Move Accelerator preparation into strategy

1 files changed, 21 insertions, 1 deletions

diff --git a/training/strategy/ti.py b/training/strategy/ti.py
index e922954..6a76f98 100644
--- a/training/strategy/ti.py
+++ b/training/strategy/ti.py
@@ -5,6 +5,7 @@ from contextlib import contextmanager, nullcontext
 from pathlib import Path
 
 import torch
+import torch.nn as nn
 from torch.utils.data import DataLoader
 
 from accelerate import Accelerator
@@ -94,7 +95,7 @@ def textual_inversion_strategy_callbacks(
             return nullcontext()
 
     def on_model():
-        return text_encoder
+        return text_encoder.text_model.embeddings.temp_token_embedding
 
     def on_prepare():
         text_encoder.text_model.embeddings.temp_token_embedding.requires_grad_(True)
@@ -163,6 +164,25 @@ def textual_inversion_strategy_callbacks(
     )
 
 
+def textual_inversion_prepare(
+    accelerator: Accelerator,
+    text_encoder: CLIPTextModel,
+    unet: UNet2DConditionModel,
+    *args
+):
+    weight_dtype = torch.float32
+    if accelerator.state.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif accelerator.state.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
+    prep = [text_encoder] + list(args)
+    text_encoder, optimizer, train_dataloader, val_dataloader, lr_scheduler = accelerator.prepare(*prep)
+    unet.to(accelerator.device, dtype=weight_dtype)
+    return text_encoder, unet, optimizer, train_dataloader, val_dataloader, lr_scheduler
+
+
 textual_inversion_strategy = TrainingStrategy(
     callbacks=textual_inversion_strategy_callbacks,
+    prepare=textual_inversion_prepare,
 )