# 5. Changing the Number of GPUs

Let's rerun the fine-tuning task with a different number of GPUs. MoAI Platform abstracts GPU resources into a single accelerator and automatically performs parallel processing. Therefore, there is no need to modify the PyTorch script even when changing the number of GPUs.

# Changing the Accelerator type

Switch the accelerator type using the moreh-switch-model tool. For instructions on changing the accelerator, please refer to the 3. Model fine-tuning

$ moreh-switch-model

Please contact your infrastructure provider and choose one of the following options before proceeding.

AMD MI250 GPU with 32 units
- When using Moreh's trial container: select 8xlarge
- When using KT Cloud's Hyperscale AI Computing: select 8xLarge.4096GB
AMD MI210 GPU with 64 units
AMD MI300X GPU with 16 units

# Training Parameters

Run the train_baichuan2_13b.py script again.

~/moreh-quickstart$ python tutorial/train_baichuan2_13b.py --batch-size 512

Since the available GPU memory has doubled, let's increase the batch size to 512 and run the training.

f the training proceeds normally, you should see the following log:

...
[info] Got DBs from backend for auto config.
[info] Requesting resources for MoAI Accelerator from the server...
[info] Initializing the worker daemon for MoAI Accelerator
[info] [1/8] Connecting to resources on the server (192.168.110.32:24174)...
[info] [2/8] Connecting to resources on the server (192.168.110.33:24174)...
[info] [3/8] Connecting to resources on the server (192.168.110.35:24174)...
[info] [4/8] Connecting to resources on the server (192.168.110.67:24174)...
[info] [5/8] Connecting to resources on the server (192.168.110.73:24174)...
[info] [6/8] Connecting to resources on the server (192.168.110.75:24174)...
[info] [7/8] Connecting to resources on the server (192.168.110.97:24174)...
[info] [8/8] Connecting to resources on the server (192.168.110.98:24174)...
[info] Establishing links to the resources...
[info] MoAI Accelerator is ready to use.
[info] The number of candidates is 65.
[info] Parallel Graph Compile start...
[info] Elapsed Time to compile all candidates = 11476 [ms]
[info] Parallel Graph Compile finished.
[info] The number of possible candidates is 6.
[info] SelectBestGraphFromCandidates start...
[info] Elapsed Time to compute cost for survived candidates = 1472 [ms]
[info] SelectBestGraphFromCandidates finished.
[info] Configuration for parallelism is selected.
[info] No PP, No TP, recomputation : 1, distribute_param : true, distribute_low_prec_param : true
[info] train: true

| INFO     | __main__:main:143 - [Step 1/52] Throughput : 1167.211504009616tokens/sec
| INFO     | __main__:main:143 - [Step 2/52] Throughput : 358524.96263602894tokens/sec
| INFO     | __main__:main:143 - [Step 3/52] Throughput : 380980.5659610025tokens/sec
| INFO     | __main__:main:143 - [Step 4/52] Throughput : 382460.244826232tokens/sec
| INFO     | __main__:main:143 - [Step 5/52] Throughput : 377403.73612910055tokens/sec
| INFO     | __main__:main:143 - [Step 6/52] Throughput : 382224.183245965tokens/sec
| INFO     | __main__:main:143 - [Step 7/52] Throughput : 380014.4669324378tokens/sec
...

Upon comparison with the results obtained when the number of GPUs was halved, you'll notice that the training progresses similarly, with an improvement in throughput.

When using AMD MI250 GPU 16 → 32 : approximately 198,000 tokens/sec → 370,000 tokens/sec

tutorial baichuan