GitHub Repo
Ollama community investigates rnj-1 loading failure; sliding-window attention metadata may trigger assertion
An October 10 report to Ollama said rnj-1 may abort during loading in newer runtimes, returning API error 500. The following day, a contributor traced the cause to an attention-check change in llama.cpp and offered an alternative GGUF, but an official fix and the full scope of impact remain unconfirmed.
The Ollama community received a report on October 10 that rnj-1 failed to load. On the 11th, a contributor added a technical diagnosis: the model file declares sliding-window attention (SWA) settings but does not actually use any SWA layers, causing a newer underlying runtime to trigger an assertion. As a result, a model that had previously worked aborts before generating an answer. Public report and contributor analysis
The case involved Linux, an Intel CPU, and Ollama 0.40.0. Logs show that after the request reached llama-server, the model memory initialization path hit GGML_ASSERT(hparams.is_swa_any()) failed. The process then exited, and /api/generate returned HTTP 500. The reporter said deleting and redownloading the model did not help. A contributor said the problem began with 0.34.4 and may be related to a llama.cpp upgrade. That upgrade was merged on September 22; the developments over the past two days are the failure report and its diagnosis. Underlying version update
According to the contributor, the old logic discarded an SWA layer list when all its entries were zero, while the new logic keeps the list and reports an error if no SWA layers are found. The rnj-1 file includes both window settings and a layer pattern whose values are all false, which meets this condition. The Unsloth model card also explains that the rnj-1 architecture is similar to Gemma 3 but uses only global attention and extends context with YaRN; this is consistent with the layer settings in the report. Model architecture details
The alternative proposed by the contributor is to use an Unsloth GGUF that does not declare SWA settings; they said this file loads successfully. The Hugging Face page provides the command ollama run hf.co/unsloth/rnj-1-instruct-GGUF:UD-Q4_K_XL. After switching quantized files, however, outputs and performance should be revalidated; the alternative should not be treated as an equivalent fix for the original model. Alternative model and instructions
For engineering teams maintaining local inference services, this case shows that upgrade validation needs to cover GGUF metadata compatibility as well as the actual loading and generation workflows. As of this review, the issue remains open, with no official fix delivered. Follow-up should track whether upstream changes how it handles all-zero SWA lists and whether other similar files are affected.