Tag: #chat-template6 results
Unsloth has refreshed all Gemma-4 uploads on Hugging Face with the updated chat template from Google's latest commit and possible other fixes. Users should redownload for improvements. Community thanks the team for rapid support post-Gemma-4 release.
A developer traced excessive cache misses in Qwen 3.5 to its chat template emitting empty historical <think> blocks, causing prompt drift after tool calls. The one-line fix prevents reprocessing of large contexts. PRs opened on Hugging Face repos.
Chat template for Gemma 4 GGUFs has been fixed recently. Users should update from recommended Hugging Face dealers like bartowski and unsloth. Various sizes available: 31B, 26B-A4B, E4B, E2B.
Qwen3.6 reliably maintains context like chosen numbers during Chain-of-Thought reasoning across iterations. It sometimes struggles to stick to numbers but works well with specific flag. Use --chat-template-kwargs '{"preserve_thinking": true}' to enable.
Provides a custom Jinja chat template to fix prompt reprocessing in Qwen 3.5 instruct mode (thinking disabled) on llama.cpp. Issue stems from empty think blocks being stripped from history, causing full reprocessing. Template now preserves empty tags while removing content-filled ones.
Users report no thinking/reasoning in Unsloth GGUF quants for Qwen3.5 small dense models regardless of parameters. Unsloth docs reveal reasoning is disabled by default for 0.8B, 2B, 4B, 9B models; enable via --chat-template-kwargs '{"enable_thinking":true}'. Bartowski quants work normally without tweaks.