DeepseekMoE support with Fused MoE kernel (#2453)
Co-authored-by: roy <jasonailu87@gmail.com>
This commit is contained in:
@@ -121,3 +121,12 @@ std::pair<std::vector<uint8_t>, std::vector<int64_t>> get_graph_buffer_ipc_meta(
|
||||
void register_graph_buffers(fptr_t _fa, const std::vector<std::string> &handles,
|
||||
const std::vector<std::vector<int64_t>> &offsets);
|
||||
#endif
|
||||
|
||||
void moe_align_block_size(
|
||||
torch::Tensor topk_ids,
|
||||
int num_experts,
|
||||
int block_size,
|
||||
torch::Tensor sorted_token_ids,
|
||||
torch::Tensor experts_ids,
|
||||
torch::Tensor num_tokens_post_pad
|
||||
);
|
||||
|
||||
Reference in New Issue
Block a user