@@ -207,22 +207,16 @@ def stubbed_get_state(ep_size, with_prefill, is_deepseek_v3_r1):
207207 # use greedy sampler to make sure the generated results are fix
208208 vllm_output = vllm_model .generate_greedy (example_prompts , 5 )
209209
210- # NOTE: vllm-ascend/pangu-pro-moe-pruing is only part of PanguProMoE
211- # with 2 hidden layers, thus the golden results seems inaccurate.
212- # This will only change if accuracy changes with the official weights
213- # of PanguProMoE.
214210 golden_results = [
215- 'Hello, my name is Remempondeprecatedmiot忱 ' ,
216- 'The president of the United States is Remem下的一个 rever ceremoni Segnali ' ,
217- 'The capital of France is Rememvoud administrativ Remem投 ' ,
218- 'The future of AI isotope Segnali Zoeken精细化 supus' ,
211+ 'Hello, my name is qwen3 ' ,
212+ 'The president of the United States is Trump ' ,
213+ 'The capital of France is Pairs ' ,
214+ 'The future of AI isotope Segnali Zoeken supus' ,
219215 ]
220216
221- assert len (golden_results ) == len (vllm_output )
222217 for i in range (len (vllm_output )):
218+ assert golden_results [i ] == vllm_output [i ][1 ]
223219 print (f"Generated text: { vllm_output [i ][1 ]!r} " )
224220
225- def test_e2e_qwen2_with_torchair ():
226- _qwen_moe_torchair_test_fixture ("Qwen/Qwen2.5-0.5B-Instruct" , 2 , False )
227221def test_e2e_qwen3_moe_with_torchair ():
228222 _qwen_moe_torchair_test_fixture ("Qwen/Qwen3-30B-A3B" , 2 , True )
0 commit comments