AMD Instinct ROCm 7.2+ MI300X Optimized

IndoFineTune

Multi-agent fine-tuning pipeline for Indonesian LLMs on AMD Instinct MI300X

📦

DataPrep

Clean, augment, validate

→

🏋️

Trainer

LoRA/QLoRA + ROCm

→

📊

Evaluator

BLEU, ROUGE, perplexity

→

🚀

Deployer

vLLM + ROCm serve

192 GB

HBM3 Memory

5.3 TB/s

Bandwidth

1,307

TFLOPS FP16

            $ python -m src.main --dataset data/train.jsonl --augment

            📦 DataPrep: 1,247 samples processed

            🏋️ Trainer: 3 epochs complete (47m)

            📊 Evaluator: loss=1.23, RECOMMENDED

            🚀 Deployer: http://0.0.0.0:8000

View on GitHub →