chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Migrate performance, accuracy, and quantization tests to CI registry (#17177)

Browse Source 
Co-authored-by: Kangyan-Zhou <zky314343421@gmail.com>
This commit is contained in:
Alison Shao
2026-01-18 23:25:24 -08:00
committed by GitHub
co-authored by Kangyan-Zhou
parent a3d9a21882
commit 8916b9d080
20 changed files with 1050 additions and 1020 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/test_eval_accuracy_large.py → test/registered/eval/test_eval_accuracy_large.py
+3
View File
@@ -7,6 +7,7 @@ import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
@@ -18,6 +19,8 @@ from sglang.test.test_utils import (
write_github_step_summary,
)
register_cuda_ci(est_time=300, suite="stage-b-test-small-1-gpu-accuracy")
class TestEvalAccuracyLarge(CustomTestCase):
@classmethod
test/srt/test_moe_eval_accuracy_large.py → test/registered/eval/test_moe_eval_accuracy_large.py
+3
View File
@@ -7,6 +7,7 @@ import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MOE_MODEL_NAME_FOR_TEST,
@@ -18,6 +19,8 @@ from sglang.test.test_utils import (
write_github_step_summary,
)
register_cuda_ci(est_time=500, suite="stage-b-test-large-2-gpu-accuracy")
class TestMoEEvalAccuracyLarge(CustomTestCase):
@classmethod
test/registered/perf/test_bench_one_batch_1gpu.py
+39
View File
@@ -0,0 +1,39 @@
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
CustomTestCase,
is_in_ci,
run_bench_offline_throughput,
run_bench_one_batch,
write_github_step_summary,
)
register_cuda_ci(est_time=120, suite="stage-b-test-large-1-gpu-performance")
class TestBenchOneBatch1GPU(CustomTestCase):
def test_bs1_small(self):
_, output_throughput, _ = run_bench_one_batch(
DEFAULT_SMALL_MODEL_NAME_FOR_TEST, ["--cuda-graph-max-bs", "2"]
)
self.assertGreater(output_throughput, 50)
def test_bs1_default(self):
output_throughput = run_bench_offline_throughput(
DEFAULT_MODEL_NAME_FOR_TEST, ["--cuda-graph-max-bs", "2"]
)
if is_in_ci():
write_github_step_summary(
f"### test_bs1_default (llama-3.1-8b)\n"
f"output_throughput: {output_throughput:.2f} token/s\n"
)
self.assertGreater(output_throughput, 135)
if __name__ == "__main__":
unittest.main()
test/srt/test_bench_one_batch.py → test/registered/perf/test_bench_one_batch_2gpu.py
+3 -23
View File
@@ -1,40 +1,20 @@
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_MOE_MODEL_NAME_FOR_TEST,
DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
CustomTestCase,
is_in_amd_ci,
is_in_ci,
run_bench_offline_throughput,
run_bench_one_batch,
write_github_step_summary,
)
# We use `run_bench_offline_throughput`` instead of `run_bench_one_batch` for most cases
# because `run_bench_offline_throughput`` has overlap scheduler.
register_cuda_ci(est_time=180, suite="stage-b-test-large-2-gpu-performance")
class TestBenchOneBatch(CustomTestCase):
def test_bs1_small(self):
_, output_throughput, _ = run_bench_one_batch(
DEFAULT_SMALL_MODEL_NAME_FOR_TEST, ["--cuda-graph-max-bs", "2"]
)
self.assertGreater(output_throughput, 50)
def test_bs1_default(self):
output_throughput = run_bench_offline_throughput(
DEFAULT_MODEL_NAME_FOR_TEST, ["--cuda-graph-max-bs", "2"]
)
if is_in_ci():
write_github_step_summary(
f"### test_bs1_default (llama-3.1-8b)\n"
f"output_throughput: {output_throughput:.2f} token/s\n"
)
self.assertGreater(output_throughput, 135)
class TestBenchOneBatch2GPU(CustomTestCase):
def test_moe_tp2_bs1(self):
output_throughput = run_bench_offline_throughput(
test/registered/perf/test_bench_serving_1gpu_large.py
+81
View File
@@ -0,0 +1,81 @@
"""
Performance tests for single GPU that need H200 (80GB) - FP8 and EAGLE tests.
"""
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import (
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_MODEL_NAME_FOR_TEST_FP8,
DEFAULT_TARGET_MODEL_EAGLE,
CustomTestCase,
is_in_amd_ci,
is_in_ci,
run_bench_serving,
write_github_step_summary,
)
register_cuda_ci(est_time=300, suite="stage-b-test-large-1-gpu-performance")
class TestBenchServing1GPULarge(CustomTestCase):
def test_offline_throughput_default_fp8(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST_FP8,
num_prompts=500,
request_rate=float("inf"),
other_server_args=[],
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_default_fp8\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 3500)
else:
self.assertGreater(res["output_throughput"], 4300)
def test_online_latency_eagle(self):
res = run_bench_serving(
model=DEFAULT_TARGET_MODEL_EAGLE,
num_prompts=300,
request_rate=8,
sharegpt_context_len=3072,
disable_ignore_eos=True,
dataset_name="sharegpt",
other_server_args=[
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
DEFAULT_DRAFT_MODEL_EAGLE,
"--speculative-num-steps",
"5",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"16",
"--mem-fraction-static",
"0.7",
],
need_warmup=True,
seed=42,
)
if is_in_ci():
write_github_step_summary(
f"### test_online_latency_eagle\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
f"accept_length: {res['accept_length']:.2f} \n"
)
if is_in_amd_ci():
self.assertLess(res["median_e2e_latency_ms"], 1800)
else:
self.assertLess(res["median_e2e_latency_ms"], 900)
self.assertGreater(res["accept_length"], 3.0)
if __name__ == "__main__":
unittest.main()
test/registered/perf/test_bench_serving_1gpu_part1.py
+258
View File
@@ -0,0 +1,258 @@
"""
Performance tests for single GPU - LLM throughput/latency and LoRA tests.
Works on 5090 (32GB).
"""
import asyncio
import itertools
import unittest
import requests
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST,
CustomTestCase,
is_in_amd_ci,
is_in_ci,
run_bench_serving,
write_github_step_summary,
)
register_cuda_ci(est_time=1000, suite="stage-b-test-large-1-gpu-performance")
class TestBenchServing1GPUPart1(CustomTestCase):
def test_offline_throughput_default(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=500,
request_rate=float("inf"),
other_server_args=[],
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_default\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 3050)
else:
self.assertGreater(res["output_throughput"], 3800)
def test_offline_throughput_non_stream_small_batch_size(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=200,
request_rate=float("inf"),
other_server_args=["--max-running-requests", "10"],
dataset_name="sharegpt",
random_input_len=None,
random_output_len=None,
disable_stream=True,
need_warmup=True,
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_non_stream_small_batch_size\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 1000)
else:
self.assertGreater(res["output_throughput"], 1050)
def test_offline_throughput_without_radix_cache(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=500,
request_rate=float("inf"),
other_server_args=["--disable-radix-cache"],
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_without_radix_cache\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 3050)
else:
self.assertGreater(res["output_throughput"], 3800)
def test_offline_throughput_without_chunked_prefill(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=500,
request_rate=float("inf"),
other_server_args=["--chunked-prefill-size", "-1"],
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_without_chunked_prefill\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
self.assertGreater(res["output_throughput"], 2600)
def test_offline_throughput_with_triton_attention_backend(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=500,
request_rate=float("inf"),
other_server_args=[
"--attention-backend",
"triton",
"--context-length",
"8192",
],
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_with_triton_attention_backend\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 3500)
else:
self.assertGreater(res["output_throughput"], 3700)
def test_online_latency_default(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=100,
request_rate=1,
other_server_args=[],
)
if is_in_ci():
write_github_step_summary(
f"### test_online_latency_default\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
)
self.assertLess(res["median_e2e_latency_ms"], 11000)
if is_in_amd_ci():
self.assertLess(res["median_ttft_ms"], 115)
else:
self.assertLess(res["median_ttft_ms"], 86)
self.assertLess(res["median_itl_ms"], 10)
def test_lora_online_latency(self):
if is_in_amd_ci():
pass
res = self._run_lora_latency_test(enable_background_task=False)
if is_in_ci():
write_github_step_summary(
f"### test_lora_online_latency\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n"
)
self.assertLess(res["median_e2e_latency_ms"], 2400)
self.assertLess(res["median_ttft_ms"], 58)
def test_lora_online_latency_with_concurrent_adapter_updates(self):
if is_in_amd_ci():
pass
res = self._run_lora_latency_test(enable_background_task=True)
if is_in_ci():
write_github_step_summary(
f"### test_lora_online_latency\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n"
)
self.assertLess(res["median_e2e_latency_ms"], 4000)
self.assertLess(res["median_ttft_ms"], 80)
def _run_lora_latency_test(self, enable_background_task: bool):
"""
Run a latency test for LoRA with the specified background task setting.
"""
async def lora_loader_unloader_task(
base_url: str,
start_event: asyncio.Event,
stop_event: asyncio.Event,
):
"""
A background task that repeatedly loads and unloads a LoRA adapter.
"""
await start_event.wait()
path_cycler = itertools.cycle(
[
"pbevan11/llama-3.1-8b-ocr-correction",
"faridlazuarda/valadapt-llama-3.1-8B-it-chinese",
"philschmid/code-llama-3-1-8b-text-to-sql-lora",
]
)
load_url = f"{base_url}/load_lora_adapter"
unload_url = f"{base_url}/unload_lora_adapter"
num_updates = 0
while not stop_event.is_set():
lora_path = next(path_cycler)
response = await asyncio.to_thread(
requests.post,
load_url,
json={"lora_name": lora_path, "lora_path": lora_path},
)
self.assertTrue(
response.ok, f"Failed to load LoRA adapter: {response.text}"
)
num_updates += 1
if stop_event.is_set():
break
await asyncio.sleep(1)
response = await asyncio.to_thread(
requests.post,
unload_url,
json={"lora_name": lora_path},
)
self.assertTrue(
response.ok, f"Failed to unload LoRA adapter: {response.text}"
)
num_updates += 1
await asyncio.sleep(1)
background_task = lora_loader_unloader_task if enable_background_task else None
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=400,
request_rate=8,
other_server_args=[
"--enable-lora",
"--max-loras-per-batch",
"1",
"--disable-radix-cache",
"--random-seed",
"42",
"--mem-fraction-static",
"0.8",
"--lora-paths",
"Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16",
"--max-lora-rank",
"256",
],
dataset_name="random",
random_input_len=256,
random_output_len=256,
lora_name=["Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16"],
background_task=background_task,
)
return res
if __name__ == "__main__":
unittest.main()
test/registered/perf/test_bench_serving_1gpu_part2.py
+186
View File
@@ -0,0 +1,186 @@
"""
Performance tests for single GPU - VLM, Score API, and Embeddings API tests.
Works on 5090 (32GB).
"""
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import (
DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST,
DEFAULT_SMALL_MODEL_NAME_FOR_TEST_SCORE,
DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
CustomTestCase,
is_in_amd_ci,
is_in_ci,
run_bench_serving,
run_embeddings_benchmark,
run_score_benchmark,
write_github_step_summary,
)
register_cuda_ci(est_time=900, suite="stage-b-test-large-1-gpu-performance")
class TestBenchServing1GPUPart2(CustomTestCase):
def test_vlm_offline_throughput(self):
res = run_bench_serving(
model=DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
num_prompts=200,
request_rate=float("inf"),
other_server_args=[
"--mem-fraction-static",
"0.7",
],
dataset_name="mmmu",
)
if is_in_ci():
write_github_step_summary(
f"### test_vlm_offline_throughput\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 2000)
else:
self.assertGreater(res["output_throughput"], 2500)
def test_vlm_online_latency(self):
res = run_bench_serving(
model=DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
num_prompts=250,
request_rate=1,
other_server_args=[
"--mem-fraction-static",
"0.7",
],
dataset_name="mmmu",
)
if is_in_ci():
write_github_step_summary(
f"### test_vlm_online_latency\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
)
self.assertLess(res["median_e2e_latency_ms"], 16500)
if is_in_amd_ci():
self.assertLess(res["median_ttft_ms"], 150)
else:
self.assertLess(res["median_ttft_ms"], 100)
self.assertLess(res["median_itl_ms"], 8)
def test_score_api_latency_throughput(self):
"""Test score API latency and throughput performance"""
res = run_score_benchmark(
model=DEFAULT_SMALL_MODEL_NAME_FOR_TEST_SCORE,
num_requests=1000,
batch_size=10,
other_server_args=[],
need_warmup=True,
)
if is_in_ci():
write_github_step_summary(
f"### test_score_api_throughput\n"
f"Average latency: {res['avg_latency_ms']:.2f} ms\n"
f"P95 latency: {res['p95_latency_ms']:.2f} ms\n"
f"Score API throughput: {res['throughput']:.2f} req/s\n"
f"Successful requests: {res['successful_requests']}/{res['total_requests']}\n"
)
self.assertEqual(res["successful_requests"], res["total_requests"])
self.assertLess(res["avg_latency_ms"], 48)
self.assertLess(res["p95_latency_ms"], 50)
self.assertGreater(res["throughput"], 20)
def test_score_api_batch_scaling(self):
"""Test score API performance with different batch sizes"""
batch_sizes = [10, 25, 50]
for batch_size in batch_sizes:
res = run_score_benchmark(
model=DEFAULT_SMALL_MODEL_NAME_FOR_TEST_SCORE,
num_requests=500,
batch_size=batch_size,
)
if is_in_ci():
write_github_step_summary(
f"### test_score_api_batch_scaling_size_{batch_size}\n"
f"Batch size: {batch_size}\n"
f"Average latency: {res['avg_latency_ms']:.2f} ms\n"
f"P95 latency: {res['p95_latency_ms']:.2f} ms\n"
f"Throughput: {res['throughput']:.2f} req/s\n"
f"Successful requests: {res['successful_requests']}/{res['total_requests']}\n"
)
self.assertEqual(res["successful_requests"], res["total_requests"])
bounds = {
10: (45, 50),
25: (50, 60),
50: (60, 65),
}
avg_latency_bound, p95_latency_bound = bounds.get(batch_size, (60, 65))
self.assertLess(res["avg_latency_ms"], avg_latency_bound)
self.assertLess(res["p95_latency_ms"], p95_latency_bound)
def test_embeddings_api_latency_throughput(self):
"""Test embeddings API latency and throughput performance"""
res = run_embeddings_benchmark(
model=DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST,
num_requests=1000,
batch_size=1,
input_tokens=500,
other_server_args=[],
need_warmup=True,
)
if is_in_ci():
write_github_step_summary(
f"### test_embeddings_api_throughput\n"
f"Average latency: {res['avg_latency_ms']:.2f} ms\n"
f"P95 latency: {res['p95_latency_ms']:.2f} ms\n"
f"Embeddings API throughput: {res['throughput']:.2f} req/s\n"
f"Successful requests: {res['successful_requests']}/{res['total_requests']}\n"
)
self.assertEqual(res["successful_requests"], res["total_requests"])
self.assertLess(res["avg_latency_ms"], 20)
self.assertLess(res["p95_latency_ms"], 25)
self.assertGreater(res["throughput"], 60)
def test_embeddings_api_batch_scaling(self):
"""Test embeddings API performance with different batch sizes"""
batch_sizes = [10, 25, 50]
for batch_size in batch_sizes:
res = run_embeddings_benchmark(
model=DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST,
num_requests=500,
batch_size=batch_size,
input_tokens=500,
)
if is_in_ci():
write_github_step_summary(
f"### test_embeddings_api_batch_scaling_size_{batch_size}\n"
f"Batch size: {batch_size}\n"
f"Average latency: {res['avg_latency_ms']:.2f} ms\n"
f"P95 latency: {res['p95_latency_ms']:.2f} ms\n"
f"Throughput: {res['throughput']:.2f} req/s\n"
f"Successful requests: {res['successful_requests']}/{res['total_requests']}\n"
)
self.assertEqual(res["successful_requests"], res["total_requests"])
bounds = {
10: (60, 65),
25: (115, 120),
50: (190, 195),
}
avg_latency_bound, p95_latency_bound = bounds.get(batch_size, (250, 250))
self.assertLess(res["avg_latency_ms"], avg_latency_bound)
self.assertLess(res["p95_latency_ms"], p95_latency_bound)
if __name__ == "__main__":
unittest.main()
test/registered/perf/test_bench_serving_2gpu.py
+107
View File
@@ -0,0 +1,107 @@
"""
Performance tests for 2-GPU that need large GPUs (H200 80GB) - MoE and Pipeline Parallel tests.
"""
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import (
DEFAULT_MOE_MODEL_NAME_FOR_TEST,
CustomTestCase,
is_in_amd_ci,
is_in_ci,
run_bench_serving,
write_github_step_summary,
)
register_cuda_ci(est_time=600, suite="stage-b-test-large-2-gpu-performance")
class TestBenchServing2GPU(CustomTestCase):
def test_moe_offline_throughput_default(self):
res = run_bench_serving(
model=DEFAULT_MOE_MODEL_NAME_FOR_TEST,
num_prompts=300,
request_rate=float("inf"),
other_server_args=["--tp", "2"],
)
if is_in_ci():
write_github_step_summary(
f"### test_moe_offline_throughput_default\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
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(
model=DEFAULT_MOE_MODEL_NAME_FOR_TEST,
num_prompts=300,
request_rate=float("inf"),
other_server_args=["--tp", "2", "--disable-radix-cache"],
)
if is_in_ci():
write_github_step_summary(
f"### test_moe_offline_throughput_without_radix_cache\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 2100)
else:
self.assertGreater(res["output_throughput"], 2200)
def test_pp_offline_throughput_default_decode(self):
res = run_bench_serving(
model=DEFAULT_MOE_MODEL_NAME_FOR_TEST,
num_prompts=1000,
request_rate=float("inf"),
random_input_len=1,
random_output_len=1024,
other_server_args=["--pp-size", "2"],
need_warmup=True,
seed=42,
)
if is_in_ci():
write_github_step_summary(
f"### test_pp_offline_throughput_default_decode\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
self.assertGreater(res["output_throughput"], 6700)
def test_pp_long_context_prefill(self):
res = run_bench_serving(
model="meta-llama/Llama-3.3-70B-Instruct",
num_prompts=4,
request_rate=float("inf"),
random_input_len=128000,
random_output_len=1,
dataset_name="random",
other_server_args=[
"--quantization",
"fp8",
"--pp-size",
"2",
]
+ (["--mem-fraction-static", "0.7"] if is_in_amd_ci() else []),
need_warmup=False,
seed=42,
)
if is_in_ci():
write_github_step_summary(
f"### test_pp_long_context_latency_prefill\n"
f"input_throughput: {res['input_throughput']:.2f} ms\n"
)
if is_in_amd_ci():
self.assertGreater(res["input_throughput"], 3000)
else:
self.assertGreater(res["input_throughput"], 4000)
if __name__ == "__main__":
unittest.main()
test/registered/perf/test_vlm_perf_5090.py
+62
View File
@@ -0,0 +1,62 @@
"""
VLM Performance tests that work on 5090 (32GB) - VLM offline throughput and online latency tests.
"""
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import (
DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
CustomTestCase,
is_in_ci,
run_bench_serving,
write_github_step_summary,
)
register_cuda_ci(est_time=600, suite="stage-b-test-small-1-gpu-performance")
class TestVLMPerf5090(CustomTestCase):
def test_vlm_offline_throughput(self):
res = run_bench_serving(
model=DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
num_prompts=200,
request_rate=float("inf"),
other_server_args=[
"--mem-fraction-static",
"0.7",
],
dataset_name="mmmu",
)
if is_in_ci():
write_github_step_summary(
f"### test_vlm_offline_throughput (5090)\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
self.assertGreater(res["output_throughput"], 2000)
def test_vlm_online_latency(self):
res = run_bench_serving(
model=DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
num_prompts=250,
request_rate=1,
other_server_args=[
"--mem-fraction-static",
"0.7",
],
dataset_name="mmmu",
)
if is_in_ci():
write_github_step_summary(
f"### test_vlm_online_latency (5090)\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
)
self.assertLess(res["median_e2e_latency_ms"], 16500)
self.assertLess(res["median_ttft_ms"], 150)
self.assertLess(res["median_itl_ms"], 8)
if __name__ == "__main__":
unittest.main()
test/srt/quant/test_awq.py → test/registered/quant/test_awq.py
+3
View File
@@ -2,6 +2,7 @@ import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_AWQ_MOE_MODEL_NAME_FOR_TEST,
@@ -11,6 +12,8 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=163, suite="stage-b-test-large-1-gpu")
class TestAWQ(CustomTestCase):
@classmethod
test/srt/test_bnb.py → test/registered/quant/test_bnb.py
+3
View File
@@ -12,6 +12,7 @@ from types import SimpleNamespace
import openai
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
@@ -21,6 +22,8 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=5, suite="stage-b-test-small-1-gpu")
VISION_MODELS = [
"unsloth/Qwen2.5-VL-7B-Instruct-bnb-4bit",
"unsloth/Qwen2-VL-7B-Instruct-bnb-4bit",
test/srt/test_gguf.py → test/registered/quant/test_gguf.py
+3
View File
@@ -3,8 +3,11 @@ import unittest
from huggingface_hub import hf_hub_download
import sglang as sgl
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import CustomTestCase
register_cuda_ci(est_time=96, suite="stage-b-test-small-1-gpu")
class TestGGUF(CustomTestCase):
def test_models(self):
test/srt/test_gptqmodel_dynamic.py → test/registered/quant/test_gptqmodel_dynamic.py
+3
View File
@@ -6,6 +6,7 @@ import torch
from sglang.srt.server_args import set_global_server_args_for_scheduler
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
@@ -13,6 +14,8 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=102, suite="stage-b-test-large-1-gpu")
def check_quant_method(model_path: str, use_marlin_kernel: bool):
from sglang.srt.configs.device_config import DeviceConfig
test/srt/quant/test_marlin_moe.py → test/registered/quant/test_marlin_moe.py
+3
View File
@@ -8,9 +8,12 @@ from sgl_kernel.scalar_type import scalar_types
from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.moe.fused_moe_triton.fused_marlin_moe import fused_marlin_moe
from sglang.srt.server_args import ServerArgs, set_global_server_args_for_scheduler
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_marlin_utils import awq_marlin_quantize, marlin_quantize
from sglang.test.test_utils import CustomTestCase
register_cuda_ci(est_time=200, suite="stage-b-test-small-1-gpu")
set_global_server_args_for_scheduler(object.__new__(ServerArgs))
test/srt/test_quantization.py → test/registered/quant/test_quantization.py
+3
View File
@@ -4,6 +4,7 @@ import warnings
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_QUANT_TP1,
@@ -15,6 +16,8 @@ from sglang.test.test_utils import (
write_results_to_json,
)
register_cuda_ci(est_time=185, suite="stage-b-test-large-1-gpu")
MODEL_SCORE_THRESHOLDS = {
"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4": 0.825,
"hugging-quants/Meta-Llama-3.1-8B-Instruct-GPTQ-INT4": 0.825,
test/run_suite.py
+4
View File
@@ -28,8 +28,12 @@ PER_COMMIT_SUITES = {
HWBackend.CUDA: [
"stage-a-test-1",
"stage-b-test-small-1-gpu",
"stage-b-test-small-1-gpu-performance",
"stage-b-test-small-1-gpu-accuracy",
"stage-b-test-large-1-gpu",
"stage-b-test-large-1-gpu-performance",
"stage-b-test-large-2-gpu",
"stage-b-test-large-2-gpu-performance",
"stage-c-test-large-4-gpu",
"stage-b-test-4-gpu-b200",
"stage-c-test-large-4-gpu-b200",
test/srt/run_suite.py
+1 -12
View File
@@ -53,21 +53,10 @@ suites = {
# "per-commit-8-gpu-h200-deepep": [
# TestFile("ep/test_deepep_large.py", 563),
# ],
"quantization_test": [
TestFile("quant/test_awq.py", 163),
TestFile("quant/test_marlin_moe.py", 200),
TestFile("test_bnb.py", 5),
TestFile("test_gptqmodel_dynamic.py", 102),
TestFile("test_quantization.py", 185),
TestFile("test_gguf.py", 96),
],
# quantization_test suite migrated to test/registered/quant/
"__not_in_ci__": [
TestFile("test_release_memory_occupation.py", 200), # Temporarily disabled
TestFile("models/test_dummy_grok_models.py"),
TestFile("test_bench_one_batch.py"),
TestFile("test_bench_serving.py"),
TestFile("test_eval_accuracy_large.py"),
TestFile("test_moe_eval_accuracy_large.py"),
TestFile("test_profile_v2.py"),
TestFile("models/test_ministral3_models.py"),
TestFile("test_mistral_large3_basic.py"),
test/srt/test_bench_serving.py
-566
View File
@@ -1,566 +0,0 @@
import asyncio
import itertools
import unittest
import requests
from sglang.test.test_utils import (
DEFAULT_DRAFT_MODEL_EAGLE,
DEFAULT_MODEL_NAME_FOR_TEST,
DEFAULT_MODEL_NAME_FOR_TEST_FP8,
DEFAULT_MOE_MODEL_NAME_FOR_TEST,
DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST,
DEFAULT_SMALL_MODEL_NAME_FOR_TEST_SCORE,
DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
DEFAULT_TARGET_MODEL_EAGLE,
CustomTestCase,
is_in_amd_ci,
is_in_ci,
run_bench_serving,
run_embeddings_benchmark,
run_score_benchmark,
write_github_step_summary,
)
class TestBenchServing(CustomTestCase):
def test_offline_throughput_default(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=500,
request_rate=float("inf"),
other_server_args=[],
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_default\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 3050)
else:
self.assertGreater(res["output_throughput"], 3800)
def test_offline_throughput_non_stream_small_batch_size(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=200,
request_rate=float("inf"),
other_server_args=["--max-running-requests", "10"],
dataset_name="sharegpt",
random_input_len=None,
random_output_len=None,
disable_stream=True,
need_warmup=True,
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_non_stream_small_batch_size\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 1000)
else:
self.assertGreater(res["output_throughput"], 1050)
def test_offline_throughput_without_radix_cache(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=500,
request_rate=float("inf"),
other_server_args=["--disable-radix-cache"],
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_without_radix_cache\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 3050)
else:
self.assertGreater(res["output_throughput"], 3800)
def test_offline_throughput_without_chunked_prefill(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=500,
request_rate=float("inf"),
other_server_args=["--chunked-prefill-size", "-1"],
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_without_chunked_prefill\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
self.assertGreater(res["output_throughput"], 2600)
def test_offline_throughput_with_triton_attention_backend(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=500,
request_rate=float("inf"),
other_server_args=[
"--attention-backend",
"triton",
"--context-length",
"8192",
],
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_with_triton_attention_backend\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 3500)
else:
self.assertGreater(res["output_throughput"], 3700)
def test_offline_throughput_default_fp8(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST_FP8,
num_prompts=500,
request_rate=float("inf"),
other_server_args=[],
)
if is_in_ci():
write_github_step_summary(
f"### test_offline_throughput_default_fp8\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 3500)
else:
self.assertGreater(res["output_throughput"], 4300)
def test_online_latency_default(self):
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=100,
request_rate=1,
other_server_args=[],
)
if is_in_ci():
write_github_step_summary(
f"### test_online_latency_default\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
)
self.assertLess(res["median_e2e_latency_ms"], 11000)
if is_in_amd_ci():
self.assertLess(res["median_ttft_ms"], 115)
else:
self.assertLess(res["median_ttft_ms"], 86)
self.assertLess(res["median_itl_ms"], 10)
def test_vlm_offline_throughput(self):
res = run_bench_serving(
model=DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
num_prompts=200,
request_rate=float("inf"),
other_server_args=[
"--mem-fraction-static",
"0.7",
],
dataset_name="mmmu",
)
if is_in_ci():
write_github_step_summary(
f"### test_vlm_offline_throughput\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 2000)
# TODO: not set yet, need AMD machine
else:
self.assertGreater(res["output_throughput"], 2500)
def test_vlm_online_latency(self):
res = run_bench_serving(
model=DEFAULT_SMALL_VLM_MODEL_NAME_FOR_TEST,
num_prompts=250,
request_rate=1,
other_server_args=[
"--mem-fraction-static",
"0.7",
],
dataset_name="mmmu",
)
if is_in_ci():
write_github_step_summary(
f"### test_vlm_online_latency\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
)
self.assertLess(res["median_e2e_latency_ms"], 16500)
if is_in_amd_ci():
self.assertLess(res["median_ttft_ms"], 150)
# TODO: not set yet, need AMD machine
else:
self.assertLess(res["median_ttft_ms"], 100)
self.assertLess(res["median_itl_ms"], 8)
def test_lora_online_latency(self):
# TODO (lifuhuang): verify LoRA support in AMD.
if is_in_amd_ci():
pass
res = self._run_lora_latency_test(enable_background_task=False)
if is_in_ci():
write_github_step_summary(
f"### test_lora_online_latency\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n"
)
self.assertLess(res["median_e2e_latency_ms"], 2400)
self.assertLess(res["median_ttft_ms"], 58)
def test_lora_online_latency_with_concurrent_adapter_updates(self):
# TODO (lifuhuang): verify LoRA support in AMD.
if is_in_amd_ci():
pass
res = self._run_lora_latency_test(enable_background_task=True)
if is_in_ci():
write_github_step_summary(
f"### test_lora_online_latency\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n"
)
self.assertLess(res["median_e2e_latency_ms"], 4000)
self.assertLess(res["median_ttft_ms"], 80)
def _run_lora_latency_test(self, enable_background_task: bool):
"""
Run a latency test for LoRA with the specified background task setting.
"""
async def lora_loader_unloader_task(
base_url: str,
start_event: asyncio.Event,
stop_event: asyncio.Event,
):
"""
A background task that repeatedly loads and unloads a LoRA adapter.
"""
await start_event.wait()
path_cycler = itertools.cycle(
[
"pbevan11/llama-3.1-8b-ocr-correction",
"faridlazuarda/valadapt-llama-3.1-8B-it-chinese",
"philschmid/code-llama-3-1-8b-text-to-sql-lora",
]
)
load_url = f"{base_url}/load_lora_adapter"
unload_url = f"{base_url}/unload_lora_adapter"
num_updates = 0
while not stop_event.is_set():
# 1. Load the LoRA adapter
lora_path = next(path_cycler)
response = await asyncio.to_thread(
requests.post,
load_url,
json={"lora_name": lora_path, "lora_path": lora_path},
)
self.assertTrue(
response.ok, f"Failed to load LoRA adapter: {response.text}"
)
num_updates += 1
if stop_event.is_set():
break
# Yield control to allow other tasks to run.
await asyncio.sleep(1)
# 2. Unload the LoRA adapter
response = await asyncio.to_thread(
requests.post,
unload_url,
json={"lora_name": lora_path},
)
self.assertTrue(
response.ok, f"Failed to unload LoRA adapter: {response.text}"
)
num_updates += 1
# Yield control to allow other tasks to run.
await asyncio.sleep(1)
background_task = lora_loader_unloader_task if enable_background_task else None
res = run_bench_serving(
model=DEFAULT_MODEL_NAME_FOR_TEST,
num_prompts=400,
request_rate=8,
other_server_args=[
"--enable-lora",
"--max-loras-per-batch",
"1",
"--disable-radix-cache",
"--random-seed",
"42",
"--mem-fraction-static",
"0.8",
"--lora-paths",
"Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16",
"--max-lora-rank",
"256",
],
dataset_name="random",
random_input_len=256,
random_output_len=256,
lora_name=["Nutanix/Meta-Llama-3.1-8B-Instruct_lora_4_alpha_16"],
background_task=background_task,
)
return res
def test_online_latency_eagle(self):
res = run_bench_serving(
model=DEFAULT_TARGET_MODEL_EAGLE,
num_prompts=300,
request_rate=8,
sharegpt_context_len=3072,
disable_ignore_eos=True,
dataset_name="sharegpt",
other_server_args=[
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
DEFAULT_DRAFT_MODEL_EAGLE,
"--speculative-num-steps",
"5",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"16",
"--mem-fraction-static",
"0.7",
],
need_warmup=True,
seed=42,
)
if is_in_ci():
write_github_step_summary(
f"### test_online_latency_eagle\n"
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
f"accept_length: {res['accept_length']:.2f} \n"
)
if is_in_amd_ci():
self.assertLess(res["median_e2e_latency_ms"], 1800)
else:
self.assertLess(res["median_e2e_latency_ms"], 900)
self.assertGreater(res["accept_length"], 3.0)
def test_moe_offline_throughput_default(self):
res = run_bench_serving(
model=DEFAULT_MOE_MODEL_NAME_FOR_TEST,
num_prompts=300,
request_rate=float("inf"),
other_server_args=["--tp", "2"],
)
if is_in_ci():
write_github_step_summary(
f"### test_moe_offline_throughput_default\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
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(
model=DEFAULT_MOE_MODEL_NAME_FOR_TEST,
num_prompts=300,
request_rate=float("inf"),
other_server_args=["--tp", "2", "--disable-radix-cache"],
)
if is_in_ci():
write_github_step_summary(
f"### test_moe_offline_throughput_without_radix_cache\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 2100)
else:
self.assertGreater(res["output_throughput"], 2200)
def test_pp_offline_throughput_default_decode(self):
res = run_bench_serving(
model=DEFAULT_MOE_MODEL_NAME_FOR_TEST,
num_prompts=1000,
request_rate=float("inf"),
random_input_len=1,
random_output_len=1024,
other_server_args=["--pp-size", "2"],
need_warmup=True,
seed=42,
)
if is_in_ci():
write_github_step_summary(
f"### test_pp_offline_throughput_default_decode\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n"
)
self.assertGreater(res["output_throughput"], 6700)
def test_pp_long_context_prefill(self):
res = run_bench_serving(
model="meta-llama/Llama-3.3-70B-Instruct",
num_prompts=4,
request_rate=float("inf"),
random_input_len=128000,
random_output_len=1,
dataset_name="random",
other_server_args=[
"--quantization",
"fp8",
"--pp-size",
"2",
]
+ (["--mem-fraction-static", "0.7"] if is_in_amd_ci() else []),
need_warmup=False,
seed=42,
)
if is_in_ci():
write_github_step_summary(
f"### test_pp_long_context_latency_prefill\n"
f"input_throughput: {res['input_throughput']:.2f} ms\n"
)
if is_in_amd_ci():
self.assertGreater(res["input_throughput"], 3000)
else:
self.assertGreater(res["input_throughput"], 4000)
def test_score_api_latency_throughput(self):
"""Test score API latency and throughput performance"""
res = run_score_benchmark(
model=DEFAULT_SMALL_MODEL_NAME_FOR_TEST_SCORE,
num_requests=1000,
batch_size=10,
other_server_args=[],
need_warmup=True,
)
if is_in_ci():
write_github_step_summary(
f"### test_score_api_throughput\n"
f"Average latency: {res['avg_latency_ms']:.2f} ms\n"
f"P95 latency: {res['p95_latency_ms']:.2f} ms\n"
f"Score API throughput: {res['throughput']:.2f} req/s\n"
f"Successful requests: {res['successful_requests']}/{res['total_requests']}\n"
)
self.assertEqual(res["successful_requests"], res["total_requests"])
self.assertLess(res["avg_latency_ms"], 48)
self.assertLess(res["p95_latency_ms"], 50)
self.assertGreater(res["throughput"], 20)
def test_score_api_batch_scaling(self):
"""Test score API performance with different batch sizes"""
batch_sizes = [10, 25, 50]
for batch_size in batch_sizes:
res = run_score_benchmark(
model=DEFAULT_SMALL_MODEL_NAME_FOR_TEST_SCORE,
num_requests=500,
batch_size=batch_size,
)
if is_in_ci():
write_github_step_summary(
f"### test_score_api_batch_scaling_size_{batch_size}\n"
f"Batch size: {batch_size}\n"
f"Average latency: {res['avg_latency_ms']:.2f} ms\n"
f"P95 latency: {res['p95_latency_ms']:.2f} ms\n"
f"Throughput: {res['throughput']:.2f} req/s\n"
f"Successful requests: {res['successful_requests']}/{res['total_requests']}\n"
)
self.assertEqual(res["successful_requests"], res["total_requests"])
bounds = {
10: (45, 50),
25: (50, 60),
50: (60, 65),
}
avg_latency_bound, p95_latency_bound = bounds.get(batch_size, (60, 65))
self.assertLess(res["avg_latency_ms"], avg_latency_bound)
self.assertLess(res["p95_latency_ms"], p95_latency_bound)
def test_embeddings_api_latency_throughput(self):
"""Test embeddings API latency and throughput performance"""
res = run_embeddings_benchmark(
model=DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST,
num_requests=1000,
batch_size=1,
input_tokens=500,
other_server_args=[],
need_warmup=True,
)
if is_in_ci():
write_github_step_summary(
f"### test_embeddings_api_throughput\n"
f"Average latency: {res['avg_latency_ms']:.2f} ms\n"
f"P95 latency: {res['p95_latency_ms']:.2f} ms\n"
f"Embeddings API throughput: {res['throughput']:.2f} req/s\n"
f"Successful requests: {res['successful_requests']}/{res['total_requests']}\n"
)
self.assertEqual(res["successful_requests"], res["total_requests"])
# Bounds based on actual performance on 1xH100: avg=15ms, p95=15ms, throughput=67req/s
self.assertLess(res["avg_latency_ms"], 20)
self.assertLess(res["p95_latency_ms"], 25)
self.assertGreater(res["throughput"], 60)
def test_embeddings_api_batch_scaling(self):
"""Test embeddings API performance with different batch sizes"""
batch_sizes = [10, 25, 50]
for batch_size in batch_sizes:
res = run_embeddings_benchmark(
model=DEFAULT_SMALL_EMBEDDING_MODEL_NAME_FOR_TEST,
num_requests=500,
batch_size=batch_size,
input_tokens=500,
)
if is_in_ci():
write_github_step_summary(
f"### test_embeddings_api_batch_scaling_size_{batch_size}\n"
f"Batch size: {batch_size}\n"
f"Average latency: {res['avg_latency_ms']:.2f} ms\n"
f"P95 latency: {res['p95_latency_ms']:.2f} ms\n"
f"Throughput: {res['throughput']:.2f} req/s\n"
f"Successful requests: {res['successful_requests']}/{res['total_requests']}\n"
)
self.assertEqual(res["successful_requests"], res["total_requests"])
bounds = {
10: (60, 65),
25: (115, 120),
50: (190, 195),
}
avg_latency_bound, p95_latency_bound = bounds.get(batch_size, (250, 250))
self.assertLess(res["avg_latency_ms"], avg_latency_bound)
self.assertLess(res["p95_latency_ms"], p95_latency_bound)
if __name__ == "__main__":
unittest.main()