fp16 models are not working on WebGPU on gLinux Machine #51
Description
It looks like that it is a known issue that fp16 is not working on WebGPU on Linux machines. In our benchmark, for the sentence similarity workload, according to observations in issue #11, the fp16 version of Alibaba-NLP/gte-base-en-v1.5 is the best option in terms of size and correctness of the result. But unfortunately, it is not working on gLinux. So we decided to use fp32 model.
Added this issue here to track the issue and once the problem solved, we can use fp16 model for this benchmark.