diff --git a/crates/ml/build.rs b/crates/ml/build.rs
index 9b5a11afb..88e62c7cd 100644
--- a/crates/ml/build.rs
+++ b/crates/ml/build.rs
@@ -93,7 +93,7 @@ fn main() {
         }
     }
     let passed = kernels_with_common.len() - failed.len();
-    eprintln!("  Precompiled {passed}/{} CUDA kernels ({arch}) — all BF16",
+    eprintln!("  Precompiled {passed}/{} CUDA kernels ({arch}) — f32/TF32",
               kernels_with_common.len());
     if !failed.is_empty() {
         eprintln!("  FAILED: {}", failed.join(", "));
diff --git a/crates/ml/src/cuda_pipeline/gpu_attention.rs b/crates/ml/src/cuda_pipeline/gpu_attention.rs
index 4cb3025dd..9f1f89798 100644
--- a/crates/ml/src/cuda_pipeline/gpu_attention.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_attention.rs
@@ -217,17 +217,17 @@ impl GpuAttention {
 
         // ── Load cubin kernels ──────────────────────────────────────────
 
-        // Forward cubin — SDP kernel lives here
-        let module = context.load_cubin(ATTENTION_CUBIN.to_vec())
+        // Forward cubin — old multihead_feature_attention (legacy, not used by cuBLAS path)
+        let _module = context.load_cubin(ATTENTION_CUBIN.to_vec())
             .map_err(|e| MLError::ModelError(format!("attention module: {e}")))?;
-        let sdp_fwd_kernel = module.load_function("attn_sdp_fwd")
-            .map_err(|e| MLError::ModelError(format!("attn_sdp_fwd kernel load: {e}")))?;
-        let layer_norm_fwd_kernel = module.load_function("attn_layer_norm_fwd")
-            .map_err(|e| MLError::ModelError(format!("attn_layer_norm_fwd kernel load: {e}")))?;
 
-        // Backward cubin — SDP bwd, LayerNorm bwd, bias grad, grad_norm, adam
+        // Backward cubin — contains ALL element-wise kernels (fwd + bwd) for cuBLAS attention
         let bwd_module = context.load_cubin(ATTENTION_BACKWARD_CUBIN.to_vec())
             .map_err(|e| MLError::ModelError(format!("attention backward module: {e}")))?;
+        let sdp_fwd_kernel = bwd_module.load_function("attn_sdp_fwd")
+            .map_err(|e| MLError::ModelError(format!("attn_sdp_fwd kernel load: {e}")))?;
+        let layer_norm_fwd_kernel = bwd_module.load_function("attn_layer_norm_fwd")
+            .map_err(|e| MLError::ModelError(format!("attn_layer_norm_fwd kernel load: {e}")))?;
         let sdp_bwd_kernel = bwd_module.load_function("attn_sdp_bwd")
             .map_err(|e| MLError::ModelError(format!("attn_sdp_bwd kernel load: {e}")))?;
         let layer_norm_bwd_kernel = bwd_module.load_function("attn_layer_norm_bwd")