Fix

2 files changed, 8 insertions, 6 deletions

diff --git a/training/strategy/lora.py b/training/strategy/lora.py
index d51a2f3..6730dc9 100644
--- a/training/strategy/lora.py
+++ b/training/strategy/lora.py
@@ -85,15 +85,16 @@ def lora_strategy_callbacks(
         )
 
         if use_emb_decay:
-            return torch.stack([
+            params = [
                 p
                 for p in text_encoder.text_model.embeddings.token_override_embedding.params
                 if p.grad is not None
-            ])
+            ]
+            return torch.stack(params) if len(params) != 0 else None
 
     @torch.no_grad()
     def on_after_optimize(w, lr: float):
-        if use_emb_decay:
+        if use_emb_decay and w is not None:
             lambda_ = emb_decay * lr
 
             if lambda_ != 0:
diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index 9df160a..55e9934 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -107,18 +107,19 @@ def textual_inversion_strategy_callbacks(
     @torch.no_grad()
     def on_before_optimize(lr: float, epoch: int):
         if use_emb_decay:
-            return torch.stack([
+            params = [
                 p
                 for p in text_encoder.text_model.embeddings.token_override_embedding.params
                 if p.grad is not None
-            ])
+            ]
+            return torch.stack(params) if len(params) != 0 else None
 
     @torch.no_grad()
     def on_after_optimize(w, lr: float):
         if ema_embeddings is not None:
             ema_embeddings.step(text_encoder.text_model.embeddings.token_override_embedding.params.parameters())
 
-        if use_emb_decay:
+        if use_emb_decay and w is not None:
             lambda_ = emb_decay * lr
 
             if lambda_ != 0: