Update TensorRT-LLM

[kaiyux]

• Model Support
  • Add example for multimodal models (BLIP with OPT or T5, LlaVA)
• Features
  • Smooth Quantization support for ChatGLM2-6B / ChatGLM3-6B / ChatGLM2-6B-32K
  • Out-of-the-box support for the QWEN model
  • Support for returning context and/or generation logits in the Triton backend
• API
  • Add a set of High-level APIs for end-to-end generation tasks, the features are as below
    • ModelConfig() as a clean configuration interface for LLM tasks
    • LLM() for LLM pipelines, it will trigger the necessary engine building or model quantization silently in the background
    • generate() API for batched offline inference, both single-GPU and multi-GPU supported
    • generate_async() API for asynchronous offline inference on a single GPU, streaming mode is supported
• Bug fixes
  • Add pickle support for InferenceRequest GptManager pybind 2/4TP run demo #701
  • Fix Mixtral-8x7b build failure with custom_all_reduce Mixtral-8x7b build fails with custom_all_reduce #825
• Performance
  • Performance optimization of beam search kernel
  • Increase default freeGpuMemoryFraction parameter from 0.85 to 0.9 for higher throughput
• Documentation
  • Add documentation for best practices for tuning the performance of TensorRT-LLM (See docs/source/perf_best_practices.md)
  • Add documentation for Falcon AWQ support (See examples/falcon/README.md)