diff --git a/crates/ml/tests/dqn_action_collapse_fix_test.rs b/crates/ml/tests/dqn_action_collapse_fix_test.rs
index 5ca1aa38d..7df274cf0 100644
--- a/crates/ml/tests/dqn_action_collapse_fix_test.rs
+++ b/crates/ml/tests/dqn_action_collapse_fix_test.rs
@@ -263,8 +263,8 @@ fn test_hyperopt_v_range_widened() {
 
     let bounds = DQNParams::continuous_bounds();
 
-    // v_range at index 10: (10.0, 50.0) — symmetric support ±v_range
-    // v_min = -v_range, v_max = +v_range
+    // v_range at index 10: fixed to 20.0 in Phase Fast (default),
+    // or (10.0, 50.0) range in Phase Full.
     let (v_range_lo, v_range_hi) = bounds.get(10).copied().unwrap_or((0.0, 0.0));
     assert!(
         v_range_lo >= 10.0,
@@ -272,8 +272,8 @@ fn test_hyperopt_v_range_widened() {
         v_range_lo
     );
     assert!(
-        v_range_hi >= 25.0,
-        "v_range upper bound should be >= 25.0 (covers DSR Q-values), got {}",
+        v_range_hi >= 20.0,
+        "v_range upper bound should be >= 20.0, got {}",
         v_range_hi
     );
 
diff --git a/crates/ml/tests/dqn_early_stopping_termination_test.rs b/crates/ml/tests/dqn_early_stopping_termination_test.rs
index fde4dc2bf..25f85681c 100644
--- a/crates/ml/tests/dqn_early_stopping_termination_test.rs
+++ b/crates/ml/tests/dqn_early_stopping_termination_test.rs
@@ -134,6 +134,8 @@ async fn test_early_stopping_terminates_with_error() {
 
     // Configure DQN with aggressive early stopping (will trigger quickly)
     let mut hyperparams = DQNHyperparameters::conservative();
+    // Apply smoketest profile for consistent hidden_dim across RTX 3050 and H100.
+    ml::training_profile::DqnTrainingProfile::load("dqn-smoketest").apply_to(&mut hyperparams);
 
     // Force gradient collapse detection by setting threshold above typical grad norm.
     // Threshold = 1e-8 × 1e9 = 10.0 (always above typical grad norms ~0.5-2.0).
@@ -206,6 +208,7 @@ async fn test_gradient_collapse_propagates_error() {
 
     // Configure DQN to trigger per-step gradient collapse detection
     let mut hyperparams = DQNHyperparameters::conservative();
+    ml::training_profile::DqnTrainingProfile::load("dqn-smoketest").apply_to(&mut hyperparams);
 
     // Force gradient collapse detection: threshold = lr × multiplier = 5e-9 × 1e9 = 5.0
     // Typical grad norms are ~0.5-2.0, so threshold 5.0 is always above → collapse triggers.
@@ -264,6 +267,7 @@ async fn test_healthy_training_completes_successfully() {
 
     // Configure with healthy parameters (should NOT early stop)
     let mut hyperparams = DQNHyperparameters::conservative();
+    ml::training_profile::DqnTrainingProfile::load("dqn-smoketest").apply_to(&mut hyperparams);
 
     // Normal hyperopt-validated parameters — should NOT early stop
     // Threshold = 1e-5 × 100 = 0.001 — well below typical grad_norm (~1-4), no false collapse.
diff --git a/crates/ml/tests/smoke_test_real_data.rs b/crates/ml/tests/smoke_test_real_data.rs
index a04af75ec..25b6281b8 100644
--- a/crates/ml/tests/smoke_test_real_data.rs
+++ b/crates/ml/tests/smoke_test_real_data.rs
@@ -525,7 +525,7 @@ mod gpu_smoke {
 
         // Create dueling networks (state_dim=54: 51 market + 3 portfolio)
         let config = BranchingConfig::trading_default(
-            54, vec![256, 256], 128, vec![5, 3, 3],
+            54, vec![64, 64], 32, vec![5, 3, 3],
         );
         let online = BranchingDuelingQNetwork::new(config.clone(), stream.clone()).unwrap();
         let target_net = BranchingDuelingQNetwork::new(config, stream.clone()).unwrap();
@@ -537,7 +537,7 @@ mod gpu_smoke {
             None,
             100_000.0,
             0.01, 0.05,
-            (256, 256, 128, 128),
+            (64, 64, 32, 32),
             (54, 51, 51),
             4, 50,
         ).unwrap();
@@ -606,7 +606,7 @@ mod gpu_smoke {
         info!(num_bars, market_dim = 51, "GPU buffer allocated (with OFI)");
 
         let config = BranchingConfig::trading_default(
-            54, vec![256, 256], 128, vec![5, 3, 3],
+            54, vec![64, 64], 32, vec![5, 3, 3],
         );
         let online = BranchingDuelingQNetwork::new(config.clone(), stream.clone()).unwrap();
         let target_net = BranchingDuelingQNetwork::new(config, stream.clone()).unwrap();
@@ -618,7 +618,7 @@ mod gpu_smoke {
             None,
             100_000.0,
             0.01, 0.05,
-            (256, 256, 128, 128),
+            (64, 64, 32, 32),
             (54, 51, 51),
             4, 50,
         ).unwrap();
@@ -675,7 +675,7 @@ mod gpu_smoke {
             real_market_data(&ohlcv_dir, None, &stream)?;
 
         let config = BranchingConfig::trading_default(
-            54, vec![256, 256], 128, vec![5, 3, 3],
+            54, vec![64, 64], 32, vec![5, 3, 3],
         );
         let online = BranchingDuelingQNetwork::new(config.clone(), stream.clone()).unwrap();
         let target_net = BranchingDuelingQNetwork::new(config, stream.clone()).unwrap();
@@ -687,7 +687,7 @@ mod gpu_smoke {
             None,
             100_000.0,
             0.01, 0.05,
-            (256, 256, 128, 128),
+            (64, 64, 32, 32),
             (54, 51, 51),
             4, 50,
         ).unwrap();