Hugging Face's logo

ZeroWw
/
gemma-2-9b-SBS

English
Model card Files
xet
 Community
gemma-2-9b-SBS
File size: 6,472 Bytes
055eba4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
// Copyright 2024 Google LLC
// SPDX-License-Identifier: Apache-2.0
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#ifndef THIRD_PARTY_GEMMA_CPP_GEMMA_TEST_UTIL_H_
#define THIRD_PARTY_GEMMA_CPP_GEMMA_TEST_UTIL_H_

#include <stddef.h>

#include <array>
#include <complex>

#include "gtest/gtest.h"
#include "gemma/weights_raw.h"

namespace gcpp {

template<typename T, typename U, size_t kLen>
void Complexify(const std::array<T, kLen>& x,
                std::array<std::complex<U>, kLen>& c_x) {
  for (size_t i = 0; i < kLen; ++i) {
    c_x[i] = std::complex<U>(x[i], 0.0);
  }
}


template<typename T, typename U, typename TConfig>
void Complexify(const Layer<T, TConfig>& w,
                Layer<std::complex<U>, TConfig>& c_w) {
  Complexify(w.pre_attention_norm_scale, c_w.pre_attention_norm_scale);
  Complexify(w.attn_vec_einsum_w, c_w.attn_vec_einsum_w);
  Complexify(w.qkv_einsum_w, c_w.qkv_einsum_w);
  Complexify(w.pre_ffw_norm_scale, c_w.pre_ffw_norm_scale);
  Complexify(w.gating_einsum_w, c_w.gating_einsum_w);
  Complexify(w.linear_w, c_w.linear_w);
}

template<typename T, typename U, typename TConfig>
void Complexify(const Weights<T, TConfig>& w,
                Weights<std::complex<U>, TConfig>& c_w) {
  static constexpr size_t kLayers = TConfig::kLayers;
  Complexify(w.embedder_input_embedding, c_w.embedder_input_embedding);
  Complexify(w.final_norm_scale, c_w.final_norm_scale);
  for (size_t i = 0; i < kLayers; ++i) {
    Complexify(*w.GetLayer(i), *c_w.GetLayer(i));
  }
}

template<typename T, typename U, size_t N>
void TestNear(const std::array<T, N>& actual, const std::array<U, N>& expected,
              double max_abs_err, double max_rel_err, int line) {
  double sum0 = 0;
  double sum1 = 0;
  double sum01 = 0;
  for (size_t i = 0; i < N; ++i) {
    sum0 += actual[i] * actual[i];
    sum1 += expected[i] * expected[i];
    sum01 += actual[i] * expected[i];
    ASSERT_NEAR(actual[i], expected[i],
                std::max(max_abs_err, std::abs(expected[i]) * max_rel_err))
        << "line: " << line << " dim=" << N << " i=" << i;
  }
  if (sum0 > 1e-40) {
    double norm_dot = sum01 / std::sqrt(sum0) / std::sqrt(sum1);
    ASSERT_NEAR(norm_dot, 1.0, 1e-7)
        << "line: " << line << " sum0: " << sum0  << " sum1: " << sum1
        << " sum01: " << sum01;
  }
}

// Compute gradient with the finite difference method in the complex plane.
// If f : R->R is the tested function and F : C->C is its extension on the
// complex plane so that F is complex differentiable in x, then
//
//   F(x + ih) = F(x) + ih F'(x) + O(h^2) F''(x)
//
// which means that
//
//   F'(x) ~= Imag(F(x + ih)) / h
//
// This method is more numerically stable than the real-valued finite difference
// method since we don't need to subtract floating point numbers that are near
// to each other.
template<typename T, typename U, size_t N, typename FUNC>
void TestGradient(const std::array<T, N>& grad,
                  std::array<std::complex<U>, N>& x, FUNC func,
                  U step, T max_abs_err, T max_rel_err, int line) {
  std::array<T, N> exp_grad;
  const U inv_step = 1.0 / step;
  for (size_t i = 0; i < N; ++i) {
    const U x0 = std::real(x[i]);
    const std::complex<U> x1 = std::complex<U>(x0, step);
    x[i] = x1;
    const std::complex<U> f1 = func();
    exp_grad [i] = std::imag(f1) * inv_step;
    x[i] = x0;
  }
  TestNear(grad, exp_grad, max_abs_err, max_rel_err, line);
}

template<size_t N, typename FUNC>
void TestGradient(const std::array<float, N>& grad,
                  std::array<std::complex<float>, N>& x, FUNC func,
                  float max_abs_err, float max_rel_error, int line) {
  TestGradient(grad, x, func, 1e-30f, max_abs_err, max_rel_error, line);
}

template<size_t N, typename FUNC>
void TestGradient(const std::array<float, N>& grad,
                  std::array<std::complex<double>, N>& x, FUNC func,
                  float max_abs_err, float max_rel_error, int line) {
  TestGradient(grad, x, func, 1e-50, max_abs_err, max_rel_error, line);
}

template<size_t N, typename FUNC>
void TestGradient(const std::array<double, N>& grad,
                  std::array<std::complex<double>, N>& x, FUNC func,
                  double max_abs_err, double max_rel_error, int line) {
  TestGradient(grad, x, func, 1e-50, max_abs_err, max_rel_error, line);
}

template<typename T, typename U, typename TConfig, typename FUNC>
void TestGradient(const Layer<T, TConfig>& grad,
                  Layer<std::complex<U>, TConfig>& c_weights,
                  FUNC func, T max_err) {
  TestGradient(grad.pre_attention_norm_scale,
               c_weights.pre_attention_norm_scale,
               func, max_err, max_err, __LINE__);
  TestGradient(grad.attn_vec_einsum_w, c_weights.attn_vec_einsum_w,
               func, max_err, max_err, __LINE__);
  TestGradient(grad.qkv_einsum_w, c_weights.qkv_einsum_w,
               func, max_err, max_err, __LINE__);
  TestGradient(grad.pre_ffw_norm_scale, c_weights.pre_ffw_norm_scale,
               func, max_err, max_err, __LINE__);
  TestGradient(grad.gating_einsum_w, c_weights.gating_einsum_w,
               func, max_err, max_err, __LINE__);
  TestGradient(grad.linear_w, c_weights.linear_w,
               func, max_err, max_err, __LINE__);
}

template<typename T, typename U, typename TConfig, typename FUNC>
void TestGradient(const Weights<T, TConfig>& grad,
                  Weights<std::complex<U>, TConfig>& c_weights,
                  FUNC func, T max_err) {
  TestGradient(grad.embedder_input_embedding,
                 c_weights.embedder_input_embedding,
                 func,  2 * max_err, max_err, __LINE__);
  TestGradient(grad.final_norm_scale, c_weights.final_norm_scale,
               func, max_err, max_err, __LINE__);
  for (int i = 0; i < TConfig::kLayers; ++i) {
    TestGradient(*grad.GetLayer(i), *c_weights.GetLayer(i), func, max_err);
  }
}

}  // namespace gcpp

#endif  // THIRD_PARTY_GEMMA_CPP_GEMMA_TEST_UTIL_H_