中文 ｜ English

Hy3 preview provides processes related to model training. This section details how to process training data for model training purposes.

Hy3 preview supports both "slow thinking" and "fast thinking" modes. By default, the model outputs in slow thinking mode. If you wish the model to use fast thinking, you can control it via the reasoning_effort parameter (options: high, low, no_think).

The training data should be formatted as a list of messages. By default, the system prompt for both training and inference is empty, but you may customize it as needed.

# Fast thinking pattern (no_think)
{"reasoning_effort": "no_think", "messages": [{"content": "You are a helpful assistant.\nThe current time is 2026-01-01 13:26:12 Thursday", "role": "system"}, {"content": "1+1=?", "role": "user"}, {"role": "assistant", "content": "1+1=2"}]}

# Slow thinking pattern (high)
{"reasoning_effort": "high", "messages": [{"content": "You are a helpful assistant.\nThe current time is 2026-01-01 13:26:12 Thursday", "role": "system"}, {"content": "1+1=?", "role": "user"}, {"role": "assistant", "content": "1+1=2", "reasoning_content": "The user is asking for the result of 1 + 1. In basic decimal arithmetic, 1 + 1 equals 2."}]}

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("./models", use_fast=False, trust_remote_code=True)
ids = tokenizer.apply_chat_template(messages, is_training=True)

The original Hy3 preview checkpoint stores each expert's weights independently. Before training, it needs to be converted to the HuggingFace-compatible format (fusing multiple experts per layer into 3D tensors with unified key naming) to improve loading and training speed. We provide a conversion script convert_ckpt_to_outer.py and a validation script check_converted.py, located in the train/tools directory.

python convert_ckpt_to_outer.py \
    --input_dir <original_checkpoint_dir> \
    --output_dir <output_dir> \
    --workers 8

Parameters:

• --input_dir: Path to the original checkpoint directory (required)
• --output_dir: Path to the converted checkpoint output directory (required)
• --workers: Number of parallel worker processes, default is 8 (optional)

The conversion script performs the following steps:

1. Pre-scans model.safetensors.index.json to detect cross-shard expert groups
2. Converts weights shard-by-shard in parallel (key renaming + expert fusion)
3. Post-processes cross-shard expert groups (merges data from multiple shards)
4. Copies config.json, tokenizer, and other files
5. Rebuilds model.safetensors.index.json

After conversion, it is recommended to validate the result using the validation script:

python check_converted.py <converted_checkpoint_dir> --spot-check 3

Parameters:

• First argument: Path to the converted checkpoint directory (required)
• --spot-check: Number of shard files to spot-check by loading tensors and verifying shape, dtype, NaN/Inf, etc. Default is 3 (optional)

The validation script checks the following:

1. Completeness of config.json
2. Whether all expected keys in model.safetensors.index.json are present (including regular layers and MTP layers)
3. Whether all referenced shard files exist and are non-empty
4. Spot-checks tensor shape, dtype, and NaN/Inf in selected shard files
5. Detects orphan empty shard files (cross-shard merge residues, safe to delete)

You can quickly get started by following the instructions in the Quick Start Guide.

Based on testing, when make_moe_param_leaf_module and zero3+offload are disabled and max_seq_length is set to 4096, full fine-tuning with LoRA requires at least a single machine with 8 GPUs (each with at least 80GB of memory).

Without LoRA, at least 4 machines with 32 GPUs (each with at least 80GB of memory) are required.

Reference: HuggingFace Transformers Trainer

In the train directory, execute:

pip install -r requirements.txt
bash train.sh

To launch training across multiple machines, please follow the steps below and ensure all machines are within the same cluster.

The following instructions use two machines as an example, with their IPs denoted as ${ip1} and ${ip2}. All steps should be performed inside the Docker container.

First, configure passwordless SSH for each container on every machine:

ssh-keygen                  # Generate id_rsa and id_rsa.pub for passwordless login
ssh-keygen -t rsa -A        # Generate /etc/ssh/ssh_host_rsa_key and ssh_host_ecdsa_key for SSH listening
/usr/sbin/sshd -p 36005 -o ListenAddress=0.0.0.0        # Start SSH listening
echo "Port 36005" > ~/.ssh/config   # Set SSH connection port to 36005
passwd root    # Set the root password to avoid monitoring platform alerts

Note: 36005 is an example port. You may use any available port, but ensure it is open and not occupied by other processes.

Next, in each machine's container, execute:

cat ~/.ssh/id_rsa.pub

Copy the output SSH public key and paste it into the ~/.ssh/authorized_keys file, one key per line. This must be done on every machine. In the end, the ~/.ssh/authorized_keys file on each machine should be identical and contain the public keys of all machines.

Please note that for multi-node training, the code executed on each node must be identical. It is recommended to mount a shared network drive. If this is not possible, you must manually copy the dataset, scripts, and code to the same directory on each machine.

Once the above preparations are complete and all dependencies are installed (if not, run pip install -r requirements.txt), add the following configuration at the beginning of train.sh:

export HOST_GPU_NUM=8
# IP list, comma separated. e.g. "192.168.1.1,192.168.1.2" or single node "192.168.1.1"
export IP_LIST=${IP_LIST:-"127.0.0.1"}

Note: If the IP_LIST environment variable is not set, replace IP_LIST with the IP list! The format is:

For a single IP:
IP_LIST=${ip_1}

For multiple IPs:
IP_LIST=${ip_1},${ip_2}

Replace ${ip_1} and ${ip_2} with the actual IP addresses.

Then, on the machine with ${ip1}, execute bash train.sh in the train/ directory. On first launch, you may see the following output:

The authenticity of host '[ip]:36005 ([ip]:36005)' can't be established.
ECDSA key fingerprint is xxxxxx.
ECDSA key fingerprint is MD5:xxxxxx.
Are you sure you want to continue connecting (yes/no)?

Type yes to continue.

The key parameters in the script are as follows:

• --deepspeed: Path to the DeepSpeed configuration file. Three default DeepSpeed configuration files are provided in the train folder: ds_zero2_no_offload.json, ds_zero3_no_offload.json, and ds_zero3_offload.json, with decreasing memory requirements in that order.
• --model_name_or_path: Path to the Hy3 preview HF pre-trained model weights to load, otherwise loading will fail.
• --tokenizer_name_or_path: Path to the tokenizer folder, otherwise loading will fail.
• --train_data_file: Path to the training file, which should be a jsonl file.
• --output_dir: Output directory where logs, tensorboard files, and model weights will be stored.
• --per_device_train_batch_size: Batch size per GPU.
• --gradient_accumulation_steps: Number of gradient accumulation steps. The global batch size is per_device_train_batch_size * gradient_accumulation_steps * dp_size.
• --max_steps: Total number of training steps.
• --save_steps: Number of steps between saving checkpoints.
• --use_lora: Whether to use LoRA training. Also accepts --lora_rank, --lora_alpha, and --lora_dropout parameters. By default, LoRA is applied to "q_proj", "k_proj", "v_proj", and "o_proj". To change this, modify the code. Note: ** When using LoRA training, only the LoRA weights are saved, not the base model weights. ** To merge LoRA weights, see the "LoRA Weight Merging" section below.
• --make_moe_param_leaf_module: When using zero3 and MoE training, treat the MoE module as a leaf module, i.e., its parameters are not partitioned by zero3. This option is expected to significantly increase memory usage.
• --gradient_checkpointing: Enable gradient checkpointing.
• --train_attention_params_only: Whether to train only attention parameters.
• --learning_rate: Maximum learning rate during training.
• --min_lr: Minimum learning rate during training.
• --use_flash_attn: Enable flash-attention for accelerated training.

Notes:

• To resume training from a previously saved checkpoint rather than loading pre-trained weights, specify --resume_from_checkpoint with the path to the checkpoint. Do not specify --model_name_or_path, this will load only the weights, not the training state.
• When resuming from a checkpoint, there may be minor differences in loss due to the randomness of some non-deterministic algorithms. This is normal. See: HuggingFace Transformers Trainer Randomness
• When --model_name_or_path is specified, all model-related parameters will be ignored.
• Samples within a batch are padded to the length of the longest sample in the batch, but the maximum length of each sample is max_seq_length. Any excess will be truncated.
• If you see a warning about bias weights not being loaded, you can ignore it. Hunyuan-Large does not use bias.

Reference: DeepSpeed Configuration

You can try modifying the DeepSpeed configuration by removing the auto attribute from the following parameters and reducing their values:

• stage3_param_persistence_threshold
• stage3_prefetch_bucket_size
• stage3_max_reuse_distance

LoRA weights saved during training cannot be merged into the zero3 model at runtime, as zero3 partitions model weights across data parallel ranks. To merge LoRA weights into the base model, you can do so offline to obtain a merged weight file. Run merge_lora_weight.sh to merge the LoRA and base model weights. The parameters are:

• --base_model_path: Directory of the base model weights
• --adapter_model_path: Directory of the LoRA weights
• --output_path: Directory to save the merged weights
• --save_dtype: Data type for saving the merged weights; options are: fp16, bf16, fp32

If you are familiar with LLaMA-Factory, you may use it for fine-tuning. All scripts, code, and configuration files are archived in the ./train/llama_factory_support directory. Unless otherwise specified, all files mentioned below are located in this directory.

You can install LLaMA-Factory by downloading the source code from https://github.com/hiyouga/LLaMA-Factory/tree/main and following the instructions on the website.

We provide sample LLaMA-Factory training configuration files: hy_v3_lora_sft.yaml and hy_v3_full_sft.yaml, corresponding to LoRA training and full fine-tuning respectively.

Key parameters in the configuration files are as follows:

Model:

• model_name_or_path: Path to the Hy3 preview HF format pre-trained model weights
• trust_remote_code: Whether to trust remote code; Hy3 preview requires this to be set to true

Training Method:

• stage: Training stage, currently sft (supervised fine-tuning)
• finetuning_type: Fine-tuning type, either full (full fine-tuning) or lora (LoRA fine-tuning)
• deepspeed: DeepSpeed configuration file path; ds_zero3_offload_hy.json is recommended for full fine-tuning, ds_zero2_offload_lora.json for LoRA fine-tuning

LoRA Parameters (only effective during LoRA fine-tuning):

• lora_rank: LoRA rank, default 64
• lora_alpha: LoRA alpha coefficient, default 128
• lora_dropout: LoRA dropout ratio, default 0.05
• lora_target: Target modules for LoRA, default q_proj,k_proj,v_proj,o_proj

Dataset:

• dataset_dir: Dataset directory path
• dataset: Dataset name, must be registered in dataset_info.json under dataset_dir
• template: Chat template; Hy3 preview uses hy_v3
• cutoff_len: Maximum sequence length; sequences exceeding this will be truncated. For full fine-tuning, can be set to 262144 (262K); for LoRA fine-tuning, 8192 is recommended to save memory
• max_samples: Maximum number of samples per dataset
• overwrite_cache: Whether to overwrite cached preprocessed datasets

Output:

• output_dir: Output directory where logs, TensorBoard files, and weights will be stored
• logging_steps: Number of steps between logging
• save_steps: Number of steps between saving checkpoints
• plot_loss: Whether to plot the training loss curve
• overwrite_output_dir: Whether to overwrite the existing output directory
• save_only_model: Whether to save only model weights (excluding optimizer states, etc.)
• report_to: Logging tool, options: none, wandb, tensorboard, swanlab, mlflow

Training Hyperparameters:

• per_device_train_batch_size: Batch size per GPU
• gradient_accumulation_steps: Gradient accumulation steps; per_device_train_batch_size * gradient_accumulation_steps * dp_size equals the global batch size
• learning_rate: Maximum learning rate; 1.0e-5 recommended for full fine-tuning, 2.0e-4 for LoRA fine-tuning
• num_train_epochs: Number of training epochs
• lr_scheduler_type: Learning rate scheduler type; cosine_with_min_lr is recommended
• lr_scheduler_kwargs.min_lr_rate: Ratio of minimum to maximum learning rate; e.g., 0.1 means the minimum learning rate is 10% of the maximum
• warmup_ratio: Proportion of total training steps used for warmup
• bf16: Whether to use BFloat16 mixed precision training
• gradient_checkpointing: Whether to enable gradient checkpointing to save memory
• ddp_timeout: Distributed training timeout (milliseconds)
• flash_attn: Attention implementation; fa2 (FlashAttention-2) is recommended, sdpa is also available; using fa2 requires the flash-attn package
• resume_from_checkpoint: Resume training from a specified checkpoint path; set to null to start from scratch

Please first configure passwordless SSH login between machines following the Configure Passwordless SSH Login Between Machines section above.

Modify the following configuration at the beginning of train_lf.sh:

export HOST_GPU_NUM=8
# IP list, comma separated. e.g. "192.168.1.1,192.168.1.2" or single node "192.168.1.1"
export IP_LIST=${IP_LIST:-"127.0.0.1"}

Note: If the IP_LIST environment variable is not set, replace IP_LIST with the IP list! The format is:

For a single IP:
IP_LIST=${ip_1}

For multiple IPs:
IP_LIST=${ip_1},${ip_2}

Replace ${ip_1} and ${ip_2} with the actual IP addresses.

Then, on each machine, run bash train_lf.sh in the train/llama_factory_support/ directory.