Container Compute Service:Deploy a distributed DeepSeek-R1 inference service with ACS GPU

División del modelo

DeepSeek-R1 cuenta con 671 mil millones de parámetros, lo cual supera los 96 GB de memoria de una sola GPU. Esta guía divide el modelo entre dos instancias de contenedor de GPU con paralelismo de modelo (PP=2) y paralelismo de datos (TP=8).

El paralelismo de modelo (PP=2) divide el modelo en dos etapas, cada una ejecutada en una instancia de contenedor de GPU. La primera etapa procesa la entrada y pasa los resultados intermedios a la segunda.

El paralelismo de datos (TP=8) distribuye el cálculo dentro de cada etapa entre 8 GPU. Así, los datos se procesan en paralelo y los resultados se combinan.

Arquitectura de despliegue distribuido

Esta solución despliega un servicio de inference distribuido para la versión completa de DeepSeek en ACS mediante vLLM y Ray. LWS administra el patrón de despliegue líder-trabajador, y Fluid proporciona una caché distribuida para acelerar la carga del modelo. vLLM se ejecuta en dos instancias de pod de GPU de ACS con 8 GPU cada una. Cada pod actúa como un Ray Group (un Ray head y un Ray worker) para dividir el modelo. Las distintas arquitecturas de despliegue afectan variables como tensor-parallel-size y LWS_GROUP_SIZE en el archivo YAML.

Paso 1: Preparar los archivos del modelo DeepSeek-R1

Los modelos de lenguaje grandes requieren mucho espacio en disco. Cree un volumen de NAS u OSS para almacenar de forma persistente los archivos del modelo. Esta guía utiliza OSS.

1. Ejecute el siguiente comando para descargar el modelo DeepSeek-R1 desde ModelScope.

   Nota

   Asegúrese de que git-lfs esté instalado. Si no lo está, ejecute yum install git-lfs o apt-get install git-lfs. Encontrará métodos adicionales en Installing Git Large File Storage.

   git lfs install
   GIT_LFS_SKIP_SMUDGE=1 git clone https://www.modelscope.cn/deepseek-ai/DeepSeek-R1.git
   cd DeepSeek-R1/
   git lfs pull

2. Cree un directorio en OSS y cargue el modelo en él.

   Nota

   Para instalar y configurar ossutil, siga Install ossutil.

   ossutil mkdir oss://<your-bucket-name>/models/DeepSeek-R1
   ossutil cp -r ./DeepSeek-R1 oss://<your-bucket-name>/models/DeepSeek-R1

3. Una vez almacenado el modelo en OSS, puede cargarlo de dos maneras.

   1. Montar el modelo directamente mediante un PVC y un PV: Más indicado para modelos pequeños que no tienen requisitos estrictos de velocidad de inicio o de carga.

      Console

      A continuación se muestran los detalles de configuración básicos del PV de ejemplo:

      Parámetro	Descripción
      Tipo de volumen de almacenamiento	OSS
      Name	llm-model
      Access key	Configure el accesskey ID y el accesskey secret para acceder a OSS.
      Bucket ID	Seleccione el bucket de OSS que creó en el paso anterior.
      OSS path	Seleccione la ruta donde se encuentra el modelo, por ejemplo, /models/DeepSeek-R1.

      A continuación se muestran los detalles de configuración básicos del PVC de ejemplo:

      Parámetro	Descripción
      Tipo de solicitud de almacenamiento	OSS
      Name	llm-model
      Modo de asignación	Seleccione Existing Volume.
      Volumen existente	Haga clic en el enlace Select Existing Volume y seleccione el PV que creó.

      kubectl

      A continuación se muestra un archivo YAML de ejemplo:

      apiVersion: v1
      kind: Secret
      metadata:
        name: oss-secret
      stringData:
        akId: <your-oss-ak> # Your AccessKey ID for accessing OSS.
        akSecret: <your-oss-sk> # Your AccessKey Secret for accessing OSS.
      ---
      apiVersion: v1
      kind: PersistentVolume
      metadata:
        name: llm-model
        labels:
          alicloud-pvname: llm-model
      spec:
        capacity:
          storage: 30Gi 
        accessModes:
          - ReadOnlyMany
        persistentVolumeReclaimPolicy: Retain
        csi:
          driver: ossplugin.csi.alibabacloud.com
          volumeHandle: llm-model
          nodePublishSecretRef:
            name: oss-secret
            namespace: default
          volumeAttributes:
            bucket: <your-bucket-name> # The bucket name.
            url: <your-bucket-endpoint> # The endpoint, such as oss-cn-hangzhou-internal.aliyuncs.com.
            otherOpts: "-o umask=022 -o max_stat_cache_size=0 -o allow_other"
            path: <your-model-path> # In this example, the path is /models/DeepSeek-R1/.
      ---
      apiVersion: v1
      kind: PersistentVolumeClaim
      metadata:
        name: llm-model
      spec:
        accessModes:
          - ReadOnlyMany
        resources:
          requests:
            storage: 30Gi
        selector:
          matchLabels:
            alicloud-pvname: llm-model

   2. Acelerar la carga del modelo con Fluid: Recomendado para modelos más grandes que requieren un inicio y una carga rápidos. Use Fluid to accelerate data access.

      1. Instale el componente ack-fluid (versión 1.0.11-* o posterior) desde el ACS App Marketplace mediante Helm. Create an application by using Helm.

      2. Habilite el modo privilegiado para el pod de ACS. Para habilitarlo, envíe un ticket.

      3. Cree un Secret para acceder a OSS.

         apiVersion: v1
         kind: Secret
         metadata:
           name: mysecret
         stringData:
           fs.oss.accessKeyId: xxx
           fs.oss.accessKeySecret: xxx

         Establezca fs.oss.accessKeyId y fs.oss.accessKeySecret con las credenciales del bucket de OSS creado anteriormente.

      4. Cree un Dataset y un JindoRuntime.

         apiVersion: data.fluid.io/v1alpha1
         kind: Dataset
         metadata:
           name: deepseek
         spec:
           mounts:
             - mountPoint:  oss://<your-bucket-name>       # Replace with your actual bucket name.
               options:
                 fs.oss.endpoint: <your-bucket-endpoint>    # Replace with your actual bucket endpoint.
               name: deepseek
               path: "/"
               encryptOptions:
                 - name: fs.oss.accessKeyId
                   valueFrom:
                     secretKeyRef:
                       name: mysecret
                       key: fs.oss.accessKeyId
                 - name: fs.oss.accessKeySecret
                   valueFrom:
                     secretKeyRef:
                       name: mysecret
                       key: fs.oss.accessKeySecret
         ---
         apiVersion: data.fluid.io/v1alpha1
         kind: JindoRuntime
         metadata:
           name: deepseek
         spec:
           replicas: 16    # Adjust as needed.
           master:
             podMetadata:
               labels:
                 alibabacloud.com/compute-class: performance
                 alibabacloud.com/compute-qos: default
           worker:
             podMetadata:
               labels:
                 alibabacloud.com/compute-class: performance
                 alibabacloud.com/compute-qos: default
               annotations:
                 kubernetes.io/resource-type: serverless
             resources:
               requests:
                 cpu: 16
                 memory: 128Gi
               limits:
                 cpu: 16
                 memory: 128Gi
           tieredstore:
             levels:
               - mediumtype: MEM
                 path: /dev/shm
                 volumeType: emptyDir
                 ## Adjust as needed.
                 quota: 128Gi
                 high: "0.99"
                 low: "0.95"

         Después de crearlos, ejecute el comando kubectl get pod | grep jindo para comprobar si los pods están en estado Running. Resultado esperado:

         deepseek-jindofs-master-0    1/1     Running   0          3m29s
         deepseek-jindofs-worker-0    1/1     Running   0          2m52s
         deepseek-jindofs-worker-1    1/1     Running   0          2m52s
         ...

      5. Almacene el modelo en la caché mediante la creación de un recurso DataLoad.

         apiVersion: data.fluid.io/v1alpha1
         kind: DataLoad
         metadata:
           name: deepseek
         spec:
           dataset:
             name: deepseek
             namespace: default
           loadMetadata: true

      6. Ejecute el siguiente comando para comprobar el estado de la caché.

         kubectl get dataload

         Resultado esperado:

         NAME       DATASET    PHASE       AGE     DURATION
         deepseek   deepseek   Executing   4m30s   Unfinished

         El campo PHASE muestra Executing mientras el almacenamiento en caché está en curso. Espere unos 20 minutos y vuelva a ejecutar el comando. El valor Complete indica que el proceso finalizó correctamente. Siga el progreso con kubectl logs $(kubectl get pods --selector=job-name=deepseek-loader-job -o jsonpath='{.items[0].metadata.name}') | grep progress.

         Parámetros del recurso DataLoad de Fluid

         Parámetro	Descripción	Ejemplo
         Name	Nombre de la tarea de carga de datos.	deepseek
         Dataset	Nombre del conjunto de datos asociado.	deepseek
         Phase	Estado de la tarea de carga de datos, por ejemplo, Complete.	Executing, Complete
         Age	Hora de creación de la tarea de carga de datos.	4m30s
         Duration	Tiempo que tomó la tarea de carga de datos.	Unfinished, 16m29s

      7. Ejecute el siguiente comando para comprobar el recurso Dataset.

         kubectl get datasets

         Resultado esperado:

         NAME       UFS TOTAL SIZE   CACHED    CACHE CAPACITY   CACHED PERCENTAGE   PHASE   AGE
         deepseek   1.25TiB          1.25TiB   2.00TiB          100.0%              Bound   21h

         Parámetros del recurso Dataset de Fluid

         Parámetro	Descripción	Ejemplo
         Name	Nombre del conjunto de datos.	deepseek
         UFS total size	Tamaño total del conjunto de datos en el almacenamiento subyacente.	1.25TiB
         Cached	Cantidad de datos actualmente en la caché.	1.25TiB
         Cache capacity	Capacidad total de la caché.	2.00TiB
         Cached %	Porcentaje de almacenamiento en caché completado.	100.0%
         Phase	Estado del conjunto de datos, por ejemplo, Bound.	Bound
         Age	Hora de creación del recurso del conjunto de datos.	21h

Paso 2: Desplegar el modelo con ACS GPU

1. Instale el componente lws desde el ACS App Marketplace mediante Helm. Create an application by using Helm.

2. Despliegue el modelo mediante un LeaderWorkerSet.

   Nota

   • En el archivo YAML, reemplace alibabacloud.com/gpu-model-series: <example-model> por un modelo de GPU específico compatible con ACS. Para consultar la lista de modelos de GPU compatibles actualmente, comuníquese con su PDSA o envíe un ticket.

   • En comparación con TCP/IP, RDMA networking de alto rendimiento utiliza copia cero y omisión del kernel para evitar la duplicación de datos y los cambios frecuentes de contexto. Esto da como resultado menor latencia, mayor rendimiento y menor uso de CPU. ACS admite RDMA si configura la etiqueta alibabacloud.com/hpn-type: "rdma" en el archivo YAML. Para consultar la lista de GPU models que admiten RDMA, comuníquese con su PDSA o envíe un ticket.

   • Si utiliza Fluid para cargar el modelo, cambie el valor de claimName en ambas secciones de PVC por el nombre del conjunto de datos de Fluid.

   • Las distintas distributed deployment architectures afectan los valores de variables como tensor-parallel-size y LWS_GROUP_SIZE en el archivo YAML.

   Standard deployment

   apiVersion: leaderworkerset.x-k8s.io/v1
   kind: LeaderWorkerSet
   metadata:
     name: deepseek-r1-671b-fp8-distrubution
   spec:
     replicas: 1
     leaderWorkerTemplate:
       size: 2 # Total number of leader and worker pods.
       restartPolicy: RecreateGroupOnPodRestart
       leaderTemplate:
         metadata:
           labels: 
             role: leader
             alibabacloud.com/compute-class: gpu  # Specify the GPU type.
             alibabacloud.com/compute-qos: default # Specify the ACS QoS level.
             alibabacloud.com/gpu-model-series: <example-model> ## Specify the GPU model.
         spec:
           volumes:
             - name: llm-model
               persistentVolumeClaim:
                 ## If using Fluid, enter the Fluid dataset name here, for example: deepseek
                 claimName: llm-model
             - name: shm
               emptyDir:
                 medium: Memory
                 sizeLimit: 32Gi
           containers:
             - name: deepseek-r1-671b-leader
               image: registry-cn-hangzhou.ack.aliyuncs.com/ack-demo/vllm:v0.7.2
               env:
                 - name: NCCL_SOCKET_IFNAME # Specify the network interface card (NIC).
                   value: eth0
               command:
                 - sh
                 - -c
                 - "/vllm-workspace/ray_init.sh leader --ray_cluster_size=$(LWS_GROUP_SIZE);vllm serve /models/DeepSeek-R1/ --port 8000 --trust-remote-code --served-model-name ds --max-model-len 2048 --gpu-memory-utilization 0.95 --tensor-parallel-size 8 --pipeline-parallel-size 2 --enforce-eager"
   # Set tensor-parallel-size to the number of GPUs per pod.
               resources:
                 limits:
                   nvidia.com/gpu: "8"
                   cpu: "64"
                   memory: 512G
                 requests:
                   nvidia.com/gpu: "8"
                   cpu: "64"
                   memory: 512G           
               ports:
                 - containerPort: 8000
               volumeMounts:
                 - mountPath: /models/DeepSeek-R1
                   name: llm-model
                 - mountPath: /dev/shm
                   name: shm
       workerTemplate:
         metadata:
           labels: 
             alibabacloud.com/compute-class: gpu  # Specify the GPU type.
             alibabacloud.com/compute-qos: default # Specify the ACS QoS level.
             alibabacloud.com/gpu-model-series: <example-model> ## Specify the GPU model.
         spec:
           volumes:
             - name: llm-model
               persistentVolumeClaim:
                 ## If using Fluid, enter the Fluid dataset name here, for example: deepseek
                 claimName: llm-model
             - name: shm
               emptyDir:
                 medium: Memory
                 sizeLimit: 32Gi
           containers:
             - name: deepseek-r1-671b-worker
               image: registry-cn-hangzhou.ack.aliyuncs.com/ack-demo/vllm:v0.7.2
               env:
                 - name: NCCL_SOCKET_IFNAME # Specify the network interface card (NIC).
                   value: eth0
               command:
                 - sh
                 - -c
                 - "/vllm-workspace/ray_init.sh worker --ray_address=$(LWS_LEADER_ADDRESS)"
               resources:
                 limits:
                   nvidia.com/gpu: "8"
                   cpu: "64"
                   memory: 512G
                 requests:
                   nvidia.com/gpu: "8"
                   cpu: "64"
                   memory: 512G
               ports:
                 - containerPort: 8000
               volumeMounts:
                 - mountPath: /models/DeepSeek-R1
                   name: llm-model
                 - mountPath: /dev/shm
                   name: shm

   RDMA acceleration

   Si utiliza una imagen base de código abierto (como vLLM), agregue las siguientes variables de entorno al archivo YAML:

   Nombre	Valor
   NCCL_SOCKET_IFNAME	eth0
   NCCL_IB_TC	136
   NCCL_IB_SL	5
   NCCL_IB_GID_INDEX	3
   NCCL_DEBUG	INFO
   NCCL_IB_HCA	mlx5
   NCCL_NET_PLUGIN	none

   apiVersion: leaderworkerset.x-k8s.io/v1
   kind: LeaderWorkerSet
   metadata:
     name: deepseek-r1-671b-fp8-distrubution
   spec:
     replicas: 1
     leaderWorkerTemplate:
       size: 2 # Total number of leader and worker pods.
       restartPolicy: RecreateGroupOnPodRestart
       leaderTemplate:
         metadata:
           labels: 
             role: leader
             alibabacloud.com/compute-class: gpu  # Specify the GPU type.
             alibabacloud.com/compute-qos: default # Specify the ACS QoS level.
             alibabacloud.com/gpu-model-series: <example-model> ## Specify the GPU model.
             # Specify that the application runs in a high-performance RDMA network. Submit a ticket for information on supported GPU models.
             alibabacloud.com/hpn-type: "rdma"
         spec:
           volumes:
             - name: llm-model
               persistentVolumeClaim:
                 ## If using Fluid, enter the Fluid dataset name here, for example: deepseek
                 claimName: llm-model
             - name: shm
               emptyDir:
                 medium: Memory
                 sizeLimit: 32Gi
           containers:
             - name: deepseek-r1-671b-leader
               image: registry-cn-hangzhou.ack.aliyuncs.com/ack-demo/vllm:v0.7.2
               env:
                 - name: NCCL_SOCKET_IFNAME # Specify the network interface card (NIC).
                   value: eth0
                 - name: NCCL_IB_TC
                   value: "136"
                 - name: NCCL_IB_SL
                   value: "5"
                 - name: NCCL_IB_GID_INDEX
                   value: "3"
                 - name: NCCL_DEBUG
                   value: "INFO"
                 - name: NCCL_IB_HCA
                   value: "mlx5"
                 - name: NCCL_NET_PLUGIN
                   value: "none"                
               command:
                 - sh
                 - -c
                 - "/vllm-workspace/ray_init.sh leader --ray_cluster_size=$(LWS_GROUP_SIZE);vllm serve /models/DeepSeek-R1/ --port 8000 --trust-remote-code --served-model-name ds --max-model-len 2048 --gpu-memory-utilization 0.95 --tensor-parallel-size 8 --pipeline-parallel-size 2 --enforce-eager"
   # Set tensor-parallel-size to the number of GPUs per pod.
               resources:
                 limits:
                   nvidia.com/gpu: "8"
                   cpu: "64"
                   memory: 512G
                 requests:
                   nvidia.com/gpu: "8"
                   cpu: "64"
                   memory: 512G           
               ports:
                 - containerPort: 8000
               volumeMounts:
                 - mountPath: /models/DeepSeek-R1
                   name: llm-model
                 - mountPath: /dev/shm
                   name: shm
       workerTemplate:
         metadata:
           labels: 
             alibabacloud.com/compute-class: gpu  # Specify the GPU type.
             alibabacloud.com/compute-qos: default # Specify the ACS QoS level.
             alibabacloud.com/gpu-model-series: <example-model> ## Specify the GPU model.
             # Specify that the application runs in a high-performance RDMA network. Submit a ticket for information on supported GPU models.
             alibabacloud.com/hpn-type: "rdma"
         spec:
           volumes:
             - name: llm-model
               persistentVolumeClaim:
                 ## If using Fluid, enter the Fluid dataset name here, for example: deepseek
                 claimName: llm-model
             - name: shm
               emptyDir:
                 medium: Memory
                 sizeLimit: 32Gi
           containers:
             - name: deepseek-r1-671b-worker
               image: registry-cn-hangzhou.ack.aliyuncs.com/ack-demo/vllm:v0.7.2
               env:
                 - name: NCCL_SOCKET_IFNAME # Specify the network interface card (NIC).
                   value: eth0
                 - name: NCCL_IB_TC
                   value: "136"
                 - name: NCCL_IB_SL
                   value: "5"
                 - name: NCCL_IB_GID_INDEX
                   value: "3"
                 - name: NCCL_DEBUG
                   value: "INFO"
                 - name: NCCL_IB_HCA
                   value: "mlx5"
                 - name: NCCL_NET_PLUGIN
                   value: "none"      
               command:
                 - sh
                 - -c
                 - "/vllm-workspace/ray_init.sh worker --ray_address=$(LWS_LEADER_ADDRESS)"
               resources:
                 limits:
                   nvidia.com/gpu: "8"
                   cpu: "64"
                   memory: 512G
                 requests:
                   nvidia.com/gpu: "8"
                   cpu: "64"
                   memory: 512G
               ports:
                 - containerPort: 8000
               volumeMounts:
                 - mountPath: /models/DeepSeek-R1
                   name: llm-model
                 - mountPath: /dev/shm
                   name: shm

3. Exponga el servicio de inference con un recurso Service.

   apiVersion: v1
   kind: Service
   metadata:
     name: ds-leader
   spec:
     ports:
       - name: http
         port: 8000
         protocol: TCP
         targetPort: 8000
     selector:
       leaderworkerset.sigs.k8s.io/name: deepseek-r1-671b-fp8-distrubution
       role: leader
     type: ClusterIP

Paso 3: Compruebe el servicio de inference

1. Reenvíe un puerto local al servicio de inference.

   Nota

   El comando kubectl port-forward se usa únicamente para desarrollo y depuración. Para la red en producción, utilice Manage Ingresses.

   kubectl port-forward svc/ds-leader 8000:8000

   Resultado esperado:

   Forwarding from 127.0.0.1:8000 -> 8000
   Forwarding from [::1]:8000 -> 8000

2. Envíe una solicitud de inference al modelo.

   curl http://localhost:8000/v1/chat/completions \
     -H "Content-Type: application/json" \
     -d '{
       "model": "ds",
       "messages": [
         {
           "role": "system", 
           "content": "You are a friendly AI assistant."
         },
         {
           "role": "user",
           "content": "Introduce deep learning."
         }
       ],
       "max_tokens": 1024,
       "temperature": 0.7,
       "top_p": 0.9,
       "seed": 10
     }'

   Resultado esperado:

   {"id":"chatcmpl-4bc78b66e2a4439f8362bd434a60be57","object":"chat.completion","created":1739501401,"model":"ds","choices":[{"index":0,"message":{"role":"assistant","reasoning_content":null,"content":"<think>\nOkay, the user wants me to introduce deep learning. I need to think about how to answer this well. First, I should clarify the basic definition of deep learning—it's a branch of machine learning, right? Then I should compare it to traditional machine learning methods to highlight its advantages, like automatic feature extraction. I'll probably need to mention neural networks, especially the structure of deep neural networks with multiple hidden layers.\n\nNext, I should talk about the core components of deep learning, like activation functions, loss functions, and optimizers. The user might not be familiar with these terms, so I'll need to briefly explain the role of each part. I can use examples like ReLU as an activation function, Adam as an optimizer, and cross-entropy loss.\n\nThen, I should cover application areas. Computer vision and natural language processing are common examples. I should provide some concrete examples like image recognition and machine translation to make it easier for the user to understand. I could also mention industry applications, like in healthcare and finance, as the user might be interested in these practical use cases.\n\nI should also mention popular frameworks like TensorFlow and PyTorch, which make it easier to implement deep learning. And the importance of hardware acceleration, like GPUs, which explains why deep learning has developed so rapidly.\n\nAdditionally, I should probably discuss the challenges and limitations of deep learning, such as its dependency on data, high demand for computational resources, and poor interpretability. This will give the user a well-rounded view of its pros and cons. I might also mention future development directions, like efficient training algorithms and research into interpretability.\n\nThe user likely wants a basic understanding of deep learning. They might have some technical background, but not a deep one. They probably want to quickly grasp the key points and applications, so the answer needs to be well-structured and highlight the main points without getting too deep into technical details, but it can't be too brief either. I need to balance technical accuracy with readability.\n\nI should avoid using too much jargon, or explain it briefly when I do, for terms like \"neural network\" or \"convolutional neural network.\" The user could be a student or someone new to the field, so I should use plain language. I'll also use examples to help them connect the concepts to real-world applications, which makes them easier to remember.\n\nI also need to be clear about the relationship between deep learning and machine learning, explaining that deep learning is a subfield of machine learning, but it's deeper and handles more complex problems. I might also mention the backpropagation algorithm as one of the key training techniques.\n\nFinally, I'll provide a summary, emphasizing the impact and potential of deep learning and its future directions, to give the user a comprehensive overview. I'll check if I've missed any important points, like common model architectures such as CNNs and RNNs, which might be worth a brief mention to show the diversity.\n\nSometimes users are interested in the principles, but it's important to keep it concise here. The focus should be on an overview rather than in-depth technical details. I'll ensure the answer flows logically, starting from the definition, then moving to core components, applications, frameworks, challenges, and future directions. That structure seems reasonable.\n</think>\n\nDeep learning is a subfield of machine learning that aims to simulate the learning mechanism of the human brain by building multi-layered neural networks (known as \"deep\" networks). It automatically learns complex features and patterns from large amounts of data and is widely used in fields like image recognition, speech processing, and natural language processing.\n\n### Key concepts\n1.  **Artificial Neural Networks (ANNs)**:\n    *   Composed of an input layer, multiple hidden layers, and an output layer, with each layer containing multiple neurons.\n    *   Processes information by simulating the activation and transmission of signals between neurons.\n\n2.  **Automatic Feature Extraction**:\n    *   Traditional machine learning relies on manually engineered features. In contrast, deep learning uses multi-layer networks to automatically extract abstract features from data (such as edges and shapes from pixels in an image).\n\n3.  **Core Components**:\n    *   **Activation Function** (e.g., ReLU, Sigmoid): Introduces non-linearity, enhancing the model's expressive power.\n    *   **Loss Function** (e.g., Cross-Entropy, Mean Squared Error): Measures the difference between the model's predictions and the actual values.\n    *   **Optimizer** (e.g., SGD, Adam): Optimizes the network's parameters through backpropagation to minimize the loss.\n\n---\n\n### Typical models\n- **Convolutional Neural Network (CNN)**：\n  Designed for images, it uses convolutional kernels to extract spatial features. Classic models include ResNet and VGG.\n- **Recurrent Neural Network (RNN)**：\n  Processes sequential data (text, speech) by introducing a memory mechanism. Improved versions include LSTM and GRU.\n- **Transformer**：\n  Based on a self-attention mechanism, it has significantly advanced natural language processing performance (e.g., BERT, GPT series).\n\n---\n\n### Use cases\n- **Computer Vision**: Face recognition, medical image analysis (e.g., detecting lesions in lung CT scans).\n- **Natural Language Processing**: Intelligent chatbots, document summarization, translation (e.g., DeepL).\n- **Speech Technology**: Voice assistants (e.g., Siri), real-time captioning.\n- **Reinforcement Learning**: Game AI (AlphaGo), robotics control.\n\n---\n\n### Advantages and challenges\n- **Advantages**:\n  - Automatically learns complex features, reducing the need for manual intervention.\n  - Far outperforms traditional methods when given large amounts of data and high computing power.\n- **Challenges**:\n  - Relies on massive amounts of labeled data (e.g., requires tens of thousands of labeled medical images).\n  - High model training costs (e.g., training GPT-3 cost over ten million USD).\n  - \"Black box\" nature leads to poor interpretability, limiting its application in high-risk fields like medicine.\n\n---\n\n### Tools and trends\n- **Mainstream Frameworks**: TensorFlow (industry-deployment friendly), PyTorch (preferred for research).\n- **Research Directions**:\n  - Lightweight models (e.g., MobileNet for mobile devices).\n  - Self-supervised learning (reducing dependency on labeled data).\n  - Enhanced interpretability (e.g., visualizing the model's decision-making basis).\n\nDeep learning is pushing the boundaries of artificial intelligence, from generative AI (like Stable Diffusion for image generation) to autonomous driving, and continues to transform the tech ecosystem. Future advancements will likely focus on reducing computational costs and improving efficiency and interpretability.","tool_calls":[]},"logprobs":null,"finish_reason":"stop","stop_reason":null}],"usage":{"prompt_tokens":17,"total_tokens":1131,"completion_tokens":1114,"prompt_tokens_details":null},"prompt_logprobs":null}

   Parámetros del modelo

   Esta sección describe únicamente los parámetros que se utilizan para la comprobación. Consulte la referencia completa de parámetros en la documentación de la API de DeepSeek.

   Parámetro	Descripción	Ejemplo
   model	Modelo que se va a utilizar.	ds
   messages	Lista de mensajes de la conversación. role: rol del autor del mensaje. content: contenido del mensaje.	-
   max_tokens	Número máximo de tokens por solicitud. Valores válidos: Un número entero entre 1 y 8192. Si no se especifica el parámetro max_tokens, el valor predeterminado es 4096.	1024
   temperature	Controla la aleatoriedad del resultado. Los valores altos (por ejemplo, 1.0) aumentan la aleatoriedad; los valores bajos (por ejemplo, 0.2) incrementan el determinismo. Rango: de 0 a 2.	0.7