Enable Runtime Selection of Attention Functions (Comfy-Org#9639)

* Looking into a @wrap_attn decorator to look for 'optimized_attention_override' entry in transformer_options

* Created logging code for this branch so that it can be used to track down all the code paths where transformer_options would need to be added

* Fix memory usage issue with inspect

* Made WAN attention receive transformer_options, test node added to wan to test out attention override later

* Added **kwargs to all attention functions so transformer_options could potentially be passed through

* Make sure wrap_attn doesn't make itself recurse infinitely, attempt to load SageAttention and FlashAttention if not enabled so that they can be marked as available or not, create registry for available attention

* Turn off attention logging for now, make AttentionOverrideTestNode have a dropdown with available attention (this is a test node only)

* Make flux work with optimized_attention_override

* Add logs to verify optimized_attention_override is passed all the way into attention function

* Make Qwen work with optimized_attention_override

* Made hidream work with optimized_attention_override

* Made wan patches_replace work with optimized_attention_override

* Made SD3 work with optimized_attention_override

* Made HunyuanVideo work with optimized_attention_override

* Made Mochi work with optimized_attention_override

* Made LTX work with optimized_attention_override

* Made StableAudio work with optimized_attention_override

* Made optimized_attention_override work with ACE Step

* Made Hunyuan3D work with optimized_attention_override

* Make CosmosPredict2 work with optimized_attention_override

* Made CosmosVideo work with optimized_attention_override

* Made Omnigen 2 work with optimized_attention_override

* Made StableCascade work with optimized_attention_override

* Made AuraFlow work with optimized_attention_override

* Made Lumina work with optimized_attention_override

* Made Chroma work with optimized_attention_override

* Made SVD work with optimized_attention_override

* Fix WanI2VCrossAttention so that it expects to receive transformer_options

* Fixed Wan2.1 Fun Camera transformer_options passthrough

* Fixed WAN 2.1 VACE transformer_options passthrough

* Add optimized to get_attention_function

* Disable attention logs for now

* Remove attention logging code

* Remove _register_core_attention_functions, as we wouldn't want someone to call that, just in case

* Satisfy ruff

* Remove AttentionOverrideTest node, that's something to cook up for later

Author: Kosinkadink

comfy/ldm/ace/attention.py

@@ -133,13 +133,15 @@ def forward(
         hidden_states: torch.Tensor,
         encoder_hidden_states: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None,
+        transformer_options={},
         **cross_attention_kwargs,
     ) -> torch.Tensor:
         return self.processor(
             self,
             hidden_states,
             encoder_hidden_states=encoder_hidden_states,
             attention_mask=attention_mask,
+            transformer_options=transformer_options,
             **cross_attention_kwargs,
         )
 
@@ -366,6 +368,7 @@ def __call__(
         encoder_attention_mask: Optional[torch.FloatTensor] = None,
         rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
         rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        transformer_options={},
         *args,
         **kwargs,
     ) -> torch.Tensor:
@@ -433,7 +436,7 @@ def __call__(
 
         # the output of sdp = (batch, num_heads, seq_len, head_dim)
         hidden_states = optimized_attention(
-            query, key, value, heads=query.shape[1], mask=attention_mask, skip_reshape=True,
+            query, key, value, heads=query.shape[1], mask=attention_mask, skip_reshape=True, transformer_options=transformer_options,
         ).to(query.dtype)
 
         # linear proj
@@ -697,6 +700,7 @@ def forward(
         rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
         rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
         temb: torch.FloatTensor = None,
+        transformer_options={},
     ):
 
         N = hidden_states.shape[0]
@@ -720,6 +724,7 @@ def forward(
                 encoder_attention_mask=encoder_attention_mask,
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=rotary_freqs_cis_cross,
+                transformer_options=transformer_options,
             )
         else:
             attn_output, _ = self.attn(
@@ -729,6 +734,7 @@ def forward(
                 encoder_attention_mask=None,
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=None,
+                transformer_options=transformer_options,
             )
 
         if self.use_adaln_single:
@@ -743,6 +749,7 @@ def forward(
                 encoder_attention_mask=encoder_attention_mask,
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=rotary_freqs_cis_cross,
+                transformer_options=transformer_options,
             )
             hidden_states = attn_output + hidden_states
 

comfy/ldm/ace/model.py

@@ -314,6 +314,7 @@ def decode(
         output_length: int = 0,
         block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None,
         controlnet_scale: Union[float, torch.Tensor] = 1.0,
+        transformer_options={},
     ):
         embedded_timestep = self.timestep_embedder(self.time_proj(timestep).to(dtype=hidden_states.dtype))
         temb = self.t_block(embedded_timestep)
@@ -339,6 +340,7 @@ def decode(
                 rotary_freqs_cis=rotary_freqs_cis,
                 rotary_freqs_cis_cross=encoder_rotary_freqs_cis,
                 temb=temb,
+                transformer_options=transformer_options,
             )
 
         output = self.final_layer(hidden_states, embedded_timestep, output_length)
@@ -393,6 +395,7 @@ def _forward(
 
         output_length = hidden_states.shape[-1]
 
+        transformer_options = kwargs.get("transformer_options", {})
         output = self.decode(
             hidden_states=hidden_states,
             attention_mask=attention_mask,
@@ -402,6 +405,7 @@ def _forward(
             output_length=output_length,
             block_controlnet_hidden_states=block_controlnet_hidden_states,
             controlnet_scale=controlnet_scale,
+            transformer_options=transformer_options,
         )
 
         return output

comfy/ldm/audio/dit.py

@@ -298,7 +298,8 @@ def forward(
         mask = None,
         context_mask = None,
         rotary_pos_emb = None,
-        causal = None
+        causal = None,
+        transformer_options={},
     ):
         h, kv_h, has_context = self.num_heads, self.kv_heads, context is not None
 
@@ -363,7 +364,7 @@ def forward(
             heads_per_kv_head = h // kv_h
             k, v = map(lambda t: t.repeat_interleave(heads_per_kv_head, dim = 1), (k, v))
 
-        out = optimized_attention(q, k, v, h, skip_reshape=True)
+        out = optimized_attention(q, k, v, h, skip_reshape=True, transformer_options=transformer_options)
         out = self.to_out(out)
 
         if mask is not None:
@@ -488,7 +489,8 @@ def forward(
         global_cond=None,
         mask = None,
         context_mask = None,
-        rotary_pos_emb = None
+        rotary_pos_emb = None,
+        transformer_options={}
     ):
         if self.global_cond_dim is not None and self.global_cond_dim > 0 and global_cond is not None:
 
@@ -498,12 +500,12 @@ def forward(
             residual = x
             x = self.pre_norm(x)
             x = x * (1 + scale_self) + shift_self
-            x = self.self_attn(x, mask = mask, rotary_pos_emb = rotary_pos_emb)
+            x = self.self_attn(x, mask = mask, rotary_pos_emb = rotary_pos_emb, transformer_options=transformer_options)
             x = x * torch.sigmoid(1 - gate_self)
             x = x + residual
 
             if context is not None:
-                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask)
+                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask, transformer_options=transformer_options)
 
             if self.conformer is not None:
                 x = x + self.conformer(x)
@@ -517,10 +519,10 @@ def forward(
             x = x + residual
 
         else:
-            x = x + self.self_attn(self.pre_norm(x), mask = mask, rotary_pos_emb = rotary_pos_emb)
+            x = x + self.self_attn(self.pre_norm(x), mask = mask, rotary_pos_emb = rotary_pos_emb, transformer_options=transformer_options)
 
             if context is not None:
-                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask)
+                x = x + self.cross_attn(self.cross_attend_norm(x), context = context, context_mask = context_mask, transformer_options=transformer_options)
 
             if self.conformer is not None:
                 x = x + self.conformer(x)
@@ -606,7 +608,8 @@ def forward(
         return_info = False,
         **kwargs
     ):
-        patches_replace = kwargs.get("transformer_options", {}).get("patches_replace", {})
+        transformer_options = kwargs.get("transformer_options", {})
+        patches_replace = transformer_options.get("patches_replace", {})
         batch, seq, device = *x.shape[:2], x.device
         context = kwargs["context"]
 
@@ -645,13 +648,13 @@ def forward(
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = layer(args["img"], rotary_pos_emb=args["pe"], global_cond=args["vec"], context=args["txt"])
+                    out["img"] = layer(args["img"], rotary_pos_emb=args["pe"], global_cond=args["vec"], context=args["txt"], transformer_options=args["transformer_options"])
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": global_cond, "pe": rotary_pos_emb}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": global_cond, "pe": rotary_pos_emb, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 x = out["img"]
             else:
-                x = layer(x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, context=context)
+                x = layer(x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, context=context, transformer_options=transformer_options)
             # x = checkpoint(layer, x, rotary_pos_emb = rotary_pos_emb, global_cond=global_cond, **kwargs)
 
             if return_info:

comfy/ldm/aura/mmdit.py

@@ -85,7 +85,7 @@ def __init__(self, dim, n_heads, mh_qknorm=False, dtype=None, device=None, opera
         )
 
     #@torch.compile()
-    def forward(self, c):
+    def forward(self, c, transformer_options={}):
 
         bsz, seqlen1, _ = c.shape
 
@@ -95,7 +95,7 @@ def forward(self, c):
         v = v.view(bsz, seqlen1, self.n_heads, self.head_dim)
         q, k = self.q_norm1(q), self.k_norm1(k)
 
-        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True)
+        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True, transformer_options=transformer_options)
         c = self.w1o(output)
         return c
 
@@ -144,7 +144,7 @@ def __init__(self, dim, n_heads, mh_qknorm=False, dtype=None, device=None, opera
 
 
     #@torch.compile()
-    def forward(self, c, x):
+    def forward(self, c, x, transformer_options={}):
 
         bsz, seqlen1, _ = c.shape
         bsz, seqlen2, _ = x.shape
@@ -168,7 +168,7 @@ def forward(self, c, x):
             torch.cat([cv, xv], dim=1),
         )
 
-        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True)
+        output = optimized_attention(q.permute(0, 2, 1, 3), k.permute(0, 2, 1, 3), v.permute(0, 2, 1, 3), self.n_heads, skip_reshape=True, transformer_options=transformer_options)
 
         c, x = output.split([seqlen1, seqlen2], dim=1)
         c = self.w1o(c)
@@ -207,7 +207,7 @@ def __init__(self, dim, heads=8, global_conddim=1024, is_last=False, dtype=None,
         self.is_last = is_last
 
     #@torch.compile()
-    def forward(self, c, x, global_cond, **kwargs):
+    def forward(self, c, x, global_cond, transformer_options={}, **kwargs):
 
         cres, xres = c, x
 
@@ -225,7 +225,7 @@ def forward(self, c, x, global_cond, **kwargs):
         x = modulate(self.normX1(x), xshift_msa, xscale_msa)
 
         # attention
-        c, x = self.attn(c, x)
+        c, x = self.attn(c, x, transformer_options=transformer_options)
 
 
         c = self.normC2(cres + cgate_msa.unsqueeze(1) * c)
@@ -255,13 +255,13 @@ def __init__(self, dim, heads=8, global_conddim=1024, dtype=None, device=None, o
         self.mlp = MLP(dim, hidden_dim=dim * 4, dtype=dtype, device=device, operations=operations)
 
     #@torch.compile()
-    def forward(self, cx, global_cond, **kwargs):
+    def forward(self, cx, global_cond, transformer_options={}, **kwargs):
         cxres = cx
         shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.modCX(
             global_cond
         ).chunk(6, dim=1)
         cx = modulate(self.norm1(cx), shift_msa, scale_msa)
-        cx = self.attn(cx)
+        cx = self.attn(cx, transformer_options=transformer_options)
         cx = self.norm2(cxres + gate_msa.unsqueeze(1) * cx)
         mlpout = self.mlp(modulate(cx, shift_mlp, scale_mlp))
         cx = gate_mlp.unsqueeze(1) * mlpout
@@ -473,13 +473,14 @@ def block_wrap(args):
                         out = {}
                         out["txt"], out["img"] = layer(args["txt"],
                                                        args["img"],
-                                                       args["vec"])
+                                                       args["vec"],
+                                                       transformer_options=args["transformer_options"])
                         return out
-                    out = blocks_replace[("double_block", i)]({"img": x, "txt": c, "vec": global_cond}, {"original_block": block_wrap})
+                    out = blocks_replace[("double_block", i)]({"img": x, "txt": c, "vec": global_cond, "transformer_options": transformer_options}, {"original_block": block_wrap})
                     c = out["txt"]
                     x = out["img"]
                 else:
-                    c, x = layer(c, x, global_cond, **kwargs)
+                    c, x = layer(c, x, global_cond, transformer_options=transformer_options, **kwargs)
 
         if len(self.single_layers) > 0:
             c_len = c.size(1)
@@ -488,13 +489,13 @@ def block_wrap(args):
                 if ("single_block", i) in blocks_replace:
                     def block_wrap(args):
                         out = {}
-                        out["img"] = layer(args["img"], args["vec"])
+                        out["img"] = layer(args["img"], args["vec"], transformer_options=args["transformer_options"])
                         return out
 
-                    out = blocks_replace[("single_block", i)]({"img": cx, "vec": global_cond}, {"original_block": block_wrap})
+                    out = blocks_replace[("single_block", i)]({"img": cx, "vec": global_cond, "transformer_options": transformer_options}, {"original_block": block_wrap})
                     cx = out["img"]
                 else:
-                    cx = layer(cx, global_cond, **kwargs)
+                    cx = layer(cx, global_cond, transformer_options=transformer_options, **kwargs)
 
             x = cx[:, c_len:]
 

comfy/ldm/cascade/common.py

@@ -32,12 +32,12 @@ def __init__(self, c, nhead, dropout=0.0, dtype=None, device=None, operations=No
 
         self.out_proj = operations.Linear(c, c, bias=True, dtype=dtype, device=device)
 
-    def forward(self, q, k, v):
+    def forward(self, q, k, v, transformer_options={}):
         q = self.to_q(q)
         k = self.to_k(k)
         v = self.to_v(v)
 
-        out = optimized_attention(q, k, v, self.heads)
+        out = optimized_attention(q, k, v, self.heads, transformer_options=transformer_options)
 
         return self.out_proj(out)
 
@@ -47,13 +47,13 @@ def __init__(self, c, nhead, dropout=0.0, dtype=None, device=None, operations=No
         self.attn = OptimizedAttention(c, nhead, dtype=dtype, device=device, operations=operations)
         # self.attn = nn.MultiheadAttention(c, nhead, dropout=dropout, bias=True, batch_first=True, dtype=dtype, device=device)
 
-    def forward(self, x, kv, self_attn=False):
+    def forward(self, x, kv, self_attn=False, transformer_options={}):
         orig_shape = x.shape
         x = x.view(x.size(0), x.size(1), -1).permute(0, 2, 1)  # Bx4xHxW -> Bx(HxW)x4
         if self_attn:
             kv = torch.cat([x, kv], dim=1)
         # x = self.attn(x, kv, kv, need_weights=False)[0]
-        x = self.attn(x, kv, kv)
+        x = self.attn(x, kv, kv, transformer_options=transformer_options)
         x = x.permute(0, 2, 1).view(*orig_shape)
         return x
 
@@ -114,9 +114,9 @@ def __init__(self, c, c_cond, nhead, self_attn=True, dropout=0.0, dtype=None, de
             operations.Linear(c_cond, c, dtype=dtype, device=device)
         )
 
-    def forward(self, x, kv):
+    def forward(self, x, kv, transformer_options={}):
         kv = self.kv_mapper(kv)
-        x = x + self.attention(self.norm(x), kv, self_attn=self.self_attn)
+        x = x + self.attention(self.norm(x), kv, self_attn=self.self_attn, transformer_options=transformer_options)
         return x
 
 

comfy/ldm/cascade/stage_b.py

@@ -173,7 +173,7 @@ def gen_c_embeddings(self, clip):
         clip = self.clip_norm(clip)
         return clip
 
-    def _down_encode(self, x, r_embed, clip):
+    def _down_encode(self, x, r_embed, clip, transformer_options={}):
         level_outputs = []
         block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers)
         for down_block, downscaler, repmap in block_group:
@@ -187,7 +187,7 @@ def _down_encode(self, x, r_embed, clip):
                     elif isinstance(block, AttnBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   AttnBlock)):
-                        x = block(x, clip)
+                        x = block(x, clip, transformer_options=transformer_options)
                     elif isinstance(block, TimestepBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   TimestepBlock)):
@@ -199,7 +199,7 @@ def _down_encode(self, x, r_embed, clip):
             level_outputs.insert(0, x)
         return level_outputs
 
-    def _up_decode(self, level_outputs, r_embed, clip):
+    def _up_decode(self, level_outputs, r_embed, clip, transformer_options={}):
         x = level_outputs[0]
         block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers)
         for i, (up_block, upscaler, repmap) in enumerate(block_group):
@@ -216,7 +216,7 @@ def _up_decode(self, level_outputs, r_embed, clip):
                     elif isinstance(block, AttnBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   AttnBlock)):
-                        x = block(x, clip)
+                        x = block(x, clip, transformer_options=transformer_options)
                     elif isinstance(block, TimestepBlock) or (
                             hasattr(block, '_fsdp_wrapped_module') and isinstance(block._fsdp_wrapped_module,
                                                                                   TimestepBlock)):
@@ -228,7 +228,7 @@ def _up_decode(self, level_outputs, r_embed, clip):
             x = upscaler(x)
         return x
 
-    def forward(self, x, r, effnet, clip, pixels=None, **kwargs):
+    def forward(self, x, r, effnet, clip, pixels=None, transformer_options={}, **kwargs):
         if pixels is None:
             pixels = x.new_zeros(x.size(0), 3, 8, 8)
 
@@ -245,8 +245,8 @@ def forward(self, x, r, effnet, clip, pixels=None, **kwargs):
             nn.functional.interpolate(effnet, size=x.shape[-2:], mode='bilinear', align_corners=True))
         x = x + nn.functional.interpolate(self.pixels_mapper(pixels), size=x.shape[-2:], mode='bilinear',
                                           align_corners=True)
-        level_outputs = self._down_encode(x, r_embed, clip)
-        x = self._up_decode(level_outputs, r_embed, clip)
+        level_outputs = self._down_encode(x, r_embed, clip, transformer_options=transformer_options)
+        x = self._up_decode(level_outputs, r_embed, clip, transformer_options=transformer_options)
         return self.clf(x)
 
     def update_weights_ema(self, src_model, beta=0.999):