SDNQ revert device_map = gpu

CHANGELOG.md

@@ -15,7 +15,6 @@
   - Implement better layer hijacks  
   - Fix Conv quant  
   - Fix lora weight change  
-  - Use device_map = gpu  
 - **IPEX**  
   - Disabe Dynamic Attention by default on PyTorch 2.7  
   - Remove GradScaler hijack and use torch.amp.GradScaler instead  

modules/model_quant.py

@@ -529,9 +529,6 @@ def get_dit_args(load_config:dict={}, module:str=None, device_map:bool=False, al
             config['device_map'] = 'cpu'
         elif shared.opts.device_map == 'gpu':
             config['device_map'] = devices.device
-        elif shared.opts.device_map == 'default' and module not in {'TE', 'LLM'}: # Transformers loads the entire model into GPU at once and OOMs
-            if 'Model' in shared.opts.sdnq_quantize_weights or (module is not None and module in shared.opts.sdnq_quantize_weights) or module == 'any':
-                config['device_map'] = devices.device
     if allow_quant:
         quant_args = create_config(module=module)
     else: