diff --git a/.github/scripts/verify_documented_env_vars.py b/.github/scripts/verify_documented_env_vars.py
index 6723b99da0..aa060b00c0 100644
--- a/.github/scripts/verify_documented_env_vars.py
+++ b/.github/scripts/verify_documented_env_vars.py
@@ -74,6 +74,9 @@
     "NXRT_SQNBIT_PREFILL_MIN": "referenced in prose about a proposed tuning knob",
     "ONNX_GENAI_BASE_URL": "client-side variable read by external tooling, not by this workspace",
     "ONNX_GENAI_SD_PACKAGE": "referenced in prose about a proposed packaging layout",
+    "ONNX_GENAI_DECODE_GEMV_PROBE_ROWS": "throwaway ceiling-probe flag in "
+    "2026-08-18-multirow-gemv-ceiling-probe.md, whose method section states the "
+    "probe was reverted after measuring and is not shipped",
 }
 
 # Uppercase only: the lowercase `onnx_genai_*` names in the docs are Prometheus
diff --git a/crates/onnx-runtime-ep-cuda/src/kernels/standard_attention.rs b/crates/onnx-runtime-ep-cuda/src/kernels/standard_attention.rs
index eb50ce92c4..1c0bd9a457 100644
--- a/crates/onnx-runtime-ep-cuda/src/kernels/standard_attention.rs
+++ b/crates/onnx-runtime-ep-cuda/src/kernels/standard_attention.rs
@@ -2125,12 +2125,10 @@ impl StandardAttentionKernel {
                 .saturating_mul(q_heads as u64)
                 .saturating_mul(q_seq as u64);
             let split_bytes = match split_cfg {
-                Some((num_splits, _)) => attention_split_scratch_floats(
-                    num_splits,
-                    total_rows_u,
-                    v_head_size as u64,
-                )
-                .saturating_mul(std::mem::size_of::<f32>()),
+                Some((num_splits, _)) => {
+                    attention_split_scratch_floats(num_splits, total_rows_u, v_head_size as u64)
+                        .saturating_mul(std::mem::size_of::<f32>())
+                }
                 None => 0,
             };
             let mut ws = if capture_workspace_eligible {