dockerfile edit

app.py

@@ -23,12 +23,22 @@ import joblib
 import torch
 import gradio as gr
 import spaces
+from huggingface_hub import snapshot_download
 
 warnings.filterwarnings("ignore")
 
 # ── Paths ─────────────────────────────────────────────────────────────────────
 BASE_DIR = Path(__file__).parent
 
+# ── ProtT5 prefetch ────────────────────────────────────────────────────────────
+# This SDK-gradio Space builds from requirements.txt, not the repo's Dockerfile,
+# so there is no build-time prefetch step. Download the weights here instead,
+# at module import (app startup, plain CPU context) — before any @spaces.GPU
+# call — so extract_t5_feature() never blocks on a network download while
+# holding a ZeroGPU allocation (which has a short time budget).
+T5_REPO_ID = "Rostlab/prot_t5_xl_half_uniref50-enc"
+snapshot_download(T5_REPO_ID)
+
 # ── Physical constants (from preprocess/misc.py) ──────────────────────────────
 MAX_TEMP   = 60.0
 MAX_CONC   = 1000.0
@@ -114,11 +124,11 @@ def load_t5():
         _modeling_utils.check_torch_load_is_safe = _noop
 
         try:
-            # Pre-fetched into the HF cache at Docker build time (see Dockerfile's
-            # snapshot_download step), so ZeroGPU calls never hit the network here —
-            # a cold download would blow past the GPU-allocation time budget and
-            # get aborted. local_files_only=True enforces that guarantee.
-            repo_id = "Rostlab/prot_t5_xl_half_uniref50-enc"
+            # Already fetched by the module-level snapshot_download() above, so
+            # ZeroGPU calls never hit the network here — a cold download would
+            # blow past the GPU-allocation time budget and get aborted.
+            # local_files_only=True enforces that guarantee.
+            repo_id = T5_REPO_ID
             dev = get_device()
             dtype = torch.float16 if dev == "cuda" else torch.float32