[Kernel] Porting the TRTLLM minimax_allreduce_rms kernels (#37045)

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
Jee Jee Li
2026-04-11 00:20:20 +08:00
committed by GitHub
parent 8f121f7879
commit ecd1ea1363
14 changed files with 1861 additions and 4 deletions
+12
View File
@@ -308,4 +308,16 @@ int64_t qr_max_size();
#ifndef USE_ROCM
void dsv3_fused_a_gemm(torch::Tensor& output, torch::Tensor const& mat_a,
torch::Tensor const& mat_b);
#endif
#ifndef USE_ROCM
torch::Tensor minimax_allreduce_rms(torch::Tensor const& input,
torch::Tensor const& norm_weight,
torch::Tensor workspace, int64_t const rank,
int64_t const nranks, double const eps);
std::tuple<torch::Tensor, torch::Tensor> minimax_allreduce_rms_qk(
torch::Tensor qkv, torch::Tensor const& norm_weight_q,
torch::Tensor const& norm_weight_k, torch::Tensor workspace,
int64_t const q_size, int64_t const kv_size, int64_t const rank,
int64_t const nranks, double const eps);
#endif