speeding_up_training.md

Speeding Up Training

This guide covers various methods to accelerate training in TRL. Each technique includes minimal examples with links to more comprehensive documentation.

vLLM for fast generation in online methods

Online methods such as GRPO or Online DPO require the model to generate completions, which is often a slow process and can significantly impact training time. To speed up generation, you can use vLLM, a library that enables fast generation through, among other things, PagedAttention. TRL's online trainers support vLLM, greatly improving training speed. For more details, see vLLM Integration.

To use vLLM, first install it using:

pip install trl[vllm]

First, start a vLLM server by running:

trl vllm-serve --model <model_name>

Then, run the training script and pass use_vllm=True in the training arguments.

from trl.experimental.online_dpo import OnlineDPOConfig

training_args = OnlineDPOConfig(..., use_vllm=True, vllm_mode="server")

First, start a vLLM server by running:

trl vllm-serve --model <model_name>

Then, run the training script and pass use_vllm=True in the training arguments.

from trl import GRPOConfig

training_args = GRPOConfig(..., use_vllm=True, vllm_mode="server")

You can customize the server configuration by passing additional arguments. For more information, see vLLM integration.

Warning

When using vLLM, ensure that the GPUs assigned for training and generation are separate to avoid resource conflicts. For instance, if you plan to use 4 GPUs for training and another 4 for vLLM generation, you can specify GPU allocation using CUDA_VISIBLE_DEVICES.

Set GPUs 0-3 for vLLM generation:

CUDA_VISIBLE_DEVICES=0,1,2,3 trl vllm-serve --model <model_name>

And GPUs 4-7 for training:

CUDA_VISIBLE_DEVICES=4,5,6,7 accelerate launch train.py

First, start a vLLM server by running:

trl vllm-serve --model <model_name>

Then, run the training script and pass use_vllm=True in the training arguments.

from trl import RLOOConfig

training_args = RLOOConfig(..., use_vllm=True, vllm_mode="server")

You can customize the server configuration by passing additional arguments. For more information, see vLLM integration.

Warning

When using vLLM, ensure that the GPUs assigned for training and generation are separate to avoid resource conflicts. For instance, if you plan to use 4 GPUs for training and another 4 for vLLM generation, you can specify GPU allocation using CUDA_VISIBLE_DEVICES.

Set GPUs 0-3 for vLLM generation:

CUDA_VISIBLE_DEVICES=0,1,2,3 trl vllm-serve --model <model_name>

And GPUs 4-7 for training:

CUDA_VISIBLE_DEVICES=4,5,6,7 accelerate launch train.py

Optimized attention implementations

TRL supports various optimized attention implementations that can significantly speed up training while reducing memory usage. You can use either a pre-optimized kernels directly from the Kernels Hub or a manually built attention backend.

You can use pre-optimized attention kernels from the Hub without manual compilation:

from trl import SFTConfig

training_args = SFTConfig(..., model_init_kwargs={"attn_implementation": "kernels-community/flash-attn2"})

Other options include kernels-community/vllm-flash-attn3 and kernels-community/paged-attention.

Optimized attention works across all TRL trainers. For more details, see Kernels Hub Integration.

Warning

Manually building optimized attention backends is complex and time-consuming. It's never recommended unless absolutely necessary. Consider using Kernels from the Hub instead, as described in the previous section.

If you have manually installed an optimized attention backend like Flash Attention 2, you can specify it in the training arguments:

from trl import SFTConfig

training_args = SFTConfig(..., model_init_kwargs={"attn_implementation": "flash_attention_2"})

Liger Kernel for memory optimization

Liger Kernel is a collection of Triton kernels designed for LLM training that can increase throughput by 20% and reduce memory usage by 60%.

from trl import SFTConfig

training_args = SFTConfig(..., use_liger_kernel=True)

from trl import DPOConfig

training_args = DPOConfig(..., use_liger_kernel=True)

from trl import GRPOConfig

training_args = GRPOConfig(..., use_liger_kernel=True)

from trl.experimental.kto import KTOConfig

training_args = KTOConfig(..., use_liger_kernel=True)

from trl.experimental.gkd import GKDConfig

training_args = GKDConfig(..., use_liger_kernel=True)

For more information, see Liger Kernel Integration.

Mixed precision training

Mixed precision training using bf16 or fp16 can speed up training and reduce memory usage with minimal impact on model quality.

from trl import SFTConfig

training_args = SFTConfig(..., bf16=True)  # or fp16=True for older GPUs

Use bf16=True for Ampere GPUs (A100, RTX 30xx) or newer, and fp16=True for older GPUs. Mixed precision training is supported across all TRL trainers.