Only request VK_KHR_shader_bfloat16 extension if supported

[madsmtm]

Overview

My friend's Windows machine has an AMD Radeon(TM) 890M Graphics iGPU which doesn't support bf16.

In some cases though the code will currently still set device->coopmat_bf16_support = true, which later on causes it to request the VK_KHR_shader_bfloat16 extension (even though it isn't reported as supported), and the following error to be reported:

llama_model_load: error loading model: vk::PhysicalDevice::createDevice: ErrorExtensionNotPresent

This PR adds a bfloat16_support check to ensure that we don't try to do this if VK_KHR_shader_bfloat16 isn't supported. The fix has been tested by building with this diff applied, and successfully loading the model in this case.

Fixes #28161.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, used for debugging the issue, I've written the code and PR myself.

[madsmtm]

CC @0cc4m @jeffbolznv

[ggml-gh-bot]

Hi @madsmtm, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• Maintainers cannot push to this PR: Please enable Allow edits by maintainers. If this PR comes from an organization-owned fork, that option is not available on GitHub; please re-open the PR from a fork owned by your personal account.

• Multiple open PRs from a new contributor: We limit new contributors (those without a previously merged PR) to 1 open PR at a time. You currently have 2 open PRs.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[0cc4m]

Open an issue with logs and description of what goes wrong instead.

[madsmtm]

Sure, I've opened #28161.

[0cc4m]

Thank you, now I understood the problem. Your fix is basically correct, the one thing missing is to remove the duplicated device_extensions.push_back("VK_KHR_shader_bfloat16"); block that is added if coopmat_bfloat16_support is set.

[madsmtm]

Ofc, I missed that the extension is already added earlier