lazy loading so server startup is not delayed

.gitignore

@@ -2,5 +2,7 @@ __pycache__/
 *.pyc
 pretrained/
 AnyAccomp-repo/
+.venv-test/
+test_gradio_health.py
 DEPLOY.md
 instructions.md

app.py

@@ -3,6 +3,7 @@ sys.stdout.reconfigure(line_buffering=True)
 
 import torch
 import os
+import threading
 # import spaces  # uncomment when running on ZeroGPU
 import librosa
 import soundfile as sf
@@ -14,20 +15,7 @@ from anyaccomp.inference_utils import Sing2SongInferencePipeline
 
 repo_id = "amphion/anyaccomp"
 base_dir = os.path.dirname(os.path.abspath(__file__))
-
-checkpoint_marker = os.path.join(base_dir, "pretrained", "flow_matching")
-if not os.path.exists(checkpoint_marker):
-    print(f"Downloading model files from {repo_id}...", flush=True)
-    model_dir = snapshot_download(repo_id=repo_id, local_dir=base_dir)
-    print(f"Model files downloaded to: {model_dir}", flush=True)
-else:
-    model_dir = base_dir
-    print(f"Model files already present, skipping download.", flush=True)
-
-CFG_PATH = os.path.join(base_dir, "config/flow_matching.json")
-VOCODER_CFG_PATH = os.path.join(base_dir, "config/vocoder.json")
-CHECKPOINT_PATH = os.path.join(model_dir, "pretrained/flow_matching")
-VOCODER_CHECKPOINT_PATH = os.path.join(model_dir, "pretrained/vocoder")
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 
 INFER_DST = os.path.join(base_dir, "output_gradio")
 os.makedirs(INFER_DST, exist_ok=True)
@@ -36,28 +24,55 @@ mixture_dst = os.path.join(INFER_DST, "mixture")
 os.makedirs(acc_dst, exist_ok=True)
 os.makedirs(mixture_dst, exist_ok=True)
 
-DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+# Model loading state — populated by the background thread.
+inference_pipeline = None
+model_loading = True
+model_error = None
 
-print("Initializing AnyAccomp InferencePipeline...")
-try:
-    inference_pipeline = Sing2SongInferencePipeline(
-        CHECKPOINT_PATH,
-        CFG_PATH,
-        VOCODER_CHECKPOINT_PATH,
-        VOCODER_CFG_PATH,
-        device=DEVICE,
-    )
-    inference_pipeline.sample_rate = 24000
-    print("Model loaded successfully.")
-except Exception as e:
-    print(f"Error loading model: {e}")
-    inference_pipeline = None
+
+def load_model():
+    global inference_pipeline, model_loading, model_error
+    try:
+        checkpoint_marker = os.path.join(base_dir, "pretrained", "flow_matching")
+        if not os.path.exists(checkpoint_marker):
+            print(f"Downloading model files from {repo_id}...", flush=True)
+            model_dir = snapshot_download(repo_id=repo_id, local_dir=base_dir)
+            print(f"Model files downloaded to: {model_dir}", flush=True)
+        else:
+            model_dir = base_dir
+            print("Model files already present, skipping download.", flush=True)
+
+        cfg_path = os.path.join(base_dir, "config/flow_matching.json")
+        vocoder_cfg_path = os.path.join(base_dir, "config/vocoder.json")
+        checkpoint_path = os.path.join(model_dir, "pretrained/flow_matching")
+        vocoder_checkpoint_path = os.path.join(model_dir, "pretrained/vocoder")
+
+        print("Initializing AnyAccomp InferencePipeline...", flush=True)
+        pipeline = Sing2SongInferencePipeline(
+            checkpoint_path, cfg_path,
+            vocoder_checkpoint_path, vocoder_cfg_path,
+            device=DEVICE,
+        )
+        pipeline.sample_rate = 24000
+        inference_pipeline = pipeline
+        print("Model loaded successfully.", flush=True)
+    except Exception as e:
+        model_error = str(e)
+        print(f"Error loading model: {e}", flush=True)
+    finally:
+        model_loading = False
+
+
+# Start model loading in the background so the server can start immediately.
+threading.Thread(target=load_model, daemon=True).start()
 
 
 # @spaces.GPU  # uncomment when running on ZeroGPU
 def sing2song_inference(vocal_filepath, n_timesteps, cfg_scale):
+    if model_loading:
+        raise gr.Error("Model is still loading, please wait a moment and try again.")
     if inference_pipeline is None:
-        raise gr.Error("Model could not be loaded. Please check paths and environment configuration.")
+        raise gr.Error(f"Model failed to load: {model_error}")
 
     if vocal_filepath is None:
         raise gr.Error("Please upload a vocal audio file.")