Fixed the accuracy fluctuation bug

scripts/run_wan_t2v_enhancer.sh

@@ -25,7 +25,7 @@ fi
 export TOKENIZERS_PARALLELISM=false
 
 export PYTHONPATH=${lightx2v_path}:$PYTHONPATH
-
+export DTYPE=BF16
 export ENABLE_PROFILING_DEBUG=true
 export ENABLE_GRAPH_MODE=false
 
@@ -36,5 +36,5 @@ python -m lightx2v.infer \
 --config_json ${lightx2v_path}/configs/wan_t2v_enhancer.json \
 --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage." \
 --use_prompt_enhancer \
---negative_prompt 色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走 \
+--negative_prompt "镜头晃动，色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" \
 --save_video_path ${lightx2v_path}/save_results/output_lightx2v_wan_t2v.mp4

scripts/run_wan_t2v_sparge.sh

@@ -25,7 +25,7 @@ fi
 export TOKENIZERS_PARALLELISM=false
 
 export PYTHONPATH=${lightx2v_path}:$PYTHONPATH
-
+export DTYPE=BF16
 export ENABLE_PROFILING_DEBUG=true
 export ENABLE_GRAPH_MODE=false
 
@@ -35,5 +35,5 @@ python -m lightx2v.infer \
 --model_path $model_path \
 --config_json ${lightx2v_path}/configs/wan_t2v_sparge.json \
 --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage." \
---negative_prompt 色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走 \
+--negative_prompt "镜头晃动，色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" \
 --save_video_path ${lightx2v_path}/save_results/output_lightx2v_wan_t2v.mp4

scripts/run_wan_t2v_tea.sh

@@ -24,7 +24,7 @@ fi
 export TOKENIZERS_PARALLELISM=false
 
 export PYTHONPATH=${lightx2v_path}:$PYTHONPATH
-
+export DTYPE=BF16
 export ENABLE_PROFILING_DEBUG=true
 export ENABLE_GRAPH_MODE=false
 
@@ -34,5 +34,5 @@ python -m lightx2v.infer \
 --model_path $model_path \
 --config_json ${lightx2v_path}/configs/caching/wan_t2v_Tea.json \
 --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage." \
---negative_prompt 色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走 \
+--negative_prompt "镜头晃动，色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" \
 --save_video_path ${lightx2v_path}/save_results/output_lightx2v_wan_t2v_tea.mp4

scripts/start_api_with_dist_inference.sh

@@ -26,6 +26,7 @@ export TOKENIZERS_PARALLELISM=false
 export PYTHONPATH=${lightx2v_path}:$PYTHONPATH
 export ENABLE_PROFILING_DEBUG=true
 export ENABLE_GRAPH_MODE=false
+export DTYPE=BF16
 
 echo "=========================================="
 echo "启动分布式推理API服务器"

scripts/start_multi_servers.sh

@@ -31,6 +31,7 @@ export TOKENIZERS_PARALLELISM=false
 export PYTHONPATH=${lightx2v_path}:$PYTHONPATH
 export ENABLE_PROFILING_DEBUG=true
 export ENABLE_GRAPH_MODE=false
+export DTYPE=BF16
 
 # Start multiple servers
 python -m lightx2v.api_multi_servers \

scripts/start_server.sh

@@ -27,6 +27,7 @@ export PYTHONPATH=${lightx2v_path}:$PYTHONPATH
 
 export ENABLE_PROFILING_DEBUG=true
 export ENABLE_GRAPH_MODE=false
+export DTYPE=BF16
 
 python -m lightx2v.api_server \
 --model_cls wan2.1 \

tools/convert/converter.py

@@ -339,7 +339,6 @@ def quantize_model(
     weights,
     w_bit=8,
     target_keys=["attn", "ffn"],
-    min_params=1e6,
     key_idx=2,
     ignore_key=None,
     dtype=torch.int8,
@@ -351,7 +350,6 @@ def quantize_model(
         weights: Model state dictionary
         w_bit: Quantization bit width
         target_keys: List of module names to quantize
-        min_params: Minimum parameter count to process tensor
 
     Returns:
         Modified state dictionary with quantized weights and scales
@@ -371,7 +369,7 @@ def quantize_model(
             tensor = weights[key]
 
             # Skip non-tensors, small tensors, and non-2D tensors
-            if not isinstance(tensor, torch.Tensor) or tensor.numel() < min_params or tensor.dim() != 2:
+            if not isinstance(tensor, torch.Tensor) or tensor.dim() != 2:
                 continue
 
             # Check if key matches target modules
@@ -442,7 +440,6 @@ def convert_weights(args):
             converted_weights,
             w_bit=args.bits,
             target_keys=args.target_keys,
-            min_params=args.min_params,
             key_idx=args.key_idx,
             ignore_key=args.ignore_key,
             dtype=args.dtype,
@@ -575,12 +572,6 @@ def main():
     # Quantization
     parser.add_argument("--quantized", action="store_true")
     parser.add_argument("--bits", type=int, default=8, choices=[8], help="Quantization bit width")
-    parser.add_argument(
-        "--min_params",
-        type=int,
-        default=1000000,
-        help="Minimum parameters to consider for quantization",
-    )
     parser.add_argument(
         "--device",
         type=str,