fix: simplify msgs format for minicpm_v .chat() to prevent empty hyphen outputs

fix: refactor minicpm_v to use official .chat() method to prevent hallucinated OCR text

fix: use eager attention for MiniCPM4.1-8B to prevent SDPA crash, fix Nemotron Parse processor format

fix: switch Core Router to AutoModelForCausalLM for generate support, fix Nemotron Parse inference with AutoProcessor

fix: block flash_attn at runtime with sys.modules=None to prevent FA2 on HF ZeroGPU where flash_attn is pre-installed

fix: resolve infinite recursion in flash_attn patch by capturing original get_imports before patching

fix: replace flash_attn mock with get_imports patch to properly disable FA2 for MiniCPM4.1-8B and Nemotron Parse

fix: use correct AutoModelForImageTextToText+AutoProcessor for MiniCPM-V-4.6, eager attn for Core Router

fix: use AutoModelForCausalLM for vision, add sdpa attn for Core Router, add open_clip_torch dep

fix: correct vision model dtype, Nemotron Parse model ID, and patch missing is_torch_fx_available

fix: correct MiniCPM-V-4.6 model ID and switch Core Router to AutoModel to resolve ModelWrapper crash

feat: add Parallel Vision Engine — MiniCPM-V and Nemotron Parse fire async with confidence-score arbiter

fix: add __spec__ to the mocked flash-attn module

fix: bypass flash-attn compilation by mocking it and using sdpa

fix: pass HF_TOKEN to access gated MiniCPM-V model

feat: Integrate MiniCPM-V 2.6 OCR model to vision tool

feat: Fix Gradio warnings and implement ZeroGPU startup hook for vision pipeline