diff --git a/Cargo.lock b/Cargo.lock
index 2f2d5456a..a011f1331 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1432,7 +1432,7 @@ dependencies = [
  "bitflags 2.9.4",
  "cexpr",
  "clang-sys",
- "itertools 0.12.1",
+ "itertools 0.10.5",
  "log",
  "prettyplease",
  "proc-macro2",
diff --git a/ml/Cargo.toml b/ml/Cargo.toml
index eabf58dff..6e88e939d 100644
--- a/ml/Cargo.toml
+++ b/ml/Cargo.toml
@@ -63,8 +63,8 @@ risk = { path = "../risk" }
 storage = { path = "../storage" }
 
 
-# Essential ML frameworks for HFT inference - CUDA OPTIONAL (enable via 'cuda' feature)
-candle-core = { version = "0.9" }  # CPU-only by default, add 'cuda' feature for GPU
+# Essential ML frameworks for HFT inference - CUDA ENABLED
+candle-core = { version = "0.9", features = ["cuda"] }  # GPU acceleration enabled
 candle-nn = { version = "0.9" }
 candle-optimisers = { version = "0.9" }
 
diff --git a/ml/src/inference.rs b/ml/src/inference.rs
index caf28d198..150844da3 100644
--- a/ml/src/inference.rs
+++ b/ml/src/inference.rs
@@ -997,9 +997,10 @@ mod tests {
     }
 
     #[tokio::test]
+    #[ignore] // Slow test: GPU initialization + 3 model loads can take 60+ seconds
     async fn test_model_loading_multiple_models() -> Result<(), Box<dyn std::error::Error>> {
         let safety_manager = Arc::new(MLSafetyManager::new(MLSafetyConfig::default()));
-        let config = RealInferenceConfig::default();
+        let config = RealInferenceConfig::default(); // Use GPU with CUDA enabled
         let engine = RealMLInferenceEngine::new(config, safety_manager);
 
         // Load multiple models