Name and Version

ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = AMD Radeon RX 6900 XT (AMD proprietary driver) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 0 | matrix cores: none
build: 7518 (c675292e3) with Clang 15.0.3 for Windows AMD64

Operating systems

Windows

Which llama.cpp modules do you know to be affected?

ggml-vulkan.dll

Command line

llama-mtmd-cli -m Qwen3-VL-8B-Instruct-UD-Q8_K_XL.gguf -mm mmproj-F16.gguf -sys "Describe provided images and translate any text contents." -p "Proceed" --no-warmup --image 512x4096.jpg -n 128 -fitt 2048

Problem description & steps to reproduce

The assert added in #18180 always trips when trying to clip a large image. Every model is affected. Reverting fd05c51 restores normal function and doesn't seem to break anything, nor was anything broken before the merge. Only images above a certain mass break: the hard limit is barely above 1024x1024 (interchangeable with 2048x512 and so on)

Tested with multiple Qwen3-VL and GLM models.

First Bad Commit

fd05c51

Relevant log output

ggml/src/ggml-vulkan/ggml-vulkan.cpp:5926: GGML_ASSERT(wg0 <= ctx->device->properties.limits.maxComputeWorkGroupCount[0] && wg1 <= ctx->device->properties.limits.maxComputeWorkGroupCount[1] && wg2 <= ctx->device->properties
.limits.maxComputeWorkGroupCount[2]) failed