Windows CUDA discovery fails (0 B VRAM / CPU fallback) on NVIDIA RTX 50-Series (Blackwell) with Driver 616.92

What is the issue?

Description

On Windows, Ollama fails to discover the VRAM pool or initialize native CUDA acceleration for NVIDIA's newer Blackwell architecture when running on modern graphics drivers. The engine records total_vram="0 B" and falls back strictly to the CPU.

Regression Note: This dual-GPU setup was working perfectly under native CUDA acceleration previously, achieving 40–50+ tokens/sec on qwen3.8:27b (a 21 GB, Q4_K_M layout model running with a 65,536 context length allocation). The failure loop was triggered directly by either the upgrade to the Ollama v0.34.x branch or installing the NVIDIA September 2026 driver update.

However, manually forcing the backend to utilize Vulkan successfully maps and groups the dual-card memory pool.

Environment

• Ollama Version: v0.33.3 and v0.34.2
• GPU: Dual NVIDIA GeForce RTX 5060 Ti 16GB (Mixed array: 1x ASUS PRIME, 1x ASUS DUAL)
• Display Setup: Internal Intel UHD Graphics 770 iGPU is driving both physical monitors. The dual discrete RTX 5060 Ti cards have no displays attached (Compute-only).
• Driver Version: Studio Driver 616.92 (CUDA UMD Version 13.4)

Steps to Reproduce

1. Inject standard CUDA variables: $env:OLLAMA_LLM_LIBRARY="cuda" and $env:CUDA_VISIBLE_DEVICES="0,1"
2. Execute ollama serve
3. Observe the hardware scan log lines.

Temporary Working Bypass (Vulkan Layer)

The cards can be forced into action by bypassing CUDA and shifting the index paths away from the integrated processor to target the discrete Vulkan nodes:

\$env:OLLAMA_LLM_LIBRARY="vulkan"
\$env:GGML_VK_VISIBLE_DEVICES="1,2" # Skips Intel Vulkan0 node
ollama serve

Relevant log output

time=2026-09-21T23:29:57.028-06:00 level=INFO source=runner.go:60 msg="discovering available GPUs..."
time=2026-09-21T23:29:57.041-06:00 level=WARN source=runner.go:722 msg="user overrode visible devices" CUDA_VISIBLE_DEVICES=0,1
time=2026-09-21T23:29:57.041-06:00 level=WARN source=runner.go:726 msg="if GPUs are not correctly discovered, unset and try again"
time=2026-09-21T23:29:57.041-06:00 level=INFO source=types.go:50 msg="inference compute" id=cpu library=cpu compute="" name=cpu description=cpu libdirs=ollama driver="" pci_id="" type="" total="31.8 GiB" available="19.5 GiB"
time=2026-09-21T23:29:57.041-06:00 level=INFO source=routes.go:2050 msg="vram-based default context" total_vram="0 B" default_num_ctx=4096

OS

Windows

GPU

Nvidia

CPU

Intel

Ollama version

0.34.2