Status

● Connected

ROCm

6.1.2

System

MI300X · 8 GPUs

Project

Retail-LLM-Inference

AK

Admin User

Administrator

End-to-End Agent Pipeline

Workflow

YOUSUN Axion turns broken or slow AI workloads into optimized, benchmarked, AMD-ready deployments through an autonomous multi-agent workflow.

1

Upload Workload

User uploads AI code, repo, or ROCm error log.

2

Project Scan

Axion detects framework, precision, CUDA usage, and readiness risks.

3

ROCm Doctor

Diagnoses HIP memory, CUDA compatibility, and ROCm runtime problems.

4

Optimizer

Generates AMD GPU optimization plan with impact and effort level.

5

Patch Generator

Creates actual code changes such as BF16, inference mode, and device abstraction.

6

Benchmark

Compares before/after latency, tokens/sec, memory, and readiness score.

7

Reports

Exports AMD Readiness Report with business impact and final verdict.

8

AMD-Ready Output

Developer gets optimized, benchmarked, AMD-ready workload.

Optimization Loop

Optimizer, Patch Generator, Benchmark, and Kernel Lab can run repeatedly until target latency, memory, and readiness goals are reached.

Optimizer

Patch

Benchmark

Kernel Lab