fix: LD_PRELOAD NVML stub to suppress CUDA detection on hybrid AMD+NVIDIA systems

[bong-water-water-bong]

Problem

On laptops with both AMD iGPU and NVIDIA dGPU, vLLM's platform detection finds both CUDA (pynvml sees the dGPU via NVML) and ROCm and raises:

RuntimeError: Only one platform plugin can be activated, but got: ['cuda', 'rocm']

Why env var approaches don't work

• CUDA_VISIBLE_DEVICES="" — NVML bypasses the CUDA runtime. It talks directly to the NVIDIA kernel driver via libnvidia-ml.so, so CUDA_VISIBLE_DEVICES has no effect on nvmlDeviceGetCount().
• CUDA_VISIBLE_DEVICES="" + HIP_VISIBLE_DEVICES=0 — Same problem; NVML doesn't respect either variable.
• Python monkey-patching — Fragile to vLLM internal API changes (function renames, module restructures).

Solution: LD_PRELOAD NVML stub

A 25-line C file compiled to a shared library. When LD_PRELOADed, it intercepts:

NVML function	Stub behavior
nvmlInit()	Returns error
nvmlInit_v2()	Returns error
nvmlInitWithFlags()	Returns error
nvmlDeviceGetCount()	Returns 0 devices

This makes pynvml see zero NVIDIA GPUs, so the CUDA platform plugin never activates — while ROCm probes through HIP/ROCm libraries normally.

Why this is robust

• Does not modify upstream wheel contents — respects the repo's inviolable principles
• NVML C ABI is stable — these function signatures haven't changed in years
• Works at the dynamic linker level — can't be bypassed by any Python code path
• Minimal — 25 lines of C, one gcc -shared invocation

How it works end-to-end

1. Lemonade sets VLLM_TARGET_DEVICE=rocm in the vLLM subprocess environment (PR #2459)
2. Launcher detects VLLM_TARGET_DEVICE=rocm → LD_PRELOADs the stub
3. vLLM starts → pynvml sees zero NVIDIA GPUs → CUDA platform inactive → ROCm activates ✓

Companion PR

• lemonade: fix: pick best AMD GPU (dGPU > iGPU) in get_rocm_arch() instead of first available lemonade#2459

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

[bong-water-water-bong]

Companion to lemonade#2598 — same class of hybrid AMD+NVIDIA GPU crash, applied at the llama.cpp backend level there and at the vLLM platform-detection level here. Both fixes use the same strategy: hide NVIDIA GPUs from the ROCm path rather than asking users to blacklist kernel modules or disable hardware.

Rebased and ready for review — would appreciate a look from the lemonade-sdk team.