我是A800-3000推理服务器+4*Atlas 300I pro 卡 ，拉的下面的MindIE_1.0.RC2 的镜像。

swr.cn-central-221.ovaijisuan.com/wh-aicc-fae/mindie:310p3-ascend_24.1.rc2-cann_8.0.rc2-py_3.10-ubuntu_22.04-aarch64-mindie_1.0.RC2.02

cd /usr/local/Ascend/atb-models/tests/modeltest

bash run.sh pa_fp16 performance [[512,512]] 16 qwen /home/qwen/Qwen1___5-7B-Chat 2 执行性能分析测试正确 。

起服务 ./bin/mindieservice_daemon 报如下错误

 

./bin/mindieservice_daemon 

Daemon wait pid with 12173, status 02025-02-14 07:31:28,725 [INFO] [pid: 12193] env.py-55: {'use_ascend': True, 'max_memory_gb': None, 'reserved_memory_gb': 3, 'skip_warmup': False, 'visible_devices': None, 'use_host_chooser': True, 'bind_cpu': True} 

2025-02-14 07:31:29,874 [INFO] [pid: 12193] logging.py-53: Skip binding cpu. 

The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. 

2025-02-14 07:31:30,392 [INFO] [pid: 12193] logging.py-53: model_runner.quantize: None 

, model_runner.kv_quant: None 

, model_runner.dytpe: torch.float16 

2025-02-14 07:31:30,392 [INFO] [pid: 12193] logging.py-53: Rank table file location: 

[W compiler_depend.ts:623] Warning: expandable_segments currently defaults to false. You can enable this feature by `export PYTORCH_NPU_ALLOC_CONF = expandable_segments:True`. (function operator()) 

2025-02-14 07:31:36,630 [INFO] [pid: 12193] dist.py-94: initialize_distributed has been Set 

2025-02-14 07:31:36,643 [INFO] [pid: 12193] logging.py-53: init tokenizer done: Qwen2TokenizerFast(name_or_path='/home/hulingjing/models/qwen/Qwen1___5-7B-Chat', vocab_size=151643, model_max_length=32768, is_fast=True, padding_side='left', truncation_side='right', special_tokens={'eos_token': '<|im_end|>', 'pad_token': '<|endoftext|>', 'additional_special_tokens': ['<|im_start|>', '<|im_end|>']}, clean_up_tokenization_spaces=False),  added_tokens_decoder={ 

        151643: AddedToken("<|endoftext|>", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True), 

        151644: AddedToken("<|im_start|>", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True), 

        151645: AddedToken("<|im_end|>", rstrip=False, lstrip=False, single_word=False, normalized=False, special=True), 

} 

2025-02-14 07:31:36,829 [INFO] [pid: 12193] logging.py-53: NPUSocInfo(soc_name='', soc_version=202, need_nz=True) 

2025-02-14 07:31:36,874 [INFO] [pid: 12193] flash_causal_qwen2.py-52: >>>> qwen_DecoderModel is called. 

2025-02-14 07:31:46,541 [INFO] [pid: 12193] logging.py-53: model: 

 FlashQwen2ForCausalLM( 

  (rotary_embedding): PositionRotaryEmbedding() 

  (attn_mask): AttentionMask() 

  (transformer): FlashQwenModel( 

    (wte): TensorParallelEmbedding() 

    (h): ModuleList( 

      (0-31): 32 x FlashQwenLayer( 

        (attn): FlashQwenAttention( 

          (rotary_emb): PositionRotaryEmbedding() 

          (c_attn): TensorParallelColumnLinear( 

            (linear): FastLinear() 

          ) 

          (c_proj): TensorParallelRowLinear( 

            (linear): FastLinear() 

          ) 

        ) 

        (mlp): QwenMLP( 

          (act): SiLU() 

          (w2_w1): TensorParallelColumnLinear( 

            (linear): FastLinear() 

          ) 

          (c_proj): TensorParallelRowLinear( 

            (linear): FastLinear() 

          ) 

        ) 

        (ln_1): QwenRMSNorm() 

        (ln_2): QwenRMSNorm() 

      ) 

    ) 

    (ln_f): QwenRMSNorm() 

  ) 

  (lm_head): TensorParallelHead( 

    (linear): FastLinear() 

  ) 

) 

Fatal Python error: PyThreadState_Get: the function must be called with the GIL held, but the GIL is released (the current Python thread state is NULL) 

Python runtime state: finalizing (tstate=0x0000fffbe001e350) 

 

Daemon wait pid with 12193, status 134Daemon is killing... 

Killed 

请大佬帮忙分析下是什么原因。另外有些疑惑为啥我拉的镜像没有/usr/local/Ascend/llm_model 这个路径。 还有我的目的是想部署多模态理解模型，MindIE_1.0.RC2是不是不支持多模态理解模型，是不是要升级到MindIE_1.0.RC3？