[Bug]: Qwen3 @jit(capture_mode="trace") fails on Ascend

### Checklist - [x] 1. 我已经搜索过相关问题，但没有得到预期的帮助. (https://gitee.com/mindspore/mindspore/issues) - [ ] 2. 最新版本中该错误尚未修复. - [ ] 3. 请注意，如果您提交的Bug描述缺少相应的环境信息和最小可复现的demo，我们将很难复现和解决该问题，从而降低收到反馈的可能性，甚至该问题将被关闭. ### 🐞 问题详细描述 测试脚本： #!/usr/bin/env python3 # -*- coding: utf-8 -*- import os import mindspore as ms from mindspore import jit, JitConfig from transformers import AutoTokenizer from mindone.transformers.models.qwen3.modeling_qwen3 import Qwen3ForCausalLM def main(): ms.set_context( mode=ms.PYNATIVE_MODE, device_target="Ascend", pynative_synchronize=True, save_graphs=1, save_graphs_path="./qwen3_profile/graphs", ) model_name = "./Qwen3-0.6B" model = Qwen3ForCausalLM.from_pretrained( model_name, mindspore_dtype=ms.bfloat16, attn_implementation="flash_attention_2", # eager / flash_attention_2 都会复现 ) model.set_train(False) model.config.use_cache = False model.set_jit_config(JitConfig(jit_level="O0", infer_boost="on")) tokenizer = AutoTokenizer.from_pretrained(model_name) prompt = "Hi" enc = tokenizer([prompt], return_tensors="np") input_ids = ms.Tensor(enc.input_ids, ms.int32) attention_mask = ms.Tensor((input_ids.asnumpy() != tokenizer.pad_token_id).astype("int32")) class TraceForward(ms.nn.Cell): def __init__(self, m): super().__init__() self.m = m @jit(capture_mode="trace") def construct(self, input_ids, attention_mask): return self.m( input_ids, attention_mask=attention_mask, use_cache=False, ) net = TraceForward(model) out = net(input_ids, attention_mask) print("OK, logits shape:", out[0].shape if isinstance(out, (tuple, list)) else out.shape) if __name__ == "__main__": main() 在 Ascend 环境下，使用 MindSpore PyNative + @jit(capture_mode="trace") 对 mindone 的 Qwen3ForCausalLM 做一次前向推理时，触发如下报错并最终出现 segmentation fault： Traceback …