package tzrec.protos
Get desktop application:
View/edit binary Protocol Buffers messages
message
metric.proto:4Used in: ,
optional uint32 thresholds = 1
message
metric.proto:25Used in: ,
optional float threshold = 1
probs threshold
optional uint32 top_k = 2
top k accuracy when num_class > 1
message
optimizer.proto:190Used in: ,
required float lr = 1
optional float rho = 2
optional float eps = 3
optional float weight_decay = 4
message
optimizer.proto:162Used in: ,
required float lr = 1
optional float weight_decay = 2
optional float initial_accumulator_value = 3
optional float eps = 4
optional bool fused = 5
message
optimizer.proto:170Used in: ,
required float lr = 1
optional float beta1 = 2
optional float beta2 = 3
optional float weight_decay = 4
optional float eps = 5
optional bool amsgrad = 6
optional bool fused = 7
message
optimizer.proto:180Used in: ,
required float lr = 1
optional float beta1 = 2
optional float beta2 = 3
optional float weight_decay = 4
optional float eps = 5
optional bool amsgrad = 6
optional bool fused = 7
message
feature.proto:49Used in: , , , , , , ,
required uint32 num_channels = 2
number of embedding channels
optional float temperature = 3
temperature coefficient for softmax
optional float keep_prob = 4
message
module.proto:39Used in:
optional uint32 max_k = 1
max number of high capsules Default: 5
required uint32 max_seq_len = 2
max behaviour sequence length
required uint32 high_dim = 3
high capsule embedding vector dimension
optional uint32 num_iters = 4
dynamic routing iterations, Default: 3
optional float routing_logits_scale = 5
routing logits scale Default: 20
optional float routing_logits_stddev = 6
routing logits initial stddev Default: 1
optional float squash_pow = 7
squash power Default: 1
optional bool const_caps_num = 8
whether to use constant capsule number, Default: false
optional string routing_init_method = 9
the initialization method for routing logits, Default: normal, available: zeros.
message
tower.proto:59Used in:
required string tower_name = 1
task name for the task tower
optional string label_name = 2
label for the task, default is label_fields by order
repeated metrics = 3
metrics for the task
repeated train_metrics = 20
log train merics for task
repeated losses = 4
loss for the task
optional uint32 num_class = 5
num_class for multi-class classification loss
optional mlp = 6
task specific mlp
optional float weight = 7
training loss weights
repeated string relation_tower_names = 8
related tower names
optional relation_mlp = 9
relation mlp
optional string sample_weight_name = 10
sample weight for the task
optional string task_space_indicator_label = 11
label name for indicating the sample space for the task tower
optional float in_task_space_weight = 12
the loss weight for sample in the task space
optional float out_task_space_weight = 13
the loss weight for sample out the task space
optional float pareto_min_loss_weight = 14
use pareto front minimal loss weight, ge 0 and lt 1
message
loss.proto:14Used in:
optional float label_smoothing = 1
message
loss.proto:30Used in:
optional float gamma = 1
optional float alpha = 2
message
feature.proto:907Used in: ,
required string feature_name = 1
feature name.
repeated string expression = 2
feature input, e.g. item:item_id
optional string embedding_name = 3
embedding name, feature with same embedding name will share embedding
required uint32 embedding_dim = 4
embedding dimension
optional uint64 hash_bucket_size = 5
number of hash size
optional uint64 num_buckets = 6
number of id enumerators
repeated string vocab_list = 7
id vocabulary list
map<string, uint64> vocab_dict = 8
id vocabulary dict
optional uint32 value_dim = 9
id value dimensions, default = 0, when use in seq, default = 1 if value_dim = 0, it supports id with multi-value
optional string pooling = 10
embedding pooling type, available is {sum | mean}
optional string default_value = 11
fg default value, default value before bucktize
optional string separator = 12
fg multi-value separator
optional string init_fn = 14
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional bool use_mask = 15
mask value in training progress
optional zch = 16
zero collision hash
optional string vocab_file = 17
id vocabulary file path
optional string asset_dir = 18
vocab file relative directory
optional dynamicemb = 19
dynamic embedding
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional uint64 default_bucketize_value = 31
out-of-vocab(OOV) id bucketize value when use vocab_list or vocab_dict when use default_bucketize_value, we will not add additional bucketize_value of `default_value`=0, bucketize_value of <OOV>=1 into vocab_list or vocab_dict
optional string fg_value_type = 32
value_type after fg before bucketize, you can specify it for better performance. e.g. fg_value_type = int64 when use num_buckets
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
module.proto:94Used in:
repeated uint32 cin_layer_size = 1
every layer size
message
feature.proto:307Used in: ,
required string feature_name = 1
feature name, e.g. tag_feat
required string expression = 2
feature input, e.g. user:tag
optional string embedding_name = 3
embedding name, feature with same embedding name will share embedding
optional uint32 embedding_dim = 4
embedding dimension
map<string, float> value_map = 6
value map for mapping input string values to float values
repeated float boundaries = 7
boundaries for bucktize numeric combine value
optional uint64 num_buckets = 8
number of id enumerators for sparse combine value
optional string pooling = 11
embedding pooling type, available is {sum | mean}. Controls embedding bag aggregation. NOTE: distinct from 'combiner' which controls FG-level multi-value aggregation.
optional string default_value = 12
fg default value, default value before bucktize
optional string separator = 13
fg multi-value separator
optional string normalizer = 14
fg normalizer, e.g. method=log10,threshold=1e-10,default=-10 method=zscore,mean=0.0,standard_deviation=10.0 method=minmax,min=2.1,max=2.2 method=expression,expr=sign(x)
optional string init_fn = 15
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional bool use_mask = 16
mask value in training progress
optional string combiner = 22
combine value combiner type, available is {sum | mean | min | max} Controls FG-level multi-value aggregation before bucketization.
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
oneof dense_emb
autodis = 40
autodis embedding
mlp = 41
mlp embedding
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
feature.proto:245Used in: ,
required string feature_name = 1
feature name, e.g. os_and_cate
repeated string expression = 2
feature input, e.g. [user:os, item:cate]
optional string embedding_name = 3
embedding name, feature with same embedding name will share embedding
optional uint32 embedding_dim = 4
embedding dimension
optional uint64 hash_bucket_size = 5
number of hash size
repeated string vocab_list = 7
id vocabulary list
map<string, uint64> vocab_dict = 8
id vocabulary dict
optional uint32 value_dim = 9
id value dimensions, if value_dim = 0, it supports id with multi-value
optional string pooling = 10
embedding pooling type, available is {sum | mean}
optional string default_value = 11
fg default value
optional string separator = 12
fg multi-value separator
optional string init_fn = 13
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional bool use_mask = 14
mask value in training progress
optional zch = 15
zero collision hash
optional string vocab_file = 16
id vocabulary file path
optional string asset_dir = 17
vocab file relative directory
optional dynamicemb = 18
dynamic embedding
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional uint64 default_bucketize_value = 31
out-of-vocab(OOV) id bucketize value when use vocab_list or vocab_dict when use default_bucketize_value, we will not add additional bucketize_value of `default_value`=0, bucketize_value of <OOV>=1 into vocab_list or vocab_dict
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
optimizer.proto:204Used in: , ,
(message has no fields)
message
optimizer.proto:236Used in: , ,
optional uint32 T_max = 1
total number of steps or epochs for cosine annealing
optional float min_learning_rate = 2
minimum learning rate
optional float warmup_learning_rate = 3
warmup start learning rate
optional uint32 warmup_size = 4
warmup steps or epochs
optional bool by_epoch = 5
schedule by epoch or by step.
message
optimizer.proto:249Used in: , ,
optional uint32 T_0 = 1
number of steps or epochs for the first cosine annealing period
optional uint32 T_mult = 2
factor to grow period length after each restart (1 = fixed period)
optional float min_learning_rate = 3
minimum learning rate
optional float warmup_learning_rate = 4
warmup start learning rate
optional uint32 warmup_size = 5
warmup steps or epochs
optional bool by_epoch = 6
schedule by epoch or by step.
message
module.proto:82Used in:
optional uint32 cross_num = 1
number of cross layers
message
module.proto:87Used in:
optional uint32 cross_num = 1
number of cross layers
optional uint32 low_rank = 2
Matrix decomposition with minimal rank.
message
feature.proto:821Used in: ,
required string feature_name = 1
feature name.
required string operator_name = 2
custom operator name.
required string operator_lib_file = 3
custom operator lib file name.
optional operator_params = 4
operator custom params.
optional bool is_op_thread_safe = 5
custom operator is thread safe or not.
repeated string expression = 6
feature input, e.g. user:os
optional string embedding_name = 7
embedding name, feature with same embedding name will share embedding
optional uint32 embedding_dim = 8
embedding dimension
repeated float boundaries = 9
boundaries for bucktize numeric value
optional uint64 hash_bucket_size = 10
number of hash size for sparse value
optional uint64 num_buckets = 11
number of id enumerators for sparse value
repeated string vocab_list = 12
id vocabulary list for sparse value
map<string, uint64> vocab_dict = 13
id vocabulary dict
optional string pooling = 14
embedding pooling type, available is {sum | mean}
optional string default_value = 15
fg default value
optional string separator = 16
fg multi-value separator
optional string normalizer = 17
fg normalizer, e.g. method=log10,threshold=1e-10,default=-10 method=zscore,mean=0.0,standard_deviation=10.0 method=minmax,min=2.1,max=2.2 method=expression,expr=sign(x)
optional string init_fn = 18
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional uint32 value_dim = 19
value dimensions
optional bool use_mask = 20
mask value in training progress
optional zch = 21
zero collision hash
optional string vocab_file = 22
id vocabulary file path
optional string asset_dir = 23
vocab file relative directory
optional dynamicemb = 24
dynamic embedding
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional uint64 default_bucketize_value = 31
out-of-vocab(OOV) id bucketize value when use vocab_list or vocab_dict when use default_bucketize_value, we will not add additional bucketize_value of `default_value`=0, bucketize_value of <OOV>=1 into vocab_list or vocab_dict
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
oneof dense_emb
autodis = 40
autodis embedding
mlp = 41
mlp embedding
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
match_model.proto:56Used in:
required user_tower = 1
required item_tower = 2
required int32 output_dim = 3
user and item tower output dimension
optional Similarity similarity = 4
similarity method
optional float temperature = 5
similarity scaling factor
optional bool in_batch_negative = 6
use in batch items as negative items.
required float amm_i_weight = 7
loss weight for amm_i
required float amm_u_weight = 8
loss weight for amm_u
message
tower.proto:164Used in:
required string input = 1
input feature group name
required string augment_input = 2
augmented feature group name
required mlp = 3
mlp config
message
multi_task_rank.proto:22Used in:
optional mask_net = 6
shared bottom MaskNet module
optional bottom_mlp = 1
shared bottom mlp layer
optional expert_mlp = 2
mmoe expert mlp layer definition
optional gate_mlp = 3
mmoe gate module definition
optional uint32 num_expert = 4
number of mmoe experts
repeated task_towers = 5
bayes task tower
message
multi_task_rank.proto:37Used in:
optional bottom_mlp = 1
shared bottom mlp layer
optional expert_mlp = 2
mmoe expert mlp layer definition
optional gate_mlp = 3
mmoe gate module definition
optional uint32 num_expert = 4
number of mmoe experts
repeated task_towers = 5
task tower
message
rank_model.proto:50Used in:
required cross = 1
required deep = 2
required final = 3
message
rank_model.proto:57Used in:
optional backbone = 1
required cross = 2
optional deep = 3
required final = 4
message
seq_encoder.proto:6Used in:
optional string name = 1
seq encoder name
required string input = 2
sequence feature name
required attn_mlp = 3
mlp config for target attention score
optional int32 max_seq_length = 6
maximum sequence length
message
tower.proto:23Used in:
required string input = 1
input feature group name
required attn_mlp = 2
mlp config for target attention score
message
rank_model.proto:36Used in:
optional dense_mlp = 1
if has dense feature group,must has dense_mlp
optional bool arch_with_sparse = 2
whether to include sparse features after interaction
required final = 3
message
match_model.proto:9Used in:
required user_tower = 1
required item_tower = 2
required int32 output_dim = 3
user and item tower output dimension
optional Similarity similarity = 4
similarity method
optional float temperature = 5
similarity scaling factor
optional bool in_batch_negative = 6
use in batch items as negative items.
message
match_model.proto:37Used in:
required user_tower = 1
required item_tower = 2
required int32 output_dim = 3
user and item tower output dimension
optional Similarity similarity = 4
similarity method
optional float temperature = 5
similarity scaling factor
optional bool in_batch_negative = 6
use in batch items as negative items.
message
data.proto:69Used in:
optional uint32 batch_size = 1
mini batch size to use for training and evaluation.
required DatasetType dataset_type = 2
dataset type.
optional bool fg_encoded = 3
[deprecated] please use fg_mode. input data is feature generate encoded or not. if fg_encoded = true, you should do fg offline first, and set fg_encoded_multival_sep for split multi-val feature
optional string fg_encoded_multival_sep = 4
separator for multi-val feature in fg encoded input data
repeated string label_fields = 5
labels
optional uint32 num_workers = 6
number of workers for parallel processing raw data
optional bool pin_memory = 7
pin memory for fast cudaMemCopy
repeated input_fields = 8
the input fields must be the same number and in the same order as data in csv files
optional string delimiter = 9
delimiter of column features, only used for CsvDataset
optional bool with_header = 10
for csv files, with header or not.
optional uint32 eval_batch_size = 11
mini batch size to use for and evaluation.
optional bool drop_remainder = 12
drop last batch less than batch_size
optional uint32 fg_threads = 13
fg threads for each worker, if fg_threads = 0, will disable fg dag handler, use python run.
optional bool is_orderby_partition = 14
when use OdpsDataset, read data orderby table partitions or not.
optional string odps_data_quota_name = 15
maxcompute storage api & tunnel quota name
optional float sample_mask_prob = 16
mask probability for samples in training progress
optional float negative_sample_mask_prob = 17
mask probability for sampled negatives in training progress
optional bool force_base_data_group = 18
force padding data into same data group with same batch_size
repeated string sample_weight_fields = 19
sample weights
optional FgMode fg_mode = 20
fg run mode.
optional bool shuffle = 22
whether to shuffle data
optional uint32 shuffle_buffer_size = 23
shufffle buffer for better performance, even shuffle buffer is set, it is suggested to do full data shuffle before training especially when the performance of models is not good.
optional string odps_data_compression = 24
maxcompute storage api data compression type, LZ4_FRAME | ZSTD | UNCOMPRESSED
optional string sample_cost_field = 25
sample cost field name
optional uint64 batch_cost_size = 26
batch cost limit size
optional string input_fields_str = 27
simplified input fields string format: input_name1:input_type1;input_name2:input_type2; type names follow ODPS conventions with aliases: BIGINT->INT64, INT->INT32
oneof sampler
negative sampler
negative_sampler = 101
negative_sampler_v2 = 102
hard_negative_sampler = 103
hard_negative_sampler_v2 = 104
tdm_sampler = 105
Used in:
OdpsDataset = 1
ParquetDataset = 2
CsvDataset = 3
OdpsDatasetV1 = 4
KafkaDataset = 5
message
rank_model.proto:16Used in:
required deep = 1
optional final = 2
optional uint32 wide_embedding_dim = 3
optional string wide_init_fn = 4
wide embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
message
train.proto:32Used in:
optional uint32 dump_interval_steps = 1
MC/ZCH features are not supported; use dynamicemb for delta dump. dump touched ids and their latest embedding every N training steps. Larger intervals retain a longer id window in memory; auto compaction reduces per-batch tensor buildup but unique ids still scale with the interval.
optional string output_dir = 2
output directory. default is ${model_dir}/delta_embedding_dump
optional string file_prefix = 3
parquet file prefix
message
optimizer.proto:26Used in:
oneof optimizer
sgd_optimizer = 1
adagrad_optimizer = 2
adam_optimizer = 3
adamw_optimizer = 4
adadelta_optimizer = 5
rmsprop_optimizer = 6
oneof learning_rate
constant_learning_rate = 101
exponential_decay_learning_rate = 102
manual_step_learning_rate = 103
cosine_annealing_learning_rate = 104
cosine_annealing_warm_restarts_learning_rate = 105
repeated part_optimizers = 201
message
feature.proto:26DistanceLFU: evict_score = access_cnt / pow((current_iter - last_access_iter), decay_exponent)
Used in:
optional float decay_exponent = 1
decay rate is access step
message
multi_task_rank.proto:60Used in:
required hstu = 1
hstu config
required fusion_mtl_tower = 2
multi task tower config
required uint32 max_seq_len = 3
max sequence length
optional uint32 item_embedding_hidden_dim = 4
item embedding mlp hidden dimension
optional bool enable_global_average_loss = 5
enables loss averaging computation globally across all ranks (total rank) instead of locally (local rank).
optional bool sequence_timestamp_is_ascending = 6
timestamp of sequence is ascending or descending
optional bool concat_contextual_features = 7
concat all contextual features on channel dim as one token
message
feature.proto:79Used in:
required uint64 threshold = 1
minimum frequency threshold for admission
optional initializer_args = 2
determine how to initialize the embedding if the key is not admitted.
optional uint64 counter_capacity = 3
kv counter capacity, if not set, use embedding max_capacity
optional uint64 counter_bucket_capacity = 4
kv counter capacity for each bucket.
message
feature.proto:61Used in: ,
optional string mode = 1
the mode of initialization. NORMAL | TRUNCATED_NORMAL | UNIFORM | CONSTANT
optional float mean = 2
the mean value for (truncated) normal distributions
optional float std_dev = 3
the standard deviation for (truncated) normal distributions. default is sqrt(1 / embedding_dim)
optional float lower = 4
the lower bound for uniform/truncated_normal distribution. default is -sqrt(1 / max_capacity)
optional float upper = 5
the upper bound for uniform/truncated_normal distribution. default is sqrt(1 / max_capacity)
optional float value = 6
the constant value for constant initialization.
message
feature.proto:90Used in: , , , , ,
optional initializer_args = 1
arguments for initializing dynamic embedding vector values. default is uniform distribution, and absolute values of upper and lower bound are sqrt(1 / embedding_dim).
optional eval_initializer_args = 2
the initializer args for evaluation mode. default is default is constant initialization with value 0.0.
optional string score_strategy = 4
strategy to set the score for each indices in forward and backward per table. TIMESTAMP | STEP | CUSTOMIZED | LFU | NO_EVICTION
required uint64 max_capacity = 5
max number of embedding rows
optional float cache_load_factor = 6
percentage of embedding rows caching on gpu
optional uint64 init_capacity_per_rank = 7
init number of capacity
optional string init_table = 8
init table path
optional uint64 bucket_capacity = 9
hash-table bucket capacity. default 128 (matches dynamicemb DEFAULT_BUCKET_CAPACITY). larger buckets trade probe cost for higher load factor.
oneof admission_strategy
frequency_admission_strategy = 100
message
pipeline.proto:11required string train_input_path = 1
required string eval_input_path = 2
required string model_dir = 3
optional train_config = 4
optional eval_config = 5
optional export_config = 6
optional data_config = 7
repeated feature_configs = 8
optional model_config = 9
message
eval.proto:4Used in:
optional uint32 num_steps = 1
number of steps to evaluate.
optional uint32 log_step_count_steps = 2
the frequency progress be logged during eval
message
optimizer.proto:207Used in: , ,
optional uint32 decay_size = 1
decay steps or epochs
optional float decay_factor = 2
decay rate
optional bool staircase = 3
if true, decay the learning rate at discrete intervals
optional float warmup_learning_rate = 4
warmup start learning rate
optional uint32 warmup_size = 5
warmup steps or epochs
optional float min_learning_rate = 6
minimum learning rate
optional bool by_epoch = 7
schedule by epoch or by step.
message
export.proto:4Used in:
optional string exporter_type = 1
type of exporter [latest | best] when train_and_evaluation latest: regularly exports the serving graph and checkpoints best: export the best model according to best_exporter_metric
optional string best_exporter_metric = 2
the metric used to determine the best checkpoint
optional bool metric_larger_is_better = 3
metric value the bigger the best
optional string mixed_precision = 4
mixed precision mode for inference/export [BF16 | FP16 | ""]. The export-time AMP intent is taken verbatim from this field; if it disagrees with train_config.mixed_precision a warning is logged. When set, the dense sub-graph is wrapped in torch.autocast before torch.export so that AOT Inductor captures dtype-promoting casts as a wrap_with_autocast HOP.
optional bool cudnn_allow_tf32 = 5
whether to use torch.backends.cudnn.allow_tf32
optional bool cuda_matmul_allow_tf32 = 6
whether to use torch.backends.cuda.matmul.allow_tf32
message
feature.proto:560Used in: ,
required string feature_name = 1
feature name, e.g. kv_os_click_count
required string expression = 2
expression, e.g. sigmoid(pv/(1+click))
repeated string variables = 3
variables in expression, e,g. ["item:pv", "item:click"]
optional string embedding_name = 4
embedding name, feature with same embedding name will share embedding
optional uint32 embedding_dim = 5
embedding dimension
repeated float boundaries = 6
boundaries for bucktize numeric expr value
optional string separator = 7
fg multi-value separator
optional float fill_missing = 8
fill value when vector length mismatch, default is NaN.
optional string pooling = 10
embedding pooling type, available is {sum | mean}
optional string default_value = 11
fg default value
optional string init_fn = 12
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional bool use_mask = 13
mask value in training progress
optional uint32 value_dim = 14
if value_dim = 0, it supports multi-value
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
oneof dense_emb
autodis = 40
autodis embedding
mlp = 41
mlp embedding
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
module.proto:19Used in:
required string network_name = 1
required uint32 expert_num_per_task = 2
number of experts per task
optional uint32 share_num = 3
number of experts for share
required task_expert_net = 4
mlp network of experts per task
optional share_expert_net = 5
mlp network of experts for share
message
sid_model.proto:8Strictly-typed subset of faiss.Kmeans(D, K, **kwargs) knobs. Unset fields fall back to faiss's own defaults (so it is safe to leave partially set). ``gpu`` is intentionally omitted — the fit is CPU-only (SidRqkmeans refuses a visible CUDA device).
Used in:
optional uint32 niter = 1
optional uint32 nredo = 2
optional uint32 seed = 3
optional uint32 max_points_per_centroid = 4
optional uint32 min_points_per_centroid = 5
optional bool spherical = 6
optional bool verbose = 7
message
feature.proto:989Used in:
oneof feature
id_feature = 1
raw_feature = 2
combo_feature = 3
lookup_feature = 4
match_feature = 5
sequence_feature = 6
expr_feature = 7
overlap_feature = 8
tokenize_feature = 9
custom_feature = 10
kv_dot_product = 11
bool_mask_feature = 12
combine_feature = 13
sequence_id_feature = 101
sequence_raw_feature = 102
sequence_combo_feature = 103
sequence_lookup_feature = 104
sequence_match_feature = 105
sequence_expr_feature = 107
sequence_overlap_feature = 108
sequence_tokenize_feature = 109
sequence_custom_feature = 110
sequence_kv_dot_product = 111
sequence_bool_mask_feature = 112
sequence_combine_feature = 113
message
model.proto:29Used in:
required string group_name = 1
repeated string feature_names = 2
required FeatureGroupType group_type = 3
repeated sequence_groups = 4
repeated sequence_encoders = 5
optional string embedding_name_suffix = 6
Suffix appended to each feature's embedding_name so groups with different suffixes use independent embedding tables. Empty == disabled.
Used in:
DEEP = 0
WIDE = 1
SEQUENCE = 2
JAGGED_SEQUENCE = 3
enum FgMode
data.proto:47Used in:
FG_NONE = 1
input data is feature generate encoded, we do not do fg
FG_NORMAL = 2
input data is raw feature, we use python to run feature generate
FG_DAG = 3
input data is raw feature, we use fg_handler to run feature generate
FG_BUCKETIZE = 4
input data is after feature generate but before do bucketize, we do bucketize only
message
data.proto:62Used in:
required string input_name = 1
optional FieldType input_type = 2
only need specify it when use CsvDataset and value dtype can not be inferred (all values in the column are null)
Used in:
INT32 = 0
INT64 = 1
STRING = 2
FLOAT = 3
DOUBLE = 4
ARRAY_INT32 = 5
ARRAY_INT64 = 6
ARRAY_STRING = 7
ARRAY_FLOAT = 8
ARRAY_DOUBLE = 9
ARRAY_ARRAY_INT32 = 10
ARRAY_ARRAY_INT64 = 11
ARRAY_ARRAY_STRING = 12
ARRAY_ARRAY_FLOAT = 13
ARRAY_ARRAY_DOUBLE = 14
MAP_STRING_INT32 = 15
MAP_STRING_INT64 = 16
MAP_STRING_STRING = 17
MAP_STRING_FLOAT = 18
MAP_STRING_DOUBLE = 19
MAP_INT64_INT32 = 20
MAP_INT64_INT64 = 21
MAP_INT64_STRING = 22
MAP_INT64_FLOAT = 23
MAP_INT64_DOUBLE = 24
MAP_INT32_INT32 = 25
MAP_INT32_INT64 = 26
MAP_INT32_STRING = 27
MAP_INT32_FLOAT = 28
MAP_INT32_DOUBLE = 29
message
optimizer.proto:135Used in:
required float lr = 1
optional float rho = 2
optional float eps = 3
optional float weight_decay = 4
optional bool gradient_clipping = 5
optional float max_gradient = 6
message
optimizer.proto:76Used in:
required float lr = 1
optional bool gradient_clipping = 2
optional float max_gradient = 3
optional float initial_accumulator_value = 4
message
optimizer.proto:83Used in:
required float lr = 1
optional float beta1 = 2
optional float beta2 = 3
optional float weight_decay = 4
optional bool gradient_clipping = 5
optional float max_gradient = 6
message
optimizer.proto:100Used in:
required float lr = 1
optional float beta1 = 2
optional float beta2 = 3
optional float weight_decay = 4
optional bool gradient_clipping = 5
optional float max_gradient = 6
message
optimizer.proto:92Used in:
required float lr = 1
optional float momentum = 2
optional float weight_decay = 3
optional bool gradient_clipping = 4
optional float max_gradient = 5
message
optimizer.proto:118Used in:
required float lr = 1
optional float beta1 = 2
optional float beta2 = 3
optional float weight_decay = 4
optional bool gradient_clipping = 5
optional float max_gradient = 6
message
optimizer.proto:109Used in:
required float lr = 1
optional float beta1 = 2
optional float beta2 = 3
optional float weight_decay = 4
optional bool gradient_clipping = 5
optional float max_gradient = 6
message
optimizer.proto:144Used in:
required float lr = 1
optional float alpha = 2
optional float eps = 3
optional float weight_decay = 4
optional bool gradient_clipping = 5
optional float max_gradient = 6
message
optimizer.proto:127Used in:
required float lr = 1
optional float weight_decay = 2
optional WeightDecayMode weight_decay_mode = 3
optional bool gradient_clipping = 4
optional float max_gradient = 5
message
optimizer.proto:70Used in:
required float lr = 1
optional bool gradient_clipping = 2
optional float max_gradient = 3
message
tower.proto:148Used in: ,
optional mlp = 1
task tower mlp
repeated task_configs = 2
sub task configs
message
tower.proto:128Used in:
required string task_name = 1
task name for the task
required string label_name = 2
label for the task
optional uint64 task_bitmask = 3
bitmask for get actual label for binary classification task
repeated losses = 4
loss for the task
optional uint32 num_class = 5
support multi-class classification loss
repeated metrics = 6
metrics for the task
optional float weight = 7
training loss weight for the task
repeated train_metrics = 20
log train merics for task
message
module.proto:123Used in: , ,
oneof action_encoder
simple_action_encoder = 1
message
module.proto:142Used in: ,
oneof content_encoder
slice_content_encoder = 1
slice candidate dim to uih dim
pad_content_encoder = 2
padding candidate dim to uih dim
mlp_content_encoder = 3
linear transform uih and candidate to same dim
message
module.proto:185Used in:
optional action_encoder = 3
action encoder config
optional bool enable_interleaving = 4
enable interleave target or not
optional action_mlp = 5
action embedding mlp config
required content_encoder = 6
content encoder config
required content_mlp = 7
content embedding mlp config
message
module.proto:174Used in:
optional action_encoder = 3
action encoder config
optional action_mlp = 4
action embedding mlp config
required content_encoder = 5
content encoder config
required content_mlp = 6
content embedding mlp config
message
module.proto:165Used in: , ,
oneof contextualized_mlp
simple_mlp = 10
mlp for sequence embedding
parameterized_mlp = 11
mlp for sequence and contextual embedding
message
module.proto:205Used in:
oneof input_preprocessor
contextual_preprocessor = 20
input preprocessor with contextual features
contextual_interleave_preprocessor = 21
input preprocessor with interleave targets
uih_preprocessor = 22
input preprocessor for sequence-only models (no candidate concat)
message
module.proto:216Used in:
(message has no fields)
message
module.proto:219Used in:
(message has no fields)
message
module.proto:135Used in:
required uih_mlp = 1
mlp for uih seq embedding
required target_mlp = 2
mlp for candidate seq embedding
message
module.proto:229Used in:
oneof output_postprocessor
l2norm_postprocessor = 31
l2 norm postprocessor
layernorm_postprocessor = 32
layer norm postprocessor
timestamp_layernorm_postprocessor = 33
timestamp layer norm postprocessor
message
module.proto:132Used in:
(message has no fields)
message
module.proto:158Used in:
required uint32 hidden_dim = 1
mlp hidden dimension
optional float contextual_dropout_ratio = 2
dropout ratio for contextual embedding
message
module.proto:277Used in:
required uint32 num_position_buckets = 1
buckets for position embedding
optional uint32 num_time_buckets = 2
buckets for timestamp embedding
optional bool use_time_encoding = 3
use timestamp encoding or not.
optional string time_bucket_fn = 4
transform function for timestamp gap. sqrt | log
optional float time_bucket_increments = 5
timestamp gap will div by time_bucket_increments
message
module.proto:110Used in:
optional uint32 action_embedding_dim = 1
action embedding dim
repeated uint32 action_weights = 3
bitmask of each action
repeated uint32 watchtime_to_action_thresholds = 4
thresholds for watch time to actions
repeated uint32 watchtime_to_action_weights = 5
bitmask for watch time to actions
optional float embedding_init_std = 6
action embedding weights init std
message
module.proto:153Used in:
required uint32 hidden_dim = 1
mlp hidden dimension
message
module.proto:129Used in:
(message has no fields)
message
module.proto:222Used in:
repeated uint32 time_duration_period_units = 1
time duration period units, e.g. 60 * 60 for hour of day.
repeated uint32 time_duration_units_per_period = 2
time duration units per period, e.g. 24 for hour of day.
message
module.proto:198Used in:
optional action_encoder = 1
action encoder config (optional - for models with action info)
optional action_mlp = 2
action embedding mlp config (required if action_encoder is set)
message
train.proto:21Used in:
optional string clipping_type = 1
Clipping type: "norm", "value", or "none"
optional float max_gradient = 2
Max gradient value/norm threshold
optional float norm_type = 3
Norm type for gradient norm clipping (2.0 for L2, inf for max)
optional bool enable_global_grad_clip = 4
Enable global gradient clipping for distributed training
message
train.proto:7Used in:
optional float init_scale = 1
Initial scale factor
optional float growth_factor = 2
Factor by which the scale is multiplied during update if no inf/NaN gradients occur for ``growth_interval`` consecutive iterations.
optional float backoff_factor = 3
Factor by which the scale is multiplied during update if inf/NaN gradients occur in an iteration.
optional uint32 growth_interval = 4
Number of consecutive iterations without inf/NaN gradients that must occur for the scale to be multiplied by ``growth_factor``.
message
metric.proto:32Used in:
required string grouping_key = 1
message
metric.proto:42Used in:
required string grouping_key = 1
optional uint64 max_pairs_per_group = 2
message
module.proto:290Used in: , ,
required stu = 1
stu config
optional float input_dropout_ratio = 2
dropout ratio after preprocessor
optional uint32 attn_num_layers = 3
num stu layers
required positional_encoder = 4
position encoder
required input_preprocessor = 5
input preprocessor
required output_postprocessor = 6
output postprocessor
optional uint32 attn_truncation_split_layer = 7
Attention truncation: after this many full-sequence layers, drop UIH prefix tokens to keep only the last attn_truncation_tail_len per sample. Contextual prefix and targets survive. Must be in (0, attn_num_layers); 0 disables.
optional uint32 attn_truncation_tail_len = 8
Trailing UIH cap; both fields must be > 0 to enable truncation.
optional string name = 9
MoT channel name. When non-empty, replaces the default `uih` prefix on UIH-side keys read from grouped_features (e.g. name="uih_click" -> uih_click.sequence, uih_click_action.sequence, uih_click_watchtime.sequence, uih_click_timestamp.sequence). Empty preserves the original uih.* keys (DlrmHSTU behavior). Channels with the same `embedding_name` on a feature share the underlying embedding table via EmbeddingGroup dedupe; per-channel tables multiply sparse-param + TBE + all-to-all cost by N.
message
match_model.proto:22Used in:
required user_tower = 1
required item_tower = 6
optional int32 output_dim = 7
user and item tower output dimension; when 0 (default), no output Linear is applied -- the caller must size the user tower's STU output and the item tower's MLP output to match.
optional Similarity similarity = 3
similarity method
optional float temperature = 4
similarity scaling factor
optional bool in_batch_negative = 5
use in batch items as negative items.
message
tower.proto:14Used in:
required string input = 1
input feature group name (uih group)
required hstu = 2
HSTU config (STU, positional_encoder, input_preprocessor, output_postprocessor)
required uint32 max_seq_len = 3
max sequence length
message
sampler.proto:54Weighted Random Sampling ItemID not in Batch and Sampling Hard Edge
Used in:
required string user_input_path = 1
user data path schema => userid:int64 | weight:float
required string item_input_path = 2
item data path schema => itemid:int64 | weight:float | attrs:string
required string hard_neg_edge_input_path = 3
hard negative edge path schema => userid:int64 | itemid:int64 | weight:float
required uint32 num_sample = 4
number of negative sample
required uint32 num_hard_sample = 5
max number of hard negative sample
repeated string attr_fields = 6
field names of attrs in train data or eval data
required string item_id_field = 7
field name of item_id in train data or eval data
required string user_id_field = 8
field name of user_id in train data or eval data
optional string attr_delimiter = 9
attribute delimiter of attrs string
optional uint32 num_eval_sample = 10
number of negative samples for evaluator
optional string field_delimiter = 11
only works on local
message
sampler.proto:83Weighted Random Sampling ItemID not with Edge and Sampling Hard Edge
Used in:
required string user_input_path = 1
user data path schema => userid:int64 | weight:float
required string item_input_path = 2
item data path schema => itemid:int64 | weight:float | attrs:string
required string pos_edge_input_path = 3
positive edge path schema => userid:int64 | itemid:int64 | weight:float
required string hard_neg_edge_input_path = 4
hard negative edge path schema => userid:int64 | itemid:int64 | weight:float
required uint32 num_sample = 5
number of negative sample
required uint32 num_hard_sample = 6
max number of hard negative sample
repeated string attr_fields = 7
field names of attrs in train data or eval data
required string item_id_field = 8
field name of item_id in train data or eval data
required string user_id_field = 9
field name of user_id in train data or eval data
optional string attr_delimiter = 10
attribute delimiter of attrs string
optional uint32 num_eval_sample = 11
number of negative samples for evaluator
optional string field_delimiter = 12
field delimiter of input data
message
feature.proto:116Used in: ,
required string feature_name = 1
feature name, e.g. item_id
required string expression = 2
feature input, e.g. item:item_id
optional string embedding_name = 3
embedding name, feature with same embedding name will share embedding
required uint32 embedding_dim = 4
embedding dimension
optional uint64 hash_bucket_size = 5
number of hash size
optional uint64 num_buckets = 6
number of id enumerators
repeated string vocab_list = 7
id vocabulary list
map<string, uint64> vocab_dict = 8
id vocabulary dict
optional uint32 value_dim = 9
id value dimensions, default = 0, when use in seq, default = 1 if value_dim = 0, it supports id with multi-value
optional string pooling = 10
embedding pooling type, available is {sum | mean}
optional string default_value = 11
fg default value, default value before bucktize
optional string separator = 12
fg multi-value separator
optional bool weighted = 13
fg multi-value with whether has weight
optional string init_fn = 14
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional bool use_mask = 15
mask value in training progress
optional zch = 16
zero collision hash
optional string vocab_file = 17
id vocabulary file path
optional string asset_dir = 18
vocab file relative directory
optional dynamicemb = 19
dynamic embedding
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional uint64 default_bucketize_value = 31
out-of-vocab(OOV) id bucketize value when use vocab_list or vocab_dict when use default_bucketize_value, we will not add additional bucketize_value of `default_value`=0, bucketize_value of <OOV>=1 into vocab_list or vocab_dict
optional string fg_value_type = 32
value_type after fg before bucketize, you can specify it for better performance. e.g. fg_value_type = int64 when use num_buckets
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
tower.proto:93Used in:
required string tower_name = 1
task name for the task tower
optional string label_name = 2
label for the task, default is label_fields by order
repeated metrics = 3
metrics for the task
repeated train_metrics = 20
log train merics for task
repeated losses = 4
loss for the task
optional uint32 num_class = 5
num_class for multi-class classification loss
optional mlp = 6
task specific mlp
optional float weight = 7
training loss weights
repeated string intervention_tower_names = 8
intervention tower names
required uint32 low_rank_dim = 9
low_rank_dim
optional float dropout_ratio = 10
dropout_ratio
optional string task_space_indicator_label = 11
label name for indicating the sample space for the task tower
optional float in_task_space_weight = 12
the loss weight for sample in the task space
optional float out_task_space_weight = 13
the loss weight for sample out the task space
optional float pareto_min_loss_weight = 14
use pareto front minimal loss weight, ge 0 and lt 1
message
loss.proto:25Used in:
required string session_name = 1
optional float alpha = 2
enum Kernel
model.proto:40Used in:
TRITON = 0
PYTORCH = 1
CUTLASS = 2
message
feature.proto:761Used in: ,
required string feature_name = 1
feature name, e.g. kv_os_click_count
required string query = 2
query, e.g. ["a:0.5", "b:0.5"]
required string document = 3
document, e,g. ["d:0.5", "b:0.5"]
optional string embedding_name = 4
embedding name, feature with same embedding name will share embedding
optional uint32 embedding_dim = 5
embedding dimension
repeated float boundaries = 6
boundaries for bucktize numeric expr value
optional string separator = 7
fg multi-value separator
optional float kv_delimiter = 8
fg kv separator, default is :.
optional string normalizer = 9
fg normalizer, e.g. method=log10,threshold=1e-10,default=-10 method=zscore,mean=0.0,standard_deviation=10.0 method=minmax,min=2.1,max=2.2 method=expression,expr=sign(x)
optional string pooling = 10
embedding pooling type, available is {sum | mean}
optional string default_value = 11
fg default value
optional string init_fn = 12
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional bool use_mask = 13
mask value in training progress
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
oneof dense_emb
autodis = 40
autodis embedding
mlp = 41
mlp embedding
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
loss.proto:22Used in:
(message has no fields)
message
feature.proto:16LFU: evict_score = access_cnt
Used in:
(message has no fields)
message
feature.proto:20LRU: evict_score = 1 / pow((current_iter - last_access_iter), decay_exponent)
Used in:
optional float decay_exponent = 1
decay rate is access step
message
feature.proto:371Used in: ,
required string feature_name = 1
feature name, e.g. kv_os_click_count
required string map = 2
map input, e.g. item:kv_os_click_count
required string key = 3
key input, e.g. user:os
optional string embedding_name = 4
embedding name, feature with same embedding name will share embedding
optional uint32 embedding_dim = 5
embedding dimension
repeated float boundaries = 6
boundaries for bucktize numeric lookup value
optional uint64 hash_bucket_size = 7
number of hash size for sparse lookup value
optional uint64 num_buckets = 8
number of id enumerators for sparse lookup value
repeated string vocab_list = 9
id vocabulary list for sparse lookup value
map<string, uint64> vocab_dict = 10
id vocabulary dict
optional string pooling = 11
embedding pooling type, available is {sum | mean}
optional string combiner = 12
lookup value combiner type, available is {sum | mean | min | max | count}
optional string default_value = 13
fg default value
optional string separator = 14
fg multi-value separator
optional bool need_discrete = 15
lookup map value is sparse or numeric, when need_discrete is true, combiner will be empty string
optional bool need_key = 16
lookup value need key as prefix or not.
optional string normalizer = 17
fg normalizer, e.g. method=log10,threshold=1e-10,default=-10 method=zscore,mean=0.0,standard_deviation=10.0 method=minmax,min=2.1,max=2.2 method=expression,expr=sign(x)
optional string init_fn = 18
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional uint32 value_dim = 19
lookup value dimensions
optional string value_separator = 20
numeric lookup value separator
optional bool use_mask = 21
mask value in training progress
optional zch = 22
zero collision hash
optional string vocab_file = 23
id vocabulary file path
optional string asset_dir = 24
vocab file relative directory
optional dynamicemb = 25
dynamic embedding
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional uint64 default_bucketize_value = 31
out-of-vocab(OOV) id bucketize value when use vocab_list or vocab_dict when use default_bucketize_value, we will not add additional bucketize_value of `default_value`=0, bucketize_value of <OOV>=1 into vocab_list or vocab_dict
optional string fg_value_type = 32
value_type after fg before bucketize, you can specify it for better performance. e.g. fg_value_type = int64 when use num_buckets
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
oneof dense_emb
autodis = 40
autodis embedding
mlp = 41
mlp embedding
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
loss.proto:4Used in: , , , ,
oneof loss
binary_cross_entropy = 1
softmax_cross_entropy = 2
l2_loss = 3
jrc_loss = 4
binary_focal_loss = 5
message
match_model.proto:75Used in:
required user_tower = 1
required item_tower = 2
optional float simi_pow = 3
optional Similarity similarity = 4
optional bool in_batch_negative = 5
optional float temperature = 6
required int32 output_dim = 7
message
tower.proto:174Used in:
required string input = 1
user feature group name
required string history_input = 2
user history group name
required user_mlp = 3
optional hist_seq_mlp = 4
optional MINDUserTower.UserSeqCombineMethod user_seq_combine = 5
required capsule_config = 6
capsule config
required concat_mlp = 7
concat mlp config for user interests vector
Used in:
CONCAT = 0
SUM = 1
message
module.proto:4Used in: , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
repeated uint32 hidden_units = 1
hidden units for each layer
repeated float dropout_ratio = 2
ratio of dropout
optional string activation = 3
activation function
optional bool use_bn = 4
use batch normalization
optional bool bias = 5
use bias
optional bool use_ln = 6
use layer normalization
message
feature.proto:57Used in: , , , , , , ,
(message has no fields)
message
multi_task_rank.proto:11Used in:
required expert_mlp = 1
mmoe expert module definition
optional gate_mlp = 2
mmoe gate module definition
required uint32 num_expert = 3
number of mmoe experts
repeated task_towers = 4
task tower
message
optimizer.proto:224Used in: , ,
repeated uint32 schedule_sizes = 1
a list of global steps or epochs at which to switch learning
repeated float learning_rates = 2
a list of learning rates corresponding to intervals
optional bool warmup = 3
Whether to linearly interpolate learning rates for steps in [0, schedule_steps[0]].
optional bool by_epoch = 4
schedule by epoch or by step.
message
module.proto:62Used in:
optional float reduction_ratio = 1
the ratio between aggregation dim and masked input dim
optional uint32 aggregation_dim = 2
the dim of aggregation layer
required uint32 hidden_dim = 3
the dim of hidden ffn layer
message
rank_model.proto:45Used in:
required mask_net_module = 1
message
module.proto:71Used in: ,
required uint32 n_mask_blocks = 1
number of mask blocks
required mask_block = 2
mask block
optional top_mlp = 3
mlp layer on top of mask blocks
optional bool use_parallel = 4
use parallel or serial mask blocks
message
feature.proto:466Used in: ,
required string feature_name = 1
feature name, e.g. match_cate_brand_click_count
required string nested_map = 2
nested map input, e.g. user:match_cate_brand_click_count
required string pkey = 3
first layer (primary) key input, e.g. item:cate or ALL
required string skey = 4
second layer (secondary) key input, e.g. item:brand or ALL
optional string embedding_name = 5
embedding name, feature with same embedding name will share embedding
optional uint32 embedding_dim = 6
embedding dimension
repeated float boundaries = 7
boundaries for bucktize numeric match value
optional uint64 hash_bucket_size = 8
number of hash size for sparse match value
optional uint64 num_buckets = 9
number of id enumerators for sparse match value
repeated string vocab_list = 10
id vocabulary list for sparse match value
map<string, uint64> vocab_dict = 11
id vocabulary dict
optional string pooling = 12
embedding pooling type, available is {sum | mean}
optional string default_value = 13
match value combiner type, available is {sum | mean | min | max | count} optional string combiner = 12 [default = "sum"]; fg default value
optional string separator = 14
fg multi-value separator
optional bool need_discrete = 15
match map value is sparse or numeric, when need_discrete is true, combiner will be empty string
optional bool show_pkey = 16
match value need pkey value as prefix or not.
optional bool show_skey = 17
match value need skey valueas prefix or not.
optional string normalizer = 18
fg normalizer, e.g. method=log10,threshold=1e-10,default=-10 method=zscore,mean=0.0,standard_deviation=10.0 method=minmax,min=2.1,max=2.2 method=expression,expr=sign(x)
optional string init_fn = 19
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional uint32 value_dim = 20
match value dimensions
optional bool use_mask = 21
mask value in training progress
optional zch = 22
zero collision hash
optional string vocab_file = 23
id vocabulary file path
optional string asset_dir = 24
vocab file relative directory
optional dynamicemb = 25
dynamic embedding
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional uint64 default_bucketize_value = 31
out-of-vocab(OOV) id bucketize value when use vocab_list or vocab_dict when use default_bucketize_value, we will not add additional bucketize_value of `default_value`=0, bucketize_value of <OOV>=1 into vocab_list or vocab_dict
optional string fg_value_type = 32
value_type after fg before bucketize, you can specify it for better performance. e.g. fg_value_type = int64 when use num_buckets
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
oneof dense_emb
autodis = 40
autodis embedding
mlp = 41
mlp embedding
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
metric.proto:19Used in: ,
(message has no fields)
message
metric.proto:22Used in: ,
(message has no fields)
message
metric.proto:53Used in: , , , ,
oneof metric
auc = 1
multiclass_auc = 2
recall_at_k = 3
mean_absolute_error = 4
mean_squared_error = 5
accuracy = 6
grouped_auc = 7
xauc = 8
grouped_xauc = 9
normalized_entropy = 10
message
model.proto:46Used in:
repeated feature_groups = 1
oneof model
dlrm = 100
deepfm = 101
multi_tower = 102
multi_tower_din = 103
mask_net = 104
wide_and_deep = 105
dcn_v1 = 106
dcn_v2 = 107
xdeepfm = 108
wukong = 109
simple_multi_task = 200
mmoe = 201
dbmtl = 202
ple = 203
dc2vr = 204
dlrm_hstu = 205
pepnet = 206
ultra_hstu = 207
dssm = 301
dssm_v2 = 302
dat = 303
hstu_match = 304
mind = 305
tdm = 400
rocket_launching = 500
sid_rqkmeans = 601
SID generation models (600 is reserved for SidRqvae, arriving in the follow-up PR)
optional uint32 num_class = 2
repeated losses = 3
repeated metrics = 4
repeated train_metrics = 5
optional variational_dropout = 11
optional Kernel kernel = 12
optional bool use_pareto_loss_weight = 13
whether use pareto loss weight
message
rank_model.proto:24Used in:
repeated towers = 1
required final = 2
message
rank_model.proto:29Used in:
repeated towers = 1
repeated din_towers = 2
required final = 3
message
seq_encoder.proto:52Used in:
optional string name = 1
seq encoder name
required string input = 2
sequence feature name
required attn_mlp = 3
time windows len
repeated uint32 windows_len = 4
mlp config for target attention score
message
tower.proto:156Used in:
repeated uint32 windows_len = 1
time windows len
required attn_mlp = 2
mlp config for target attention score
message
metric.proto:8Used in:
optional uint32 thresholds = 1
optional string average = 2
macro: calculate score for each class and average them weighted: calculates score for each class and computes weighted average using their support
message
sampler.proto:5Weighted Random Sampling ItemID not in Batch
Used in:
required string input_path = 1
sample data path schema => id:int64 | weight:float | attrs:string
required uint32 num_sample = 2
number of negative sample
repeated string attr_fields = 3
field names of attrs in train data or eval data
required string item_id_field = 4
field name of item_id in train data or eval data
optional string attr_delimiter = 5
attribute delimiter of attrs string
optional uint32 num_eval_sample = 6
number of negative samples for evaluator
optional string field_delimiter = 7
field delimiter of input data
optional string item_id_delim = 8
item id delimiter
message
sampler.proto:27Weighted Random Sampling ItemID not with Edge
Used in:
required string user_input_path = 1
user data path schema => userid:int64 | weight:float
required string item_input_path = 2
item data path schema => itemid:int64 | weight:float | attrs:string
required string pos_edge_input_path = 3
positive edge path schema => userid:int64 | itemid:int64 | weight:float
required uint32 num_sample = 4
number of negative sample
repeated string attr_fields = 5
field names of attrs in train data or eval data
required string item_id_field = 6
field name of item_id in train data or eval data
required string user_id_field = 7
field name of user_id in train data or eval data
optional string attr_delimiter = 8
attribute delimiter of attrs string
optional uint32 num_eval_sample = 9
number of negative samples for evaluator
optional string field_delimiter = 10
field delimiter of input data
message
metric.proto:48Used in:
optional float eta = 1
small epsilon clamping the population mean label rate away from {0, 1}.
message
feature.proto:616Used in: ,
required string feature_name = 1
feature name, e.g. overlap_ratio
required string query = 2
query input name, e.g. user:query
required string title = 3
title input name, e,g. item:title
required string method = 4
overlap calculate method, available is {query_common_ratio | title_common_ratio | is_contain | is_equal}
optional string embedding_name = 5
embedding name, feature with same embedding name will share embedding
optional uint32 embedding_dim = 6
embedding dimension
repeated float boundaries = 7
boundaries for bucktize numeric expr value
optional string normalizer = 8
fg normalizer, e.g. method=log10,threshold=1e-10,default=-10 method=zscore,mean=0.0,standard_deviation=10.0 method=minmax,min=2.1,max=2.2 method=expression,expr=sign(x)
optional string pooling = 10
embedding pooling type, available is {sum | mean}
optional string separator = 12
fg default value optional string default_value = 11 [default = "0"]; fg multi-value separator
optional string init_fn = 13
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional bool use_mask = 14
mask value in training progress
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
oneof dense_emb
autodis = 40
autodis embedding
mlp = 41
mlp embedding
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
multi_task_rank.proto:101Used in:
optional uint32 epnet_hidden_unit = 1
epnet hidden units
optional float epnet_gamma = 2
activation function for epnet
repeated uint32 ppnet_hidden_units = 3
ppnet hidden units
optional string ppnet_activation = 4
activation function for ppnet
repeated float ppnet_dropout_ratio = 5
ratio of dropout
optional float ppnet_gamma = 6
optional string domain_input_name = 7
domain feature name, must is num bucket
optional uint32 task_domain_num = 8
domain number for each task
repeated task_towers = 9
task tower
message
multi_task_rank.proto:52Used in:
repeated extraction_networks = 1
extraction network
repeated task_towers = 2
task tower
message
feature.proto:6Used in: , , , , , , , , , , , ,
repeated string sharding_types = 1
embedding sharding type constraints data_parallel | table_wise | column_wise | row_wise | table_row_wise | table_column_wise | grid_shard
repeated string compute_kernels = 2
embedding compute kernel constraints dense | fused | fused_uvm | fused_uvm_caching | key_value
message
optimizer.proto:45Used in:
oneof optimizer
sgd_optimizer = 1
adagrad_optimizer = 2
adam_optimizer = 3
adamw_optimizer = 4
adadelta_optimizer = 6
rmsprop_optimizer = 7
required string regex_pattern = 5
oneof learning_rate
constant_learning_rate = 101
exponential_decay_learning_rate = 102
manual_step_learning_rate = 103
cosine_annealing_learning_rate = 104
cosine_annealing_warm_restarts_learning_rate = 105
message
seq_encoder.proto:26Used in:
optional string name = 1
seq encoder name
required string input = 2
sequence feature name
optional string pooling_type = 3
pooling type, sum or mean
optional int32 max_seq_length = 6
maximum sequence length
message
optimizer.proto:197Used in: ,
required float lr = 1
optional float alpha = 2
optional float eps = 3
optional float weight_decay = 4
message
feature.proto:186Used in: ,
required string feature_name = 1
feature name, e.g. click_count
required string expression = 2
feature input, e.g. item:click_count
optional string embedding_name = 3
embedding name, feature with same embedding name will share embedding
optional uint32 embedding_dim = 4
embedding dimension
repeated float boundaries = 5
boundaries for bucktize numeric feature
optional uint32 value_dim = 6
raw feature of multiple dimensions
optional string normalizer = 7
fg normalizer, e.g. method=log10,threshold=1e-10,default=-10 method=zscore,mean=0.0,standard_deviation=10.0 method=minmax,min=2.1,max=2.2 method=expression,expr=sign(x)
optional string pooling = 10
embedding pooling type, available is {sum | mean}
optional string default_value = 11
fg default value
optional string separator = 12
fg multi-value separator
optional string init_fn = 13
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional bool use_mask = 14
mask value in training progress
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
oneof dense_emb
autodis = 40
autodis embedding
mlp = 41
mlp embedding
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
message
metric.proto:15Used in: ,
optional uint32 top_k = 1
message
general_rank_model.proto:8Used in:
optional share_mlp = 1
required booster_mlp = 2
required light_mlp = 3
optional bool feature_based_distillation = 5
optional Similarity feature_distillation_function = 6
COSINE = 0; EUCLID = 1;
message
optimizer.proto:153Used in: ,
required float lr = 1
optional float momentum = 2
optional float weight_decay = 3
optional float dampening = 4
optional bool nesterov = 5
optional bool fused = 6
message
module.proto:240Used in:
required uint32 embedding_dim = 1
dimension of input embeddings
required uint32 num_heads = 2
number of attention heads
required uint32 hidden_dim = 3
dimension of hidden linear layers
required uint32 attention_dim = 4
dimension of attention mechanism
optional float output_dropout_ratio = 5
dropout probability for linear layers
optional uint32 max_attn_len = 6
maximum length of attention window
optional float attn_alpha = 7
alpha for mha attention
optional bool use_group_norm = 8
use group normalization or layer normalization.
optional bool recompute_normed_x = 9
whether to recompute normed_x in backward
optional bool recompute_uvqk = 10
whether to recompute_uvqk in backward
optional bool recompute_y = 11
whether to recompute y in backward
optional bool sort_by_length = 12
whether to sort by sequence length when forwarding
optional int32 contextual_seq_len = 13
sequence length of contextual feature. Sentinel: < 0 (default) = use input_preprocessor.contextual_seq_len().
optional uint32 sla_k1 = 14
Semi-Local Attention: local causal window size (0 = disabled)
optional uint32 sla_k2 = 15
Semi-Local Attention: global prefix length (0 = disabled)
optional int32 scaling_seqlen = 16
attention output scaling divisor (denominator of the SiLU(QK)/N term). Sentinel: < 0 (default) = use runtime max_seq_len.
message
seq_encoder.proto:37Used in:
optional string name = 1
seq encoder name
required string input = 2
sequence feature name
optional int32 multihead_attn_dim = 3
multihead_attn_dim must be divisible by num_heads
optional int32 num_heads = 4
self attention num heads
optional float dropout = 5
dropout for attn_output_weights
optional int32 max_seq_length = 6
maximum sequence length
message
seq_encoder.proto:63Used in:
oneof seq_module
din_encoder = 1
simple_attention = 2
pooling_encoder = 3
multi_window_din_encoder = 4
self_attention_encoder = 6
message
feature.proto:1021Used in:
oneof feature
id_feature = 1
raw_feature = 2
combo_feature = 3
lookup_feature = 4
match_feature = 5
expr_feature = 7
overlap_feature = 8
tokenize_feature = 9
custom_feature = 10
kv_dot_product = 11
bool_mask_feature = 12
combine_feature = 13
message
model.proto:21Used in:
optional string group_name = 1
repeated string feature_names = 2
optional string embedding_name_suffix = 3
Suffix appended to each feature's embedding_name; same semantics as FeatureGroupConfig.embedding_name_suffix. Empty == disabled.
message
feature.proto:975Used in:
required string sequence_name = 1
sequence name
required uint32 sequence_length = 2
max sequence length, only take effect in fg
required string sequence_delim = 3
sequence delimiter
optional string sequence_pk = 4
sequence primary key name for serving, default will be user:{sequence_name}
repeated features = 5
sub feature config
message
sid_model.proto:18Used in:
optional uint32 input_dim = 1
Input embedding dimension (K-Means runs directly on raw embeddings, no encoder).
repeated uint32 codebook = 3
Per-layer cluster counts, e.g. [256, 256, 256]. List length is the number of residual quantization layers. Entries may differ per layer (non-uniform codebooks such as [256, 512, 1024] are supported — the FAISS backend fits a separate ``faiss.Kmeans`` per layer).
optional bool normalize_residuals = 4
L2-normalize residuals before each layer.
optional faiss_kmeans_kwargs = 5
Strictly-typed extra kwargs forwarded to faiss.Kmeans(D, K, **kwargs).
optional uint32 train_sample_size = 6
Target number of embeddings to reservoir-sample for the FAISS fit. Bounds host memory regardless of corpus size. 0 (the default) auto-derives it as max(K) * max_points_per_centroid (the largest per-layer codebook, for non-uniform codebooks) — exactly what FAISS subsamples to internally (default 256), so no training points are wasted.
optional string embedding_feature_name = 40
Name of the item embedding feature inside the input Batch.
Used in: , , , , ,
COSINE = 0
INNER_PRODUCT = 1
EUCLID = 2
message
seq_encoder.proto:17Used in:
optional string name = 1
seq encoder name
required string input = 2
sequence feature name
optional int32 max_seq_length = 6
maximum sequence length
message
multi_task_rank.proto:7Used in:
repeated task_towers = 1
message
loss.proto:18Used in:
optional float label_smoothing = 1
message
optimizer.proto:4Used in:
oneof optimizer
sgd_optimizer = 1
adagrad_optimizer = 2
adam_optimizer = 3
lars_sgd_optimizer = 4
lamb_optimizer = 5
partial_rowwise_lamb_optimizer = 6
partial_rowwise_adam_optimizer = 7
rowwise_adagrad_optimizer = 8
adadelta_optimizer = 9
rmsprop_optimizer = 10
oneof learning_rate
constant_learning_rate = 101
exponential_decay_learning_rate = 102
manual_step_learning_rate = 103
cosine_annealing_learning_rate = 104
cosine_annealing_warm_restarts_learning_rate = 105
message
match_model.proto:50Used in:
required multiwindow_din = 1
required final = 2
message
sampler.proto:114Used in:
required string item_input_path = 1
schema => itemid:int64 | weight:float | attrs:string
required string edge_input_path = 2
scheme => src_id:int64 | dst_id:int64 | weight:float edge for train.
required string predict_edge_input_path = 3
scheme => src_id:int64 | dst_id:int64 | weight:float edge for retrieval beam search.
repeated string attr_fields = 4
field names of attrs in train data or eval data
required string item_id_field = 5
field name of item_id in train data or eval data
repeated uint32 layer_num_sample = 6
the number of negative samples per layer
optional string attr_delimiter = 7
attribute delimiter of attrs string
optional uint32 num_eval_sample = 8
number of negative samples for evaluator
optional string field_delimiter = 9
field delimiter of input data
optional float remain_ratio = 10
The training process only trains a randomly selected proportion of nodes in the middle layers of the tree
optional string probability_type = 11
The type of probability for selecting and retaining each layer in the middle layers of the tree
message
tower.proto:30Used in: , , ,
required string tower_name = 1
task name for the task tower
required string label_name = 2
label for the task
repeated metrics = 3
metrics for the task
repeated train_metrics = 20
log train merics for task
repeated losses = 4
loss for the task
optional uint32 num_class = 5
num_class for multi-class classification loss
optional mlp = 6
task specific mlp
optional float weight = 7
training loss weights
optional string sample_weight_name = 8
sample weight for the task
optional string task_space_indicator_label = 9
label name for indicating the sample space for the task tower
optional float in_task_space_weight = 10
the loss weight for sample in the task space
optional float out_task_space_weight = 11
the loss weight for sample out the task space
optional float pareto_min_loss_weight = 12
use pareto front minimal loss weight, ge 0 and lt 1
Used in:
TEXT_LOWER2UPPER = 0
lower case to upper case
TEXT_UPPER2LOWER = 1
upper case to lower case
TEXT_SBC2DBC = 2
sbc case to dbc case
TEXT_CHT2CHS = 3
traditional chinese to simple chinese
TEXT_FILTER = 4
filter speicial chars
TEXT_SPLITCHRS = 5
chinese split to chars with blanks
TEXT_REMOVE_SPACE = 6
remove space
message
feature.proto:694Used in:
optional uint32 max_length = 1
if text_length greater than max_length, will not do normalize
optional string stop_char_file = 2
stop char file path, default will use built-in stop char
repeated TextNormalizeOption norm_options = 3
text normalize options, default is TEXT_LOWER2UPPER & TEXT_SBC2DBC & TEXT_CHT2CHS & TEXT_FILTER
message
feature.proto:703Used in: ,
required string feature_name = 1
feature name, e.g. title_token
required string expression = 2
feature input, e.g. item:title
optional string embedding_name = 3
embedding name, feature with same embedding name will share embedding
required uint32 embedding_dim = 4
embedding dimension
optional text_normalizer = 6
text normalizer
required string vocab_file = 7
tokenizer vocabulary file path
optional string asset_dir = 8
vocab file relative directory
optional string pooling = 10
embedding pooling type, available is {sum | mean}
optional string default_value = 11
fg default value, default value before bucktize
optional string tokenizer_type = 12
tokenizer_type type, available is {bpe | sentencepiece}
optional string init_fn = 14
embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
optional bool use_mask = 15
mask value in training progress
optional string fg_encoded_default_value = 30
default value when fg_mode = FG_NONE, when use pai-fg, you do not need to set the param. when use own fg and data contain null value, you can set the param for fill null
optional bool trainable = 33
embedding param trainable or not
optional bool stub_type = 34
only used as fg dag intermediate result or not
optional string data_type = 35
embedding data type
optional embedding_constraints = 50
embedding param constraints
optional uint32 sequence_length = 101
max sequence length, only take effect when use it as sequence
optional string sequence_delim = 102
sequence delimiter, only take effect when use it as sequence
repeated string sequence_fields = 103
specify sequence type fields in inputs. default is item side inputs.
optional bool tokens_as_sequence = 104
When true, treat the tokenized output of a single text as a sequence of tokens (each token is one sequence element, value_dim=1) so the feature can be fed into sequence_encoder modules (pooling_encoder, self_attention_encoder, etc.) via EmbeddingCollection instead of being pooled by EmbeddingBagCollection. This is distinct from the `sequence_tokenize_feature` oneof entry, which interprets the input as a `sequence_delim`-separated list of texts.
message
tower.proto:7Used in: , , , , ,
required string input = 1
input feature group name
optional mlp = 2
mlp config (optional; when unset the tower applies no projection)
message
train.proto:44Used in:
required sparse_optimizer = 1
embedding part optimizer
required dense_optimizer = 2
dense part optimizer
optional uint32 num_steps = 3
number of steps to train models
optional uint32 num_epochs = 4
number of epochs to train models
optional uint32 save_checkpoints_steps = 5
step interval for saving checkpoint
optional string fine_tune_checkpoint = 6
checkpoint to restore parameters from
optional string fine_tune_ckpt_param_map = 7
checkpoint to restore parameters mapping, each line is {param name in current model}\\t{param name in old ckpt}
optional uint32 log_step_count_steps = 8
the frequency the loss and lr will be logged during training
optional bool is_profiling = 9
profiling or not
optional bool use_tensorboard = 10
use tensorboard or not.
optional uint32 save_checkpoints_epochs = 11
epoch interval for saving checkpoint
repeated string tensorboard_summaries = 12
the summaries to be saved in tensorboard, activated only when use_tensorboard=true, possible values are: "loss", "learning_rate", "parameter", "global_gradient_norm", "gradient_norm", "gradient" default values are ["loss", "learning_rate"]
optional bool cudnn_allow_tf32 = 13
where to use torch.backends.cudnn.allow_tf32
optional bool cuda_matmul_allow_tf32 = 14
where to use torch.backends.cuda.matmul.allow_tf32
optional global_embedding_constraints = 15
global embedding param constraints
optional string mixed_precision = 16
mixed precision dtype.
optional grad_scaler = 17
grad_scaler dynamically estimates the scale factor each iteration.
optional uint32 gradient_accumulation_steps = 18
gradient accumulation steps
optional grad_clipping = 19
dense gradient clipping config
optional uint32 keep_checkpoint_max = 20
maximum number of recent checkpoints to keep; 0 keeps all.
optional uint32 save_checkpoints_timestamp_interval = 21
save every N seconds of consumed event-time (e.g. kafka message timestamp), aligned to the Unix epoch (not training epochs). 0 disables.
repeated uint64 save_checkpoints_timestamps = 22
absolute event-time targets (Unix-epoch seconds); save once when consumed data crosses each. empty disables.
optional float save_checkpoints_timestamp_quorum = 23
fraction of workers (0,1] that must pass a boundary/target before a timestamp checkpoint fires; default 0.5 (1.0 = all). outlier-robust.
optional delta_embedding_dump_config = 24
Configuring this field dumps changed sparse embedding rows during CUDA training. Multi-GPU training writes one parquet shard per rank; column-wise embedding sharding is not supported.
TBD: qcomm config
message
metric.proto:69Used in: , , , ,
oneof metric
auc = 1
recall_at_k = 3
mean_absolute_error = 4
mean_squared_error = 5
accuracy = 6
xauc = 8
optional float decay_rate = 20
metric decay rate
optional uint32 decay_step = 21
train_config.log_step_count_steps can divide decay_steps evenly.
message
multi_task_rank.proto:78Used in:
repeated hstu = 1
Mixture of Transducers: one HSTU per channel; per-candidate outputs are concatenated on the embedding dim. >= 2 entries requires every entry to set a unique non-empty `name` plus the matching `<name>` / `<name>_action` / `<name>_watchtime` / `<name>_timestamp` feature_groups. Candidate-side and contextual groups are shared across all channels.
required fusion_mtl_tower = 2
multi task tower config
required uint32 max_seq_len = 3
max sequence length
optional uint32 item_embedding_hidden_dim = 4
item embedding mlp hidden dimension
optional bool enable_global_average_loss = 5
enables loss averaging computation globally across all ranks (total rank) instead of locally (local rank).
optional bool sequence_timestamp_is_ascending = 6
timestamp of sequence is ascending or descending
optional bool concat_contextual_features = 7
concat all contextual features on channel dim as one token
message
module.proto:31Used in:
optional float regularization_lambda = 1
regularization coefficient lambda
optional bool embedding_wise_variational_dropout = 2
variational_dropout dimension
Used in:
NONE = 0
L2 = 1
DECOUPLE = 2
message
rank_model.proto:8Used in:
required deep = 1
optional final = 2
optional uint32 wide_embedding_dim = 3
optional string wide_init_fn = 4
wide embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"
message
rank_model.proto:75Used in:
optional dense_mlp = 1
repeated wukong_layers = 2
required final = 3
message
module.proto:99Used in:
required uint32 lcb_feature_num = 1
LinearCompressBlock output feature num
required uint32 fmb_feature_num = 2
FactorizationMachineBlock output feature num
optional uint32 compressed_feature_num = 3
number of compressed features in optimized FM.
required feature_num_mlp = 4
feature num mlp
message
metric.proto:36Used in: ,
optional float sample_ratio = 1
optional uint64 max_pairs = 2
optional bool in_batch = 3
message
feature.proto:31Used in: , , , , ,
required uint64 zch_size = 1
zero collision size
optional uint32 eviction_interval = 2
evict interval steps
oneof eviction_policy
evict policy
lfu = 101
lru = 102
distance_lfu = 103
optional string threshold_filtering_func = 3
lambda function string used to filter incoming ids before update/eviction. experimental feature. [input: Tensor] the function takes as input a 1-d tensor of unique id counts. [output1: Tensor] the function returns a boolean_mask or index array of corresponding elements in the input tensor that pass the filter. [output2: float, Tensor] the function returns the threshold that will be used to filter ids before update/eviction. all values <= this value will be filtered out.
message
rank_model.proto:65Used in:
required cin = 1
required deep = 2
required final = 3
optional uint32 wide_embedding_dim = 4
optional string wide_init_fn = 5
wide embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01"