[aittalam]

What

When Vulkan is explicitly requested (--gpu vulkan) but no usable device is found, llamafile already exits with a fatal error instead of silently downgrading to CPU. This keeps that behavior (a silent downgrade would surprise users with ~10× slower inference on an explicit GPU request) but replaces the dead-end message:

fatal error: support for --gpu vulkan was explicitly requested, but it wasn't available

with an actionable one that names the two real causes and the way forward:

fatal error: --gpu vulkan was explicitly requested but Vulkan is not usable on this system
  - no Vulkan-capable device was detected, or the driver failed/crashed during probe and was skipped for safety
  - retry with --gpu auto to use the best available backend (falls back to CPU), or --gpu disabled to force CPU

Why

The "driver failed/crashed during probe and was skipped" case is real on some Windows configurations (e.g. certain Intel iGPU setups, issue #938), where the out-of-process probe / crash guard (#988) correctly skips a faulting backend so the process stays alive. In --gpu auto that already falls back to CPU silently and correctly; only the explicit-request path surfaced a message that left users with no next step.

Scope

• Messaging only — no behavior change. --gpu auto fallback and the exit(1) on explicit request are unchanged.
• Touches llamafile/vulkan.c only (11 lines). The sibling cuda.c/metal.c messages could get the same treatment in a follow-up if desired.
• Compiles clean under cosmocc.

🤖 Generated with Claude Code