diff --git a/crates/ml/src/cuda_pipeline/batched_backward.rs b/crates/ml/src/cuda_pipeline/batched_backward.rs
index 6d1d12080..4537ac701 100644
--- a/crates/ml/src/cuda_pipeline/batched_backward.rs
+++ b/crates/ml/src/cuda_pipeline/batched_backward.rs
@@ -237,7 +237,7 @@ impl CublasBackward {
                 c as *mut std::ffi::c_void,
                 cublas_sys::cudaDataType_t::CUDA_R_16BF,
                 ldc,
-                cublas_sys::cublasComputeType_t::CUBLAS_COMPUTE_32F,
+                cublas_sys::cublasComputeType_t::CUBLAS_COMPUTE_32F_FAST_TF32,
                 cublas_sys::cublasGemmAlgo_t::CUBLAS_GEMM_DEFAULT_TENSOR_OP,
             );
             if status != cublas_sys::cublasStatus_t::CUBLAS_STATUS_SUCCESS {
@@ -281,7 +281,7 @@ impl CublasBackward {
                 c as *mut std::ffi::c_void,
                 cublas_sys::cudaDataType_t::CUDA_R_32F,
                 ldc,
-                cublas_sys::cublasComputeType_t::CUBLAS_COMPUTE_32F,
+                cublas_sys::cublasComputeType_t::CUBLAS_COMPUTE_32F_FAST_TF32,
                 cublas_sys::cublasGemmAlgo_t::CUBLAS_GEMM_DEFAULT_TENSOR_OP,
             );
             if status != cublas_sys::cublasStatus_t::CUBLAS_STATUS_SUCCESS {