fix: pre-download cvssp/audioldm2 at startup to avoid GPU budget drain

app.py

@@ -80,6 +80,14 @@ print("Pre-downloading MMAudio CLIP model (apple/DFN5B-CLIP-ViT-H-14-384)…")
 snapshot_download(repo_id="apple/DFN5B-CLIP-ViT-H-14-384")
 print("MMAudio CLIP model pre-downloaded.")
 
+# Pre-download TARO's AudioLDM2 VAE + vocoder (cvssp/audioldm2).
+# AutoencoderKL.from_pretrained() and SpeechT5HifiGan.from_pretrained() fetch
+# this repo inside the GPU window on every cold worker start, burning GPU budget
+# before inference even begins.  Pre-fetching here ensures the cache is warm.
+print("Pre-downloading AudioLDM2 (cvssp/audioldm2)…")
+snapshot_download(repo_id="cvssp/audioldm2")
+print("AudioLDM2 pre-downloaded.")
+
 # ================================================================== #
 #                     SHARED CONSTANTS / HELPERS                      #
 # ================================================================== #