diff --git a/crates/ml/src/cuda_pipeline/batched_forward.rs b/crates/ml/src/cuda_pipeline/batched_forward.rs
index 0dcc79747..4b5f9debc 100644
--- a/crates/ml/src/cuda_pipeline/batched_forward.rs
+++ b/crates/ml/src/cuda_pipeline/batched_forward.rs
@@ -501,9 +501,9 @@ impl CublasForward {
         let distinct_branches = h_b0_ptr != h_b1_ptr && h_b1_ptr != h_b2_ptr && h_b0_ptr != h_b2_ptr
             && h_b0_ptr != h_b3_ptr && h_b1_ptr != h_b3_ptr && h_b2_ptr != h_b3_ptr;
 
-        if distinct_branches {
-            // Multi-stream branch dispatch (f32 path, experience collection).
-            // cublasLtMatmul takes stream per-call — no cublasSetStream needed.
+        if false && distinct_branches {
+            // Multi-stream branch dispatch DISABLED for H100 cublasLtMatmul debugging.
+            // cublasLtMatmul returns NOT_SUPPORTED on branch streams on H100.
             self.trunk_done_event.record(stream)
                 .map_err(|e| MLError::ModelError(format!("f32 trunk event record: {e}")))?;