5d12446573
### What this PR does / why we need it?
2nd PR for https://github.com/vllm-project/vllm-ascend/issues/5712,
extend SP to VL MoE models.
### Does this PR introduce _any_ user-facing change?
remove `sp_threshold` in additional config and reuse `sp_min_token_num`
from vLLM.
### How was this patch tested?
- Model: Qwen3-VL-30B-A3B,
- TP4 DP2
- 100 reqs
- max concurrency 1
| Seq length | Mean TTFT (ms) main | Mean TTFT (ms) this PR |
|------------|---------------------|------------------------|
| 4k | 429.40 | 323.3 |
| 16k | 1297.01 | 911.74 |
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
---------
Signed-off-by: realliujiaxu <realliujiaxu@163.com>
62 lines
1.9 KiB
Python
62 lines
1.9 KiB
Python
import pytest
|
|
from vllm import SamplingParams
|
|
|
|
from tests.e2e.conftest import VllmRunner
|
|
from tests.e2e.model_utils import check_outputs_equal
|
|
|
|
MODELS = [
|
|
"Qwen/Qwen3-VL-2B-Instruct",
|
|
]
|
|
|
|
|
|
@pytest.mark.parametrize("model", MODELS)
|
|
def test_qwen3_vl_sp_tp2(model: str) -> None:
|
|
prompts = [
|
|
"Hello, my name is",
|
|
"The capital of the United States is",
|
|
"The capital of France is",
|
|
"The future of AI is",
|
|
]
|
|
sampling_params = SamplingParams(max_tokens=10, temperature=0.0)
|
|
|
|
with VllmRunner(
|
|
model,
|
|
max_model_len=1024,
|
|
tensor_parallel_size=2,
|
|
compilation_config={
|
|
"cudagraph_capture_sizes": [2, 4],
|
|
"cudagraph_mode": "FULL_DECODE_ONLY",
|
|
"pass_config": {"enable_sp": False},
|
|
},
|
|
additional_config={"ascend_compilation_config": {"enable_npugraph_ex": False}},
|
|
) as runner:
|
|
no_sp_outputs = runner.model.generate(prompts, sampling_params)
|
|
|
|
with VllmRunner(
|
|
model,
|
|
max_model_len=1024,
|
|
tensor_parallel_size=2,
|
|
compilation_config={
|
|
"cudagraph_capture_sizes": [2, 4],
|
|
"cudagraph_mode": "FULL_DECODE_ONLY",
|
|
"pass_config": {"enable_sp": True, "sp_min_token_num": 10},
|
|
},
|
|
additional_config={"ascend_compilation_config": {"enable_npugraph_ex": False}},
|
|
) as runner:
|
|
sp_outputs = runner.model.generate(prompts, sampling_params)
|
|
|
|
no_sp_outputs_list = []
|
|
for output in no_sp_outputs:
|
|
no_sp_outputs_list.append((output.outputs[0].index, output.outputs[0].text))
|
|
|
|
sp_outputs_list = []
|
|
for output in sp_outputs:
|
|
sp_outputs_list.append((output.outputs[0].index, output.outputs[0].text))
|
|
|
|
check_outputs_equal(
|
|
outputs_0_lst=no_sp_outputs_list,
|
|
outputs_1_lst=sp_outputs_list,
|
|
name_0="no_sp_outputs",
|
|
name_1="sp_outputs",
|
|
)
|