sglang_v0.5.2/pytorch_2.8.0/third_party/fbgemm/test/EmbeddingSpMDM8BitTest.cc

563 lines
21 KiB
C++

/*
* Copyright (c) Meta Platforms, Inc. and affiliates.
* All rights reserved.
*
* This source code is licensed under the BSD-style license found in the
* LICENSE file in the root directory of this source tree.
*/
#include <algorithm>
#include <numeric>
#include <ostream>
#include <random>
#include <stdexcept>
#include <gtest/gtest.h>
#include "./EmbeddingSpMDMTestUtils.h"
#include "fbgemm/Fbgemm.h"
#include "src/RefImplementations.h"
using namespace std;
using namespace fbgemm;
static vector<vector<int>> GetInputs_() {
vector<vector<int>> input_dims = {
// batch size, number of rows of table, emb dim , avg lengthl
{1, 8, 8, 4},
{2, 8, 16, 4},
{10, 4000, 32, 100},
{100, 4000, 32, 100},
{10, 4000, 64, 100},
{10, 4000, 128, 100},
{4, 400, 256, 10},
{10, 4000, 48, 100},
{10, 4000, 40, 100},
{10, 4000, 56, 100},
{10, 4000, 1, 100},
{10, 4000, 4, 100},
// These were from C2 tests
{10, 40, 16, 10},
{10, 40, 85, 10},
{10, 40, 8, 10},
{10, 40, 96, 10},
{10, 40, 163, 10},
};
return input_dims;
}
vector<int> prefetch_distances{0, 16, 1000000};
namespace {
class Fused8BitRowwiseEmbeddingLookupTest
: public testing::TestWithParam<tuple<
int,
EmbeddingSpMDMWeightChoice,
EmbeddingSpMDMCornerCase,
EmbeddingSpMDMOutputDtypeChoice>> {};
}; // namespace
INSTANTIATE_TEST_CASE_P(
InstantiationName,
Fused8BitRowwiseEmbeddingLookupTest,
::testing::Combine(
::testing::ValuesIn(prefetch_distances),
::testing::Values(
UNWEIGHTED,
WEIGHTED,
POSITIONAL_WEIGHTED), // use_weight
::testing::Values(
NONE,
EMPTY_INDICES,
OUT_OF_BOUND_INDICES,
UNMATCHED_NUM_INDICES_AND_LENGTHS_SUM),
::testing::Values(FLOAT, FLOAT16, BFLOAT16)));
TEST_P(Fused8BitRowwiseEmbeddingLookupTest, basicTest) {
vector<vector<int>> inputs(GetInputs_());
random_device r;
default_random_engine generator(r());
uniform_int_distribution<> bool_dist(0, 1);
bool isIndex64b = bool_dist(generator);
bool isOffset64b = bool_dist(generator);
bool normalize_by_lengths = bool_dist(generator);
bool use_offsets = bool_dist(generator);
bool scale_bias_last = bool_dist(generator);
int prefetch;
EmbeddingSpMDMWeightChoice weight_choice;
EmbeddingSpMDMCornerCase corner_case;
EmbeddingSpMDMOutputDtypeChoice out_type;
tie(prefetch, weight_choice, corner_case, out_type) = GetParam();
bool is_wt_positional = weight_choice == POSITIONAL_WEIGHTED;
bool use_weight = weight_choice != UNWEIGHTED;
if (corner_case != NONE || weight_choice == POSITIONAL_WEIGHTED) {
// Check corner case only for subset of tests.
if (normalize_by_lengths || out_type != FLOAT || !scale_bias_last) {
return;
}
}
if (is_wt_positional && !use_weight) {
// weight positional only makes sense when use_weight is true
return;
}
for (auto input : inputs) {
int batch_size = input[0];
int num_rows = input[1];
int embedding_dim = input[2];
int average_len = input[3];
// Create embedding table
normal_distribution<float> embedding_distribution;
uniform_int_distribution<int> entries(0, 16);
int fused_embedding_dim =
embedding_dim + 2 * (scale_bias_last ? sizeof(float) : sizeof(float16));
vector<uint8_t> fused_embedding_table(num_rows * fused_embedding_dim);
for (int i = 0; i < num_rows; i++) {
for (int ii = 0; ii < embedding_dim; ii++) {
fused_embedding_table
[i * fused_embedding_dim + ii +
(scale_bias_last ? 0 : 2 * sizeof(float16))] = entries(generator);
}
float* scale_bias = reinterpret_cast<float*>(
fused_embedding_table.data() + i * fused_embedding_dim +
(scale_bias_last ? embedding_dim : 0));
if (scale_bias_last) {
scale_bias[0] = embedding_distribution(generator);
scale_bias[1] = embedding_distribution(generator);
} else {
reinterpret_cast<float16*>(scale_bias)[0] =
cpu_float2half_rn(embedding_distribution(generator));
reinterpret_cast<float16*>(scale_bias)[1] =
cpu_float2half_rn(embedding_distribution(generator));
}
}
vector<int64_t> lengths, offsets, indices;
vector<int32_t> lengths_32, offsets_32, indices_32;
vector<float> weights;
int lengths_sum = GenerateLengthsIndicesWeights(
lengths,
lengths_32,
offsets,
offsets_32,
indices,
indices_32,
weights,
batch_size,
num_rows,
average_len,
corner_case);
const int64_t* offsets_or_lengths =
(use_offsets ? offsets : lengths).data();
const int32_t* offsets_or_lengths_32 =
(use_offsets ? offsets_32 : lengths_32).data();
if (!scale_bias_last && use_weight) {
// When scale_bias_last == false, assume this is for table batched
// embedding (TBE) that can get -1 for pruned rows.
uniform_int_distribution<int> pruned_indices_distribution(
0, indices.size() - 1);
constexpr float PRUNED_INDICES_PROPORTION = 0.1;
for (int i = 0; i < indices.size() * PRUNED_INDICES_PROPORTION; ++i) {
auto idx = pruned_indices_distribution(generator);
indices[idx] = -1;
indices_32[idx] = -1;
}
}
// Sentries at the end to make sure masking is done correctly not to write
// out of bounds.
constexpr int num_sentries = 10;
const float sentry_value = 1.0f;
int output_size_wo_sentries = batch_size * embedding_dim;
vector<float> output_ref(output_size_wo_sentries + num_sentries);
vector<float> output(output_ref.size());
vector<uint16_t> output_ref_16b(output.size()), output_16b(output.size());
for (size_t i = output_size_wo_sentries; i < output.size(); ++i) {
output_ref[i] = sentry_value;
output[i] = sentry_value;
output_ref_16b[i] =
convert_from_float_ref<uint16_t>(sentry_value, out_type == BFLOAT16);
output_16b[i] =
convert_from_float_ref<uint16_t>(sentry_value, out_type == BFLOAT16);
}
bool success, success_ref;
#define TEST_BASE( \
indices, \
offsets_or_lengths, \
output_ref, \
output, \
IndexType, \
OffsetType, \
OutType) \
success_ref = EmbeddingSpMDM_ref<uint8_t>( \
embedding_dim, \
batch_size, \
lengths_sum, \
num_rows, \
fused_embedding_table.data(), \
corner_case == EMPTY_INDICES ? nullptr : indices.data(), \
offsets_or_lengths, \
use_weight ? weights.data() : nullptr, \
normalize_by_lengths, \
output_ref.data(), \
is_wt_positional, \
use_offsets, \
/*output_stride=*/-1, \
/*input_stride=*/-1, \
scale_bias_last, \
/*is_bf16_out=*/out_type == BFLOAT16, \
/*is_bf16_in=*/false); \
\
auto kernel = GenerateEmbeddingSpMDMWithStrides< \
uint8_t, \
IndexType, \
OffsetType, \
OutType>( \
embedding_dim, \
use_weight, \
normalize_by_lengths, \
prefetch, \
is_wt_positional, \
use_offsets, \
/*output_stride=*/-1, \
/*input_stride=*/-1, \
scale_bias_last, \
/*is_bf16_out=*/out_type == BFLOAT16, \
/*is_bf16_in=*/false); \
success = kernel( \
batch_size, \
lengths_sum, \
num_rows, \
fused_embedding_table.data(), \
corner_case == EMPTY_INDICES ? nullptr : indices.data(), \
offsets_or_lengths, \
use_weight ? weights.data() : nullptr, \
output.data());
#define TEST_OUT_TYPE(indices, offsets_or_lengths, IndexType, OffsetType) \
if (out_type == FLOAT) { \
TEST_BASE( \
indices, \
offsets_or_lengths, \
output_ref, \
output, \
IndexType, \
OffsetType, \
float); \
} else { \
TEST_BASE( \
indices, \
offsets_or_lengths, \
output_ref_16b, \
output_16b, \
IndexType, \
OffsetType, \
float16); \
}
#define TEST_OFFSET_TYPE(indices, IndexType) \
if (isOffset64b) { \
TEST_OUT_TYPE(indices, offsets_or_lengths, IndexType, int64_t); \
} else { \
TEST_OUT_TYPE(indices, offsets_or_lengths_32, IndexType, int32_t); \
}
if (isIndex64b) {
TEST_OFFSET_TYPE(indices, int64_t);
} else {
TEST_OFFSET_TYPE(indices_32, int32_t);
}
#undef TEST_OFFSET_TYPE
#undef TEST_OUT_TYPE
#undef TEST_BASE
// Check correctness
EXPECT_EQ(success, success_ref)
<< "Reference and JIT impl did not both succeed";
if (corner_case == OUT_OF_BOUND_INDICES ||
corner_case == UNMATCHED_NUM_INDICES_AND_LENGTHS_SUM) {
EXPECT_EQ(success, false);
}
if (success) {
for (size_t i = 0; i < output.size(); ++i) {
float actual = (out_type == FLOAT)
? output[i]
: convert_to_float_ref(output_16b[i], out_type == BFLOAT16);
float expected = (out_type == FLOAT)
? output_ref[i]
: convert_to_float_ref(output_ref_16b[i], out_type == BFLOAT16);
EXPECT_EQ(actual, expected)
<< "results differ at (" << i << ") reference: " << expected
<< ", FBGEMM: " << actual << " emb dim :" << embedding_dim;
}
for (int offset = output_size_wo_sentries;
offset < output_size_wo_sentries + num_sentries;
++offset) {
float actual = (out_type == FLOAT)
? output[offset]
: convert_to_float_ref(output_16b[offset], out_type == BFLOAT16);
float expected = (out_type == FLOAT)
? output_ref[offset]
: convert_to_float_ref(
output_ref_16b[offset], out_type == BFLOAT16);
EXPECT_EQ(actual, expected)
<< "results differ at (" << offset << ") reference: " << expected
<< ", FBGEMM: " << actual << " emb dim :" << embedding_dim;
}
}
} // end for input
}
TEST_P(Fused8BitRowwiseEmbeddingLookupTest, rowwiseSparseTest) {
vector<vector<int>> inputs(GetInputs_());
random_device r;
default_random_engine generator(r());
uniform_int_distribution<> bool_dist(0, 1);
bool isIndex64b = bool_dist(generator);
bool isOffset64b = bool_dist(generator);
bool normalize_by_lengths = bool_dist(generator);
bool use_offsets = bool_dist(generator);
bool scale_bias_last = bool_dist(generator);
int prefetch;
EmbeddingSpMDMWeightChoice weight_choice;
EmbeddingSpMDMCornerCase corner_case;
EmbeddingSpMDMDtypeChoice out_type;
tie(prefetch, weight_choice, corner_case, out_type) = GetParam();
bool is_wt_positional = weight_choice == POSITIONAL_WEIGHTED;
bool use_weight = weight_choice != UNWEIGHTED;
if (out_type != FLOAT || !scale_bias_last) {
return;
}
constexpr float sparsity = 0.7;
for (auto input : inputs) {
int batch_size = input[0];
int num_rows = input[1];
int embedding_dim = input[2];
int average_len = input[3];
// Create mapping table for rowwise sparsity
vector<int32_t> mapping_table;
int num_compressed_rows =
CreateMappingTableForRowWiseSparsity(mapping_table, num_rows, sparsity);
// Create embedding table
normal_distribution<float> embedding_distribution;
uniform_int_distribution<int> entries(0, 16);
int fused_embedding_dim = embedding_dim + 2 * sizeof(float);
vector<uint8_t> fused_embedding_table(
num_compressed_rows * fused_embedding_dim);
for (int i = 0; i < num_compressed_rows; i++) {
for (int ii = 0; ii < embedding_dim; ii++) {
fused_embedding_table[i * fused_embedding_dim + ii] =
entries(generator);
}
float* scale_bias = reinterpret_cast<float*>(
fused_embedding_table.data() + i * fused_embedding_dim +
embedding_dim);
scale_bias[0] = embedding_distribution(generator);
scale_bias[1] = embedding_distribution(generator);
}
vector<int64_t> lengths, offsets, indices;
vector<int32_t> lengths_32, offsets_32, indices_32;
vector<float> weights;
int lengths_sum = GenerateLengthsIndicesWeights(
lengths,
lengths_32,
offsets,
offsets_32,
indices,
indices_32,
weights,
batch_size,
num_rows,
average_len,
corner_case);
const int64_t* offsets_or_lengths =
(use_offsets ? offsets : lengths).data();
const int32_t* offsets_or_lengths_32 =
(use_offsets ? offsets_32 : lengths_32).data();
vector<float> output_sls_ref(batch_size * embedding_dim);
vector<float> output_slws_ref(output_sls_ref.size()),
output_sls(output_sls_ref.size()), output_slws(output_sls_ref.size());
vector<float>& output_ref = use_weight ? output_slws_ref : output_sls_ref;
vector<float>& output = use_weight ? output_slws : output_sls;
bool success, success_ref;
if (isOffset64b) {
if (isIndex64b) {
success_ref = EmbeddingSpMDMRowWiseSparse_ref<uint8_t, int64_t>(
embedding_dim,
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
corner_case == EMPTY_INDICES ? nullptr : indices.data(),
mapping_table.data(),
offsets_or_lengths,
use_weight ? weights.data() : nullptr,
normalize_by_lengths,
output_ref.data(),
is_wt_positional,
use_offsets);
auto kernel =
GenerateEmbeddingSpMDMRowWiseSparse<uint8_t, int64_t, int64_t>(
embedding_dim,
use_weight,
normalize_by_lengths,
prefetch,
is_wt_positional,
use_offsets);
success = kernel(
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
corner_case == EMPTY_INDICES ? nullptr : indices.data(),
offsets_or_lengths,
use_weight ? weights.data() : nullptr,
output.data(),
mapping_table.data());
} else {
success_ref = EmbeddingSpMDMRowWiseSparse_ref<uint8_t, int32_t>(
embedding_dim,
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
corner_case == EMPTY_INDICES ? nullptr : indices_32.data(),
mapping_table.data(),
offsets_or_lengths,
use_weight ? weights.data() : nullptr,
normalize_by_lengths,
output_ref.data(),
is_wt_positional,
use_offsets);
auto kernel =
GenerateEmbeddingSpMDMRowWiseSparse<uint8_t, int32_t, int64_t>(
embedding_dim,
use_weight,
normalize_by_lengths,
prefetch,
is_wt_positional,
use_offsets);
success = kernel(
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
corner_case == EMPTY_INDICES ? nullptr : indices_32.data(),
offsets_or_lengths,
use_weight ? weights.data() : nullptr,
output.data(),
mapping_table.data());
}
} else {
if (isIndex64b) {
success_ref = EmbeddingSpMDMRowWiseSparse_ref<uint8_t, int64_t>(
embedding_dim,
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
corner_case == EMPTY_INDICES ? nullptr : indices.data(),
mapping_table.data(),
offsets_or_lengths,
use_weight ? weights.data() : nullptr,
normalize_by_lengths,
output_ref.data(),
is_wt_positional,
use_offsets);
auto kernel = GenerateEmbeddingSpMDMRowWiseSparse<uint8_t, int64_t>(
embedding_dim,
use_weight,
normalize_by_lengths,
prefetch,
is_wt_positional,
use_offsets);
success = kernel(
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
corner_case == EMPTY_INDICES ? nullptr : indices.data(),
offsets_or_lengths_32,
use_weight ? weights.data() : nullptr,
output.data(),
mapping_table.data());
} else {
success_ref = EmbeddingSpMDMRowWiseSparse_ref<uint8_t, int32_t>(
embedding_dim,
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
corner_case == EMPTY_INDICES ? nullptr : indices_32.data(),
mapping_table.data(),
offsets_or_lengths,
use_weight ? weights.data() : nullptr,
normalize_by_lengths,
output_ref.data(),
is_wt_positional,
use_offsets);
auto kernel = GenerateEmbeddingSpMDMRowWiseSparse<uint8_t, int32_t>(
embedding_dim,
use_weight,
normalize_by_lengths,
prefetch,
is_wt_positional,
use_offsets);
success = kernel(
batch_size,
lengths_sum,
num_rows,
fused_embedding_table.data(),
corner_case == EMPTY_INDICES ? nullptr : indices_32.data(),
offsets_or_lengths_32,
use_weight ? weights.data() : nullptr,
output.data(),
mapping_table.data());
}
}
// Check correctness
EXPECT_EQ(success, success_ref)
<< "Reference and JIT impl did not both succeed";
if (corner_case == OUT_OF_BOUND_INDICES ||
corner_case == UNMATCHED_NUM_INDICES_AND_LENGTHS_SUM) {
EXPECT_EQ(success, false);
}
if (success) {
for (size_t i = 0; i < output.size(); ++i) {
EXPECT_EQ(output[i], output_ref[i])
<< "results differ at (" << i << ") reference: " << output_ref[i]
<< ", FBGEMM: " << output[i] << " emb dim :" << embedding_dim;
}
}
} // end for input
}