Triton Inference Server CUDA Shared Memory API

openapi: 3.1.0
info:
  title: Triton Inference Server NVIDIA Triton Inference Server HTTP/REST CUDA Shared Memory API
  description: RESTful API for the NVIDIA Triton Inference Server, implementing the KServe V2 inference protocol with Triton-specific extensions. Provides endpoints for model inference, health checks, server and model metadata, model repository management, statistics, tracing, logging, and system shared memory management.
  version: '2.0'
  contact:
    name: NVIDIA Triton Team
    url: https://github.com/triton-inference-server/server
    email: triton@nvidia.com
  license:
    name: BSD 3-Clause
    url: https://github.com/triton-inference-server/server/blob/main/LICENSE
servers:
- url: http://localhost:8000
  description: Triton HTTP endpoint (default)
- url: http://{host}:{port}
  description: Custom Triton HTTP endpoint
  variables:
    host:
      default: localhost
      description: Triton server hostname or IP
    port:
      default: '8000'
      description: Triton HTTP port
tags:
- name: CUDA Shared Memory
  description: CUDA shared memory region management
paths:
  /v2/cudasharedmemory/status:
    get:
      operationId: cudaSharedMemoryStatus
      summary: Triton Inference Server Get CUDA Shared Memory Status
      description: Retrieve the status of all registered CUDA shared memory regions. This is a Triton extension to the KServe protocol.
      tags:
      - CUDA Shared Memory
      responses:
        '200':
          description: CUDA shared memory status returned successfully
          content:
            application/json:
              schema:
                type: array
                items:
                  $ref: '#/components/schemas/CudaSharedMemoryRegion'
        '400':
          description: Error retrieving status
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ErrorResponse'
  /v2/cudasharedmemory/region/{region_name}/register:
    post:
      operationId: cudaSharedMemoryRegister
      summary: Triton Inference Server Register a CUDA Shared Memory Region
      description: Register a CUDA shared memory region for use with inference requests.
      tags:
      - CUDA Shared Memory
      parameters:
      - $ref: '#/components/parameters/regionName'
      requestBody:
        required: true
        content:
          application/json:
            schema:
              type: object
              required:
              - raw_handle
              - device_id
              - byte_size
              properties:
                raw_handle:
                  type: object
                  description: CUDA IPC memory handle serialized as a base64-encoded object
                  properties:
                    b64:
                      type: string
                      format: byte
                      description: Base64-encoded CUDA IPC memory handle
                device_id:
                  type: integer
                  description: GPU device ID where the memory is allocated
                byte_size:
                  type: integer
                  description: Size of the CUDA shared memory region in bytes
      responses:
        '200':
          description: Region registered successfully
        '400':
          description: Error registering region
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ErrorResponse'
  /v2/cudasharedmemory/region/{region_name}/unregister:
    post:
      operationId: cudaSharedMemoryUnregister
      summary: Triton Inference Server Unregister a CUDA Shared Memory Region
      description: Unregister a previously registered CUDA shared memory region.
      tags:
      - CUDA Shared Memory
      parameters:
      - $ref: '#/components/parameters/regionName'
      responses:
        '200':
          description: Region unregistered successfully
        '400':
          description: Error unregistering region
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ErrorResponse'
  /v2/cudasharedmemory/unregister:
    post:
      operationId: cudaSharedMemoryUnregisterAll
      summary: Triton Inference Server Unregister All CUDA Shared Memory Regions
      description: Unregister all registered CUDA shared memory regions.
      tags:
      - CUDA Shared Memory
      responses:
        '200':
          description: All regions unregistered successfully
        '400':
          description: Error unregistering regions
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ErrorResponse'
components:
  parameters:
    regionName:
      name: region_name
      in: path
      required: true
      description: Name of the shared memory region
      schema:
        type: string
  schemas:
    CudaSharedMemoryRegion:
      type: object
      properties:
        name:
          type: string
          description: Name of the CUDA shared memory region
        device_id:
          type: integer
          description: GPU device ID
        byte_size:
          type: integer
          description: Size of the region in bytes
    ErrorResponse:
      type: object
      properties:
        error:
          type: string
          description: Error message describing what went wrong
externalDocs:
  description: Triton Inference Server Protocol Documentation
  url: https://github.com/triton-inference-server/server/blob/main/docs/protocol/README.md