diff --git a/crates/ml/src/cuda_pipeline/gpu_backtest_evaluator.rs b/crates/ml/src/cuda_pipeline/gpu_backtest_evaluator.rs
index 01a3d5be9..75f9ef1ab 100644
--- a/crates/ml/src/cuda_pipeline/gpu_backtest_evaluator.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_backtest_evaluator.rs
@@ -221,7 +221,7 @@ pub struct GpuBacktestEvaluator {
     actions_history_buf: CudaSlice<i32>, // [n_windows * max_len]
 
     // Gather kernel output buffer (overwritten every step)
-    states_buf: CudaSlice<f32>, // [n_windows * (feature_dim + 3)]
+    states_buf: CudaSlice<f32>, // [n_windows * state_dim] (8-aligned)
 
     // Output buffer (written by metrics kernel)
     metrics_buf: CudaSlice<f32>, // [n_windows * 10]
@@ -232,6 +232,8 @@ pub struct GpuBacktestEvaluator {
     feature_dim: usize,
     /// Portfolio feature dimension used during construction (always 3).
     portfolio_dim: usize,
+    /// Aligned state dimension (feature_dim + portfolio_dim, 8-aligned for tensor cores).
+    state_dim: usize,
     config: GpuBacktestConfig,
 
     /// Forward kernel for pure-CUDA DQN evaluation (compiled lazily via NVRTC).
@@ -426,9 +428,10 @@ impl GpuBacktestEvaluator {
             .map_err(|e| MLError::ModelError(format!("metrics alloc: {e}")))?;
 
         // Portfolio dimension is always 3: (normalised value, position, spread_cost).
-        // state_dim = feature_dim + 3, zero-padded to align to tensor cores if needed.
+        // state_dim = feature_dim + 3, then 8-aligned for H100 tensor core HMMA dispatch.
+        // The CUDA gather_states kernel zero-pads positions [feat_dim+3 .. state_dim).
         const PORTFOLIO_DIM: usize = 3;
-        let state_dim = feature_dim + PORTFOLIO_DIM;
+        let state_dim = (feature_dim + PORTFOLIO_DIM + 7) & !7; // gpu-ok: constructor alignment
         let states_buf = stream
             .alloc_zeros::<f32>(n_windows * state_dim)
             .map_err(|e| MLError::ModelError(format!("states_buf alloc: {e}")))?;
@@ -469,6 +472,7 @@ impl GpuBacktestEvaluator {
             max_len,
             feature_dim,
             portfolio_dim: PORTFOLIO_DIM,
+            state_dim,
             config,
             forward_kernel: None,
             forward_actions_buf,
@@ -524,7 +528,7 @@ impl GpuBacktestEvaluator {
             )));
         }
 
-        let state_dim = self.feature_dim + self.portfolio_dim;
+        let state_dim = self.state_dim;
 
         // Launch the gather kernel — one thread per window
         let grid = ((self.n_windows + 255) / 256) as u32;
@@ -746,7 +750,7 @@ impl GpuBacktestEvaluator {
             self.forward_kernel = Some(kernel);
         }
 
-        let state_dim = self.feature_dim + self.portfolio_dim;
+        let state_dim = self.state_dim;
         let (shared_h1, shared_h2, _, _) = network_dims;
         let shmem_max_in = state_dim.max(shared_h1).max(shared_h2);
         let shmem_tile_rows = {
@@ -957,7 +961,7 @@ impl GpuBacktestEvaluator {
         online_weights: &DuelingWeightSet,
         network_dims: (usize, usize, usize, usize),
     ) -> Result<(), MLError> {
-        let state_dim = self.feature_dim + self.portfolio_dim;
+        let state_dim = self.state_dim;
         let (shared_h1, shared_h2, _, _) = network_dims;
         let shmem_max_in = state_dim.max(shared_h1).max(shared_h2);
         let shmem_tile_rows = {
@@ -1262,7 +1266,7 @@ impl GpuBacktestEvaluator {
         network_dims: (usize, usize, usize, usize),
     ) -> Result<CudaFunction, MLError> {
         let (shared_h1, shared_h2, value_h, adv_h) = network_dims;
-        let state_dim = self.feature_dim + self.portfolio_dim;
+        let state_dim = self.state_dim;
         let shmem_max_in_dim = state_dim.max(shared_h1).max(shared_h2);
         let shmem_tile_rows = {
             let gpu_caps = crate::gpu::capabilities::cached_capabilities();
@@ -1325,7 +1329,7 @@ impl GpuBacktestEvaluator {
     /// Does NOT produce a Candle Tensor — the pure-CUDA `evaluate_dqn` and
     /// `evaluate_ppo` paths read directly from `states_buf`.
     fn launch_gather(&self, step: usize) -> Result<(), MLError> {
-        let state_dim = self.feature_dim + self.portfolio_dim;
+        let state_dim = self.state_dim;
         let grid = ((self.n_windows + 255) / 256) as u32;
         let launch_cfg = LaunchConfig {
             grid_dim: (grid.max(1), 1, 1),