fix(qwen): preserve validated runtime on checkpoint reload

app.py

@@ -78,7 +78,7 @@ from obliteratus.model_load_settings import (
     resolve_model_load_settings,
 )
 from obliteratus.persistence_contracts import validate_reloadable_checkpoint
-from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
 
 _gpu_lifecycle = from_environment()
 
@@ -156,7 +156,33 @@ def _load_model_to_device(
     if local_files_only:
         kwargs["local_files_only"] = True
 
-    if dev.supports_device_map_auto():
+    config_kwargs = {
+        key: value
+        for key, value in {
+            "trust_remote_code": trust_remote_code or None,
+            "token": token,
+            "local_files_only": local_files_only or None,
+        }.items()
+        if value is not None
+    }
+    config = AutoConfig.from_pretrained(pretrained_path, **config_kwargs)
+    from obliteratus.models.loader import qwen_hybrid_runtime_overrides
+
+    quantization = None
+    if quantization_config is not None:
+        if getattr(quantization_config, "load_in_4bit", False):
+            quantization = "4bit"
+        elif getattr(quantization_config, "load_in_8bit", False):
+            quantization = "8bit"
+    qwen_overrides = qwen_hybrid_runtime_overrides(
+        config,
+        torch_dtype or torch.float32,
+        quantization,
+    )
+    if qwen_overrides:
+        kwargs["config"] = config
+        kwargs.update(qwen_overrides)
+    elif dev.supports_device_map_auto():
         kwargs["device_map"] = "auto"
 
     try: