From 97c809448ff106ca8084a8775f14e2f8af71f325 Mon Sep 17 00:00:00 2001
From: JohannesGaessler <johannesg@5d6.de>
Date: Tue, 8 Aug 2023 19:34:31 +0200
Subject: [PATCH] add plotting files

---
 benchmark.sh       |  6 ++++++
 plot_ts_per_ngl.py | 23 +++++++++++++++++++++++
 2 files changed, 29 insertions(+)
 create mode 100755 benchmark.sh
 create mode 100644 plot_ts_per_ngl.py

diff --git a/benchmark.sh b/benchmark.sh
new file mode 100755
index 000000000..67d628698
--- /dev/null
+++ b/benchmark.sh
@@ -0,0 +1,6 @@
+#!/usr/bin/env sh
+
+for ngl in {0..35}
+do
+    ./main --model models/nvme/llama-7b-ggml-q4_0.bin --seed 1337 --ignore-eos --n-predict 128 --ctx-size 2048 --threads 8 -ngl $ngl -mmq
+done
diff --git a/plot_ts_per_ngl.py b/plot_ts_per_ngl.py
new file mode 100644
index 000000000..b922cee7b
--- /dev/null
+++ b/plot_ts_per_ngl.py
@@ -0,0 +1,23 @@
+#!/usr/bin/env python3
+
+import sqlite3
+import matplotlib.pyplot as plt
+
+con = sqlite3.connect("llama.sqlite")
+cur = con.cursor()
+
+ts = []
+
+for ngl in range(0, 36):
+    res = cur.execute(f"SELECT t_eval_us,n_eval FROM llama_runs WHERE n_gpu_layers={ngl};")
+    t_eval_us, n_eval = res.fetchone()
+    ts.append(n_eval * 1000000/t_eval_us)
+
+plt.plot(ts)
+plt.xlim(0, 35)
+plt.ylim(0, 130)
+plt.title("7b q4_0, 3700X, 3200 MHz dual-channel RAM, RTX 3090")
+plt.xlabel("-ngl")
+plt.ylabel("Generated t/s")
+plt.savefig("benchmark.png", dpi=240)
+plt.show()