Skip to content

Commit a054f0f

Browse files
authored
[CI] change to new ds model (#1513)
Previous, the DeepSeek V3 Pruning weight is not correct, the moe layer is not tested. We update a new Pruning model to enable moe layer compute. This PR fix the CI to address the new weight. --------- Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
1 parent 8013634 commit a054f0f

File tree

3 files changed

+7
-7
lines changed

3 files changed

+7
-7
lines changed

tests/multicard/test_fused_moe_allgather_ep.py renamed to tests/e2e/multicard/test_fused_moe_allgather_ep.py

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -41,7 +41,7 @@ def test_generate_with_allgather():
4141
sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
4242

4343
with VllmRunner(snapshot_download("vllm-ascend/DeepSeek-V3-Pruning"),
44-
tensor_parallel_size=16,
44+
tensor_parallel_size=4,
4545
enforce_eager=True,
4646
max_model_len=1024,
4747
dtype="auto",
@@ -67,7 +67,7 @@ def test_generate_with_alltoall():
6767
sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
6868

6969
with VllmRunner(snapshot_download("vllm-ascend/DeepSeek-V3-Pruning"),
70-
tensor_parallel_size=16,
70+
tensor_parallel_size=4,
7171
enforce_eager=True,
7272
max_model_len=1024,
7373
dtype="auto",

tests/e2e/multicard/test_offline_inference_distributed.py

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -56,7 +56,7 @@ def test_models_distributed_DeepSeek_multistream_moe():
5656
with VllmRunner(
5757
"vllm-ascend/DeepSeek-V3-Pruning",
5858
dtype=dtype,
59-
tensor_parallel_size=2,
59+
tensor_parallel_size=4,
6060
distributed_executor_backend="mp",
6161
additional_config={
6262
"torchair_graph_config": {

tests/e2e/multicard/test_torchair_graph_mode.py

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -66,10 +66,10 @@ def _deepseek_torchair_test_fixture(
6666
# inaccurate. This will only change if accuracy improves with the
6767
# official weights of DeepSeek-V3.
6868
golden_results = [
69-
'Hello, my name is feasibility伸 spazio debtor添',
70-
'The president of the United States is begg"""\n杭州风和 bestimm',
71-
'The capital of France is frequentlyশามalinkAllowed',
72-
'The future of AI is deleting俯احت怎么样了حراف',
69+
'Hello, my name is下载早点向前很有่อง',
70+
'The president of the United States isSender)## physiological Albany',
71+
'The capital of France is Rocky转角 hospitalizedinterval sparked',
72+
'The future of AI is её asegο BIOS一扫',
7373
]
7474

7575
assert len(golden_results) == len(vllm_output)

0 commit comments

Comments
 (0)