• Apply flash attention on vision encoder for lower first-token latency.
• Fix metal compilation error on Apple silicon chips.

• Support GLM4V, the first vision language model in GLM series
• Fix nan/inf logits by rescheduling attention scaling

• Dynamic memory allocation on demand to fully utilize device memory. No preset scratch size or memory size any more.
• Drop Baichuan/InternLM support since they were integrated in llama.cpp.
• API change:
  • CMake CUDA option: -DGGML_CUBLAS changed to -DGGML_CUDA
  • CMake CUDA architecture: -DCUDA_ARCHITECTURES changed to -DCMAKE_CUDA_ARCHITECTURES
  • num_threads in GenerationConfig was removed: the optimal thread settings will be automatically selected.

• Fix regex negative lookahead for code input tokenization
• Fix OpenAI API server by using apply_chat_template to calculate tokens

Support ChatGLM4 conversation mode

• Support p-tuning v2 finetuned models for ChatGLM family
• Fix convert.py for lora models & chatglm3-6b-128k
• Fix RoPE theta config for 32k/128k sequence length
• Better cuda cmake script respecting nvcc version

• Support function calling in OpenAI api server
• Faster repetition penalty sampling
• Support max_new_tokens generation option

• Full functionality of ChatGLM3 including system prompt, function call and code interpreter
• Brand new OpenAI-style chat API
• Add token usage information in OpenAI api server to be compatible with LangChain frontend
• Fix conversion error for chatglm3-6b-32k

• Support ChatGLM3 in conversation mode.
• Coming soon: new prompt format for system message and function call.

• Support InternLM 7B & 20B model architectures