BUG: Excessive RAM usage on large file

Describe the issue:

Hi, I have a 2 hr recording of 384 channels (1 probe, ~200GB file size). I'm trying to run run_kilosort with this file (using DEFAULT_SETTINGS). During the "Computing drift correction" step ("Detecting spikes" sub-step), the (CPU) RAM usage grows steadily up to ~120 GB at 40% progress in the routine. Can you please help me understand if there's a memory link or why so much is being held in memory and not released?

I quickly tried to call run_kilosort with clear_cache=True, but the high RAM usage persisted. Even a smaller recording (~20 GB) used ~80 GB RAM, so the file size doesn't seem to be the issue.

I've seen this comment in #766, but unless run_kilosort is calling spikeinterface internally, I'm not using spikeinterface. Based on this other comment in that thread, I'm happy to change the data dtype if that prevents high memory usage, though I'm not sure where to change that when using run_kilosort.

(Tagging my collaborator: @jeffjennings)

Reproduce the bug:

My full script is:

from kilosort import run_kilosort, DEFAULT_SETTINGS
from kilosort.utils import download_probes
download_probes()

settings = DEFAULT_SETTINGS
settings['n_chan_bin'] = 385
data_file = <path>
probe_name = 'neuropixPhase3B2_kilosortChanMap.mat'
clear_cache = False

ops, st, clu, tF, Wall, similar_templates, is_ref, est_contam_rate, kept_spikes = \
    run_kilosort(
        settings=settings, 
        filename=data_file,
        probe_name=probe_name, 
        clear_cache=clear_cache
        )

Error message:

No response

Version information:

Python 3.11.10 Kilosort 4.0.20 (using CUDA) Windows 10 (GPU: Nvidia RTX 3090) CUDA 11.8.89

MouseLand / Kilosort