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

17
comfy/ldm/cosmos/predict2.py
View File

@@ -44,7 +44,7 @@ class GPT2FeedForward(nn.Module):
return x
def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H_D: torch.Tensor) -> torch.Tensor:
def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H_D: torch.Tensor, transformer_options: Optional[dict] = {}) -> torch.Tensor:
"""Computes multi-head attention using PyTorch's native implementation.
This function provides a PyTorch backend alternative to Transformer Engine's attention operation.
@@ -71,7 +71,7 @@ def torch_attention_op(q_B_S_H_D: torch.Tensor, k_B_S_H_D: torch.Tensor, v_B_S_H
q_B_H_S_D = rearrange(q_B_S_H_D, "b ... h k -> b h ... k").view(in_q_shape[0], in_q_shape[-2], -1, in_q_shape[-1])
k_B_H_S_D = rearrange(k_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
v_B_H_S_D = rearrange(v_B_S_H_D, "b ... h v -> b h ... v").view(in_k_shape[0], in_k_shape[-2], -1, in_k_shape[-1])
return optimized_attention(q_B_H_S_D, k_B_H_S_D, v_B_H_S_D, in_q_shape[-2], skip_reshape=True)
return optimized_attention(q_B_H_S_D, k_B_H_S_D, v_B_H_S_D, in_q_shape[-2], skip_reshape=True, transformer_options=transformer_options)
class Attention(nn.Module):
@@ -180,8 +180,8 @@ class Attention(nn.Module):
return q, k, v
def compute_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
result = self.attn_op(q, k, v) # [B, S, H, D]
def compute_attention(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, transformer_options: Optional[dict] = {}) -> torch.Tensor:
result = self.attn_op(q, k, v, transformer_options=transformer_options) # [B, S, H, D]
return self.output_dropout(self.output_proj(result))
def forward(
@@ -189,6 +189,7 @@ class Attention(nn.Module):
x: torch.Tensor,
context: Optional[torch.Tensor] = None,
rope_emb: Optional[torch.Tensor] = None,
transformer_options: Optional[dict] = {},
) -> torch.Tensor:
"""
Args:
@@ -196,7 +197,7 @@ 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.compute_qkv(x, context, rope_emb=rope_emb)
return self.compute_attention(q, k, v)
return self.compute_attention(q, k, v, transformer_options=transformer_options)
class Timesteps(nn.Module):
@@ -459,6 +460,7 @@ class Block(nn.Module):
rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
adaln_lora_B_T_3D: Optional[torch.Tensor] = None,
extra_per_block_pos_emb: Optional[torch.Tensor] = None,
transformer_options: Optional[dict] = {},
) -> torch.Tensor:
if extra_per_block_pos_emb is not None:
x_B_T_H_W_D = x_B_T_H_W_D + extra_per_block_pos_emb
@@ -512,6 +514,7 @@ class Block(nn.Module):
rearrange(normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
None,
rope_emb=rope_emb_L_1_1_D,
transformer_options=transformer_options,
),
"b (t h w) d -> b t h w d",
t=T,
@@ -525,6 +528,7 @@ class Block(nn.Module):
layer_norm_cross_attn: Callable,
_scale_cross_attn_B_T_1_1_D: torch.Tensor,
_shift_cross_attn_B_T_1_1_D: torch.Tensor,
transformer_options: Optional[dict] = {},
) -> torch.Tensor:
_normalized_x_B_T_H_W_D = _fn(
_x_B_T_H_W_D, layer_norm_cross_attn, _scale_cross_attn_B_T_1_1_D, _shift_cross_attn_B_T_1_1_D
@@ -534,6 +538,7 @@ class Block(nn.Module):
rearrange(_normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
crossattn_emb,
rope_emb=rope_emb_L_1_1_D,
transformer_options=transformer_options,
),
"b (t h w) d -> b t h w d",
t=T,
@@ -547,6 +552,7 @@ class Block(nn.Module):
self.layer_norm_cross_attn,
scale_cross_attn_B_T_1_1_D,
shift_cross_attn_B_T_1_1_D,
transformer_options=transformer_options,
)
x_B_T_H_W_D = result_B_T_H_W_D * gate_cross_attn_B_T_1_1_D + x_B_T_H_W_D
@@ -865,6 +871,7 @@ class MiniTrainDIT(nn.Module):
"rope_emb_L_1_1_D": rope_emb_L_1_1_D.unsqueeze(1).unsqueeze(0),
"adaln_lora_B_T_3D": adaln_lora_B_T_3D,
"extra_per_block_pos_emb": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
"transformer_options": kwargs.get("transformer_options", {}),
}
for block in self.blocks:
x_B_T_H_W_D = block(