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

[EPLB][CI] Add dynamic EPLB CI for qwen3-moe (#5179)

Browse Source 
### What this PR does / why we need it?
Add dynamic EPLB CI for qwen3-moe-30B-W8A8

- vLLM version: v0.12.0
- vLLM main:
ad32e3e19c

Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
This commit is contained in:
LI SHENGYONG
2025-12-23 11:31:00 +08:00
committed by GitHub
parent 449f8f65a7
commit 2e010e12dd
1 changed files with 56 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

73
tests/e2e/multicard/test_qwen3_moe.py
View File

@@ -21,12 +21,16 @@
Run `pytest tests/e2e/multicard/test_qwen3_moe.py`.
"""
import json
import os
from unittest.mock import patch
import openai
import pytest
from modelscope import snapshot_download # type: ignore
from vllm.utils import get_open_port
from tests.e2e.conftest import VllmRunner
from tests.e2e.conftest import RemoteOpenAIServer, VllmRunner
@patch.dict(os.environ, {"HCCL_BUFFSIZE": "1024"})
@@ -58,22 +62,6 @@ def test_qwen3_moe_w8a8_distributed_tp2():
vllm_model.generate_greedy(example_prompts, max_tokens)
@patch.dict(os.environ, {"HCCL_BUFFSIZE": "1024"})
def test_qwen3_moe_w8a8_distributed_tp2_ep():
example_prompts = [
"Hello, my name is",
]
max_tokens = 5
with VllmRunner(
snapshot_download("vllm-ascend/Qwen3-30B-A3B-W8A8"),
max_model_len=8192,
tensor_parallel_size=2,
enable_expert_parallel=True,
quantization="ascend",
) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)
def test_qwen3_moe_distributed_aiv_tp2():
os.environ['HCCL_OP_EXPANSION_MODE'] = 'AIV'
example_prompts = [
@@ -87,3 +75,54 @@ def test_qwen3_moe_distributed_aiv_tp2():
tensor_parallel_size=2,
) as vllm_model:
vllm_model.generate_greedy(example_prompts, max_tokens)
@pytest.mark.asyncio
async def test_qwen3_moe_w8a8_distributed_tp2_ep_dynamic_eplb():
model = "vllm-ascend/Qwen3-30B-A3B-W8A8"
port = get_open_port()
server_args = [
"--max_model_len", "8192", "--tensor_parallel_size", "2",
"--enable_expert_parallel", "--quantization", "ascend", "--port",
str(port), "--enforce_eager"
]
env_dict = {"HCCL_BUFFSIZE": "1024"}
with RemoteOpenAIServer(model,
server_args,
server_port=port,
auto_port=False,
env_dict=env_dict) as server:
client = server.get_async_client()
batch = await client.completions.create(model=model,
prompt="What is deeplearning?",
max_tokens=300,
temperature=0,
top_p=1.0,
n=1)
gt_choices: list[openai.types.CompletionChoice] = batch.choices
# dynamic eplb test
# Since pytest runs as a daemon, it conflicts with the dynamic eplb manager
# during initialization in offline mode, so the online mode is used instead.
env_dict.update({"DYNAMIC_EPLB": "true"})
additional_config = {
"dynamic_eplb": True,
"num_iterations_eplb_update": 100,
"num_wait_worker_iterations": 20
}
server_args.extend(["--additional-config", json.dumps(additional_config)])
with RemoteOpenAIServer(model,
server_args,
server_port=port,
auto_port=False,
env_dict=env_dict) as server:
client = server.get_async_client()
batch = await client.completions.create(model=model,
prompt="What is deeplearning?",
max_tokens=300,
temperature=0,
top_p=1.0,
n=1)
eplb_choices: list[openai.types.CompletionChoice] = batch.choices
assert gt_choices[0].text == eplb_choices[
0].text, f"{gt_choices[0].text=} \n {eplb_choices[0].text=}"