EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Add TP2 MOE benchmarks for AMD. (#5909)

Browse Source
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
Download Patch File Download Diff File Expand all files Collapse all files

22
.github/workflows/pr-test-amd.yml vendored
View File

@@ -141,11 +141,31 @@ jobs:
mkdir -p dummy-grok && wget https://sharkpublic.blob.core.windows.net/sharkpublic/sglang/dummy_grok.json -O dummy-grok/config.json
docker cp ./dummy-grok ci_sglang:/
- name: Evaluate Benchmark
- name: Benchmark dummy grok (TP=2)
timeout-minutes: 20
run: |
docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 models/test_dummy_grok_models.py
- name: Benchmark single latency (TP=2)
timeout-minutes: 20
run: |
docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 -e SGLANG_AMD_CI=1 ci_sglang python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_moe_tp2_bs1
- name: Benchmark single latency + torch.compile (TP=2)
timeout-minutes: 20
run: |
docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 ci_sglang python3 -m unittest test_bench_one_batch.TestBenchOneBatch.test_torch_compile_tp2_bs1
- name: Benchmark offline throughput (TP=2)
timeout-minutes: 20
run: |
docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 -e SGLANG_AMD_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_moe_offline_throughput_default
- name: Benchmark offline throughput (w/o RadixAttention) (TP=2)
timeout-minutes: 20
run: |
docker exec -w /sglang-checkout/test/srt -e SGLANG_IS_IN_CI=1 -e SGLANG_AMD_CI=1 ci_sglang python3 -m unittest test_bench_serving.TestBenchServing.test_moe_offline_throughput_without_radix_cache
finish:
if: always()
needs: [

6
test/srt/test_bench_one_batch.py
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(

11
test/srt/test_bench_serving.py
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__":