I'm using edgeR to calculate CPM values for bulk RNA-seq samples. I have >100 samples from a diverse cohort, with each sample collected, library prepped, and sequenced at different time points throughout the year. For example 1-10 samples could had been prepped in Jan, and 3 more in Feb, 2 mid-Feb, etc... Moreover I do not have information as to when these samples were processed per se.

Given this situation, should I merge all the data together and use TMM normalization to calculate the normalization factors before producing CPM counts? Or should I calculate CPM individually for each sample without using TMM normalization?

thanks in advance.

If you wish to perform DE analyses using the samples then you should perform TMM normalization. Whether or not the samples are diverse or sequenced at different times is irrelevant, unless that prevents you from undertaking a DE analysis with all the samples.

In addition to Gordon's point, in my experience batch effects from different batches at the same sequencing center are usually quite minor. We have analyzed large data sets that were run at different times over the years due to the logistics of processing the samples. While it is important to ensure that you randomize to the batches appropriately so your batch effects are orthogonal to the coefficient of interest, running samples in several batches is not likely to be problematic.