[Bug]: [MMAP] expert = viste zero-copy nei file (page cache = cache) when running in cluster

[lapchenkovbyu]

Commit or version

v1.10.2

Environment

PC -- Win11 CUDA build, 3060ti, 32ram, Ryzen 5 5600X
Laptop -- Linux Mint, 32gb RAM, intel core i7-10610U

Reproduction steps

./coli cluster coordinator --host 0.0.0.0 --port 8765 at Win host -- colibri cluster coordinator listening on http://0.0.0.0:8765
So, I expect to connect a worker normally, but running
./coli cluster worker --model "D:\models" --port 9100 --coordinator http://<WIN_HOST_IP>:8765 --advertise-host <WORKER_IP> results in
[MMAP] expert = viste zero-copy nei file (page cache = cache) cache/layer: expected a whole number, got "None"
at both Linux and Windows (which is the same-machine call).

Expected behavior

Both workers sucessfully connected to coordinator, creating a cluster

Actual behavior and logs

puts following into the console
`[MMAP] expert = viste zero-copy nei file (page cache = cache)
cache/layer: expected a whole number, got "None"`
and stops execution

[lapchenkovbyu]

Traced it down to here

colibri/c/glm53.c

Line 2600 in fd93c41

else if (argv[i][0] >= '0' && argv[i][0] <= '9') {

This one expects --cap int flag to be passed, and when nothing is passed, the Python script passes that var as "None", which is not a valid whole number, thus throwing cache/layer: expected a whole number, got "None"
To force through this, pass --cap 0 to allow automatic configuration.

[JustVugg]

Reproduced by reading, and you found a real one: coli cluster worker could never start without an explicit --cap.

--cap comes from the shared argument parser and defaults to None, which means "auto" everywhere else. The worker path stringified it straight into the engine's argv, and str(None) is the literal string "None", which the engine's own argument check refuses before it does anything at all:

return subprocess.call([engine, str(a.cap), str(a.ebits), str(a.dbits)], env=e)

That is why you got the identical message on the Linux laptop and the Windows box: it is not a cluster, network or model problem, it happens before any of that. Every other direct-engine launcher resolves the value through cap_for_launch first; this one did not.

Fixed in #1461, with a regression test that fails if the old line comes back.

One thing to know before you retry, because it will save you a second puzzle: the expert-worker path in colibri.c is guarded by #if !defined(_WIN32). With #1461 the Windows worker will start and then not take the worker branch, so plan on the coordinator being the Windows box and the workers being Linux, which is the shape your setup already has on the laptop side.

Also, [MMAP] expert = viste zero-copy nei file (page cache = cache) in your output is not an error, it is an informational line about the mmap reader. Sorry it reads like a fault; it is on the list to make it say so.

The fix will be in the next release. If you can build from dev before then, coli cluster worker with no --cap should get past this immediately.

[JustVugg]

Fixed by #1461 and shipped in v1.11.0: coli cluster worker passes a resolved cap instead of the literal string None. Closing; reopen if the worker still refuses to start on 1.11.0.