2 files changed, 15 insertions, 11 deletions

diff --git a/training/strategy/dreambooth.py b/training/strategy/dreambooth.py
index fcf5c0d..0290327 100644
--- a/training/strategy/dreambooth.py
+++ b/training/strategy/dreambooth.py
@@ -115,14 +115,13 @@ def dreambooth_strategy_callbacks(
             yield
 
     def on_before_optimize(lr: float, epoch: int):
-        if accelerator.sync_gradients:
-            params_to_clip = [unet.parameters()]
-            if epoch < train_text_encoder_epochs:
-                params_to_clip.append(text_encoder.parameters())
-            accelerator.clip_grad_norm_(itertools.chain(*params_to_clip), max_grad_norm)
+        params_to_clip = [unet.parameters()]
+        if epoch < train_text_encoder_epochs:
+            params_to_clip.append(text_encoder.parameters())
+        accelerator.clip_grad_norm_(itertools.chain(*params_to_clip), max_grad_norm)
 
     @torch.no_grad()
-    def on_after_optimize(lr: float):
+    def on_after_optimize(_, lr: float):
         if ema_unet is not None:
             ema_unet.step(unet.parameters())
 
diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index 09beec4..732cd74 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -116,6 +116,15 @@ def textual_inversion_strategy_callbacks(
     @torch.no_grad()
     def on_before_optimize(lr: float, epoch: int):
         if use_emb_decay:
+            w = text_encoder.text_model.embeddings.temp_token_embedding.weight
+            return torch.all(w.grad == 0, dim=1)
+
+    @torch.no_grad()
+    def on_after_optimize(zero_ids, lr: float):
+        if ema_embeddings is not None:
+            ema_embeddings.step(text_encoder.text_model.embeddings.temp_token_embedding.parameters())
+
+        if use_emb_decay:
             lambda_ = emb_decay * lr
 
             if lambda_ != 0:
@@ -123,15 +132,11 @@ def textual_inversion_strategy_callbacks(
 
                 mask = torch.zeros(w.size(0), dtype=torch.bool)
                 mask[text_encoder.text_model.embeddings.temp_token_ids] = True
-                mask[torch.all(w.grad == 0, dim=1)] = False
+                mask[zero_ids] = False
 
                 norm = w[mask, :].norm(dim=-1, keepdim=True)
                 w[mask].add_((w[mask] / norm.clamp_min(1e-12)) * lambda_ * (emb_decay_target - norm))
 
-    def on_after_optimize(lr: float):
-        if ema_embeddings is not None:
-            ema_embeddings.step(text_encoder.text_model.embeddings.temp_token_embedding.parameters())
-
     def on_log():
         if ema_embeddings is not None:
             return {"ema_decay": ema_embeddings.decay}