Only import mamba libraries if cuda available

This should remove the cpu only inference errors.

variations/attention_variations.py

@@ -9,9 +9,11 @@
 from quantization.quant_utils import set_variant, create_activation_buffers
 from variations.softmax_variations import softmax_dictionary
 from variations.position_encoding_variations import QuantizedEmbedding, RotaryEmbedding, SymmetricalOverlapAngularPositions, FIRE
-from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
 
-from mamba_ssm.ops.selective_scan_interface import selective_scan_fn
+# Mamba related imports
+if torch.cuda.is_available():
+    from causal_conv1d import causal_conv1d_fn, causal_conv1d_update
+    from mamba_ssm.ops.selective_scan_interface import selective_scan_fn
 
 class CausalSelfAttention(nn.Module):
     def __init__(self, config, fire_pos_enc=None):