The team continues to refine our core infrastructure and boost performance across Gemma3 / zkML Interface key modules. Here’s a quick look at what’s been built and improved this week.
2/
Gemma3 Performance: Quantized Gemma3 model currently includes nearly 10 000 nodes; kernelized execution shows limited performance due to excessive node granularity.
3/
Gemma3 Refactor: Analyzed model structure and found most nodes are shape-related and redundant—potentially removable. In the ideal case, over 90 % of nodes can be eliminated.
4/
zkML Iface Latency Optimization: Refactored zkmlface codebase, cutting inference latency down to tens of milliseconds. The interface is not yet connected to the TEE environment.
5/
Next Steps:
Deploy the optimized zkmlface on a GPU TEE-enabled machine once available.
Compile the pruned Gemma3 graph into high-efficiency GPU kernels for integration testing.
Stay tuned for more updates
3,939
23
本页面内容由第三方提供。除非另有说明，欧易不是所引用文章的作者，也不对此类材料主张任何版权。该内容仅供参考，并不代表欧易观点，不作为任何形式的认可，也不应被视为投资建议或购买或出售数字资产的招揽。在使用生成式人工智能提供摘要或其他信息的情况下，此类人工智能生成的内容可能不准确或不一致。请阅读链接文章，了解更多详情和信息。欧易不对第三方网站上的内容负责。包含稳定币、NFTs 等在内的数字资产涉及较高程度的风险，其价值可能会产生较大波动。请根据自身财务状况，仔细考虑交易或持有数字资产是否适合您。