Ollama Community Reports Dual-Model Loading Crash, GPU Allocation Error During Qwen3-VL Image Encoding

On October 6, the Ollama community reported a multi-model inference failure: with another large model already loaded, qwen3-vl:8b crashed the underlying llama-server three times in about 45 seconds while processing an image. The same request completed successfully on the fourth launch. At the time of review, the issue remained open and listed no related fix. Original report

The case used Ollama 0.33.3 on Arch Linux with two RTX 3090 GPUs, each with 24 GB of VRAM. The 27B model loaded first was using about 12 GB on one GPU and 16 GB on the other. An image-enabled Qwen3-VL request then launched a second inference process. Two failures were recorded as segmentation faults from reading address zero; another explicitly reported a CUDA resource allocation failure. The stack trace passes through clip_encode and ends at ggml_gallocr_alloc_graph, pointing the investigation toward allocation for the image-encoding computation graph. Error logs and stack traces

The reporter hypothesized that VRAM pressure could cause a buffer allocation to fail, followed by a null-pointer dereference. But the author did not disassemble the faulting instruction, so this remains unverified. The same request succeeding after a restart is also insufficient to establish a root cause or rule out other environmental factors. The evidence comes from a single setup and does not show that the issue affects all Qwen3-VL deployments. Case limitations

The technical question to watch is whether multi-model scheduling accounts for resource needs during image encoding. Ollama’s official documentation says that when there is not enough memory to load a new model, requests should be queued until an existing model is unloaded; loading models concurrently on GPUs is also subject to available VRAM. This documentation provides a reference for expected behavior, but does not confirm that this case violates scheduling rules or identify the location of a defect. Official concurrency documentation

Engineers deploying text and vision models can first compare single-model and dual-model scenarios, while recording image dimensions, context settings, and per-GPU VRAM usage. They can also use OLLAMA_MAX_LOADED_MODELS to limit the number of models loaded at once as a diagnostic experiment. These are troubleshooting suggestions based on the report, not verified fixes. Next, watch for whether maintainers can reproduce the issue, whether allocation failures can be returned safely as errors, and whether a later version provides a clear fix. Configuration documentation