Make split mode graph work with vision enabled

[ikawrakow]

Closes #1353

[ubergarm]

It pulled tip of main this morning, but now with opencode client or through the built in web-ui I'm getting gibberish inference output (from tip of main@cda15bf1):

from main@14492bfd (this PR):

I tested with and without --mmproj but was the same.

I rolled back and tested each recent commit until it worked again at 666ea0e:

$ git log --oneline
cda15bf1 (HEAD -> main, upstream/main, upstream/HEAD) Discard very first compute graph for recurrent models (#1393)
f90b4c2f Full graph parallel for Qwen3.5 (dense and MoE) (#1388)
14492bfd Make split mode graph work with vision enabled (#1392) <--- first broken here
666ea0e9 Revise build instructions for ik_llama.cpp <--- working

Here is my command:

# full offload on 2x RTX A6000 48GB VRAM each
./build/bin/llama-server \
  --alias Qwen3.5-122B-A10B \
  --model "$model" \
  -fa on \
  -c 262144 \
  -sm graph \
  -ngl 99 \
  -ub 4096 -b 4096 \
  --parallel 1 \
  --threads 1 \
  --host 127.0.0.1 \
  --port 8080 \
  --jinja \
  --no-mmap

Not sure if anyone else is seeing this? I'll keep testing to see if I can narrow it down any more.

[ubergarm]

Removing --no-mmap fixes the issue as discussed in the linked PR ☝️

I tested and --mmproj is working and reading images correctly!

[ikawrakow]

@ubergarm

#1397 should fix the --no-mmap issue.