cassdlcm/mediapipe
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

No public description

Browse Source 
PiperOrigin-RevId: 581469194
This commit is contained in:
MediaPipe Team 2023-11-10 23:57:18 -08:00 committed by Copybara-Service
parent ad4da8c9cc
commit 939a9c2a37
3 changed files with 0 additions and 116 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
mediapipe/tasks/cc/components/processors/proto/BUILD
View File

@ -93,14 +93,3 @@ mediapipe_proto_library(
"//mediapipe/framework:calculator_proto", "//mediapipe/framework:calculator_proto",
], ],
) )
mediapipe_proto_library(
name = "transformer_params_proto",
srcs = ["transformer_params.proto"],
)
mediapipe_proto_library(
name = "llm_params_proto",
srcs = ["llm_params.proto"],
deps = [":transformer_params_proto"],
)

41
mediapipe/tasks/cc/components/processors/proto/llm_params.proto
View File

@ -1,41 +0,0 @@
/* Copyright 2023 The MediaPipe 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.
==============================================================================*/
syntax = "proto3";
package mediapipe.tasks.components.processors.proto;
import "mediapipe/tasks/cc/components/processors/proto/transformer_params.proto";
option java_package = "com.google.mediapipe.tasks.components.processors.proto";
option java_outer_classname = "LLMParametersProto";
// Parameters for Large Language Models (LLM).
message LLMParameters {
TransformerParameters transformer_parameters = 1;
// Size of vocabulary.
int32 vocab_size = 2;
// Whether or not to disable KV cache, which is also referred as state
// somewhere else.
bool disable_kv_cache = 3;
// Id of the start token.
int32 start_token_id = 4;
// Token to determine the end of output stream.
string stop_token = 5;
}

64
mediapipe/tasks/cc/components/processors/proto/transformer_params.proto
View File

@ -1,64 +0,0 @@
/* Copyright 2023 The MediaPipe 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.
==============================================================================*/
syntax = "proto3";
package mediapipe.tasks.components.processors.proto;
option java_package = "com.google.mediapipe.tasks.components.processors.proto";
option java_outer_classname = "TransformerParametersProto";
// The parameters of transformer (https://arxiv.org/pdf/1706.03762.pdf)
message TransformerParameters {
// Batch size of tensors.
int32 batch_size = 1;
// Maximum sequence length of the input/output tensor.
int32 max_seq_length = 2;
// Embedding dimension (or model dimension), `d_model` in the paper.
// `d_k` == `d_v` == `d_model`/`h`.
int32 embedding_dim = 3;
// Hidden dimension used in the feedforward layer, `d_ff` in the paper.
int32 hidden_dimension = 4;
// Head dimension, `d_k` or `d_v` in the paper.
int32 head_dimension = 5;
// Number of heads, `h` in the paper.
int32 num_heads = 6;
// Number of stacked transformers, `N` in the paper.
int32 num_stacks = 7;
// Deprecated: bool use_mqa. Use num_kv_heads below.
reserved 8;
// Number of kv heads. 0 means Multi-Head-Attention (MHA), key and value have
// same number of heads as query; 1 means Multi-Query-Attention (MQA), key and
// value have one head; otherwise, this specifies the number of heads for key
// and value, and Grouped-Query-Attention (GQA) will be used. See
// https://arxiv.org/pdf/2305.13245.pdf for details.
int32 num_kv_heads = 9;
// Different types of attention mask type.
enum AttentionMaskType {
UNSPECIFIED = 0;
CAUSAL = 1;
PREFIX = 2;
}
AttentionMaskType attention_mask_type = 10;
}