[LoRA] Make sure LoRA can be disabled after it's run (huggingface#2128)

patrickvonplaten · web-flow · commit f653ded7eda6 · 2023-01-26T21:26:11.000+01:00
diff --git a/src/diffusers/models/cross_attention.py b/src/diffusers/models/cross_attention.py
@@ -17,9 +17,13 @@
 import torch.nn.functional as F
 from torch import nn
 
+from ..utils import logging
 from ..utils.import_utils import is_xformers_available
 
 
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
 if is_xformers_available():
     import xformers
     import xformers.ops
@@ -151,6 +155,16 @@ def set_attention_slice(self, slice_size):
         self.set_processor(processor)
 
     def set_processor(self, processor: "AttnProcessor"):
+        # if current processor is in `self._modules` and if passed `processor` is not, we need to
+        # pop `processor` from `self._modules`
+        if (
+            hasattr(self, "processor")
+            and isinstance(self.processor, torch.nn.Module)
+            and not isinstance(processor, torch.nn.Module)
+        ):
+            logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}")
+            self._modules.pop("processor")
+
         self.processor = processor
 
     def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, **cross_attention_kwargs):
diff --git a/tests/models/test_models_unet_2d_condition.py b/tests/models/test_models_unet_2d_condition.py
@@ -20,7 +20,7 @@
 import torch
 
 from diffusers import UNet2DConditionModel
-from diffusers.models.cross_attention import LoRACrossAttnProcessor
+from diffusers.models.cross_attention import CrossAttnProcessor, LoRACrossAttnProcessor
 from diffusers.utils import (
     floats_tensor,
     load_hf_numpy,
@@ -40,6 +40,34 @@
 torch.backends.cuda.matmul.allow_tf32 = False
 
 
+def create_lora_layers(model):
+    lora_attn_procs = {}
+    for name in model.attn_processors.keys():
+        cross_attention_dim = None if name.endswith("attn1.processor") else model.config.cross_attention_dim
+        if name.startswith("mid_block"):
+            hidden_size = model.config.block_out_channels[-1]
+        elif name.startswith("up_blocks"):
+            block_id = int(name[len("up_blocks.")])
+            hidden_size = list(reversed(model.config.block_out_channels))[block_id]
+        elif name.startswith("down_blocks"):
+            block_id = int(name[len("down_blocks.")])
+            hidden_size = model.config.block_out_channels[block_id]
+
+        lora_attn_procs[name] = LoRACrossAttnProcessor(
+            hidden_size=hidden_size, cross_attention_dim=cross_attention_dim
+        )
+        lora_attn_procs[name] = lora_attn_procs[name].to(model.device)
+
+        # add 1 to weights to mock trained weights
+        with torch.no_grad():
+            lora_attn_procs[name].to_q_lora.up.weight += 1
+            lora_attn_procs[name].to_k_lora.up.weight += 1
+            lora_attn_procs[name].to_v_lora.up.weight += 1
+            lora_attn_procs[name].to_out_lora.up.weight += 1
+
+    return lora_attn_procs
+
+
 class UNet2DConditionModelTests(ModelTesterMixin, unittest.TestCase):
     model_class = UNet2DConditionModel
 
@@ -336,30 +364,7 @@ def test_lora_save_load(self):
         with torch.no_grad():
             old_sample = model(**inputs_dict).sample
 
-        lora_attn_procs = {}
-        for name in model.attn_processors.keys():
-            cross_attention_dim = None if name.endswith("attn1.processor") else model.config.cross_attention_dim
-            if name.startswith("mid_block"):
-                hidden_size = model.config.block_out_channels[-1]
-            elif name.startswith("up_blocks"):
-                block_id = int(name[len("up_blocks.")])
-                hidden_size = list(reversed(model.config.block_out_channels))[block_id]
-            elif name.startswith("down_blocks"):
-                block_id = int(name[len("down_blocks.")])
-                hidden_size = model.config.block_out_channels[block_id]
-
-            lora_attn_procs[name] = LoRACrossAttnProcessor(
-                hidden_size=hidden_size, cross_attention_dim=cross_attention_dim
-            )
-            lora_attn_procs[name] = lora_attn_procs[name].to(model.device)
-
-            # add 1 to weights to mock trained weights
-            with torch.no_grad():
-                lora_attn_procs[name].to_q_lora.up.weight += 1
-                lora_attn_procs[name].to_k_lora.up.weight += 1
-                lora_attn_procs[name].to_v_lora.up.weight += 1
-                lora_attn_procs[name].to_out_lora.up.weight += 1
-
+        lora_attn_procs = create_lora_layers(model)
         model.set_attn_processor(lora_attn_procs)
 
         with torch.no_grad():
@@ -380,6 +385,33 @@ def test_lora_save_load(self):
         # LoRA and no LoRA should NOT be the same
         assert (sample - old_sample).abs().max() > 1e-4
 
+    def test_lora_on_off(self):
+        # enable deterministic behavior for gradient checkpointing
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+
+        init_dict["attention_head_dim"] = (8, 16)
+
+        torch.manual_seed(0)
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+
+        with torch.no_grad():
+            old_sample = model(**inputs_dict).sample
+
+        lora_attn_procs = create_lora_layers(model)
+        model.set_attn_processor(lora_attn_procs)
+
+        with torch.no_grad():
+            sample = model(**inputs_dict, cross_attention_kwargs={"scale": 0.0}).sample
+
+        model.set_attn_processor(CrossAttnProcessor())
+
+        with torch.no_grad():
+            new_sample = model(**inputs_dict).sample
+
+        assert (sample - new_sample).abs().max() < 1e-4
+        assert (sample - old_sample).abs().max() < 1e-4
+
 
 @slow
 class UNet2DConditionModelIntegrationTests(unittest.TestCase):
