中文版

AKG CLI

⚠️ Deprecated: The CLI module (akg_cli) is no longer actively maintained and will not receive future updates.

1. Overview

akg_cli is the command-line interface for AKG Agents, providing interactive access to all framework capabilities.

For installation and basic configuration, see README.

2. Commands

Command	Description
akg_cli op	Kernel Agent — multi-backend, multi-DSL kernel generation. See Kernel Agent.
akg_cli common	Common Agent — general-purpose ReAct agent (demo).
akg_cli worker --start	Start Worker Service for distributed execution.
akg_cli worker --stop	Stop Worker Service.
akg_cli sessions	List all sessions with trace history.
akg_cli resume <session_id>	Resume a previous session.
akg_cli list	List supported subcommands.

3. akg_cli op

The primary command for AI kernel code generation.

Required Parameters

Parameter	Description	Examples
--framework	Compute framework	torch, mindspore
--backend	Hardware backend	ascend, cuda, cpu
--arch	Hardware architecture	ascend910b2, ascend910b4, a100, x86_64
--dsl	Target DSL	triton_ascend, triton_cuda, cuda, tilelang_cuda, cpp

Optional Parameters

Parameter	Default	Description
--intent	None	Provide requirement text directly (skip interactive prompt)
--task-file	None	Read task description file (KernelBench format)
--devices	None	Local device list, comma-separated (e.g., 0,1,2,3)
--worker-url	None	Worker Service URLs, comma-separated
--stream/--no-stream	--stream	Enable/disable LLM streaming output
--rag/--no-rag	--no-rag	Enable/disable RAG retrieval
--resume	None	Resume a previous session by ID
-y	False	Auto-confirm all prompts

Examples

# Ascend 910B2 with Triton
akg_cli op --framework torch --backend ascend --arch ascend910b2 \
  --dsl triton_ascend --devices 0,1,2,3,4,5,6,7

# CUDA A100 with Triton
akg_cli op --framework torch --backend cuda --arch a100 \
  --dsl triton_cuda --devices 0,1,2,3,4,5,6,7

# With direct intent
akg_cli op --framework torch --backend cuda --arch a100 \
  --dsl triton_cuda --devices 0 --intent "Generate a relu kernel"

# With KernelBench task file
akg_cli op --framework torch --backend cuda --arch a100 \
  --dsl triton_cuda --devices 0 --task-file path/to/task.py

4. Worker Service

For distributed execution across multiple machines:

# Start worker on each machine
akg_cli worker --start --port 9001

# Connect from client
akg_cli op --framework torch --backend cuda --arch a100 \
  --dsl triton_cuda --worker-url machine1:9001,machine2:9002

5. Session Management

# List all sessions
akg_cli sessions list

# Resume a session
akg_cli resume <session_id>