Protocol Documentation

mini batch size to use for training and evaluation. Default: 1024

dataset type. Default: OdpsDataset

[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 Default: true

separator for multi-val feature in fg encoded input data Default: 

labels 

number of workers for parallel processing raw data Default: 8

pin memory for fast cudaMemCopy Default: true

the input fields must be the same number and in the
same order as data in csv files 

delimiter of column features, only used for CsvDataset Default: ,

for csv files, with header or not. Default: false

mini batch size to use for and evaluation. 

drop last batch less than batch_size Default: false

fg threads for each worker,
  if fg_threads = 0, will disable fg dag handler, use python run. Default: 1

when use OdpsDataset, read data orderby table partitions or not. Default: false

maxcompute storage api & tunnel quota name Default: pay-as-you-go

mask probability for samples in training progress Default: 0

mask probability for sampled negatives in training progress Default: 0

force padding data into same data group with same batch_size Default: false

sample weights 

fg run mode. Default: FG_NONE

whether to shuffle data Default: false

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. Default: 32

maxcompute storage api data compression type, LZ4_FRAME | ZSTD | UNCOMPRESSED Default: LZ4_FRAME

sample cost field name 

batch cost limit size 

simplified input fields string format: input_name1:input_type1;input_name2:input_type2;
type names follow ODPS conventions with aliases: BIGINT->INT64, INT->INT32 

only need specify it when use CsvDataset and
value dtype can not be inferred (all values in the column are null) 

number of steps to evaluate. 

the frequency progress be logged during eval Default: 10

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 Default: latest

the metric used to determine the best checkpoint Default: auc

metric value the bigger the best Default: true

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. 

whether to use torch.backends.cudnn.allow_tf32 Default: true

whether to use torch.backends.cuda.matmul.allow_tf32 Default: false

number of embedding channels 

temperature coefficient for softmax Default: 0.1

 Default: 0.8

feature name. 

feature input, e.g. item:item_id 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

number of hash size 

number of id enumerators 

id vocabulary list 

id vocabulary dict 

id value dimensions, default = 0, when use in seq, default = 1
if value_dim = 0, it supports id with multi-value 

embedding pooling type, available is {sum | mean} Default: sum

fg default value, default value before bucktize 

fg multi-value separator Default: 

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

mask value in training progress 

zero collision hash 

id vocabulary file path 

vocab file relative directory 

dynamic embedding 

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 

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 

value_type after fg before bucketize, you can specify it for better performance.
e.g. fg_value_type = int64 when use num_buckets 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

feature name, e.g. tag_feat 

feature input, e.g. user:tag 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

value map for mapping input string values to float values 

boundaries for bucktize numeric combine value 

number of id enumerators for sparse combine value 

embedding pooling type, available is {sum | mean}. Controls embedding bag aggregation.
NOTE: distinct from 'combiner' which controls FG-level multi-value aggregation. Default: sum

fg default value, default value before bucktize 

fg multi-value separator Default: 

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) 

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

mask value in training progress 

combine value combiner type, available is {sum | mean | min | max}
Controls FG-level multi-value aggregation before bucketization. Default: sum

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 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

autodis embedding 

mlp embedding 

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

feature name, e.g. os_and_cate 

feature input, e.g. [user:os, item:cate] 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

number of hash size 

id vocabulary list 

id vocabulary dict 

id value dimensions, if value_dim = 0, it supports id with multi-value Default: 0

embedding pooling type, available is {sum | mean} Default: sum

fg default value 

fg multi-value separator Default: 

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

mask value in training progress 

zero collision hash 

id vocabulary file path 

vocab file relative directory 

dynamic embedding 

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 

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 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

feature name. 

custom operator name. 

custom operator lib file name. 

operator custom params. 

custom operator is thread safe or not. Default: false

feature input, e.g. user:os 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

boundaries for bucktize numeric value 

number of hash size for sparse value 

number of id enumerators for sparse value 

id vocabulary list for sparse value 

id vocabulary dict 

embedding pooling type, available is {sum | mean} Default: sum

fg default value Default: 0

fg multi-value separator Default: 

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) 

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

value dimensions 

mask value in training progress 

zero collision hash 

id vocabulary file path 

vocab file relative directory 

dynamic embedding 

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 

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 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

autodis embedding 

mlp embedding 

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

decay rate is access step Default: 1

minimum frequency threshold for admission 

determine how to initialize the embedding if the key is not admitted. 

kv counter capacity, if not set, use embedding max_capacity 

kv counter capacity for each bucket. Default: 1024

the mode of initialization. NORMAL | TRUNCATED_NORMAL | UNIFORM | CONSTANT 

the mean value for (truncated) normal distributions Default: 0

the standard deviation for (truncated) normal distributions.
default is sqrt(1 / embedding_dim) 

the lower bound for uniform/truncated_normal distribution.
default is -sqrt(1 / max_capacity) 

the upper bound for uniform/truncated_normal distribution.
default is sqrt(1 / max_capacity) 

the constant value for constant initialization. Default: 0

arguments for initializing dynamic embedding vector values.
default is uniform distribution, and absolute values of upper and lower bound are sqrt(1 / embedding_dim). 

the initializer args for evaluation mode. default is
default is constant initialization with value 0.0. 

strategy to set the score for each indices in forward and backward per table. TIMESTAMP | STEP | CUSTOMIZED | LFU | NO_EVICTION Default: STEP

max number of embedding rows 

percentage of embedding rows caching on gpu 

init number of capacity 

init table path 

hash-table bucket capacity. default 128 (matches dynamicemb
DEFAULT_BUCKET_CAPACITY). larger buckets trade probe cost for
higher load factor. 

feature name, e.g. kv_os_click_count 

expression, e.g. sigmoid(pv/(1+click)) 

variables in expression, e,g. ["item:pv", "item:click"] 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

boundaries for bucktize numeric expr value 

fg multi-value separator Default: 

fill value when vector length mismatch, default is NaN. 

embedding pooling type, available is {sum | mean} Default: sum

fg default value Default: 0

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

mask value in training progress 

if value_dim = 0, it supports multi-value Default: 0

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 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

autodis embedding 

mlp embedding 

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

feature name, e.g. item_id 

feature input, e.g. item:item_id 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

number of hash size 

number of id enumerators 

id vocabulary list 

id vocabulary dict 

id value dimensions, default = 0, when use in seq, default = 1
if value_dim = 0, it supports id with multi-value 

embedding pooling type, available is {sum | mean} Default: sum

fg default value, default value before bucktize 

fg multi-value separator Default: 

fg multi-value with whether has weight Default: false

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

mask value in training progress 

zero collision hash 

id vocabulary file path 

vocab file relative directory 

dynamic embedding 

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 

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 

value_type after fg before bucketize, you can specify it for better performance.
e.g. fg_value_type = int64 when use num_buckets 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

feature name, e.g. kv_os_click_count 

query, e.g. ["a:0.5", "b:0.5"] 

document, e,g. ["d:0.5", "b:0.5"] 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

boundaries for bucktize numeric expr value 

fg multi-value separator Default: 

fg kv separator, default is :. 

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) 

embedding pooling type, available is {sum | mean} Default: sum

fg default value Default: 0

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

mask value in training progress 

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 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

autodis embedding 

mlp embedding 

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

decay rate is access step Default: 1

feature name, e.g. kv_os_click_count 

map input, e.g. item:kv_os_click_count 

key input, e.g. user:os 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

boundaries for bucktize numeric lookup value 

number of hash size for sparse lookup value 

number of id enumerators for sparse lookup value 

id vocabulary list for sparse lookup value 

id vocabulary dict 

embedding pooling type, available is {sum | mean} Default: sum

lookup value combiner type, available is {sum | mean | min | max | count} Default: sum

fg default value Default: 0

fg multi-value separator Default: 

lookup map value is sparse or numeric,
when need_discrete is true, combiner will be empty string Default: false

lookup value need key as prefix or not. Default: false

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) 

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

lookup value dimensions 

numeric lookup value separator Default: ,

mask value in training progress 

zero collision hash 

id vocabulary file path 

vocab file relative directory 

dynamic embedding 

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 

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 

value_type after fg before bucketize, you can specify it for better performance.
e.g. fg_value_type = int64 when use num_buckets 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

autodis embedding 

mlp embedding 

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

feature name, e.g. match_cate_brand_click_count 

nested map input, e.g. user:match_cate_brand_click_count 

first layer (primary) key input, e.g. item:cate or ALL 

second layer (secondary) key input, e.g. item:brand or ALL 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

boundaries for bucktize numeric match value 

number of hash size for sparse match value 

number of id enumerators for sparse match value 

id vocabulary list for sparse match value 

id vocabulary dict 

embedding pooling type, available is {sum | mean} Default: sum

match value combiner type, available is {sum | mean | min | max | count}
optional string combiner = 12 [default = "sum"];
fg default value Default: 0

fg multi-value separator Default: 

match map value is sparse or numeric,
when need_discrete is true, combiner will be empty string Default: false

match value need pkey value as prefix or not. Default: false

match value need skey  valueas prefix or not. Default: false

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) 

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

match value dimensions 

mask value in training progress 

zero collision hash 

id vocabulary file path 

vocab file relative directory 

dynamic embedding 

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 

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 

value_type after fg before bucketize, you can specify it for better performance.
e.g. fg_value_type = int64 when use num_buckets 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

autodis embedding 

mlp embedding 

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

feature name, e.g. overlap_ratio 

query input name, e.g. user:query 

title input name, e,g. item:title 

overlap calculate method, available is {query_common_ratio | title_common_ratio | is_contain | is_equal} 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

boundaries for bucktize numeric expr value 

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) 

embedding pooling type, available is {sum | mean} Default: sum

fg default value
optional string default_value = 11 [default = "0"];
fg multi-value separator Default: 

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

mask value in training progress 

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 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

autodis embedding 

mlp embedding 

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

embedding sharding type constraints
data_parallel | table_wise | column_wise | row_wise | table_row_wise | table_column_wise | grid_shard 

embedding compute kernel constraints
dense | fused | fused_uvm | fused_uvm_caching | key_value 

feature name, e.g. click_count 

feature input, e.g. item:click_count 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

boundaries for bucktize numeric feature 

raw feature of multiple dimensions Default: 1

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) 

embedding pooling type, available is {sum | mean} Default: sum

fg default value Default: 0

fg multi-value separator Default: 

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

mask value in training progress 

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 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

autodis embedding 

mlp embedding 

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

sequence name 

max sequence length, only take effect in fg 

sequence delimiter Default: ;

sequence primary key name for serving,
default will be user:{sequence_name} 

sub feature config 

if text_length greater than max_length, will not do normalize 

stop char file path, default will use built-in stop char 

text normalize options, default is TEXT_LOWER2UPPER & TEXT_SBC2DBC & TEXT_CHT2CHS & TEXT_FILTER 

feature name, e.g. title_token 

feature input, e.g. item:title 

embedding name, feature with same embedding name will share embedding 

embedding dimension 

text normalizer 

tokenizer vocabulary file path 

vocab file relative directory 

embedding pooling type, available is {sum | mean} Default: sum

fg default value, default value before bucktize 

tokenizer_type type, available is {bpe | sentencepiece} Default: bpe

embedding init function, e.g. "nn.init.uniform_,a=-0.01,b=0.01" 

mask value in training progress 

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 

embedding param trainable or not Default: true

only used as fg dag intermediate result or not Default: false

embedding data type Default: FP32

embedding param constraints 

max sequence length, only take effect when use it as sequence 

sequence delimiter, only take effect when use it as sequence Default: ;

specify sequence type fields in inputs. default is item side inputs. 

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. Default: false

zero collision size 

evict interval steps Default: 5

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. 

 Default: 0

 Default: 2

 Default: 0.5

 Default: 0.5

 Default: 0

 Default: 200

probs threshold Default: 0.5

top k accuracy when num_class > 1 Default: 1

 Default: 100

 Default: 200

macro: calculate score for each class and average them
weighted: calculates score for each class and computes weighted average using their support Default: macro

small epsilon clamping the population mean label rate away from {0, 1}. Default: 1e-12

 Default: 5

metric decay rate Default: 0.9

train_config.log_step_count_steps can divide decay_steps evenly. Default: 100

 Default: 0.001

 Default: false

 Default: DEEP

Suffix appended to each feature's embedding_name so groups with
different suffixes use independent embedding tables. Empty == disabled. 

 Default: 1

 Default: PYTORCH

whether use pareto loss weight Default: false

Suffix appended to each feature's embedding_name; same semantics as
FeatureGroupConfig.embedding_name_suffix. Empty == disabled. 

max number of high capsules Default: 5 Default: 5

max behaviour sequence length 

high capsule embedding vector dimension 

dynamic routing iterations, Default: 3 Default: 3

routing logits scale Default: 20 Default: 20

routing logits initial stddev Default: 1 Default: 1

squash power Default: 1 Default: 1

whether to use constant capsule number, Default: false Default: false

the initialization method for routing logits,
 Default: normal, available: zeros. Default: normal

every layer size 

number of cross layers Default: 3

number of cross layers Default: 3

Matrix decomposition with minimal rank. Default: 32

number of experts per task 

number of experts for share 

mlp network of experts per task 

mlp network of experts for share 

slice candidate dim to uih dim 

padding candidate dim to uih dim 

linear transform uih and candidate to same dim 

action encoder config 

enable interleave target or not Default: true

action embedding mlp config 

content encoder config 

content embedding mlp config 

action encoder config 

action embedding mlp config 

content encoder config 

content embedding mlp config 

mlp for sequence embedding 

mlp for sequence and contextual embedding 

input preprocessor with contextual features 

input preprocessor with interleave targets 

input preprocessor for sequence-only models (no candidate concat) 

mlp for uih seq embedding 

mlp for candidate seq embedding 

l2 norm postprocessor 

layer norm postprocessor 

timestamp layer norm postprocessor 

mlp hidden dimension 

dropout ratio for contextual embedding Default: 0.3

buckets for position embedding 

buckets for timestamp embedding 

use timestamp encoding or not. Default: false

transform function for timestamp gap. sqrt | log Default: sqrt

timestamp gap will div by time_bucket_increments Default: 60

action embedding dim 

bitmask of each action 

thresholds for watch time to actions 

bitmask for watch time to actions 

action embedding weights init std Default: 0.1

mlp hidden dimension 

time duration period units, e.g. 60 * 60 for hour of day. 

time duration units per period, e.g. 24 for hour of day. 

action encoder config (optional - for models with action info) 

action embedding mlp config (required if action_encoder is set) 

stu config 

dropout ratio after preprocessor Default: 0.2

num stu layers Default: 3

position encoder 

input preprocessor 

output postprocessor 

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. 

Trailing UIH cap; both fields must be > 0 to enable truncation. 

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. 

hidden units for each layer 

ratio of dropout 

activation function Default: nn.ReLU

use batch normalization Default: false

use bias Default: true

use layer normalization Default: false

the ratio between aggregation dim and masked input dim Default: 1

the dim of aggregation layer 

the dim of hidden ffn layer 

number of mask blocks 

mask block 

mlp layer on top of mask blocks 

use parallel or serial mask blocks Default: true

dimension of input embeddings 

number of attention heads 

dimension of hidden linear layers 

dimension of attention mechanism 

dropout probability for linear layers Default: 0.3

maximum length of attention window 

alpha for mha attention 

use group normalization or layer normalization. Default: false

whether to recompute normed_x in backward Default: true

whether to recompute_uvqk in backward Default: true

whether to recompute y in backward Default: true

whether to sort by sequence length when forwarding Default: true

sequence length of contextual feature.
Sentinel: < 0 (default) = use input_preprocessor.contextual_seq_len(). Default: -1

Semi-Local Attention: local causal window size (0 = disabled) 

Semi-Local Attention: global prefix length (0 = disabled) 

attention output scaling divisor (denominator of the SiLU(QK)/N term).
Sentinel: < 0 (default) = use runtime max_seq_len. Default: -1

regularization coefficient lambda Default: 0.01

variational_dropout dimension Default: false

LinearCompressBlock output feature num 

FactorizationMachineBlock output feature num 

number of compressed features in optimized FM. Default: 16

feature num mlp 

 Default: 0.002

 Default: 0.95

 Default: 1e-06

 Default: 0

 Default: 0.002

 Default: 0

 Default: 0

 Default: 1e-10

 Default: false

 Default: 0.002

 Default: 0.9

 Default: 0.999

 Default: 0

 Default: 1e-08

 Default: false

 Default: false

 Default: 0.002

 Default: 0.9

 Default: 0.999

 Default: 0

 Default: 1e-08

 Default: false

 Default: false

total number of steps or epochs for cosine annealing 

minimum learning rate Default: 0

warmup start learning rate Default: 0

warmup steps or epochs Default: 0

schedule by epoch or by step. Default: false

number of steps or epochs for the first cosine annealing period 

factor to grow period length after each restart (1 = fixed period) Default: 1

minimum learning rate Default: 0

warmup start learning rate Default: 0

warmup steps or epochs Default: 0

schedule by epoch or by step. Default: false

decay steps or epochs 

decay rate Default: 0.95

if true, decay the learning rate at discrete intervals Default: true

warmup start learning rate Default: 0

warmup steps or epochs Default: 0

minimum learning rate Default: 0

schedule by epoch or by step. Default: false

 Default: 0.002

 Default: 0.95

 Default: 1e-06

 Default: 0

 Default: false

 Default: 1

 Default: 0.002

 Default: false

 Default: 1

 Default: 0

 Default: 0.002

 Default: 0.9

 Default: 0.999

 Default: 0

 Default: false

 Default: 1

 Default: 0.002

 Default: 0.9

 Default: 0.999

 Default: 0

 Default: false

 Default: 1

 Default: 0.002

 Default: 0.9

 Default: 0

 Default: false

 Default: 1

 Default: 0.002

 Default: 0.9

 Default: 0.999

 Default: 0

 Default: false

 Default: 1

 Default: 0.002

 Default: 0.9

 Default: 0.999

 Default: 0

 Default: false

 Default: 1

 Default: 0.002

 Default: 0.99

 Default: 1e-08

 Default: 0

 Default: false

 Default: 1

 Default: 0.002

 Default: 0

 Default: NONE

 Default: false

 Default: 1

 Default: 0.002

 Default: false

 Default: 1

a list of global steps or epochs at which to switch learning 

a list of learning rates corresponding to intervals 

Whether to linearly interpolate learning rates for steps in
[0, schedule_steps[0]]. Default: false

schedule by epoch or by step. Default: false

 Default: 0.002

 Default: 0.99

 Default: 1e-08

 Default: 0

 Default: 0.002

 Default: 0.9

 Default: 0

 Default: 0

 Default: false

 Default: false

user data path
schema => userid:int64 | weight:float 

item data path
schema => itemid:int64 | weight:float | attrs:string 

hard negative edge path
schema => userid:int64 | itemid:int64 | weight:float 

number of negative sample 

max number of hard negative sample 

field names of attrs in train data or eval data 

field name of item_id in train data or eval data 

field name of user_id in train data or eval data 

attribute delimiter of attrs string Default: :

number of negative samples for evaluator Default: 0

only works on local 

user data path
schema => userid:int64 | weight:float 

item data path
schema => itemid:int64 | weight:float | attrs:string 

positive edge path
schema => userid:int64 | itemid:int64 | weight:float 

hard negative edge path
schema => userid:int64 | itemid:int64 | weight:float 

number of negative sample 

max number of hard negative sample 

field names of attrs in train data or eval data 

field name of item_id in train data or eval data 

field name of user_id in train data or eval data 

attribute delimiter of attrs string Default: :

number of negative samples for evaluator Default: 0

field delimiter of input data 

sample data path
schema => id:int64 | weight:float | attrs:string 

number of negative sample 

field names of attrs in train data or eval data 

field name of item_id in train data or eval data 

attribute delimiter of attrs string Default: :

number of negative samples for evaluator Default: 0

field delimiter of input data 

item id delimiter Default: ;

user data path
schema => userid:int64 | weight:float 

item data path
schema => itemid:int64 | weight:float | attrs:string 

positive edge path
schema => userid:int64 | itemid:int64 | weight:float 

number of negative sample 

field names of attrs in train data or eval data 

field name of item_id in train data or eval data 

field name of user_id in train data or eval data 

attribute delimiter of attrs string Default: :

number of negative samples for evaluator Default: 0

field delimiter of input data 

schema => itemid:int64 | weight:float | attrs:string 

scheme => src_id:int64 | dst_id:int64 | weight:float
edge for train. 

scheme => src_id:int64 | dst_id:int64 | weight:float
edge for retrieval beam search. 

field names of attrs in train data or eval data 

field name of item_id in train data or eval data