Production-grade distributed training framework optimized for AMD Instinct GPUs.
- FSDP2 with activation checkpointing
- Pipeline parallelism for model sharding
- Tensor parallelism for intra-node distribution
- Mixed precision training (FP16/BF16/FP8)
- Gradient accumulation with dynamic loss scaling
| Model | MI300X | H100 | vs A100 |
|---|---|---|---|
| Llama-3-70B | 1,847 tok/s | 1,721 tok/s | 2.1x |
| Llama-3-8B | 14,200 tok/s | 13,800 tok/s | 2.3x |
pip install -r requirements.txt
torchrun --nproc_per_node=8 train.py \
--model meta-llama/Llama-3-70B \
--strategy fsdp --precision bf16 --batch_size 4