AKG CLI

⚠️ Deprecated: The CLI module (akg_cli) is no longer actively maintained and will not receive future updates.

AKG CLI (akg_cli) is the command-line interface for AI Kernel Generator. This document focuses on the two most common flows:

• Start a Worker Service: akg_cli worker --start
• Generate an operator: akg_cli op ...

How to get help:

akg_cli --help
akg_cli op --help
akg_cli worker --help

0. Installation

First, install akg_cli by running the following command in the aikg directory:

pip install -e .

1. Start a Worker Service

A Worker Service provides the backend runtime for kernel generation. Start it before running akg_cli op.

# Ascend 910B2
akg_cli worker --start --backend ascend --arch ascend910b2 --devices 0 --host 127.0.0.1 --port 9001

# CUDA A100: --backend cuda --arch a100
# akg_cli worker --start --backend cuda --arch a100 --devices 0 --host 127.0.0.1 --port 9001

Notes:

• --backend supports cuda or ascend. If omitted, defaults to WORKER_BACKEND or cuda.
• --arch defaults to WORKER_ARCH or a100.
• --devices is a comma-separated list of local device IDs (e.g., 0 or 0,1,2). It must be non-empty, contain no duplicates, and have no negative numbers.
• --host defaults to 0.0.0.0, --port defaults to 9001. For IPv6-only machines, use --host :: (dual-stack/IPv6 bind).

Stop the service when needed:

akg_cli worker --stop --port 9001

2. Generate an Operator

Use akg_cli op to start operator generation. The target config must be provided explicitly from CLI.

# Ascend 910B2
akg_cli op --framework torch --backend ascend --arch ascend910b2 --dsl triton_ascend --worker-url 127.0.0.1:9001

# CUDA A100: --backend cuda --arch a100 --dsl triton_cuda
# akg_cli op --framework torch --backend cuda --arch a100 --dsl triton_cuda --worker-url 127.0.0.1:9001

Optional parameters:

• --intent "..." to provide requirements directly (skip interactive prompt).
• --worker-url accepts multiple worker addresses separated by commas. The CLI also accepts both --worker-url and --worker_url.
• Choose exactly one of the following:
  • Use --worker-url/--worker_url when you want to run with remote Worker Service(s).
  • Use --devices when you want to run on local devices.
  • --devices and --worker-url/--worker_url are mutually exclusive.
• --output-path sets the base directory for saved_verifications (default: current working directory where akg_cli is started).
• --stream/--no-stream to enable/disable streaming output (default: --stream).
• --rag/--no-rag to enable/disable RAG (default: --no-rag).
• --yes/-y to auto-confirm all prompts.

Tip: for a mostly non-interactive run, use --intent "..." --yes (best-effort; the CLI may still ask for input in some cases).

Examples:

# Ascend 910B2 with intent
akg_cli op --framework torch --backend ascend --arch ascend910b2 --dsl triton_ascend --worker-url 127.0.0.1:9001 --intent "implement fused softmax kernel for input [batch, head, seq, dim]"

# CUDA A100 with intent
# akg_cli op --framework torch --backend cuda --arch a100 --dsl triton_cuda --worker-url 127.0.0.1:9001 --intent "implement fused softmax kernel for input [batch, head, seq, dim]"

Multiple workers example:

# Ascend 910B2 with multiple workers
akg_cli op --framework mindspore --backend ascend --arch ascend910b2 --dsl triton_ascend --worker-url 127.0.0.1:9001,127.0.0.1:9002

Local devices example (no worker URL needed):

# Ascend 910B2 with local devices
akg_cli op --framework torch --backend ascend --arch ascend910b2 --dsl triton_ascend --devices 0

# CUDA A100 with local devices
# akg_cli op --framework torch --backend cuda --arch a100 --dsl triton_cuda --devices 0

IPv6-only example (note the brackets):

akg_cli worker --start --host :: --port 9001
akg_cli op --framework torch --backend ascend --arch ascend910b2 --dsl triton_ascend --worker-url [::1]:9001

3. Interactive CLI Basics

akg_cli op runs an interactive prompt loop. You can iterate with multiple rounds of input and generation in the same session.

Input:

• Enter: submit the current input.
• Ctrl+J: insert a newline (multi-line input is supported).

Exit / cancel:

• When idle, press Ctrl+C to exit.
• While generating, press Ctrl+C to cancel the current round. You can then continue with the next round, or press Ctrl+C again to exit.

Panel:

• F2: show/hide the information panel.
• The panel is used to show the current task status, output/save path, and recent top implementation summaries.

History:

• Use the Up/Down arrows to browse input history.
• History is saved to ~/.akg_cli_history.