Fix: use direct WhisperProcessor/WhisperForConditionalGeneration imports

Auto classes (AutoProcessor, AutoModelForSpeechSeq2Seq) are not reliably
exported from transformers __init__.py across all 4.x/5.x versions.
Direct module paths (transformers.models.whisper) are stable in all versions.
Also allow _ensure_whisper_loaded to retry after an import error.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -82,35 +82,30 @@ if HF_TOKEN:
 def _do_load_whisper():
     global _whisper_model, _whisper_processor, _adapter_manager, _model_status
     import torch
-    from transformers import AutoProcessor
-    # AutoModelForSpeechSeq2Seq is not always exported at the transformers top level
-    # depending on the version; fall back through known-good alternatives.
+    from src.engine.adapter_manager import AdapterManager
+
+    # Import concrete Whisper classes directly — bypasses transformers __init__.py
+    # Auto-class exports differ between transformers 4.x and 5.x; direct paths are stable.
     try:
-        from transformers import AutoModelForSpeechSeq2Seq as _ModelCls
+        from transformers.models.whisper import WhisperProcessor, WhisperForConditionalGeneration
     except ImportError:
-        try:
-            from transformers import AutoModelForConditionalGeneration as _ModelCls
-        except ImportError:
-            from transformers.models.whisper.modeling_whisper import WhisperForConditionalGeneration as _ModelCls
-    from src.engine.adapter_manager import AdapterManager
+        from transformers.models.whisper.processing_whisper import WhisperProcessor
+        from transformers.models.whisper.modeling_whisper import WhisperForConditionalGeneration
 
     _model_status = "loading…"
     try:
-        _whisper_processor = AutoProcessor.from_pretrained(
+        _whisper_processor = WhisperProcessor.from_pretrained(
             WHISPER_MODEL_ID, token=HF_TOKEN
         )
-        # `dtype` is the transformers 5.x name; `torch_dtype` works on 4.x.
-        # Try the new name first, fall back if the version doesn't accept it.
         try:
-            _whisper_model = _ModelCls.from_pretrained(
+            _whisper_model = WhisperForConditionalGeneration.from_pretrained(
                 WHISPER_MODEL_ID,
-                dtype=torch.float32,
+                torch_dtype=torch.float32,
                 token=HF_TOKEN,
             )
         except TypeError:
-            _whisper_model = _ModelCls.from_pretrained(
+            _whisper_model = WhisperForConditionalGeneration.from_pretrained(
                 WHISPER_MODEL_ID,
-                torch_dtype=torch.float32,
                 token=HF_TOKEN,
             )
         _whisper_model.eval()
@@ -160,7 +155,9 @@ def _ensure_whisper_loaded():
     """Load Whisper to CPU in a background thread on first call. Non-blocking."""
     global _model_status
     with _model_lock:
-        if _whisper_model is None and "loading" not in _model_status and "error" not in _model_status:
+        # Retry if previous attempt errored (e.g. import failed on first try)
+        if _whisper_model is None and "loading" not in _model_status:
+            _model_status = "loading…"
             t = threading.Thread(target=_do_load_whisper, daemon=True)
             t.start()
     return _model_status