minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 10 Packages Projects Releases Wiki Activity

[AMD] Add Qwen3.5-397B FP8 nightly perf benchmarks for MI30x and MI35x (#21669)

Browse Source
This commit is contained in:
Michael
2026-04-06 23:46:00 -07:00
committed by GitHub
parent 3f7dfba419
commit ba78f6e0ef
6 changed files with 408 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-amd-rocm720.yml
+26 -2
View File
@@ -621,7 +621,7 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# 8-GPU Qwen 3.5 (Accuracy) ROCm 7.2
# 8-GPU Qwen 3.5 (Accuracy + Performance combined) ROCm 7.2
nightly-8-gpu-qwen35-rocm720:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-qwen35-rocm720,'))
runs-on: linux-mi325-8gpu-sglang
@@ -653,6 +653,18 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
- name: Performance Test ROCm 7.2 (8-GPU Qwen 3.5 FP8)
timeout-minutes: 120
continue-on-error: true
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-qwen35-fp8 --nightly --timeout-per-file 5400 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# 8-GPU GLM-5 (Accuracy) ROCm 7.2
nightly-8-gpu-glm5-rocm720:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-glm5-rocm720,'))
@@ -1219,7 +1231,7 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU Qwen 3.5 (Accuracy) ROCm 7.2
# MI35x 8-GPU Qwen 3.5 (Accuracy + Performance combined) ROCm 7.2
nightly-8-gpu-mi35x-qwen35-rocm720:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-qwen35-rocm720,'))
runs-on: linux-mi35x-gpu-8
@@ -1252,6 +1264,18 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
- name: Performance Test MI35x ROCm 7.2 (8-GPU Qwen 3.5 FP8)
timeout-minutes: 120
continue-on-error: true
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-mi35x-qwen35-fp8 --nightly --timeout-per-file 5400 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
nightly-8-gpu-mi35x-glm5-rocm720:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-glm5-rocm720,'))
runs-on: linux-mi35x-gpu-8
.github/workflows/nightly-test-amd.yml
+26 -2
View File
@@ -624,7 +624,7 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# 8-GPU Qwen 3.5 (Accuracy)
# 8-GPU Qwen 3.5 (Accuracy + Performance combined)
nightly-8-gpu-qwen35:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-qwen35,'))
runs-on: linux-mi325-8gpu-sglang
@@ -656,6 +656,18 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
- name: Performance Test (8-GPU Qwen 3.5 FP8)
timeout-minutes: 120
continue-on-error: true
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-qwen35-fp8 --nightly --timeout-per-file 5400 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
nightly-8-gpu-glm5:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-glm5,'))
runs-on: linux-mi325-8gpu-sglang
@@ -1224,7 +1236,7 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# MI35x 8-GPU Qwen 3.5 (Accuracy)
# MI35x 8-GPU Qwen 3.5 (Accuracy + Performance combined)
nightly-8-gpu-mi35x-qwen35:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-qwen35,'))
runs-on: linux-mi35x-gpu-8
@@ -1257,6 +1269,18 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
- name: Performance Test MI35x (8-GPU Qwen 3.5 FP8)
timeout-minutes: 120
continue-on-error: true
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-mi35x-qwen35-fp8 --nightly --timeout-per-file 5400 ${{ inputs.continue_on_error && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
nightly-8-gpu-mi35x-glm5:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-glm5,'))
runs-on: linux-mi35x-gpu-8
test/registered/amd/accuracy/mi30x/test_qwen35_eval_amd.py
+42 -1
View File
@@ -8,6 +8,10 @@ Registry: nightly-amd-accuracy-8-gpu-qwen35 suite
import os
import unittest
from pathlib import Path
import numpy as np
import yaml
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci
@@ -15,7 +19,9 @@ from sglang.test.kits.lm_eval_kit import LMEvalMixin
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
register_amd_ci(est_time=3600, suite="nightly-amd-accuracy-8-gpu-qwen35", nightly=True)
@@ -38,7 +44,7 @@ class TestQwen35EvalAMD(LMEvalMixin, CustomTestCase):
"--tp",
str(TP_SIZE),
"--attention-backend",
"triton",
"aiter",
"--trust-remote-code",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
@@ -59,6 +65,41 @@ class TestQwen35EvalAMD(LMEvalMixin, CustomTestCase):
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_lm_eval(self):
"""Override to write accuracy results to GitHub step summary."""
import requests
requests.get(self.base_url + "/flush_cache")
eval_config = yaml.safe_load(
Path(self.model_config_name).read_text(encoding="utf-8")
)
results = self.launch_lm_eval(eval_config)
rtol = eval_config.get("rtol", self.default_rtol)
model_name = eval_config.get("model_name", self.model)
success = True
summary = f"### lm-eval accuracy ({model_name})\n"
summary += "| task | metric | expected | measured | status |\n"
summary += "| ---- | ------ | -------- | -------- | ------ |\n"
for task in eval_config["tasks"]:
for metric in task["metrics"]:
expected = metric["value"]
measured = results["results"][task["name"]][metric["name"]]
passed = bool(np.isclose(expected, measured, rtol=rtol))
status = "✅" if passed else "❌"
summary += f"| {task['name']} | {metric['name']} | {expected:.4f} | {measured:.4f} | {status} |\n"
print(
f"{task['name']} | {metric['name']}: "
f"expected={expected:.3f} | measured={measured:.3f} | rtol={rtol}"
)
success = success and passed
if is_in_ci():
write_github_step_summary(summary)
self.assertTrue(success, "lm-eval validation failed")
if __name__ == "__main__":
unittest.main()
test/registered/amd/accuracy/mi35x/test_qwen35_eval_mi35x.py
+36 -3
View File
@@ -8,8 +8,11 @@ Registry: nightly-amd-accuracy-8-gpu-mi35x-qwen35 suite
import os
import unittest
from pathlib import Path
import numpy as np
import requests
import yaml
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci
@@ -17,7 +20,9 @@ from sglang.test.kits.lm_eval_kit import LMEvalMixin
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
register_amd_ci(
@@ -40,12 +45,12 @@ class TestQwen35EvalMI35x(LMEvalMixin, CustomTestCase):
cls.base_url = DEFAULT_URL_FOR_TEST
def test_lm_eval(self):
"""Override to handle server lifecycle within test method (MI35x pattern)."""
"""Override to handle server lifecycle and write results to summary."""
other_args = [
"--tp",
str(TP_SIZE),
"--attention-backend",
"triton",
"aiter",
"--trust-remote-code",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
@@ -65,7 +70,35 @@ class TestQwen35EvalMI35x(LMEvalMixin, CustomTestCase):
try:
requests.get(self.base_url + "/flush_cache")
super().test_lm_eval()
eval_config = yaml.safe_load(
Path(self.model_config_name).read_text(encoding="utf-8")
)
results = self.launch_lm_eval(eval_config)
rtol = eval_config.get("rtol", self.default_rtol)
model_name = eval_config.get("model_name", self.model)
success = True
summary = f"### lm-eval accuracy ({model_name})\n"
summary += "| task | metric | expected | measured | status |\n"
summary += "| ---- | ------ | -------- | -------- | ------ |\n"
for task in eval_config["tasks"]:
for metric in task["metrics"]:
expected = metric["value"]
measured = results["results"][task["name"]][metric["name"]]
passed = bool(np.isclose(expected, measured, rtol=rtol))
status = "✅" if passed else "❌"
summary += f"| {task['name']} | {metric['name']} | {expected:.4f} | {measured:.4f} | {status} |\n"
print(
f"{task['name']} | {metric['name']}: "
f"expected={expected:.3f} | measured={measured:.3f} | rtol={rtol}"
)
success = success and passed
if is_in_ci():
write_github_step_summary(summary)
self.assertTrue(success, "lm-eval validation failed")
finally:
kill_process_tree(process.pid)
test/registered/amd/perf/mi30x/test_qwen35_fp8_perf_amd.py
+139
View File
@@ -0,0 +1,139 @@
"""Nightly performance benchmark for Qwen3.5-397B-A17B FP8.
Tests Qwen3.5-397B-A17B-FP8 (MoE, Hybrid Attention with Gated Delta Networks)
on 8 GPUs with triton attention backend.
Model path can be configured via environment variable:
- QWEN35_FP8_MODEL_PATH: Path to Qwen3.5-FP8 model
(default: Qwen/Qwen3.5-397B-A17B-FP8)
Example usage:
python -m pytest test_qwen35_fp8_perf_amd.py -v
"""
import os
import unittest
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, _parse_int_list_env
register_amd_ci(est_time=5400, suite="nightly-perf-8-gpu-qwen35-fp8", nightly=True)
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", "MI325")
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"
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
QWEN35_FP8_MODEL_PATH = os.environ.get(
"QWEN35_FP8_MODEL_PATH", "Qwen/Qwen3.5-397B-A17B-FP8"
)
PROFILE_DIR = "performance_profiles_qwen35_fp8"
class TestNightlyQwen35Fp8Performance(unittest.TestCase):
"""Nightly performance benchmark for Qwen3.5-397B-A17B FP8.
Tests Qwen3.5 FP8 with triton attention backend on TP=8.
Runtime: ~90 minutes
"""
@classmethod
def setUpClass(cls):
cls.base_url = DEFAULT_URL_FOR_TEST
cls.batch_sizes = [1, 8, 16, 64]
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
cls.model_config = {
"name": "qwen35-fp8",
"model_path": QWEN35_FP8_MODEL_PATH,
"other_args": [
"--trust-remote-code",
"--tp",
"8",
"--attention-backend",
"aiter",
"--mem-fraction-static",
"0.8",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
"--watchdog-timeout",
"1200",
],
"env_vars": {
"SGLANG_USE_AITER": "1",
},
}
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_qwen35_fp8(self):
"""Run benchmark for Qwen3.5-397B-A17B FP8."""
old_env = {}
for key, value in self.model_config.get("env_vars", {}).items():
old_env[key] = os.environ.get(key)
os.environ[key] = value
try:
result_tuple = self.runner.run_benchmark_for_model(
model_path=self.model_config["model_path"],
batch_sizes=self.batch_sizes,
input_lens=self.input_lens,
output_lens=self.output_lens,
other_args=self.model_config["other_args"],
variant=self.model_config["name"],
extra_bench_args=["--trust-remote-code"],
enable_profile=False,
timeout=5400,
)
results = result_tuple[0]
success = result_tuple[1]
if results:
self.runner.full_report += (
generate_simple_markdown_report(results) + "\n"
)
self.assertTrue(success, f"Benchmark failed for {QWEN35_FP8_MODEL_PATH}")
finally:
for key, value in old_env.items():
if value is None:
os.environ.pop(key, None)
else:
os.environ[key] = value
self.runner.write_final_report()
if __name__ == "__main__":
unittest.main()
test/registered/amd/perf/mi35x/test_qwen35_fp8_perf_mi35x.py
+139
View File
@@ -0,0 +1,139 @@
"""MI35x Nightly performance benchmark for Qwen3.5-397B-A17B FP8.
Tests Qwen3.5-397B-A17B-FP8 (MoE, Hybrid Attention with Gated Delta Networks)
on 8 GPUs with triton attention backend.
Registry: nightly-perf-8-gpu-mi35x-qwen35-fp8 suite
"""
import os
os.environ.setdefault("HF_HOME", "/data2/models/huggingface")
os.environ.setdefault("HF_HUB_CACHE", "/data2/models/huggingface/hub")
import unittest
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, _parse_int_list_env
register_amd_ci(
est_time=5400, suite="nightly-perf-8-gpu-mi35x-qwen35-fp8", nightly=True
)
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", "MI35x")
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"
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
QWEN35_FP8_MODEL_PATH = os.environ.get(
"QWEN35_FP8_MODEL_PATH", "Qwen/Qwen3.5-397B-A17B-FP8"
)
PROFILE_DIR = "performance_profiles_qwen35_fp8_mi35x"
class TestQwen35Fp8PerfMI35x(unittest.TestCase):
"""Test suite for Qwen3.5-397B-A17B FP8 performance benchmarks on MI35x."""
@classmethod
def setUpClass(cls):
cls.base_url = DEFAULT_URL_FOR_TEST
cls.batch_sizes = [1, 8, 16, 64]
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
cls.model_config = {
"name": "qwen35-fp8-mi35x",
"model_path": QWEN35_FP8_MODEL_PATH,
"other_args": [
"--trust-remote-code",
"--tp",
"8",
"--attention-backend",
"aiter",
"--mem-fraction-static",
"0.8",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
"--watchdog-timeout",
"1200",
],
"env_vars": {
"SGLANG_USE_AITER": "1",
},
}
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_qwen35_fp8_perf(self):
"""Run Qwen3.5-397B-A17B FP8 performance benchmark on MI35x."""
old_env = {}
for key, value in self.model_config.get("env_vars", {}).items():
old_env[key] = os.environ.get(key)
os.environ[key] = value
try:
result_tuple = self.runner.run_benchmark_for_model(
model_path=self.model_config["model_path"],
batch_sizes=self.batch_sizes,
input_lens=self.input_lens,
output_lens=self.output_lens,
other_args=self.model_config["other_args"],
variant=self.model_config["name"],
extra_bench_args=["--trust-remote-code"],
enable_profile=False,
timeout=5400,
)
results = result_tuple[0]
success = result_tuple[1]
if results:
self.runner.full_report += (
generate_simple_markdown_report(results) + "\n"
)
self.assertTrue(
success,
f"Benchmark failed for {QWEN35_FP8_MODEL_PATH} on MI35x",
)
finally:
for key, value in old_env.items():
if value is None:
os.environ.pop(key, None)
else:
os.environ[key] = value
self.runner.write_final_report()
if __name__ == "__main__":
unittest.main()