Add TP2 MOE benchmarks for AMD. (#5909)

This commit is contained in:
Sai Enduri
2025-04-30 11:38:20 -07:00
committed by GitHub
parent e330f2b86c
commit 2afba1b1c1
3 changed files with 35 additions and 4 deletions

View File

@@ -1,3 +1,4 @@
import os
import unittest
from sglang.test.test_utils import (
@@ -45,7 +46,10 @@ class TestBenchOneBatch(CustomTestCase):
f"### test_moe_tp2_bs1 (Mixtral-8x7B)\n"
f"output_throughput: {output_throughput:.2f} token/s\n"
)
self.assertGreater(output_throughput, 125)
if os.getenv("SGLANG_AMD_CI") == "1":
self.assertGreater(output_throughput, 85)
else:
self.assertGreater(output_throughput, 125)
def test_torch_compile_tp2_bs1(self):
output_throughput = run_bench_offline_throughput(

View File

@@ -1,3 +1,4 @@
import os
import unittest
from sglang.test.test_utils import (
@@ -180,7 +181,10 @@ class TestBenchServing(CustomTestCase):
f"### test_moe_offline_throughput_default\n"
f'Output throughput: {res["output_throughput"]:.2f} token/s\n'
)
self.assertGreater(res["output_throughput"], 2200)
if os.getenv("SGLANG_AMD_CI") == "1":
self.assertGreater(res["output_throughput"], 2100)
else:
self.assertGreater(res["output_throughput"], 2200)
def test_moe_offline_throughput_without_radix_cache(self):
res = run_bench_serving(
@@ -195,7 +199,10 @@ class TestBenchServing(CustomTestCase):
f"### test_moe_offline_throughput_without_radix_cache\n"
f'Output throughput: {res["output_throughput"]:.2f} token/s\n'
)
self.assertGreater(res["output_throughput"], 2200)
if os.getenv("SGLANG_AMD_CI") == "1":
self.assertGreater(res["output_throughput"], 2100)
else:
self.assertGreater(res["output_throughput"], 2200)
if __name__ == "__main__":