[FEATURE]: Integrate GaLore into Colossalai Optimizer(Gemini/Hybrid)

[airlsyn]

Describe the feature

A recent paper titled "GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection" (https://arxiv.org/pdf/2403.03507.pdf) demonstrates a remarkable memory-efficient approach during the training of large language models (LLMs).

Can we integrate this memory-efficient technique into the Colossalai framework?

FYI

• GaLore Adamw: https://github.com/jiaweizzhao/GaLore/blob/master/galore_torch/adamw.py
• 8bit-GaLore Adamw: https://github.com/jiaweizzhao/GaLore/blob/master/galore_torch/adamw8bit.py

[airlsyn]

Any ColossalAI-er could take a look?

[ver217]

Thanks! We will take a look.

[Edenzzzz]

I will take multiple looks

[airlsyn]

I see the MR, that's awesome, when can we use it?

[Edenzzzz]

I plan to release it next week