diff --git a/unsloth/kernels/geglu.py b/unsloth/kernels/geglu.py
index d5a69aa67f..1ece87c080 100644
--- a/unsloth/kernels/geglu.py
+++ b/unsloth/kernels/geglu.py
@@ -45,9 +45,10 @@ pass
 def geglu_exact_forward_kernel(gate, up):
     batch, seq_len, hd = gate.shape
     n_elements = gate.numel()
-    out = torch.empty((batch, seq_len, hd), dtype = gate.dtype, device = "cuda:0")
+    device = gate.device
+    out = torch.empty((batch, seq_len, hd), dtype = gate.dtype, device = device)
     grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),)
-    with torch_cuda_device(gate.device):
+    with torch_cuda_device(device):
         _exact_forward_kernel[grid](gate, up, out, n_elements, BLOCK_SIZE = 1024,)
     return out
 pass
@@ -139,9 +140,10 @@ pass
 def geglu_approx_forward_kernel(gate, up):
     batch, seq_len, hd = gate.shape
     n_elements = gate.numel()
-    out = torch.empty((batch, seq_len, hd), dtype = gate.dtype, device = "cuda:0")
+    device = gate.device
+    out = torch.empty((batch, seq_len, hd), dtype = gate.dtype, device = device)
     grid = lambda meta: (triton.cdiv(n_elements, meta['BLOCK_SIZE']),)
-    with torch_cuda_device(gate.device):
+    with torch_cuda_device(device):
         _approx_forward_kernel[grid](gate, up, out, n_elements, BLOCK_SIZE = 1024,)
     return out
 pass
diff --git a/unsloth/kernels/utils.py b/unsloth/kernels/utils.py
index eb3a2e38cc..2c4edf334b 100644
--- a/unsloth/kernels/utils.py
+++ b/unsloth/kernels/utils.py
@@ -460,7 +460,9 @@ def matmul_lora(X, W, W_quant, A, B, s, out = None):
     else:
         reshape = False
     pass
-    
+
+    if X.device != W.device:
+        print(X.device, W.device, torch.cuda.current_device())
     out = torch_matmul(X, W, out = out)
     if W_quant is not None: del W
 
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index fe0627f8d7..7f475869c8 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -385,6 +385,7 @@ def LlamaAttention_fast_forward(
     head_dim   = self.head_dim
     assert(n_kv_heads * n_groups == n_heads)
 
+    print(hidden_states.device, torch.cuda.current_device())
     Q, K, V = self.apply_qkv(self, hidden_states)
     Q = Q.view(bsz, q_len, n_heads,    head_dim).transpose(1, 2)
     K = K.view(bsz, q_len, n_kv_heads, head_dim).transpose(1, 2)