147 lines
5.3 KiB
Python
147 lines
5.3 KiB
Python
"""AMD Nightly performance benchmark for VLM models (2-GPU).
|
|
|
|
This test benchmarks Vision-Language Models on AMD MI30x/MI35x with 2 GPUs.
|
|
|
|
Registry: nightly-amd-perf-vlm-2-gpu suite
|
|
|
|
Example usage:
|
|
python -m pytest test_vlms_perf_amd.py -v
|
|
"""
|
|
|
|
import os
|
|
import unittest
|
|
import warnings
|
|
from typing import List
|
|
|
|
from sglang.test.ci.ci_register import register_amd_ci
|
|
from sglang.test.nightly_bench_utils import BenchmarkResult
|
|
from sglang.test.nightly_utils import NightlyBenchmarkRunner
|
|
from sglang.test.test_utils import (
|
|
DEFAULT_URL_FOR_TEST,
|
|
ModelLaunchSettings,
|
|
_parse_int_list_env,
|
|
parse_models,
|
|
)
|
|
|
|
# Register for AMD CI - VLM models benchmark (~120 min)
|
|
register_amd_ci(est_time=7200, suite="nightly-amd-perf-vlm-2-gpu", nightly=True)
|
|
|
|
PROFILE_DIR = "performance_profiles_vlms_amd"
|
|
|
|
# VLM models suitable for AMD
|
|
MODEL_DEFAULTS = [
|
|
ModelLaunchSettings(
|
|
"Qwen/Qwen2.5-VL-7B-Instruct",
|
|
extra_args=["--mem-fraction-static=0.7"],
|
|
),
|
|
ModelLaunchSettings(
|
|
"Qwen/Qwen3-VL-30B-A3B-Instruct",
|
|
tp_size=2,
|
|
),
|
|
]
|
|
|
|
|
|
def generate_simple_markdown_report(results: List[BenchmarkResult]) -> str:
|
|
"""Generate a simplified markdown report without traces and cost columns.
|
|
|
|
Skips the first result if it's a warmup run (duplicate batch_size).
|
|
"""
|
|
model_header = results[0].model_path
|
|
if results[0].run_name and results[0].run_name != "default":
|
|
model_header += f" ({results[0].run_name})"
|
|
|
|
gpu_config = os.getenv("GPU_CONFIG", "AMD")
|
|
if gpu_config:
|
|
model_header += f" [{gpu_config}]"
|
|
|
|
summary = f"### {model_header}\n"
|
|
summary += "| batch size | input len | latency (s) | input throughput (tok/s) | output throughput (tok/s) | ITL (ms) |\n"
|
|
summary += "| ---------- | --------- | ----------- | ------------------------ | ------------------------- | -------- |\n"
|
|
|
|
# Skip first result if it's a warmup (same batch_size as second result)
|
|
report_results = (
|
|
results[1:]
|
|
if len(results) > 1 and results[0].batch_size == results[1].batch_size
|
|
else results
|
|
)
|
|
|
|
for result in report_results:
|
|
itl = 1 / (result.output_throughput / result.batch_size) * 1000
|
|
summary += f"| {result.batch_size} | {result.input_len} | {result.latency:.2f} | {result.input_throughput:.2f} | {result.output_throughput:.2f} | {itl:.2f} |\n"
|
|
|
|
return summary
|
|
|
|
|
|
class TestNightlyVLMsPerfAMD(unittest.TestCase):
|
|
"""AMD Nightly performance benchmark for VLM models (2-GPU)."""
|
|
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
warnings.filterwarnings(
|
|
"ignore", category=ResourceWarning, message="unclosed.*socket"
|
|
)
|
|
|
|
nightly_vlm_models_str = os.environ.get("NIGHTLY_VLM_MODELS")
|
|
if nightly_vlm_models_str:
|
|
cls.models = []
|
|
model_paths = parse_models(nightly_vlm_models_str)
|
|
for model_path in model_paths:
|
|
cls.models.append(ModelLaunchSettings(model_path))
|
|
else:
|
|
cls.models = MODEL_DEFAULTS
|
|
|
|
cls.base_url = DEFAULT_URL_FOR_TEST
|
|
# First batch_size=1 is warmup (standalone job, no accuracy test to warm up)
|
|
cls.batch_sizes = _parse_int_list_env("NIGHTLY_VLM_BATCH_SIZES", "1,1,2,8,16")
|
|
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_VLM_INPUT_LENS", "4096"))
|
|
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_VLM_OUTPUT_LENS", "512"))
|
|
cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
|
|
cls.runner.setup_profile_directory()
|
|
cls.runner.full_report = f"## {cls.__name__}\n"
|
|
|
|
def test_bench_one_batch(self):
|
|
"""Run benchmark for all configured VLM models."""
|
|
all_model_succeed = True
|
|
|
|
try:
|
|
for model_setup in self.models:
|
|
with self.subTest(model=model_setup.model_path):
|
|
other_args = list(model_setup.extra_args or [])
|
|
if model_setup.tp_size and model_setup.tp_size > 1:
|
|
other_args.extend(["--tp", str(model_setup.tp_size)])
|
|
|
|
# VLMs need additional benchmark args for dataset and trust-remote-code
|
|
extra_bench_args = [
|
|
"--trust-remote-code",
|
|
"--dataset-name=mmmu",
|
|
]
|
|
|
|
result_tuple = self.runner.run_benchmark_for_model(
|
|
model_path=model_setup.model_path,
|
|
batch_sizes=self.batch_sizes,
|
|
input_lens=self.input_lens,
|
|
output_lens=self.output_lens,
|
|
other_args=other_args,
|
|
extra_bench_args=extra_bench_args,
|
|
enable_profile=False, # Disable profiling for AMD tests
|
|
)
|
|
results = result_tuple[0]
|
|
success = result_tuple[1]
|
|
|
|
if not success:
|
|
all_model_succeed = False
|
|
|
|
if results:
|
|
self.runner.full_report += (
|
|
generate_simple_markdown_report(results) + "\n"
|
|
)
|
|
finally:
|
|
self.runner.write_final_report()
|
|
|
|
if not all_model_succeed:
|
|
raise AssertionError("Some models failed the perf tests.")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|