Upgrade transformers at runtime for OmniVoice HiggsAudioV2TokenizerModel support

handler.py

@@ -12,7 +12,13 @@ logger = logging.getLogger(__name__)
 
 def _install_deps():
     """Install OmniVoice and voicetut-tts with --no-deps to avoid conflicts
-    with the HF Inference Endpoint base image (which pins transformers<5)."""
+    with the HF Inference Endpoint base image.
+
+    Then upgrade transformers (WITH deps) because OmniVoice needs
+    HiggsAudioV2TokenizerModel which only exists in transformers>=4.50.
+    The HF Inference Toolkit already loaded successfully with the old
+    transformers at this point, so upgrading is safe."""
+    # 1. Install OmniVoice and voicetut-tts without their dep trees
     pkgs = [
         "git+https://github.com/k2-fsa/OmniVoice.git",
         "voicetut-tts>=0.1.0",
@@ -27,10 +33,22 @@ def _install_deps():
             logger.info(f"Installed {pkg}")
         except subprocess.CalledProcessError as e:
             stderr = e.stderr.decode() if e.stderr else ""
-            # If already installed, that's fine
             if "already satisfied" not in stderr.lower():
                 logger.warning(f"Failed to install {pkg}: {stderr[:200]}")
 
+    # 2. Upgrade transformers to get HiggsAudioV2TokenizerModel (needed by OmniVoice)
+    try:
+        subprocess.check_call(
+            [sys.executable, "-m", "pip", "install", "--upgrade",
+             "transformers>=4.50.0", "tokenizers>=0.21.0"],
+            stdout=subprocess.DEVNULL,
+            stderr=subprocess.PIPE,
+        )
+        logger.info("Upgraded transformers for OmniVoice compatibility")
+    except subprocess.CalledProcessError as e:
+        stderr = e.stderr.decode() if e.stderr else ""
+        logger.warning(f"Failed to upgrade transformers: {stderr[:200]}")
+
 
 # Install deps before importing voicetut_tts
 _install_deps()