diff --git a/crates/ml/src/trainers/tlob.rs b/crates/ml/src/trainers/tlob.rs
index 33a3eb01f..09eb43568 100644
--- a/crates/ml/src/trainers/tlob.rs
+++ b/crates/ml/src/trainers/tlob.rs
@@ -416,9 +416,13 @@ impl TLOBTrainer {
     }
 
     /// Train one epoch
+    ///
+    /// Loss is accumulated on-device as a GPU tensor to avoid per-batch
+    /// `to_scalar()` synchronisation. A single scalar extraction happens
+    /// after the loop completes, with a NaN guard every 100 batches.
     async fn train_epoch(&mut self, sequences: &[OrderBookSequence]) -> Result<f64> {
-        let mut total_loss = 0.0;
-        let mut num_batches = 0;
+        let mut loss_accum = Tensor::zeros((), DType::F32, &self.device)?;
+        let mut num_batches: usize = 0;
 
         // Process in batches
         for batch_sequences in sequences.chunks(self.hyperparams.batch_size) {
@@ -437,18 +441,37 @@ impl TLOBTrainer {
             // Gradient clipping
             self.clip_gradients()?;
 
-            total_loss += loss.to_scalar::<f32>()? as f64;
+            // Accumulate loss on-device (detach to prevent graph growth)
+            loss_accum = loss_accum.add(&loss.detach())?;
             num_batches += 1;
+
+            // NaN guard every 100 batches — single cheap sync
+            if num_batches % 100 == 0 {
+                let check = loss_accum.to_scalar::<f32>()?;
+                if check.is_nan() || check.is_infinite() {
+                    warn!(batch = num_batches, "TLOB train loss diverged (NaN/Inf)");
+                    return Ok(f64::NAN);
+                }
+            }
         }
 
-        Ok(total_loss / num_batches as f64)
+        if num_batches == 0 {
+            return Ok(0.0);
+        }
+
+        // Single scalar extraction after all batches
+        let total = loss_accum.to_scalar::<f32>()? as f64;
+        Ok(total / num_batches as f64)
     }
 
     /// Validate one epoch
+    ///
+    /// Loss is accumulated on-device to avoid per-batch `to_scalar()`.
+    /// MAE already returns `f64` (CPU scalar) so it stays as f64 sum.
     async fn validate_epoch(&self, sequences: &[OrderBookSequence]) -> Result<(f64, f64)> {
-        let mut total_loss = 0.0;
+        let mut loss_accum = Tensor::zeros((), DType::F32, &self.device)?;
         let mut total_mae = 0.0;
-        let mut num_batches = 0;
+        let mut num_batches: usize = 0;
 
         // Process in batches (no gradient computation)
         for batch_sequences in sequences.chunks(self.hyperparams.batch_size) {
@@ -461,11 +484,27 @@ impl TLOBTrainer {
             let loss = self.calculate_mse_loss(&predictions, &target_tensor)?.detach();
             let mae = self.calculate_mae(&predictions, &target_tensor)?;
 
-            total_loss += loss.to_scalar::<f32>()? as f64;
+            // Accumulate loss on-device
+            loss_accum = loss_accum.add(&loss)?;
             total_mae += mae;
             num_batches += 1;
+
+            // NaN guard every 100 batches
+            if num_batches % 100 == 0 {
+                let check = loss_accum.to_scalar::<f32>()?;
+                if check.is_nan() || check.is_infinite() {
+                    warn!(batch = num_batches, "TLOB val loss diverged (NaN/Inf)");
+                    return Ok((f64::NAN, f64::NAN));
+                }
+            }
         }
 
+        if num_batches == 0 {
+            return Ok((0.0, 0.0));
+        }
+
+        // Single scalar extraction after all batches
+        let total_loss = loss_accum.to_scalar::<f32>()? as f64;
         Ok((
             total_loss / num_batches as f64,
             total_mae / num_batches as f64,