Case note
Train a promotions model on H100 SXM
ALPHV fine-tuned Qwen2.5-14B-Instruct with LoRA and Flash Attention on H100 SXM, through FPT AI Factory, to recommend promotions from historical sales and promotions data.
Partner
FPT AI Factory
Industry
Retail
Stack
Qwen2.5-14B · LoRA · Flash Attention · H100 SXM
The situation
Promotions are planned from historical sales and promotions data. ALPHV set out to train a specialised model that recommends promotions from that data.
What ALPHV built
ALPHV fine-tuned Qwen2.5-14B-Instruct using LoRA with Flash Attention, producing a model that recommends promotions from historical sales and promotions data.
Base model: Qwen2.5-14B-Instruct
Fine-tuning: LoRA
Attention: Flash Attention
Hardware: H100 SXM through FPT AI Factory
How it runs
Training ran on H100 SXM through the FPT AI Factory partnership. LoRA trains a small set of adapter weights instead of the full model, and Flash Attention speeds up attention on the GPU.
Outcome
A specialised promotions model built on Qwen2.5-14B-Instruct. Access to AI hardware through FPT AI Factory shortens training time and lowers the cost of AI R&D and production for ALPHV’s clients.
Planning something similar?
The engineers who scope your AI are the ones who ship and improve it. No hand-offs.
Talk to an AI engineer
Keep exploring
Service
AI engineering
The capability behind this work, from first model to go-live.
Explore the service
Approach
How we work
One team from business case to production: discover, design, implement and improve.
See how we work
Let’s build what’s next
Tell us where your business needs to be. We’ll map the path to get there.