fix loras on nf4 models when activate "loras in fp16"

backend/operations.py

@@ -342,7 +342,7 @@ class ForgeOperations:
 
 
 try:
-    from backend.operations_bnb import ForgeLoader4Bit, ForgeParams4bit, functional_linear_4bits
+    from backend.operations_bnb import ForgeLoader4Bit, ForgeParams4bit, functional_linear_4bits, functional_dequantize_4bit
 
     class ForgeOperationsBNB4bits(ForgeOperations):
         class Linear(ForgeLoader4Bit):
@@ -356,6 +356,11 @@ try:
                     # And it only invokes one time, and most linear does not have bias
                     self.bias = utils.tensor2parameter(self.bias.to(x.dtype))
 
+                if hasattr(self, 'forge_online_loras'):
+                    weight, bias, signal = weights_manual_cast(self, x, weight_fn=functional_dequantize_4bit, bias_fn=None, skip_bias_dtype=True)
+                    with main_stream_worker(weight, bias, signal):
+                        return torch.nn.functional.linear(x, weight, bias)
+
                 if not self.parameters_manual_cast:
                     return functional_linear_4bits(x, self.weight, self.bias)
                 elif not self.weight.bnb_quantized:

modules_forge/main_entry.py

@@ -25,9 +25,11 @@ forge_unet_storage_dtype_options = {
     'Automatic': (None, False),
     'Automatic (fp16 LoRA)': (None, True),
     'bnb-nf4': ('nf4', False),
+    'bnb-nf4 (fp16 LoRA)': ('nf4', True),
     'float8-e4m3fn': (torch.float8_e4m3fn, False),
     'float8-e4m3fn (fp16 LoRA)': (torch.float8_e4m3fn, True),
     'bnb-fp4': ('fp4', False),
+    'bnb-fp4 (fp16 LoRA)': ('fp4', True),
     'float8-e5m2': (torch.float8_e5m2, False),
     'float8-e5m2 (fp16 LoRA)': (torch.float8_e5m2, True),
 }
@@ -195,6 +197,7 @@ def refresh_model_loading_parameters():
     )
 
     print(f'Model selected: {model_data.forge_loading_parameters}')
+    processing.need_global_unload = True
 
     return