possible fix for NaNs when using LoRA and fp8

2024-10-12 02:24:53 +03:00
parent ac5daa7148
commit 276b3b86d9
1 changed files with 4 additions and 1 deletions
@@ -372,7 +372,10 @@ class DownloadAndLoadCogVideoModel:
                    if "patch_embed" not in name:
                        param.data = param.data.to(torch.float8_e4m3fn)
            else:
-                transformer.to(torch.float8_e4m3fn)
+                #transformer.to(torch.float8_e4m3fn)
+                for name, param in transformer.named_parameters():
+                    if "lora" not in name:
+                        param.data = param.data.to(torch.float8_e4m3fn)
        
            if fp8_transformer == "fastmode":
                from .fp8_optimization import convert_fp8_linear