| 1 | #include "core/registration.h" |
| 2 | #include "moe_ops.h" |
| 3 | |
| 4 | #include <torch/csrc/stable/library.h> |
| 5 | |
| 6 | STABLE_TORCH_LIBRARY_FRAGMENT(_moe_C, m) { |
| 7 | // Apply topk softmax to the gating outputs. |
| 8 | m.def( |
| 9 | "topk_softmax(Tensor! topk_weights, Tensor! topk_indices, Tensor! " |
| 10 | "token_expert_indices, Tensor gating_output, bool renormalize, Tensor? " |
| 11 | "bias, Tensor? is_padding) -> ()"); |
| 12 | |
| 13 | // Apply topk sigmoid to the gating outputs. |
| 14 | m.def( |
| 15 | "topk_sigmoid(Tensor! topk_weights, Tensor! topk_indices, Tensor! " |
| 16 | "token_expert_indices, Tensor gating_output, bool renormalize, " |
| 17 | "Tensor? bias, float routed_scaling_factor, Tensor? is_padding) -> ()"); |
| 18 | |
| 19 | m.def( |
| 20 | "topk_softplus_sqrt(Tensor! topk_weights, Tensor! topk_indices, Tensor! " |
| 21 | "token_expert_indices, Tensor gating_output, bool renormalize, float " |
| 22 | "routed_scaling_factor, Tensor? " |
| 23 | "bias, Tensor? input_ids, Tensor? tid2eid, Tensor? is_padding, " |
| 24 | "Tensor? bias_vl=None, int image_sentinel_lo=0) -> ()"); |
| 25 | |
| 26 | // Calculate the result of moe by summing up the partial results |
| 27 | // from all selected experts. topk_ids/expert_map are optional and, when |
| 28 | // both given, enable pad-aware reduce that skips (token, expert) |
| 29 | // slots that were never actually computed (unrouted, or routed to an |
| 30 | // expert not owned by this rank under expert parallelism). |
| 31 | m.def( |
| 32 | "moe_sum(Tensor input, Tensor! output, Tensor? topk_ids=None, " |
| 33 | "Tensor? expert_map=None) -> ()"); |
| 34 | |
| 35 | // Aligning the number of tokens to be processed by each expert such |
| 36 | // that it is divisible by the block size. |
| 37 | m.def( |
| 38 | "moe_align_block_size(Tensor topk_ids, int num_experts," |
| 39 | " int block_size, Tensor! sorted_token_ids," |
| 40 | " Tensor! experts_ids," |
| 41 | " Tensor! num_tokens_post_pad," |
| 42 | " Tensor? maybe_expert_map," |
| 43 | " Tensor(a!)? scatter_idx=None) -> ()"); |
| 44 | |
| 45 | // Aligning the number of tokens to be processed by each expert such |
| 46 | // that it is divisible by the block size, but for the batched case. |
| 47 | m.def( |
| 48 | "batched_moe_align_block_size(int max_tokens_per_batch," |
| 49 | " int block_size, Tensor expert_num_tokens," |
| 50 | " Tensor! sorted_token_ids," |
| 51 | " Tensor! experts_ids," |
| 52 | " Tensor! num_tokens_post_pad) -> ()"); |
| 53 | |
| 54 | // Aligning the number of tokens to be processed by each expert such |
| 55 | // that it is divisible by the block size. |
| 56 | m.def( |
| 57 | "moe_lora_align_block_size(Tensor topk_ids," |
| 58 | " Tensor token_lora_mapping," |
| 59 | " int num_experts," |
| 60 | " int block_size, int max_loras, " |
| 61 | " int max_num_tokens_padded, " |
| 62 | " int max_num_m_blocks, " |
| 63 | " Tensor !sorted_token_ids," |
| 64 | " Tensor !experts_ids," |
| 65 | " Tensor !num_tokens_post_pad," |
| 66 | " Tensor !adapter_enabled," |
| 67 | " Tensor !lora_ids," |
| 68 | " Tensor? maybe_expert_map) -> () "); |
| 69 | #ifndef USE_ROCM |
| 70 | m.def( |
| 71 | "moe_wna16_gemm(Tensor input, Tensor! output, Tensor b_qweight, " |
| 72 | "Tensor b_scales, Tensor? b_qzeros, " |
| 73 | "Tensor? topk_weights, Tensor sorted_token_ids, " |
| 74 | "Tensor expert_ids, Tensor num_tokens_post_pad, " |
| 75 | "int top_k, int BLOCK_SIZE_M, int BLOCK_SIZE_N, int BLOCK_SIZE_K, " |
| 76 | "int bit) -> Tensor"); |
| 77 | |
| 78 | m.def( |
| 79 | "moe_wna16_marlin_gemm(Tensor! a, Tensor? c_or_none," |
| 80 | "Tensor! b_q_weight, Tensor? b_bias_or_none," |
| 81 | "Tensor! b_scales, Tensor? a_scales, Tensor? global_scale, Tensor? " |
| 82 | "b_zeros_or_none," |
| 83 | "Tensor! workspace," |
| 84 | "Tensor sorted_token_ids," |
| 85 | "Tensor! expert_ids, Tensor! num_tokens_past_padded," |
| 86 | "Tensor! topk_weights, int moe_block_size, int top_k, " |
| 87 | "bool mul_topk_weights, int b_type_id," |
| 88 | "int size_m, int size_n, int size_k, bool use_atomic_add," |
| 89 | "bool use_fp32_reduce, bool is_zp_float," |
| 90 | "int thread_k, int thread_n, int blocks_per_sm) -> Tensor"); |
| 91 | |
| 92 | m.def( |
| 93 | "moe_permute(Tensor input, Tensor topk_ids," |
| 94 | "Tensor token_expert_indices, Tensor? expert_map, int n_expert," |
| 95 | "int n_local_expert," |
| 96 | "int topk, Tensor! permuted_input, Tensor! " |
| 97 | "expert_first_token_offset, Tensor! inv_permuted_idx, Tensor! " |
| 98 | "permuted_idx)->()"); |
| 99 | |
| 100 | m.def( |
| 101 | "moe_permute_with_scratch(Tensor input, Tensor topk_ids," |
| 102 | "Tensor token_expert_indices, Tensor? expert_map, int n_expert," |
| 103 | "int n_local_expert," |
| 104 | "int topk, Tensor! permuted_input, Tensor! " |
| 105 | "expert_first_token_offset, Tensor! inv_permuted_idx, Tensor! " |
| 106 | "permuted_idx, Tensor! sort_workspace, Tensor! permuted_experts_id, " |
| 107 | "Tensor! sorted_row_idx, Tensor! topk_ids_for_sort)->()"); |
| 108 | |
| 109 | m.def( |
| 110 | "moe_prepare_scatter(Tensor topk_ids, Tensor token_expert_indices, " |
| 111 | "Tensor? expert_map, int n_expert, int n_local_expert, " |
| 112 | "Tensor! expert_first_token_offset, Tensor! scatter_idx, " |
| 113 | "Tensor! sort_workspace, Tensor! sorted_experts, Tensor! sorted_rows, " |
| 114 | "Tensor! topk_ids_for_sort)->()"); |
| 115 | |
| 116 | m.def( |
| 117 | "moe_unpermute(Tensor permuted_hidden_states, Tensor topk_weights," |
| 118 | "Tensor inv_permuted_idx, Tensor? expert_first_token_offset, " |
| 119 | "int topk, Tensor! hidden_states)->()"); |
| 120 | |
| 121 | m.def("moe_permute_unpermute_supported() -> bool"); |
| 122 | m.def( |
| 123 | "moe_permute_sort_workspace_size(int num_expanded_rows, int n_expert) -> " |
| 124 | "int"); |
| 125 | |
| 126 | // Row shuffle for MoE |
| 127 | m.def( |
| 128 | "shuffle_rows(Tensor input_tensor, Tensor dst2src_map, Tensor! " |
| 129 | "output_tensor) -> ()"); |
| 130 | |
| 131 | // Apply grouped topk routing to select experts. |
| 132 | m.def( |
| 133 | "grouped_topk(Tensor scores, int n_group, int " |
| 134 | "topk_group, int topk, bool renormalize, float " |
| 135 | "routed_scaling_factor, Tensor bias, int scoring_func) -> (Tensor, " |
| 136 | "Tensor)"); |
| 137 | #endif |
| 138 | } |
| 139 | |
| 140 | STABLE_TORCH_LIBRARY_IMPL(_moe_C, CUDA, m) { |
| 141 | m.impl("topk_softmax", TORCH_BOX(&topk_softmax)); |
| 142 | m.impl("topk_sigmoid", TORCH_BOX(&topk_sigmoid)); |
| 143 | m.impl("topk_softplus_sqrt", TORCH_BOX(&topk_softplus_sqrt)); |
| 144 | m.impl("moe_sum", TORCH_BOX(&moe_sum)); |
| 145 | m.impl("moe_align_block_size", TORCH_BOX(&moe_align_block_size)); |
| 146 | m.impl("batched_moe_align_block_size", |
| 147 | TORCH_BOX(&batched_moe_align_block_size)); |
| 148 | m.impl("moe_lora_align_block_size", TORCH_BOX(&moe_lora_align_block_size)); |
| 149 | #ifndef USE_ROCM |
| 150 | m.impl("moe_wna16_gemm", TORCH_BOX(&moe_wna16_gemm)); |
| 151 | m.impl("shuffle_rows", TORCH_BOX(&shuffle_rows)); |
| 152 | m.impl("grouped_topk", TORCH_BOX(&grouped_topk)); |
| 153 | #endif |
| 154 | } |
| 155 | |
| 156 | #ifndef USE_ROCM |
| 157 | // Primitive-only ops have no tensor to dispatch on. |
| 158 | STABLE_TORCH_LIBRARY_IMPL(_moe_C, CompositeExplicitAutograd, m) { |
| 159 | m.impl("moe_permute_unpermute_supported", |
| 160 | TORCH_BOX(&moe_permute_unpermute_supported)); |
| 161 | m.impl("moe_permute_sort_workspace_size", |
| 162 | TORCH_BOX(&moe_permute_sort_workspace_size)); |
| 163 | } |
| 164 | #endif |
| 165 | |
| 166 | REGISTER_EXTENSION(_moe_C_stable_libtorch) |