biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Kernel] moe wna16 marlin kernel (#14447)

Browse Source 
Signed-off-by: Jinzhen Lin <linjinzhen@hotmail.com>
Co-authored-by: Michael Goin <michael@neuralmagic.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Jinzhen Lin
2025-04-15 11:05:22 +08:00
committed by GitHub
parent 6b40996ae8
commit d06ba4ed3f
16 changed files with 3477 additions and 329 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
csrc/moe/torch_bindings.cpp
View File

@@ -43,14 +43,17 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
m.impl("moe_wna16_gemm", torch::kCUDA, &moe_wna16_gemm);
m.def(
"marlin_gemm_moe(Tensor! a, Tensor! b_q_weights, Tensor! sorted_ids, "
"Tensor! topk_weights, Tensor! topk_ids, Tensor! b_scales, Tensor! "
"b_zeros, Tensor! g_idx, Tensor! perm, Tensor! workspace, "
"int b_q_type, SymInt size_m, "
"SymInt size_n, SymInt size_k, bool is_k_full, int num_experts, int "
"topk, "
"int moe_block_size, bool replicate_input, bool apply_weights)"
" -> Tensor");
"moe_wna16_marlin_gemm(Tensor! a, Tensor? c_or_none,"
"Tensor! b_q_weight, Tensor! b_scales, Tensor? b_zeros_or_none,"
"Tensor? g_idx_or_none, Tensor? perm_or_none, Tensor! workspace,"
"Tensor sorted_token_ids,"
"Tensor! expert_ids, Tensor! num_tokens_past_padded,"
"Tensor! topk_weights, int moe_block_size, int top_k, "
"bool mul_topk_weights, bool is_ep, int b_q_type_id,"
"int size_m, int size_n, int size_k,"
"bool is_full_k, bool use_atomic_add,"
"bool use_fp32_reduce, bool is_zp_float) -> Tensor");
// conditionally compiled so impl registration is in source file
#endif