andreas/ComfyUI
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Enable Runtime Selection of Attention Functions (#9639)

Browse Source 
* 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
This commit is contained in:
Jedrzej Kosinski
2025-09-12 15:07:38 -07:00
committed by GitHub
parent b149e2e1e3
commit d7f40442f9
26 changed files with 316 additions and 179 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
comfy/ldm/cosmos/blocks.py
View File

@@ -176,6 +176,7 @@ class Attention(nn.Module):
context=None,
mask=None,
rope_emb=None,
transformer_options={},
**kwargs,
):
"""
@@ -184,7 +185,7 @@ class Attention(nn.Module):
context (Optional[Tensor]): The key tensor of shape [B, Mk, K] or use x as context [self attention] if None
"""
q, k, v = self.cal_qkv(x, context, mask, rope_emb=rope_emb, **kwargs)
out = optimized_attention(q, k, v, self.heads, skip_reshape=True, mask=mask, skip_output_reshape=True)
out = optimized_attention(q, k, v, self.heads, skip_reshape=True, mask=mask, skip_output_reshape=True, transformer_options=transformer_options)
del q, k, v
out = rearrange(out, " b n s c -> s b (n c)")
return self.to_out(out)
@@ -546,6 +547,7 @@ class VideoAttn(nn.Module):
context: Optional[torch.Tensor] = None,
crossattn_mask: Optional[torch.Tensor] = None,
rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
transformer_options: Optional[dict] = {},
) -> torch.Tensor:
"""
Forward pass for video attention.
@@ -571,6 +573,7 @@ class VideoAttn(nn.Module):
context_M_B_D,
crossattn_mask,
rope_emb=rope_emb_L_1_1_D,
transformer_options=transformer_options,
)
x_T_H_W_B_D = rearrange(x_THW_B_D, "(t h w) b d -> t h w b d", h=H, w=W)
return x_T_H_W_B_D
@@ -665,6 +668,7 @@ class DITBuildingBlock(nn.Module):
crossattn_mask: Optional[torch.Tensor] = None,
rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
adaln_lora_B_3D: Optional[torch.Tensor] = None,
transformer_options: Optional[dict] = {},
) -> torch.Tensor:
"""
Forward pass for dynamically configured blocks with adaptive normalization.
@@ -702,6 +706,7 @@ class DITBuildingBlock(nn.Module):
adaln_norm_state(self.norm_state, x, scale_1_1_1_B_D, shift_1_1_1_B_D),
context=None,
rope_emb_L_1_1_D=rope_emb_L_1_1_D,
transformer_options=transformer_options,
)
elif self.block_type in ["cross_attn", "ca"]:
x = x + gate_1_1_1_B_D * self.block(
@@ -709,6 +714,7 @@ class DITBuildingBlock(nn.Module):
context=crossattn_emb,
crossattn_mask=crossattn_mask,
rope_emb_L_1_1_D=rope_emb_L_1_1_D,
transformer_options=transformer_options,
)
else:
raise ValueError(f"Unknown block type: {self.block_type}")
@@ -784,6 +790,7 @@ class GeneralDITTransformerBlock(nn.Module):
crossattn_mask: Optional[torch.Tensor] = None,
rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
adaln_lora_B_3D: Optional[torch.Tensor] = None,
transformer_options: Optional[dict] = {},
) -> torch.Tensor:
for block in self.blocks:
x = block(
@@ -793,5 +800,6 @@ class GeneralDITTransformerBlock(nn.Module):
crossattn_mask,
rope_emb_L_1_1_D=rope_emb_L_1_1_D,
adaln_lora_B_3D=adaln_lora_B_3D,
transformer_options=transformer_options,
)
return x