ANE prediction failed: MultiArray shape (324 x 768) does not match the shape (256 x 768) specified in the model description

[yujonglee]

Reproduce

images.zip

APPLE_MODEL: lfm2-vl-450m-int8-apple.zip

CPU_MODEL: lfm2-vl-450m-int8.zip

---

export \
  APPLE_MODEL=/tmp/char-vlm-test/apple \
  CPU_MODEL=/tmp/char-vlm-test/non_ane \
  IMG_PASS=~/Downloads/black-264.png \
  IMG_FAIL=~/Downloads/black-272.png \
  VLM_PROMPT="Describe this image."

run_case(){ name=$1 model=$2 image=$3; printf '== %s ==\nmodel=%s\nimage=%s\n' "$name" "$model" "$image"; cargo run -q -p llm-cactus --example complete -- --model "$model" --prompt "$VLM_PROMPT" --image "$image" >"/tmp/${name}.out" 2>"/tmp/${name}.err"; rc=$?; printf 'exit=%s\n' "$rc"; tail -n 4 "/tmp/${name}.err"; echo; }; run_case apple_pass "$APPLE_MODEL" "$IMG_PASS"; run_case apple_fail "$APPLE_MODEL" "$IMG_FAIL"; run_case cpu_pass_small "$CPU_MODEL" "$IMG_PASS"; run_case cpu_pass_large "$CPU_MODEL" "$IMG_FAIL"

Result

== apple_pass ==
model=/tmp/char-vlm-test/apple
image=/Users/yujonglee/Downloads/black-264.png
exit=0

== apple_fail ==
model=/tmp/char-vlm-test/apple
image=/Users/yujonglee/Downloads/black-272.png
exit=1
[ERROR] [npu] ANE prediction failed: MultiArray shape (324 x 768) does not match the shape (256 x 768) specified in the model description
[ERROR] [complete] Exception: NPU encoder failed
error: request failed: HTTP 500 Internal Server Error
inference failed: cactus_complete failed (-1)

== cpu_pass_small ==
model=/tmp/char-vlm-test/non_ane
image=/Users/yujonglee/Downloads/black-264.png
exit=0
[WARN] [npu] not using NPU

== cpu_pass_large ==
model=/tmp/char-vlm-test/non_ane
image=/Users/yujonglee/Downloads/black-272.png
exit=0
[WARN] [npu] not using NPU