1c5900327b
### What this PR does / why we need it?
This PR deletes ~2K lines of code about deepseek modeling. It falls back
CustomDeepseekV2 modules to original vllm implementations and adapts
some modifications in vllm about deepseek and moe.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
E2E vllm serving with torchair graph mode and eager mode.
- vLLM version: v0.10.2
- vLLM main:
759ef49b15
---------
Signed-off-by: linfeng-yuan <1102311262@qq.com>
Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
Co-authored-by: yiz-liu <136800916+yiz-liu@users.noreply.github.com>
Co-authored-by: Yizhou Liu <liu_yizhou@outlook.com>
43 lines
1.3 KiB
Python
43 lines
1.3 KiB
Python
import pytest
|
|
|
|
from tests.e2e.conftest import VllmRunner
|
|
from tests.e2e.model_utils import check_outputs_equal
|
|
|
|
|
|
@pytest.mark.parametrize("model_name", ["deepseek-ai/DeepSeek-V2-Lite-Chat"])
|
|
def test_e2e_ep_correctness(model_name):
|
|
example_prompts = [
|
|
"Hello, my name is",
|
|
"The president of the United States is",
|
|
"The capital of France is",
|
|
"The future of AI is",
|
|
]
|
|
max_tokens = 5
|
|
|
|
# FIXME: Really strange that chunked prefill might lead to different results, investigate further
|
|
with VllmRunner(
|
|
model_name,
|
|
tensor_parallel_size=2,
|
|
additional_config={"ascend_scheduler_config": {
|
|
"enabled": True
|
|
}},
|
|
enforce_eager=True) as vllm_model:
|
|
tp_output = vllm_model.generate_greedy(example_prompts, max_tokens)
|
|
|
|
with VllmRunner(
|
|
model_name,
|
|
tensor_parallel_size=2,
|
|
enable_expert_parallel=True,
|
|
additional_config={"ascend_scheduler_config": {
|
|
"enabled": True
|
|
}},
|
|
enforce_eager=True) as vllm_model:
|
|
ep_output = vllm_model.generate_greedy(example_prompts, max_tokens)
|
|
|
|
check_outputs_equal(
|
|
outputs_0_lst=ep_output,
|
|
outputs_1_lst=tp_output,
|
|
name_0="ep_output",
|
|
name_1="tp_output",
|
|
)
|