{"slug": "gemma-4-bug-fixes-and-research-request", "title": "Gemma 4 bug fixes and Research Request", "summary": "A critical bug in Google's Gemma 4 causes it to malform tool calls under real load, affecting vLLM, llama.cpp, Ollama, and oobabooga. A developer open-sourced a diagnosis, repair, and experimental LoRA that reduces but doesn't eliminate the issue, calling for community help with richer data.", "body_md": "Gemma-4 has an ecosystem-wide agentic bug: under real load (long context + reasoning + generating\n\ncontent) it malforms its own tool calls and then loops on the broken output — reported across vLLM,\n\nllama.cpp, Ollama, and oobabooga, and unfixed by Google.\n\nI open-sourced a full diagnosis + a reproduction harness + a working parser-level repair + a stock\n\nNVFP4 quant & recipe + an experimental format LoRA — with all raw data. Honest findings: the LoRA\n\n*reduces* the slip (I ruled out quantization as the cause via a BF16 control) but doesn’t fully\n\neliminate it yet — it overfit my synthetic data, so a bigger/more-diverse dataset is likely the real fix.\n\nEverything’s here, Built with Gemma — looking for help pushing the LoRA further with richer data", "url": "https://wpnews.pro/news/gemma-4-bug-fixes-and-research-request", "canonical_source": "https://discuss.huggingface.co/t/gemma-4-bug-fixes-and-research-request/176979#post_1", "published_at": "2026-06-19 10:44:10+00:00", "updated_at": "2026-06-19 11:13:41.403223+00:00", "lang": "en", "topics": ["large-language-models", "ai-agents", "ai-tools", "ai-research", "ai-safety"], "entities": ["Google", "Gemma 4", "vLLM", "llama.cpp", "Ollama", "oobabooga", "NVFP4", "LoRA"], "alternates": {"html": "https://wpnews.pro/news/gemma-4-bug-fixes-and-research-request", "markdown": "https://wpnews.pro/news/gemma-4-bug-fixes-and-research-request.md", "text": "https://wpnews.pro/news/gemma-4-bug-fixes-and-research-request.txt", "jsonld": "https://wpnews.pro/news/gemma-4-bug-fixes-and-research-request.jsonld"}}