From 1de285d3a15139d067e4ace0ad3d195c7257f438 Mon Sep 17 00:00:00 2001
From: "Albert Cheng (Engrg-Hardware 1)"
 <albecheng@login-bia02.bia.clusters.nvidia.com>
Date: Sat, 30 May 2026 00:07:32 -0700
Subject: [PATCH] Add SPEED-Bench reference AL values for DeepSeek-V4-Pro MTP
 1-8

Measured with SPEED-Bench coding dataset, temperature=1.0, thinking=true.
Values used for synthetic acceptance rate configuration in MTP benchmarks.
---
 benchmarks/speedbench-reference-al.yaml | 13 +++++++++++++
 1 file changed, 13 insertions(+)
 create mode 100644 benchmarks/speedbench-reference-al.yaml

diff --git a/benchmarks/speedbench-reference-al.yaml b/benchmarks/speedbench-reference-al.yaml
new file mode 100644
index 000000000..fe9f0e7cc
--- /dev/null
+++ b/benchmarks/speedbench-reference-al.yaml
@@ -0,0 +1,13 @@
+# Acceptance Length (AL) reference values measured with SPEED-Bench
+# dataset: coding | temperature: 1.0 | thinking: true
+# key = num_speculative_tokens (MTP level), value = golden AL
+
+deepseek-v4-pro:
+  1: 1.90
+  2: 2.60
+  3: 2.97
+  4: 3.04
+  5: 3.13
+  6: 3.08
+  7: 3.13
+  8: 3.12
\ No newline at end of file