feat: add Modal deployment for FLUX.2 klein 9B image generation endpoint

untrack test image, ignore generated images

fix: improve weight directory validation to ensure critical shards exist before initialization and add test asset

feat: add FLUX.2 Klein image generation service and enable flash-attention for MioTTS

chore: update Modal GPU requirement from T4 to A10G for MioTTS inference

refactor: standardize ASR endpoint to /transcribe and update MiniCPM-V dependency and video decoding logic

feat: add MiniCPM-V 4.6 vision inference endpoint and update Nemotron-ASR to Python 3.11 with regex-based transcription parsing.

feat: transition to cache-aware streaming inference by cloning NeMo and running external inference script

feat: add Nemotron ASR backend for speech transcription and update persona skills to include Japanese teaching

refactor: increase llama context window, remove redundant output format parameter, and add error logging to MioTTS requests

refactor: rename searxng file reference and clarify persona emoji requirements

refactor: scope memory management and system prompts to individual user_id sessions throughout the chat pipeline

feat: add Fish Speech TTS backend support and improve MioTTS preset management and UI layout

refactor: switch to prebuilt llama.cpp CUDA binaries and remove obsolete build logic

fix: resolve llama.cpp build failure via CMake no-op and simplify login modal CSS styles

fix: patch llama.cpp build to remove ui dependencies and add CSS styles for login modal

fix: disable llama-server webui build to prevent compilation errors from empty build.json assets

refactor: decouple ASR processing from HTTP endpoint and update Modal environment to CUDA 12.3 registry image

Add missing CUDA

Add missing CUDA

refactor: move FastAPI application logic to subprocess for improved Modal container lifecycle management

feat: add huggingface secret to ASRServer container configuration

feat: replace search implementation with ASR service and modularized SearXNG backend

feat: implement Modal deployment script for MioTTS inference backend

feat: implement Modal-based SearXNG search backend with custom configuration

refactor: optimize token streaming logic in llm.py for improved response latency

Add tool call jinja

Switch to Ministral3-8B model

Add debug print

fix: add ddgs to Modal image pip_install

feat: add Modal search endpoint

Modal backend to inference LLM

Add Modal backend to inference LLM