main
Restrict CORS to localhost by default to prevent drive-by API access
Add missing __init__.py to modules/grammar
Remove ExLlamaV2 backend
Minor cleanup
Sanitize character name in load_character to prevent path traversal
Make user_data configurable: add --user-data-dir flag, auto-detect ../user_data
Show live context size while generating
Reduce VRAM peak in prompt logprobs forward pass
Several small code simplifications
Minor fixes
Fix image generation: default to SDPA attention backend
Security: Fix SSRF vulnerabilities in URL fetching
Bind llama-server to parent process lifetime on Windows (closes #7574)
UI: Add MTP support, reorganize the speculative decoding menu
Electron: fix missing log colors on Windows
Logits display improvements (#7486)
fix: replace bare except clauses with except Exception (#7400)
Rename "truncation length" to "context length" in logs
Auto-enable MTP for MTP GGUFs
UI: refinements after c6a57a4 & electron user_data path
Move top_p and top_k higher up in the UI and CLI help
fix(security): sanitize filenames in all prompt file operations (CWE-22) (#7462)
Fix Gemma 4 thinking tags not hidden after tool calls (closes #7509)
Add adaptive-p sampler and n-gram speculative decoding support
Disable uncommonly used indented codeblocks (#7401)
UI: Fix token count not being set in non-streaming mode
Lint
Add tool call confirmation with inline approve/reject buttons
Remove obsolete DeepSpeed inference code (2023 relic)
Support .jinja/.jinja2 instruction template files in the UI (closes #7517)
Don't pass torch_dtype to transformers, autodetect from model config
Add an info message to the "Activate web search" checkbox
Major UI overhaul
Use `logger.exception()` instead of `traceback.print_exc()` for error messages
Electron: Refresh models dropdown when a new models directory is selected
Remove `backend="duckduckgo"` from ddgs (https://github.com/oobabooga/textgen/issues/7548#issuecomment-4433225429)