"optimizer": {
"type": "OneBitAdam",
"params": {
"lr": 0.001,
"betas": [
0.8,
0.999
"eps": 1e-8,
"weight_decay": 3e-7,
"freeze_step": 400,
"cuda_aware": false,
"comm_backend_name": "nccl"
The 1-bit Adam optimizer supports the following three params keys/values in addition to the standard Adam (learn more in our tutorial):
freeze_step
Number of warm up steps before 1-bit compression gets applied to the communication
100000
cuda_aware
To indicate that the underlying MPI library supports CUDA-Aware communication
false
comm_backend_name
To indicate which backend implementation to use
“nccl”
A variant optimizer for 1-bit Adam is 0/1 Adam, which further optimizes 1-bit Adam via adaptive variance freezing and 1-bit synchronization over optimizer states.
"optimizer": {
"type": "ZeroOneAdam",
"params": {
"lr": 1e-3,
"weight_decay": 0.01,
"bias_correction": false,
"var_freeze_step": 1000,
"var_update_scaler": 16,
"local_step_scaler": 1000,
"local_step_clipper": 16,
"cuda_aware": false,
"comm_backend_name": "nccl"
0/1 Adam supports the following params key/values in addition to standard Adam (learn more in our tutorial.)
local_step_scaler
The interval to scale the local steps interval according to the learning rate policy
32678
local_step_clipper
The largest interval for local steps with learning rate policy
cuda_aware
To indicate that the underlying MPI library supports CUDA-Aware communication
false
comm_backend_name
To indicate which backend implementation to use
“nccl”
max_coeff
Scaling coefficient upper bound for original LAMB algorithm and 1-bit LAMB’s warmup stage
min_coeff
Scaling coefficient lower bound for original LAMB algorithm and 1-bit LAMB’s warmup stage
freeze_step
Number of warm up steps before 1-bit compression gets applied to the communication
100000
cuda_aware
To indicate that the underlying MPI library supports CUDA-Aware communication
false
comm_backend_name
To indicate which backend implementation to use
“nccl”
coeff_beta
Coefficient used for computing running averages of lamb coefficient
factor_max
Maximum value of scaling factor to the frozen lamb coefficient during compression stage
factor_min
Minimum value of scaling factor to the frozen lamb coefficient during compression stage
factor_threshold
Threshold of how much the scaling factor can fluctuate between steps
Scheduler Parameters
DeepSpeed calls the step() method of the scheduler at every training step when model_engine.step() is executed.
scheduler: [dictionary]
params
Dictionary of parameters to instantiate scheduler. The parameter names should match scheduler constructor signature.
{"warmup_min_lr": 0, "warmup_max_lr": 0.001}
During gradient averaging perform communication with selected data type. By default it will be determined by selected regime
Before gradient averaging predivide gradients by a specified factor, can sometimes help with fp16 stability when scaling to large numbers of GPUs
Enable sparse compression of torch.nn.Embedding gradients. This feature is essentially deprecated as we don’t see use cases for it as much anymore. It should be noted that this feature is not compatible with torch.sparse related features.
false
FP16 training options
Note: this mode cannot be combined with the amp mode described below.
fp16: [dictionary]
Configuration for using mixed precision/FP16 training that leverages NVIDIA’s Apex package. An example, including the available dictionary keys is illustrated below. NOTE: this does not use Apex’s AMP mode that allows for more flexibility in mixed precision training modes, this mode is similar to AMP’s O2 mode. Please see AMP support below if you want to use more complex mixed precision modes. If you want to use ZeRO (currently) you must use this mode.
loss_scale is a fp16 parameter representing the loss scaling value for FP16 training. The default value of 0.0 results in dynamic loss scaling, otherwise the value will be used for static fixed loss scaling.
initial_scale_power is a fp16 parameter representing the power of the initial dynamic loss scale value. The actual loss scale is computed as 2initial_scale_power.
loss_scale_window is a fp16 parameter representing the window over which to raise/lower the dynamic loss scale value.
consecutive_hysteresis is a fp16 parameter representing whether to refill the hysteresis if we reach an iteration that doesn’t overflow
false
BFLOAT16 training options
Note: this mode cannot be combined with the amp mode described below.
Note: this mode cannot be combined with the fp16 mode described above.
bf16: [dictionary]
Configuration for using bfloat16 floating-point format as an alternative to FP16. BFLOAT16 requires hardware support (e.g., NVIDIA A100). An example, including the available dictionary keys is illustrated below. Training with bfloat16 does not require loss scaling.
Automatic mixed precision (AMP) training options
Note: this mode cannot be combined with the fp16 mode described above. In addition this mode is not currently compatible with ZeRO.
amp: [dictionary]
Configuration for using automatic mixed precision (AMP) training that leverages NVIDIA’s Apex AMP package. An example, including the available dictionary keys is illustrated below. Is not compatible with fp16 mode above or ZeRO. Any parameters outside of “enabled” will be passed to AMP’s initialize call, see the API and descriptions here at the apex.amp.initialize documentation.
Any parameters outside of “enabled” will be passed to AMP’s initialize call, see the API and descriptions here at the apex.amp.initialize documentation.
Enabling and configuring ZeRO memory optimizations
"zero_optimization": {
"stage": [0|1|2|3],
"allgather_partitions": [true|false],
"allgather_bucket_size": 5e8,
"overlap_comm": false,
"reduce_scatter": [true|false],
"reduce_bucket_size": 5e8,
"contiguous_gradients" : [true|false],
"offload_param": {
"offload_optimizer": {
"stage3_max_live_parameters" : 1e9,
"stage3_max_reuse_distance" : 1e9,
"stage3_prefetch_bucket_size" : 5e8,
"stage3_param_persistence_threshold" : 1e6,
"sub_group_size" : 1e12,
"elastic_checkpoint" : [true|false],
"stage3_gather_16bit_weights_on_model_save": [true|false],
"ignore_unused_parameters": [true|false]
"round_robin_gradients": [true|false]
"zero_hpz_partition_size": 1
"zero_quantized_weights": [true|false]
"zero_quantized_gradients": [true|false]
zero_optimization: [dictionary]
Chooses different stages of ZeRO Optimizer. Stage 0, 1, 2, and 3 refer to disabled, optimizer state partitioning, and optimizer+gradient state partitioning, and optimizer+gradient+parameter partitioning, respectively.
Chooses between allgather collective or a series of broadcast collectives to gather updated parameters from all the GPUs at the end of each step
Number of elements allgathered at a time. Limits the memory required for the allgather for large model sizes
Number of elements reduced/allreduced at a time. Limits the memory required for the allgather for large model sizes
Copies the gradients to a contiguous buffer as they are produced. Avoids memory fragmentation during backward pass.
Initialize fp32 master weights from fp32 copies in checkpoint (no precision loss) or from model’s fp16 copies (with precision loss). This can be used to initialize optimizer state even when checkpoint is missing optimizer state.
For use with ZeRO stage 1, enable backward hooks to reduce gradients during the backward pass or wait until the end of the backward pass.
Stage 1 and 2 optimization for CPU offloading that parallelizes gradient copying to CPU memory among ranks by fine-grained gradient partitioning. Performance benefit grows with gradient accumulation steps (more copying between optimizer steps) or GPU count (increased parallelism).
False
Enable offloading of model parameters to CPU or NVMe. This frees up GPU memory for larger models or batch sizes. Valid only with stage 3. See here for more details.
False
Enable offloading of optimizer state to CPU or NVMe, and optimizer computation to CPU. This frees up GPU memory for larger models or batch sizes. Valid for ZeRO stage 1, 2, 3. See here for more details.
False
The maximum number of parameters resident per GPU before releasing. Smaller values use less memory, but perform more communication.
Do not release a parameter if it will be reused within this threshold of parameters. Smaller values use less memory, but perform more communication.
The size of the fixed buffer for prefetching parameters. Smaller values use less memory, but can increase stalls due to communication.
Do not partition parameters smaller than this threshold. Smaller values use less memory, but can greatly increase communication (especially latency-bound messages).
Consolidate the weights before saving the model by save_16bit_model(). Since the weights are partitioned across GPUs, they aren’t part of state_dict, so this function automatically gathers the weights when this option is enabled and then saves the fp16 model weights.
False
Number of ranks in hiearchical partitioning ZeRO (hpZ) secondary tensor group of ZeRO++, default is 1 meaning no hpZ, ideal is number of ranks (gpus) per node.
Boolean indicating whether to enable communication efficient quantized weights of ZeRO++.
False
Boolean indicating whether to enable communication efficient quantized gradients of ZeRO++.
False
Enable offloading of optimizer memory and computation to CPU. This frees up GPU memory for larger models or batch sizes. Valid with stage 1 and 2.
False
Parameter offloading
Enabling and configuring ZeRO optimization of parameter offloading to CPU/NVMe. Available only with ZeRO stage 3.
Note that if the value of “device” is not specified or not supported, an assertion will be triggered.
"offload_param": {
"device": "[cpu|nvme]",
"nvme_path": "/local_nvme",
"pin_memory": [true|false],
"buffer_count": 5,
"buffer_size": 1e8,
"max_in_cpu": 1e9
device: [string]
Offload to page-locked CPU memory. This could boost throughput at the cost of extra memory overhead.
false
Number of parameter elements to maintain in CPU memory when offloading to NVMe is enabled.
Optimizer offloading
Enabling and configuring ZeRO optimization of offloading optimizer computation to CPU and state to CPU/NVMe. CPU offloading is available with ZeRO stage 1, 2, 3. NVMe offloading is available only with ZeRO stage 3.
Note that if the value of “device” is not specified or not supported, an assertion will be triggered.
"offload_optimizer": {
"device": "[cpu|nvme]",
"nvme_path": "/local_nvme",
"pin_memory": [true|false],
"ratio": 0.3,
"buffer_count": 4,
"fast_init": false
device: [string]
Device memory to offload optimizer state. Supported options are cpu and nvme. Optimizer computation is offload to CPU regardless of device option.
Offload to page-locked CPU memory. This could boost throughput at the cost of extra memory overhead.
false
Number of buffers in buffer pool for optimizer state offloading to NVMe. This should be at least the number of states maintained per parameter by the optimizer. For example, Adam optimizer has 4 states (parameter, gradient, momentum, and variance).
Asynchronous I/O
Configuring the asynchronous I/O module for offloading parameter and optimizer states to persistent (NVMe) storage. This module uses Linux native asynchronous I/O (libaio).
"aio": {
"block_size": 1048576,
"queue_depth": 8,
"thread_count": 1,
"single_submit": false,
"overlap_events": true
block_size: [integer]
Submit requests to storage device as multiple individual requests as opposed to one block of requests.
false
Submit requests to storage device in an overlapped fashion without waiting for completion of earlier requests.
Unused parameters in modules may be unexpected in static networks, but could be normal in dynamic networks. This controls whether or not training should terminate with an error message when unused parameters are detected. This is set to True by default, which means unused parameters are ignored and training continues. Now is just used in stage 2.
Print progress report every N training steps. The report includes the number of training steps, number of skipped optimizer updates (likely due to overflows in mixed-precision training), current learning rate, and current momentum.
"enabled": false,
"results_dir": "autotuning_results",
"exps_dir": "autotuning_exps",
"overwrite": false,
"metric": "throughput",
"start_profile_step": 3,
"end_profile_step": 5,
"fast": true,
"max_train_batch_size": null,
"mp_size": 1,
"num_tuning_micro_batch_sizes": 3,
"tuner_type": "model_based",
"tuner_early_stopping": 5,
"tuner_num_trials": 50,
"arg_mappings": null
enabled: [boolean]
Path to the autotuning experiment results directory. The default appears in the working directory from which Deepspeed was launched.
“autotuning_results”
Path to the auotuning experiment descriptions directory. The default appears in the working directory from which Deepspeed was launched.
“autotuning_exps”
Whether to run autotuning experiments whose results already exist. Setting it to true would overwrite the existing result.
false
The performance metric to use for ranking autotuning experiments. latency, throughput, and FLOPS are currently supported, referring to training step latency, training samples per second, and floating-point operations per second achieved per GPU respectively.
throughput
The global training step at which to start profiling in an autotuning experiment. Note that warm-up is needed for accurate performance measurement.
The global training step at which to end profiling in an autotuning experiment. Must not be less than start_profile_step.
Enables fast-model autotuning where only Zero stages and micro-batch sizes per GPU are tuned.
The number of experiments to run beyond the current best experiment. If no better experiment is found within that number, the Autotuner stops the exploration.
The global training step at which to profile. Note that warm up steps are needed for accurate time measurement.
The depth of the model at which to print the aggregated module information. When set to -1, it prints information from the top module to the innermost modules (the maximum depth).
Activation Checkpointing
"activation_checkpointing": {
"partition_activations": false,
"cpu_checkpointing": false,
"contiguous_memory_optimization": false,
"number_checkpoints": null,
"synchronize_checkpoint_boundary": false,
"profile": false
partition_activations: [boolean]
Total number of activation checkpoints used to allocate memory buffer for contiguous_memory_optimization
A string determining sparsity structure type. Deepspeed currently supports "dense", "fixed", "bigbird", "bslongformer", and "variable".
"fixed"
block
An integer determining the block size. Current implementation of sparse self-attention is based on blocked sparse matrices. In which this parameter defines size of such blocks, Block X Block.
different_layout_per_head
A boolean determining if each head should be assigned a different sparsity layout; this will be satisfied based on availability.
false
num_local_blocks
An integer determining the number of random blocks in each block row; only used in "fixed" mode.
num_global_blocks
An integer determining how many consecutive blocks in a local window is used as the representative of the window for global attention; used in "fixed" and "bigbird" modes.
attention
A string determining attention type. Attention can be "unidirectional", such as autoregressive models, in which tokens attend only to tokens appear before them in the context. Considering that, the upper triangular of attention matrix is empty. Or it can be "bidirectional", such as BERT, in which tokens can attend to any other tokens before or after them. Then, the upper triangular part of the attention matrix is mirror of the lower triangular; used in "fixed" and "variable" modes.
"bidirectional"
horizontal_global_attention
A boolean determining if blocks that are global representative of a local window, also attend to all other blocks. This is valid only if attention type is "bidirectional". Looking at the attention matrix, that means global attention not only includes the vertical blocks, but also horizontal blocks; used in "fixed" and "variable" modes.
false
num_different_global_patterns
An integer determining number of different global attentions layouts. While global attention can be fixed by which block/s are representative of any local window, since there are multi-heads, each head can use a different global representative; used only in "fixed" mode.
num_random_blocks
An integer determining the number of random blocks in each block row; used in "variable" and "bigbird" modes.
local_window_blocks
A list of integers determining the number of blocks in each local attention window. It assumes first number determines # of blocks in the first local window, second the second window, …, and the last number determines the number of blocks in the remaining local windows; only used in "variable" mode.
global_block_indices
A list of integers determining which blocks are considered as global attention. Given indices, determine the blocks that all other token blocks attend to and they attend to all other token blocks. Notice that if global_block_end_indices parameter is set, this parameter is used as starting index of each global window; used in "variable" and "bslongformer" modes.
global_block_end_indices
A list of integers determining end indices of global window blocks. By default this is not used. But if it is set, it must have the same size of global_block_indices parameter, and combining this two parameters, for each index i, blocks from global_block_indices[i] to global_block_end_indices[i], exclusive, are considered as global attention; used in "variable" and "bslongformer" modes.
num_sliding_window_blocks
An integer determining the number of blocks in sliding local attention window; used in "bigbird" and "bslongformer" modes.
"num_global_blocks": 1,
"attention": "bidirectional",
"horizontal_global_attention": false,
"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
Data Efficiency
DeepSpeed Data Efficiency Library includes two techniques: curriculum learning and random layerwise token dropping (random-LTD). Read more about how to use the DeepSpeed Data Efficiency Library in our tutorial.
"data_efficiency": {
"enabled": true,
"seed": 1234,
"data_routing": {
"enabled": true,
"random_ltd":{
"enabled": true,
"total_layer_num": 24,
"random_ltd_layer_num": 22,
"random_ltd_layer_id": [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22],
"model_mask_name": "attention_mask",
"model_type": "decoder",
"hidden_state_order": "seq_batch_dim",
"random_ltd_schedule": {
"min_value": 128,
"max_value": 2048,
"schedule_type":"fixed_linear",
"schedule_config": {
"require_steps": 200000,
"seq_per_step": 16
"data_sampling": {
"enabled": true,
"num_epochs": 1,
"num_workers": 0,
"curriculum_learning": {
"enabled": true,
"data_cluster_path": "/path/to/data_clusters",
"curriculum_metrics": {
"vocabularyrarity": {
"index_to_sample_path": "/path/to/index_to_sample",
"index_to_metric_path": "/path/to/index_to_metric",
"difficulty_type": "percentile",
"clustering_type": "schedule_based",
"min_difficulty": 1,
"max_difficulty": 100,
"schedule_type": "fixed_root",
"schedule_config": {
"total_curriculum_step": 110000,
"difficulty_step": 1,
"root_degree": 2
data_efficiency: [dictionary]
num_epochs: [integer]
At most how many epoches of the original dataset will be iterated.
num_workers: [integer]
Data loader number of workers.
curriculum_learning: [dictionary]
Configs for curriculum learing technique.
total_layer_num: [integer]
The number of layer (or the depth) for the pretraining/fine-tuning model.
random_ltd_layer_num: [integer]
The number of layers that will be applied with random-LTD.
random_ltd_layer_id: [list]
The exact layer_id that will be applied with random-LTD. The length of this list must be the same as random_ltd_layer_num.
model_mask_name: [str]
The variable name of the attention_mask. Different libraries have different names, such as att_mask. For huggingface model, it’s named “attention_mask”. Users need to check the forward function in the original model files. If the attention mask input in the original model’s forward function is not a keyword/named argument (e.g., attention_mask=None), user would need to change it to a keyword/named argument and provide that keyword as model_mask_name.
model_type: [str]
Users need to identify whether the model is decoder or encoder. Currently we only support these two.
hidden_state_order: [str]
Users need to know the input order of the hidden state tensor. Normally, it’s batch, sequence and then the hidden dimension, which is batch_seq_dim. Somethings, the order between batch and sequence will be switch like seq_batch_dim. Currently, we support these two.
random_ltd_schedule: [dictionary]
The schedule of the effective sequence length after token dropping. It’s a linear function where random-LTD gradually drops less tokens and increases effective sequence length.
min_value: [integer]
The initial effective sequence length (after token dropping) at step/iteration 0.
max_value: [integer]
The max effective sequence length (usually the case without any token dropping). Usually this is set as baseline’s seqlen.
schedule_type: [str]
The sequence length follows a linear increasing function starting from min_value and reaching max_value. We currently only support this type.
schedule_config: [dictionary]
Configs for the linear increasing function.
require_steps: [integer]
How many iterations will be needed to reach max_value from min_value.
seq_per_step: [integer]
At any time, the effective sequence length be multiple of this seq_per_step. Set this to multiple of 8 (for FP16 data) or 16 (for INT8 data) to enable NVIDIA Tensor Core acceleration.
data_cluster_path: [str]
Path to directory where curriculum learning will store the indexes of data samples within the same difficulty ranges.
curriculum_metrics: [dictionary]
This dictionary includes all desired curriculum metrics and their configs. Each metric will be a separate sub-dictionary, where the key is the metric name and the values are configs below.
index_to_sample_path: [str]
Path to the index_to_sample file generated during offline data analysis. Note that data analysis will generate two kinds of index_to_sample files: The metric_name_index_to_sample_percentile_merged file is a concatenated index for perf improvement, but it only works when you set difficulty_type=percentile. If you use difficulty_type=value, you need to change this to use the metric_name_index_to_sample file.
index_to_metric_path: [str]
Path to the index_to_metric_path file generated during offline data analysis.
difficulty_type: [str]
During training, how to increase the max accepted difficulty. Currently support value (increase by absolute value) and percentile (increase by difficulty percentile).
clustering_type: [str]
Currently support schedule_based (cluster data based on the difficulty schedule (pacing function) below) and single_cluster (no clustering required and probably CL is achieved by data postprocessing, such as sequence length truncation).
min_difficulty: [integer]
Starting difficulty at first step. When difficulty_type=value the min_difficulty is an absolute difficulty value. When difficulty_type=percentile the min_difficulty is a difficulty percentile value.
max_difficulty: [integer]
Final max difficulty. When difficulty_type=value the max_difficulty is an absolute difficulty value. When difficulty_type=percentile the max_difficulty is a difficulty percentile value.
schedule_type: [str]
The difficulty schedule (pacing function) that defines how the max accepted difficulty increases from min_difficulty to max_difficulty during training. Currently support fixed_linear, fixed_root, fixed_discrete, and custom.
schedule_config: [dictionary]
Configs for the pacing function. When schedule_type=custom this dictionary is not necessary. Instead user needs to provide a callback function (via the set_custom_curriculum_learning_schedule API in deepspeed/runtime/engine.py) which will update the max accepted difficulty during training. Configs below are all belongs to schedule_config.
total_curriculum_step: [integer]
How many steps the curriculum learning takes to go from min difficulty to max difficulty. Used by fixed_linear and fixed_root schedule.
difficulty_step: [integer]
The max accepted difficulty level determined every step must be a multiple of this difficulty_step. This is used to ensure the use of NVIDIA Tensor Core acceleration (requires multiple of 8 (FP16) or 16 (INT8)). Used by fixed_linear and fixed_root schedule.
root_degree: [integer]
The degree of the root function. Degree of 2 means square root and degree of 3 means cube root. Degree of 1 is equivalent to linear. Used by fixed_root schedule.
difficulty: [list]
List of max accepted difficulty levels to be used during schedule. Used by fixed_discrete schedule.
max_step: [list]
List of which step to change max accepted difficulty level. Used by fixed_discrete schedule.
Curriculum Learning
Note: On 12/12/2022, we released DeepSpeed Data Efficiency Library which provides a more general curriculum learning support. This legacy curriculum learning feature below is still supported but we recommend to use the Data Efficiency Library.
"curriculum_learning": {
"enabled": true,
"curriculum_type": "seqlen",
"min_difficulty": 8,
"max_difficulty": 1024,
"schedule_type": "fixed_linear",
"schedule_config": {
"total_curriculum_step": 40000,
"difficulty_step": 8
enabled: [boolean]
Type of curriculum schedule. Currently support fixed_linear, fixed_root, and fixed_discrete.
Total number of steps for the curriculum learning. One of the schedule_config when the fixed_linear and fixed_root schedule_type are used.
At any time, the curriculum learning difficulty must be multiple of this difficulty_step. Set this to multiple of 8 (for FP16 data) or 16 (for INT8 data) to enable NVIDIA Tensor Core acceleration. One of the schedule_config when the fixed_linear and fixed_root schedule_type are used.
Root degree of the curriculum schedule function. One of the schedule_config when the fixed_root schedule_type is used.
List of difficulty levels to be used during schedule. One of the schedule_config when the fixed_discrete schedule_type is used.
List of which step to change difficulty level. One of the schedule_config when the fixed_discrete schedule_type is used.
Monitoring Module
Note: Deepspeed logs to TensorBoard through PyTorch. Logging to TensorBoard requires that the tensorboard package is installed (read more in the PyTorch documentation).
Note: Logging to WandB requires that the wandb package is installed (read more in the WandB documentation).
Note: Logging to Comet requires that the comet_ml package is installed (read more in the Comet documentation).
Deepspeed’s Monitor module can log training details into a Tensorboard-compatible file, to WandB, to Comet or to simple CSV files. Below is an overview of what DeepSpeed will log automatically.
Train/Eigenvalues/ModelBlockParam_{i}
Eigen values per param block.
eigenvalue must be enabled.
Train/Samples/elapsed_time_ms_forward
The global duration of the forward pass.
flops_profiler.enabled or wall_clock_breakdown.
Train/Samples/elapsed_time_ms_backward
The global duration of the forward pass.
flops_profiler.enabled or wall_clock_breakdown.
Train/Samples/elapsed_time_ms_backward_inner
The backward time that does not include the gradient reduction time. Only in cases where the gradient reduction is not overlapped, if it is overlapped then the inner time should be about the same as the entire backward time.
flops_profiler.enabled or wall_clock_breakdown.
Train/Samples/elapsed_time_ms_backward_allreduce
The global duration of the allreduce operation.
flops_profiler.enabled or wall_clock_breakdown.
Train/Samples/elapsed_time_ms_step
The optimizer step time
flops_profiler.enabled or wall_clock_breakdown.
output_path
Path to where the Tensorboard logs will be written. If None, the output path is set under the training script’s launching path.
job_name
Name for the current job. This will become a new directory inside output_path.
"DeepSpeedJobName"
samples_log_interval
Metrics will be submitted to Comet after processing every samples_log_intervas samples.
experiment_name
The name for comet experiment to be used for logging.
api_key
Comet API key. It’s not recommended to save the Comet API Key in code.
experiment_key
The key for comet experiment to be used for logging. Must be an alphanumeric string whose length is between 32 and 50 characters.
online
If True, the data will be logged to Comet server, otherwise it will be stored locally in offline experiment. Default is True.
Control how the Comet experiment is started. “get”: Continue logging to an existing experiment identified by the experiment_key value. “create”: Always creates of a new experiment, useful for HPO sweeps. “get_or_create” (default): Starts a fresh experiment if required, or persists logging to an existing one.
"project": "my_project",
"samples_log_interval": 50,
"experiment_name": "llama-fine-tuning",
"experiment_key": "0c4a1c4a90664f2a8084e600b19a9d7",
"online": false,
"mode": "get",
csv_monitor: [dictionary]
output_path
Path to where the csv files will be written. If None, the output path is set under the training script’s launching path.
job_name
Name for the current job. This will become a new directory inside output_path
"DeepSpeedJobName"
"version": 8,
"ignore_non_elastic_batch_info": 1024,
"num_gpus_per_node": "fixed_linear",
"model_parallel_size": MODEL_PARALLEL_SIZE
micro_batch_sizes
Acceptable micro batch sizes, same as train_micro_batch_size_per_gpu
[2,4,6]
min_gpus
Min number of GPUs to search over when computing highly composite batch size in v0.1 and v0.2.
max_gpus
Max number of GPUs to search over when computing highly composite batch size in v0.1 and v0.2.
10000
min_time
Minimum running time (minutes) before the scheduler will scale again (only used in v0.1). 0 implies it’s unknown
prefer_large_batch
When finding a suitable batch size, attempt to find one that is closest to the max train batch size given.
version
Version of elastic logic to use.
ignore_non_elastic_batch_info
Ignore all batch info provided outside the elastic config. To reduce confusion, we require all batch related info to be given in elastic config only.
false
num_gpus_per_node
Number of GPUs per node. This information is used by v0.2 to support model-parallel training (only used by v0.2)
model_parallel_size
Tensor or model parallel size (only used by v0.2)
Communication Logging
DeepSpeed provides a flexible communication logging tool which can automatically detect and record communication operations launched via deepspeed.comm. NOTE: All logging communication calls are synchronized in order to provide accurate timing information. This may hamper performance if your model heavily uses asynchronous communication operations.
Once the logs are populated, they can be summarized with deepspeed.comm.log_summary(). For more detail and example usage, see the tutorial
comms_logger: [dictionary]
prof_ops
A list of communication operations to log (only the specified ops will be profiled).
Example of comms_logger configuration for logging specific operations only:
"comms_logger": {
"enabled": true,
"verbose": false,
"prof_all": false,
"debug": false,
"prof_ops": ["all_reduce", "all_gather"]
Compression
Note: Compression has seven different components, including layer reduction, weight quantization, activation quantization, sparse pruning, row pruning, head pruning, and channel pruning. We explain them one by one with simple json examples. Read more about how to use the DeepSpeed Compression library in our tutorial.
Layer Reduction
Note: Layer reduction works much better when using knowledage distillation (learn more in our tutorial):
"compression_training": {
"layer_reduction": {
"enabled": true,
"keep_number_layer": 5,
"module_name_prefix": "bert.encoder.layer",
"teacher_layer": [
"other_module_name": [
"bert.pooler",
"bert.embeddings",
"classifier"
layer_reduction: [dictionary]
module_name_prefix: [str]
The (uniform) name prefix of the model’s modules of which the associated weight parameters are to be reinitialized.
teacher_layer: [list]
The layer of the weight parameters are to be reinitialized. The length of the list equals to ‘keep_number_layer’.
other_module_name: [list]
The name of modules of which the associated weight parameters are to be reinitialized. It is an complemenatory or alternative of module_name_prefix. For instance, “other_module_name”: [“bert.encoder.layer.2”,”bert.encoder.layer.4”] equals to “module_name_prefix”:”bert.encoder.layer” and “teacher_layer”: [2,4].
"quantize_groups": 1,
"quantize_verbose": false,
"quantization_type": "symmetric",
"rounding": "nearest",
"quantize_weight_in_forward": false,
"fp16_mixed_quantize":{
"enabled": false,
"quantize_change_ratio": 0.001
"different_groups":{
"wq1": {
"params": {
"start_bits": 8,
"target_bits": 8,
"quantization_period": 50
"modules": [
"attention.self",
"intermediate"
"wq2": {
"params": {
"start_bits": 4,
"target_bits": 4,
"quantization_period": 50
"modules": [
"attention.output"
shared_parameters: [dictionary]
Shared parameters for all weight quantization groups.
quantizer_kernel: [boolean]
Use DeepSpeed quantization kernel for >=4 bit quantization. This can only be enabled when using DeepSpeed FP16 optimizer.
false
schedule_offset: [integer]
Enable weight quantization after scheduled steps (can be treated as warmup steps).
quantize_groups: [integer]
Split the weight matrix into different number of groups, and each of them has its own scaling factor.
quantize_verbose: [boolean]
Print the quantization related logs.
false
quantization_type: [string]
Choose the quantization algorithm, symmetric or asymmetric.
"symmetric"
rounding: [string]
Rounding algorithm associated with quantization, nearest or stochastic.
"nearest"
quantize_weight_in_forward: [boolean]
Quantize weight in optimizer or forward step, must set to be true for FP32 optimizer training.
false
fp16_mixed_quantize: [dictionary]
Using the value mixed by FP16 value and the quantized value.
enabled: [boolean]
Whether fp16 mixed quantization is enabled.
false
quantize_change_ratio: [float]
Initial quantize value ratio, will gradually increase to 1.
0.001
quantization_period: [integer]
For every n steps, the quantization bits will be reduce by 1.
modules: [list]
Scope of weight parameters associated to the params setting.
"All Linear and CONV2D layers"
"enabled": true,
"quantization_type": "asymmetric",
"range_calibration": "dynamic",
"schedule_offset": 50
"different_groups":{
"aq1": {
"params": {
"bits": 8
"modules": [
"attention.output"
shared_parameters: [dictionary]
Shared parameters for all activation quantization groups.
quantization_type: [string]
Choose the quantization algorithm, symmetric or asymmetric.
"symmetric"
range_calibration: [string]
Using dynamic (per token or per image) or static (fixed min/max using momentum) for inference.
"static"
schedule_offset: [integer]
Enable activation quantization after scheduled steps (can be treated as warmup steps).
"schedule_offset_end": 90,
"schedule_offset_stride": 15,
"method": "snip_momentum",
"block_pattern": "4x1",
"dense_ratio": 0.4,
"excluded_modules": ['classifier', 'pooler']
"different_groups":{
shared_parameters: [dictionary]
Shared parameters for all sparse pruning groups.
schedule_offset: [integer]
Enable sparse pruning after scheduled steps (can be treated as warmup steps).
schedule_offset_end: [integer]
Disable sparse pruning after scheduled steps, mandotory for snip_momentum.
schedule_offset_stride: [integer]
The stride of pruning on training steps, mandotory for snip_momentum.
method: [string]
Choose different pruning methods, l1 (static, magnitude based), topk (dynamic, learnable) or snip_momentum (structured pruning).
block_pattern: [string]
Choose different structured pruning block patterns, NxM or N:M (N and M are integers). For instance, “4x1” or “2:4” are common block patterns, mandotory for snip_momentum.
"4x1"
dense_ratio: [float]
Used to get the targeted global sparsity ratio, mandotory for snip_momentum.
"0.1"
excluded_modules: [list]
Excluded pruning scope on some special modules like output layer.
different_groups: [dictionary]
Different pruning sets, this is used for different pruning parameters. In this example, we give one set. In practice, you can choose the number of sets based on your requirements.
Note for snip_momentum method, you can leave it as empty.
Row Pruning
Note: Row Pruning is a feature designed for two back-to-back linear layers (e.g., Feed Forward Network in Transformers). As such, we suggested use row pruning for the first linear layer (i.e., the intermediate.dense layer for BERT). Reducing the row dimension of this matrix can help reducing the column of the follow-up matrix (i.e., layer.\\w+.output.dense layer for BERT). It should also work for other linear layers as well.
"compression_training": {
"row_pruning":{
"shared_parameters":{
"enabled": true,
"schedule_offset": 20,
"method": "topk"
"different_groups":{
"rp1": {
"params": {
"dense_ratio": 0.5
"modules": [
"intermediate.dense"
"related_modules":[
["layer.\\w+.output.dense"]
shared_parameters: [dictionary]
Shared parameters for all row pruning groups.
schedule_offset: [integer]
Enable row pruning after scheduled steps (can be treated as warmup steps).
method: [string]
Choose different pruning methods, l1 (static, magnitude based) or topk (dynamic, learnable).
modules: [list]
Scope of weight parameters associated to the params setting.
"All Linear and CONV2D layers"
related_modules: [list[list]]
Related module to the row pruned module, which can be performed column pruning.
Head Pruning
Note: Head Pruning is a feature designed for two attention layers (e.g., Multi Head Attention in Transformers). For now, it can only be applied to output matrix of the Transformer (i.e., attention.output.dense in BERT). Pruning the output matrix can lead to the pruning of Query/Key/Value matrix as well.
"compression_training": {
"head_pruning":{
"shared_parameters":{
"enabled": true,
"schedule_offset": 10,
"method": "topk",
"num_heads": 12
"different_groups":{
"rp1": {
"params": {
"dense_ratio": 0.5
"modules": [
"attention.output.dense"
"related_modules":[
["self.query", "self.key", "self.value"]
shared_parameters: [dictionary]
Shared parameters for all head pruning groups.
schedule_offset: [integer]
Enable head pruning after scheduled steps (can be treated as warmup steps).
method: [string]
Choose different pruning methods. For now, we only support topk (dynamic, learnable).
"topk"
num_heads: [int]
Number of heads (must be provided by user).
modules: [list]
Scope of weight parameters associated to the params setting.
"All Linear and CONV2D layers"
related_modules: [list[list]]
Related module (Usually Q/K/V) to the head pruned module (i.e., the output matrix). For now, this feature only works for BERT.
Channel Pruning
Note: Channel Pruning is a feature designed for two back-to-back CONV2d layers (e.g., residual connection in ResNet). As such, we suggested use channel pruning for the first CONV2d layer. Reducing the number of output channels of this layer can help reducing the number of input channels the follow-up layer. It should also work for other CONV2d layers as well.
"compression_training": {
"channel_pruning":{
"shared_parameters":{
"enabled": true,
"schedule_offset": 0,
"method": "topk"
"different_groups":{
"cp1": {
"params": {
"dense_ratio": 0.5
"modules": [
"layer....conv1"
"related_modules": [
["layer....conv2", "layer....bn1"]
shared_parameters: [dictionary]
Shared parameters for all channel pruning groups.
schedule_offset: [integer]
Enable channel pruning after scheduled steps (can be treated as warmup steps).
method: [string]
Choose different pruning methods, l1 (static, magnitude based) or topk (dynamic, learnable).
modules: [list]
Scope of weight parameters associated to the params setting.
"All CONV2D layers"
related_modules: [list[list]]
Related module to the channel pruned module.
Enables level of checking to ensure checkpoint tags are consistent across all ranks. Useful when restoring with different world sizes.
“Warn”
If true DeepSpeed will store model parameter states and checkpoint states based on local rank allowing checkpoints to be loaded without access to a shared filesystem.
false
Specifies the data type in which to do gradient accumulation. If None the default is to match the model type.
