openPangu-Embedded-7B-model

开源盘古 Embedded-7B

中文 | English

1. 简介

openPangu-Embedded-7B 是基于昇腾 NPU 从零训练的高效大语言模型，参数量为 7B（不含词表Embedding）。openPangu-Embedded-7B 训练了约 19T tokens，具备快慢思考融合能力。

2. 模型架构

3. 测评结果

注： 评测过程中system prompt 为空，且不添加任何额外的思维链（CoT）提示。评测采用 128k 的序列长度进行。

4. 部署和使用

4.1 环境准备

硬件规格

Atlas 800T A2 (64GB)，驱动与固件安装包获取请参照 [Atlas 800T A2]。

软件环境

• 操作系统：Linux（推荐 openEuler>=24.03）
• CANN==8.1.RC1，安装准备及流程请参照 [CANN Install]
• python==3.10
• torch==2.1.0
• torch-npu==2.1.0.post12
• transformers==4.53.2

以上软件配套经过验证，理论可以支持更高版本，如有疑问，可以提交 issue。

4.2 权重完整性校验

请参考以下方法对下载内容进行完整性校验，hash 值存储在 checklist.chk 文件中。

#!/usr/bin/env bash
ARCH=$(uname -m)
MODEL_PATH="${TARGET_FOLDER}/${MODEL_FOLDER_PATH}"
cd "$MODEL_PATH" || exit 1
if [ "$ARCH" = "arm64" ]; then
    sha256sum checklist.chk
else
    sha256sum -c checklist.chk
fi

4.3 推理样例

下述内容提供 openPangu-Embedded-7B 在 transformers 框架上进行推理的一个简单示例：

运行前请修改 generate.py，添加模型路径。

cd inference
python generate.py

openPangu-Embedded-7B 模型默认为慢思考模式，可以通过以下手段切换至快思考模式：

• 在代码实例generate.py中，no_thinking_prompt变量的定义展示了切换至快思考模式的具体实现：通过在用户输入末尾添加/no_think标记，可将当前轮次切换至快思考模式。处于该模式时，thinking_content将为空值。

4.4 使用推理框架

vllm_ascend：参考[vllm_ascend_for_openpangu_embedded_7b.zh]

5. 模型许可证

除文件中对开源许可证另有约定外，openPangu-Embedded-7B 模型根据 OPENPANGU MODEL LICENSE AGREEMENT VERSION 1.0 授权，旨在允许使用并促进人工智能技术的进一步发展。有关详细信息，请参阅模型存储库根目录中的 LICENSE 文件。

6. 免责声明

由于 openPangu-Embedded-7B（“模型”）所依赖的技术固有的技术限制，以及人工智能生成的内容是由盘古自动生成的，华为无法对以下事项做出任何保证：

• 尽管该模型的输出由 AI 算法生成，但不能排除某些信息可能存在缺陷、不合理或引起不适的可能性，生成的内容不代表华为的态度或立场；
• 无法保证该模型 100% 准确、可靠、功能齐全、及时、安全、无错误、不间断、持续稳定或无任何故障；
• 该模型的输出内容不构成任何建议或决策，也不保证生成的内容的真实性、完整性、准确性、及时性、合法性、功能性或实用性。生成的内容不能替代医疗、法律等领域的专业人士回答您的问题。生成的内容仅供参考，不代表华为的任何态度、立场或观点。您需要根据实际情况做出独立判断，华为不承担任何责任。

7. 反馈

如果有任何意见和建议，请提交issue或联系 [email protected]。

<details> <summary>相关论文清单</summary>

Depth-Breadth Synergy in RLVR: Unlocking LLM Reasoning Gains with Adaptive Exploration <br> TiFRe: Text-guided Video Frame Reduction for Efficient Video Multi-modal Large Language Models <br> Searching Meta Reasoning Skeleton to Guide LLM Reasoning <br> Post-Training Quantization of OpenPangu Models for Efficient Deployment on Atlas A2 <br> OSUM-Pangu: An Open-Source Multidimension Speech Understanding Foundation Model Built upon OpenPangu on Ascend NPUs <br> Structured Episodic Event Memory <br> Quasar: Quantized Self-Speculative Acceleration for Rapid Inference via Memory-Efficient Verification <br> CSSBench: Evaluating the Safety of Lightweight LLMs against Chinese-Specific Adversarial Patterns <br> Beyond Static Question Banks: Dynamic Knowledge Expansion via LLM-Automated Graph Construction and Adaptive Generation <br> Your Models Have Thought Enough: Training Large Reasoning Models to Stop Overthinking <br> EvoOpt-LLM: Evolving industrial optimization models with large language models <br> Accelerating OpenPangu Inference on NPU via Speculative Decoding <br> V-CAGE: Context-Aware Generation and Verification for Scalable Long-Horizon Embodied Tasks <br> Scheduling LLM Inference with Uncertainty-Aware Output Length Predictions <br> Differentially Private and Communication Efficient Large Language Model Split Inference via Stochastic Quantization and Soft Prompt <br> A-IO: Adaptive Inference Orchestration for Memory-Bound NPUs <br> Multi-source Heterogeneous Public Opinion Analysis via Collaborative Reasoning and Adaptive Fusion: A Systematically Integrated Approach <br> Characterize LSM-tree Compaction Performance via On-Device LLM Inference <br> Traceable Cross-Source RAG for Chinese Tibetan Medicine Question Answering <br> A Comprehensive Evaluation of LLM Reasoning: From Single-Model to Multi-Agent Paradigms <br> Structured Self-Consistency: A Multi-Task Evaluation of LLMs on VirtualHome <br> ArkEval: Benchmarking and Evaluating Automated CodeRepair for ArkTS <br> FMBench: Adaptive Large Language Model Output Formatting <br> DScheLLM: Enabling Dynamic Scheduling through a Fine-Tuned Dual-System Large language Model <br> Holmes: An Evidence-Grounded LLM Agent for Auditable DDoS Investigation in Cloud Networks <br> Cochain: Balancing Insufficient and Excessive Collaboration in LLM Agent Workflows <br> How Few-shot Demonstrations Affect Prompt-based Defenses Against LLM Jailbreak Attacks <br> Focus-then-Context: Subject-Centric Progressive Visual Token Reduction for Vision-Language Models <br> PGN: Design and Implementation of a Vision-Language Navigation System Based on Pangu Multimodal Foundation Model <br> Behavioral Controllability of Agentic Models for Information Extraction: From Fixed Workflows to Reflective Agents <br> Fine-tuning Large Language Model for Automated Algorithm Design <br> Streaming-dLLM: Accelerating Diffusion LLMs via Suffix Pruning and Dynamic Decoding <br> From Rigid to Dynamic: Entropy-Guided Adaptive Inference for Long-Context LLMs <br>

</details>