1. 20 Aug, 2024 1 commit
  2. 16 Aug, 2024 2 commits
  3. 12 Aug, 2024 1 commit
  4. 06 Aug, 2024 1 commit
  5. 05 Aug, 2024 3 commits
  6. 02 Aug, 2024 1 commit
  7. 31 Jul, 2024 3 commits
  8. 30 Jul, 2024 4 commits
  9. 29 Jul, 2024 2 commits
  10. 27 Jul, 2024 2 commits
  11. 26 Jul, 2024 1 commit
  12. 22 Jul, 2024 1 commit
  13. 21 Jul, 2024 1 commit
  14. 20 Jul, 2024 1 commit
  15. 18 Jul, 2024 1 commit
  16. 14 Jul, 2024 1 commit
  17. 03 Jul, 2024 1 commit
  18. 28 Jun, 2024 1 commit
  19. 26 Jun, 2024 1 commit
  20. 23 Jun, 2024 1 commit
  21. 20 Jun, 2024 3 commits
  22. 18 Jun, 2024 1 commit
  23. 14 Jun, 2024 2 commits
  24. 13 Jun, 2024 1 commit
  25. 12 Jun, 2024 1 commit
    • Cody Yu's avatar
      [Kernel] Vectorized FP8 quantize kernel (#5396) · 5985e342
      Cody Yu authored
      Inspired by #5146, this PR improves FP8 quantize kernel by vectorizing data transfer to better utilize memory bandwidth. Microbenchmark shows that this improved kernel can achieve 1.0x-1.5x speedup (especially when hidden size is large).
      
      In details, we applied 3 optimizations:
      
      - Use inverted scale so that most divisions are changed to multiplications.
      - Unroll the loop by 4 times to improve ILP.
      - Use vectorized 4 to transfer data between HBM and SRAM.
      5985e342
  26. 09 Jun, 2024 1 commit
  27. 07 Jun, 2024 1 commit