表格批量嵌入算子

std::tuple<at::Tensor, at::Tensor, std::optional<at::Tensor>> get_unique_indices_cuda(const at::Tensor &linear_indices, const int64_t max_indices, const bool compute_count)

去重索引。

std::tuple<at::Tensor, at::Tensor, std::optional<at::Tensor>, std::optional<at::Tensor>> get_unique_indices_with_inverse_cuda(const at::Tensor &linear_indices, const int64_t max_indices, const bool compute_count, const bool compute_inverse_indices)

去重索引。

std::tuple<at::Tensor, at::Tensor, std::optional<at::Tensor>> lru_cache_find_uncached_cuda(at::Tensor unique_indices, at::Tensor unique_indices_length, int64_t max_indices, at::Tensor lxu_cache_state, int64_t time_stamp, at::Tensor lru_state, bool gather_cache_stats, at::Tensor uvm_cache_stats, bool lock_cache_line, at::Tensor lxu_cache_locking_counter, const bool compute_inverse_indices)

查找 LRU 缓存以找到未缓存的索引，然后根据集合对其进行排序。

int64_t host_lxu_cache_slot(int64_t h_in, int64_t C)

将索引映射到 cache_set。h_in：linear_indices；C：#cache_sets。

at::Tensor linearize_cache_indices_cuda(const at::Tensor &cache_hash_size_cumsum, const at::Tensor &indices, const at::Tensor &offsets, const std::optional<at::Tensor> &B_offsets, const int64_t max_B, const int64_t indices_base_offset)

线性化所有表的索引，使其唯一

at::Tensor linearize_cache_indices_from_row_idx_cuda(at::Tensor cache_hash_size_cumsum, at::Tensor update_table_indices, at::Tensor update_row_indices)

线性化所有表的索引，使其唯一。请注意，update_table_indices 和 update_row_indices 来自行索引格式，用于就地更新。

at::Tensor direct_mapped_lxu_cache_lookup_cuda(at::Tensor linear_cache_indices, at::Tensor lxu_cache_state, int64_t invalid_index, bool gather_cache_stats, std::optional<at::Tensor> uvm_cache_stats)

LRU 缓存：从 weights 中获取与 linear_cache_indices 对应的行，并将它们插入到时间步 time_stamp 的缓存中

. void lru_cache_populate_cuda(

at::Tensor weights,

at::Tensor hash_size_cumsum,

int64_t total_cache_hash_size,

at::Tensor cache_index_table_map,

at::Tensor weights_offsets,

at::Tensor D_offsets,

at::Tensor linear_cache_indices,

at::Tensor lxu_cache_state,

at::Tensor lxu_cache_weights,

int64_t time_stamp,

at::Tensor lru_state,

bool stochastic_rounding,

bool gather_cache_stats,

std::optional<at::Tensor> uvm_cache_stats,

bool lock_cache_line,

std::optional<at::Tensor> lxu_cache_locking_counter);

/ / LRU 缓存：从 /weights 中获取与 linear_cache_indices 对应的行，并将它们插入到时间步 time_stamp 的缓存中

. / weights 和 lxu_cache_weights 具有 “uint8_t” 字节元素 void lru_cache_populate_byte_cuda(

at::Tensor weights,

at::Tensor hash_size_cumsum,

int64_t total_cache_hash_size,

at::Tensor cache_index_table_map,

at::Tensor weights_offsets,

at::Tensor weights_tys,

at::Tensor D_offsets,

at::Tensor linear_cache_indices,

at::Tensor lxu_cache_state,

at::Tensor lxu_cache_weights,

int64_t time_stamp,

at::Tensor lru_state,

int64_t row_alignment,

bool gather_cache_stats,

std::optional<at::Tensor> uvm_cache_stats);

/ / direct_mapped (assoc=1) 版本的 lru_cache_populate_byte_cuda void direct_mapped_lru_cache_populate_byte_cuda(

at::Tensor weights,

at::Tensor hash_size_cumsum,

int64_t total_cache_hash_size,

at::Tensor cache_index_table_map,

at::Tensor weights_offsets,

at::Tensor weights_tys,

at::Tensor D_offsets,

at::Tensor linear_cache_indices,

at::Tensor lxu_cache_state,

at::Tensor lxu_cache_weights,

int64_t time_stamp,

at::Tensor lru_state,

at::Tensor lxu_cache_miss_timestamp,

int64_t row_alignment,

bool gather_cache_stats,

std::optional<at::Tensor> uvm_cache_stats);

/ / LFU 缓存：从 /weights 中获取与 linear_cache_indices 对应的行

, 并将它们插入到缓存中。void lfu_cache_populate_cuda(

at::Tensor weights,

at::Tensor cache_hash_size_cumsum,

int64_t total_cache_hash_size,

at::Tensor cache_index_table_map,

at::Tensor weights_offsets,

at::Tensor D_offsets,

at::Tensor linear_cache_indices,

at::Tensor lxu_cache_state,

at::Tensor lxu_cache_weights,

at::Tensor lfu_state,

bool stochastic_rounding);

/ / LFU 缓存：从 /weights 中获取与 linear_cache_indices 对应的行

, 并将它们插入到缓存中。/ weights 和 lxu_cache_weights 具有 “uint8_t” 字节元素 void lfu_cache_populate_byte_cuda(

at::Tensor weights,

at::Tensor cache_hash_size_cumsum,

int64_t total_cache_hash_size,

at::Tensor cache_index_table_map,

at::Tensor weights_offsets,

at::Tensor weights_tys,

at::Tensor D_offsets,

at::Tensor linear_cache_indices,

at::Tensor lxu_cache_state,

at::Tensor lxu_cache_weights,

at::Tensor lfu_state,

int64_t row_alignment);

/ / 查找 LRU/LFU 缓存：找到所有索引的缓存权重位置。/ 查找缓存中与 linear_cache_indices 对应的槽位

, 对于缺失值使用 / 哨兵值。at::Tensor lxu_cache_lookup_cuda(

at::Tensor linear_cache_indices,

at::Tensor lxu_cache_state,

int64_t invalid_index,

bool gather_cache_stats,

std::optional<at::Tensor> uvm_cache_stats,

std::optional<at::Tensor> num_uniq_cache_indices,

std::optional<at::Tensor> lxu_cache_locations_output);

at::Tensor emulate_cache_miss(

at::Tensor lxu_cache_locations,

const int64_t enforced_misses_per_256,

const bool gather_cache_stats,

at::Tensor uvm_cache_stats);

/ / 查找 LRU/LFU 缓存：找到所有索引的缓存权重位置。/ 查找缓存中与 linear_cache_indices 对应的槽位，对于缺失值使用哨兵值。

void lxu_cache_flush_cuda(at::Tensor uvm_weights, at::Tensor cache_hash_size_cumsum, at::Tensor cache_index_table_map, at::Tensor weights_offsets, at::Tensor D_offsets, int64_t total_D, at::Tensor lxu_cache_state, at::Tensor lxu_cache_weights, bool stochastic_rounding)

刷新缓存：将缓存中的权重存储到后备存储。

void reset_weight_momentum_cuda(at::Tensor dev_weights, at::Tensor uvm_weights, at::Tensor lxu_cache_weights, at::Tensor weights_placements, at::Tensor weights_offsets, at::Tensor momentum1_dev, at::Tensor momentum1_uvm, at::Tensor momentum1_placements, at::Tensor momentum1_offsets, at::Tensor D_offsets, at::Tensor pruned_indices, at::Tensor pruned_indices_offsets, at::Tensor logical_table_ids, at::Tensor buffer_ids, at::Tensor cache_hash_size_cumsum, at::Tensor lxu_cache_state, int64_t total_cache_hash_size)

void lxu_cache_locking_counter_decrement_cuda(at::Tensor lxu_cache_locking_counter, at::Tensor lxu_cache_locations)

根据 lxu_cache_locations 递减 LRU/LFU 缓存计数器。

void lxu_cache_locations_update_cuda(at::Tensor lxu_cache_locations, at::Tensor lxu_cache_locations_new, std::optional<at::Tensor> num_uniq_cache_indices)

就地更新 lxu_cache_locations 到新的值，只有当 lxu_cache_locations[i] == -1 且 lxu_cache_locations_new[i] >= 0 时才应更新