Resource Guide for Large Model Training
Last updated: 2025-05-09 15:54:30
This document aims to introduce the configuration resources that can guarantee the normal running of the model when performing large-scale model training on the TI-ONE platform, for your reference only.
The following are recommended resources for training the platform built-in open-source large model.
| Recommended Resources (SFT-FULL) BatchSize=1,MaxSequenceLength=2048 | Recommended Resources (SFT-LORA) BatchSize=1,MaxSequenceLength=2048 |
Models below 7B | HCCPNV6 Model: 1 card for models below 3b; 2 cards for 7b/8b models; | HCCPNV6 model: 1 GPU |
13b model | HCCPNV6 model: 4 GPUs | HCCPNV6 model: 1 card |
32b model | HCCPNV6 model: 8 GPUs | HCCPNV6 model: 2 GPUs |
70b model | HCCPNV6 model: 2 machines with 16 GPUs | HCCPNV6 model: 4 GPUs |
DeepSeek-R1-671b/DeepSeek-V3-671b | HCCPNV6 model: 32 machines with 256 GPUs | Not supported. |
Hunyuan-large | HCCPNV6 model: 8 machines with 64 GPUs | HCCPNV6 model: 8 GPUs |
The platform built-in open-source large model uses the LORA fine-tuning method by default, which can be configured through the FinetuningType parameter.
The 7b model requires 100 cores and 500g of memory on a single node; the 13b and 70b models require 150 cores and 1t of memory on a single node. It is recommended to use the complete machine resources for larger models.
Some models use tilearn acceleration technology, which can achieve about 30% acceleration effect when training on recommended resources.
Was this page helpful?
You can also Contact Sales or Submit a Ticket for help.
Yes
No
Feedback