docs/source_en/Instruction/Megatron-SWIFT-Training.md

Megatron-SWIFT Training

SWIFT incorporates Megatron's parallelization techniques to accelerate the training of large models, including data parallelism, tensor parallelism, pipeline parallelism, sequence parallelism, context parallelism, and expert parallelism. It supports the pre-training and fine-tuning of models such as Qwen3, Qwen3-MoE, Qwen2.5, Llama3, and the Deepseek-R1 distillation series. For a complete list of supported models, please refer to the Supported Models and Datasets documentation.

Environment Setup

To use Megatron-SWIFT, in addition to installing the swift dependencies, you also need to install the following:

# Recommended PyTorch version: 2.5 / 2.6
pip install pybind11
# transformer_engine
# If an installation error occurs, you can refer to this issue for resolution: https://github.com/modelscope/ms-swift/issues/3793
pip install git+https://github.com/NVIDIA/TransformerEngine.git@stable

# apex
git clone https://github.com/NVIDIA/apex
cd apex
pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" ./

# megatron-core
pip install git+https://github.com/NVIDIA/Megatron-LM.git@core_r0.12.0

Alternatively, you can also use the image:

modelscope-registry.cn-hangzhou.cr.aliyuncs.com/modelscope-repo/modelscope:ubuntu22.04-cuda12.4.0-py311-torch2.6.0-vllm0.8.3-modelscope1.25.0-swift3.3.0.post1
modelscope-registry.us-west-1.cr.aliyuncs.com/modelscope-repo/modelscope:ubuntu22.04-cuda12.4.0-py311-torch2.6.0-vllm0.8.3-modelscope1.25.0-swift3.3.0.post1

The training module in the dependent library Megatron-LM will be cloned and installed by swift via git clone. Alternatively, you can use the environment variable MEGATRON_LM_PATH to point to the path of an already downloaded repository (in offline environments, use the core_r0.12.0 branch).

Quick Start Example

This section introduces a quick start example for fine-tuning the self-awareness of the Qwen2.5-7B-Instruct model using two 80GiB A100 GPUs. The following best practices can be completed within 10 minutes.

First, we need to convert the weights from HF (Hugging Face) format to Megatron format:

CUDA_VISIBLE_DEVICES=0 \
swift export \
    --model Qwen/Qwen2.5-7B-Instruct \
    --to_mcore true \
    --torch_dtype bfloat16 \
    --output_dir Qwen2.5-7B-Instruct-mcore

Next, use the following script to start training. The required GPU memory resources are 2*80GiB:

NPROC_PER_NODE=2 \
CUDA_VISIBLE_DEVICES=0,1 \
megatron sft \
    --load Qwen2.5-7B-Instruct-mcore \
    --dataset 'AI-ModelScope/alpaca-gpt4-data-zh#500' \
              'AI-ModelScope/alpaca-gpt4-data-en#500' \
              'swift/self-cognition#500' \
    --tensor_model_parallel_size 2 \
    --micro_batch_size 4 \
    --global_batch_size 16 \
    --recompute_granularity selective \
    --train_iters 100 \
    --eval_iters 5 \
    --finetune true \
    --cross_entropy_loss_fusion true \
    --lr 1e-5 \
    --lr_warmup_iters 10 \
    --min_lr 1e-6 \
    --save megatron_output/Qwen2.5-7B-Instruct \
    --save_interval 100 \
    --max_length 2048 \
    --system 'You are a helpful assistant.' \
    --num_workers 4 \
    --no_save_optim true \
    --no_save_rng true \
    --dataset_num_proc 4 \
    --model_author swift \
    --model_name swift-robot

Finally, convert the Megatron format weights back to HF format:

CUDA_VISIBLE_DEVICES=0 \
swift export \
    --mcore_model megatron_output/Qwen2.5-7B-Instruct/vx-xxx \
    --to_hf true \
    --torch_dtype bfloat16 \
    --output_dir megatron_output/Qwen2.5-7B-Instruct/vx-xxx-hf

We then perform inference on the generated HF format weights:

CUDA_VISIBLE_DEVICES=0 \
swift infer \
    --model megatron_output/Qwen2.5-7B-Instruct/vx-xxx-hf \
    --stream true \
    --temperature 0 \
    --max_new_tokens 2048

The inference results are as follows:

<<< who are you?
I am a language model developed by swift, you can call me swift-robot. How can I assist you?

• For pretraining, you can use megatron pt instead of megatron sft, which will use a generative template for training.
• More examples: Including packing, multi-node training, 32K context, MoE models, and pre-training, can be found here.

Benchmark

The speed comparison of full-parameter training for Dense/MoE models using megatron sft and swift sft on a single machine with eight A800 GPUs is shown below. The corresponding scripts can be found here.

Dense Qwen2.5-14B:

	Megatron-LM	Deepspeed-ZeRO2	Deepspeed-ZeRO3
Training Speed	9.04s/it	10.32s/it	10.56s/it
GPU Memory Usage	8*64GB	8*80GB	8*58GB

MoE Qwen1.5-MoE-A2.7B:

	Megatron-LM	Deepspeed-ZeRO2	Deepspeed-ZeRO3
Training Speed	2.93s/it	6.02s/it	24.30s/it
GPU Memory Usage	8*66GB	8*72GB	8*50GB

Command Line Arguments

Megatron Parameters

Training Parameters:

• 🔥micro_batch_size: Batch size per device, default is 1.
• 🔥global_batch_size: Total batch size, equivalent to micro_batch_size * data parallel size * gradient accumulation steps. Default is 16.
• 🔥recompute_granularity: Granularity of activation recomputation, options are 'full', 'selective'. 'full' means recomputing the entire transformer layer, while 'selective' means only recomputing the core attention part of the transformer layer. 'selective' is generally recommended. Default is 'selective'.
• 🔥recompute_method: This parameter takes effect only when recompute_granularity is set to 'full', options are 'uniform', 'block'. Default is None.
• 🔥recompute_num_layers: This parameter takes effect only when recompute_granularity is set to 'full'. Default is None. If recompute_method is set to uniform, this parameter specifies the number of transformer layers in each uniformly divided recomputation unit. For example, you can specify --recompute_granularity full --recompute_method uniform --recompute_num_layers 4. The larger the recompute_num_layers, the smaller the memory usage but higher computation cost. Default is None.
• recompute_modules: Options include "core_attn", "moe_act", "layernorm", "mla_up_proj", "mlp", and "moe". The default value is ["core_attn"]. For example, during MoE training, you can reduce memory usage by specifying --recompute_granularity selective --recompute_modules core_attn moe. Among these, "core_attn", "mlp", and "moe" use normal checkpointing, while "moe_act", "layernorm", and "mla_up_proj" use output-discarding checkpointing.
  • "core_attn": Recomputes the core attention part of the Transformer layer.
  • "mlp": Recomputes the dense MLP layer.
  • "moe": Recomputes the MoE layer.
  • "moe_act": Recomputes the MLP activation function part in the MoE module.
  • "layernorm": Recomputes the input_layernorm and pre_mlp_layernorm.
  • "mla_up_proj": Recomputes the MLA up-projection and RoPE application parts.
• deterministic_mode: Deterministic mode, which may lead to slower training speed, default is False.
• 🔥train_iters: Total number of training iterations, default is None.
• 🔥log_interval: Log interval (unit: iters), default is 5.
• tensorboard_dir: Directory where TensorBoard logs are written. Default is None, meaning logs will be stored in the f'{save}/runs' directory.
• no_masked_softmax_fusion: Default is False. Disables scaling, masking, and softmax fusion for query_key_value.
• no_bias_dropout_fusion: Default is False. Disables bias and dropout fusion.
• no_bias_swiglu_fusion: Default is False. Specify --no_bias_dropout_fusion true to disable bias and swiglu fusion.
• no_rope_fusion: Default is False. Specify --no_rope_fusion true to disable rope fusion.
• no_gradient_accumulation_fusion: Default is False. Specify --no_gradient_accumulation_fusion true to disable gradient accumulation fusion.
• 🔥cross_entropy_loss_fusion: Enables cross-entropy loss calculation fusion. Default is False.
• calculate_per_token_loss: Scales the cross-entropy loss according to the number of non-padded tokens in the global batch. Default is True.
• 🔥attention_backend: The attention backend to use (flash, fused, unfused, local, auto). Defaults to auto.
• optimizer: Optimizer type, options are 'adam', 'sgd'. Default is adam.
• dataloader_type: Default is 'cyclic', options are 'single', 'cyclic', 'external'. If --streaming is enabled, set it to external.
• manual_gc: Disables the default garbage collector and manually triggers garbage collection. Default is False.
• manual_gc_interval: Interval at which garbage collection is triggered. Default is 0.
• seed: Random seed for python, numpy, pytorch, and cuda, default is 42.
• 🔥num_workers: Number of workers for the dataloader, default is 4.
  • Note: If --streaming true is set, it will be set to 1. seq_length: Defaults to None, meaning it is set to max_length. To restrict the dataset length, please use the --max_length parameter in the basic arguments; there is no need to set this parameter.
• use_cpu_initialization: Initializes weights on the CPU, default is False. Used during HF and MCore weight conversion.
• no_create_attention_mask_in_dataloader: Does not create an attention mask in the dataloader, default is True.

Learning Rate Parameters:

• 🔥lr: Initial learning rate, which will ultimately determine the learning rate for each iteration based on the warm-up and decay strategy, default is 1e-5.
• lr_decay_style: Learning rate decay strategy, default is 'cosine'. Commonly set to 'cosine', 'linear', or 'constant'.
• 🔥lr_decay_iters: Number of iterations for learning rate decay. Default is None, meaning it will be set to --train_iters.
• 🔥lr_warmup_iters: Number of iterations for linear learning rate warm-up, default is 0.
• 🔥min_lr: Minimum value of the learning rate, clipping any learning rate below this threshold to this value, default is 0.

Regularization Parameters:

• 🔥weight_decay: Default is 0.1.
• 🔥clip_grad: L2 gradient clipping, default is 1.0.
• adam_beta1: Default is 0.9.
• adam_beta2: Default is 0.95.
• adam_eps: Default is 1e-8.
• sgd_momentum: Default is 0.9.

Checkpoint Parameters:

• 🔥save: Output directory for checkpoints, default is None. During training, if this parameter is not set, it defaults to f'megatron_output/{model_suffix}', e.g., 'megatron_output/Qwen2.5-7B-Instruct'.
  • Note: When training on multiple machines, ensure that the save paths on each node point to the same location. Otherwise, you will need to manually consolidate these weights after training.
• 🔥save_interval: Checkpoint saving interval (steps), default is 500.
  • Note: Weights will always be saved at the end of training.
• 🔥no_save_optim: Do not save optimizer, default is False.
• 🔥no_save_rng: Do not save RNG, default is False.
• 🔥load: Directory of the checkpoint to load, default is None.
• 🔥no_load_optim: Do not load optimizer, default is False.
• 🔥no_load_rng: Do not load RNG, default is False.
• 🔥finetune: Load the model and fine-tune. Does not load the optimizer and random seed states from the checkpoint and resets the iteration count to 0. Default is False.
• ckpt_format: Format of the checkpoint. Options are 'torch', 'torch_dist', 'zarr'. Default is 'torch_dist'.
• no_initialization: Do not initialize weights, default is True.
• auto_detect_ckpt_format: Automatically detect whether the checkpoint format is legacy or distributed. Default is True.
• exit_on_missing_checkpoint: If --load is set but no checkpoint is found, exit directly instead of initializing. Default is True.

Distributed Parameters:

• distributed_backend: Distributed backend, options are 'nccl', 'gloo'. Default is nccl.
• 🔥use_distributed_optimizer: Use a distributed optimizer. Default is True.
• 🔥tensor_model_parallel_size: TP (Tensor Parallelism) size, default is 1.
• 🔥pipeline_model_parallel_size: PP (Pipeline Parallelism) size, default is 1.
• decoder_first_pipeline_num_layers: The number of Transformer layers in the first pipeline stage of the decoder. Default is None, which means the Transformer layers are evenly distributed across all pipeline stages.
• decoder_last_pipeline_num_layers: The number of Transformer layers in the last pipeline stage of the decoder. Default is None, which means the Transformer layers are evenly distributed across all pipeline stages.
• 🔥sequence_parallel: Enable sequence parallel optimization. Default is False.
• 🔥context_parallel_size: CP (Context Parallelism) size, default is 1.
• tp_comm_overlap: Overlap tensor parallel communication with GEMM (General Matrix Multiplication) kernels (to reduce communication time). Default is False.
• overlap_grad_reduce: Overlap grad reduction operations in DDP (to reduce DP communication time). Default is False.
• overlap_param_gather: Overlap all-gather of parameters in the distributed optimizer (to reduce DP communication time). Default is False.
• distributed_timeout_minutes: Timeout duration for torch.distributed (in minutes), default is 60 minutes.

Logging Parameters:

• log_params_norm: Logs the norm of parameters. Default is False.
• log_throughput: Logs throughput per GPU. Default is True.
  • Note: In non-packing scenarios, log_throughput is not accurate because seq_length does not equal the actual sequence length.
• tensorboard_log_interval: Interval (steps) for logging to TensorBoard, default is 1.
• tensorboard_queue_size: Queue length (related to disk I/O), similar to write intervals. Default is 50.
• log_timers_to_tensorboard: Logs timers to TensorBoard. Default is True.
• no_log_learning_rate_to_tensorboard: Do not log learning rate to TensorBoard. Default is False.
• log_validation_ppl_to_tensorboard: Writes validation perplexity to TensorBoard. Default is True.
• log_memory_to_tensorboard: Writes memory logs to TensorBoard. Default is True.
• logging_level: Logging level. Default is None.
• wandb_project: The name of the wandb project. Defaults to '', which means ignoring wandb.
• wandb_exp_name: The name of the wandb experiment. Defaults to ''.
• wandb_save_dir: The local path to save wandb results. Defaults to ''.

Evaluation Parameters:

• 🔥eval_iters: Number of evaluation iterations, default is 100.
• 🔥eval_interval: Evaluation interval (steps), default is None, meaning it will be set to save_interval.

Mixed Precision Parameters:

• fp16: FP16 mode. The default is None, and it will be set according to the model's torch_dtype. The torch_dtype is read from the config.json by default.
• bf16: BF16 mode. The default is None, and it will be set according to the model's torch_dtype.
• apply_query_key_layer_scaling: Scales Q * K^T by 1 / layer number (e.g., divide by layer_num for layer_num-th layer). This is helpful for FP16 training. Default is None, meaning that if --fp16 is used, it will be set to True.
• attention_softmax_in_fp32: Uses FP32 for computations in attention_mask and softmax. Default is True.

Model Parameters: (The following parameters typically do not need to be set as they will be configured based on the HF model’s config.json; users don’t need to worry about them)

• num_layers: Number of transformer layers, default is None.
• hidden_size: Transformer hidden size, default is None.
• ffn_hidden_size: Hidden size of the FFN layer in the transformer. Default is None, set to 4*hidden_size.
• num_attention_heads: Number of transformer attention heads, default is None.
• group_query_attention: Default is None. If num_query_groups > 1, group_query_attention is set to True, otherwise False.
• num_query_groups: Default is 1.
• max_position_embeddings: Maximum length of positional embeddings, default is None.
• position_embedding_type: Type of positional embedding, options are 'learned_absolute', 'rope', 'relative', and 'none'. Default is 'rope'.
• rotary_base: Default is 10000.
• rotary_percent: Default is 1.
• normalization: Options are 'LayerNorm', 'RMSNorm'. Default is RMSNorm.
• norm_epsilon: Default is 1e-5.
• swiglu: Uses swiglu instead of the default gelu. Default is True.
• untie_embeddings_and_output_weights: Unties embedding and output weights. Default is True.
• disable_bias_linear: Disables bias in linear layers. Default is True.
• add_qkv_bias: Adds bias only to QKV linear layers. Default is True.
• attention_dropout: Default is 0.
• hidden_dropout: Default is 0.
• kv_channels: Defaults to None, set to args.hidden_size // args.num_attention_heads.
• qk_layernorm: Whether to apply layer normalization to Q and K.
• transformer_impl: Which transformer implementation to use, options are 'local' and 'transformer_engine'. Default is transformer_engine.
• padded_vocab_size: Full vocabulary size, default is None.
• rope_scaling: Related parameters for rope_scaling, default is None. Refer to the format in llama3.1 config.json. Pass the value as a JSON string.
• model_type: The model_type in the config.json of the Huggingface model weights.

MoE Parameters:

• num_experts: The number of experts in MoE, default is None. Automatically read from config.json.
• moe_ffn_hidden_size: The hidden layer size of the feed-forward network (ffn) for each expert. Default is None, set to ffn_hidden_size. Automatically read from config.json.
• moe_shared_expert_intermediate_size: The total FFN hidden layer size for shared experts. If there are multiple shared experts, it should equal num_shared_experts * ffn_size_of_each_shared_expert. Default is None. Automatically read from config.json.
• moe_router_topk: The number of experts each token is routed to. Default is None. Automatically read from config.json.
• moe_router_pre_softmax: Enable pre-softmax routing for MoE, meaning that softmax will be applied before top-k selection. Default is None. Automatically read from config.json.
• 🔥moe_aux_loss_coeff: Scaling coefficient for the auxiliary loss: the recommended initial value is 1e-2. Default is None. Automatically read from config.json.
• 🔥expert_model_parallel_size: The degree of expert parallelism, default is 1.
• moe_token_dispatcher_type: The type of token dispatcher to use. Options include 'allgather', 'alltoall', and 'alltoall_seq'. Default is 'alltoall'.
• moe_grouped_gemm: When each rank contains multiple experts, improve utilization and performance by launching multiple local GEMM kernels across multiple streams using GroupedLinear in TransformerEngine. Default is False.
• moe_router_load_balancing_type: Determines the load balancing strategy for the router. Options are "aux_loss", "seq_aux_loss", "sinkhorn", "none". Default is "aux_loss".
• moe_z_loss_coeff: Scaling coefficient for z-loss. Default is None.
• moe_expert_capacity_factor: Capacity factor for each expert, None means no tokens will be dropped. Default is None.
• moe_shared_expert_overlap: Enable overlapping of shared expert computation with scheduler communication. If this option is not enabled, shared experts will execute after the routing experts. Only effective when moe_shared_expert_intermediate_size is set. Default is False.

Megatron Training Parameters

Megatron training parameters inherit from Megatron parameters and basic parameters. For information on basic parameters, see here. Additionally, the following parameters are included:

• add_version: Adds a directory <version>-<timestamp> to save to prevent overwriting weights, default is True.
• 🔥packing: Whether to use sequence packing, defaults to False.
• 🔥streaming: Stream reading and processing of the dataset, default is False. It is typically set to True when handling large datasets. For more information on streaming parameters, refer to the command-line parameters documentation.
• lazy_tokenize: Default is False. If this parameter is set to False, all dataset samples are tokenized before training (this avoids errors during training); if set to True, tokenization occurs during training (this saves memory).
• dataloader_persistent_workers: A parameter passed directly to the dataloader, with a default value of True.
• dataloader_prefetch_factor: A parameter passed directly to the dataloader, with a default value of 10.
• max_epochs: Forces the training to exit after reaching max_epochs, and performs validation and saving of the model weights. This parameter is especially useful when using a streaming dataset. Default is None.