Intel iGPU (Vulkan) detected but VRAM reported as 0 GB, node never gets "inference" capability

[carloalbertogiordano]

Summary

Nodes with an Intel integrated GPU (and likely any GPU that is only reachable via Vulkan rather than CUDA/ROCm) are permanently treated as GPU-less, even though the underlying hardware is fully capable of serving inference.

Two separate gaps combine to cause this:

1. Install script (agents.hyper.space/api/install) GPU detection only checks for nvidia-smi, rocm-smi, or macOS. There is no fallback check for a generic Vulkan-capable device, so on any Linux box with only an Intel (or non-ROCm AMD) GPU it prints No GPU detected — node will participate as relay only regardless of the actual hardware.
2. hyperspace system-info correctly identifies the Intel GPU (vendor + model name) but always reports VRAM: 0.0 GB for it. Because the effective-VRAM computation is fed from this field, the node is placed in "Recommended Tier: 0", and the inference capability is never added to hyperspace status — even after installing/configuring Ollama with Vulkan support enabled and confirming the exact same GPU can serve models at 100% GPU offload.

So the node is stuck as CPU-only/relay-only purely because of a detection gap, not an actual hardware limitation.

Impact

Any machine whose only GPU is an Intel iGPU (Iris Xe, Arc, etc. — a very common configuration on modern laptops) is silently excluded from:

• The Inference network capability (+10% weight, per the README's capability table) and its associated points.
• Local model sharding (hyperspace pod shard), since pod resources/sharding math is driven by the same VRAM figure.

This likely affects a large fraction of laptop nodes on the network, not just this one machine.

Steps to Reproduce

1. On a Linux machine with only an Intel integrated GPU (no NVIDIA/AMD), run the standard installer:

   curl -fsSL https://agents.hyper.space/api/install | bash

   Output includes:

   warn "No GPU detected — node will participate as relay only"
   

   (install script only branches on command -v nvidia-smi, command -v rocm-smi, or $OS == darwin).

2. hyperspace status afterward shows capabilities without inference:

   Capabilities: embedding, storage, memory, relay, validation, orchestration, caching, proxy, crawl, index, search, rank, matrix
   

3. Confirm the GPU is in fact Vulkan-capable and can fully serve inference:

   vulkaninfo --summary
   # → Intel(R) Iris(R) Xe Graphics (RPL-P) (Mesa driver), Vulkan 1.4

   Install/update Ollama (≥0.31) and enable its (experimental) Vulkan + iGPU flags via a systemd drop-in:

   # /etc/systemd/system/ollama.service.d/override.conf
   [Service]
   Environment="OLLAMA_VULKAN=1"
   Environment="OLLAMA_IGPU_ENABLE=1"

   sudo systemctl daemon-reload && sudo systemctl restart ollama
   ollama pull smollm2:135m
   ollama run smollm2:135m "Say hello in one short sentence."
   ollama ps
   # NAME            ... PROCESSOR    ...
   # smollm2:135m    ... 100% GPU     ...

   Ollama's own log confirms the device and correct memory figures:

   level=INFO msg="inference compute" id=0 library=Vulkan compute=0.0 name=Vulkan0 \
     description="Intel(R) Iris(R) Xe Graphics (RPL-P)" type=iGPU total="23.3 GiB" available="18.7 GiB"
   

4. Restart the hyperspace daemon so it re-detects Ollama and its models — hyperspace status now lists the Ollama models (ollama:smollm2:135m, ollama:llama3.2:latest) under "Models", proving the daemon does see Ollama. But capabilities still do not include inference, and the startup log still prints:

   [SYSTEM] Started — coordinator=false, gpu=false, model=Qwen/Qwen2.5-0.5B-Instruct, aggregator=false, checkpoint=false
   

   (note also the unsubstituted %s placeholders in this log line — looks like a format-string bug independent of the GPU issue).

5. Run hyperspace system-info directly — this is where the actual bug is visible:

   ┌─ GPU ────────────────────────────────┐
   │ GPU 0:      Intel(R) Iris(R) Xe Graphics (RPL-P)
   │   VRAM:     0.0 GB
   │   Vendor:   intel
   ├─ Allocation ─────────────────────────┤
   │ Effective VRAM: 0.0 GB
   │ Recommended Tier: 0 (0 GB)
   └──────────────────────────────────────┘
   

   Vendor and model name are detected correctly — only the VRAM figure is wrong.

Environment

• OS: Fedora Linux, kernel 7.0.11-200.fc44.x86_64
• CPU: 13th Gen Intel(R) Core(TM) i7-1360P (Raptor Lake-P)
• GPU: Intel(R) Iris(R) Xe Graphics (RPL-P), Mesa driver, Vulkan 1.4.341
• Hyperspace CLI: v2 (5.85.32)
• Bundled inference engine reported by daemon: prism-b9605-linux-x64-source
• Ollama: updated to v0.31.1 specifically to get Vulkan support

Additional Context

• Secondary finding, same root cause family: ~/.hyperspace/bin/prism/llama-server is compiled without any GPU backend at all (empty --list-devices, no vulkan/cuda strings in its libs), unlike upstream ggml-org/llama.cpp releases which already ship llama-b8148-bin-ubuntu-vulkan-x64.tar.gz. Dropping that vanilla Vulkan build + libs in place of prism/llama-server (same CLI surface — looks like the same/compatible source tree) fully offloaded a test model to Vulkan0: Intel(R) Iris(R) Xe Graphics (31/31 layers) and served /v1/chat/completions correctly. Not the actual gate on the inference capability, but means prism itself would need a Vulkan-enabled build to work standalone without Ollama.
• Suggested fix: when GPU vendor isn't NVIDIA/AMD, fall back to reading VRAM/available-memory from Ollama's own correctly-reported Vulkan figures, or query Vulkan heap size directly, instead of defaulting to 0.