add fp8 version

convert_fp8_simple.py

@@ -0,0 +1,61 @@
+import torch
+from safetensors import safe_open
+from safetensors.torch import save_file
+from tqdm import tqdm
+import argparse
+import os
+
+def convert_to_fp8_simple(input_path, output_path=None):
+    """
+    LoRAファイルを通常のFP8（非スケール版）に変換
+    """
+    if output_path is None:
+        # 出力ファイル名を自動生成
+        base_name = os.path.splitext(input_path)[0]
+        output_path = f"{base_name}_fp8.safetensors"
+    
+    print(f"入力ファイル: {input_path}")
+    print(f"出力ファイル: {output_path}")
+    
+    # メタデータを読み込む
+    metadata = {}
+    with safe_open(input_path, framework="pt", device="cpu") as f:
+        if f.metadata() is not None:
+            metadata = f.metadata()
+    
+    # テンソルを変換
+    converted_tensors = {}
+    print("テンソルを変換中...")
+    
+    with safe_open(input_path, framework="pt", device="cpu") as f:
+        for key in tqdm(f.keys(), desc="変換中"):
+            tensor = f.get_tensor(key)
+            # シンプルにFP8に変換
+            converted_tensors[key] = tensor.to(torch.float8_e4m3fn)
+    
+    # メタデータに形式情報を追加
+    metadata["format"] = "pt"
+    metadata["fp8_type"] = "simple"  # fp8-scaledではないことを明示
+    
+    # 保存
+    print(f"変換したモデルを保存中: {output_path}")
+    save_file(converted_tensors, output_path, metadata=metadata)
+    
+    # ファイルサイズの比較
+    original_size = os.path.getsize(input_path) / (1024**3)  # GB
+    converted_size = os.path.getsize(output_path) / (1024**3)  # GB
+    
+    print(f"\n✅ 変換完了！")
+    print(f"元のサイズ: {original_size:.2f} GB")
+    print(f"変換後: {converted_size:.2f} GB")
+    print(f"削減率: {(1 - converted_size/original_size)*100:.1f}%")
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="LoRAファイルを通常のFP8に変換")
+    parser.add_argument("input", type=str, help="入力LoRAファイルのパス")
+    parser.add_argument("--output", "-o", type=str, default=None, 
+                       help="出力ファイルのパス（省略時は自動生成）")
+    
+    args = parser.parse_args()
+    
+    convert_to_fp8_simple(args.input, args.output)

wan2.1/ChronoEdit_480_extracted_lora_rank2048_adaptive_fro_0.95_fp8.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cbc1245f0a6224b401eae00f5d6a8fe6356a65c0eb21286bd9c108aadf5d51a8
+size 10563734344

wan2.1/README.md

@@ -1,6 +1,7 @@
 # **LoRA Files**
 ## ChronoEdit
-**ChronoEdit_480_extracted_lora_rank2048_adaptive_fro_0.95_fp16.safetensors** <BR>
+**ChronoEdit_480_extracted_lora_rank2048_adaptive_fro_0.95_fp16.safetensors**
+**ChronoEdit_480_extracted_lora_rank2048_adaptive_fro_0.95_fp8.safetensors**  <BR>
 This is a LoRA created using the LoraExtractKJ node, representing the difference between the weights of ChronoEdit and wan2.1_i2v_480p_14B_fp16.safetensors. (Accuracy is approximately 80-95% at Rank 2048)
 It is very large and therefore not suitable for typical LoRA use, but because it has been converted into a LoRA, it can be applied to T2V models, and I have confirmed that it can be used when using VACE. (There may be some degradation in quality.)