The Vulkan backend does not support the UPSCALE operator.

[tdtgi]

Pre-checks

• I searched existing issues and discussions

What problem are you trying to solve?

The Vulkan backend does not support the UPSCALE operator, which results in slow inference speeds for certain models (such as Qwen3VL).
warmup: *****************************************************************
warmup: WARNING: the CLIP graph uses unsupported operators by the backend
warmup:          the performance will be suboptimal
warmup:          list of unsupported ops (backend=Vulkan0):
warmup:          UPSCALE: type = f32, ne = [92 92 1024 1]
warmup: flash attention is enabled
warmup: please report this on github as an issue
warmup: ref: https://github.com/ggml-org/llama.cpp/pull/16837#issuecomment-3461676118
warmup: *****************************************************************

What would you like NexaSDK to do?

Add support for the UPSCALE operator in Vulkan backend.

Alternatives you've considered

Who does this help, and how much?

Additional context

Refer to this link:https://github.com/ggml-org/llama.cpp/pull/18327#event-21729873367

[iwr-redmond]

As this is an upstream issue, it may be appropriate to mark it here as wontfix.