Two RTX 3090s still struggle to fit Qwen Image Edit alongside a 27B text model
Civil_Fee_7862 · reddit · 2026-07-30
Running Qwen Image Edit on two 3090s still hits VRAM limits
A user describes attempts to support image-generation tasks on a dual-RTX-3090 setup while also running a 4-bit Qwen3.6-27B text model with an 8-bit KV cache.
What they tried
- Reducing concurrent streams to one, which lowered GPU utilization enough to keep full 256k context length and left about 7 GB VRAM per card.
- Splitting component models across cards with tensor or pipeline parallelism, but the image-generation workload still pushed one card over 7 GB and triggered OOM.
- CPU offloading, which reduced VRAM usage from about 18.5 GB to around 12.8 MB under load, but only if the image model could also be split evenly across both GPUs.
Where they ended up
The author is considering buying a third RTX 3090 dedicated to image generation, because even if a 4-bit Qwen Image Edit model can be squeezed into roughly 14 GB, an 8-bit version likely will not fit. They ask whether anyone has built a similar multi-GPU setup, and whether there are PyTorch settings that make tensor or pipeline parallelism behave reliably for stable-diffusion-like models such as Qwen Image Edit.
More from Infra
- Baseten Launches Model Labs Platform for Closed-Model Monetization — baseten · 2026-07-30
- Baseten Launches Model Labs Platform for Commercializing Closed Models — baseten · 2026-07-30
- Replacing Cloud Vision APIs Locally with Nvidia Nemotron on DGX Spark — JFPuget · 2026-07-30
- Laguna XS Breaks 140 TPS on Apple Machines with New FAST Mode — gajesh · 2026-07-30
- $50B+ in AI Data Center Leases Signed in July as Bitcoin Miners Pivot — abhiadesai · 2026-07-30
- 26B-Parameter Gemma 4 Runs on Mac with 2GB RAM via SSD Streaming — petrusenko_max · 2026-07-30