SQL 2016 SQL Server Assertion: File: <pageref.cpp>, line=951 Failed Assertion

Looking at the KB you have a couple of options / workarounds:

1. Switch to FULL recovery model. You say "this isn't much of an option for a warehouse" but it's really just a matter of setting up the transaction log backups on a regular basis eg 15 minutes and then disposing of them. SSIS / Maintenance Plans have stock tasks for doing this. You will lose bulk-logged transactions, but I've never found these have made a great difference in runtimes, just log size. You can even backup log to nul which I won't describe here. If you're not sure what to do, ask you local DBA. Disk space and transaction log backup retention are easier problems to solve than fatal errors. When this problem eventually gets solved you can switch back.
2. The KB mentions "multiple BULK INSERT statements in a single-distributed transaction". It's not clear from your question how your bulk inserts are set up. Are you using SSIS to run 'Execute SQL' tasks which use the MERGE command? What does 'multiple BULK INSERTs' mean here? Is there a way to convert your approach to single BULK INSERTs, one at a time for example? In SSIS you can set the 'MaxConcurrentExecutables' to 1 temporarily for example, see if that helps. Tie it in to a config variable so you can change it back later. Obviously it will slow things down but you prefer your ETL to finish rather than fail quickly. Doing things in parallel is a good pattern and a real strength of SSIS, but you can only go as fast as your slowest component; say you have 10 dimensions which take a minute and one fact which takes an hour, your ETL finishes in an hour running parallel or 1 hour 10 minutes running serially.
3. MERGE is nice but does have a few issues. You could consider converting back to INSERT/UPDATE. Also you should be using HOLDLOCK with MERGE according to here. Do you use that hint? Does it make any difference to this issue if you do? We had an issue in an early SQL 2014 build where using MERGE with composable DML (OUTPUT clause) into columnstore caused this kind of assertion - I made them take the columnstore indexes off the dimensions, which they had added without telling me.
4. What kind of transaction handling are you doing? Sometimes with ETL the position is repeatable by just re-running. Sometimes you need it to fail and rollback. How have you implemented this? Can be altered so it's not "single distributed transaction"?

Good luck.

I believe this has been resolved in 2016 SP1 CU1.

Specifically by KB3205964