diff --git a/RWKV-v7/rwkv_v8_rc00_demo.py b/RWKV-v7/rwkv_v8_rc00_demo.py
index 186426eb4..94fb2b660 100644
--- a/RWKV-v7/rwkv_v8_rc00_demo.py
+++ b/RWKV-v7/rwkv_v8_rc00_demo.py
@@ -106,6 +106,7 @@ def __init__(self, args):
         z['blocks.0.att.v0'] = z['blocks.0.att.a0'] # actually ignored
         z['blocks.0.att.v1'] = z['blocks.0.att.a1'] # actually ignored
         z['blocks.0.att.v2'] = z['blocks.0.att.a2'] # actually ignored
+        self.deepembs = {}
 
     def forward(self, idx, state, full_output=False):
         if state == None:
@@ -178,7 +179,15 @@ def forward_seq(self, idx:List[int], state:List[torch.Tensor], full_output:bool=
 
                 xx = F.layer_norm(x, (self.n_embd,), weight=z[bbb+'ln2.weight'], bias=z[bbb+'ln2.bias'])
 
-                xx, state[i*3+2] = RWKV_x080_CMix_seq(xx, state[i*3+2], z[ffn+'x_k'], z[ffn+'key.weight'], z[ffn+'value.weight'], z[ffn+'enn.weight'][idx])
+                # Disk space offloading
+                if not (ffn+'enn.weight' in self.deepembs):
+                    z[ffn+'enn.weight'].numpy().tofile(ffn+'enn.weight_storage'+'.pt')
+                    emb_size = z[ffn+'enn.weight'].numel()
+                    emb_dtype = z[ffn+'enn.weight'].dtype
+                    z[ffn+'enn.weight'] = torch.from_file(ffn+'enn.weight_storage'+'.pt', size=emb_size, dtype=emb_dtype, device='cpu')
+                    self.deepembs[ffn+'enn.weight'] = True
+                
+                xx, state[i*3+2] = RWKV_x080_CMix_seq(xx, state[i*3+2], z[ffn+'x_k'], z[ffn+'key.weight'], z[ffn+'value.weight'], z[ffn+'enn.weight'][idx].cuda())
                 x = x + xx
             
             if not full_output: x = x[-1,:]