Enable Multi-vector distance function kernel for f32, f16, i8, u8 and MinMax quantization. #883

Activity

1. Alex Razumov (arrayka)

   added this to DiskANN backlogon Apr 3, 2026
2. Alex Razumov (arrayka)

   added

   enhancementNew feature or requestNew feature or request

   on Apr 3, 2026
3. Alex Razumov (arrayka)

   added sub-issues on Apr 3, 2026

   • Add multi-vector distance kernels #884
   • Validate recall stability for multi-vector distances #885
4. Suryansh Gupta (suri-kumkaran)

   added sub-issues on Apr 28, 2026

   • RFC: Multi-Vector Distance Function Architecture #985
   • Implement exact/baseline multi-vector distance #986
   • Add tiled_reduce function with Kernel and ConvertTo abstraction (f16 and f32 impls with hardcoded reduce) #988
   • Add Reduce abstraction to support quantizations and modify tiled_reduce accordingly #989
   • Add MicroKernels to support f32, f16, u8, i8 and MinMax quantizations. #990
   • Add block transposed memory layout type #991
   • Define storage & serialization for quantization types MinMax. #992
5. Suryansh Gupta (suri-kumkaran)

   self-assigned this

   on Apr 28, 2026
6. Alex Razumov (arrayka)

   added this to the 2026-05 milestone on May 5, 2026
7. Alex Razumov (arrayka)

   changed the title [-]Enable Multi-vector support for Colbert re-ranking[/-] [+]Enable Multi-vector distance function kernel[/+] on May 5, 2026
8. Suryansh Gupta (suri-kumkaran)

   removed a sub-issue on May 8, 2026

   • Add Reduce abstraction to support quantizations and modify tiled_reduce accordingly #989
9. 3 remaining items

   Load more
10. Suryansh Gupta (suri-kumkaran)

    removed a sub-issue on May 8, 2026

    • Add Reduce abstraction to support quantizations and modify tiled_reduce accordingly #989
11. Suryansh Gupta (suri-kumkaran)

    added a sub-issue on May 8, 2026

    • Add Reduce abstraction to support quantizations and modify tiled_reduce accordingly #989
12. Suryansh Gupta (suri-kumkaran)

    removed a sub-issue on May 8, 2026

    • Add Reduce abstraction to support quantizations and modify tiled_reduce accordingly #989
13. Suryansh Gupta (suri-kumkaran)

    added a sub-issue on May 8, 2026

    • Add Reduce abstraction to support quantizations and modify tiled_reduce accordingly #989
14. Suryansh Gupta (suri-kumkaran)

    removed a sub-issue on May 8, 2026

    • Add Reduce abstraction to support quantizations and modify tiled_reduce accordingly #989
15. Suryansh Gupta (suri-kumkaran)

    added sub-issues on May 13, 2026

    • Add Reduce abstraction to support quantizations and modify tiled_reduce accordingly #989
    • Add support for reading L1/L2 cache sizes for different platforms(Windows, Linux, MacOS) for our efficient cache aware multi-vector distance functions. #1062
16. Aditya Krishnan (arkrishn94)

    added a sub-issue on Jun 19, 2026

    • Support optimized eigenprojection scoring kernel #1190
17. Harsha Simhadri (harsha-simhadri)

    moved this to In Progress in DiskANN backlogon Jun 30, 2026
18. Harsha Simhadri (harsha-simhadri)

    modified the milestones: 2026-05, 2026-07 on Jun 30, 2026
19. Suryansh Gupta (suri-kumkaran)

    changed the title [-]Enable Multi-vector distance function kernel[/-] [+]Enable Multi-vector distance function kernel for f32, f16 and MinMax quantization.[/+] on Jul 14, 2026
20. Suryansh Gupta (suri-kumkaran)

    changed the title [-]Enable Multi-vector distance function kernel for f32, f16 and MinMax quantization.[/-] [+]Enable Multi-vector distance function kernel for f32, f16, i8, u8 and MinMax quantization.[/+] on Jul 14, 2026
21. Magdalen Dobson Manohar (magdalendobson)

    modified the milestones: 2026-07, 2026-09 on Sep 8, 2026