update config and fix wan22 lora bug (#396)

configs/wan22/wan_moe_i2v_distil_with_lora.json

+{
+    "infer_steps": 4,
+    "target_video_length": 81,
+    "text_len": 512,
+    "target_height": 720,
+    "target_width": 1280,
+    "self_attn_1_type": "sage_attn2",
+    "cross_attn_1_type": "sage_attn2",
+    "cross_attn_2_type": "sage_attn2",
+    "sample_guide_scale": [
+        3.5,
+        3.5
+    ],
+    "sample_shift": 5.0,
+    "enable_cfg": false,
+    "cpu_offload": true,
+    "offload_granularity": "block",
+    "t5_cpu_offload": false,
+    "vae_cpu_offload": false,
+    "use_image_encoder": false,
+    "boundary_step_index": 2,
+    "denoising_step_list": [
+        1000,
+        750,
+        500,
+        250
+    ],
+    "lora_configs": [
+        {
+            "name": "high_noise_model",
+            "path": "lightx2v/Wan2.2-Distill-Loras/wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_1022.safetensors",
+            "strength": 1.0
+        },
+        {
+            "name": "low_noise_model",
+            "path": "lightx2v/Wan2.2-Distill-Loras/wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_1022.safetensors",
+            "strength": 1.0
+        }
+    ]
+}

configs/wan22/wan_moe_i2v_distill_quant.json

@@ -28,5 +28,5 @@
     "dit_quantized": true,
     "dit_quant_scheme": "fp8-sgl",
     "t5_quantized": true,
-    "t5_quant_scheme": "fp8"
+    "t5_quant_scheme": "fp8-sgl"
 }

lightx2v/models/networks/wan/model.py

@@ -107,7 +107,7 @@ class WanModel(CompiledMethodsMixin):
         return False
 
     def _should_init_empty_model(self):
-        if self.config.get("lora_configs") and self.config.lora_configs:
+        if self.config.get("lora_configs") and self.config["lora_configs"]:
             if self.model_type in ["wan2.1"]:
                 return True
             if self.model_type in ["wan2.2_moe_high_noise"]:

tools/convert/converter.py

@@ -715,6 +715,7 @@ def convert_weights(args):
                 index["metadata"]["total_size"] += os.path.getsize(output_path)
 
         # Save index file
+        if not args.single_file:
             index_path = os.path.join(args.output, "diffusion_pytorch_model.safetensors.index.json")
             with open(index_path, "w", encoding="utf-8") as f:
                 json.dump(index, f, indent=2)
@@ -783,7 +784,7 @@ def main():
     parser.add_argument(
         "--device",
         type=str,
-        default="cpu",
+        default="cuda",
         help="Device to use for quantization (cpu/cuda)",
     )
     parser.add_argument(

tools/convert/readme.md

@@ -48,7 +48,7 @@ A powerful model weight conversion tool that supports format conversion, quantiz
   - `torch.bfloat16`: BF16
   - `torch.float16`: FP16
   - `torch.float32`: FP32 (default)
-- `--device`: Device for quantization, `cpu` (default) or `cuda`
+- `--device`: Device for quantization, `cpu` or `cuda` (default)
 - `--comfyui_mode`: ComfyUI compatible mode
 - `--full_quantized`: Full quantization mode (effective in ComfyUI mode)
 
@@ -391,7 +391,8 @@ python converter.py \
     --source /path/to/model.pth \
     --output /path/to/output \
     --output_ext .safetensors \
-    --output_name model
+    --output_name model \
+    --single_file
 ```
 
 #### 4.2 Multiple .safetensors → Single File

tools/convert/readme_zh.md

@@ -48,7 +48,7 @@
   - `torch.bfloat16`: BF16
   - `torch.float16`: FP16
   - `torch.float32`: FP32（默认）
-- `--device`: 量化使用的设备，可选 `cpu`（默认）或 `cuda`
+- `--device`: 量化使用的设备，可选 `cpu` 或 `cuda`（默认）
 - `--comfyui_mode`: ComfyUI 兼容模式
 - `--full_quantized`: 全量化模式（ComfyUI 模式下有效）
 
@@ -391,7 +391,8 @@ python converter.py \
     --source /path/to/model.pth \
     --output /path/to/output \
     --output_ext .safetensors \
-    --output_name model
+    --output_name model \
+    --single_file
 ```
 
 #### 4.2 多个 .safetensors → 单文件