The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Disable sage attention in stable audio dit and VAE. (#14148)
comfyanonymous85a403d1ea86
diff --git a/comfy/ldm/audio/dit.py b/comfy/ldm/audio/dit.pyindex a6258b755..c28be5b49 100644--- a/comfy/ldm/audio/dit.py+++ b/comfy/ldm/audio/dit.py@@ -433,11 +433,11 @@ class Attention(nn.Module):if self.differential:q, q_diff = q.unbind(dim=1)k, k_diff = k.unbind(dim=1)- out = optimized_attention(q, k, v, h, skip_reshape=True, transformer_options=transformer_options)- out_diff = optimized_attention(q_diff, k_diff, v, h, skip_reshape=True, transformer_options=transformer_options)+ out = optimized_attention(q, k, v, h, skip_reshape=True, low_precision_attention=False, transformer_options=transformer_options)+ out_diff = optimized_attention(q_diff, k_diff, v, h, skip_reshape=True, low_precision_attention=False, transformer_options=transformer_options)out = out - out_diffelse:- out = optimized_attention(q, k, v, h, skip_reshape=True, transformer_options=transformer_options)+ out = optimized_attention(q, k, v, h, skip_reshape=True, low_precision_attention=False, transformer_options=transformer_options)out = self.to_out(out)diff --git a/comfy/ldm/audio/vae_sa3.py b/comfy/ldm/audio/vae_sa3.pyindex 276846444..8be36d6ee 100644--- a/comfy/ldm/audio/vae_sa3.py+++ b/comfy/ldm/audio/vae_sa3.py@@ -138,11 +138,11 @@ class Attention(nn.Module):k_diff = _apply_rotary_pos_emb(k_diff.float(), freqs).to(k_dtype)if self.differential:- out = (optimized_attention(q, k, v, h, mask=mask, skip_reshape=True)- - optimized_attention(q_diff, k_diff, v, h, mask=mask, skip_reshape=True))+ out = (optimized_attention(q, k, v, h, mask=mask, skip_reshape=True, low_precision_attention=False)+ - optimized_attention(q_diff, k_diff, v, h, mask=mask, skip_reshape=True, low_precision_attention=False))del q, k, v, q_diff, k_diffelse:- out = optimized_attention(q, k, v, h, mask=mask, skip_reshape=True)+ out = optimized_attention(q, k, v, h, mask=mask, skip_reshape=True, low_precision_attention=False)del q, k, vreturn self.to_out(out)