# Vendor Deployment Guide Source: https://docs.rminte.com/essentials/deploy Comprehensive guide for vendors to deploy RM-01 Portable AI Supercomputer solutions for enterprise customers, achieving a true “plug-and-play” AI experience. ## Overview This guide covers the complete deployment process from preparation to after-sales support: Model preparation, environment assessment, and device verification Installation, initialization, and network configuration Ongoing support, maintenance, and customer service ## Pre-deployment Preparation ### Model and Application Preparation Prepare suitable models and applications based on enterprise requirements: Includes common open-source models such as: * DeepSeek * Zhipu * Qwen * Other mainstream models Standard models are pre-configured and ready for immediate deployment across various enterprise use cases. Vertical domain models tailored for specific industries: * Healthcare and medical AI * Financial services models * Manufacturing optimization * Legal and compliance models Industry-specific models often provide better performance for specialized enterprise workflows. Custom trained models specific to enterprise needs: * Convert models to RM-01 compatible formats * Store in specified directory on CFexpress card * Ensure proper model validation and testing Custom models require thorough validation to ensure compatibility with RM-01's hardware specifications. ### Enterprise Environment Assessment Conduct thorough environment assessment before deployment: Assess the feasibility of connecting RM-01 to the enterprise internal network: * Evaluate network topology and security requirements * Identify firewall and proxy configurations * Plan IP addressing and network segmentation * Test network bandwidth and latency requirements Document network requirements and obtain necessary approvals from IT security team. Ensure stable power supply at the deployment location: * Verify power outlet availability and specifications * Check for uninterruptible power supply (UPS) requirements * Assess power consumption impact on facility * Plan for power redundancy if needed RM-01 requires 100W maximum power consumption with 60W TDP for optimal performance. Assess physical security of device placement: * Evaluate physical access controls * Review environmental conditions (temperature, humidity) * Plan for device mounting and cable management * Assess theft prevention measures Ensure deployment location meets enterprise security standards and environmental requirements. Understand existing IT infrastructure: * Review current AI/ML infrastructure * Assess integration with existing systems * Plan for user access management * Evaluate monitoring and logging requirements ### Device and Accessories Verification Before delivery, ensure all components are complete and functional: **Core Components:** * RM-01 Host device * CFexpress Type B Storage Card (pre-installed with models and applications) * MicroSD Card for system storage * Quick Start Guide with setup instructions * Warranty Card and documentation Verify all components are present and undamaged before shipment. **Essential Accessories:** * USB-C Power Adapter (PD3.1, up to 140W) * USB-C to Ethernet Adapter for network connectivity * Additional cables as needed Keep spare accessories available for immediate replacement if needed. ## Hardware Deployment Process ### Device Installation Prepare the installation location: 1. Choose optimal placement (desktop or dedicated cabinet) 2. Ensure adequate ventilation around the device 3. Verify power and network connectivity 4. Prepare cable management solutions Ensure the device is placed in a well-ventilated environment, avoiding obstruction of heat dissipation vents. Install the RM-01 device: 1. Insert CFexpress storage card with pre-installed models and applications 2. Insert MicroSD card for system storage 3. Connect USB-C power adapter 4. Connect USB-C to Ethernet adapter if wired network is required Handle storage cards carefully to avoid damage to connectors and data corruption. ### System Initialization Power on and initialize the system: 1. Press the power button to start the device 2. Monitor initialization process via connected laptop (USB-C) 3. Wait for automatic model and application loading (\~5 minutes) 4. Verify system reaches ready state During initialization, you can connect a laptop via USB-C port to monitor device status and troubleshoot if needed. Confirm successful initialization: * Check system status indicators * Verify model loading completion * Test basic device responsiveness * Validate storage card recognition System should display ready status after approximately 5 minutes of initialization. ### Network Configuration Configure network connectivity based on enterprise requirements: Configure Ethernet connection: ```bash theme={null} # Static IP Configuration sudo ip addr add 192.168.1.100/24 dev eth0 sudo ip route add default via 192.168.1.1 # DNS Configuration echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf ``` Verify network connectivity with ping test to external servers. Enable automatic IP assignment: ```bash theme={null} # Enable DHCP client sudo dhclient eth0 # Verify IP assignment ip addr show eth0 ``` DHCP is recommended for simplified network management in most enterprise environments. ## Verification and Testing ### Comprehensive Testing Protocol Verify core functionality: **Model Performance:** * Test model loading and execution * Verify inference speed and accuracy * Check resource utilization **Application Testing:** * Test basic application functionality * Verify user interface responsiveness * Check integration with enterprise systems Document test results and compare against expected performance benchmarks. Conduct performance evaluation: **Metrics to Test:** * Model inference latency * Concurrent processing capability * Memory and storage utilization * Network throughput ```bash theme={null} # Example performance test command ./benchmark_tool --model deepseek --iterations 100 --concurrent 4 ``` Use standardized benchmarking tools to ensure consistent performance evaluation across deployments. Verify security measures: * Test user authentication and authorization * Verify data encryption at rest and in transit * Check network access controls * Validate audit logging functionality Ensure all security configurations meet enterprise compliance requirements before deployment approval. ## Training and Delivery ### Administrator Training Program Train administrators on basic device management: **Core Topics:** * Device startup and shutdown procedures * Hardware maintenance and troubleshooting * Storage card management * Network configuration updates Demonstrate system management capabilities: * Web-based management console * System monitoring and alerts * User account management * Configuration backup and restore Record training sessions for future reference and onboarding new administrators. Cover advanced management topics: * Model updates and version control * Application deployment procedures * Performance optimization techniques * Troubleshooting common issues ### End User Training Train users on fundamental operations: * Application access and login procedures * Basic AI model interaction * File upload and processing * Result interpretation and export Customize training content based on specific departmental use cases for maximum effectiveness. Cover sophisticated functionality: * Multi-model workflow creation * Custom prompt engineering * Batch processing operations * Integration with external tools Provide hands-on practice sessions with real enterprise data to enhance learning effectiveness. ### Documentation Delivery Provide comprehensive documentation package: Hardware specifications and basic operation guide System management and troubleshooting procedures Application usage steps and best practices Common issues and resolution procedures All documentation is provided in both electronic and printed formats, with the latest versions available through the management interface. ## After-sales Support ### Support Service Levels **Included Services:** * 12 months of remote support * Business hours response (9 AM - 6 PM) * Email and phone support * System health monitoring **Response Times:** * 4-hour response during business hours * 24-hour resolution for standard issues Standard support covers most enterprise deployment needs with comprehensive coverage. **Enhanced Services:** * 24/7 technical support availability * 2-hour response time guarantee * Priority issue resolution * Dedicated support engineer **Additional Features:** * Proactive system monitoring * Quarterly performance reviews * Priority access to updates Premium support is recommended for mission-critical deployments and high-availability requirements. **Comprehensive Services:** * 30-minute emergency response * On-site support when needed * Custom model development assistance * Integration consulting services **Value-Added Services:** * Performance optimization consulting * Custom application development * Staff augmentation services Enterprise support provides the highest level of service for complex deployments and specialized requirements. ### Support Contact Information **Phone:** 158-8200-8185 **Hours:** Monday - Friday, 9:00 AM - 9:00 PM Direct access to technical support engineers **Email:** [support@rminte.com](mailto:support@rminte.com) **Response:** 24-hour guarantee Detailed technical inquiries and documentation **Website:** rminte.com **Chat:** Real-time during business hours Instant support and resource access ## Frequently Asked Questions No. RM-01 is designed to operate completely offline, ensuring data security and independence from internet connectivity. All processing occurs locally on the device. RM-01 has a maximum power consumption of 100W with a TDP of 60W, saving 98.6% energy compared to traditional AI servers. Annual electricity cost for 24/7 operation is approximately \$480. RM-01 supports models ranging from 0.5B to 235B parameters (GPTQ Int4 quantization), covering all mainstream open-source models including DeepSeek, Qwen, and Llama variants. RM-01 uses hardware-level asymmetric encryption with all data processed locally. No data is ever uploaded to the cloud, ensuring complete enterprise data sovereignty and security. Compared to traditional AI deployment, RM-01 saves 80% in initial investment, reduces operational costs by 98%, and achieves 99% total cost of ownership savings over 3 years. Yes, we offer a 7-day free trial service for enterprise customers with professional technical staff providing on-site support and evaluation assistance. ## Technical Support Resources Comprehensive technical documentation and API references Technical forums and best practice sharing Discover and deploy professional AI applications Video tutorials and certification programs ***
© 2025 Panidea (Chengdu) Artificial Intelligence Technology Co., Ltd. All rights reserved.
# RM-01 Developer Guide Source: https://docs.rminte.com/essentials/develop Comprehensive guide to RM-01 system architecture, module configuration, network setup and model deployment ## Overview This guide provides developers with complete technical documentation for the RM-01 portable supercomputer, covering system architecture, network configuration, model deployment and other core content: Configure network connections for data interaction between device and host Understand how the inference module, application module and management chip work together Master AI model deployment, configuration and optimization methods **Read Before Use** RM-01 consists of an **Inference Module**, an **Application Module**, and an **Encryption and Management Chip** (hereinafter referred to as the Management Module), interconnected via an **onboard Ethernet switch chip**, forming an internal LAN subnet. When a user connects to a host (e.g., PC, smartphone, iPad) via the **USB Type-C** interface, RM-01 will virtualize an Ethernet interface for the host through USB Ethernet functionality. The host will then obtain an IP address and automatically join the subnet for data interaction. After the device is powered on and connected to the host via the **USB Type-C** interface, the system will automatically configure the local network subnet. The **user host** will be assigned a static IP address `10.10.99.100`, and the **Out-of-Band Management Chip** will have a static IP address of `10.10.99.97`. The **Inference Module** (IP: `10.10.99.98`) and the **Application Module** (IP: `10.10.99.99`)—both deploy independent SSH services, allowing users to access them directly via standard SSH clients (e.g., OpenSSH, PuTTY). The Management Module, however, requires access via a serial port tool. ## Network Configuration ### About the Out-of-Band Management Chip **Network Configuration** * IP Address: `10.10.99.97` * Access Method: Web browser In addition to handling certain encryption tasks, the Out-of-Band Management Chip also hosts the **RM-01's real-time system performance monitoring dashboard** — **RobOS**. You can access `10.10.99.97` through a web browser to monitor the **connection status** and **operational status** of each module in real-time. ### How to Provide Internet Access to RM-01 from the Host (Using macOS as an Example) After connecting the **user host** via USB Type-C, RM-01 will appear in the network interface list as: * **`AX88179A`** (Developer Version) * **`RMinte RM-01`** (Commercial Release Version) Open **System Settings** Go to **Network** → **Sharing** Enable **Internet Sharing** Click the **"i" icon** next to the sharing settings to enter the configuration interface: * Set **"Share your connection from"** to: **Wi-Fi** * In **"To computers using"**, select: **AX88179A** or **RMinte RM-01** (depending on the device model) Click **Done** Return to the **Network** settings page and manually configure the RM-01 network interface: * **IP Address**: `10.10.99.100` * **Subnet Mask**: `255.255.255.0` * **Router**: `10.10.99.100` (i.e., the host's own IP) This configuration sets the host as a gateway, providing NAT network access for RM-01. The default gateway and DNS for RM-01 are automatically assigned by the host via DHCP. Manually setting the IP ensures that it remains within the `10.10.99.0/24` subnet, consistent with the device's internal service communication. ## System Architecture ### About the CFexpress Type-B Storage Card The **CFexpress Type-B** storage card is one of the core components of the RM-01 device, responsible for system boot, deployment of the inference framework, and key functions such as ISV/SV software distribution and authorization authentication. The storage card is divided into three independent partitions: **System Partition** The operating system and core runtime environment of the Inference Module are installed in this partition. Users or developers are strictly prohibited from accessing, modifying, or deleting the contents of this partition. Any unauthorized changes may cause the Inference Module to fail to boot or render inference functions inoperable, and any resulting hardware or software damage is not covered by any warranty services. **Application Partition** This partition is used to temporarily store `Docker` image files submitted by users or developers. After the image is written to `rm01app`, the RM-01 system will automatically migrate it to the device's built-in **NVMe SSD** storage and complete containerized deployment. Do not directly run or modify application files in this partition. **Model Partition** Dedicated to storing large-scale AI models (e.g., LLMs, multimodal models, etc.) loaded by users or developers. For details on model formats, size limitations, loading procedures, and compatibility requirements, refer to the "Model Deployment" section below. ### About the Application Module **Network Configuration** * IP Address: `10.10.99.99` * Port Range: `59000-59299` #### Application Module Hardware Specifications ```text theme={null} Processor: Intel Core i3-N305 (8 cores, 8 threads, base frequency 1.8 GHz, max turbo frequency 3.8 GHz) Memory: 16 GB / 24 GB LPDDR5-4800MT/s (onboard, non-expandable) Storage: 512 GB / 1 TB / 2 TB (optional) NVMe SSD ``` #### Application Module SSH Access Credentials ```bash SSH Login Command theme={null} ssh rm01@10.10.99.99 # Default Username: rm01 # Default Password: rm01 (factory preset, for initial login only) ``` ```bash Change Password theme={null} # Execute immediately after first login passwd ``` **Security Notice** To ensure system security, immediately use the `passwd` command to change the default password after the first SSH login. The default password is only for initial configuration and must not be used in production or deployment environments. #### Pre-installed Software The Application Module has **Open WebUI** pre-installed on port `80` to facilitate simple model debugging and conversational work. You can access Open WebUI by navigating to `10.10.99.99` in your web browser. ### About the Inference Module **Network Configuration** * IP Address: `10.10.99.98` * Service Port Range: `58000–58999` The **Inference Module** is the core computing unit of RM-01, supporting various high-performance AI inference configurations. Users can select the appropriate model based on model scale and performance requirements. #### Hardware Configuration Options | Memory | Memory Bandwidth | Compute Power | Tensor Core Count | | :----: | :--------------: | :----------------: | :---------------: | | 32 GB | 204.8 GB/s | 200 TOPS (INT8) | 56 | | 64 GB | 204.8 GB/s | 275 TOPS (INT8) | 64 | | 64 GB | 273 GB/s | 1,200 TFLOPS (FP4) | 64 | | 128 GB | 273 GB/s | 2,070 TFLOPS (FP4) | 96 | #### Pre-installed Inference Frameworks RM-01 comes pre-installed with the following two inference frameworks on the **CFexpress Type-B** storage card, both running on the Inference Module: * **Status**: Automatically starts * **Default Port**: 58000 * **Function**: Provides OpenAI-compatible API interfaces * **Supported Requests**: Standard POST `/v1/chat/completions` etc. * **Status**: Requires manual startup * **Function**: Text embedding services #### API Access Method After successfully loading a model, the **vLLM** inference service can be accessed via the following address: ```bash theme={null} http://10.10.99.98:58000/v1/chat/completions ``` Supports direct calls using standard OpenAI clients (e.g., openai-python, curl, Postman). **Security Notice** To ensure system security and stability, the Inference Module does not provide SSH access permissions. Users and developers cannot directly log in or interactively operate the underlying operating system of this module. Any attempts to bypass security policies or directly access the Inference Module may result in system anomalies, data corruption, or service interruptions, which are not covered by warranty services. ## Model Deployment ### About Models RM-01 supports inference for various AI models, including but not limited to: Large Language Models Multimodal Models Vision-Language Models Text Embedding Models Reranker Models All model files must be stored on the device's built-in **CFexpress Type-B** storage card, and users need to use a compatible **CFexpress Type-B** card reader to upload, manage, and update models on the host side. When the **CFexpress Type-B** storage card is connected to RM-01, the system mounts it as a read-only data volume named `models` at the path `/home/rm01/models`. Its standard file structure is as follows: ```bash theme={null} models/ ├── auto/ # Directory for automatic model loading (production-grade deployment) │ ├── embedding/ # Embedding models (not automatically loaded, see below) │ ├── llm/ # Large language models (weight files stored directly, see below) │ └── reranker/ # Reranker models (not automatically loaded, see below) └── dev/ # Developer-defined model directory (higher priority) ├── embedding/ ├── llm/ └── reranker/ ``` * The `auto/` directory is used for **lightweight, standardized model deployment**, automatically recognized by the system * The `dev/` directory is used for **fine-grained control of model behavior by developers**, with higher priority than `auto/`. The system will ignore models in `auto/` if `dev/` is used ### Deployment Mode Selection Simplified mode suitable for quick verification and standardized deployment. Place the **complete weight files** of the model (e.g., `.safetensors`, `.bin`, `.pt`, `.awq`, etc.) **directly in the `auto/llm/` directory**, **without nesting in subfolders**. ```text ❌ Incorrect Example theme={null} auto/llm/Qwen3-30B-A3B-Instruct-2507-AWQ/model.safetensors ``` ```text ✅ Correct Example theme={null} auto/llm/model-001-of-006.safetensors auto/llm/config.json auto/llm/tokenizer.json auto/llm/vocab.json ``` * Upon device startup, the system scans the `auto/llm/` directory and automatically loads models in compatible formats * **Automatic loading is not supported for embedding or reranker models**, only LLMs * After loading, the model enables **basic inference capabilities by default** and **does not enable the following advanced features**: * Speculative Decoding * Prefix Caching * Chunked Prefill * The maximum context length (`max_model_len`) is restricted to the system's safe threshold (typically ≤ 8192 tokens) * **Limited performance optimization**: To ensure system stability and multitasking concurrency, models in automatic mode use a conservative memory allocation strategy (`gpu_memory_utilization` ≤ 0.8) **Important Note** Automatic mode is suitable for **quickly verifying model compatibility** or **standardized deployment scenarios**, but it is **not suitable for high-performance inference in production environments**. For full performance, use **manual mode (dev)**. Developer mode, providing complete configuration control and performance optimization capabilities. **Manual mode has higher priority than automatic mode**. When any `.yaml` configuration file exists in the `embedding`, `llm`, or `reranker` directories under `dev/`, the system will **completely ignore all models in `auto/`**. #### Configuration File Structure The `embedding`, `llm`, and `reranker` directories under `dev/` must contain the following three YAML configuration files to start the corresponding model services: | File Name | Purpose | | -------------------- | ---------------------------------- | | `embedding_run.yaml` | Start text embedding model service | | `llm_run.yaml` | Start large language model service | | `reranker_run.yaml` | Start reranker model service | All files must be located in the corresponding folders under `dev/`, and **file names must match exactly**, case-sensitive. #### Example Configuration File ```yaml llm_run.yaml theme={null} model: /home/rm01/models/dev/llm/Qwen3-30B-A3B-Instruct-2507-AWQ port: 58000 gpu_memory_utilization: 0.85 max_model_len: 24576 served_model_name: RM-01 LLM enable_prefix_caching: true enable_chunked_prefill: true max_num_batched_tokens: 512 block_size: 16 tensor_parallel_size: 1 dtype: auto ``` **Configuration Notes** * The `model` path must be an **absolute path** pointing to the model file directory (not a compressed file) * The `port` must be within the `58000–58999` range and must not conflict with other services * The `gpu_memory_utilization` is recommended to be set between 0.7–0.9 to maximize throughput * All parameters follow the [vLLM official documentation](https://docs.vllm.ai/en/latest/index.html). Ensure compatibility with the vLLM version currently used on the **CFexpress Type-B** storage card #### Deployment Steps Copy the model files (complete directory) to `dev/llm/` (or `embedding/`, `reranker/`) Create and place the corresponding `.yaml` configuration file Insert the **CFexpress Type-B** storage card and restart RM-01 The system will automatically load the configuration in `dev/` and start the corresponding service Once the service is started, it can be accessed via interfaces like `http://10.10.99.98:58000/v1/chat/completions` **Recommendation**: For initial deployment, use the `--load-format auto` and `--dtype auto` options provided by vLLM to automatically adapt to the model format. ### Security and Maintenance Notes * **Prohibited SSH Login to Inference Module**: All model management must be done via the **CFexpress Type-B** storage card * **Model Files Must Be Raw Weights**: Do not use compressed files (.zip/.tar.gz), encrypted packages, or non-standard formats * **File Permissions**: All model files must be readable (`chmod 644`), and directories must be executable (`chmod 755`) * **Version Control**: It is recommended to use Git or file naming conventions (e.g., `Qwen3-30B-A3B-Instruct-v1.2-20250930`) to manage model versions * **Backup Recommendation**: Back up the `dev/` and `auto/` directories before updating models to avoid configuration loss ### Mode Selection Recommendations | Scenario | Recommended Mode | Description | | ------------------------------------------- | -------------------------------------------- | ---------------------------------------- | | Quickly verify model compatibility | Automatic Mode (auto) | No configuration required, plug and play | | High-performance inference in production | Manual Mode (dev) + fine-tuned configuration | Full performance optimization | | Multi-model parallel deployment | Manual Mode (dev) + multiple `.yaml` files | Flexible service orchestration | | Development debugging, prototype validation | Manual Mode (dev) | Complete control | ## Technical Support Complete API reference and technical documentation Sample code and open-source tools Developer community and technical discussions Professional technical support services ***
© 2025 Panidea (Chengdu) Artificial Intelligence Technology Co., Ltd. All rights reserved.
# Enterprise Quickstart Source: https://docs.rminte.com/essentials/quickstart Get your team up and running with the RM-01 AI supercomputer quickly with this comprehensive enterprise guide ## Device Overview The RM-01 AI supercomputer is a portable AI solution that combines high performance, security, and ease of use, specifically designed for enterprise-level AI applications. It operates offline with local computation, significantly reducing data security risks and operational costs. ### System Architecture RM-01 consists of three core modules interconnected through a built-in Ethernet switch chip, forming an independent internal network: **IP Address**: `10.10.99.98` Handles AI model inference computation, providing high-performance computing power **IP Address**: `10.10.99.99` Runs pre-installed applications and business systems, such as Open WebUI **IP Address**: `10.10.99.97` Provides RobOS real-time monitoring and device management functions When you connect RM-01 to your host device (computer, phone, or tablet) via USB Type-C, the host will automatically be assigned the static IP address `10.10.99.100` and join RM-01's internal network. ### Key Features * Supports 300+ mainstream open-source models * Up to 2070 TFLOPS computing power * 128GB LPDDR5 high-speed memory * Hardware-level asymmetric encryption protection * Data processed locally, never transmitted externally * Anti-tamper physical security design * Deployment setup completed within 5 minutes * Plug-and-play, no professional IT support needed * Maximum power consumption only 140W (TDP 60W) ### Device Interfaces * **Power Button**: Short press to turn on/off, long press for 5 seconds to force shutdown * **Status Indicator**: White (Running), Blue (Starting), Red (Error) * **CFexpress Slot**: For model and application storage, supports hot-swapping * **MicroSD Slot**: For logo customization and key storage * **Expansion Port**: For connecting external expansion devices (supports multimodal input) * **USB-C Power Port**: Connect to power adapter (supports PD3.1) * **USB-C Network Port**: Connect to personal devices or enterprise switches * **Ventilation**: Ensure good ventilation, do not block ## Quick Start ### First-time Setup Confirm the package contains the following items: * RM-01 Main Unit * CFexpress Card with pre-installed models and applications * MicroSD Card * Warranty Card * Quick Start Guide * USB-C Power Adapter * USB-C to Ethernet Adapter * Connect the USB-C power adapter to the RM-01 * Connect the power adapter to a power outlet * Press the power button * Wait for the device to boot (about 1-2 minutes) * A steady white indicator light means the device is running normally Use a USB Type-C cable to connect RM-01 to your host device (computer, phone, or tablet). After the device is powered on, it will automatically configure the internal network, and your host will be assigned the static IP address: `10.10.99.100` After successful connection, your host's network interface will display as **AX88179A** or **RMinte RM-01** (depending on version) If you need RM-01 to access the internet, you can share the network through your host device. For macOS example: 1. Open **System Settings** > **Network** > **Sharing** 2. Enable **Internet Sharing**, select **Wi-Fi** as the sharing source 3. In "Share to devices", check the RM-01 network interface 4. Manually set the RM-01 network interface parameters: * **IP Address**: `10.10.99.100` * **Subnet Mask**: `255.255.255.0` * **Router**: `10.10.99.100` 5. Click **Done** This configuration makes the host act as a gateway, and RM-01 will automatically obtain the default gateway and DNS. If internet access is not needed, you can skip this step. ### Accessing the Management Interface RM-01 provides multiple management entry points, each corresponding to different functional modules: Access the real-time monitoring dashboard of the management module: ``` http://10.10.99.97 ``` View the connection status and operation of each RM-01 module here, keeping track of device health in real-time. Access the pre-installed Open WebUI on the application module: ``` http://10.10.99.99 ``` **Default login credentials**: * Username: `rm01` * Password: `rm01` After first login, please use SSH to log in (address: `10.10.99.99`) and immediately change the password to ensure security. Directly access the API service of the inference module: ``` http://10.10.99.98:58000/v1/chat/completions ``` Supports standard OpenAI client calls, using tools like `openai-python`, Postman, or curl. If you have purchased or deployed other enterprise applications, they typically run on the application module (`10.10.99.99`). Please contact your service provider for specific port numbers and access methods. ## Model Management ### Loading Open-Source Models (Automatic Mode) RM-01 supports quick loading of open-source large language models (LLM), allowing you to easily verify model compatibility. Using a **CFexpress Type-B** card reader, place the complete model weight files into the `auto/llm/` directory of the storage card. Supported formats include: `.safetensors`, `.bin`, `.pt`, `.awq`, etc. **Correct example**: ``` auto/llm/model-001-of-006.safetensors auto/llm/config.json auto/llm/tokenizer.json auto/llm/vocab.json ``` Do not use subfolders! Files must be placed directly in the `auto/llm/` directory, not in subdirectories. **Incorrect example**: ``` auto/llm/Qwen3-30B-A3B-Instruct-2507-AWQ/model.safetensors ``` Insert the **CFexpress Type-B** storage card into RM-01 and restart the device. The system will automatically scan the `auto/llm/` directory and load compatible models. After the model is loaded, you can access the inference service through the inference service endpoint (`http://10.10.99.98:58000/v1/chat/completions`). **Automatic Mode Limitations**: * Only supports large language models (LLM), does not support embedding models or reranker models * Suitable for quick verification of model compatibility, performance is limited (maximum context length ≤ 8192 tokens, memory usage rate ≤ 0.8) * For higher performance or multi-model parallel deployment, please contact your service provider for **manual mode (dev)** technical support ## Using AI Applications ### View and Launch Applications 1. Log in to the management interface 2. Click the "App Library" tab 3. Browse the list of installed AI applications Click on the application name to view details, including functionality introduction, usage scenarios, and required resources 1. Click the "Launch" button 2. The application will open in a new tab 3. Follow the application prompts to operate ## Data Management ### Data Import 1. Find the "Upload" or "Import" button in the application 2. Select the file to upload 3. Wait for the upload to complete 1. Prepare the files to be imported 2. Use the provided batch import tool 3. Select the folder or multiple files 4. Start the import process ### Data Security Guarantees RM-01 uses multiple security measures to protect your data All data is processed within the device, never uploaded to the cloud Uses asymmetric encryption technology to protect stored data Dust and moisture-proof design, prevents physical disassembly Granular user permission management system ### Security Usage Notes **The inference module does not support SSH login**. All model management must be completed through the **CFexpress Type-B** storage card. This design ensures the stability and security of the inference service. Do not attempt to directly access the SSH service of the inference module (`10.10.99.98`). All model deployments should be performed through the storage card. Before updating models or configurations, it is recommended to back up the `auto/` and `dev/` directories on the storage card to prevent configuration loss or the need to roll back to a previous version. * Change all default passwords immediately after first login * Use strong passwords (at least 12 characters, including uppercase and lowercase letters, numbers, and special characters) * Update passwords regularly (recommended every 90 days) * Do not use the same password across multiple systems ## User Management ### Adding Users Use an administrator account to access the management interface Click on the "User Management" option in the left navigation bar Click the "Add User" button Enter username, email, initial password, and permission level Click the "Create" button to complete user addition ### Permission Settings RM-01 supports granular permission control: Full access permissions, including system settings Can use all applications, manage their own data Can use designated applications, restricted data access Can only use specific demo applications **Adjusting User Permissions** 1. Enter "User Management" 2. Select the user to modify 3. Click "Edit Permissions" 4. Set application access permissions and data access scope 5. Save changes ## Regular Maintenance ### Status Monitoring Regular system status checks ensure smooth operation After logging into the management interface, check the system status information on the dashboard: * CPU usage * Memory usage * Storage space usage * Data read/write speed * Operating temperature * Network connection status ### Update Management Update models through the CFexpress Type-B storage card: 1. Obtain new model files from your service provider 2. Use a card reader to place the model files into the `auto/llm/` directory of the storage card 3. Ensure the file structure is correct (refer to the Model Management section) 4. Reinsert the storage card and restart the device It is recommended to back up existing model files before updating Contact your service provider for application update packages and specific update procedures. Typically requires: 1. Log in to the application module (`10.10.99.99`) 2. Upload the update package 3. Complete the update process following the prompts Contact your service provider for system update support. System updates may require professional technical support. Do not attempt unauthorized system update operations on your own. ### Performance Optimization If you notice device performance degradation, try the following steps Go to "Application Management" and stop any applications you don't need Go to "System Maintenance" and select the "Clear Cache" option Select "Restart" in the management interface, or short press the power button and select "Restart" ### Troubleshooting **Possible causes and solutions**: * **Power issue**: Check power adapter connection, ensure using original power supply * **Storage card issue**: Try removing and reinserting the storage card * **System issue**: Hold the power button for 10 seconds to force shutdown, then restart **Possible causes and solutions**: * **Insufficient resources**: Close other running applications * **Corrupted application**: Reinstall the application * **Model issue**: Check if the required models are correctly loaded **Possible causes and solutions**: * **Network adapter issue**: Check the USB-C network adapter connection, confirm the host recognizes the AX88179A or RMinte RM-01 network interface * **IP address configuration error**: Confirm the host IP address is `10.10.99.100` with subnet mask `255.255.255.0` * **Firewall restrictions**: Confirm the enterprise firewall is not blocking the following addresses: * `10.10.99.97` (Management Module) * `10.10.99.98` (Inference Module) * `10.10.99.99` (Application Module) * **Network interface not enabled**: Confirm the RM-01 network interface is enabled in network settings **Possible causes and solutions**: * **Multiple applications running simultaneously**: Reduce the number of running applications * **Large model load**: Switch to a smaller model * **High temperature**: Ensure good device ventilation, avoid stacking * **Insufficient storage space**: Clear unnecessary data * **Automatic mode limitations**: If using automatic mode to load models, performance is limited to a maximum context of 8192 tokens. Contact your service provider to upgrade to manual mode for higher performance **Possible causes and solutions**: * **Incorrect file structure**: Ensure model files are placed directly in the `auto/llm/` directory, do not use subfolders * **Incompatible file format**: Check if supported formats are used (`.safetensors`, `.bin`, `.pt`, `.awq`, etc.) * **Missing required files**: Ensure complete model files are included, including `config.json`, `tokenizer.json`, and other configuration files * **Storage card not properly inserted**: Reinsert the CFexpress Type-B storage card and restart the device * **Storage card failure**: Try using another CFexpress Type-B storage card ## Getting Support ### Service Provider Contact Information 400-XXX-XXXX (Weekdays 9:00-18:00) [support@rminte.com](mailto:support@rminte.com) [https://support.rminte.com](https://support.rminte.com) ### Learning Resources Comprehensive documentation and guides Step-by-step visual learning materials Connect with other developers and experts Community discussions and Q\&A Congratulations! You have completed the basic setup and usage overview of RM-01! Now you can: * Use RobOS at `http://10.10.99.97` to monitor device status * Access Open WebUI at `http://10.10.99.99` for model debugging * Load your own open-source models to the `auto/llm/` directory via CFexpress Type-B storage card * Use the inference service endpoint at `http://10.10.99.98:58000/v1/chat/completions` for AI inference Start exploring the endless possibilities that AI technology brings to your enterprise! If you have any questions, please contact your service provider or visit our support center for assistance. For in-depth development or optimizing RM-01 performance, please refer to the RM-01 Developer Guide or contact your service provider for additional technical support. ***
© 2025 Panidea (Chengdu) Artificial Intelligence Technology Co., Ltd. All rights reserved.
# 企业部署指南 Source: https://docs.rminte.com/essentials_cn/deploy_cn 为企业客户部署搭载AI解决方案的 RM-01 便携式超级计算机,实现真正的即插即用AI体验 ## 概述 本指南涵盖从准备到售后支持的完整部署流程: 模型准备、环境评估和设备验证 安装、初始化和网络配置 持续支持、维护和客户服务 ## 部署前准备 ### 模型和应用准备 根据企业需求准备合适的模型和应用: 包含常见的开源模型,如: * DeepSeek * 智谱 * 通义千问 * 其他主流模型 标准模型已预先配置,可立即部署用于各种企业用例。 针对特定行业定制的垂直领域模型: * 医疗健康 AI * 金融服务模型 * 制造业优化 * 法律合规模型 行业专用模型通常为特定企业工作流程提供更好的性能。 根据企业需求定制训练的模型: * 将模型转换为 RM-01 兼容格式 * 存储在 CFexpress 卡的指定目录中 * 确保正确的模型验证和测试 定制模型需要彻底验证以确保与 RM-01 的硬件规格兼容。 ### 企业环境评估 部署前进行全面的环境评估: 评估将 RM-01 连接到企业内部网络的可行性: * 评估网络拓扑和安全要求 * 识别防火墙和代理配置 * 规划 IP 地址分配和网络分段 * 测试网络带宽和延迟要求 记录网络需求并获得 IT 安全团队的必要批准。 确保部署位置的稳定电源供应: * 验证电源插座的可用性和规格 * 检查不间断电源(UPS)需求 * 评估功耗对设施的影响 * 如需要,规划电源冗余 RM-01 最大功耗为 100W,TDP 为 60W,可实现最佳性能。 评估设备放置的物理安全性: * 评估物理访问控制 * 审查环境条件(温度、湿度) * 规划设备安装和线缆管理 * 评估防盗措施 确保部署位置符合企业安全标准和环境要求。 了解现有 IT 基础设施: * 审查当前的 AI/ML 基础设施 * 评估与现有系统的集成 * 规划用户访问管理 * 评估监控和日志记录要求 ### 设备和配件验证 交付前,确保所有组件齐全且功能正常: **核心组件:** * RM-01 主机设备 * CFexpress Type B 存储卡(预装模型和应用) * MicroSD 卡用于系统存储 * 快速入门指南及设置说明 * 保修卡和文档 发货前验证所有组件是否齐全且未损坏。 **必要配件:** * USB-C 电源适配器(PD3.1,最高 140W) * USB-C 转以太网适配器用于网络连接 * 根据需要的其他线缆 准备备用配件以便在需要时立即更换。 ## 硬件部署流程 ### 设备安装 准备安装位置: 1. 选择最佳放置位置(桌面或专用机柜) 2. 确保设备周围有充足的通风 3. 验证电源和网络连接 4. 准备线缆管理解决方案 确保设备放置在通风良好的环境中,避免阻塞散热孔。 安装 RM-01 设备: 1. 插入预装模型和应用的 CFexpress 存储卡 2. 插入用于系统存储的 MicroSD 卡 3. 连接 USB-C 电源适配器 4. 如需有线网络,连接 USB-C 转以太网适配器 小心处理存储卡,避免损坏连接器和数据损坏。 ### 系统初始化 开机并初始化系统: 1. 按下电源按钮启动设备 2. 通过连接的笔记本电脑(USB-C)监控初始化过程 3. 等待自动加载模型和应用(约 5 分钟) 4. 验证系统达到就绪状态 在初始化期间,您可以通过 USB-C 端口连接笔记本电脑来监控设备状态并在需要时进行故障排除。 确认初始化成功: * 检查系统状态指示器 * 验证模型加载完成 * 测试基本设备响应性 * 验证存储卡识别 系统应在初始化约 5 分钟后显示就绪状态。 ### 网络配置 根据企业要求配置网络连接: 配置以太网连接: ```bash theme={null} # 静态 IP 配置 sudo ip addr add 192.168.1.100/24 dev eth0 sudo ip route add default via 192.168.1.1 # DNS 配置 echo "nameserver 8.8.8.8" | sudo tee /etc/resolv.conf ``` 通过 ping 测试外部服务器验证网络连接。 启用自动 IP 分配: ```bash theme={null} # 启用 DHCP 客户端 sudo dhclient eth0 # 验证 IP 分配 ip addr show eth0 ``` 在大多数企业环境中,建议使用 DHCP 以简化网络管理。 ## 验证和测试 ### 全面测试协议 验证核心功能: **模型性能:** * 测试模型加载和执行 * 验证推理速度和准确性 * 检查资源利用率 **应用测试:** * 测试基本应用功能 * 验证用户界面响应性 * 检查与企业系统的集成 记录测试结果并与预期性能基准进行比较。 进行性能评估: **要测试的指标:** * 模型推理延迟 * 并发处理能力 * 内存和存储利用率 * 网络吞吐量 ```bash theme={null} # 示例性能测试命令 ./benchmark_tool --model deepseek --iterations 100 --concurrent 4 ``` 使用标准化基准测试工具确保跨部署的一致性能评估。 验证安全措施: * 测试用户身份验证和授权 * 验证静态和传输中的数据加密 * 检查网络访问控制 * 验证审计日志功能 确保所有安全配置在部署批准前符合企业合规要求。 ## 培训和交付 ### 管理员培训计划 培训管理员基本设备管理: **核心主题:** * 设备启动和关闭程序 * 硬件维护和故障排除 * 存储卡管理 * 网络配置更新 演示系统管理功能: * 基于 Web 的管理控制台 * 系统监控和警报 * 用户账户管理 * 配置备份和恢复 录制培训课程以供将来参考和新管理员入职培训。 涵盖高级管理主题: * 模型更新和版本控制 * 应用部署程序 * 性能优化技术 * 常见问题故障排除 ### 最终用户培训 培训用户基础操作: * 应用访问和登录程序 * 基本 AI 模型交互 * 文件上传和处理 * 结果解释和导出 根据特定部门用例定制培训内容以获得最大效果。 涵盖高级功能: * 多模型工作流创建 * 自定义提示工程 * 批处理操作 * 与外部工具集成 提供使用真实企业数据的实践课程以增强学习效果。 ### 文档交付 提供全面的文档包: 硬件规格和基本操作指南 系统管理和故障排除程序 应用使用步骤和最佳实践 常见问题和解决程序 所有文档均以电子和印刷格式提供,最新版本可通过管理界面获取。 ## 售后支持 ### 支持服务级别 **包含服务:** * 12 个月远程支持 * 工作时间响应(上午 9 点 - 下午 6 点) * 电子邮件和电话支持 * 系统健康监控 **响应时间:** * 工作时间内 4 小时响应 * 标准问题 24 小时解决 标准支持通过全面覆盖满足大多数企业部署需求。 **增强服务:** * 24/7 技术支持可用性 * 2 小时响应时间保证 * 优先问题解决 * 专属支持工程师 **附加功能:** * 主动系统监控 * 季度性能审查 * 优先获取更新 建议关键任务部署和高可用性要求使用高级支持。 **全面服务:** * 30 分钟紧急响应 * 需要时现场支持 * 定制模型开发协助 * 集成咨询服务 **增值服务:** * 性能优化咨询 * 定制应用开发 * 人员增强服务 企业支持为复杂部署和特殊要求提供最高级别的服务。 ### 支持联系信息 **电话:** 158-8200-8185 **时间:** 周一至周五,上午 9:00 - 晚上 9:00 直接联系技术支持工程师 **邮箱:** [support@rminte.com](mailto:support@rminte.com) **响应:** 24 小时保证 详细技术咨询和文档 **网站:** rminte.com **聊天:** 工作时间内实时支持 即时支持和资源访问 ## 常见问题解答 不需要。RM-01 设计为完全离线运行,确保数据安全并独立于互联网连接。所有处理都在设备本地进行。 RM-01 最大功耗为 100W,TDP 为 60W,与传统 AI 服务器相比节能 98.6%。全天候运行的年电费约为 480 美元。 RM-01 支持从 5 亿到 2350 亿参数的模型(GPTQ Int4 量化),涵盖所有主流开源模型,包括 DeepSeek、通义千问和 Llama 变体。 RM-01 使用硬件级非对称加密,所有数据都在本地处理。没有任何数据上传到云端,确保完整的企业数据主权和安全。 与传统 AI 部署相比,RM-01 节省 80% 的初始投资,降低 98% 的运营成本,3 年总拥有成本节省达 99%。 是的,我们为企业客户提供 7 天免费试用服务,专业技术人员提供现场支持和评估协助。 ## 技术支持资源 全面的技术文档和 API 参考 技术论坛和最佳实践分享 发现和部署专业 AI 应用 视频教程和认证计划 ***
© 2025 泛灵(成都)人工智能科技有限公司 保留所有权利
# RM-01开发者使用指南 Source: https://docs.rminte.com/essentials_cn/develop_cn 全面了解 RM-01 的系统架构、模块配置、网络设置及模型部署方法 ## 概述 本指南为开发者提供 RM-01 便携式超级计算机的完整技术文档,涵盖系统架构、网络配置、模型部署等核心内容: 配置网络连接,实现设备与主机的数据交互 了解推理模组、应用模组和管理芯片的协同工作机制 掌握 AI 模型的部署、配置和优化方法 **使用前必读** RM-01 由一个**推理模组**、一个**应用模组**和一颗**加密与管理芯片**(以下简称管理模组)构成,三者通过**板载以太网交换芯片**互联,形成内部局域网子网。当用户通过 **USB Type-C** 接口连接主机(如 PC、手机、iPad)时,RM-01 会通过 USB Ethernet 功能为主机虚拟出一个以太网接口,主机随即获得 IP 地址并自动加入该子网实现数据交互。 设备上电启动并经 **USB Type-C** 接口连接主机后,系统将自动配置本地网络子网,**用户主机**将被分配静态 IP 地址 `10.10.99.100`,**带外管理芯片**静态 IP 地址 `10.10.99.97`。**推理模组**(IP: `10.10.99.98`)和**应用模组**(IP: `10.10.99.99`)——均部署独立的 SSH 服务,支持用户通过标准 SSH 客户端(如 OpenSSH、PuTTY)直接访问,而管理模组则需通过串口工具访问。 ## 网络配置 ### 如何通过主机为 RM-01 提供互联网访问(以 macOS 为例) 在通过 USB Type-C 连接**用户主机**后,RM-01 将在网络接口列表中显示为: * **`AX88179A`**(开发者版本) * **`RMinte RM-01`**(商业发行版本) 打开 **系统设置**(System Settings) 进入 **网络**(Network) → **共享**(Sharing) 启用 **互联网共享**(Internet Sharing) 点击共享设置旁的 **"i"图标**,进入配置界面: * 将 **"共享以下来源的连接"**(Share your connection from)设置为:**Wi-Fi** * 在 **"使用以下端口共享给设备"**(To computers using)中,勾选:**AX88179A** 或 **RMinte RM-01**(根据设备型号选择) 点击 **完成**(Done) 返回 **网络**(Network)设置页面,手动配置 RM-01 的网络接口: * **IP 地址**:`10.10.99.100` * **子网掩码**:`255.255.255.0` * **路由器**(Router):`10.10.99.100`(即主机自身 IP) 此配置将主机作为网关,为 RM-01 提供 NAT 网络访问。RM-01 的默认网关和 DNS 均由主机通过 DHCP 服务自动分配,手动设置 IP 可确保其始终位于 `10.10.99.0/24` 子网内,与设备内部服务通信一致。 ### 关于带外管理芯片 带外管理芯片除了肩负一定的加密工作,还安装了 **RM-01** 的**实时系统性能监控面板** —— **RobOS**。 用户可以通过浏览器访问 `http://10.10.99.97`,实时监看各模组的**连接情况**和**运行情况**。 RobOS 提供直观的 Web 界面,无需额外安装软件即可监控设备状态。 ## 系统架构 ### 关于 CFexpress Type-B 存储卡 **CFexpress Type-B** 存储卡是 RM-01 设备的核心组件之一,承担系统引导、模型推理框架部署及 ISV/SV 软件分发与授权认证的关键功能。 该存储卡划分为三个独立分区: **系统分区** 推理模组的操作系统与核心运行环境安装于该分区。 严禁用户或开发者访问、修改或删除该分区内容。任何未经授权的更改均可能导致推理模组无法启动或推理功能失效,且由此造成的硬件或软件损坏不在任何保修服务范围内。 **应用分区** 该分区用于暂存用户或开发者提交的 `Docker` 镜像文件。当镜像被写入 `rm01app` 后,RM-01 系统将自动将其迁移至主机内置的 **NVMe SSD** 存储中,并完成容器化部署。 请勿直接在该分区中运行或修改应用文件。 **模型分区** 该分区专用于存储用户或开发者加载的大规模人工智能模型(如 LLM、多模态模型等)。 关于模型格式、大小限制、加载流程及兼容性要求,请参阅下文"模型部署"章节。 ### 关于应用模组 **网络配置** * IP地址: `10.10.99.99` * 端口范围: `59000-59299` #### 应用模组硬件规格 ```text theme={null} 处理器:Intel Core i3-N305(8 核 8 线程,基础频率 1.8 GHz,最大睿频 3.8 GHz) 内存:16 GB / 24 GB LPDDR5-4800MT/s(板载,不可扩展) 存储:512 GB / 1 TB / 2 TB (可选) NVMe SSD ``` #### 应用模组 SSH 访问凭证 ```bash SSH 登录命令 theme={null} ssh rm01@10.10.99.99 # 默认用户名:rm01 # 默认密码:rm01(出厂预设,仅用于首次登录) ``` ```bash 修改密码 theme={null} # 首次登录后立即执行 passwd ``` **安全须知** 为保障系统安全,首次通过 SSH 登录后,请立即使用 `passwd` 命令修改默认密码;默认密码仅用于初始配置,严禁在生产及交付环境中使用。 #### 预装服务 应用模组的 `80` 端口下,预装了 **Open WebUI** 以方便用户进行简单的模型调试和对话工作。 用户可通过浏览器访问 `http://10.10.99.99` 使用 Open WebUI 进行模型测试和对话。 Open WebUI 提供友好的图形界面,无需编写代码即可快速测试和体验 AI 模型的对话能力。 ### 关于推理模组 **网络配置** * IP 地址:`10.10.99.98` * 服务端口范围:`58000–58999` **推理模组**是 RM-01 的核心计算单元,支持多种高性能 AI 推理配置,用户可根据模型规模与性能需求选择对应型号。 #### 硬件配置选项 | 显存 | 显存带宽 | 算力 | Tensor Core 数量 | | :----: | :--------: | :-----------------: | :------------: | | 32 GB | 204.8 GB/s | 200 TOPS (INT8) | 56 | | 64 GB | 204.8 GB/s | 275 TOPS (INT8) | 64 | | 64 GB | 273 GB/s | 1,200 TFLOPS (FP4) | 64 | | 128 GB | 273 GB/s | 2,070 TFLOPS (FP4) | 96 | #### 预装推理框架 RM-01 出厂时,**CFexpress Type-B** 存储卡中预装以下两个推理框架,均运行于推理模组: * **状态**: 自动启动 * **默认端口**: 58000 * **功能**: 提供 OpenAI 兼容 API 接口 * **支持请求**: 标准 POST `/v1/chat/completions` 等 * **状态**: 需手动启动 * **功能**: 文本嵌入(Embedding)服务 #### API 访问方式 成功加载模型后,可通过以下地址访问 **vLLM** 推理服务: ```bash theme={null} http://10.10.99.98:58000/v1/chat/completions ``` 支持标准 OpenAI 客户端(如 openai-python、curl、Postman)直接调用。 **安全须知** 为保障系统安全与稳定性,推理模组不开放 SSH 访问权限,用户与开发者无法通过任何方式直接登录或交互式操作该模块的底层操作系统。 任何试图绕过安全策略、直接访问推理模组的操作,均可能导致系统异常、数据损坏或服务中断,且不在保修服务范围内。 ## 模型部署 ### 关于模型 RM-01 支持推理多种人工智能模型,包括但不限于: 大语言模型 多模态模型 视觉语言模型 文本嵌入模型 重排序模型 所有模型文件均需存储于设备内置的 **CFexpress Type-B** 存储卡中,用户需使用兼容的 **CFexpress Type-B** 读卡器在主机端进行模型的上传、管理与更新。 当 **CFexpress Type-B** 存储卡接入 RM-01 后,系统会将其挂载为名为 `models` 的只读数据卷,路径为 `/home/rm01/models`。其标准文件结构如下: ```bash theme={null} models/ ├── auto/ # 系统自动加载模型目录(生产级部署) │ ├── embedding/ # 嵌入模型(系统不自动加载,见下文说明) │ ├── llm/ # 大语言模型(权重文件直接存放,见下文) │ └── reranker/ # 重排序模型(系统不自动加载,见下文) └── dev/ # 开发者自定义模型目录(高优先级) ├── embedding/ ├── llm/ └── reranker/ ``` * `auto/` 目录用于**轻量级、标准化模型部署**,由系统自动识别 * `dev/` 目录用于**开发者精细控制模型行为**,优先级高于 `auto/`,系统将忽略 `auto/` 中的模型 ### 部署模式选择 适用于快速验证和标准化部署的简化模式。 将模型的**完整权重文件**(如 `.safetensors`、`.bin`、`.pt`、`.awq` 等)**直接放置于 `auto/llm/` 目录下**,**禁止使用子文件夹嵌套**。 ```text ❌ 错误示例 theme={null} auto/llm/Qwen3-30B-A3B-Instruct-2507-AWQ/model.safetensors ``` ```text ✅ 正确示例 theme={null} auto/llm/model-001-of-006.safetensors auto/llm/config.json auto/llm/tokenizer.json auto/llm/vocab.json ``` * 设备开机后,系统将扫描 `auto/llm/` 目录,自动加载符合兼容格式的模型 * **不支持嵌入模型(embedding)与重排序模型(reranker)的自动加载**,仅支持 LLM * 模型加载后,**默认启用基础推理能力**,**不开启以下高级功能**: * Speculative Decoding(推测解码) * Prefix Caching(前缀缓存) * Chunked Prefill(分块预填充) * 最大上下文长度(`max_model_len`)将被限制为系统安全阈值(通常 ≤ 8192 tokens) * **性能优化受限**:为保障系统稳定性与多任务并发能力,自动模式下的模型将使用保守的显存分配策略(`gpu_memory_utilization` ≤ 0.8) **重要提示** 自动模式适用于**快速验证模型兼容性**或**标准化部署场景**,**不适用于生产级高性能推理**。如需完整性能,请使用**手动模式(dev)**。 开发者模式,提供完整的配置控制和性能优化能力。 **手动模式优先级高于自动模式**。当 `dev/` 的 `embedding`、`llm`、`reranker` 三个目录下存在任意 `.yaml` 配置文件时,系统将**完全忽略 `auto/` 中的所有模型**。 #### 配置文件结构 在 `dev/` 的 `embedding`、`llm`、`reranker` 三个目录下,需分别放置以下三个 YAML 配置文件,分别用于启动对应模型服务: | 文件名 | 用途 | | -------------------- | ------------------------ | | `embedding_run.yaml` | 启动嵌入模型服务(Text Embedding) | | `llm_run.yaml` | 启动大语言模型服务(LLM) | | `reranker_run.yaml` | 启动重排序模型服务(Reranker) | 所有文件必须位于 `dev/` 目录下相对应的文件夹内,**文件名必须完全匹配**,大小写敏感。 #### 示例配置文件 ```yaml llm_run.yaml theme={null} model: /home/rm01/models/dev/llm/Qwen3-30B-A3B-Instruct-2507-AWQ port: 58000 gpu_memory_utilization: 0.85 max_model_len: 24576 served_model_name: RM-01 LLM enable_prefix_caching: true enable_chunked_prefill: true max_num_batched_tokens: 512 block_size: 16 tensor_parallel_size: 1 dtype: auto ``` **配置说明** * `model` 路径必须为**绝对路径**,指向模型文件所在目录(非压缩包) * `port` 必须位于 `58000–58999` 范围内,且不可与其他服务冲突 * `gpu_memory_utilization` 建议设置为 0.7–0.9,以最大化吞吐量 * 所有参数均遵循 [vLLM 官方文档规范](https://docs.vllm.ai/en/latest/index.html) #### 部署步骤 将模型文件(完整目录)复制至 `dev/llm/`(或 `embedding/`、`reranker/`) 编写并放置对应的 `.yaml` 配置文件 插入 **CFexpress Type-B** 存储卡并重启 RM-01 系统将自动加载 `dev/` 中的配置,并启动对应服务 服务启动后,可通过 `http://10.10.99.98:58000/v1/chat/completions` 等接口访问 **建议**:首次部署时,建议使用 `vLLM` 官方提供的 `--load-format auto` 和 `--dtype auto`,以自动适配模型格式。 ### 安全与维护须知 * **禁止直接 SSH 登录推理模组**:所有模型管理必须通过 `CFexpress Type-B` 存储卡完成 * **模型文件必须为原始权重**:禁止使用压缩包(.zip/.tar.gz)、加密包或非标准格式 * **文件权限**:所有模型文件需为可读(`chmod 644`),目录需可执行(`chmod 755`) * **版本控制**:建议使用 Git 或文件命名规范(如 `Qwen3-30B-A3B-Instruct-v1.2-20250930`)管理模型版本 * **备份建议**:每次更新模型前,请备份 `dev/` 和 `auto/` 目录,避免配置丢失 ### 模式选择建议 | 场景 | 推荐模式 | 说明 | | ---------- | ------------------------- | --------- | | 快速验证模型是否兼容 | 自动模式(auto) | 无需配置,即插即用 | | 生产环境高性能推理 | 手动模式(dev) + 精细配置 | 完整性能优化 | | 多模型并行部署 | 手动模式(dev) + 多个 `.yaml` 文件 | 灵活的服务编排 | | 开发调试、原型验证 | 手动模式(dev) | 完全控制权 | ## 技术支持 完整的 API 参考和技术文档 示例代码和开源工具 开发者社区和技术讨论 专业技术支持服务 ***
© 2025 泛灵(成都)人工智能科技有限公司 版权所有
# 企业快速入门 Source: https://docs.rminte.com/essentials_cn/quickstart_cn 通过这份全面的企业指南,快速让您的团队开始使用 RM-01 AI 超级计算机 ## 设备概述 RM-01 AI 超级计算机是一款便携式 AI 解决方案,结合了高性能、安全性和易用性,专为企业级 AI 应用而设计。它在本地进行离线运算,显著降低了数据安全风险和运营成本。 ### 系统架构 RM-01 由三个核心模组组成,通过内置的以太网交换芯片互联,形成独立的内部网络: **IP地址**: `10.10.99.98` 负责 AI 模型推理计算,提供高性能算力支持 **IP地址**: `10.10.99.99` 运行预装应用和业务系统,如 Open WebUI **IP地址**: `10.10.99.97` 提供 RobOS 实时监控和设备管理功能 当您通过 USB Type-C 连接 RM-01 到主机(电脑、手机或平板)时,主机将自动分配静态 IP 地址 `10.10.99.100` 并加入 RM-01 的内部网络。 ### 主要特性 * 支持 300+ 主流开源模型 * 高达 2070 TFLOPS 算力 * 128GB LPDDR5 高速显存 * 硬件级非对称加密保护 * 数据本地处理,永不对外传输 * 防篡改物理安全设计 * 5 分钟内完成部署设置 * 即插即用,无需专业 IT 支持 * 最大功耗仅 140W(TDP 60W) ### 设备接口 * **电源按钮**:短按开/关机,长按 5 秒强制关机 * **状态指示灯**:白色(运行中)、蓝色(启动中)、红色(错误) * **CFexpress 插槽**:用于模型和应用存储,支持热插拔 * **MicroSD 插槽**:用于Logo定制和密钥存储 * **扩展端口**:用于连接外部扩展设备(支持多模态输入) * **USB-C 电源端口**:连接电源适配器(支持 PD3.1) * **USB-C 网络端口**:连接个人设备或企业交换机 * **通风口**:确保通风良好,请勿遮挡 ## 快速开始 ### 首次设置 确认包装内包含以下物品: * RM-01 主机 * 预装模型和应用的 CFexpress 卡 * MicroSD 卡 * 保修卡 * 快速入门指南 * USB-C 电源适配器 * USB-C 转以太网适配器 * 将 USB-C 电源适配器连接到 RM-01 * 将电源适配器连接到电源插座 * 按下电源按钮 * 等待设备启动(约 1-2 分钟) * 稳定的白色指示灯表示设备正常运行 使用 USB Type-C 线将 RM-01 连接到您的主机(电脑、手机或平板)。设备上电后会自动配置内部网络,您的主机将被分配静态 IP 地址:`10.10.99.100` 连接成功后,您的主机网络接口将显示为 **AX88179A** 或 **RMinte RM-01**(视版本而定) 如需让 RM-01 访问互联网,可通过主机共享网络。以 macOS 为例: 1. 打开 **系统设置** > **网络** > **共享** 2. 启用 **互联网共享**,选择 **Wi-Fi** 作为共享来源 3. 在"共享给设备"中,勾选 RM-01 的网络接口 4. 手动设置 RM-01 网络接口参数: * **IP 地址**:`10.10.99.100` * **子网掩码**:`255.255.255.0` * **路由器**:`10.10.99.100` 5. 点击 **完成** 此配置让主机充当网关,RM-01 会自动获取默认网关和 DNS。如不需互联网访问,可跳过此步骤。 ### 访问管理界面 RM-01 提供多个管理入口,分别对应不同的功能模组: 访问管理模组的实时监控面板: ``` http://10.10.99.97 ``` 在此查看 RM-01 各模组的连接状态和运行情况,实时掌握设备健康状态。 访问应用模组的预装 Open WebUI: ``` http://10.10.99.99 ``` **默认登录信息**: * 用户名:`rm01` * 密码:`rm01` 首次登录后,请使用 SSH 登录(地址:`10.10.99.99`)并立即更改密码以确保安全。 直接访问推理模组的 API 服务: ``` http://10.10.99.98:58000/v1/chat/completions ``` 支持标准 OpenAI 客户端调用,可使用 `openai-python`、Postman 或 curl 等工具。 如果您购买或部署了其他企业应用,它们通常运行在应用模组(`10.10.99.99`)上。具体端口号和访问方式请联系您的服务提供商。 ## 模型管理 ### 加载开源模型(自动模式) RM-01 支持快速加载开源大语言模型(LLM),让您轻松验证模型兼容性。 使用 **CFexpress Type-B** 读卡器,将模型的完整权重文件放入存储卡的 `auto/llm/` 目录。 支持的格式包括:`.safetensors`、`.bin`、`.pt`、`.awq` 等 **正确示例**: ``` auto/llm/model-001-of-006.safetensors auto/llm/config.json auto/llm/tokenizer.json auto/llm/vocab.json ``` 不要使用子文件夹!文件必须直接放在 `auto/llm/` 目录下,而不是放在其子目录中。 **错误示例**: ``` auto/llm/Qwen3-30B-A3B-Instruct-2507-AWQ/model.safetensors ``` 将 **CFexpress Type-B** 存储卡插入 RM-01,重新启动设备。系统会自动扫描 `auto/llm/` 目录并加载兼容的模型。 模型加载完成后,您可以通过推理服务端点(`http://10.10.99.98:58000/v1/chat/completions`)访问推理服务。 **自动模式限制**: * 仅支持大语言模型(LLM),不支持嵌入模型(embedding)或重排序模型(reranker) * 适合快速验证模型兼容性,性能会受限(最大上下文长度 ≤ 8192 tokens,显存使用率 ≤ 0.8) * 如需更高性能或多模型并行部署,请联系服务提供商获取 **手动模式(dev)** 的技术支持 ## 使用 AI 应用 ### 查看和启动应用 1. 登录管理界面 2. 点击"应用库"标签 3. 浏览已安装的 AI 应用列表 点击应用名称查看详细信息,包括功能介绍、使用场景和所需资源 1. 点击"启动"按钮 2. 应用将在新标签页中打开 3. 按照应用提示进行操作 ## 数据管理 ### 数据导入 1. 在应用中找到"上传"或"导入"按钮 2. 选择要上传的文件 3. 等待上传完成 1. 准备要导入的文件 2. 使用提供的批量导入工具 3. 选择文件夹或多个文件 4. 启动导入过程 ### 数据安全保障 RM-01 使用多重安全措施保护您的数据 所有数据在设备内处理,永不上传到云端 使用非对称加密技术保护存储数据 防尘防潮设计,防止物理拆解 细粒度用户权限管理系统 ### 安全使用须知 **推理模组不支持 SSH 登录**。所有模型管理需通过 **CFexpress Type-B** 存储卡完成。这一设计确保推理服务的稳定性和安全性。 不要尝试直接访问推理模组(`10.10.99.98`)的 SSH 服务,所有模型部署应通过存储卡进行。 在更新模型或配置前,建议备份存储卡中的 `auto/` 和 `dev/` 目录,以防配置丢失或需要回滚到之前的版本。 * 首次登录后立即更改所有默认密码 * 使用强密码(至少 12 位,包含大小写字母、数字和特殊字符) * 定期更新密码(建议每 90 天更新一次) * 不要在多个系统中使用相同密码 ## 用户管理 ### 添加用户 使用管理员账户访问管理界面 点击左侧导航栏中的"用户管理"选项 点击"添加用户"按钮 输入用户名、邮箱、初始密码和权限级别 点击"创建"按钮完成用户添加 ### 权限设置 RM-01 支持细粒度权限控制: 完全访问权限,包括系统设置 可以使用所有应用,管理自己的数据 可以使用指定应用,受限的数据访问 仅能使用特定演示应用 **调整用户权限** 1. 进入"用户管理" 2. 选择要修改的用户 3. 点击"编辑权限" 4. 设置应用访问权限和数据访问范围 5. 保存更改 ## 定期维护 ### 状态监控 定期系统状态检查确保稳定运行 登录管理界面后,在仪表板上查看系统状态信息: * CPU 使用率 * 内存使用率 * 存储空间使用情况 * 数据读写速度 * 运行温度 * 网络连接状态 ### 更新管理 通过 CFexpress Type-B 存储卡更新模型: 1. 从服务提供商获取新模型文件 2. 使用读卡器将模型文件放入存储卡的 `auto/llm/` 目录 3. 确保文件结构正确(参考模型管理章节) 4. 重新插入存储卡并重启设备 建议在更新前备份现有模型文件 联系您的服务提供商获取应用更新包和具体更新步骤。 通常需要: 1. 登录应用模组(`10.10.99.99`) 2. 上传更新包 3. 按照提示完成更新流程 联系您的服务提供商获取系统更新支持。 系统更新可能需要专业技术支持,请勿自行尝试未经授权的系统更新操作。 ### 性能优化 如果您注意到设备性能下降,请尝试以下步骤 前往"应用管理"并停止任何不需要的应用 前往"系统维护"并选择"清理缓存"选项 在管理界面选择"重启",或短按电源按钮并选择"重启" ### 故障排除 **可能原因和解决方案**: * **电源问题**:检查电源适配器连接,确保使用原装电源 * **存储卡问题**:尝试取出并重新插入存储卡 * **系统问题**:按住电源按钮 10 秒强制关机,然后重新启动 **可能原因和解决方案**: * **资源不足**:关闭其他运行的应用 * **应用损坏**:重新安装应用 * **模型问题**:检查所需模型是否正确加载 **可能原因和解决方案**: * **网络适配器问题**:检查 USB-C 网络适配器连接,确认主机识别到 AX88179A 或 RMinte RM-01 网络接口 * **IP 地址配置错误**:确认主机 IP 地址为 `10.10.99.100`,子网掩码为 `255.255.255.0` * **防火墙限制**:确认企业防火墙未阻止以下地址: * `10.10.99.97`(管理模组) * `10.10.99.98`(推理模组) * `10.10.99.99`(应用模组) * **网络接口未启用**:在网络设置中确认 RM-01 网络接口已启用 **可能原因和解决方案**: * **多个应用同时运行**:减少运行的应用数量 * **大型模型加载**:切换到较小的模型 * **温度过高**:确保设备通风良好,避免堆叠 * **存储空间不足**:清理不必要的数据 * **自动模式限制**:如使用自动模式加载模型,性能受限于最大上下文 8192 tokens。联系服务提供商升级到手动模式以获得更高性能 **可能原因和解决方案**: * **文件结构错误**:确保模型文件直接放在 `auto/llm/` 目录下,不要使用子文件夹 * **文件格式不兼容**:检查是否使用了支持的格式(`.safetensors`、`.bin`、`.pt`、`.awq` 等) * **必要文件缺失**:确保包含完整的模型文件,包括 `config.json`、`tokenizer.json` 等配置文件 * **存储卡未正确插入**:重新插入 CFexpress Type-B 存储卡并重启设备 * **存储卡故障**:尝试使用其他 CFexpress Type-B 存储卡 ## 获取支持 ### 服务提供商联系信息 400-XXX-XXXX(工作日 9:00-18:00) [support@rminte.com](mailto:support@rminte.com) [https://support.rminte.com](https://support.rminte.com) ### 学习资源 全面的文档和指南 分步骤的可视化学习材料 与其他开发者和专家交流 社区讨论和问答 恭喜!您已完成 RM-01 的基本设置和使用概览!现在您可以: * 通过 `http://10.10.99.97` 使用 RobOS 监控设备状态 * 访问 `http://10.10.99.99` 使用 Open WebUI 进行模型调试 * 通过 CFexpress Type-B 存储卡加载自己的开源模型到 `auto/llm/` 目录 * 使用推理服务端点 `http://10.10.99.98:58000/v1/chat/completions` 进行 AI 推理 开始探索 AI 技术为您的企业带来的无限可能!如有任何问题,请联系您的服务提供商或访问我们的支持中心获取帮助。 如需深入开发或优化 RM-01 的性能,请参考《RM-01 开发者使用指南》或联系服务提供商获取更多技术支持。 ***
© 2025 泛灵(成都)人工智能科技有限公司 版权所有
# Introduction Source: https://docs.rminte.com/index Welcome to the RM-01 AI Supercomputer Documentation - your comprehensive guide to deploying and managing enterprise AI solutions