cd /news/machine-learning/flux-1-schnell-on-8gb-vram-amd-no-cu… · home topics machine-learning article
[ARTICLE · art-35270] src=discuss.huggingface.co ↗ pub= topic=machine-learning verified=true sentiment=↓ negative

FLUX.1-schnell on 8GB VRAM (AMD, no CUDA): the GGUF format mismatch that wastes hours

A developer discovered that FLUX.1-schnell GGUF models from city96's repository fail to load in stable-diffusion.cpp on AMD RX 580 8GB via Vulkan, while leejet's builds work. The error message 'new_sd_ctx_t failed' misleads users into thinking it's a VRAM issue rather than a quantization-source mismatch. Switching to leejet's GGUF files and using specific offloading flags (CLIP on VRAM, T5XXL and VAE on CPU, VAE tiling enabled) enables inference on low-VRAM AMD cards.

read2 min views1 publishedJun 21, 2026

Sharing this because it cost me real time and I haven’t seen it documented clearly anywhere: if you’re trying to run FLUX.1-schnell through stable-diffusion.cpp (the C++ inference engine, not ComfyUI) on a low-VRAM AMD card via Vulkan, GGUF source matters and the error message doesn’t tell you why.

The setup: AMD RX 580 8GB (Polaris/GCN4, 2017 card — no ROCm support, no CUDA obviously), running stable-diffusion.cpp compiled with -DGGML_VULKAN=ON, no DirectML, no cloud.

The gotcha:

city96’s FLUX GGUF builds on HF (the ones most tutorials link, since they’re the most popular) — only work inside ComfyUI with the ComfyUI-GGUF custom node. They will NOT load in sd-server / the stable-diffusion.cpp CLI.

leejet’s FLUX GGUF builds — these are the ones built for stable-diffusion.cpp specifically and actually load.

Using a city96 file in sd-server just gives:

[ERROR] main.cpp:92 - new_sd_ctx_t failed

No further explanation, no hint that it’s a packaging/quantization-method mismatch rather than a VRAM or flag problem. I spent a while assuming it was a memory issue before realizing the file itself was the wrong build.

Once I switched to leejet’s FLUX.1-schnell-gguf repo, it loaded fine. For an 8GB card, the practical split that works:

Diffusion model on VRAM (~6.5GB for q4_k)

CLIP_L on VRAM (~235MB)

T5XXL and VAE offloaded to system RAM (–clip-on-cpu --vae-on-cpu)

–vae-tiling is NOT optional — without it, VAE decode OOMs even with everything else offloaded correctly.

Full command and timing breakdown (T5XXL conditioning ~11s, sampling ~14min at 4 steps/1024x1024, VAE decode ~40s) here if useful: GitHub - aivisionslab-studios/rx580-local-ai-guide: Complete guide to running local AI on AMD RX 580 8GB via Vulkan — llama.cpp, Ollama, OpenWebUI, Stable Diffusion. No CUDA. No cloud. Free. · GitHub

Posting mainly because “GGUF doesn’t load” is a generic enough error that it’s easy to misdiagnose as a VRAM or driver problem instead of a quantization-source mismatch. Curious if anyone’s hit the same wall with other GGUF-quantized diffusion models from different sources.

── more in #machine-learning 4 stories · sorted by recency
── more on @amd 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/flux-1-schnell-on-8g…] indexed:0 read:2min 2026-06-21 ·