fix(server): use GTT memory for Strix Halo VRAM monitoring

[jane-alesi]

What

Fix false VRAM pressure readings on AMD Strix Halo (gfx1105) and Strix Point (gfx1102) systems.

The VRAM monitor only reads mem_info_vram_* (2GB LPDDR5 chip) and ignores mem_info_gtt_* (124GB GTT). On Strix Halo, models run entirely in GTT — the LPDDR5 is a tiny buffer.

Symptoms Before Fix

• 600+ log lines/min from nvidia-smi subprocess spawning
• "VRAM pressure critical (99.0% >= 90.0%)" every 2 seconds
• False eviction triggers when LPDDR5 has transient load spikes
• Models evicted incorrectly despite 0.13% actual memory pressure

Symptoms After Fix

• 0 VRAM-related log lines/min (on AMD systems)
• Correct memory pressure reading: 0.13% (was 99%)
• No false evictions
• nvidia-smi subprocess skipped entirely on AMD

Technical Details

Strix Halo uses a unified memory architecture. The kernel exposes two memory pools:

• mem_info_vram_* — 2GB LPDDR5 chip (tiny buffer)
• mem_info_gtt_* — 124GB GTT (where models actually live)

The fix detects unified-memory GPUs by checking KFD gfx_target_version (gfx1102, gfx1105) and computes pressure as (vram_used + gtt_used) / (vram_total + gtt_total).

Discrete AMD GPUs (7900 XTX, etc.) and NVIDIA GPUs use the existing path unchanged.

nvidia-smi skip: Uses PCI vendor ID scanning (/sys/bus/pci/devices/*/vendor == 0x10de) — more robust than KFD-only check since KFD may not be present on all AMD systems.

Files Changed

• src/cpp/include/lemon/system_info.h — header-only has_unified_memory() inline function
• src/cpp/include/lemon/system_info.h — is_strix_apu() declaration in SystemInfo class
• src/cpp/server/system_info.cpp — GTT-aware VRAM calculation + PCI-based nvidia-smi skip
• src/cpp/server/system_info.cpp — is_strix_apu() implementation (KFD gfx_target_version scan)
• test/cpp/test_amd_vram.cpp — unit tests for has_unified_memory() and vram ratio calculation
• CMakeLists.txt — test target for test_amd_vram

Related Issues

• refactor(server): replace goto with structured control flow in get_global_vram_usage_pct() #2 — refactor(server): replace goto with structured control flow
• test(server): add test coverage for is_strix_apu() sysfs scanning + rename to is_apu() #3 — test(server): add test coverage for is_strix_apu() sysfs scanning
• refactor(server): future-proof APU detection beyond Strix gfx versions #4 — refactor(server): future-proof APU detection beyond Strix gfx versions