Coder Design Document

Overview

The Coder is a core component in the AI Kernel Generator that converts algorithm designs into concrete implementation code. It inherits from AgentBase and is responsible for transforming high-level algorithm sketches (generated by Designer) into executable kernel code for specific hardware platforms and frameworks.

Core Functions

• Design-to-Code Translation: Converts algorithm designs into concrete implementation code
• Doc-Driven Integration: Leverages custom reference documents for accurate code generation
• Multi-DSL Support: Supports various target languages (Triton, SWFT, etc.)
• Framework Adaptation: Generates code for different frontend frameworks (MindSpore, PyTorch, etc.)
• Error-Guided Repair: Fixes generated code based on verification feedback
• API Integration: Intelligently selects appropriate APIs based on documentation

Initialization Parameters

Parameter Name	Type/Required	Description
op_name	str (Required)	Kernel name, identifying the specific kernel
task_desc	str (Required)	Task description, detailing the kernel functional requirements
dsl	str (Required)	Target DSL: "triton_cuda", "triton_ascend", "swft", etc.
framework	str (Required)	Frontend framework: "mindspore", "torch", "numpy", etc.
backend	str (Required)	Hardware backend: "ascend", "cuda", etc.
arch	str (Required)	Hardware architecture: "ascend910b4", "a100", etc.
workflow_config_path	str (Optional)	Workflow configuration file path (usually injected by Task from the orchestration plan)
config	dict (Required)	Complete orchestration plan configuration, including log_dir, agent_model_config, docs_dir, etc. (see Task Orchestration Plan Configuration)

Related docs: Workflow details in Workflow; documentation integration in Doc-Driven Integration Guide.

Doc-Driven Integration

The Coder loads reference documents via docs_dir.coder in the orchestration plan. For document list, directory structure, and authoring guidelines, see the Doc-Driven Integration Guide. Details are not duplicated here.

Configuration Example

# In task orchestration plan file
docs_dir:
  coder: "resources/docs/triton_ascend_docs"  # Coder reference documents
  
agent_model_config:
  coder: "your_model_name"
  api_generator: "your_api_model"  # For API document compression

Execution Flow

1. Initialization Stage

   • Load workflow configuration and create parser
   • Initialize code generation templates
   • Load all reference documents using Doc-Driven Integration
   • Prepare base document structure with loaded content

2. Processing Stage

   • Extract algorithm design from task information
   • Process API documentation (compress if needed)
   • Incorporate conductor suggestions and error feedback

3. Generation Stage

   • Execute LLM generation with comprehensive context
   • Return generated code, prompt, and reasoning process