| 1 | /* |
| 2 | * Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved. |
| 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
| 16 | |
| 17 | #pragma once |
| 18 | |
| 19 | #include <cuda_bf16.h> |
| 20 | #include <cuda_fp16.h> |
| 21 | |
| 22 | #include <torch/headeronly/core/ScalarType.h> |
| 23 | |
| 24 | namespace vllm { |
| 25 | namespace tensorrt_llm { |
| 26 | |
| 27 | template <typename DType> |
| 28 | struct ElemsPerAccess; |
| 29 | |
| 30 | template <> |
| 31 | struct ElemsPerAccess<half> { |
| 32 | static constexpr int value = 8; |
| 33 | using vec_type = float4; |
| 34 | }; |
| 35 | |
| 36 | template <> |
| 37 | struct ElemsPerAccess<nv_bfloat16> { |
| 38 | static constexpr int value = 8; |
| 39 | using vec_type = float4; |
| 40 | }; |
| 41 | |
| 42 | template <> |
| 43 | struct ElemsPerAccess<float> { |
| 44 | static constexpr int value = 4; |
| 45 | using vec_type = float4; |
| 46 | }; |
| 47 | |
| 48 | template <typename DType> |
| 49 | static constexpr int kElemsPerAccess = ElemsPerAccess<DType>::value; |
| 50 | |
| 51 | struct MiniMaxReduceRMSParams { |
| 52 | int nranks{}; |
| 53 | int rank{}; |
| 54 | torch::headeronly::ScalarType dtype{torch::headeronly::ScalarType::Undefined}; |
| 55 | int size_q{}; |
| 56 | int hidden_dim{}; |
| 57 | int size_k{}; |
| 58 | int hidden_dim_k{}; |
| 59 | int stride_q{}; // row stride for q input (elements); when > hidden_dim, |
| 60 | // q is part of a wider qkv tensor |
| 61 | int stride_k{}; // row stride for k input (elements); when > hidden_dim_k, |
| 62 | // k is part of a wider qkv tensor |
| 63 | int stride_q_out{}; // row stride for q output (elements); 0 = contiguous |
| 64 | int stride_k_out{}; // row stride for k output (elements); 0 = contiguous |
| 65 | void** workspace{}; |
| 66 | void* allreduce_in{}; |
| 67 | void* rms_norm_out{}; |
| 68 | void* rms_gamma{}; |
| 69 | void* allreduce_in_k{}; |
| 70 | void* rms_norm_out_k{}; |
| 71 | void* rms_gamma_k{}; |
| 72 | float rms_eps{}; |
| 73 | cudaStream_t stream{}; |
| 74 | }; |
| 75 | |
| 76 | void minimax_reduce_rms_op(MiniMaxReduceRMSParams const& params); |
| 77 | |
| 78 | } // namespace tensorrt_llm |
| 79 | } // namespace vllm |