合入背景

请描述为什么要做这个PR内的改动。
如涉及，请关联前序PR或同特性/需求下的其他PR。
如果是修复之前PR引入的问题，请关联引入问题的PR。
注意：Fixes #ISSUE ID会自动关闭issue，如问题部分解决请不要使用Fixes，可以用Fix part of #ISSUE ID替代.

Fix part of #55
当前已支持的模型：Qwen稠密模型

修改内容

请描述修改内容的具体实现，涉及哪些组件之间进行交互，可以用1、2、3、...进行罗列。
如果是需求或者重构类的PR，需要补充详细设计文档（说明上下游组件关系、时序图、类图、DFX能力等内容）。

资料变更

请确认是否涉及资料变更。如涉及，需要在PR中体现，并简要说明修改内容。如不涉及，需填写“不涉及”。

不涉及。

接口变更

请确认是否涉及跨代码仓或者客户面可见的接口变更。如涉及，需要详细说明接口以及对应的变更内容，同时需要在资料中体现。如不涉及，需填写“不涉及”。

不涉及。

测试结果

请说明测试场景，测试方法以及测试结果。
测试用例设计时需考虑硬件、部署方式、功能、性能、精度、显存等维度。

• tie_word_embedding Qwen3 4B

torchrun --nproc_per_node 4 --master_port 20038 -m examples.run_pa --model_path  {权重路径}

[2025-12-30 01:47:51,514] [54933] [281473017825952] [llmmodels] [INFO] [run_pa.py-434] : ---------------end inference---------------
[2025-12-30 01:47:51,514] [54933] [281473017825952] [llmmodels] [INFO] [run_pa.py-611] : Answer[0]:  What are the key components of deep learning? What are the applications of deep learning? What are the
[2025-12-30 01:47:51,514] [54933] [281473017825952] [llmmodels] [INFO] [run_pa.py-612] : Generate[0] token num: (0, 20)

• Qwen3 32B W8A8 mix

torchrun --nproc_per_node 4 --master_port 20038 -m examples.run_pa --model_path {权重路径}

[2025-12-30 01:50:09,692] [58035] [281473580292768] [llmmodels] [INFO] [run_pa.py-434] : ---------------end inference---------------
[2025-12-30 01:50:09,692] [58035] [281473580292768] [llmmodels] [INFO] [run_pa.py-611] : Answer[0]:  Deep learning is a subset of machine learning that uses artificial neural networks with multiple layers to model complex patterns
[2025-12-30 01:50:09,692] [58035] [281473580292768] [llmmodels] [INFO] [run_pa.py-612] : Generate[0] token num: (0, 20)

• Qwen2.5 7B VL - 测试通过

BATCH_SIZE=1
INPUT_LEN=2800
OUTPUT_LEN=256

bash examples/models/qwen2_vl/run_pa.sh \
    --max_batch_size $BATCH_SIZE \
    --max_input_length $INPUT_LEN \
    --max_output_length $OUTPUT_LEN \
    --input_text "请介绍图片中的内容" \
    --model_path "/mnt/nfs/weight/Qwen2.5-VL-7B-Instruct/" \
    --input_image "/mnt/nfs/c00814587/textvqa_val/train_images/1080p.jpg"


[2025-12-30 18:12:41,749] [1881572] [281473079382112] [llmmodels] [INFO] [run_pa.py-393] : Question[0]: [{'image': '/mnt/nfs/c00814587/textvqa_val/train_images/1080p.jpg'}, {'text': '请介绍图片中的内容'}]
[2025-12-30 18:12:41,750] [1881572] [281473079382112] [llmmodels] [INFO] [run_pa.py-394] : Answer[0]: 这张图片展示了一位女性和一只狗在海滩上互动的场景。女性坐在沙滩上，穿着格子衬衫和牛仔裤，面带微笑，似乎在和狗进行友好互动。狗戴着彩色的项圈，坐在沙滩上，伸出前爪与女性的手相触。背景是广阔的海洋和天空，阳光洒在沙滩上，营造出一种温暖而宁静的氛围。整个画面给人一种轻松愉快的感觉。
[2025-12-30 18:12:41,750] [1881572] [281473079382112] [llmmodels] [INFO] [run_pa.py-395] : Generate[0] token num: 92
[2025-12-30 18:12:41,750] [1881572] [281473079382112] [llmmodels] [INFO] [run_pa.py-397] : Latency(s): 1.1621677875518799
[2025-12-30 18:12:41,750] [1881572] [281473079382112] [llmmodels] [INFO] [run_pa.py-398] : Throughput(tokens/s): 79.16240751587091

• 边云协同场景用例通过

• 非Qwen模型，Deepseek 大EP A2 场景验证OK

• test coverage

examples/atb_models/atb_llm/atb_llm/models/base/router.py 80.15% 60.22%
examples/atb_models/atb_llm/flash_causal_qwen2.py 83.72% 70.73%
examples/atb_models/atb_llm/atb_llm/models/qwen2/modeling_qwen2_refactor.py 93.48% 50%
examples/atb_models/atb_llm/atb_llm/models/qwen2/router_qwen2.py 90.36% 83.33%

mindie_llm/runtime/__init__.py 100% 100%
mindie_llm/runtime/config/__init__.py 100% 100%
mindie_llm/runtime/config/configuration_utils.py 81.18% 75%
mindie_llm/runtime/config/huggingface_config.py 98.28% 75%
mindie_llm/runtime/config/mindie_llm_config.py 82.5% 66.67%
mindie_llm/runtime/layers/__init__.py 100% 100%
mindie_llm/runtime/layers/custom_layer.py 85.71% 100%
mindie_llm/runtime/layers/normalization.py 95.74% 75%
mindie_llm/runtime/layers/parameter.py 87.34% 100%
mindie_llm/runtime/layers/embedding/__init__.py 100% 100%
mindie_llm/runtime/layers/embedding/embedding.py 97.87% 88.89%
mindie_llm/runtime/layers/fused_moe/fused_moe.py 100% 100%
mindie_llm/runtime/layers/fused_moe/fused_moe_method_base.py 80% 100%
mindie_llm/runtime/layers/linear/linear.py 94.66% 78.12%
mindie_llm/runtime/layers/linear/linear_method_base.py 84.62% 100%
mindie_llm/runtime/layers/quantization/quantization_config_base.py 80.95% 100%
mindie_llm/runtime/layers/quantization/quantization_method_base.py 75% 100%
mindie_llm/runtime/layers/quantization/unquantized.py 100% 100%
mindie_llm/runtime/layers/quantization/ms_model_slim/anti_outlier.py 100% 100%
mindie_llm/runtime/layers/quantization/ms_model_slim/quant_type.py 100% 100%
mindie_llm/runtime/layers/quantization/ms_model_slim/quantization_config.py 100% 100%
mindie_llm/runtime/layers/quantization/ms_model_slim/w8a8.py 97.25% 72.22%
mindie_llm/runtime/model_runner/forward_context.py 71.11% 0%
mindie_llm/runtime/utils/command_executor_utils.py 100% 100%
mindie_llm/runtime/utils/npu_utils.py 85.71% 50%
mindie_llm/runtime/utils/parameter_validators.py 92.39% 85.29%
mindie_llm/runtime/utils/torch_utils.py 100% 100%
mindie_llm/runtime/utils/distributed/__init__.py 100% 100%
mindie_llm/runtime/utils/distributed/parallel_info_manager.py 94.16% 88.46%
mindie_llm/runtime/utils/distributed/utils.py 100% 100%
mindie_llm/runtime/utils/helpers/env.py 85.25% 78.12%
mindie_llm/runtime/utils/helpers/safety/file.py 97.33% 87.5%
mindie_llm/runtime/utils/helpers/safety/path.py 100% 91.67%
mindie_llm/runtime/utils/loader/default_model_loader.py 96.88% 87.5%
mindie_llm/runtime/utils/loader/weight_utils.py 30.88% 0%

CheckList

PR提交人对以下CheckList自检项进行全量自检，自检通过或不涉及，均修改 [ ] 为 [x]。

• 代码注释完备
• 正确记录错误日志
• 进行了返回值校验 (禁止使用void屏蔽安全函数、自研函数返回值；考虑接口的异常场景；调用底层组件接口时，需要进行返回值校验)
• 进行了空指针校验
• 若存在资源申请，使用后资源被正确的释放了
• 若涉及多线程场景，考虑了并发场景，不存在死锁问题
• 按照代码仓中提供的格式模板，使用clang-format工具格式化代码
• 符合Ascend社区的编码规范。C++ 语言编程指导 | C++ 语言安全编程指导