Spaces:
Running
Running
File size: 5,625 Bytes
5f923cd | 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 | // Copyright 2025 The ODML Authors.
//
// 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
//
// http://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.
#include "runtime/util/executor_data_util.h"
#include <cstddef>
#include <optional>
#include <utility>
#include <vector>
#include "absl/status/status.h" // from @com_google_absl
#include "absl/status/statusor.h" // from @com_google_absl
#include "litert/cc/litert_layout.h" // from @litert
#include "litert/cc/litert_macros.h" // from @litert
#include "litert/cc/litert_ranked_tensor_type.h" // from @litert
#include "litert/cc/litert_tensor_buffer.h" // from @litert
#include "runtime/executor/llm_executor_io_types.h"
#include "runtime/util/status_macros.h" // IWYU pragma: keep
#include "runtime/util/tensor_buffer_util.h"
namespace litert::lm {
namespace {
template <typename T>
absl::StatusOr<T> CombineExecutorDataImpl(std::vector<T>& executor_data) {
if (executor_data.empty()) {
return absl::InvalidArgumentError("Executor data is empty.");
}
if (executor_data.size() == 1) {
// If there is only one image, we can just move it to the combined image
// data.
return std::move(executor_data[0]);
}
// If there are multiple executor data, we need to first combine them into a
// TensorBuffer, then create a single ExecutorVisionData from the
// TensorBuffer.
int num_executor_data = executor_data.size();
ASSIGN_OR_RETURN(const auto* first_tensor,
executor_data[0].GetEmbeddingsPtr());
LITERT_ASSIGN_OR_RETURN(auto first_tensor_type, first_tensor->TensorType());
auto first_tensor_dims = TensorBufferDims(*first_tensor);
int total_token_num = 0;
int total_packed_size = 0;
std::vector<int> combined_token_num;
for (const auto& executor_data : executor_data) {
ASSIGN_OR_RETURN(const auto* embeddings_ptr,
executor_data.GetEmbeddingsPtr());
auto dims = TensorBufferDims(*embeddings_ptr);
if (dims.size() != 3 && dims.size() != 4) {
return absl::InvalidArgumentError(
"The embedding tensor type must have 3 or 4 dimensions.");
}
combined_token_num.push_back(dims[dims.size() - 2]);
total_token_num += dims[dims.size() - 2];
LITERT_ASSIGN_OR_RETURN(size_t packed_size, embeddings_ptr->PackedSize());
total_packed_size += packed_size;
}
Layout combined_layout;
if constexpr (std::is_same_v<T, ExecutorAudioData>) {
combined_layout = Layout(Dimensions(
{first_tensor_dims[0], total_token_num, first_tensor_dims[2]}));
} else if (first_tensor_dims.size() == 3) {
combined_layout = Layout(Dimensions(
{first_tensor_dims[0], 1, total_token_num, first_tensor_dims[2]}));
} else if (first_tensor_dims.size() == 4) {
combined_layout =
Layout(Dimensions({first_tensor_dims[0], first_tensor_dims[1],
total_token_num, first_tensor_dims[3]}));
}
::litert::RankedTensorType combined_tensor_type(
first_tensor_type.ElementType(), std::move(combined_layout));
LITERT_ASSIGN_OR_RETURN(auto combined_tensor_buffer,
TensorBuffer::CreateManagedHostMemory(
combined_tensor_type, total_packed_size));
LITERT_ASSIGN_OR_RETURN(
auto combined_embeddings_lock_and_addr,
::litert::TensorBufferScopedLock::Create(combined_tensor_buffer,
TensorBuffer::LockMode::kWrite));
char* combined_tensor_buffer_ptr =
static_cast<char*>(combined_embeddings_lock_and_addr.second);
for (int i = 0; i < num_executor_data; ++i) {
ASSIGN_OR_RETURN(auto embeddings_ptr,
executor_data[i].GetMutableEmbeddingsPtr());
LITERT_ASSIGN_OR_RETURN(auto embeddings_size, embeddings_ptr->PackedSize());
LITERT_ASSIGN_OR_RETURN(
auto embeddings_lock_and_addr,
::litert::TensorBufferScopedLock::Create(
*embeddings_ptr, TensorBuffer::LockMode::kRead));
memcpy(combined_tensor_buffer_ptr, embeddings_lock_and_addr.second,
embeddings_size);
combined_tensor_buffer_ptr += embeddings_size;
}
if constexpr (std::is_same_v<T, ExecutorVisionData>) {
return ExecutorVisionData(std::move(combined_tensor_buffer),
/*per_layer_embeddings=*/std::nullopt);
} else if constexpr (std::is_same_v<T, ExecutorAudioData>) {
int num_audio_tokens = 0;
for (const auto& executor_data : executor_data) {
num_audio_tokens += executor_data.GetValidTokens();
}
return ExecutorAudioData(std::move(combined_tensor_buffer),
/*per_layer_embeddings=*/std::nullopt,
num_audio_tokens);
} else {
return absl::InvalidArgumentError("Executor data type is not supported.");
}
}
} // namespace
absl::StatusOr<ExecutorVisionData> CombineExecutorVisionData(
std::vector<ExecutorVisionData>& executor_data) {
return CombineExecutorDataImpl(executor_data);
}
absl::StatusOr<ExecutorAudioData> CombineExecutorAudioData(
std::vector<ExecutorAudioData>& executor_data) {
return CombineExecutorDataImpl(executor_data);
}
} // namespace litert::lm
|