Remove external logger dependency

[riccardofelluga]

What does this PR do?

After discussing with @crcrpar we thought it's better to remove the dependency to loguru and not use #2275.

The log looks like:

25-07-14 15:42 INFO Global batch size: 1 (mbs: 1, grad_acc_steps: 1, world_size: 1)
25-07-14 15:42 INFO Loading tokenizer
25-07-14 15:42 INFO Loading base model on meta device...
25-07-14 15:42 INFO Base model loaded on meta device
25-07-14 15:42 INFO Configured model for static shapes with sequence length: 4096
25-07-14 15:42 INFO Gradient checkpointing disabled
25-07-14 15:42 INFO Applying LoRA to model
25-07-14 15:42 INFO LoRA applied to model
25-07-14 15:42 INFO Verifying gradient setup...
25-07-14 15:42 INFO Applying compilation: thunder to model
25-07-14 15:42 INFO Resetting cache size for torch.dynamo
25-07-14 15:42 INFO Disabled gradient checkpointing for Thunder compilation
25-07-14 15:42 INFO Thunder used executors: ['cudnn', 'sdpa', 'torchcompile_xentropy', 'nvfuser']
25-07-14 15:42 INFO Applying Thunder compilation with 4 executors
25-07-14 15:42 INFO Using ThunderFX
25-07-14 15:42 INFO Compilation applied to model
25-07-14 15:42 INFO Trainable parameters: 2,883,584
25-07-14 15:42 INFO Total parameters: 1,558,997,504

cc @Borda

[Borda]

I am fine with trimming dependencies, but could you pls share the motivation?

[crcrpar]

I personally would find it better to remove a benchmark specific dependency if it's not that difficult to remove it and the required code change isn't overwhelming

[riccardofelluga]

@Borda

I am fine with trimming dependencies, but could you pls share the motivation?

Yess so this came up in the PR that merged the benchmark script #2254 (review) and I've told @t-vi that I would try to move from this dep to the internal Thunder logger.

However the logger is better reserved to Thunder internals and therefore I've just removed the loguru dependency.

[t-vi]

Thank you @riccardofelluga