Deprecated

LLMForge is being deprecated: The Ray Team is consolidating around open source fine-tuning solutions. Llama Factory and Axolotl provide enhanced functionality (quantization, advanced algorithms) and native Ray support for scaling. See the migration guide for transitioning your workflows.

LLMForge Releases

LLMForge version	Image URI	Ray version	Python version	Changelog
0.5.10	localhost:5555/anyscale/llm-forge:0.5.10	2.40.0	3.11.10	Link
0.5.9	localhost:5555/anyscale/llm-forge:0.5.9	2.40.0	3.11.10	Link
0.5.8	localhost:5555/anyscale/llm-forge:0.5.8	2.38.0	3.11.10	Link
0.5.7	localhost:5555/anyscale/llm-forge:0.5.7	2.35.0	3.11.9	Link
0.5.6	localhost:5555/anyscale/llm-forge:0.5.6	2.35.0	3.11.9	Link
0.5.5	localhost:5555/anyscale/llm-forge:0.5.5	2.32.0	3.11.9	Link
0.5.4	localhost:5555/anyscale/llm-forge:0.5.4	2.32.0	3.11.9	-
0.5.3	localhost:5555/anyscale/llm-forge:0.5.3	2.32.0	3.11.9	-
0.5.2	localhost:5555/anyscale/llm-forge:0.5.2	2.32.0	3.11.9	-
0.5.1	localhost:5555/anyscale/llm-forge:0.5.1	2.31.0	3.11.9	-
0.5.0.1	localhost:5555/anyscale/llm-forge:0.5.0.1-ngmM6BdcEdhWo0nvedP7janPLKS9Cdz2	2.24.0	3.10.14	-

Changelog

Version 0.5.10

• Added support for customizing the fine-tuned model's tag with custom_model_tag config option.
• Added an eval_first flag allowing for running an iteration of the evaluation loop (+ logging metrics) before training
• The assistant role is now optional in each sample. This is helpful for use cases like continued pretraining.
• General improvements in logging and bug fixes

Version 0.5.9

• Added experimental support for vision-language training (supports Pixtral architecture for now)
• Added new optimizer_config option for custom HuggingFace optimizer configurations [API Reference]
• Upgraded rayturbo to 2.40
• General improvements and bug fixes

Version 0.5.8

• Added new checkpoint_and_evaluation_frequency for customizing checkpointing and evaluation cadence.
  • Resources: [API Reference] [Example usage]
• Upgraded rayturbo to 2.38, which comes with many improvements in Ray Data
• Added support for torch.compile [API Reference]
• General improvements in data preprocessing and logging

⚠️ Deprecation notice ⚠️

• The parameters checkpoint_every_n_epochs and max_num_checkpoints are deprecated and will be removed in version 0.6. Please use checkpoint_and_evaluation_frequency instead.

Version 0.5.7

• Added support for Llama 3.1 and 3.2 with default configurations
• Improved checkpoint discovery for serving - storage_uri returned by anyscale.llm.model.get now contains forwarded checkpoint for LoRA. [User guide]
• Renamed no_gradient_checkpoint to gradient_checkpointing
• General improvements in configuration management and stability

⚠️ Deprecation notice ⚠️

• The no_gradient_checkpoint parameter was confusing to use and is now deprecated. This will be removed in version 0.6. Please use gradient_checkpointing instead.

Version 0.5.6

• Upgraded to Rayturbo 2.35 with Ray Data and Ray dashboard improvements
• Added integration with Liger Kernel.
  • Resources: [API Reference] | [Docs] | [Example Usage]
• Added support for different fine-tuning tasks with a new task parameter:
  • Instruction tuning (ignores user messages in the loss)
  • Classification (predicts only user-defined labels)
  • Direct Preference Optimization (trains on preference data)
  • Resources: [Docs] | [Task Examples] | [Preference Tuning End-to-End Example] | [Classification End-to-End Example]
• General improvements and bug fixes

Version 0.5.5

• Added configurable support for MLflow and Wandb logging.
  • Resources: [API Reference] | [Docs] | [Wandb Example] | [MLFlow Example]