Fix spec-parser cold-start hang: pre-cache weights at boot

spec_brick.py

@@ -100,7 +100,7 @@ if _should_load():
             _state["tok"] = tok
             _state["model"] = model.to("cuda").eval()
 
-        @spaces.GPU(duration=120)   # cold path = 3.4GB download + load + generate
+        @spaces.GPU(duration=180)   # warm path: load from pre-cached weights + generate
         def _generate(text: str) -> str:
             if _state["model"] is None:
                 _load()
@@ -121,6 +121,21 @@ if _should_load():
 
         _GENERATE = _generate
 
+        # Pre-cache the weights at BOOT, on CPU, with no time limit, so the first
+        # /parse call loads from a warm cache and finishes inside its GPU window.
+        # Without this a cold cache (e.g. after a redeploy) makes the first call
+        # try to pull ~3.4GB during the GPU slot, which overruns it and hangs
+        # every request. snapshot_download is a no-op once the files are cached.
+        try:
+            from huggingface_hub import snapshot_download
+            snapshot_download(BASE_ID, revision=BASE_REVISION)
+            snapshot_download(ADAPTER_ID, revision=ADAPTER_REVISION)
+            print("[FitCheck] spec parser weights pre-cached at boot",
+                  file=sys.stderr, flush=True)
+        except Exception as pe:  # noqa: BLE001
+            print(f"[FitCheck] spec parser pre-download skipped "
+                  f"({type(pe).__name__}: {pe})", file=sys.stderr, flush=True)
+
         # Eager boot path (flagged): resident-load now only if cached; never
         # downloads at boot (local_files_only=True), so a cold cache is harmless.
         if EAGER_LOAD: