fix(sgl-kernel): sm90 compile flashmla failed (#24130)
This commit is contained in:
@@ -36,11 +36,13 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
|
||||
"-> Tensor[]");
|
||||
m.impl("fwd_kvcache_mla", torch::kCUDA, &fwd_kvcache_mla);
|
||||
|
||||
#ifdef FLASHMLA_ENABLE_SM100
|
||||
m.def(
|
||||
"dense_prefill_fwd(Tensor workspace_buffer, Tensor q, Tensor k, Tensor v, Tensor cumulative_seqlen_q, Tensor "
|
||||
"cumulative_seqlen_kv, Tensor o, Tensor lse, int mask_mode_code, float softmax_scale, int max_seqlen_q, int "
|
||||
"max_seqlen_kv, bool is_varlen) -> ()");
|
||||
m.impl("dense_prefill_fwd", torch::kCUDA, &FMHACutlassSM100FwdRun);
|
||||
#endif
|
||||
|
||||
m.def("sparse_prefill_fwd(Tensor q, Tensor kv, Tensor indices, float sm_scale, int d_v) -> Tensor[]");
|
||||
m.impl("sparse_prefill_fwd", torch::kCUDA, &sparse_prefill_fwd);
|
||||
|
||||
Reference in New Issue
Block a user