Closed hhy3 closed 10 months ago
LGTM.
Was able to install and run smoothly. Results on Standard D8lds v5 (8 vcpus, 16 GiB memory)
:
zilliz,zilliz_(R48_L500)_ef90,text2image-10M,10,31148.096981044335,0.0,7589.557703971863,11115584.0,0.0,0.0,ood,0.889166
zilliz,zilliz_(R48_L500)_ef95,text2image-10M,10,29863.85022221466,0.0,7589.557703971863,11115584.0,0.0,0.0,ood,0.8938750000000001
zilliz,zilliz_(R48_L500)_ef100,text2image-10M,10,28538.10639505947,0.0,7589.557703971863,11115584.0,0.0,0.0,ood,0.897826
zilliz,zilliz_(R48_L500)_ef102,text2image-10M,10,28001.62095870227,0.0,7589.557703971863,11115584.0,0.0,0.0,ood,0.899642
zilliz,zilliz_(R48_L500)_ef104,text2image-10M,10,27585.042124975516,0.0,7589.557703971863,11115584.0,0.0,0.0,ood,0.901216
zilliz,zilliz_(R48_L500)_ef106,text2image-10M,10,27374.67205313826,0.0,7589.557703971863,11115584.0,0.0,0.0,ood,0.9026379999999999
zilliz,zilliz_(R48_L500)_ef108,text2image-10M,10,26774.042755315306,0.0,7589.557703971863,11115584.0,0.0,0.0,ood,0.9041650000000001
zilliz,zilliz_(R48_L500)_ef110,text2image-10M,10,26395.14072612212,0.0,7589.557703971863,11115584.0,0.0,0.0,ood,0.9054420000000001
zilliz,zilliz_(R48_L500)_ef115,text2image-10M,10,25240.164360895433,0.0,7589.557703971863,11115584.0,0.0,0.0,ood,0.908855
zilliz,zilliz_(R48_L500)_ef120,text2image-10M,10,24576.384006000095,0.0,7589.557703971863,11115584.0,0.0,0.0,ood,0.911992
@harsha-simhadri OK to merge
@ingberam Thanks! Results LGTM
The code is not open-source. Do you have plans to open-source the code?
Expected result: ~29k qps with 90% recall Our solution is based on vamana graph and SQ8 quantization. There are some optimizations to make it search more efficiently. First, a multilevel bitset is used to accelerate bitset lookup. Second, highly efficient instruction sets and SIMD codes are used to accelerate quantized data computation. Third, queries are clustered before the graph search to get more cache locality. Finally, fine-grained memory management is used to make the search process more cache-friendly.