zero-gpu: prefetch the checkpoint at startup, not inside the GPU slice

BUILD.json

@@ -1,6 +1,6 @@
 {
-  "built_at": "2026-10-29T16:48:34",
-  "commit": "8f3c1ee",
+  "built_at": "2026-10-29T16:58:09",
+  "commit": "ef0b6ab",
   "files": 53,
   "with_store": false,
   "with_card": true

traceforge/space/app.py

@@ -483,6 +483,23 @@ def build_app(store: Optional[TraceStore] = None,
     return app
 
 
+def _prefetch_in_background() -> None:
+    """Pull the checkpoint at startup, not inside a visitor's GPU slice.
+
+    The 24.8 GB FP8 download takes ~45 s at the ~600 MB/s the Hub gives us.
+    Doing it lazily meant the first GPU call of every container spent most of
+    its 120 s slice downloading, so the weights are fetched here instead.
+    """
+    try:
+        from traceforge.space.pipeline import _local_mode
+        if not _local_mode():
+            return
+        from traceforge.space.local_llm import prefetch
+        prefetch()
+    except Exception:
+        pass
+
+
 def main(argv: Optional[list[str]] = None) -> int:
     import argparse
 
@@ -498,6 +515,7 @@ def main(argv: Optional[list[str]] = None) -> int:
 
     app = build_app()
     app.queue(default_concurrency_limit=4)
+    _prefetch_in_background()
     app.launch(server_name=args.host, server_port=args.port,
                share=args.share, inbrowser=False, show_error=True,
                css=CSS, theme=gr.themes.Base())

traceforge/space/local_llm.py

@@ -678,20 +678,32 @@ def _volume_read_mbps(mb: int = 256) -> dict:
 
 def _ram_gb() -> dict:
     """Container memory, so we can tell whether a module-level load fits."""
+    out = {}
     try:
         import psutil
         vm = psutil.virtual_memory()
-        return {"total": round(vm.total / 1e9, 1),
-                "available": round(vm.available / 1e9, 1)}
+        out = {"total": round(vm.total / 1e9, 1),
+               "available": round(vm.available / 1e9, 1)}
     except Exception:
         try:
             with open("/proc/meminfo") as fh:
                 for line in fh:
                     if line.startswith("MemTotal:"):
-                        return {"total": round(int(line.split()[1]) / 1e6, 1)}
+                        out = {"total": round(int(line.split()[1]) / 1e6, 1)}
         except Exception:
             pass
-    return {}
+    # psutil reports the host's RAM; the cgroup limit is what we can use.
+    for path in ("/sys/fs/cgroup/memory.max",
+                 "/sys/fs/cgroup/memory/memory.limit_in_bytes"):
+        try:
+            with open(path) as fh:
+                raw = fh.read().strip()
+            if raw.isdigit() and int(raw) < (1 << 62):
+                out["cgroup"] = round(int(raw) / 1e9, 1)
+                break
+        except Exception:
+            continue
+    return out
 
 
 def health() -> dict: