Help with debugging - ValueError: optimizer got an empty parameter list

projects
1

For this NLHF code, I have a noob question where my eyes / mind could not catch where went wrong at this moment yet.

No parameters in params.

Traceback (most recent call last):

File “/Users/john/Downloads/nlhf.py”, line 567, in

optimizer_current_policy = AdamW_on_Lion_Optimizer(

File “/Users/john/Downloads/nlhf.py”, line 534, in init

self.adamW = optim.AdamW(params=params, lr=lr, betas=adam_betas,

File “/Users/john/.nlp/lib/python3.9/site-packages/torch/optim/adamw.py”, line 52, in init

super().init(params, defaults)

File “/Users/john/.nlp/lib/python3.9/site-packages/torch/optim/optimizer.py”, line 261, in init

raise ValueError(“optimizer got an empty parameter list”)

ValueError: optimizer got an empty parameter list

Note: Before params is passed into AdamW_on_Lion_Optimizer(), params is a valid non-empty parameter list.

2

How did you create params?

3

See lines 567 and 577 of the code

Using line 577 does not cause the problem, but line 567 has the above problem.

4

@ptrblck if you add the following to just before line 567 : optimizer_current_policy = AdamW_on_Lion_Optimizer( params=current_policy.parameters(), lr=1e-3 ) , you will notice that params is a valid non-empty parameter list before it is passed into AdamW_on_Lion_Optimizer()

     # Diagnostic code to check parameters
     if not list(current_policy.parameters()):
         print("No parameters in current_policy.")
     else:
         for idx, param in enumerate(current_policy.parameters()):
             print(f"Param {idx}: requires_grad={param.requires_grad}")
             if param.requires_grad:
                 print("There are parameters that require gradients.")
             else:
                 print("No parameters require gradients.")
5

params is not actually empty when before it is first being passed into the AdamW_on_Lion_Optimizer() init function as current_policy.parameters() , so I am really confused on why it became empty when it went inside the init function.

@ptrblck : I wonder if you have any idea / comments / suggestions OR experienced something similar in your past experience ?