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

[Bugfix] fix dcp_only bug and add e2e accuracy test for dcp only and pcp only (#5565)

Browse Source 
### What this PR does / why we need it?
[Bugfix] fix dcp_only bug and add e2e accuracy test for dcp only and pcp
only
this pr fix the bug of accuracy test when decode_parallel_size>1 and
prefill_context_parallel_size=1.
### Does this PR introduce _any_ user-facing change?
NO

### How was this patch tested?
- vLLM version: v0.13.0
- vLLM main:
7157596103

---------

Signed-off-by: zhenwenqi2024 <zhenwenqi_2022@qq.com>
This commit is contained in:
zhenwenqi2024
2026-01-06 22:48:21 +08:00
committed by GitHub
parent 77a029979e
commit ad9b711f89
3 changed files with 128 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

114
tests/e2e/multicard/long_sequence/test_accuracy.py
View File

@@ -96,3 +96,117 @@ def test_models_long_sequence_output_between_tp_and_cp(
name_0="vllm_eager_outputs",
name_1="vllm_context_parallel_outputs",
)
model = "vllm-ascend/DeepSeek-V2-Lite-W8A8"
@pytest.mark.parametrize("max_tokens", [10])
def test_accuracy_dcp_only_graph(max_tokens: int, ) -> None:
prompts = [
"The president of the United States is", "The capital of France is"
]
cp_kwargs = {
"tensor_parallel_size": 2,
"decode_context_parallel_size": 2,
"prefill_context_parallel_size": 1,
"enable_expert_parallel": True,
"compilation_config": {
"cudagraph_mode": "FULL_DECODE_ONLY",
"cudagraph_capture_sizes": [4, 8, 24, 48, 60]
},
"quantization": "ascend",
"max_model_len": 1024,
}
tp_kwargs = {
"tensor_parallel_size": 4,
"enable_expert_parallel": True,
"enforce_eager": True,
"quantization": "ascend",
"max_model_len": 1024,
}
with VllmRunner(model, **cp_kwargs) as runner: # type: ignore
vllm_context_parallel_outputs = runner.generate_greedy(
prompts, max_tokens)
with VllmRunner(model, **tp_kwargs) as runner: # type: ignore
vllm_eager_outputs = runner.generate_greedy(prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_eager_outputs,
outputs_1_lst=vllm_context_parallel_outputs,
name_0="vllm_eager_outputs",
name_1="vllm_dcp_only_graph_outputs",
)
@pytest.mark.parametrize("max_tokens", [10])
def test_accuracy_dcp_only_eager(max_tokens: int, ) -> None:
prompts = [
"The president of the United States is", "The capital of France is"
]
cp_kwargs = {
"tensor_parallel_size": 2,
"decode_context_parallel_size": 2,
"prefill_context_parallel_size": 1,
"enable_expert_parallel": True,
"enforce_eager": True,
"quantization": "ascend",
"max_model_len": 1024,
}
tp_kwargs = {
"tensor_parallel_size": 4,
"enable_expert_parallel": True,
"enforce_eager": True,
"quantization": "ascend",
"max_model_len": 1024,
}
with VllmRunner(model, **cp_kwargs) as runner: # type: ignore
vllm_context_parallel_outputs = runner.generate_greedy(
prompts, max_tokens)
with VllmRunner(model, **tp_kwargs) as runner: # type: ignore
vllm_eager_outputs = runner.generate_greedy(prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_eager_outputs,
outputs_1_lst=vllm_context_parallel_outputs,
name_0="vllm_eager_outputs",
name_1="vllm_dcp_only_eager_outputs",
)
@pytest.mark.parametrize("max_tokens", [10])
def test_accuracy_pcp_only(max_tokens: int, ) -> None:
prompts = [
"The president of the United States is", "The capital of France is"
]
cp_kwargs = {
"tensor_parallel_size": 2,
"decode_context_parallel_size": 1,
"prefill_context_parallel_size": 2,
"enable_expert_parallel": True,
"enforce_eager": True,
"quantization": "ascend",
"max_model_len": 1024,
}
tp_kwargs = {
"tensor_parallel_size": 4,
"enable_expert_parallel": True,
"enforce_eager": True,
"quantization": "ascend",
"max_model_len": 1024,
}
with VllmRunner(model, **cp_kwargs) as runner: # type: ignore
vllm_context_parallel_outputs = runner.generate_greedy(
prompts, max_tokens)
with VllmRunner(model, **tp_kwargs) as runner: # type: ignore
vllm_eager_outputs = runner.generate_greedy(prompts, max_tokens)
check_outputs_equal(
outputs_0_lst=vllm_eager_outputs,
outputs_1_lst=vllm_context_parallel_outputs,
name_0="vllm_eager_outputs",
name_1="vllm_pcp_only_outputs",
)