Fix #1362 by improving VRAM usage patterns when doing .swap()

commit ef3f7a26e242b73c2beb0195c7fd8f654ef47f55
Author: damian0815 <null@damianstewart.com>
Date:   Tue Nov 8 12:18:37 2022 +0100

    remove log spam

commit 7189d649622d4668b120b0dd278388ad672142c4
Author: damian0815 <null@damianstewart.com>
Date:   Tue Nov 8 12:10:28 2022 +0100

    change the way saved slicing strategy is applied

commit 01c40f751ab72955140165c16f95ae411732265b
Author: damian0815 <null@damianstewart.com>
Date:   Tue Nov 8 12:04:43 2022 +0100

    fix slicing_strategy_getter callsite

commit f8cfe25150a346958903316bc710737d99839923
Author: damian0815 <null@damianstewart.com>
Date:   Tue Nov 8 11:56:22 2022 +0100

    cleanup, consistent dim=0 also tested

commit 5bf9b1e890d48e962afd4a668a219b68271e5dc1
Author: damian0815 <null@damianstewart.com>
Date:   Tue Nov 8 11:34:09 2022 +0100

    refactored context, tested with non-sliced cross attention control

commit d58a46e39bf562e7459290d2444256e8c08ad0b6
Author: damian0815 <null@damianstewart.com>
Date:   Sun Nov 6 00:41:52 2022 +0100

    cleanup

commit 7e2c658b4c06fe239311b65b9bb16fa3adec7fd7
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 22:57:31 2022 +0100

    disable logs

commit 20ee89d93841b070738b3d8a4385c93b097d92eb
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 22:36:58 2022 +0100

    slice saved attention if necessary

commit 0a7684a22c880ec0f48cc22bfed4526358f71546
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 22:32:38 2022 +0100

    raise instead of asserting

commit 7083104c7f3a0d8fd96e94a2f391de50a3c942e4
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 22:31:00 2022 +0100

    store dim when saving slices

commit f7c0808ed383ec1dc70645288a798ed2aa4fa85c
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 22:27:16 2022 +0100

    don't retry on exception

commit 749a721e939b3fe7c1741e7998dab6bd2c85a0cb
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 22:24:50 2022 +0100

    stuff

commit 032ab90e9533be8726301ec91b97137e2aadef9a
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 22:20:17 2022 +0100

    more logging

commit 3dc34b387f033482305360e605809d95a40bf6f8
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 22:16:47 2022 +0100

    logs

commit 901c4c1aa4b9bcef695a6551867ec8149e6e6a93
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 22:12:39 2022 +0100

    actually set save_slicing_strategy to True

commit f780e0a0a7c6b6a3db320891064da82589358c8a
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 22:10:35 2022 +0100

    store slicing strategy

commit 93bb6d566fd18c5c69ef7dacc8f74ba2cf671cb7
Author: damian <git@damianstewart.com>
Date:   Sat Nov 5 20:43:48 2022 +0100

    still not it

commit 5e3a9541f8ae00bde524046963910323e20c40b7
Author: damian <git@damianstewart.com>
Date:   Sat Nov 5 17:20:02 2022 +0100

    wip offloading attention slices on-demand

commit 4c2966aa856b6f3b446216da3619ae931552ef08
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 15:47:40 2022 +0100

    pre-emptive offloading, idk if it works

commit 572576755e9f0a878d38e8173e485126c0efbefb
Author: root <you@example.com>
Date:   Sat Nov 5 11:25:32 2022 +0000

    push attention slices to cpu. slow but saves memory.

commit b57c83a68f2ac03976ebc89ce2ff03812d6d185f
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 12:04:22 2022 +0100

    verbose logging

commit 3a5dae116f110a96585d9eb71d713b5ed2bc3d2b
Author: damian0815 <null@damianstewart.com>
Date:   Sat Nov 5 11:50:48 2022 +0100

    wip fixing mem strategy crash (4 test on runpod)

commit 3cf237db5fae0c7b0b4cc3c47c81830bdb2ae7de
Author: damian0815 <null@damianstewart.com>
Date:   Fri Nov 4 09:02:40 2022 +0100

    wip, only works on cuda

ldm/models/diffusion/cross_attention_control.py

@@ -1,10 +1,13 @@
-from enum import Enum
+import enum
+from typing import Optional
 
 import torch
 
 # adapted from bloc97's CrossAttentionControl colab
 # https://github.com/bloc97/CrossAttentionControl
 
+
+
 class CrossAttentionControl:
 
     class Arguments:
@@ -27,7 +30,14 @@ class CrossAttentionControl:
                     print('warning: cross-attention control options are not working properly for >1 edit')
                 self.edit_options = non_none_edit_options[0]
 
+
     class Context:
+
+        class Action(enum.Enum):
+            NONE = 0
+            SAVE = 1,
+            APPLY = 2
+
         def __init__(self, arguments: 'CrossAttentionControl.Arguments', step_count: int):
             """
             :param arguments: Arguments for the cross-attention control process
@@ -36,14 +46,124 @@ class CrossAttentionControl:
             self.arguments = arguments
             self.step_count = step_count
 
+            self.self_cross_attention_module_identifiers = []
+            self.tokens_cross_attention_module_identifiers = []
+
+            self.saved_cross_attention_maps = {}
+
+            self.clear_requests(cleanup=True)
+
+        def register_cross_attention_modules(self, model):
+            for name,module in CrossAttentionControl.get_attention_modules(model,
+                                                                                 CrossAttentionControl.CrossAttentionType.SELF):
+                self.self_cross_attention_module_identifiers.append(name)
+            for name,module in CrossAttentionControl.get_attention_modules(model,
+                                                                                   CrossAttentionControl.CrossAttentionType.TOKENS):
+                self.tokens_cross_attention_module_identifiers.append(name)
+
+        def request_save_attention_maps(self, cross_attention_type: 'CrossAttentionControl.CrossAttentionType'):
+            if cross_attention_type == CrossAttentionControl.CrossAttentionType.SELF:
+                self.self_cross_attention_action = CrossAttentionControl.Context.Action.SAVE
+            else:
+                self.tokens_cross_attention_action = CrossAttentionControl.Context.Action.SAVE
+
+        def request_apply_saved_attention_maps(self, cross_attention_type: 'CrossAttentionControl.CrossAttentionType'):
+            if cross_attention_type == CrossAttentionControl.CrossAttentionType.SELF:
+                self.self_cross_attention_action = CrossAttentionControl.Context.Action.APPLY
+            else:
+                self.tokens_cross_attention_action = CrossAttentionControl.Context.Action.APPLY
+
+        def is_tokens_cross_attention(self, module_identifier) -> bool:
+            return module_identifier in self.tokens_cross_attention_module_identifiers
+
+        def get_should_save_maps(self, module_identifier: str) -> bool:
+            if module_identifier in self.self_cross_attention_module_identifiers:
+                return self.self_cross_attention_action == CrossAttentionControl.Context.Action.SAVE
+            elif module_identifier in self.tokens_cross_attention_module_identifiers:
+                return self.tokens_cross_attention_action == CrossAttentionControl.Context.Action.SAVE
+            return False
+
+        def get_should_apply_saved_maps(self, module_identifier: str) -> bool:
+            if module_identifier in self.self_cross_attention_module_identifiers:
+                return self.self_cross_attention_action == CrossAttentionControl.Context.Action.APPLY
+            elif module_identifier in self.tokens_cross_attention_module_identifiers:
+                return self.tokens_cross_attention_action == CrossAttentionControl.Context.Action.APPLY
+            return False
+
+        def get_active_cross_attention_control_types_for_step(self, percent_through:float=None)\
+                -> list['CrossAttentionControl.CrossAttentionType']:
+            """
+            Should cross-attention control be applied on the given step?
+            :param percent_through: How far through the step sequence are we (0.0=pure noise, 1.0=completely denoised image). Expected range 0.0..<1.0.
+            :return: A list of attention types that cross-attention control should be performed for on the given step. May be [].
+            """
+            if percent_through is None:
+                return [CrossAttentionControl.CrossAttentionType.SELF, CrossAttentionControl.CrossAttentionType.TOKENS]
+
+            opts = self.arguments.edit_options
+            to_control = []
+            if opts['s_start'] <= percent_through and percent_through < opts['s_end']:
+                to_control.append(CrossAttentionControl.CrossAttentionType.SELF)
+            if opts['t_start'] <= percent_through and percent_through < opts['t_end']:
+                to_control.append(CrossAttentionControl.CrossAttentionType.TOKENS)
+            return to_control
+
+        def save_slice(self, identifier: str, slice: torch.Tensor, dim: Optional[int], offset: int,
+                       slice_size: Optional[int]):
+            if identifier not in self.saved_cross_attention_maps:
+                self.saved_cross_attention_maps[identifier] = {
+                    'dim': dim,
+                    'slice_size': slice_size,
+                    'slices': {offset or 0: slice}
+                }
+            else:
+                self.saved_cross_attention_maps[identifier]['slices'][offset or 0] = slice
+
+        def get_slice(self, identifier: str, requested_dim: Optional[int], requested_offset: int, slice_size: int):
+            saved_attention_dict = self.saved_cross_attention_maps[identifier]
+            if requested_dim is None:
+                if saved_attention_dict['dim'] is not None:
+                    raise RuntimeError(f"dim mismatch: expected dim=None, have {saved_attention_dict['dim']}")
+                return saved_attention_dict['slices'][0]
+
+            if saved_attention_dict['dim'] == requested_dim:
+                if slice_size != saved_attention_dict['slice_size']:
+                    raise RuntimeError(
+                        f"slice_size mismatch: expected slice_size={slice_size}, have {saved_attention_dict['slice_size']}")
+                return saved_attention_dict['slices'][requested_offset]
+
+            if saved_attention_dict['dim'] == None:
+                whole_saved_attention = saved_attention_dict['slices'][0]
+                if requested_dim == 0:
+                    return whole_saved_attention[requested_offset:requested_offset + slice_size]
+                elif requested_dim == 1:
+                    return whole_saved_attention[:, requested_offset:requested_offset + slice_size]
+
+            raise RuntimeError(f"Cannot convert dim {saved_attention_dict['dim']} to requested dim {requested_dim}")
+
+        def get_slicing_strategy(self, identifier: str) -> Optional[tuple[int, int]]:
+            saved_attention = self.saved_cross_attention_maps.get(identifier, None)
+            if saved_attention is None:
+                return None, None
+            return saved_attention['dim'], saved_attention['slice_size']
+
+        def clear_requests(self, cleanup=True):
+            self.tokens_cross_attention_action = CrossAttentionControl.Context.Action.NONE
+            self.self_cross_attention_action = CrossAttentionControl.Context.Action.NONE
+            if cleanup:
+                self.saved_cross_attention_maps = {}
+
+        def offload_saved_attention_slices_to_cpu(self):
+            for key, map_dict in self.saved_cross_attention_maps.items():
+                for offset, slice in map_dict['slices'].items():
+                    map_dict[offset] = slice.to('cpu')
+
     @classmethod
     def remove_cross_attention_control(cls, model):
         cls.remove_attention_function(model)
 
     @classmethod
-    def setup_cross_attention_control(cls, model,
-                                      cross_attention_control_args: Arguments
-                                      ):
+    def setup_cross_attention_control(cls, model, context: Context):
         """
         Inject attention parameters and functions into the passed in model to enable cross attention editing.
 
@@ -53,7 +173,7 @@ class CrossAttentionControl:
         """
 
         # adapted from init_attention_edit
-        device = cross_attention_control_args.edited_conditioning.device
+        device = context.arguments.edited_conditioning.device
 
         # urgh. should this be hardcoded?
         max_length = 77
@@ -61,141 +181,82 @@ class CrossAttentionControl:
         mask = torch.zeros(max_length)
         indices_target = torch.arange(max_length, dtype=torch.long)
         indices = torch.zeros(max_length, dtype=torch.long)
-        for name, a0, a1, b0, b1 in cross_attention_control_args.edit_opcodes:
+        for name, a0, a1, b0, b1 in context.arguments.edit_opcodes:
             if b0 < max_length:
                 if name == "equal":# or (name == "replace" and a1 - a0 == b1 - b0):
                     # these tokens have not been edited
                     indices[b0:b1] = indices_target[a0:a1]
                     mask[b0:b1] = 1
 
-        cls.inject_attention_function(model)
-
-        for m in cls.get_attention_modules(model, cls.CrossAttentionType.SELF):
-            m.last_attn_slice_mask = None
-            m.last_attn_slice_indices = None
-
-        for m in cls.get_attention_modules(model, cls.CrossAttentionType.TOKENS):
-            m.last_attn_slice_mask = mask.to(device)
-            m.last_attn_slice_indices = indices.to(device)
+        context.register_cross_attention_modules(model)
+        context.cross_attention_mask = mask.to(device)
+        context.cross_attention_index_map = indices.to(device)
+        cls.inject_attention_function(model, context)
 
 
-    class CrossAttentionType(Enum):
+    class CrossAttentionType(enum.Enum):
         SELF = 1
         TOKENS = 2
 
-    @classmethod
-    def get_active_cross_attention_control_types_for_step(cls, context: 'CrossAttentionControl.Context', percent_through:float=None)\
-            -> list['CrossAttentionControl.CrossAttentionType']:
-        """
-        Should cross-attention control be applied on the given step?
-        :param percent_through: How far through the step sequence are we (0.0=pure noise, 1.0=completely denoised image). Expected range 0.0..<1.0.
-        :return: A list of attention types that cross-attention control should be performed for on the given step. May be [].
-        """
-        if percent_through is None:
-            return [cls.CrossAttentionType.SELF, cls.CrossAttentionType.TOKENS]
-
-        opts = context.arguments.edit_options
-        to_control = []
-        if opts['s_start'] <= percent_through and percent_through < opts['s_end']:
-            to_control.append(cls.CrossAttentionType.SELF)
-        if opts['t_start'] <= percent_through and percent_through < opts['t_end']:
-            to_control.append(cls.CrossAttentionType.TOKENS)
-        return to_control
-
-
     @classmethod
     def get_attention_modules(cls, model, which: CrossAttentionType):
         which_attn = "attn1" if which is cls.CrossAttentionType.SELF else "attn2"
-        return [module for name, module in model.named_modules() if
+        return [(name,module) for name, module in model.named_modules() if
                       type(module).__name__ == "CrossAttention" and which_attn in name]
 
-    @classmethod
-    def clear_requests(cls, model, clear_attn_slice=True):
-        self_attention_modules = cls.get_attention_modules(model, cls.CrossAttentionType.SELF)
-        tokens_attention_modules = cls.get_attention_modules(model, cls.CrossAttentionType.TOKENS)
-        for m in self_attention_modules+tokens_attention_modules:
-            m.save_last_attn_slice = False
-            m.use_last_attn_slice = False
-            if clear_attn_slice:
-                m.last_attn_slice = None
 
     @classmethod
-    def request_save_attention_maps(cls, model, cross_attention_type: CrossAttentionType):
-        modules = cls.get_attention_modules(model, cross_attention_type)
-        for m in modules:
-            # clear out the saved slice in case the outermost dim changes
-            m.last_attn_slice = None
-            m.save_last_attn_slice = True
-
-    @classmethod
-    def request_apply_saved_attention_maps(cls, model, cross_attention_type: CrossAttentionType):
-        modules = cls.get_attention_modules(model, cross_attention_type)
-        for m in modules:
-            m.use_last_attn_slice = True
-
-
-
-    @classmethod
-    def inject_attention_function(cls, unet):
+    def inject_attention_function(cls, unet, context: 'CrossAttentionControl.Context'):
         # ORIGINAL SOURCE CODE: https://github.com/huggingface/diffusers/blob/91ddd2a25b848df0fa1262d4f1cd98c7ccb87750/src/diffusers/models/attention.py#L276
 
-        def attention_slice_wrangler(self, attention_scores, suggested_attention_slice, dim, offset, slice_size):
+        def attention_slice_wrangler(module, suggested_attention_slice:torch.Tensor, dim, offset, slice_size):
 
-            #print("in wrangler with suggested_attention_slice shape", suggested_attention_slice.shape, "dim", dim)
+            #memory_usage = suggested_attention_slice.element_size() * suggested_attention_slice.nelement()
 
-            attn_slice = suggested_attention_slice
-            if dim is not None:
-                start = offset
-                end = start+slice_size
-                #print(f"in wrangler, sliced dim {dim} {start}-{end}, use_last_attn_slice is {self.use_last_attn_slice}, save_last_attn_slice is {self.save_last_attn_slice}")
-            #else:
-            #    print(f"in wrangler, whole, use_last_attn_slice is {self.use_last_attn_slice}, save_last_attn_slice is {self.save_last_attn_slice}")
+            attention_slice = suggested_attention_slice
 
-            if self.use_last_attn_slice:
-                if dim is None:
-                    last_attn_slice = self.last_attn_slice
-                    # print("took whole slice of shape", attn_slice.shape, "from complete shape", self.last_attn_slice.shape)
+            if context.get_should_save_maps(module.identifier):
+                #print(module.identifier, "saving suggested_attention_slice of shape",
+                #      suggested_attention_slice.shape, "dim", dim, "offset", offset)
+                slice_to_save = attention_slice.to('cpu') if dim is not None else attention_slice
+                context.save_slice(module.identifier, slice_to_save, dim=dim, offset=offset, slice_size=slice_size)
+            elif context.get_should_apply_saved_maps(module.identifier):
+                #print(module.identifier, "applying saved attention slice for dim", dim, "offset", offset)
+                saved_attention_slice = context.get_slice(module.identifier, dim, offset, slice_size)
+
+                # slice may have been offloaded to CPU
+                saved_attention_slice = saved_attention_slice.to(suggested_attention_slice.device)
+
+                if context.is_tokens_cross_attention(module.identifier):
+                    index_map = context.cross_attention_index_map
+                    remapped_saved_attention_slice = torch.index_select(saved_attention_slice, -1, index_map)
+                    this_attention_slice = suggested_attention_slice
+
+                    mask = context.cross_attention_mask
+                    saved_mask = mask
+                    this_mask = 1 - mask
+                    attention_slice = remapped_saved_attention_slice * saved_mask + \
+                                      this_attention_slice * this_mask
                 else:
-                    last_attn_slice = self.last_attn_slice[offset]
-
-                if self.last_attn_slice_mask is None:
                     # just use everything
-                    attn_slice = last_attn_slice
-                else:
-                    last_attn_slice_mask = self.last_attn_slice_mask
-                    remapped_last_attn_slice = torch.index_select(last_attn_slice, -1, self.last_attn_slice_indices)
+                    attention_slice = saved_attention_slice
 
-                    this_attn_slice = attn_slice
-                    this_attn_slice_mask = 1 - last_attn_slice_mask
-                    attn_slice = this_attn_slice * this_attn_slice_mask + \
-                                 remapped_last_attn_slice * last_attn_slice_mask
-
-            if self.save_last_attn_slice:
-                if dim is None:
-                    self.last_attn_slice = attn_slice
-                else:
-                    if self.last_attn_slice is None:
-                        self.last_attn_slice = { offset: attn_slice }
-                    else:
-                        self.last_attn_slice[offset] = attn_slice
-
-            return attn_slice
+            return attention_slice
 
         for name, module in unet.named_modules():
             module_name = type(module).__name__
             if module_name == "CrossAttention":
-                module.last_attn_slice = None
-                module.last_attn_slice_indices = None
-                module.last_attn_slice_mask = None
-                module.use_last_attn_weights = False
-                module.use_last_attn_slice = False
-                module.save_last_attn_slice = False
+                module.identifier = name
                 module.set_attention_slice_wrangler(attention_slice_wrangler)
+                module.set_slicing_strategy_getter(lambda module, module_identifier=name: \
+                                                       context.get_slicing_strategy(module_identifier))
 
     @classmethod
     def remove_attention_function(cls, unet):
+        # clear wrangler callback
         for name, module in unet.named_modules():
             module_name = type(module).__name__
             if module_name == "CrossAttention":
                 module.set_attention_slice_wrangler(None)
+                module.set_slicing_strategy_getter(None)
 

ldm/models/diffusion/shared_invokeai_diffusion.py

@@ -1,9 +1,11 @@
+import traceback
 from math import ceil
 from typing import Callable, Optional, Union
 
 import torch
 
 from ldm.models.diffusion.cross_attention_control import CrossAttentionControl
+from ldm.modules.attention import get_mem_free_total
 
 
 class InvokeAIDiffuserComponent:
@@ -34,7 +36,7 @@ class InvokeAIDiffuserComponent:
         """
         self.model = model
         self.model_forward_callback = model_forward_callback
-
+        self.cross_attention_control_context = None
 
     def setup_cross_attention_control(self, conditioning: ExtraConditioningInfo, step_count: int):
         self.conditioning = conditioning
@@ -42,11 +44,7 @@ class InvokeAIDiffuserComponent:
             arguments=self.conditioning.cross_attention_control_args,
             step_count=step_count
         )
-        CrossAttentionControl.setup_cross_attention_control(self.model,
-                                                            cross_attention_control_args=self.conditioning.cross_attention_control_args
-                                                            )
-        #todo: refactor  edited_conditioning, edit_opcodes, edit_options into a struct
-        #todo: apply edit_options using step_count
+        CrossAttentionControl.setup_cross_attention_control(self.model, self.cross_attention_control_context)
 
     def remove_cross_attention_control(self):
         self.conditioning = None
@@ -54,6 +52,7 @@ class InvokeAIDiffuserComponent:
         CrossAttentionControl.remove_cross_attention_control(self.model)
 
 
+
     def do_diffusion_step(self, x: torch.Tensor, sigma: torch.Tensor,
                                 unconditioning: Union[torch.Tensor,dict],
                                 conditioning: Union[torch.Tensor,dict],
@@ -70,12 +69,12 @@ class InvokeAIDiffuserComponent:
         :return: the new latents after applying the model to x using unscaled unconditioning and CFG-scaled conditioning.
         """
 
-        CrossAttentionControl.clear_requests(self.model)
 
         cross_attention_control_types_to_do = []
+        context: CrossAttentionControl.Context = self.cross_attention_control_context
         if self.cross_attention_control_context is not None:
             percent_through = self.estimate_percent_through(step_index, sigma)
-            cross_attention_control_types_to_do = CrossAttentionControl.get_active_cross_attention_control_types_for_step(self.cross_attention_control_context, percent_through)
+            cross_attention_control_types_to_do = context.get_active_cross_attention_control_types_for_step(percent_through)
 
         wants_cross_attention_control = (len(cross_attention_control_types_to_do) > 0)
         wants_hybrid_conditioning = isinstance(conditioning, dict)
@@ -124,7 +123,7 @@ class InvokeAIDiffuserComponent:
         return unconditioned_next_x, conditioned_next_x
 
 
-    def apply_cross_attention_controlled_conditioning(self, x, sigma, unconditioning, conditioning, cross_attention_control_types_to_do):
+    def apply_cross_attention_controlled_conditioning(self, x:torch.Tensor, sigma, unconditioning, conditioning, cross_attention_control_types_to_do):
         # print('pct', percent_through, ': doing cross attention control on', cross_attention_control_types_to_do)
         # slower non-batched path (20% slower on mac MPS)
         # We are only interested in using attention maps for conditioned_next_x, but batching them with generation of
@@ -134,32 +133,30 @@ class InvokeAIDiffuserComponent:
         # representing batched uncond + cond, but then when it comes to applying the saved attention, the
         # wrangler gets an attention tensor which only has shape[0]=8, representing just self.edited_conditionings.)
         # todo: give CrossAttentionControl's `wrangler` function more info so it can work with a batched call as well.
+        context:CrossAttentionControl.Context = self.cross_attention_control_context
 
         try:
             unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning)
 
             # process x using the original prompt, saving the attention maps
-            for type in cross_attention_control_types_to_do:
-                CrossAttentionControl.request_save_attention_maps(self.model, type)
+            #print("saving attention maps for", cross_attention_control_types_to_do)
+            for ca_type in cross_attention_control_types_to_do:
+                context.request_save_attention_maps(ca_type)
             _ = self.model_forward_callback(x, sigma, conditioning)
-            CrossAttentionControl.clear_requests(self.model, clear_attn_slice=False)
+            context.clear_requests(cleanup=False)
 
             # process x again, using the saved attention maps to control where self.edited_conditioning will be applied
-            for type in cross_attention_control_types_to_do:
-                CrossAttentionControl.request_apply_saved_attention_maps(self.model, type)
+            #print("applying saved attention maps for", cross_attention_control_types_to_do)
+            for ca_type in cross_attention_control_types_to_do:
+                context.request_apply_saved_attention_maps(ca_type)
             edited_conditioning = self.conditioning.cross_attention_control_args.edited_conditioning
             conditioned_next_x = self.model_forward_callback(x, sigma, edited_conditioning)
 
-            CrossAttentionControl.clear_requests(self.model)
+        finally:
+            context.clear_requests(cleanup=True)
 
         return unconditioned_next_x, conditioned_next_x
 
-        except RuntimeError:
-            # make sure we clean out the attention slices we're storing on the model
-            # TODO don't store things on the model
-            CrossAttentionControl.clear_requests(self.model)
-            raise
-
     def estimate_percent_through(self, step_index, sigma):
         if step_index is not None and self.cross_attention_control_context is not None:
             # percent_through will never reach 1.0 (but this is intended)

ldm/modules/attention.py

@@ -1,6 +1,6 @@
 from inspect import isfunction
 import math
-from typing import Callable
+from typing import Callable, Optional
 
 import torch
 import torch.nn.functional as F
@@ -151,6 +151,17 @@ class SpatialSelfAttention(nn.Module):
 
         return x+h_
 
+def get_mem_free_total(device):
+    #only on cuda
+    if not torch.cuda.is_available():
+        return None
+    stats = torch.cuda.memory_stats(device)
+    mem_active = stats['active_bytes.all.current']
+    mem_reserved = stats['reserved_bytes.all.current']
+    mem_free_cuda, _ = torch.cuda.mem_get_info(device)
+    mem_free_torch = mem_reserved - mem_active
+    mem_free_total = mem_free_cuda + mem_free_torch
+    return mem_free_total
 
 
 class CrossAttention(nn.Module):
@@ -173,31 +184,43 @@ class CrossAttention(nn.Module):
 
         self.mem_total_gb = psutil.virtual_memory().total // (1 << 30)
 
+        self.cached_mem_free_total = None
         self.attention_slice_wrangler = None
+        self.slicing_strategy_getter = None
 
-    def set_attention_slice_wrangler(self, wrangler:Callable[[nn.Module, torch.Tensor, torch.Tensor, int, int, int], torch.Tensor]):
+    def set_attention_slice_wrangler(self, wrangler: Optional[Callable[[nn.Module, torch.Tensor, int, int, int], torch.Tensor]]):
         '''
         Set custom attention calculator to be called when attention is calculated
-        :param wrangler: Callback, with args (self, attention_scores, suggested_attention_slice, dim, offset, slice_size),
+        :param wrangler: Callback, with args (module, suggested_attention_slice, dim, offset, slice_size),
         which returns either the suggested_attention_slice or an adjusted equivalent.
-            self is the current CrossAttention module for which the callback is being invoked.
-            attention_scores are the scores for attention
-            suggested_attention_slice is a softmax(dim=-1) over attention_scores
-            dim is -1 if the call is non-sliced, or 0 or 1 for dimension-0 or dimension-1 slicing.
-                If dim is >= 0, offset and slice_size specify the slice start and length.
+            `module` is the current CrossAttention module for which the callback is being invoked.
+            `suggested_attention_slice` is the default-calculated attention slice
+            `dim` is -1 if the attenion map has not been sliced, or 0 or 1 for dimension-0 or dimension-1 slicing.
+                If `dim` is >= 0, `offset` and `slice_size` specify the slice start and length.
 
         Pass None to use the default attention calculation.
         :return:
         '''
         self.attention_slice_wrangler = wrangler
 
+    def set_slicing_strategy_getter(self, getter: Optional[Callable[[nn.Module], tuple[int,int]]]):
+        self.slicing_strategy_getter = getter
+
+    def cache_free_memory_count(self, device):
+        self.cached_mem_free_total = get_mem_free_total(device)
+        print("free cuda memory: ", self.cached_mem_free_total)
+
+    def clear_cached_free_memory_count(self):
+        self.cached_mem_free_total = None
+
     def einsum_lowest_level(self, q, k, v, dim, offset, slice_size):
         # calculate attention scores
         attention_scores = einsum('b i d, b j d -> b i j', q, k)
-        # calculate attenion slice by taking the best scores for each latent pixel
+        # calculate attention slice by taking the best scores for each latent pixel
         default_attention_slice = attention_scores.softmax(dim=-1, dtype=attention_scores.dtype)
-        if self.attention_slice_wrangler is not None:
-            attention_slice = self.attention_slice_wrangler(self, attention_scores, default_attention_slice, dim, offset, slice_size)
+        attention_slice_wrangler = self.attention_slice_wrangler
+        if attention_slice_wrangler is not None:
+            attention_slice = attention_slice_wrangler(self, default_attention_slice, dim, offset, slice_size)
         else:
             attention_slice = default_attention_slice
 
@@ -240,17 +263,27 @@ class CrossAttention(nn.Module):
         return self.einsum_op_slice_dim1(q, k, v, max(q.shape[1] // div, 1))
 
     def einsum_op_cuda(self, q, k, v):
-        stats = torch.cuda.memory_stats(q.device)
-        mem_active = stats['active_bytes.all.current']
-        mem_reserved = stats['reserved_bytes.all.current']
-        mem_free_cuda, _ = torch.cuda.mem_get_info(q.device)
-        mem_free_torch = mem_reserved - mem_active
-        mem_free_total = mem_free_cuda + mem_free_torch
+        # check if we already have a slicing strategy (this should only happen during cross-attention controlled generation)
+        slicing_strategy_getter = self.slicing_strategy_getter
+        if slicing_strategy_getter is not None:
+            (dim, slice_size) = slicing_strategy_getter(self)
+            if dim is not None:
+                # print("using saved slicing strategy with dim", dim, "slice size", slice_size)
+                if dim == 0:
+                    return self.einsum_op_slice_dim0(q, k, v, slice_size)
+                elif dim == 1:
+                    return self.einsum_op_slice_dim1(q, k, v, slice_size)
+
+        # fallback for when there is no saved strategy, or saved strategy does not slice
+        mem_free_total = self.cached_mem_free_total or get_mem_free_total(q.device)
         # Divide factor of safety as there's copying and fragmentation
         return self.einsum_op_tensor_mem(q, k, v, mem_free_total / 3.3 / (1 << 20))
 
+
     def get_attention_mem_efficient(self, q, k, v):
         if q.device.type == 'cuda':
+            torch.cuda.empty_cache()
+            #print("in get_attention_mem_efficient with q shape", q.shape, ", k shape", k.shape, ", free memory is", get_mem_free_total(q.device))
             return self.einsum_op_cuda(q, k, v)
 
         if q.device.type == 'mps':