TensorRT-LLM 0.6.1 Releases

[kaiyux]

Hi,

We are very pleased to announce the 0.6.1 version of TensorRT-LLM. It has been an intense effort, and we hope that it will enable you to easily deploy GPU-based inference for state-of-the-art LLMs. We want TensorRT-LLM to help you run those LLMs very fast.

This update includes:

• Models
  • ChatGLM3
  • InternLM (contributed by @wangruohui)
  • Mistral 7B (developed in collaboration with Mistral.AI)
  • MQA/GQA support to MPT (and GPT) models (contributed by @bheilbrun)
  • Qwen (contributed by @Tlntin and @zhaohb)
  • Replit Code V-1.5 3B (contributed by @bheilbrun)
  • T5, mT5, Flan-T5 (Python runtime only, contributed by @mlmonk and @nqbao11)
• Features
  • Add runtime statistics related to active requests and KV cache utilization from the batch manager (see the batch manager documentation)
  • Add sequence_length tensor to support proper lengths in beam-search (when beam-width > 1 - see tensorrt_llm/batch_manager/GptManager.h)
  • BF16 support for encoder-decoder models (Python runtime - see examples/enc_dec)
  • Improvements to memory utilization (CPU and GPU - including memory leaks)
  • Improved error reporting and memory consumption
  • Improved support for stop and bad words
  • INT8 SmoothQuant and INT8 KV Cache support for the Baichuan models (see examples/baichuan)
  • INT4 AWQ Tensor Parallelism support and INT8 KV cache + AWQ/weight-only support for the GPT-J model (see examples/gptj)
  • INT4 AWQ support for the Falcon models (see examples/falcon)
  • LoRA support (functional preview only - limited to the Python runtime, only QKV support and not optimized in terms of runtime performance) for the GPT model (see the Run LoRA with the Nemo checkpoint in the GPT example)
  • Multi-GPU support for encoder-decoder models (Python runtime - see examples/enc_dec)
  • New heuristic for launching the Multi-block Masked MHA kernel (similar to FlashDecoding - see decoderMaskedMultiheadAttentionLaunch.h)
  • Prompt-Tuning support for GPT and LLaMA models (see the Prompt-tuning Section in the GPT example)
  • Performance optimizations in various CUDA kernels
  • Possibility to exclude input tokens from the output (see excludeInputInOutput in GptManager)
  • Python binding for the C++ runtime (GptSession - see pybind)
  • Support for different micro batch sizes for context and generation phases with pipeline parallelism (see GptSession::Config::ctxMicroBatchSize and GptSession::Config::genMicroBatchSize in tensorrt_llm/runtime/gptSession.h)
  • Support for "remove input padding" for encoder-decoder models (see examples/enc_dec)
  • Support for context and generation logits (see mComputeContextLogits and mComputeGenerationLogits in tensorrt_llm/runtime/gptModelConfig.h)
  • Support for logProbs and cumLogProbs (see "output_log_probs" and "cum_log_probs" in GptManager)
  • Update to CUTLASS 3.x
• Bug fixes
  • Fix for ChatGLM2 does the chatglm2-6B example support codegeex2-6b's building and running? #93 and benchmark for chatglm2-6b failed #138
  • Fix tensor names error "RuntimeError: Tensor names (host_max_kv_cache_length) in engine are not the same as expected in the main branch" RuntimeError: Tensor names (host_max_kv_cache_length) in engine are not the same as expected on main branch #369
  • Fix weights split issue in BLOOM when world_size = 2 ("array split does not result in an equal division") world_size = 2 will raise error "array split does not result in an equal division" #374
  • Fix SmoothQuant multi-GPU failure with tensor parallelism is 2 smooth-quant with tp=2, and build llama 7b with tp=2, pp=2 failed #267
  • Fix a crash in GenerationSession if stream keyword argument is not None GenerationSession crashes if stream keyword argument is not None #202
  • Fix a typo when calling PyNVML API [BUG] code bug [BUG] code bug #410
  • Fix bugs related to the improper management of the end_id for various models [C++ and Python]
  • Fix memory leaks [C++ code and Python models]
  • Fix the std::alloc error when running the gptManagerBenchmark -- issue gptManagerBenchmark std::bad_alloc error gptManagerBenchmark std::bad_alloc error #66
  • Fix a bug in pipeline parallelism when beam-width > 1
  • Fix a bug with Llama GPTQ due to improper support of GQA
  • Fix issue set max_input_len, max_output_len=4096, But The actual input cannot reach this level #88
  • Fix an issue with the Huggingface Transformers version Huggingface Transformers version should be bumped #16
  • Fix link jump in windows readme.md Fix link jump in windows readme.md #30 - by @yuanlehome
  • Fix typo in batchScheduler.h Fix typo in batchScheduler.h #56 - by @eltociear
  • Fix typo Perceived #58 - by @RichardScottOZ
  • Fix Multi-block MMHA: Difference between max_batch_size in the engine builder and max_num_sequences in TrtGptModelOptionalParams? Difference between max_batch_size in the engine builder and max_num_sequences in TrtGptModelOptionalParams? #65
  • Fix the log message to be more accurate on KV cache Smaller available space for paged KV cache compared with vLLM #224
  • Fix Windows release wheel installation: Failed to install the release wheel for Windows using pip Failed to install the release wheel for Windows using pip #261
  • Fix missing torch dependencies: [BUG] The batch_manage.a choice error in --cpp-only when torch's cxx_abi version is different with gcc [BUG] The batch_manage.a choice error in --cpp-only when torch's cxx_abi version is different with gcc #151
  • Fix linking error during compiling google-test & benchmarks Linking error during compiling google-test & benchmarks #277
  • Fix logits dtype for Baichuan and ChatGLM: segmentation fault caused by the lack of bfloat16 segment falut cause bfloat16 decoder not create #335
  • Minor bug fixes

Currently, there are two key branches in the project:

• The rel branch contains what we'd call the stable branch for the release of TensorRT-LLM. It has been QA-ed and carefully tested.
• The main branch contains what we'd call the dev branch. It is more experimental.

We are updating the main branch regularly with new features, bug fixes and performance optimizations. The stable branch will be updated less frequently. The exact frequencies depend on your feedback.

Thanks,

The TensorRT-LLM Engineering Team