Update app.py

app.py

@@ -6,16 +6,18 @@ from huggingface_hub import hf_hub_download
 import os
 
 VARIANTS = ["model_1000000"]
-models = {}
-for variant in VARIANTS:
-    models[variant] = F5TTS(
-        ckpt_file=hf_hub_download("mrfakename/openf5-v2", f"{variant}.pt", token=os.getenv("HF_TOKEN")),
-        vocab_file=hf_hub_download("mrfakename/openf5-v2", "vocab.txt", token=os.getenv("HF_TOKEN"))
-    )
+models = {}  # Global inside GPU process
 
 @spaces.GPU
-def generate(text, ref_audio, variant, progress = gr.Progress()):
+def generate(text, ref_audio, variant, progress=gr.Progress()):
+    if variant not in models:
+        print(f"⏬ Loading model for variant: {variant}")
+        ckpt_file = hf_hub_download("mrfakename/openf5-v2", f"{variant}.pt", token=os.getenv("HF_TOKEN"))
+        vocab_file = hf_hub_download("mrfakename/openf5-v2", "vocab.txt", token=os.getenv("HF_TOKEN"))
+        models[variant] = F5TTS(ckpt_file=ckpt_file, vocab_file=vocab_file)
+
     api = models[variant]
+
     with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
         api.infer(
             ref_file=ref_audio,