ggml : check for allocation failures to prevent crashes

[sanmai]

Overview

• The result of ggml_gallocr_reserve_n wasn't considered so ggml_gallocr_alloc_graph crashed on buffers that weren't allocated
• The return value from ggml_backend_sched_reserve wasn't considered either, caused SIGABRT

Related to #27817

Additional information

build/bin/llama-mtmd-cli -m qwen3.8-27b-q4_0.gguf --mmproj Qwen3.8-mmproj-BF16.gguf \
  --image apple.jpg --image tmp/big.jpg -p describe \
  --image-min-tokens 1024 -dev CUDA0 -mmdev Vulkan1 -fitt 1200

Here -fitt 1200 leaves us with 1200 MiB free for 888 MiB mmproj plus 123 MiB for the first image but not 480 MiB second image.

A command to create the second image:

magick -size 3000x3000 plasma:fractal -quality 90 tmp/big.jpg

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, I used AI to debug and find the place for the fix.

[abra-code]

upvote