diff --git a/ml/src/liquid/candle_cfc.rs b/ml/src/liquid/candle_cfc.rs
index 6253a78cc..d83d57673 100644
--- a/ml/src/liquid/candle_cfc.rs
+++ b/ml/src/liquid/candle_cfc.rs
@@ -10,7 +10,7 @@ use serde::{Deserialize, Serialize};
 use crate::MLError;
 
 /// Device configuration for CfC training
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
 pub enum DeviceConfig {
     Cpu,
     Cuda(usize),
@@ -21,8 +21,11 @@ impl DeviceConfig {
     pub fn resolve(&self) -> Result<Device, MLError> {
         match self {
             DeviceConfig::Cpu => Ok(Device::Cpu),
-            DeviceConfig::Cuda(id) => Device::cuda_if_available(*id).map_err(|e| {
-                MLError::ConfigurationError(format!("CUDA device {} error: {}", id, e))
+            DeviceConfig::Cuda(id) => Device::new_cuda(*id).map_err(|e| {
+                MLError::ConfigurationError(format!(
+                    "CUDA device {} required but unavailable: {}",
+                    id, e
+                ))
             }),
             DeviceConfig::Auto => Device::cuda_if_available(0)
                 .map_err(|e| MLError::ConfigurationError(format!("Auto device error: {}", e))),
@@ -35,7 +38,7 @@ impl DeviceConfig {
 /// This is distinct from `cells::CfCConfig` which configures the FixedPoint inference path.
 /// `CfCTrainConfig` configures the Candle-based differentiable training path with
 /// gradient descent, dropout, and market regime adaptation.
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
 pub struct CfCTrainConfig {
     pub input_size: usize,
     pub hidden_size: usize,