rmsnorm.cpp

/*
 * SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: MIT
 */

#include <catch2/catch_test_macros.hpp>
#include "../utils/helpers.h"

#include <cudnn_frontend.h>

TEST_CASE("RmsNorm Training", "[rmsnorm][graph]") {
    namespace fe = cudnn_frontend;
    fe::graph::Graph graph;
    graph.set_intermediate_data_type(fe::DataType_t::FLOAT).set_compute_data_type(fe::DataType_t::FLOAT);

    auto batch_size  = 4;
    auto seq_length  = 1024;
    auto hidden_size = 128;

    auto X     = graph.tensor(fe::graph::Tensor_attributes()
                              .set_name("X")
                              .set_data_type(fe::DataType_t::FLOAT)
                              .set_dim({batch_size, seq_length, hidden_size})
                              .set_stride({seq_length * hidden_size, hidden_size, 1}));
    auto scale = graph.tensor(fe::graph::Tensor_attributes()
                                  .set_name("scale")
                                  .set_dim({1, 1, hidden_size})
                                  .set_stride({hidden_size, hidden_size, 1})
                                  .set_data_type(fe::DataType_t::FLOAT));

    float epsilon_cpu = 1e-05f;
    auto epsilon      = graph.tensor(epsilon_cpu);

    auto rmsnorm_options =
        fe::graph::Rmsnorm_attributes().set_forward_phase(fe::NormFwdPhase_t::TRAINING).set_epsilon(epsilon);
    auto [Y, inv_variance] = graph.rmsnorm(X, scale, rmsnorm_options);
    Y->set_output(true).set_data_type(fe::DataType_t::FLOAT);
    inv_variance->set_output(true).set_data_type(fe::DataType_t::FLOAT);

#if (CUDNN_VERSION < 8906)
    SKIP("RmsNorm is not supported in cudnn versions prior to 8.9.6");
#endif
    if (check_device_arch_newer_than("ampere") == false) {
        SKIP("RMSNorm requires Ampere and up");
    }
    // Create a unique_ptr for the cuDNN handle
    auto handle_ptr = create_cudnn_handle();
    auto handle     = *handle_ptr;

    REQUIRE(graph.validate().is_good());

    REQUIRE(graph.build_operation_graph(handle).is_good());

    REQUIRE(graph.create_execution_plans({fe::HeurMode_t::A, fe::HeurMode_t::FALLBACK}).is_good());

    REQUIRE(graph.check_support().is_good());

    REQUIRE(graph.build_plans().is_good());

    REQUIRE(graph.warmup(handle).is_good());

    Surface<float> X_tensor(batch_size * seq_length * hidden_size);
    Surface<float> Inv_variance_tensor(batch_size * seq_length);
    Surface<float> Scale_tensor(hidden_size);
    Surface<float> Y_tensor(batch_size * seq_length * hidden_size);

    int64_t workspace_size = 0;
    REQUIRE(graph.get_workspace_size(workspace_size).is_good());
    Surface<int8_t> workspace(workspace_size);

    std::unordered_map<std::shared_ptr<fe::graph::Tensor_attributes>, void*> variant_pack = {
        {X, X_tensor.devPtr},
        {inv_variance, Inv_variance_tensor.devPtr},
        {scale, Scale_tensor.devPtr},
        {Y, Y_tensor.devPtr}};

    REQUIRE(graph.execute(handle, variant_pack, workspace.devPtr).is_good());
}

TEST_CASE("RmsNorm Inference", "[rmsnorm][graph]") {
    namespace fe = cudnn_frontend;
    fe::graph::Graph graph;
    graph.set_intermediate_data_type(fe::DataType_t::FLOAT).set_compute_data_type(fe::DataType_t::FLOAT);

    auto batch_size  = 4;
    auto seq_length  = 1024;
    auto hidden_size = 128;

    auto X     = graph.tensor(fe::graph::Tensor_attributes()
                              .set_name("X")
                              .set_data_type(fe::DataType_t::FLOAT)
                              .set_dim({batch_size, seq_length, hidden_size})
                              .set_stride({seq_length * hidden_size, hidden_size, 1}));
    auto scale = graph.tensor(fe::graph::Tensor_attributes()
                                  .set_name("scale")
                                  .set_dim({1, 1, hidden_size})
                                  .set_stride({hidden_size, hidden_size, 1})
                                  .set_data_type(fe::DataType_t::FLOAT));
    auto bias  = graph.tensor(fe::graph::Tensor_attributes()
                                 .set_name("bias")
                                 .set_dim({1, 1, hidden_size})
                                 .set_stride({hidden_size, hidden_size, 1})
                                 .set_data_type(fe::DataType_t::FLOAT));

    float epsilon_cpu = 1e-05f;
    auto epsilon      = graph.tensor(epsilon_cpu);

    auto rmsnorm_options = fe::graph::Rmsnorm_attributes()
                               .set_forward_phase(fe::NormFwdPhase_t::INFERENCE)
                               .set_epsilon(epsilon)
                               .set_bias(bias);
    auto [Y, inv_variance] = graph.rmsnorm(X, scale, rmsnorm_options);
    Y->set_output(true).set_data_type(fe::DataType_t::FLOAT);
    REQUIRE(inv_variance == nullptr);

#if (CUDNN_VERSION < 8906)
    SKIP("RmsNorm is not supported in cudnn versions prior to 8.9.6");
#endif
    if (check_device_arch_newer_than("ampere") == false) {
        SKIP("RmsNorm requires Ampere and up");
    }
    // Create a unique_ptr for the cuDNN handle
    auto handle_ptr = create_cudnn_handle();
    auto handle     = *handle_ptr;

    REQUIRE(graph.validate().is_good());

    REQUIRE(graph.build_operation_graph(handle).is_good());

    REQUIRE(graph.create_execution_plans({fe::HeurMode_t::A, fe::HeurMode_t::FALLBACK}).is_good());

    REQUIRE(graph.check_support().is_good());

    REQUIRE(graph.build_plans().is_good());

    Surface<float> X_tensor(batch_size * seq_length * hidden_size);
    Surface<float> Scale_tensor(hidden_size);
    Surface<float> Bias_tensor(hidden_size);
    Surface<float> Y_tensor(batch_size * seq_length * hidden_size);

    int64_t workspace_size = 0;
    REQUIRE(graph.get_workspace_size(workspace_size).is_good());
    Surface<int8_t> workspace(workspace_size);

    std::unordered_map<std::shared_ptr<fe::graph::Tensor_attributes>, void*> variant_pack = {
        {X, X_tensor.devPtr}, {scale, Scale_tensor.devPtr}, {bias, Bias_tensor.devPtr}, {Y, Y_tensor.devPtr}};

    REQUIRE(graph.execute(handle, variant_pack, workspace.devPtr).is_good());
}

TEST_CASE("RmsNorm Backward", "[rmsnorm][graph]") {
    namespace fe = cudnn_frontend;
    fe::graph::Graph graph;
    graph.set_intermediate_data_type(fe::DataType_t::FLOAT).set_compute_data_type(fe::DataType_t::FLOAT);

    auto batch_size  = 4;
    auto seq_length  = 1024;
    auto hidden_size = 128;

    auto X  = graph.tensor(fe::graph::Tensor_attributes()
                              .set_name("X")
                              .set_data_type(fe::DataType_t::FLOAT)
                              .set_dim({batch_size, seq_length, hidden_size})
                              .set_stride({seq_length * hidden_size, hidden_size, 1}));
    auto DY = graph.tensor(fe::graph::Tensor_attributes()
                               .set_name("DY")
                               .set_data_type(fe::DataType_t::FLOAT)
                               .set_dim({batch_size, seq_length, hidden_size})
                               .set_stride({seq_length * hidden_size, hidden_size, 1}));

    auto scale        = graph.tensor(fe::graph::Tensor_attributes()
                                  .set_name("scale")
                                  .set_dim({1, 1, hidden_size})
                                  .set_stride({hidden_size, hidden_size, 1})
                                  .set_data_type(fe::DataType_t::FLOAT));
    auto inv_variance = graph.tensor(fe::graph::Tensor_attributes()
                                         .set_name("inv_variance")
                                         .set_dim({batch_size, seq_length, 1})
                                         .set_stride({seq_length, 1, 1})
                                         .set_data_type(fe::DataType_t::FLOAT));

    auto DRMS_options        = fe::graph::Rmsnorm_backward_attributes().has_dbias(false);
    auto [DX, dscale, dbias] = graph.rmsnorm_backward(DY, X, scale, inv_variance, DRMS_options);
    DX->set_output(true).set_data_type(fe::DataType_t::FLOAT);
    dscale->set_output(true).set_data_type(fe::DataType_t::FLOAT);
    REQUIRE(dbias == nullptr);

#if (CUDNN_VERSION < 8906)
    SKIP("RmsNorm is not supported in cudnn versions prior to 8.9.6");
#endif
    if (check_device_arch_newer_than("ampere") == false) {
        SKIP("RmsNorm Backward requires Ampere and up");
    }
    // Create a unique_ptr for the cuDNN handle
    auto handle_ptr = create_cudnn_handle();
    auto handle     = *handle_ptr;

    REQUIRE(graph.validate().is_good());

    REQUIRE(graph.build_operation_graph(handle).is_good());

    REQUIRE(graph.create_execution_plans({fe::HeurMode_t::A, fe::HeurMode_t::FALLBACK}).is_good());

    REQUIRE(graph.check_support().is_good());

    REQUIRE(graph.build_plans().is_good());

    Surface<float> X_tensor(batch_size * seq_length * hidden_size);
    Surface<float> DY_tensor(batch_size * seq_length * hidden_size);
    Surface<float> Inv_variance_tensor(batch_size * seq_length);
    Surface<float> Scale_tensor(hidden_size);
    Surface<float> Dscale_tensor(hidden_size);
    Surface<float> DX_tensor(batch_size * seq_length * hidden_size);

    int64_t workspace_size = 0;
    REQUIRE(graph.get_workspace_size(workspace_size).is_good());
    Surface<int8_t> workspace(workspace_size);

    std::unordered_map<std::shared_ptr<fe::graph::Tensor_attributes>, void*> variant_pack = {
        {X, X_tensor.devPtr},
        {DY, DY_tensor.devPtr},
        {inv_variance, Inv_variance_tensor.devPtr},
        {scale, Scale_tensor.devPtr},
        {dscale, Dscale_tensor.devPtr},
        {DX, DX_tensor.devPtr}};

    REQUIRE(graph.execute(handle, variant_pack, workspace.devPtr).is_good());
}