diff --git a/unsloth/kernels/utils.py b/unsloth/kernels/utils.py
index 5b31d1e5ba..4f4ce41006 100644
--- a/unsloth/kernels/utils.py
+++ b/unsloth/kernels/utils.py
@@ -119,9 +119,7 @@ def fast_gemv(X, W, quant_state, out = None):
     # For fast X @ W where seq_len == 1
     # From https://github.com/TimDettmers/bitsandbytes/blob/main/bitsandbytes/functional.py#L1469
     bsz, q_len, hd = X.shape
-    device = X.device
     assert(q_len == 1)
-    assert(device == W.device)
 
     if type(quant_state) is not list:
         # https://github.com/TimDettmers/bitsandbytes/pull/763/files
@@ -144,7 +142,7 @@ def fast_gemv(X, W, quant_state, out = None):
     bout = shape[0]
 
     if out is None:
-        out = torch.empty((bsz, 1, bout,), dtype = dtype, device = device)
+        out = torch.empty((bsz, 1, bout,), dtype = dtype, device = "cuda")
     else:
         assert(out.shape == (bsz, 1, bout,))
     pass
@@ -162,7 +160,7 @@ def fast_gemv(X, W, quant_state, out = None):
     ldb = ctypes.c_int32(ldb)
     ldc = ctypes.c_int32(ldc)
 
-    df = torch.empty(absmax.shape, dtype = torch.float32, device = device)
+    df = torch.empty(absmax.shape, dtype = torch.float32, device = "cuda")
     cdequantize_blockwise_fp32(
         get_ptr(code2), get_ptr(absmax), get_ptr(absmax2), get_ptr(df),
         ctypes.c_int(blocksize2), ctypes.c_int(df.numel()),
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 9de65be04b..cc57f6ac75 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -638,7 +638,8 @@ def LlamaForCausalLM_fast_forward(
     )
 
     hidden_states = outputs[0]
-    if hidden_states.shape[0] == 1:
+    bsz, q_len, hd = hidden_states.shape
+    if bsz == 1 and q_len == 1:
         logits = torch.mv(self.lm_head.weight, hidden_states.ravel())
         logits = logits.unsqueeze(0).unsqueeze(0)
     else:
diff --git a/unsloth/models/mistral.py b/unsloth/models/mistral.py
index 08ee0fae93..c472b0f4a3 100644
--- a/unsloth/models/mistral.py
+++ b/unsloth/models/mistral.py
@@ -210,7 +210,8 @@ def MistralForCausalLM_fast_forward(
     )
 
     hidden_states = outputs[0]
-    if hidden_states.shape[0] == 1:
+    bsz, q_len, hd = hidden_states.shape
+    if bsz == 1 and q_len == 1:
         logits = torch.mv(self.lm_head.weight, hidden_states.ravel())
         logits = logits.unsqueeze(0).unsqueeze(0)
     else: