cache.h 903 B

12345678910111213141516171819202122232425262728293031323334353637383940
  1. #pragma once
  2. #include <torch/extension.h>
  3. #include <map>
  4. #include <vector>
  5. void swap_blocks(
  6. torch::Tensor& src,
  7. torch::Tensor& dst,
  8. const torch::Tensor& block_mapping);
  9. void copy_blocks(
  10. std::vector<torch::Tensor>& key_caches,
  11. std::vector<torch::Tensor>& value_caches,
  12. const torch::Tensor& block_mapping);
  13. void reshape_and_cache(
  14. torch::Tensor& key,
  15. torch::Tensor& value,
  16. torch::Tensor& key_cache,
  17. torch::Tensor& value_cache,
  18. torch::Tensor& slot_mapping,
  19. const std::string& kv_cache_dtype,
  20. const float kv_scale);
  21. void reshape_and_cache_flash(
  22. torch::Tensor& key,
  23. torch::Tensor& value,
  24. torch::Tensor& key_cache,
  25. torch::Tensor& value_cache,
  26. torch::Tensor& slot_mapping,
  27. const std::string& kv_cache_dtype);
  28. // Just for unittest
  29. void convert_fp8(
  30. torch::Tensor& dst_cache,
  31. torch::Tensor& src_cache,
  32. const float scale,
  33. const std::string& kv_cache_dtype);