diff --git a/crates/ml/tests/dqn_training_pipeline_test.rs b/crates/ml/tests/dqn_training_pipeline_test.rs
index 9f7a330f4..e76c53ca6 100644
--- a/crates/ml/tests/dqn_training_pipeline_test.rs
+++ b/crates/ml/tests/dqn_training_pipeline_test.rs
@@ -176,16 +176,17 @@ async fn test_dqn_trains_on_es_fut() -> Result<()> {
 
     let checkpoint_dir = create_checkpoint_dir()?;
 
-    // Configure hyperparameters for fast test (5 epochs)
+    // Configure hyperparameters: smoketest profile for consistent sizing across GPUs
     let mut hyperparams = DQNHyperparameters::conservative();
+    ml::training_profile::DqnTrainingProfile::load("dqn-smoketest").apply_to(&mut hyperparams);
     // GPU PER is mandatory for fused CUDA training — do not disable
     hyperparams.epochs = 3; // CI smoke: validate pipeline, not convergence
-    hyperparams.batch_size = 64;
     hyperparams.learning_rate = 0.001;
     hyperparams.epsilon_start = 0.5;
     hyperparams.epsilon_end = 0.05;
     hyperparams.checkpoint_frequency = 3;
     hyperparams.early_stopping_enabled = false; // Test all epochs
+    hyperparams.replay_buffer_vram_fraction = 0.0; // Explicit buffer_size from TOML
 
     info!(data_dir = %data_dir, checkpoint_dir = %checkpoint_dir.display(), epochs = hyperparams.epochs, "Configuration ready");
 
@@ -304,6 +305,8 @@ async fn test_dqn_loss_decreases() -> Result<()> {
 
     // Train for 3 epochs — CI validates gradient flow, not full convergence
     let mut hyperparams = DQNHyperparameters::conservative();
+    ml::training_profile::DqnTrainingProfile::load("dqn-smoketest").apply_to(&mut hyperparams);
+    hyperparams.replay_buffer_vram_fraction = 0.0;
     // GPU PER is mandatory for fused CUDA training — do not disable
     hyperparams.epochs = 3;
     hyperparams.batch_size = 64;
@@ -378,6 +381,8 @@ async fn test_dqn_checkpoint_save_load() -> Result<()> {
 
     // Train for 2 epochs and save checkpoint
     let mut hyperparams = DQNHyperparameters::conservative();
+    ml::training_profile::DqnTrainingProfile::load("dqn-smoketest").apply_to(&mut hyperparams);
+    hyperparams.replay_buffer_vram_fraction = 0.0;
     // GPU PER is mandatory for fused CUDA training — do not disable
     hyperparams.epochs = 2;
     hyperparams.batch_size = 64;
@@ -447,6 +452,8 @@ async fn test_dqn_q_value_predictions() -> Result<()> {
 
     // Train minimal model
     let mut hyperparams = DQNHyperparameters::conservative();
+    ml::training_profile::DqnTrainingProfile::load("dqn-smoketest").apply_to(&mut hyperparams);
+    hyperparams.replay_buffer_vram_fraction = 0.0;
     // GPU PER is mandatory for fused CUDA training — do not disable
     hyperparams.epochs = 2;
     hyperparams.batch_size = 32;
@@ -504,6 +511,8 @@ async fn test_dqn_epsilon_greedy() -> Result<()> {
 
     // Configure with high epsilon decay
     let mut hyperparams = DQNHyperparameters::conservative();
+    ml::training_profile::DqnTrainingProfile::load("dqn-smoketest").apply_to(&mut hyperparams);
+    hyperparams.replay_buffer_vram_fraction = 0.0;
     hyperparams.epochs = 2;
     hyperparams.epsilon_start = 1.0;
     hyperparams.epsilon_end = 0.01;
@@ -578,8 +587,10 @@ async fn test_dqn_full_production_training() -> Result<()> {
     let checkpoint_dir = create_checkpoint_dir()?;
     let production_checkpoint_path = checkpoint_dir.join("dqn_es_fut_v1.safetensors");
 
-    // Production hyperparameters
+    // Production hyperparameters (smoketest profile for GPU sizing)
     let mut hyperparams = DQNHyperparameters::conservative();
+    ml::training_profile::DqnTrainingProfile::load("dqn-smoketest").apply_to(&mut hyperparams);
+    hyperparams.replay_buffer_vram_fraction = 0.0;
     // GPU PER is mandatory for fused CUDA training — do not disable
     hyperparams.epochs = 50;
     hyperparams.batch_size = 128;