Update TensorRT-LLM

[kaiyux]

• Features
  • Support the pipeline parallelism cases when the number of layers cannot be divided by PP size.
  • Add LoRA support to Qwen2, see “Run models with LoRA” section in examples/qwen/README.md.
  • Add support for Phi-3-mini/small FP8 base + FP16/BF16 LoRA, see “Run Phi-3 with LoRA” section in examples/phi/README.md.
  • Add support for starcoder-v2 FP8 base + FP16/BF16 LoRA, see “Run StarCoder2 with LoRA” section in examples/gpt/README.md.
  • Add numQueuedRequests to the iteration stats log of the executor API.
  • Support FP8 OOTB MoE.
  • Add concurrency argument for gptManagerBenchmark.
• API
  • [BREAKING CHANGE] Remove attention_qk_half_accumulation knob from trtllm-build command.
  • [BREAKING CHANGE] Add a runtime max_num_tokens knob to the ExecutorConfig and gptManagerBenchmark.
  • [BREAKING CHANGE] The default value of max_seq_len is read from the HuggingFace mode config now.
  • [BREAKING CHANGE] Several refactors to the Python high level API, see examples/high-level-api/README.md.
  • Update the apps examples using the LLM APIs, please refer to the examples/apps/READEME.md for details.
• Bug fixes
  • Fix stop and bad words list contiguous for ModelRunner [ModelRunner] Fix stop and bad words list contiguous for offsets #1815, thanks to the contribution from @Marks101.
  • Fix missing comment for FAST_BUILD, thanks to the support from @lkm2835 in Add FAST_BUILD comment at #endif #1851.
  • Fix the issues that Top-P sampling occasionally produces invalid tokens. Top-P sampling occasionally produces invalid tokens #1590
• Infra
  • The dependent ModelOpt version is updated to v0.13.0.