diff --git a/crates/onnx-runtime-ep-cpu/benches/activation_bench.rs b/crates/onnx-runtime-ep-cpu/benches/activation_bench.rs
index 206ac68883..41152162ce 100644
--- a/crates/onnx-runtime-ep-cpu/benches/activation_bench.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/activation_bench.rs
@@ -79,6 +79,9 @@ fn time(name: &str, kernel: &dyn Kernel, ins: &[TensorView], out: &mut Tensor, e
 }
 
 fn main() {
+    // Match the decode thread topology a served session runs in (#1749).
+    common::init_decode_topology();
+
     // Decode (`[1, 1, H]`), prefill (`[1, S, H]`) and two sizes around the
     // vector dispatch threshold.
     let shapes: [(&str, Vec<usize>); 5] = [
diff --git a/crates/onnx-runtime-ep-cpu/benches/common/mod.rs b/crates/onnx-runtime-ep-cpu/benches/common/mod.rs
index bfd6ae2884..146b0120c1 100644
--- a/crates/onnx-runtime-ep-cpu/benches/common/mod.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/common/mod.rs
@@ -203,3 +203,30 @@ pub fn assert_close(actual: &[f32], expected: &[f32], tolerance: f32) {
         );
     }
 }
+
+/// Put this benchmark process into the same decode thread topology a served
+/// session runs in.
+///
+/// `CpuExecutionProvider::initialize` is the earliest per-session hook and the
+/// only place an explicit `ONNX_GENAI_CPU_DECODE_THREADS` budget becomes a
+/// process-wide bound on prefill/MLAS Rayon parallelism and, on Linux, on CPU
+/// affinity. A benchmark that never calls it runs every row on an *unbounded*
+/// process, so a `t=N` row measures N decode workers competing with a
+/// full-width Rayon pool on every core -- not the configuration production runs
+/// (#1749).
+///
+/// Deliberately the real `CpuExecutionProvider::initialize` rather than a copy
+/// of what it currently does. A reimplementation would silently stop matching
+/// the moment `initialize` grows a second responsibility, which is exactly the
+/// drift that leaves a benchmark measuring a configuration nothing ships.
+///
+/// A no-op unless a budget is set, and idempotent: the underlying bound latches
+/// on first call. Benches that install their own fixed-width Rayon pool (see
+/// `kernels.rs`, which sweeps `[1, 8]`) will oversubscribe that pool onto a
+/// smaller budget's cores -- which is what production does with the same two
+/// settings, and is why this is applied there too rather than special-cased.
+pub fn init_decode_topology() {
+    CpuExecutionProvider::new()
+        .initialize(&Default::default())
+        .expect("the CPU EP must initialize");
+}
diff --git a/crates/onnx-runtime-ep-cpu/benches/gqa_decode.rs b/crates/onnx-runtime-ep-cpu/benches/gqa_decode.rs
index b3dcd649f1..37f936e4a3 100644
--- a/crates/onnx-runtime-ep-cpu/benches/gqa_decode.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/gqa_decode.rs
@@ -71,6 +71,10 @@ fn bind_decode_pool_width(threads: usize) {
 /// `sdpa_decode_group` call over the same KV window. Single-threaded on purpose
 /// — this isolates the memory-traffic difference from the decode scheduler.
 fn bench_sdpa_group_vs_row(c: &mut Criterion) {
+    // Match the decode thread topology a served session runs in (#1749).
+    // Idempotent, and called from every group member rather than just the
+    // first so coverage does not silently ride on `criterion_group!` order.
+    common::init_decode_topology();
     let mut group_bench = c.benchmark_group("sdpa_group_vs_row");
     for (label, num_heads, kv_num_heads, head_size) in GEOMETRIES {
         let group = num_heads / kv_num_heads;
@@ -217,6 +221,10 @@ fn gqa_kernel(
 /// the bench twice (once with the variable set to `0`, once to `1`) to get the
 /// A/B. The current setting is reported in the benchmark id.
 fn bench_gqa_kernel_decode(c: &mut Criterion) {
+    // Match the decode thread topology a served session runs in (#1749).
+    // Idempotent, and called from every group member rather than just the
+    // first so coverage does not silently ride on `criterion_group!` order.
+    common::init_decode_topology();
     let fused = std::env::var("ONNX_GENAI_GQA_GROUP_FUSED").unwrap_or_else(|_| "default".into());
     let mut group_bench = c.benchmark_group("gqa_kernel_decode");
     // 8 workers is the bounded decode-pool width the runtime uses for M=1.
@@ -266,6 +274,10 @@ fn bench_gqa_kernel_decode(c: &mut Criterion) {
 /// read either way, so the ratio between arms reads directly as the fraction of
 /// cache traffic the append removes.
 fn bench_gqa_half_kv_append(c: &mut Criterion) {
+    // Match the decode thread topology a served session runs in (#1749).
+    // Idempotent, and called from every group member rather than just the
+    // first so coverage does not silently ride on `criterion_group!` order.
+    common::init_decode_topology();
     let mut group_bench = c.benchmark_group("gqa_half_kv_append");
     bind_decode_pool_width(8);
     for (label, num_heads, kv_num_heads, head_size) in GEOMETRIES {
diff --git a/crates/onnx-runtime-ep-cpu/benches/half_decode_gemv_ab.rs b/crates/onnx-runtime-ep-cpu/benches/half_decode_gemv_ab.rs
index 98e898a9c2..1d1c6777f6 100644
--- a/crates/onnx-runtime-ep-cpu/benches/half_decode_gemv_ab.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/half_decode_gemv_ab.rs
@@ -144,6 +144,9 @@ fn max_relative_deviation(got: &[f32], want: &[f32]) -> f32 {
 }
 
 fn main() {
+    // Match the decode thread topology a served session runs in (#1749).
+    common::init_decode_topology();
+
     // Shapes a decode step actually issues: a 7B-class MLP projection, a
     // vocabulary head (the widest weight in the graph), and a square control.
     let shapes: Vec<(&str, usize, usize)> = match std::env::var("PROBE_SHAPE").as_deref() {
diff --git a/crates/onnx-runtime-ep-cpu/benches/half_prefill_route_ab.rs b/crates/onnx-runtime-ep-cpu/benches/half_prefill_route_ab.rs
index dcb9ee051e..94e969e911 100644
--- a/crates/onnx-runtime-ep-cpu/benches/half_prefill_route_ab.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/half_prefill_route_ab.rs
@@ -105,6 +105,9 @@ fn build_kernel(dtype: FloatDType, m: usize, k: usize, n: usize) -> Box<dyn Kern
 }
 
 fn main() {
+    // Match the decode thread topology a served session runs in (#1749).
+    common::init_decode_topology();
+
     let shapes: Vec<(usize, usize)> = match std::env::var("PROBE_SHAPE").as_deref() {
         Ok("big") => vec![(4096, 11008)],
         Ok("small") => vec![(2048, 2048)],
diff --git a/crates/onnx-runtime-ep-cpu/benches/int4_acc0_attribution.rs b/crates/onnx-runtime-ep-cpu/benches/int4_acc0_attribution.rs
index 85ba498e6b..14a4dcee96 100644
--- a/crates/onnx-runtime-ep-cpu/benches/int4_acc0_attribution.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/int4_acc0_attribution.rs
@@ -52,6 +52,8 @@
 //! cargo bench -p onnx-runtime-ep-cpu --bench int4_acc0_attribution
 //! ```
 
+mod common;
+
 use std::time::Instant;
 
 /// Shapes taken from the projections that dominate llama/qwen decode.
@@ -207,6 +209,9 @@ fn bench(label: &str, reps: usize, mut f: impl FnMut()) -> f64 {
 }
 
 fn main() {
+    // Match the decode thread topology a served session runs in (#1749).
+    common::init_decode_topology();
+
     let reps: usize = std::env::var("REPS")
         .ok()
         .and_then(|v| v.parse().ok())
diff --git a/crates/onnx-runtime-ep-cpu/benches/int4_decode_loop_ab.rs b/crates/onnx-runtime-ep-cpu/benches/int4_decode_loop_ab.rs
index 1945747df1..35275720a0 100644
--- a/crates/onnx-runtime-ep-cpu/benches/int4_decode_loop_ab.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/int4_decode_loop_ab.rs
@@ -179,6 +179,9 @@ struct Weight {
 }
 
 fn main() {
+    // Match the decode thread topology a served session runs in (#1749).
+    common::init_decode_topology();
+
     let block_size: usize = std::env::var("PROBE_BLOCK")
         .ok()
         .and_then(|v| v.parse().ok())
diff --git a/crates/onnx-runtime-ep-cpu/benches/int4_prefill_route_ab.rs b/crates/onnx-runtime-ep-cpu/benches/int4_prefill_route_ab.rs
index 20ffd38625..fa85ecf62e 100644
--- a/crates/onnx-runtime-ep-cpu/benches/int4_prefill_route_ab.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/int4_prefill_route_ab.rs
@@ -96,6 +96,9 @@ fn build_kernel(
 }
 
 fn main() {
+    // Match the decode thread topology a served session runs in (#1749).
+    common::init_decode_topology();
+
     // `PROBE_BLOCK` exists because `int4_prefill_gebp_min_rows` returns a
     // *different* threshold for weights whose block size the column-blocked
     // kernels cannot take (`INT4_PREFILL_GEBP_MIN_ROWS_UNBLOCKED`), and with
diff --git a/crates/onnx-runtime-ep-cpu/benches/int8_prefill_route_ab.rs b/crates/onnx-runtime-ep-cpu/benches/int8_prefill_route_ab.rs
index 3b04ed8e77..e0851b9b0b 100644
--- a/crates/onnx-runtime-ep-cpu/benches/int8_prefill_route_ab.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/int8_prefill_route_ab.rs
@@ -130,6 +130,9 @@ fn oracle_row(activations: &[f32], weight: &Weight<'_>, row: usize, col: usize)
 }
 
 fn main() {
+    // Match the decode thread topology a served session runs in (#1749).
+    common::init_decode_topology();
+
     let block_size = 32usize;
     let shapes: Vec<(usize, usize)> = match std::env::var("PROBE_SHAPE").as_deref() {
         Ok("big") => vec![(4096, 11008)],
diff --git a/crates/onnx-runtime-ep-cpu/benches/kernels.rs b/crates/onnx-runtime-ep-cpu/benches/kernels.rs
index 961f26a4e6..157019e6a8 100644
--- a/crates/onnx-runtime-ep-cpu/benches/kernels.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/kernels.rs
@@ -19,6 +19,10 @@ fn thread_pool(threads: usize) -> ThreadPool {
 }
 
 fn bench_add(c: &mut Criterion) {
+    // Match the decode thread topology a served session runs in (#1749).
+    // Idempotent, and called from every group member rather than just the
+    // first so coverage does not silently ride on `criterion_group!` order.
+    common::init_decode_topology();
     let mut group = c.benchmark_group("add");
     for (size, shape) in [
         ("small", vec![1_024]),
@@ -53,6 +57,10 @@ fn bench_add(c: &mut Criterion) {
 }
 
 fn bench_reduce_mean(c: &mut Criterion) {
+    // Match the decode thread topology a served session runs in (#1749).
+    // Idempotent, and called from every group member rather than just the
+    // first so coverage does not silently ride on `criterion_group!` order.
+    common::init_decode_topology();
     let mut group = c.benchmark_group("reduce_mean");
     for (size, shape) in [
         ("small", vec![32, 128]),
@@ -90,6 +98,10 @@ fn bench_reduce_mean(c: &mut Criterion) {
 }
 
 fn bench_gather(c: &mut Criterion) {
+    // Match the decode thread topology a served session runs in (#1749).
+    // Idempotent, and called from every group member rather than just the
+    // first so coverage does not silently ride on `criterion_group!` order.
+    common::init_decode_topology();
     let mut group = c.benchmark_group("gather");
     for (size, rows, columns, index_count) in [
         ("small", 4_096, 128, 32),
@@ -156,6 +168,10 @@ fn with_gemm_backend<T>(backend: &str, f: impl FnOnce() -> T) -> T {
 }
 
 fn bench_matmul(c: &mut Criterion) {
+    // Match the decode thread topology a served session runs in (#1749).
+    // Idempotent, and called from every group member rather than just the
+    // first so coverage does not silently ride on `criterion_group!` order.
+    common::init_decode_topology();
     let mut group = c.benchmark_group("matmul");
     let backends = &[
         "generic",
@@ -267,6 +283,10 @@ fn dequantize_mxfp4_kn(n: usize, k: usize, packed: &[u8]) -> Vec<f32> {
 }
 
 fn bench_block_quantized_matmul_cache(c: &mut Criterion) {
+    // Match the decode thread topology a served session runs in (#1749).
+    // Idempotent, and called from every group member rather than just the
+    // first so coverage does not silently ride on `criterion_group!` order.
+    common::init_decode_topology();
     let mut group = c.benchmark_group("block_quantized_matmul_cached_dense");
     group.sample_size(15);
     let (m, k, n) = (1usize, 1_024usize, 1_024usize);
@@ -368,6 +388,10 @@ fn block_quantized_moe_kernel(top_k: usize) -> Box<dyn Kernel> {
 }
 
 fn bench_block_quantized_moe_cache(c: &mut Criterion) {
+    // Match the decode thread topology a served session runs in (#1749).
+    // Idempotent, and called from every group member rather than just the
+    // first so coverage does not silently ride on `criterion_group!` order.
+    common::init_decode_topology();
     let mut group = c.benchmark_group("block_quantized_moe_cached_dense");
     group.sample_size(15);
     let (rows, hidden, inter, experts, top_k) = (1usize, 256usize, 256usize, 4usize, 1usize);
diff --git a/crates/onnx-runtime-ep-cpu/benches/matmul_nbits_prefill_ab.rs b/crates/onnx-runtime-ep-cpu/benches/matmul_nbits_prefill_ab.rs
index c1f57bdebe..81536f08a3 100644
--- a/crates/onnx-runtime-ep-cpu/benches/matmul_nbits_prefill_ab.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/matmul_nbits_prefill_ab.rs
@@ -132,6 +132,9 @@ fn median(mut samples: Vec<f64>) -> f64 {
 }
 
 fn main() {
+    // Match the decode thread topology a served session runs in (#1749).
+    common::init_decode_topology();
+
     let cases = [
         Case {
             label: "int8 dense-fallback",
diff --git a/crates/onnx-runtime-ep-cpu/benches/native_vs_mlas.rs b/crates/onnx-runtime-ep-cpu/benches/native_vs_mlas.rs
index 558bf7177b..05e58f9540 100644
--- a/crates/onnx-runtime-ep-cpu/benches/native_vs_mlas.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/native_vs_mlas.rs
@@ -53,6 +53,8 @@
 //! regresses every concurrent session on the box. A large `ratio` with a small
 //! `cpu_ratio` means exactly that, and is a reason to look before graduating.
 
+mod common;
+
 use std::hint::black_box;
 use std::sync::OnceLock;
 use std::time::Instant;
@@ -410,6 +412,9 @@ fn bench_transcendentals() {
 }
 
 fn main() {
+    // Match the decode thread topology a served session runs in (#1749).
+    common::init_decode_topology();
+
     println!("# native vs MLAS, one binary, interleaved, median of {REPS} reps");
     println!(
         "# mlas_linked={}  graduation_ratio={GRADUATION_RATIO}  threads={}",
diff --git a/crates/onnx-runtime-ep-cpu/benches/sdpa_simd.rs b/crates/onnx-runtime-ep-cpu/benches/sdpa_simd.rs
index b6b745443c..ac07eecff9 100644
--- a/crates/onnx-runtime-ep-cpu/benches/sdpa_simd.rs
+++ b/crates/onnx-runtime-ep-cpu/benches/sdpa_simd.rs
@@ -11,6 +11,8 @@
 //! Shapes are real encoder/decoder geometries rather than round numbers:
 //! Whisper-base encoder, BERT-base, ViT-L/14 and a 32/8/128 GQA decode step.
 
+mod common;
+
 use criterion::{BenchmarkId, Criterion, black_box, criterion_group, criterion_main};
 use onnx_runtime_ep_cpu::kernels::sdpa::{
     NoBias, NoMask, ScaleMode, SdpaConfig, SdpaTensors, sdpa_f32, sdpa_f32_scalar,
@@ -73,6 +75,10 @@ fn values(n: usize, seed: u64) -> Vec<f32> {
 }
 
 fn bench_sdpa_simd(c: &mut Criterion) {
+    // Match the decode thread topology a served session runs in (#1749).
+    // Idempotent, and called from every group member rather than just the
+    // first so coverage does not silently ride on `criterion_group!` order.
+    common::init_decode_topology();
     let mut group = c.benchmark_group("sdpa_f32");
     for geo in &GEOMETRIES {
         let batch = 1usize;