[megatron] megatron lora_llm support no_save_optim

[Jintao-Huang]

No description provided.

[gemini-code-assist]

Code Review

This pull request relaxes validation constraints for the lora_llm tuner type and updates the trainer to handle lora_llm similarly to lora during weight loading and checkpointing. Review feedback highlights a potential issue where non-LoRA trainable parameters in multimodal models might be excluded from checkpoints when peft_format is enabled. It is also suggested to default the tuner type to lora_llm for multimodal models when adapters are present.

[gemini-code-assist]

Setting peft_format=True for lora_llm ensures that LoRA weights are saved, but it may exclude other trainable parameters in multimodal models. If freeze_vit or freeze_aligner are set to False, their weights will be filtered out of the Megatron checkpoint because they are not LoRA parameters and are not automatically added to modules_to_save. Consider ensuring that peft_format is only True when no other non-LoRA parameters are trainable, or that those modules are included in the filter.

[gemini-code-assist]

For multimodal models, when adapters are provided but tuner_type is not specified (defaulting to 'full'), it might be more appropriate to set tuner_type to 'lora_llm' instead of 'lora'. This ensures that the specialized multimodal LoRA logic is used consistently for multimodal models.

Suggested change

	if self.tuner_type == 'full':
	self.tuner_type = 'lora'
	if self.tuner_type == 'full':
	self.tuner_type = 'lora_llm' if self.is_multimodal else 'lora'