[NPU] Enable piecewise CUDA graph support on NPU (#28417)

This commit is contained in:
hanwlax
2026-09-23 17:30:30 +08:00
committed by GitHub
parent 58f622ebc4
commit 2d25767759
12 changed files with 414 additions and 35 deletions
@@ -0,0 +1,142 @@
import json
import tempfile
import unittest
from urllib.parse import urlparse
from sglang.test.ascend.e2e.test_npu_multi_node_utils import NIC_NAME
from sglang.test.ascend.e2e.test_npu_performance_utils import (
TestNpuPerformanceTestCaseBase,
run_bench_serving,
)
from sglang.test.ci.ci_register import register_npu_ci
register_npu_ci(est_time=400, suite="stage-b-test-2-npu-a3", nightly=False)
MODEL = "Qwen/Qwen3.6-35B-A3B"
INPUT_THROUGHPUT_THRESHOLD = 8500
MAX_CONCURRENCY = 32
NUM_PROMPTS = 128
INPUT_LEN = 4096
OUTPUT_LEN = 1
WARMUP_REQUESTS = 20
MAX_BENCHMARK_ATTEMPTS = 2
ENVS = {
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"SGLANG_SET_CPU_AFFINITY": "1",
"HCCL_SOCKET_IFNAME": NIC_NAME,
"GLOO_SOCKET_IFNAME": NIC_NAME,
"STREAMS_PER_DEVICE": "32",
"SGLANG_ENABLE_SPEC_V2": "1",
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
"DEEP_NORMAL_MODE_USE_INT8_QUANT": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "96",
"HCCL_BUFFSIZE": "1200",
"HCCL_OP_EXPANSION_MODE": "AIV",
"SGLANG_NPU_FORWARD_NATIVE_GEMMA_RMS_NORM": "1",
"SGLANG_NPU_ENABLE_PIECEWISE_CUDA_GRAPH": "1",
}
OTHER_ARGS = [
"--trust-remote-code",
"--mem-fraction-static",
0.7,
"--max-running-requests",
MAX_CONCURRENCY,
"--attention-backend",
"ascend",
"--disable-radix-cache",
"--cuda-graph-max-bs",
32,
"--tp-size",
2,
"--chunked-prefill-size",
2048,
"--cuda-graph-backend-prefill",
"tc_piecewise",
"--cuda-graph-bs-prefill",
128,
256,
512,
1024,
2048,
]
def load_benchmark_result(output_file):
with open(output_file, "r", encoding="utf-8") as f:
lines = [line for line in f if line.strip()]
return json.loads(lines[-1])
def benchmark_passed(metrics):
return (
metrics["completed"] == NUM_PROMPTS
and metrics["max_concurrency"] == MAX_CONCURRENCY
and metrics["input_throughput"] > INPUT_THROUGHPUT_THRESHOLD
)
class TestPiecewiseGraphPrefillTp2(TestNpuPerformanceTestCaseBase):
model = MODEL
other_args = OTHER_ARGS
envs = ENVS
max_concurrency = MAX_CONCURRENCY
num_prompts = NUM_PROMPTS
input_len = INPUT_LEN
output_len = OUTPUT_LEN
warmup_requests = WARMUP_REQUESTS
seed = 1
def run_bench_serving_once(self, parsed_url):
with tempfile.NamedTemporaryFile(suffix=".jsonl") as output_file:
run_bench_serving(
host=parsed_url.hostname,
port=parsed_url.port,
model_path=self.model,
dataset_name=self.dataset_name,
request_rate=self.request_rate,
max_concurrency=self.max_concurrency,
num_prompts=self.num_prompts,
input_len=self.input_len,
output_len=self.output_len,
random_range_ratio=self.random_range_ratio,
warmup_requests=self.warmup_requests,
seed=self.seed,
output_file=output_file.name,
)
return load_benchmark_result(output_file.name)
def test_npu_piecewise_graph_prefill_tp2(self):
print(f"##=== Testing PCG serving: {MODEL} ===##")
parsed_url = urlparse(self.base_url)
metrics_list = []
for attempt in range(MAX_BENCHMARK_ATTEMPTS):
metrics = self.run_bench_serving_once(parsed_url)
metrics_list.append(metrics)
if benchmark_passed(metrics):
break
print(
"PCG benchmark attempt "
f"{attempt + 1} failed: completed={metrics['completed']}, "
f"max_concurrency={metrics['max_concurrency']}, "
f"input_throughput={metrics['input_throughput']}"
)
passed_metrics = [
metrics for metrics in metrics_list if benchmark_passed(metrics)
]
metrics = (
passed_metrics[-1]
if passed_metrics
else max(metrics_list, key=lambda x: x["input_throughput"])
)
self.assertEqual(metrics["completed"], self.num_prompts)
self.assertEqual(metrics["max_concurrency"], self.max_concurrency)
self.assertGreater(metrics["input_throughput"], INPUT_THROUGHPUT_THRESHOLD)
if __name__ == "__main__":
unittest.main()