fix meta tensor crash: force low_cpu_mem_usage=False and float32 for CPU

app.py

@@ -339,6 +339,15 @@ try:
     torch.backends.cuda.enable_flash_sdp(False)
     os.environ["ATTN_BACKEND"] = "sdpa"
 
+    import transformers
+    _orig_from_pretrained = transformers.AutoModel.from_pretrained
+    def _cpu_from_pretrained(*args, **kwargs):
+        kwargs['low_cpu_mem_usage'] = False
+        kwargs.setdefault('torch_dtype', torch.float32)
+        return _orig_from_pretrained(*args, **kwargs)
+    transformers.AutoModel.from_pretrained = _cpu_from_pretrained
+    log("  Patched AutoModel.from_pretrained: low_cpu_mem_usage=False, dtype=float32")
+
     import torchaudio
     _orig = torchaudio.load
     def _sf(p, *a, **kw):