[BUG] eta is incorrect in step 1

[njzjz]

Bug summary

As shown below, eta in step 1 is obviously different from others.

DeePMD-kit Version

devel

Backend and its version

PyTorch 2.7

How did you download the software?

pip

Input Files, Running Commands, Error Log, etc.

See above.

Steps to Reproduce

Training any model could reproduce the result.

Further Information, Files, and Links

cc @HydrogenSulfate

[HydrogenSulfate]

This is expected, because in typical model training, the first step usually involves additional initialization operations, such as creating optimizer parameters and initializing GPU memory. Statistics collected after dozens of steps tend to be more accurate.

[njzjz]

the first step usually involves additional initialization operations, such as creating optimizer parameters and initializing GPU memory.

But eta in step 1 is shorter than others. Considering your explanation, shouldn't it become longer?

[HydrogenSulfate]

the first step usually involves additional initialization operations, such as creating optimizer parameters and initializing GPU memory.

But eta in step 1 is shorter than others. Considering your explanation, shouldn't it become longer?

Oh It’s probably because, for the first print, the denominator shouldn’t be disp_freq, but rather the interval from start_step to display_step.

# correct eta computation

eta = int(
    (self.num_steps - display_step_id) / min(self.disp_freq, display_step_id - self.start_step) * train_time
)

[HydrogenSulfate]

@njzjz fixed in #4886