EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[EPLB][Nightly] Refactor UT (#6543)

Browse Source 
### What this PR does / why we need it?
The basic configs are extracted and reused for eplb UT. This is done so
that if the basic configs are changed later, eplb UT does not need to be
modified repeatedly.
### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

- vLLM version: v0.15.0
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.15.0

Signed-off-by: bigsir007 <xujiacheng12@huawei.com>
Co-authored-by: bigsir007 <xujiacheng12@huawei.com>
This commit is contained in:
JIACHENG XU
2026-02-14 10:56:29 +08:00
committed by GitHub
parent 1e77077788
commit 64aea60f2e
4 changed files with 61 additions and 121 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
tests/e2e/nightly/single_node/models/test_deepseek_r1_0528_w8a8.py
View File

@@ -63,10 +63,7 @@ aisbench_cases = [{
}]
@pytest.mark.asyncio
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("mode", MODES)
async def test_models(model: str, mode: str) -> None:
def config():
port = get_open_port()
env_dict = {
"OMP_NUM_THREADS": "10",
@@ -85,6 +82,13 @@ async def test_models(model: str, mode: str) -> None:
"--speculative-config",
json.dumps(speculative_config)
]
return port, env_dict, additional_config, server_args
@pytest.mark.asyncio
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("mode", MODES)
async def test_models(model: str, mode: str) -> None:
port, env_dict, additional_config, server_args = config()
if mode == "single":
server_args.append("--enforce-eager")
server_args.extend(["--additional-config", json.dumps(additional_config)])

69
tests/e2e/nightly/single_node/models/test_deepseek_r1_0528_w8a8_eplb.py
View File

@@ -23,18 +23,7 @@ from vllm.utils.network_utils import get_open_port
from tests.e2e.conftest import RemoteOpenAIServer
from tools.aisbench import run_aisbench_cases
MODELS = [
"vllm-ascend/DeepSeek-R1-0528-W8A8",
]
prompts = [
"San Francisco is a",
]
api_keyword_args = {
"max_tokens": 10,
}
from .test_deepseek_r1_0528_w8a8 import *
aisbench_cases = [{
"case_type": "accuracy",
@@ -50,46 +39,23 @@ aisbench_cases = [{
@pytest.mark.asyncio
@pytest.mark.parametrize("model", MODELS)
async def test_models(model: str) -> None:
port = get_open_port()
env_dict = {
"OMP_NUM_THREADS": "100",
"OMP_PROC_BIND": "false",
"HCCL_BUFFSIZE": "200",
"VLLM_RPC_TIMEOUT": "3600000",
"VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": "3600000",
"DISABLE_L2_CACHE": "1",
"DYNAMIC_EPLB": "true",
}
speculative_config = {"num_speculative_tokens": 1, "method": "mtp"}
compilation_config = {
"cudagraph_capture_sizes": [24],
"cudagraph_mode": "FULL_DECODE_ONLY"
}
additional_config: dict[str, Any] = {
"enable_shared_expert_dp": False,
"multistream_overlap_shared_expert": False,
"eplb_config": {
"dynamic_eplb": True,
"expert_heat_collection_interval": 512,
"algorithm_execution_interval": 100,
"num_redundant_experts": 0
async def test_models_eplb(model: str) -> None:
port, env_dict, additional_config, server_args = config()
additional_config.update(
{
"eplb_config": {
"dynamic_eplb": "true",
"expert_heat_collection_interval": 1000,
"algorithm_execution_interval": 50,
"eplb_policy_type": 3,
}
}
}
server_args = [
"--quantization", "ascend", "--seed", "1024",
"--no-enable-prefix-caching", "--data-parallel-size", "4",
"--tensor-parallel-size", "4", "--enable-expert-parallel", "--port",
str(port), "--max-model-len", "40000", "--max-num-batched-tokens",
"4096", "--max-num-seqs", "12", "--trust-remote-code",
"--gpu-memory-utilization", "0.92"
]
server_args.extend(
["--speculative-config",
json.dumps(speculative_config)])
server_args.extend(
["--compilation-config",
json.dumps(compilation_config)])
)
env_dict.update(
{
"DYNAMIC_EPLB": "true",
}
)
server_args.extend(["--additional-config", json.dumps(additional_config)])
request_keyword_args: dict[str, Any] = {
**api_keyword_args,
@@ -113,3 +79,4 @@ async def test_models(model: str) -> None:
port,
aisbench_cases,
server_args=server_args)

62
tests/e2e/nightly/single_node/models/test_qwen3_235b_a22b_w8a8_eplb.py
View File

@@ -23,64 +23,30 @@ from vllm.utils.network_utils import get_open_port
from tests.e2e.conftest import RemoteOpenAIServer
from tools.aisbench import run_aisbench_cases
MODELS = [
"vllm-ascend/Qwen3-235B-A22B-W8A8",
]
prompts = [
"San Francisco is a",
]
api_keyword_args = {
"max_tokens": 10,
}
aisbench_cases = [{
"case_type": "accuracy",
"dataset_path": "vllm-ascend/gsm8k-lite",
"request_conf": "vllm_api_general_chat",
"dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_chat_prompt",
"max_out_len": 32768,
"batch_size": 32,
"top_k": 20,
"baseline": 95,
"threshold": 5
}]
from .test_qwen3_235b_w8a8 import *
@pytest.mark.asyncio
@pytest.mark.parametrize("model", MODELS)
async def test_models(model: str) -> None:
port = get_open_port()
env_dict = {
"OMP_NUM_THREADS": "10",
"OMP_PROC_BIND": "false",
"HCCL_BUFFSIZE": "1024",
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"VLLM_ASCEND_ENABLE_FLASHCOMM1": "1"
}
async def test_models_eplb(model: str) -> None:
port, aisbench_cases, env_dict, compilation_config, server_args = config()
env_dict.update(
{
"DYNAMIC_EPLB": "true",
}
)
additional_config: dict[str, Any] = {}
compilation_config = {"cudagraph_mode": "FULL_DECODE_ONLY"}
server_args = [
"--quantization", "ascend", "--async-scheduling",
"--data-parallel-size", "4", "--tensor-parallel-size", "4",
"--enable-expert-parallel", "--port",
str(port), "--max-model-len", "40960", "--max-num-batched-tokens",
"8192", "--max-num-seqs", "12", "--trust-remote-code",
"--gpu-memory-utilization", "0.9"
]
env_dict["DYNAMIC_EPLB"] = "true"
additional_config["eplb_config"] = {
"dynamic_eplb": True,
"expert_heat_collection_interval": 512,
"algorithm_execution_interval": 100,
"num_redundant_experts": 0
"dynamic_eplb": "true",
"expert_heat_collection_interval": 600,
"algorithm_execution_interval": 50,
"num_redundant_experts": 16,
"eplb_policy_type": 2,
}
server_args.extend(["--additional-config", json.dumps(additional_config)])
server_args.extend(
["--compilation-config",
json.dumps(compilation_config)])
server_args.extend(["--additional-config", json.dumps(additional_config)])
request_keyword_args: dict[str, Any] = {
**api_keyword_args,
}

39
tests/e2e/nightly/single_node/models/test_qwen3_235b_w8a8.py
View File

@@ -38,24 +38,19 @@ api_keyword_args = {
"max_tokens": 10,
}
aisbench_cases = [{
"case_type": "accuracy",
"dataset_path": "vllm-ascend/gsm8k-lite",
"request_conf": "vllm_api_general_chat",
"dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_chat_prompt",
"max_out_len": 32768,
"batch_size": 32,
"top_k": 20,
"baseline": 95,
"threshold": 5
}]
@pytest.mark.asyncio
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("mode", MODES)
async def test_models(model: str, mode: str) -> None:
def config():
port = get_open_port()
aisbench_cases = [{
"case_type": "accuracy",
"dataset_path": "vllm-ascend/gsm8k-lite",
"request_conf": "vllm_api_general_chat",
"dataset_conf": "gsm8k/gsm8k_gen_0_shot_cot_chat_prompt",
"max_out_len": 32768,
"batch_size": 32,
"top_k": 20,
"baseline": 95,
"threshold": 5
}]
env_dict = {
"OMP_NUM_THREADS": "10",
"OMP_PROC_BIND": "false",
@@ -72,11 +67,19 @@ async def test_models(model: str, mode: str) -> None:
"8192", "--max-num-seqs", "12", "--trust-remote-code",
"--gpu-memory-utilization", "0.9"
]
return port, aisbench_cases, env_dict, compilation_config, server_args
@pytest.mark.asyncio
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("mode", MODES)
async def test_models(model: str, mode: str) -> None:
port, aisbench_cases, env_dict, compilation_config, server_args = config()
if mode == "piecewise":
compilation_config["cudagraph_mode"] = "PIECEWISE"
server_args.extend(
["--compilation-config",
json.dumps(compilation_config)])
json.dumps(compilation_config)])
request_keyword_args: dict[str, Any] = {
**api_keyword_args,
}