Prozente ausgeben

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

jobengine.py

@@ -6,6 +6,7 @@ zeilenweisem Log-Capture. Keine Persistenz, kein Broker. Genutzt von allen
 Routern, die laenger laufende Shell-Befehle anstossen (Download, Update, ...).
 """
 
+import glob
 import os
 import shlex
 import subprocess
@@ -102,6 +103,38 @@ def _run_job(job_id: str, args: list[str], env: dict | None = None, stdin_data:
     job["finished_at"] = time.time()
 
 
+def attach_download_progress(job_id: str, local_dir: str, total_bytes: int) -> None:
+    """Echten Download-Fortschritt (in %) auf den Job legen. `hf` gibt im Nicht-TTY-
+    Modus keinen Fortschritt aus, schreibt aber in <local_dir>/.cache/huggingface/
+    download/*.incomplete (waechst). Wir vergleichen dessen Groesse mit total_bytes
+    (aus der HF-Tree-API). Ein Daemon-Thread aktualisiert job["progress"]."""
+    if not total_bytes or total_bytes <= 0:
+        return
+    job = JOBS.get(job_id)
+    if job is not None:
+        job["progress"] = 0
+
+    def _watch():
+        pat = os.path.join(local_dir, ".cache", "huggingface", "download", "*.incomplete")
+        while True:
+            j = JOBS.get(job_id)
+            if not j or j["state"] in ("done", "failed", "canceled"):
+                break
+            try:
+                inc = glob.glob(pat)
+                cur = sum(os.path.getsize(f) for f in inc) if inc else 0
+                if cur:
+                    j["progress"] = min(99, int(cur * 100 / total_bytes))
+            except Exception:  # noqa: BLE001
+                pass
+            time.sleep(1.0)
+        j = JOBS.get(job_id)
+        if j and j["state"] == "done":
+            j["progress"] = 100
+
+    threading.Thread(target=_watch, daemon=True).start()
+
+
 def cancel_job(job_id: str) -> bool:
     """Laufenden Job abbrechen: Prozess terminieren. Liefert False, wenn der Job
     nicht (mehr) laeuft oder unbekannt ist."""

routers/cookbook.py

@@ -14,7 +14,7 @@ from auth import auth
 from hw_math import evaluate_fit, max_ctx_for
 from config import MODELS_DIR, CMD_TEMPLATE, DEFAULT_TTL, HF_DOWNLOAD_ENV, hf_bin
 from llamaswap import read_config, write_config
-from jobengine import start_job, JOBS
+from jobengine import start_job, JOBS, attach_download_progress
 from recipes import RECIPES, UPGRADES
 
 router = APIRouter(prefix="/api/cookbook", dependencies=[Depends(auth)])
@@ -163,6 +163,7 @@ def install_recipe(req: InstallRecipeReq):
         args = [hf_bin(), "download", m["repo"], file, "--local-dir", str(target)]
         jid = start_job(args, f"download {m['name']}", env=env)
         JOBS[jid]["result_path"] = str(target / file)
+        attach_download_progress(jid, str(target), hf_file_size(m["repo"], 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)
@@ -173,6 +174,20 @@ def install_recipe(req: InstallRecipeReq):
     return {"job_ids": job_ids, "count": len(job_ids)}
 
 
+def hf_file_size(repo: str, file: str) -> int:
+    """Groesse einer Datei im HF-Repo (Bytes) fuer die Fortschrittsanzeige. 0 wenn unbekannt.
+    GGUFs sind LFS -> ggf. unter 'lfs.size'."""
+    try:
+        with httpx.Client(timeout=10.0) as c:
+            tree = c.get(f"https://huggingface.co/api/models/{repo}/tree/main").json()
+        for f in tree:
+            if isinstance(f, dict) and f.get("path") == file:
+                return int(f.get("size") or (f.get("lfs") or {}).get("size") or 0)
+    except Exception:  # noqa: BLE001
+        return 0
+    return 0
+
+
 def _pick_gguf(repo: str, quant: str = "Q4_K_M") -> str | None:
     """Beste GGUF-Datei eines Repos auflösen: bevorzugt gewünschten Quant, keine Split-Teile."""
     try:
@@ -234,6 +249,7 @@ def install_model(req: InstallModelReq):
     jid = start_job([hf_bin(), "download", req.repo, file, "--local-dir", str(target)],
                     f"download {req.repo.split('/')[-1]}", env=env)
     JOBS[jid]["result_path"] = str(target / file)
+    attach_download_progress(jid, str(target), hf_file_size(req.repo, file))
     cfg = read_config()
     ctx = min(max_ctx_for(req.params_b, req.quant, ram_gb), 32768)
     cmd = CMD_TEMPLATE.replace("{model}", str(target / file)).replace("{ctx}", str(ctx))

routers/models.py

@@ -15,7 +15,8 @@ from ruamel.yaml.scalarstring import LiteralScalarString
 from auth import auth
 from config import (CMD_TEMPLATE, CONFIG_PATH, DEFAULT_TTL, HF_DOWNLOAD_ENV, LLAMA_SWAP_URL,
                    MODELS_DIR, TOKEN, hf_bin)
-from jobengine import JOBS, start_job
+from jobengine import JOBS, start_job, attach_download_progress
+from routers.cookbook import hf_file_size
 from llamaswap import _swap_get, read_config, write_config
 from hw_math import extract_params_b, max_ctx_for, estimate_memory_gb
 import re
@@ -146,6 +147,7 @@ def download(req: DownloadReq):
         env["HF_TOKEN"] = req.hf_token
     job_id = start_job(args, f"download {req.repo}/{req.file}", env=env)
     JOBS[job_id]["result_path"] = str(target / req.file)
+    attach_download_progress(job_id, str(target), hf_file_size(req.repo, req.file))
     return {"job_id": job_id, "expected_path": str(target / req.file)}
 
 

static/js/panels/jobs.js

@@ -16,9 +16,11 @@ const MAX_HIST = 60;
 function statusBadge(s) { return s === "done" ? '<span class="badge b-run">fertig</span>' : s === "failed" ? '<span class="badge b-err">fehler</span>' : s === "canceled" ? '<span class="badge">abgebrochen</span>' : '<span class="badge b-load">läuft…</span>'; }
 function dotClass(s) { return s === "done" ? "on" : (s === "failed" || s === "canceled") ? "" : "load"; }
 
-// Fortschritt in % aus der letzten Log-Zeile ziehen (tqdm schreibt z.B. " 45%|…").
+// Fortschritt in %: bevorzugt der exakte Wert vom Backend (Download-Datei-Polling),
+// sonst Fallback auf eine %-Angabe in der letzten Log-Zeile.
 function jobPct(j) {
   if (j.state !== "running") return null;
+  if (typeof j.progress === "number") return Math.min(100, j.progress);
   const last = (j.log || [])[j.log.length - 1] || "";
   const m = last.match(/(\d+(?:\.\d+)?)\s*%/);
   return m ? Math.min(100, parseFloat(m[1])) : null;

static/js/panels/server.js

@@ -141,12 +141,20 @@ function onJobs(jobs) {
     : canceled ? '<span class="badge">abgebrochen</span>'
     : '<span class="badge b-load">läuft…</span>';
 
-  // Fortschrittsbalken aus der letzten Log-Zeile (tqdm: " 45%|…")
+  // Fortschrittsbalken: exakter Backend-Wert (Download) bevorzugt, sonst %-Angabe im Log.
+  let pval = null;
+  if (!terminal) {
+    if (typeof j.progress === "number") {
+      pval = j.progress;
+    } else {
       const last = (j.log || [])[j.log.length - 1] || "";
-  const pm = !terminal && last.match(/(\d+(?:\.\d+)?)\s*%/);
+      const pm = last.match(/(\d+(?:\.\d+)?)\s*%/);
+      if (pm) pval = parseFloat(pm[1]);
+    }
+  }
   const bar = $("#w-update-bar");
   if (bar) {
-    if (pm) { bar.style.display = ""; bar.querySelector("i").style.width = Math.min(100, parseFloat(pm[1])) + "%"; }
+    if (pval != null) { bar.style.display = ""; bar.querySelector("i").style.width = Math.min(100, pval) + "%"; }
     else bar.style.display = "none";
   }
   $("#w-update-cancel").style.display = terminal ? "none" : "";