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 issue where autocast fp32 CLIP gave different results from regular.
comfyanonymousfb3b7282034a
diff --git a/comfy/sd1_clip.py b/comfy/sd1_clip.pyindex 477d5c309..b84a38490 100644--- a/comfy/sd1_clip.py+++ b/comfy/sd1_clip.py@@ -60,6 +60,9 @@ class SD1ClipModel(torch.nn.Module, ClipTokenWeightEncoder):if dtype is not None:self.transformer.to(dtype)+ self.transformer.text_model.embeddings.token_embedding.to(torch.float32)+ self.transformer.text_model.embeddings.position_embedding.to(torch.float32)+self.max_length = max_lengthif freeze:self.freeze()@@ -138,7 +141,7 @@ class SD1ClipModel(torch.nn.Module, ClipTokenWeightEncoder):tokens = self.set_up_textual_embeddings(tokens, backup_embeds)tokens = torch.LongTensor(tokens).to(device)- if backup_embeds.weight.dtype != torch.float32:+ if self.transformer.text_model.final_layer_norm.weight.dtype != torch.float32:precision_scope = torch.autocastelse:precision_scope = lambda a, b: contextlib.nullcontext(a)