HiFloat is a low-bit-width (≤8 bit) floating-point data format designed for high-efficiency large model computing in the era of intelligent computing. The GCC-HiFloat community brings together industry partners to jointly promote its commercial deployment and technological evolution. To fully unlock the computational potential of this new data format and inspire innovative applications among intelligent computing partners, the GCC-HiFloat community is launching the HiFloat Large Model Co-Creation Initiative today, commencing with the first season of HiF8 Model Quantization Innovation Activities.
Overview of the HiF8 Model Quantization Innovation Activity
This activity invites AI developers to perform quantization validation using the HiF8 data format on the most influential mainstream open-source large models in the industry, with the goal of creating global demonstration cases for low-bit-width computing through developers' innovative practices.
Selected developers will have the opportunity to engage in in-depth discussions with the inventors of this technology and their technical teams regarding the design philosophy and application algorithms. They will receive guidance from these teams during the validation and development process, accumulating valuable experience in cutting-edge large model technologies.
For developers who complete the deliverables, the GCC-HiFloat community will provide globally verifiable certificates of completion. Developers who achieve challenge tasks will receive R&D funding support. Those who make innovative contributions will have their results included in relevant technology innovation case studies, and will be recommended/invited to share their achievements at related events.
Task List
Perform W8A8 HiF8 quantization validation for inference scenarios on the following designated open-source models based on NPU or GPU. Requirements:
-
Both Attention and Linear layers quantized to HiF8
-
Average accuracy loss on downstream tasks < 0.5%. Downstream task set includes: ARC / BoolQ / HellaS / Piqa / WinoG / Gsm8K / MMLU / Math500 / CMMLU, etc.
-
Baseline accuracy must be provided for comparison
Target Model
Developer Registration Instructions
Developers who register for the activity will subscribe to task notification emails. Detailed task descriptions, technical resources, subsequent operating instructions, and other information will be communicated to registrants via email. You are also welcome to scan the QR code to join the "GCC HiFloat Large Model Co-Creation Camp" for real-time event updates and interaction with the organizers and other developers.
[QR code placeholder – "Scan to join GCC HiFloat Large Model Co-Creation Camp"]
Important Notes on Activity Registration
-
As the tasks are challenging, it is recommended that registered developers have experience in large model quantization and fine-tuning based on NPU/GPU multi-card setups.
-
To support selected developers in completing their tasks, computing power support will be provided based on the developer's chosen task. This support requires real-name verification and signing an agreement with GCC.
-
Upon task completion, a validation report and model code must be submitted. Following expert review, R&D funding support and a certificate of completion will be provided.
High Performance Computing Development Group(HPC DG) 
