The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Support hypernetwork with mish activation function and layer norm.
comfyanonymousf8caa24bcc6a
diff --git a/comfy_extras/nodes_hypernetwork.py b/comfy_extras/nodes_hypernetwork.pyindex d16c49aeb..f692945a8 100644--- a/comfy_extras/nodes_hypernetwork.py+++ b/comfy_extras/nodes_hypernetwork.py@@ -19,6 +19,7 @@ def load_hypernetwork_patch(path, strength):"tanh": torch.nn.Tanh,"sigmoid": torch.nn.Sigmoid,"softsign": torch.nn.Softsign,+ "mish": torch.nn.Mish,}if activation_func not in valid_activation:@@ -42,7 +43,8 @@ def load_hypernetwork_patch(path, strength):linears = list(map(lambda a: a[:-len(".weight")], linears))layers = []- for i in range(len(linears)):+ i = 0+ while i < len(linears):lin_name = linears[i]last_layer = (i == (len(linears) - 1))penultimate_layer = (i == (len(linears) - 2))@@ -56,10 +58,17 @@ def load_hypernetwork_patch(path, strength):if (not last_layer) or (activate_output):layers.append(valid_activation[activation_func]())if is_layer_norm:- layers.append(torch.nn.LayerNorm(lin_weight.shape[0]))+ i += 1+ ln_name = linears[i]+ ln_weight = attn_weights['{}.weight'.format(ln_name)]+ ln_bias = attn_weights['{}.bias'.format(ln_name)]+ ln = torch.nn.LayerNorm(ln_weight.shape[0])+ ln.load_state_dict({"weight": ln_weight, "bias": ln_bias})+ layers.append(ln)if use_dropout:if (not last_layer) and (not penultimate_layer or last_layer_dropout):layers.append(torch.nn.Dropout(p=0.3))+ i += 1output.append(torch.nn.Sequential(*layers))out[dim] = torch.nn.ModuleList(output)