"Missing or incorrect absmax2 handling in q_galore_adamw8bit.py

[LLMresearcher]

When using the 8-bit AdamW optimizer in q_galore_adamw8bit.py, optimizer_update_8bit_blockwise() requires absmax2 but it is not defined. Seems like absmax2 expects passing a tensor.

[rank0]: File "/root/Q-GaLore/q_galore_torch/q_galore_adamw8bit.py", line 200, in update_step
[rank0]: F.optimizer_update_8bit_blockwise(
[rank0]: TypeError: optimizer_update_8bit_blockwise() missing 1 required positional argument: 'absmax2'