Update app.py

app.py

@@ -1,10 +1,14 @@
 # app.py
 import os, gradio as gr, spaces
+
+# ---- deactivate deepspeed CUDA build BEFORE anything else imports it
+os.environ["DS_ACCELERATOR"] = "cpu"   # force deepspeed to CPU backend
+os.environ["DS_BUILD_OPS"]   = "0"     # skip compiling custom ops
+# ------------------------------------
+
 from huggingface_hub import snapshot_download
 from indextts.infer import IndexTTS
 
-os.environ["DS_BUILD_OPS"] = "0"      # block any JIT build if deepspeed sneaks in
-
 model_dir = snapshot_download("IndexTeam/IndexTTS-1.5",
                               local_dir="checkpoints",
                               local_dir_use_symlinks=False)
@@ -14,10 +18,7 @@ tts = None
 def load():
     global tts
     if tts is None:
-        # the key bit → force plain pytorch
-        tts = IndexTTS(model_dir=model_dir,
-                       cfg_path=cfg_path,
-                       use_deepspeed=False)   # <- param exists in IndexTTS 1.5
+        tts = IndexTTS(model_dir=model_dir, cfg_path=cfg_path)  # now imports deepspeed safely
 
 @spaces.GPU
 def synth(ref_wav, prompt):
@@ -27,7 +28,7 @@ def synth(ref_wav, prompt):
     return out
 
 with gr.Blocks() as demo:
-    gr.Markdown("# index-tts 1.5 zerogpu")
+    gr.Markdown("# index-tts 1.5 – zero-gpu space")
     txt = gr.Textbox(label="text prompt")
     ref = gr.Audio(label="reference voice", type="filepath")
     gen = gr.Audio(label="generated speech", type="filepath")