common : default GPU_MAX_HW_QUEUES=1 on HIP to fix idle GPU load

[liminfei-amd]

Overview

On HIP/ROCm, llama-server with --mmproj can leave the GPU pinned at high utilization and
maximum core clock while all slots are idle. This is a known ROCm/amdgpu MES multi-queue behavior
(see #3929, #21430) and is reported for the mmproj idle case in #23965. Limiting the number of HW
queues to 1 avoids the idle busy-loop.

In common_init(), for HIP builds only, default GPU_MAX_HW_QUEUES to 1 when the user has
not already set it. An explicitly-set value is always respected; non-HIP builds are untouched:

#if defined(GGML_USE_HIP)
    if (!std::getenv("GPU_MAX_HW_QUEUES")) {
        setenv("GPU_MAX_HW_QUEUES", "1", 0);
    }
#endif

Fixes #23965.

Additional information

Validated on real AMD ROCm hardware:

• Hardware: Radeon RX 9070 (gfx1201, RDNA4) · ROCm 7.2.1
• Before: with --mmproj, an idle server holds the core clock at ~3.2 GHz.
• After: with the default in place, an idle server settles to its low idle clock for the large
  majority of samples (a single transient boost spike is normal).

Note on the approach: this sets an environment default rather than changing mmproj logic. If
maintainers prefer documenting GPU_MAX_HW_QUEUES=1 for ROCm consumer GPUs instead of a code
default, I'm glad to switch to a docs-only change.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES — AI was used in an assistive capacity only (locating the relevant
  code paths and drafting wording). I authored and reviewed the change, understand it fully, and can
  explain every line; the fix and the on-hardware validation are mine. No AI-written code was
  submitted without manual review.

[liminfei-amd]

Friendly ping on this one.

The original reporter (@zzgundam) has since tested this change and confirmed it resolves the issue: they closed #23965 as completed, noting that "#24237 just fixed this problem".

The change is a small, HIP-only default (GPU_MAX_HW_QUEUES=1, set only when the user hasn't already set it), validated on a Radeon RX 9070 (gfx1201, ROCm 7.2.1). Would a maintainer be able to take a look when convenient? As noted above, if documenting GPU_MAX_HW_QUEUES=1 for ROCm consumer GPUs is preferred over a built-in default, I'm happy to switch to a docs-only change instead.

Thanks!