Arguments for gpt-neox. All of the following can be specified in your .yml config file(s):
LR Scheduler Arguments
-
lr_decay_style: typing.Literal['constant', 'linear', 'cosine', 'exponential']
Default = linear
Learning rate decay function. Choose from 'constant', 'linear', 'cosine', 'exponential'.
-
lr_decay_iters: int
Default = None
Number of iterations to decay learning rate over, If None defaults to --train-iters
-
min_lr: float
Default = 0.0
Minumum value for learning rate. The scheduler clips values below this threshold.
-
warmup: float
Default = 0.01
Percentage of total iterations to warmup on (.01 = 1 percent of all training iters).
-
override_lr_scheduler: bool
Default = False
Reset the values of the scheduler (learning rate,warmup iterations, minimum learning rate, maximum number of iterations, and decay style from input arguments and ignore values from checkpoints. Note that all the above values will be reset.
-
use_checkpoint_lr_scheduler: bool
Default = False
Use checkpoint to set the values of the scheduler (learning rate, warmup iterations, minimum learning rate, maximum number of iterations, and decay style from checkpoint and ignore input arguments.
Logging Arguments
-
use_wandb: bool
Default = None
Flag indicating if wandb is to be used.
-
wandb_group: str
Default = None
Weights and Biases group name - used to group together "runs".
-
wandb_team: str
Default = None
Team name for Weights and Biases.
-
wandb_project: str
Default = neox
wandb project name
-
wandb_host: str
Default = https://api.wandb.ai
url of the wandb host
-
git_hash: str
Default = a593ce2
current git hash of repository
-
log_dir: str
Default = None
Directory to save logs to.
-
tensorboard_dir: str
Default = None
Write TensorBoard logs to this directory.
-
log_interval: int
Default = None
Interval between logging.
-
log_param_norm: bool
Default = False
Log the frob norm of the parameters to wandb / tensorboard (useful for debugging).
-
log_grad_norm: bool
Default = False
Log the frob norm of the gradients to wandb / tensorboard (useful for debugging). (N.B - this will only work with pp = 0 for now, as we don't have access to the gradients of the model because deepspeed.)
-
log_optimizer_states: bool
Default = False
Log the frob norm of the optimizer states to wandb / tensorboard (useful for debugging).
-
log_gradient_noise_scale: bool
Default = False
Whether to log the gradient noise scale when training (cf. https://arxiv.org/abs/1812.06162 for explanation)
-
gradient_noise_scale_n_batches: int
Default = 5
Number of batches to accumulate gradients for in the gradient noise scale logger.
-
gradient_noise_scale_cpu_offload: bool
Default = False
Whether to offload the buffered gradients to cpu when measuring gradient noise scale.
Model Arguments
-
precision: typing.Literal['fp16', 'fp32', 'bfloat16']
Default = None
description of the used precision, either one of fp16 or fp32 (and in the future bf16).
-
num_layers: int
Default = None
Number of transformer layers.
-
hidden_size: int
Default = None
Transformer hidden size.
-
num_attention_heads: int
Default = None
Number of transformer attention heads.
-
seq_length: int
Default = None
Maximum sequence length to process.
-
max_position_embeddings: int
Default = None
Maximum number of position embeddings to use. This is the size of position embedding.
-
norm: typing.Literal['layernorm', 'rmsnorm', 'scalenorm']
Default = layernorm
Normalization layer to use. Choose from "layernorm", "rmsnorm", "scalenorm".
-
layernorm_epsilon: float
Default = 1e-05
Layer norm epsilon.
-
rms_norm_epsilon: float
Default = 1e-08
Root mean squared norm epsilon
-
scalenorm_epsilon: float
Default = 1e-08
Scalenorm epsilon
-
pos_emb: typing.Literal['learned', 'rotary', 'sinusoidal', 'rpe', 'alibi', 'none']
Default = learned
Type of positional embedding to use - choose from 'learned', 'rotary', 'sinusoidal', 'rpe', 'none'
-
rpe_num_buckets: int
Default = 32
T5 relative positional encoding number of buckets, default 32.
-
rpe_max_distance: int
Default = 128
T5 relative positional encoding max distance, default 128.
-
no_weight_tying: bool
Default = False
Disables weight tying between embedding weights and final Linear layer
-
attention_config: list
Default = None
Attention configuration for gpt-neox
The first item in the list specifies the attention type(s), and should be a list of strings. The second item specifies the number of times to repeat those attention types in the full list.
attention type choices: [global, local, sparse_fixed, sparse_variable, bslongformer, bigbird]
So a 12 layer network with only global attention could be specified like: [[[
global], 12]]or a 12 layer network with alternating global / local like: [[[
global,local], 6]]If none is specified, this defaults to [[[
global], n_layers]] -
sparsity_config: dict
Default = None
Sparsity configuration dict as defined in https://www.deepspeed.ai/docs/config-json/#sparse-attention
Note that since neox is autoregressive, attention is always "unidirectional" and
horizontal_global_attentionis always false.The main difference between our sparsity config and deepspeed's is that
modeis ignored - since it is instead specified in attention_config defining each layer.An example config is given below: "sparse_attention": { "block": 16, "different_layout_per_head": true, "num_local_blocks": 4, "num_global_blocks": 1, "num_different_global_patterns": 4, "num_random_blocks": 0, "local_window_blocks": [4], "global_block_indices": [0], "global_block_end_indices": None, "num_sliding_window_blocks": 3 }
-
num_unique_layers: int
Default = None
Number of unique transformer layers. num-layers should be divisible by this value. Currently only has an effect when pipe_parallel_size=0.
-
param_sharing_style: str
Default = grouped
Ordering of the shared parameters. For example, for a num-layers=4 and --num-unique-layers=2, we will have the following ordering for two unique layers 1 and 2-: grouped: [1, 2, 1, 2] and spaced: [1, 1, 2, 2].
-
make_vocab_size_divisible_by: int
Default = 128
Pad the vocab size to be divisible by this value. This is added for computational efficiency reasons.
-
activation: typing.Literal['gelu', 'geglu', 'relu', 'softsign', 'swish', 'mish']
Default = gelu
Activation function to use - choose from ["gelu", "geglu", "relu", "softsign", "swish", "mish"]
-
scaled_upper_triang_masked_softmax_fusion: bool
Default = False
Enable fusion of query_key_value_scaling time (upper diagonal) masking and softmax.
-
scaled_masked_softmax_fusion: bool
Default = False
Enable fusion of query_key_value_scaling general masking and softmax.
-
bias_gelu_fusion: bool
Default = False
Enable bias and gelu fusion.
-
bias_dropout_fusion: bool
Default = False
Enable bias and dropout fusion.
-
fp16_lm_cross_entropy: bool
Default = False
Move the cross entropy unreduced loss calculation for lm head to fp16.
-
init_method_std: float
Default = 0.02
Standard deviation of the zero mean normal distribution used for weight initialization.
-
apply_query_key_layer_scaling: bool
Default = False
Scale Q * K^T by 1 / layer-number. If this flag is set, then it will automatically set attention-softmax-in-fp32 to true
-
use_cpu_initialization: bool
Default = False
If set, affine parallel weights initialization uses CPU
-
attention_softmax_in_fp32: bool
Default = False
Run attention masking and softmax in fp32.
-
rotary_pct: float
Default = 1.0
pct of hidden dims to apply rotary positional embedding to
-
rotary_emb_base: int
Default = 10000
Base for rotary positional embedding
-
init_method: typing.Literal['normal', 'scaled_normal', 'orthogonal', 'scaled_orthogonal', 'xavier_uniform', 'xavier_normal', 'wang_init', 'small_init']
Default = normal
Init function used on all layers except ff residual outputs - choose from ["normal", "scaled_normal", "orthogonal", "scaled_orthogonal", "xavier_uniform", "xavier_normal", "wang_init", "small_init"]
-
output_layer_init_method: typing.Literal['normal', 'scaled_normal', 'orthogonal', 'scaled_orthogonal', 'xavier_uniform', 'xavier_normal', 'wang_init', 'small_init']
Default = scaled_normal
Init function used for ff residual outputs - choose from ["normal", "scaled_normal", "orthogonal", "scaled_orthogonal", "xavier_uniform", "xavier_normal", "wang_init", "small_init"]
-
gmlp_attn_dim: int
Default = 64
the dimension of the single head self attention in gmlp model (not used in gpt models). If None - gmlp model doesn't use attention.
-
gpt_j_residual: bool
Default = False
If false, we use the conventional residual path: x = x + attn(ln1(x)) x = x + mlp(ln2(x)) Otherwise, we use the residual path from GPT-J, which offers a slight speedup: x = ln(x) x = x + attn(x) + mlp(x)
-
soft_prompt_tuning: dict
Default = None
Dictionary configuring the soft prompt tuning parameters. If enabled, will train only the soft prompt, and freezes the rest of the model. parameters in the dict are: 'enabled': bool = True # enables soft prompting 'num_tokens': int = 10 # length of the soft prompt in tokens 'init_string': str = '' # if provided, initialize the soft prompt with the word embeddings of this string 'init_range': float = 0.5 # if no init string is provided, initialize the soft prompt with a uniform distribution between -init_range and init_rang
-
output_layer_parallelism: typing.Literal['row', 'column']
Default = row
Parameter controlling whether the output layer is parallelized over the hidden dim (row) or the vocab dim (column)
Optimizer Arguments
-
optimizer_type: typing.Literal['adam', 'onebitadam', 'cpu_adam', 'cpu_torch_adam', 'sm3', 'madgrad_wd']
Default = adam
Type of optimizer to use. Choose from ['adam', 'onebitadam', 'cpu_adam', 'cpu_torch_adam', 'sm3', 'madgrad_wd]
-
use_bnb_optimizer: bool
Default = False
Whether to enable the bitsandbytes optimizers
-
zero_stage: int
Default = None
Zero Optimizer stage
-
zero_reduce_scatter: bool
Default = None
Zero: Uses reduce or reduce scatter instead of allreduce to average gradients
-
zero_contiguous_gradients: bool
Default = None
Zero: Copies the gradients to a contiguous buffer as they are produced. Avoids memory fragmentation during backward pass. Only useful when running very large models.
-
zero_reduce_bucket_size: int
Default = None
Zero: Number of elements reduced/allreduced at a time. Limits the memory required for the allgather for large model sizes
-
zero_allgather_bucket_size: int
Default = None
Zero: Number of elements allgathered at a time. Limits the memory required for the allgather for large model sizes
-
lr: float
Default = None
Max Learning rate during training
Misc. Arguments
-
distributed_backend: str
Default = nccl
Which backend to use for distributed training.
-
local_rank: int
Default = None
local rank passed from distributed launcher.
-
rank: int
Default = None
global rank of process being run (passed in via distributed launcher)
-
lazy_mpu_init: bool
Default = False
If set to True, initialize_megatron() skips DDP initialization and returns function to complete it instead. Also turns on use-cpu-initialization flag. This is for external DDP manager.
-
short_seq_prob: float
Default = 0.1
Probability of producing a short sequence.
-
eod_mask_loss: bool
Default = False
Mask loss for the end of document tokens.
-
adlr_autoresume: bool
Default = False
Enable auto-resume on adlr cluster.
-
adlr_autoresume_interval: int
Default = 1000
Intervals over which check for auto-resume termination signal
-
seed: int
Default = 1234
Random seed used for python, numpy, pytorch, and cuda.
-
onnx_safe: bool
Default = False
Use workarounds for known problems with Torch ONNX exporter
-
deepscale: bool
Default = False
(Deprecated) enable DeepSpeed (helper flag for user code, no impact on DeepSpeed backend)'
-
deepscale_config: str
Default = None
(Deprecated) deepscale json configuration file.
-
deepspeed_mpi: bool
Default = False
Run via MPI, this will attempt to discover the necessary variables to initialize torch distributed from the MPI environment
-
user_script: str
Default = None
user script to be run
-
iteration: int
Default = None
Set during training
-
do_train: int
Default = None
Set during training
-
do_valid: int
Default = None
Set during training
-
do_test: int
Default = None
Set during training
-
global_num_gpus: int
Default = None
Set during launching
Parallelism Arguments
-
pipe_parallel_size: int
Default = 0
Number of pipeline parallel stages. Disable with 0.
-
model_parallel_size: int
Default = 1
Size of the model parallelism.
-
pipe_partition_method: str
Default = type:transformer|mlp
method used to distribute model layers across pipeline stages. Choose from "parameters", which balances the number of parameters on each pipeline stage, "uniform", which naively balances the number of layers per stage, or "type:[regex]", which balances layers whose class names match [regex]
-
world_size: int
Default = None
Total world size (i.e number of gpus in cluster). Configured post-launch using distributed launcher
-
is_pipe_parallel: bool
Default = False
flag to determine whether pipeline parallelism is on - shouldn't be set by user, is automatically determined according to pipeline parallel size.
NeoXArgsTemplate()
Text Generation arguments
-
text_gen_type: str
Default = unconditional
How to generate text/sample the model. Options:
unconditional,input-file,interactive -
temperature: float
Default = 0.0
exponential scaling output distribution ("higher == more risk")
-
top_p: float
Default = 0.0
Top-p (nucleus) sampling chooses from the smallest possible set of tokens whose cumulative probability exceeds the probability top_p.
-
top_k: int
Default = 0
integer between 0 and the models vocab size. Filters out any logits with a probability less than that of the top_kth token.
-
maximum_tokens: int
Default = 64
maximum number of tokens to be generated
-
sample_input_file: str
Default = None
Get input from file instead of interactive mode, each line is an input.
-
sample_output_file: str
Default = samples.txt
Output file
-
num_samples: int
Default = 1
Number of samples to generate unconditionally, defaults to 1 and interactive conditional sampling
-
recompute: bool
Default = False
During generation recompute all attention instead of using previously computed keys/values. Should be set to true for sparse attention models
-
eval_results_prefix: str
Default =
prefix to which to save evaluation results - final fp will be {eval_results_prefix}_eval_results_yy-mm-dd-HH-MM.json
-
eval_tasks: list
Default = None
Tasks to evaluate on using lm_eval_harness
Tokenizer Arguments
-
tokenizer_type: typing.Literal['GPT2BPETokenizer', 'HFTokenizer', 'HFGPT2Tokenizer', 'SPMTokenizer', 'CharLevelTokenizer']
Default = GPT2BPETokenizer
Type of tokenizer to use - should be one of ["GPT2BPETokenizer", "HFTokenizer", "HFGPT2Tokenizer", "SPMTokenizer", "CharLevelTokenizer"]
-
padded_vocab_size: int
Default = None
Total (padded) vocabulary size of tokenizer. Configured after launching of training, as it's dependent on the parallelism size.
Training Arguments
-
data_path: str
Default = None
Path to combined dataset to split.
-
train_data_paths: list
Default = None
List of paths to train datasets.
-
test_data_paths: list
Default = None
List of paths to test datasets.
-
valid_data_paths: list
Default = None
List of paths to validation datasets.
-
train_data_weights: list
Default = None
List of 'weights' that decide how often to sample from each training dataset when blending datasets. If None, defaults to equal weighting. Should be a list the same length as
train_data_paths -
valid_data_weights: list
Default = None
List of 'weights' that decide how often to sample from each validation dataset when blending datasets. If None, defaults to equal weighting. Should be a list the same length as
valid_data_paths -
test_data_weights: list
Default = None
List of 'weights' that decide how often to sample from each test dataset when blending datasets. If None, defaults to equal weighting. Should be a list the same length as
test_data_paths -
weight_by_num_documents: bool
Default = False
If True, Builds dataset weights from a multinomial distribution over groups of data according to the number of documents in each group.
WARNING: setting this to True will override any user provided weights
We sample from a group according to the probability p(L) ∝ |L| ** α, where p(L) is the probability of sampling from a given group, |L| is the number of examples in that datapoint, and α is a coefficient that acts to upsample data from underrepresented groups
Hence α (
alpha) allows us to control how much to 'boost' the probability of training on low-resource groups.See https://arxiv.org/abs/1911.02116 for more details
-
weighted_sampler_alpha: float
Default = 0.3
Alpha value for
weight_by_num_documents. Only has an effect ifweight_by_num_documents= True.when alpha = 1, the probability of sampling from a given group = n_samples / total_samples as alpha -> 0, the probability of sampling from all groups becomes equal, and number of documents has no effect as alpha -> inf, the probability of sampling from the groups with the most samples -> 1
-
data_impl: str
Default = infer
Implementation of indexed datasets.
-
mmap_warmup: bool
Default = False
Warm up mmap files.
-
save: str
Default = None
Output directory to save checkpoints to.
-
config_files: dict
Default = None
Store of original config files mapping config filename to file contents
-
load: str
Default = None
Directory containing a model checkpoint.
-
checkpoint_validation_with_forward_pass: bool
Default = False
save input and output of a forward pass with the checkpoint and validate after load
-
save_interval: int
Default = None
Number of iterations between checkpoint saves.
-
no_save_optim: bool
Default = False
Do not save current optimizer.
-
no_save_rng: bool
Default = False
Do not save current rng state.
-
no_load_optim: bool
Default = False
Do not load optimizer when loading checkpoint.
-
no_load_rng: bool
Default = False
Do not load rng state when loading checkpoint.
-
finetune: bool
Default = False
Load model for finetuning. Do not load optimizer or rng state from checkpoint and set iteration to 0. Assumed when loading a release checkpoint.
-
batch_size: int
Default = None
training microbatch size per gpu
-
train_iters: int
Default = None
Number of iterations to run for training.
-
eval_iters: int
Default = 100
Number of iterations to run for evaluation validation/test for.
-
keep_last_n_checkpoints: int
Default = None
Number of last checkpoints to keep
-
eval_interval: int
Default = 1000
Interval between running evaluation on validation set.
-
split: str
Default = 969, 30, 1
Comma_separated list of proportions for training, validation, and test split. For example the split 90,5,5 will use 90% of data for training, 5% for validation and 5% for test.
-
vocab_file: str
Default = None
Path to the vocab file.
-
merge_file: str
Default = None
Path to the BPE merge file.
-
num_workers: int
Default = 2
Dataloader number of workers.
-
exit_interval: int
Default = None
Exit the program after the iteration is divisible by this value.
-
attention_dropout: float
Default = 0.1
Post attention dropout probability.
-
hidden_dropout: float
Default = 0.1
Dropout probability for hidden state transformer.
-
weight_decay: float
Default = 0.01
Weight decay coefficient for L2 regularization.
-
checkpoint_activations: bool
Default = False
Checkpoint activation to allow for training with larger models, sequences, and batch sizes.
-
checkpoint_num_layers: int
Default = 1
Chunk size (number of layers) for checkpointing.
-
deepspeed_activation_checkpointing: bool
Default = True
DEPRECATED - TODO: remove Uses activation checkpointing from deepspeed
-
contiguous_checkpointing: bool
Default = False
Contiguous memory checkpointing for activations.
-
checkpoint_in_cpu: bool
Default = False
Move the activation checkpoints to CPU.
-
synchronize_each_layer: bool
Default = False
does a synchronize at the beginning and end of each checkpointed layer.
-
profile_backward: bool
Default = False
Enables backward pass profiling for checkpointed layers.
-
partition_activations: bool
Default = False
Partition Activations across GPUs before checkpointing.
-
gas: int
Default = None
gradient_accumulation_steps
-
clip_grad: float
Default = None
Gradient clipping based on global L2 norm.
-
hysteresis: int
Default = 2
hysteresis for dynamic loss scaling
-
dynamic_loss_scale: bool
Default = None
flag indicating whether dynamic loss scale is used
-
loss_scale: float
Default = None
Static loss scaling, positive power of 2 values can improve fp16 convergence. If None, dynamic loss scaling is used.
-
loss_scale_window: float
Default = 1000.0
Window over which to raise/lower dynamic scale.
-
min_scale: float
Default = 1.0
Minimum loss scale for dynamic loss scale.
-
char_level_ppl: bool
Default = False
Whether to calculate character level perplexity as well as token level perplexity. (may incur a time cost)
Args for deepspeed config Every argument included here will be included in deepspeed config json #TODO this list is not complete as compared to https://www.deepspeed.ai/docs/config-json/
-
deepspeed: bool
Default = True
boolean flag to enable DeepSpeed (Always True)
-
train_batch_size: int
Default = None
The effective training batch size. This is the amount of data samples that leads to one step of model update. train_batch_size is aggregated by the batch size that a single GPU processes in one forward/backward pass (a.k.a., train_step_batch_size), the gradient accumulation steps (a.k.a., gradient_accumulation_steps), and the number of GPUs.
-
train_micro_batch_size_per_gpu: int
Default = None
Batch size to be processed by one GPU in one step (without gradient accumulation). When specified, gradient_accumulation_steps is automatically calculated using train_batch_size and number of GPUs. Should not be concurrently specified with gradient_accumulation_steps in the configuration JSON.
-
gradient_accumulation_steps: int
Default = 1
Number of training steps to accumulate gradients before averaging and applying them. This feature is sometimes useful to improve scalability since it results in less frequent communication of gradients between steps. Another impact of this feature is the ability to train with larger batch sizes per GPU. When specified, train_step_batch_size is automatically calculated using train_batch_size and number of GPUs. Should not be concurrently specified with train_step_batch_size in the configuration JSON.
-
optimizer: dict
Default = None
dict containing the keys type and params
type: The optimizer name. DeepSpeed natively supports Adam, AdamW, OneBitAdam, Lamb, and OneBitLamb optimizers (See here for details) and will import other optimizers from torch.
params: Dictionary of parameters to instantiate optimizer. The parameter names must match the optimizer constructor signature (e.g., for Adam).
-
scheduler: dict
Default = None
dict containing the keys type and params
type: The scheduler name. See here (https://deepspeed.readthedocs.io/en/latest/schedulers.html) for list of support schedulers.
params: Dictionary of parameters to instantiate scheduler. The parameter names should match scheduler constructor signature.
-
fp32_allreduce: bool
Default = False
During gradient averaging perform allreduce with 32 bit values
-
prescale_gradients: bool
Default = False
Scale gradients before doing allreduce
-
gradient_predivide_factor: float
Default = 1.0
Before gradient averaging predivide gradients by a specified factor, can sometimes help with fp16 stability when scaling to large numbers of GPUs
-
sparse_gradients: bool
Default = False
Enable sparse compression of torch.nn.Embedding gradients.
-
fp16: dict
Default = None
Configuration for using mixed precision/FP16 training that leverages NVIDIA’s Apex package.
-
amp: dict
Default = None
Dictionary as described in Deepspeed documentation: https://www.deepspeed.ai/docs/config-json/#automatic-mixed-precision-amp-training-options
-
gradient_clipping: float
Default = 0.0
Enable gradient clipping with provided value
-
zero_optimization: dict
Default = None
-
steps_per_print: int
Default = 10
Print train loss every N steps.
-
wall_clock_breakdown: bool
Default = False
Enable timing of the latency of forward/backward/update training phases.
-
dump_state: bool
Default = False
Print out state information of DeepSpeed object after initialization.
-
flops_profiler: dict
Default = None
Dictionary as described in Deepspeed documentation: https://www.deepspeed.ai/docs/config-json/#flops-profiler
-
zero_allow_untested_optimizer: bool
Default = False
Whether Deepspeed Zero Optimizer will allow an optimizer that hasn't been tested by the deepspeed team
Args for deepspeed runner (deepspeed.launcher.runner). Every argument included here will be passed as command line argument to deepspeed.launcher.runner
-
hostfile: str
Default = None
list of hostnames / ssh aliases and the number of GPUs per host
example file contents: worker-1 slots=4 worker-2 slots=4 127.0.0 slots=4 127.0.1 slots=4
-
include: str
Default = None
Specify hardware resources to use during execution. String format is
NODE_SPEC[@NODE_SPEC ...]whereNODE_SPEC=NAME[:SLOT[,SLOT ...]]. If:SLOTis omitted, include all slots on that host. Example:"worker-0@worker-1:0,2"will use all slots. onworker-0and slots[0, 2]onworker-1. -
exclude: str
Default = None
Specify hardware resources to NOT use during execution. Same format as include
-
num_nodes: int
Default = -1
Total number of worker nodes to run on, this will use the top N hosts from the given hostfile. -1 will use all.
-
num_gpus: int
Default = None
Max number of GPUs to use on each node, will use [0:N) GPU ids on each node. None / not specifying a value will use all.
-
master_port: int
Default = 29500
Port used by PyTorch distributed for communication during training.
-
master_addr: str
Default = None
IP address of node 0, will be inferred via 'hostname -I' if not specified.
-
launcher: str
Default = pdsh
Launcher backend for multi-node training. Options currently include PDSH, OpenMPI, MVAPICH.
-
detect_nvlink_pairs: bool
Default = False
If true, autodetects nvlink pairs and remaps cuda visible devices to place them next to each other. This is an Eleuther addition to deepspeed, and should speed up model parallel training on setups with nvlink pairs when mp=2.