diff --git a/fastvideo/demo/gradio_web_demo.py b/fastvideo/demo/gradio_web_demo.py
index 2c663c563..b294532d4 100644
--- a/fastvideo/demo/gradio_web_demo.py
+++ b/fastvideo/demo/gradio_web_demo.py
@@ -40,7 +40,7 @@ def load_model(args):
     
     mochi_genmo = True
     if mochi_genmo:
-        model_path = "/root/weights/dit.safetensors"
+        model_path = "/root/fastmochi_genmo/dit.safetensors"
         state_dcit = load_file(model_path)
         transformer = AsymmDiTJoint()
         transformer.load_state_dict(state_dcit)
diff --git a/fastvideo/models/mochi_hf/modeling_mochi.py b/fastvideo/models/mochi_hf/modeling_mochi.py
index e904fd8ba..9d5ff859c 100644
--- a/fastvideo/models/mochi_hf/modeling_mochi.py
+++ b/fastvideo/models/mochi_hf/modeling_mochi.py
@@ -55,11 +55,7 @@ from flash_attn import flash_attn_varlen_qkvpacked_func
 from flash_attn.bert_padding import pad_input, unpad_input
 
 from liger_kernel.ops.swiglu import LigerSiLUMulFunction
-<<<<<<< HEAD
 from sageattention import sageattn
-=======
-
->>>>>>> de1e8d868e5b7e2a462f975a03a6043c01f2a8ba
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
 
 
@@ -188,14 +184,11 @@ class MochiAttention(nn.Module):
             attention_mask=attention_mask,
             **kwargs,
         )
-<<<<<<< HEAD
             
 def sage_attn_processor(query, key, value, attn_mask):
     query = query.squeeze(2)  # B, S, H, D
     key = key.squeeze(2)
     value = value.squeeze(2)
-=======
->>>>>>> de1e8d868e5b7e2a462f975a03a6043c01f2a8ba
 
     B = query.shape[0]
     assert B == 2, "This implementation expects batch size 2"
@@ -330,17 +323,11 @@ class MochiAttnProcessor2_0:
 
         attn_mask = encoder_attention_mask[:, :].bool()
         attn_mask = F.pad(attn_mask, (sequence_length, 0), value=True)
-<<<<<<< HEAD
         sage_attn = False
         if sage_attn:
             hidden_states = sage_attn_processor(query, key, value, attn_mask)
         else:
             hidden_states = flash_attn_no_pad(qkv, attn_mask, causal=False, dropout_p=0.0, softmax_scale=None)
-=======
-        hidden_states = flash_attn_no_pad(
-            qkv, attn_mask, causal=False, dropout_p=0.0, softmax_scale=None
-        )
->>>>>>> de1e8d868e5b7e2a462f975a03a6043c01f2a8ba
 
         # hidden_states = F.scaled_dot_product_attention(query, key, value, attn_mask = None, dropout_p=0.0, is_causal=False)
 
diff --git a/fastvideo/sample/sample_t2v_mochi.py b/fastvideo/sample/sample_t2v_mochi.py
index ac93fd51d..fad2dcb38 100644
--- a/fastvideo/sample/sample_t2v_mochi.py
+++ b/fastvideo/sample/sample_t2v_mochi.py
@@ -22,7 +22,8 @@ from typing import Dict
 from diffusers import FlowMatchEulerDiscreteScheduler
 from diffusers.utils import convert_unet_state_dict_to_peft
 from fastvideo.distill.solver import PCMFMScheduler
-
+from fastvideo.models.mochi_genmo.mochi_preview.dit.joint_model.asymm_models_joint import AsymmDiTJoint
+from safetensors.torch import load_file
 
 def initialize_distributed():
     local_rank = int(os.getenv("RANK", 0))
@@ -53,12 +54,22 @@ def main(args):
             args.linear_threshold,
             args.linear_range,
         )
-    if args.transformer_path is not None:
-        transformer = MochiTransformer3DModel.from_pretrained(args.transformer_path)
+
+    mochi_genmo = True
+    if mochi_genmo:
+        model_path = "/root/fastmochi_genmo/dit.safetensors"
+        state_dcit = load_file(model_path)
+        transformer = AsymmDiTJoint()
+        transformer.load_state_dict(state_dcit)
+        # from IPython import embed
+        # embed()
+        transformer.config.in_channels = 12
+        print("load gennmo mochi successfully")
     else:
-        transformer = MochiTransformer3DModel.from_pretrained(
-            args.model_path, subfolder="transformer/"
-        )
+        if args.transformer_path:
+            transformer = MochiTransformer3DModel.from_pretrained(args.transformer_path)
+        else:
+            transformer = MochiTransformer3DModel.from_pretrained(args.model_path, subfolder='transformer/')
 
     pipe = MochiPipeline.from_pretrained(
         args.model_path, transformer=transformer, scheduler=scheduler