The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Fix case where text encoders where running on the CPU instead of GPU. (#11095)
comfyanonymousea17add3c621
diff --git a/comfy/sd.py b/comfy/sd.pyindex f9e5efab5..734bd2845 100644--- a/comfy/sd.py+++ b/comfy/sd.py@@ -193,6 +193,7 @@ class CLIP:self.cond_stage_model.set_clip_options({"projected_pooled": False})self.load_model()+ self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})all_hooks.reset()self.patcher.patch_hooks(None)if show_pbar:@@ -240,6 +241,7 @@ class CLIP:self.cond_stage_model.set_clip_options({"projected_pooled": False})self.load_model()+ self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})o = self.cond_stage_model.encode_token_weights(tokens)cond, pooled = o[:2]if return_dict:diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.pyindex 0fc9ab3db..503a51843 100644--- a/comfy/sd1_clip.py+++ b/comfy/sd1_clip.py@@ -147,6 +147,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):self.layer_norm_hidden_state = layer_norm_hidden_stateself.return_projected_pooled = return_projected_pooledself.return_attention_masks = return_attention_masks+ self.execution_device = Noneif layer == "hidden":assert layer_idx is not None@@ -163,6 +164,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):def set_clip_options(self, options):layer_idx = options.get("layer", self.layer_idx)self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled)+ self.execution_device = options.get("execution_device", self.execution_device)if isinstance(self.layer, list) or self.layer == "all":passelif layer_idx is None or abs(layer_idx) > self.num_layers:@@ -175,6 +177,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):self.layer = self.options_default[0]self.layer_idx = self.options_default[1]self.return_projected_pooled = self.options_default[2]+ self.execution_device = Nonedef process_tokens(self, tokens, device):end_token = self.special_tokens.get("end", None)@@ -258,7 +261,11 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens, embeds_infodef forward(self, tokens):- device = self.transformer.get_input_embeddings().weight.device+ if self.execution_device is None:+ device = self.transformer.get_input_embeddings().weight.device+ else:+ device = self.execution_device+embeds, attention_mask, num_tokens, embeds_info = self.process_tokens(tokens, device)attention_mask_model = None