EGYADMIN/kimi-k2-thinking-dev

Fix: Complete MODEL_NAME string (add .3")

Fix: Add missing DEFAULT_SYSTEM_PROMPT variable

Switch to Mistral 7B Instruct - compatible and ready to fine-tune

Switch to Qwen 2.5 32B Instruct - powerful and compatible model

Switch back to original moonshotai/Kimi-K2-Instruct model

Switch to quantized model RedHatAI/Kimi-K2-Instruct-quantized.w4a16

Switch to Hugging Face Inference API for efficient model access

Optimize app.py with better memory management and simplified code

Remove model pre-download from Dockerfile - model will be downloaded at runtime

Fix indentation issues in download_model.py

Update Dockerfile to pre-download model during build

Create download_model.py

Switch to Kimi-K2-Instruct model for better compatibility

Fix Gradio 6.x compatibility - remove show_copy_button and theme from Blocks

Update app.py with improved UI and error handling

Update Dockerfile with optimized settings for Kimi-K2

Add compressed-tensors and update dependencies for Kimi-K2

Apply patch BEFORE transformers import to fix is_torch_fx_available error

Fix import error by patching is_torch_fx_available function

Add tiktoken dependency for model tokenizer

Add application file to load Kimi-K2-Thinking model

Add requirements for model loading

Add Dockerfile for development environment

initial commit