Don’t rsync 400 GB of weights to a Synology

Created August 19, 2026 Last updated August 19, 2026 at 9:51 PM CDT

Same 46.05 GB GGUF · cp both ways · Forge (Linux) and M4 Max (macOS) · DS1823xs+ SMB · 10G on every NIC

For agents parking weights. A 46 GB cp write from Forge and from the M4 both held 170 MB/s (1.35–1.36 Gb) for the whole file. Reads were 17–26 MB/s. The RAID array is not the write cliff. rsync on a Mac SMB mount is still the wrong first copy. A 4.4 GB ditto at 8.65 Gb is not a 46 GB number.

Write, both hosts

170 MB/s

Read, Forge / M4

26 / 17 MB/s

File

46.05 GB

The 46 GB test

One file: GLM-5.2-UD-Q4_K_XL-00002-of-00011.gguf, 49,433,942,336 bytes. Sequential /bin/cp, dest deleted first, byte counts matched. Forge is ms01 at 192.168.1.19, CIFS 3.1.1, NIC 10G. M4 is 10Gbase-T, stock smbfs, COPYFILE_DISABLE=1.

Host	Direction	Seconds	MB/s	Gbps
Forge	NAS → local	1,902	26.0	0.21
Forge	local → NAS	290	170.3	1.36
M4 Max	NAS → local	2,942	16.8	0.13
M4 Max	local → NAS	293	169.0	1.35

Forge read wall is from process start (~8:09 PM) to dest mtime (8:40 PM). The wrapper died before JSON; the file size matched. Write and both M4 legs are timed in-process.

What that means

Two different 10G clients wrote the same 46 GB at the same 170 MB/s. That is a sustained sequential write, not a 4 GB cache splash. Earlier we guessed RAID6 disks could not hold a long write. That guess is wrong for this workload. The array can.

The slow side is read through cp. M4 was worse than Forge (16.8 vs 26.0 MB/s). Neither is the 10G link. If an agent is pulling a quant off the NAS with cp, that is the path that wastes the evening.

The smaller-file Mac numbers

Same idle miloshare mount, 4.4 GB shard, after a 1.5 TB park. Useful for command choice. Not a 46 GB rate.

Command	Best	Other pass
`rsync -a` (bulk 65–435 GB, then 4.4 GB)	166 MB/s	143–159 MB/s on the trees
`/bin/cp`	343 MB/s	253, then 205
`/usr/bin/ditto`	1,081 MB/s (8.65 Gb)	170 MB/s

ditto pass 1 matched Finder-by-eye. Pass 2 landed on the same 170 MB/s the 46 GB writes later held. Treat 8.65 Gb as a short-file / cache number. Treat 170 MB/s as the long write.

macOS rsync is openrsync 2.6.9. It writes through Apple’s SMB client. Do not pass GNU-only flags; they error and copy nothing.

What to tell your agent

export COPYFILE_DISABLE=1
ditto /path/to/model /Volumes/share/models/model   # Mac write
cp /path/to/model /mnt/share/models/model          # Linux write

Expect ~170 MB/s for a tens-of-GB write from either host we tested.
Do not use Mac rsync for the first 400 GB dump.
Do not promise Finder’s ~8 Gb on a 400 GB tree. We have not timed one.
Do not pull a large quant off the NAS with cp if you can avoid it. That read is 17–26 MB/s here.

August 19, 2026. Forge NIC enp2s0f1np1 10G full duplex. M4 en0 10Gbase-T. NAS DS1823xs+ miloshare SMB 3.1.1.