Update TensorRT-LLM

[kaiyux]

• Model Support
  • Support Neva
  • Support Kosmos-2
• Features
  • Support quantization for Nemotron models
  • Add LoRA support for Mixtral and Qwen
  • Add weight-stripping feature
    • A new command trtllm-refit is added
    • See documentation at examples/sample_weight_stripping/README.md
  • Add weight streaming feature
    • See documentation at docs/source/advanced/weight-streaming.md
  • The Python high level API
    • Add embedding parallel, embedding sharing and fused MLP support
    • Enable the usage of executor API
  • Add in-flight batching support for ChatGLM models
  • Add support to ModelRunnerCpp so that it runs with the executor API for IFB-compatible models
• API
  • [BREAKING CHANGE] Refactor scheduling configurations
    • Unify the SchedulerPolicy with the same name in batch_scheduler and executor, and rename it to CapacitySchedulerPolicy.
    • Expand the existing configuration scheduling strategy from SchedulerPolicy to SchedulerConfig to enhance extensibility. The latter also introduces a chunk-based configuration called ContextChunkingPolicy.
  • [BREAKING CHANGE] Remove use_context_fmha_for_generation argument from trtllm-build command since it’s not used anymore
  • [BREAKING CHANGE] Input prompt is removed from the generation output in the generate() and generate_async() APIs.
    • E.g. Given a prompt as A B, the original generation result could be <s>A B C D E where only C D E is the actual output, and now the result is C D E.
  • [BREAKING CHANGE] Switch default add_special_token in the TensorRT-LLM backend to True make add_special_tokens/skip_special_tokens default value is true which align with hf setting triton-inference-server/tensorrtllm_backend#446, thanks to the contribution from @XiaobingSuper , the changes are integrated in Update TensorRT-LLM backend triton-inference-server/tensorrtllm_backend#454.
  • GptSession and TrtGptModelV1 are marked as deprecated
• Bug fixes
  • Fix a NVRTC runtime ABI compatibility issue
• Performance
  • [BREAKING CHANGE] Set default tokens_per_block argument of trtllm-build command to 64 for better performance
  • Enhance the multiple profiles feature, multiple_profiles argument in trtllm-build command builds more optimization profiles now for better performance
  • Enhance the custom AllReduce by adding a heuristic: fall back to use native NCCL kernel when hardware requirements are not satisfied to get the best performance
  • Optimize the performance of checkpoint conversion process for LLaMA
• Documentation
  • Add documentation for KV cache reuse feature, see docs/source/kv_cache_reuse.md