TensorRT-LLM Release 0.18.2

Key Features and Enhancements

• This update addresses known security issues. For the latest NVIDIA Vulnerability Disclosure Information visit https://www.nvidia.com/en-us/security/.

TensorRT-LLM Release 0.18.1

Key Features and Enhancements

• The 0.18.x series of releases builds upon the 0.17.0 release, focusing exclusively on dependency updates without incorporating features from the previous 0.18.0.dev pre-releases. These features will be included in future stable releases.

Infrastructure Changes

• The dependent transformers package version is updated to 4.48.3.

TensorRT-LLM Release 0.18.0

Hi,

We are very pleased to announce the 0.18.0 version of TensorRT-LLM. This update includes:

Key Features and Enhancements

• Features that were previously available in the 0.18.0.dev pre-releases are not included in this release.
• [BREAKING CHANGE] Windows platform support is deprecated as of v0.18.0. All Windows-related code and functionality will be completely removed in future releases.

Known Issues

• The PyTorch workflow on SBSA is incompatible with bare metal environments like Ubuntu 24.04. Please use the PyTorch NGC Container for optimal support on SBSA platforms.

Infrastructure Changes

• The base Docker image for TensorRT-LLM is updated to nvcr.io/nvidia/pytorch:25.03-py3.
• The base Docker image for TensorRT-LLM Backend is updated to nvcr.io/nvidia/tritonserver:25.03-py3.
• The dependent TensorRT version is updated to 10.9.
• The dependent CUDA version is updated to 12.8.1.
• The dependent NVIDIA ModelOpt version is updated to 0.25 for Linux platform.

TensorRT-LLM Release 0.17.0

Hi,

We are very pleased to announce the 0.17.0 version of TensorRT-LLM. This update includes:

Key Features and Enhancements

• Blackwell support
  • NOTE: pip installation is not supported for TRT-LLM 0.17 on Blackwell platforms only. Instead, it is recommended that the user build from source using NVIDIA NGC 25.01 PyTorch container.
  • Added support for B200.
  • Added support for GeForce RTX 50 series using Windows Subsystem for Linux (WSL) for limited models.
  • Added NVFP4 Gemm support for Llama and Mixtral models.
  • Added NVFP4 support for the LLM API and trtllm-bench command.
  • GB200 NVL is not fully supported.
  • Added benchmark script to measure perf benefits of KV cache host offload with expected runtime improvements from GH200.
• PyTorch workflow
  • The PyTorch workflow is an experimental feature in tensorrt_llm._torch. The following is a list of supported infrastructure, models, and features that can be used with the PyTorch workflow.
  • Added support for H100/H200/B200.
  • Added support for Llama models, Mixtral, QWen, Vila.
  • Added support for FP16/BF16/FP8/NVFP4 Gemm and fused Mixture-Of-Experts (MOE), FP16/BF16/FP8 KVCache.
  • Added custom context and decoding attention kernels support via PyTorch custom op.
  • Added support for chunked context (default off).
  • Added CudaGraph support for decoding only.
  • Added overlap scheduler support to overlap prepare inputs and model forward by decoding 1 extra token.
• Added FP8 context FMHA support for the W4A8 quantization workflow.
• Added ModelOpt quantized checkpoint support for the LLM API.
• Added FP8 support for the Llama-3.2 VLM model. Refer to the “MLLaMA” section in examples/multimodal/README.md.
• Added PDL support for userbuffer based AllReduce-Norm fusion kernel.
• Added runtime support for seamless lookahead decoding.
• Added token-aligned arbitrary output tensors support for the C++ executor API.

API Changes

• [BREAKING CHANGE] KV cache reuse is enabled automatically when paged_context_fmha is enabled.
• Added --concurrency support for the throughput subcommand of trtllm-bench.

Fixed Issues

• Fixed incorrect LoRA output dimension. Thanks for the contribution from @akhoroshev in #2484.
• Added NVIDIA H200 GPU into the cluster_key for auto parallelism feature. (#2552)
• Fixed a typo in the __post_init__ function of LLmArgs Class. Thanks for the contribution from @topenkoff in #2691.
• Fixed workspace size issue in the GPT attention plugin. Thanks for the contribution from @AIDC-AI.
• Fixed Deepseek-V2 model accuracy.

Infrastructure Changes

• The base Docker image for TensorRT-LLM is updated to nvcr.io/nvidia/pytorch:25.01-py3.
• The base Docker image for TensorRT-LLM Backend is updated to nvcr.io/nvidia/tritonserver:25.01-py3.
• The dependent TensorRT version is updated to 10.8.0.
• The dependent CUDA version is updated to 12.8.0.
• The dependent ModelOpt version is updated to 0.23 for Linux platform, while 0.17 is still used on Windows platform.

Known Issues

• Need --extra-index-url https://pypi.nvidia.com when running pip install tensorrt-llm due to new third-party dependencies.
• The PYPI SBSA wheel is incompatible with PyTorch 2.5.1 due to a break in the PyTorch ABI/API, as detailed in the related GitHub issue.

TensorRT-LLM Release 0.16.0

Hi,

We are very pleased to announce the 0.16.0 version of TensorRT-LLM. This update includes:

Key Features and Enhancements

• Added guided decoding support with XGrammar backend.
• Added quantization support for RecurrentGemma. Refer to examples/recurrentgemma/README.md.
• Added ulysses context parallel support. Refer to an example on building LLaMA 7B using 2-way tensor parallelism and 2-way context parallelism at examples/llama/README.md.
• Added W4A8 quantization support to BF16 models on Ada (SM89).
• Added PDL support for the FP8 GEMM plugins.
• Added a runtime max_num_tokens dynamic tuning feature, which can be enabled by setting --enable_max_num_tokens_tuning to gptManagerBenchmark.
• Added typical acceptance support for EAGLE.
• Supported chunked context and sliding window attention to be enabled together.
• Added head size 64 support for the XQA kernel.
• Added the following features to the LLM API:
  • Lookahead decoding.
  • DeepSeek V1 support.
  • Medusa support.
  • max_num_tokens and max_batch_size arguments to control the runtime parameters.
  • extended_runtime_perf_knob_config to enable various performance configurations.
• Added LogN scaling support for Qwen models.
• Added AutoAWQ checkpoints support for Qwen. Refer to the “INT4-AWQ” section in examples/qwen/README.md.
• Added AutoAWQ and AutoGPTQ Hugging Face checkpoints support for LLaMA. (#2458)
• Added allottedTimeMs to the C++ Request class to support per-request timeout.
• [BREAKING CHANGE] Removed NVIDIA V100 GPU support.

API Changes

• [BREAKING CHANGE] Removed enable_xqa argument from trtllm-build.
• [BREAKING CHANGE] Chunked context is enabled by default when KV cache and paged context FMHA is enabled on non-RNN based models.
• [BREAKING CHANGE] Enabled embedding sharing automatically when possible and remove the flag --use_embedding_sharing from convert checkpoints scripts.
• [BREAKING CHANGE] The if __name__ == "__main__" entry point is required for both single-GPU and multi-GPU cases when using the LLM API.
• [BREAKING CHANGE] Cancelled requests now return empty results.
• Added the enable_chunked_prefill flag to the LlmArgs of the LLM API.
• Integrated BERT and RoBERTa models to the trtllm-build command.

Model Updates

• Added Qwen2-VL support. Refer to the “Qwen2-VL” section of examples/multimodal/README.md.
• Added multimodal evaluation examples. Refer to examples/multimodal.
• Added Stable Diffusion XL support. Refer to examples/sdxl/README.md. Thanks for the contribution from @Zars19 in #1514.

Fixed Issues

• Fixed unnecessary batch logits post processor calls. (#2439)
• Fixed a typo in the error message. (#2473)
• Fixed the in-place clamp operation usage in smooth quant. Thanks for the contribution from @StarrickLiu in #2485.
• Fixed sampling_params to only be setup if end_id is None and tokenizer is not None in the LLM API. Thanks to the contribution from @mfuntowicz in #2573.

Infrastructure Changes

• Updated the base Docker image for TensorRT-LLM to nvcr.io/nvidia/pytorch:24.11-py3.
• Updated the base Docker image for TensorRT-LLM Backend to nvcr.io/nvidia/tritonserver:24.11-py3.
• Updated to TensorRT v10.7.
• Updated to CUDA v12.6.3.
• Added support for Python 3.10 and 3.12 to TensorRT-LLM Python wheels on PyPI.
• Updated to ModelOpt v0.21 for Linux platform, while v0.17 is still used on Windows platform.

Known Issues

• There is a known AllReduce performance issue on AMD-based CPU platforms on NCCL 2.23.4, which can be workarounded by export NCCL_P2P_LEVEL=SYS.

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

Thanks,
The TensorRT-LLM Engineering Team

TensorRT-LLM 0.15.0 Release

Hi,

We are very pleased to announce the 0.15.0 version of TensorRT-LLM. This update includes:

Key Features and Enhancements

• Added support for EAGLE. Refer to examples/eagle/README.md.
• Added functional support for GH200 systems.
• Added AutoQ (mixed precision) support.
• Added a trtllm-serve command to start a FastAPI based server.
• Added FP8 support for Nemotron NAS 51B. Refer to examples/nemotron_nas/README.md.
• Added INT8 support for GPTQ quantization.
• Added TensorRT native support for INT8 Smooth Quantization.
• Added quantization support for Exaone model. Refer to examples/exaone/README.md.
• Enabled Medusa for Qwen2 models. Refer to “Medusa with Qwen2” section in examples/medusa/README.md.
• Optimized pipeline parallelism with ReduceScatter and AllGather for Mixtral models.
• Added support for Qwen2ForSequenceClassification model architecture.
• Added Python plugin support to simplify plugin development efforts. Refer to examples/python_plugin/README.md.
• Added different rank dimensions support for LoRA modules when using the Hugging Face format. Thanks for the contribution from @AlessioNetti in #2366.
• Enabled embedding sharing by default. Refer to "Embedding Parallelism, Embedding Sharing, and Look-Up Plugin" section in docs/source/performance/perf-best-practices.md for information about the required conditions for embedding sharing.
• Added support for per-token per-channel FP8 (namely row-wise FP8) on Ada.
• Extended the maximum supported beam_width to 256.
• Added FP8 and INT8 SmoothQuant quantization support for the InternVL2-4B variant (LLM model only). Refer to examples/multimodal/README.md.
• Added support for prompt-lookup speculative decoding. Refer to examples/prompt_lookup/README.md.
• Integrated the QServe w4a8 per-group/per-channel quantization. Refer to “w4aINT8 quantization (QServe)” section in examples/llama/README.md.
• Added a C++ example for fast logits using the executor API. Refer to “executorExampleFastLogits” section in examples/cpp/executor/README.md.
• [BREAKING CHANGE] NVIDIA Volta GPU support is removed in this and future releases.
• Added the following enhancements to the LLM API:
  • [BREAKING CHANGE] Moved the runtime initialization from the first invocation of LLM.generate to LLM.__init__ for better generation performance without warmup.
  • Added n and best_of arguments to the SamplingParams class. These arguments enable returning multiple generations for a single request.
  • Added ignore_eos, detokenize, skip_special_tokens, spaces_between_special_tokens, and truncate_prompt_tokens arguments to the SamplingParams class. These arguments enable more control over the tokenizer behavior.
  • Added support for incremental detokenization to improve the detokenization performance for streaming generation.
  • Added the enable_prompt_adapter argument to the LLM class and the prompt_adapter_request argument for the LLM.generate method. These arguments enable prompt tuning.
• Added support for a gpt_variant argument to the examples/gpt/convert_checkpoint.py file. This enhancement enables checkpoint conversion with more GPT model variants. Thanks to the contribution from @tonylek in #2352.

API Changes

• [BREAKING CHANGE] Moved the flag builder_force_num_profiles in trtllm-build command to the BUILDER_FORCE_NUM_PROFILES environment variable.
• [BREAKING CHANGE] Modified defaults for BuildConfig class so that they are aligned with the trtllm-build command.
• [BREAKING CHANGE] Removed Python bindings of GptManager.
• [BREAKING CHANGE] auto is used as the default value for --dtype option in quantize and checkpoints conversion scripts.
• [BREAKING CHANGE] Deprecated gptManager API path in gptManagerBenchmark.
• [BREAKING CHANGE] Deprecated the beam_width and num_return_sequences arguments to the SamplingParams class in the LLM API. Use the n, best_of and use_beam_search arguments instead.
• Exposed --trust_remote_code argument to the OpenAI API server. (#2357)

Model Updates

• Added support for Llama 3.2 and llama 3.2-Vision model. Refer to examples/mllama/README.md for more details on the llama 3.2-Vision model.
• Added support for Deepseek-v2. Refer to examples/deepseek_v2/README.md.
• Added support for Cohere Command R models. Refer to examples/commandr/README.md.
• Added support for Falcon 2, refer to examples/falcon/README.md, thanks to the contribution from @puneeshkhanna in #1926.
• Added support for InternVL2. Refer to examples/multimodal/README.md.
• Added support for Qwen2-0.5B and Qwen2.5-1.5B model. (#2388)
• Added support for Minitron. Refer to examples/nemotron.
• Added a GPT Variant - Granite(20B and 34B). Refer to “GPT Variant - Granite” section in examples/gpt/README.md.
• Added support for LLaVA-OneVision model. Refer to “LLaVA, LLaVa-NeXT, LLaVA-OneVision and VILA” section in examples/multimodal/README.md.

Fixed Issues

• Fixed a slice error in forward function. (#1480)
• Fixed an issue that appears when building BERT. (#2373)
• Fixed an issue that model is not loaded when building BERT. (2379)
• Fixed the broken executor examples. (#2294)
• Fixed the issue that the kernel moeTopK() cannot find the correct expert when the number of experts is not a power of two. Thanks @dongjiyingdjy for reporting this bug.
• Fixed an assertion failure on crossKvCacheFraction. (#2419)
• Fixed an issue when using smoothquant to quantize Qwen2 model. (#2370)
• Fixed a PDL typo in docs/source/performance/perf-benchmarking.md, thanks @MARD1NO for pointing it out in #2425.

Infrastructure Changes

• The base Docker image for TensorRT-LLM is updated to nvcr.io/nvidia/pytorch:24.10-py3.
• The base Docker image for TensorRT-LLM Backend is updated to nvcr.io/nvidia/tritonserver:24.10-py3.
• The dependent TensorRT version is updated to 10.6.
• The dependent CUDA version is updated to 12.6.2.
• The dependent PyTorch version is updated to 2.5.1.
• The dependent ModelOpt version is updated to 0.19 for Linux platform, while 0.17 is still used on Windows platform.

Documentation

• Added a copy button for code snippets in the documentation. (#2288)

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

Thanks,
The TensorRT-LLM Engineering Team

TensorRT-LLM 0.14.0 Release

Hi,

We are very pleased to announce the 0.14.0 version of TensorRT-LLM. This update includes:

Key Features and Enhancements

• Enhanced the LLM class in the LLM API.
  • Added support for calibration with offline dataset.
  • Added support for Mamba2.
  • Added support for finish_reason and stop_reason.
• Added FP8 support for CodeLlama.
• Added __repr__ methods for class Module, thanks to the contribution from @1ytic in #2191.
• Added BFloat16 support for fused gated MLP.
• Updated ReDrafter beam search logic to match Apple ReDrafter v1.1.
• Improved customAllReduce performance.
• Draft model now can copy logits directly over MPI to the target model's process in orchestrator mode. This fast logits copy reduces the delay between draft token generation and the beginning of target model inference.
• NVIDIA Volta GPU support is deprecated and will be removed in a future release.

API Changes

• [BREAKING CHANGE] The default max_batch_size of the trtllm-build command is set to 2048.
• [BREAKING CHANGE] Remove builder_opt from the BuildConfig class and the trtllm-build command.
• Add logits post-processor support to the ModelRunnerCpp class.
• Added isParticipant method to the C++ Executor API to check if the current process is a participant in the executor instance.

Model Updates

• Added support for NemotronNas, see examples/nemotron_nas/README.md.
• Added support for Deepseek-v1, see examples/deepseek_v1/README.md.
• Added support for Phi-3.5 models, see examples/phi/README.md.

Fixed Issues

• Fixed a typo in tensorrt_llm/models/model_weights_loader.py, thanks to the contribution from @wangkuiyi in #2152.
• Fixed duplicated import module in tensorrt_llm/runtime/generation.py, thanks to the contribution from @lkm2835 in #2182.
• Enabled share_embedding for the models that have no lm_head in legacy checkpoint conversion path, thanks to the contribution from @lkm2835 in #2232.
• Fixed kv_cache_type issue in the Python benchmark, thanks to the contribution from @qingquansong in #2219.
• Fixed an issue with SmoothQuant calibration with custom datasets. Thanks to the contribution by @Bhuvanesh09 in #2243.
• Fixed an issue surrounding trtllm-build --fast-build with fake or random weights. Thanks to @ZJLi2013 for flagging it in #2135.
• Fixed missing use_fused_mlp when constructing BuildConfig from dict, thanks for the fix from @ethnzhng in #2081.
• Fixed lookahead batch layout for numNewTokensCumSum. (#2263)

Infrastructure Changes

• The dependent ModelOpt version is updated to v0.17.

Documentation

• @Sherlock113 added a tech blog to the latest news in #2169, thanks for the contribution.

Known Issues

• Replit Code is not supported with the transformers 4.45+

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

Thanks,
The TensorRT-LLM Engineering Team

TensorRT-LLM 0.13.0 Release

Hi,

We are very pleased to announce the 0.13.0 version of TensorRT-LLM. This update includes:

Key Features and Enhancements

• Supported lookahead decoding (experimental), see docs/source/speculative_decoding.md.
• Added some enhancements to the ModelWeightsLoader (a unified checkpoint converter, see docs/source/architecture/model-weights-loader.md).
  • Supported Qwen models.
  • Supported auto-padding for indivisible TP shape in INT4-wo/INT8-wo/INT4-GPTQ.
  • Improved performance on *.bin and *.pth.
• Supported OpenAI Whisper in C++ runtime.
• Added some enhancements to the LLM class.
  • Supported LoRA.
  • Supported engine building using dummy weights.
  • Supported trust_remote_code for customized models and tokenizers downloaded from Hugging Face Hub.
• Supported beam search for streaming mode.
• Supported tensor parallelism for Mamba2.
• Supported returning generation logits for streaming mode.
• Added curand and bfloat16 support for ReDrafter.
• Added sparse mixer normalization mode for MoE models.
• Added support for QKV scaling in FP8 FMHA.
• Supported FP8 for MoE LoRA.
• Supported KV cache reuse for P-Tuning and LoRA.
• Supported in-flight batching for CogVLM models.
• Supported LoRA for the ModelRunnerCpp class.
• Supported head_size=48 cases for FMHA kernels.
• Added FP8 examples for DiT models, see examples/dit/README.md.
• Supported decoder with encoder input features for the C++ executor API.

API Changes

• [BREAKING CHANGE] Set use_fused_mlp to True by default.
• [BREAKING CHANGE] Enabled multi_block_mode by default.
• [BREAKING CHANGE] Enabled strongly_typed by default in builder API.
• [BREAKING CHANGE] Renamed maxNewTokens, randomSeed and minLength to maxTokens, seed and minTokens following OpenAI style.
• The LLM class
  • [BREAKING CHANGE] Updated LLM.generate arguments to include PromptInputs and tqdm.
• The C++ executor API
  • [BREAKING CHANGE] Added LogitsPostProcessorConfig.
  • Added FinishReason to Result.

Model Updates

• Supported Gemma 2, see "Run Gemma 2" section in examples/gemma/README.md.

Fixed Issues

• Fixed an accuracy issue when enabling remove padding issue for cross attention. (#1999)
• Fixed the failure in converting qwen2-0.5b-instruct when using smoothquant. (#2087)
• Matched the exclude_modules pattern in convert_utils.py to the changes in quantize.py. (#2113)
• Fixed build engine error when FORCE_NCCL_ALL_REDUCE_STRATEGY is set.
• Fixed unexpected truncation in the quant mode of gpt_attention.
• Fixed the hang caused by race condition when canceling requests.
• Fixed the default factory for LoraConfig. (#1323)

Infrastructure Changes

• Base Docker image for TensorRT-LLM is updated to nvcr.io/nvidia/pytorch:24.07-py3.
• Base Docker image for TensorRT-LLM Backend is updated to nvcr.io/nvidia/tritonserver:24.07-py3.
• The dependent TensorRT version is updated to 10.4.0.
• The dependent CUDA version is updated to 12.5.1.
• The dependent PyTorch version is updated to 2.4.0.
• The dependent ModelOpt version is updated to v0.15.

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

Thanks,
The TensorRT-LLM Engineering Team

TensorRT-LLM 0.12.0 Release

Hi,

We are very pleased to announce the 0.12.0 version of TensorRT-LLM. This update includes:

Key Features and Enhancements

• Supported LoRA for MoE models.
• The ModelWeightsLoader is enabled for LLaMA family models (experimental), see docs/source/architecture/model-weights-loader.md.
• Supported FP8 FMHA for NVIDIA Ada Lovelace Architecture.
• Supported GPT-J, Phi, Phi-3, Qwen, GPT, GLM, Baichuan, Falcon and Gemma models for the LLM class.
• Supported FP8 OOTB MoE.
• Supported Starcoder2 SmoothQuant. (#1886)
• Supported ReDrafter Speculative Decoding, see “ReDrafter” section in docs/source/speculative_decoding.md.
• Supported padding removal for BERT, thanks to the contribution from @Altair-Alpha in #1834.
• Added in-flight batching support for GLM 10B model.
• Supported gelu_pytorch_tanh activation function, thanks to the contribution from @ttim in #1897.
• Added chunk_length parameter to Whisper, thanks to the contribution from @MahmoudAshraf97 in #1909.
• Added concurrency argument for gptManagerBenchmark.
• Executor API supports requests with different beam widths, see docs/source/executor.md#sending-requests-with-different-beam-widths.
• Added the flag --fast_build to trtllm-build command (experimental).

API Changes

• [BREAKING CHANGE] max_output_len is removed from trtllm-build command, if you want to limit sequence length on engine build stage, specify max_seq_len.
• [BREAKING CHANGE] The use_custom_all_reduce argument is removed from trtllm-build.
• [BREAKING CHANGE] The multi_block_mode argument is moved from build stage (trtllm-build and builder API) to the runtime.
• [BREAKING CHANGE] The build time argument context_fmha_fp32_acc is moved to runtime for decoder models.
• [BREAKING CHANGE] The arguments tp_size, pp_size and cp_size is removed from trtllm-build command.
• The C++ batch manager API is deprecated in favor of the C++ executor API, and it will be removed in a future release of TensorRT-LLM.
• Added a version API to the C++ library, a cpp/include/tensorrt_llm/executor/version.h file is going to be generated.

Model Updates

• Supported LLaMA 3.1 model.
• Supported Mamba-2 model.
• Supported EXAONE model, see examples/exaone/README.md.
• Supported Qwen 2 model.
• Supported GLM4 models, see examples/chatglm/README.md.
• Added LLaVa-1.6 (LLaVa-NeXT) multimodal support, see “LLaVA, LLaVa-NeXT and VILA” section in examples/multimodal/README.md.

Fixed Issues

• Fixed wrong pad token for the CodeQwen models. (#1953)
• Fixed typo in cluster_infos defined in tensorrt_llm/auto_parallel/cluster_info.py, thanks to the contribution from @saeyoonoh in #1987.
• Removed duplicated flags in the command at docs/source/reference/troubleshooting.md, thanks for the contribution from @hattizai in #1937.
• Fixed segmentation fault in TopP sampling layer, thanks to the contribution from @akhoroshev in #2039. (#2040)
• Fixed the failure when converting the checkpoint for Mistral Nemo model. (#1985)
• Propagated exclude_modules to weight-only quantization, thanks to the contribution from @fjosw in #2056.
• Fixed wrong links in README, thanks to the contribution from @Tayef-Shah in #2028.
• Fixed some typos in the documentation, thanks to the contribution from @lfz941 in #1939.
• Fixed the engine build failure when deduced max_seq_len is not an integer. (#2018)

Infrastructure Changes

• Base Docker image for TensorRT-LLM is updated to nvcr.io/nvidia/pytorch:24.07-py3.
• Base Docker image for TensorRT-LLM Backend is updated to nvcr.io/nvidia/tritonserver:24.07-py3.
• The dependent TensorRT version is updated to 10.3.0.
• The dependent CUDA version is updated to 12.5.1.
• The dependent PyTorch version is updated to 2.4.0.
• The dependent ModelOpt version is updated to v0.15.0.

Known Issues

• On Windows, installation of TensorRT-LLM may succeed, but you might hit OSError: exception: access violation reading 0x0000000000000000 when importing the library in Python. See Installing on Windows for workarounds.

Currently, there are two key branches in the project:

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

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

Thanks,
The TensorRT-LLM Engineering Team

TensorRT-LLM 0.11.0 Release

Hi,

We are very pleased to announce the 0.11.0 version of TensorRT-LLM. This update includes:

Key Features and Enhancements

• Supported very long context for LLaMA (see “Long context evaluation” section in examples/llama/README.md).
• Low latency optimization
  • Added a reduce-norm feature which aims to fuse the ResidualAdd and LayerNorm kernels after AllReduce into a single kernel, which is recommended to be enabled when the batch size is small and the generation phase time is dominant.
  • Added FP8 support to the GEMM plugin, which benefits the cases when batch size is smaller than 4.
  • Added a fused GEMM-SwiGLU plugin for FP8 on SM90.
• LoRA enhancements
  • Supported running FP8 LLaMA with FP16 LoRA checkpoints.
  • Added support for quantized base model and FP16/BF16 LoRA.
    • SQ OOTB (- INT8 A/W) + FP16/BF16/FP32 LoRA​
    • INT8/ INT4 Weight-Only (INT8 /W) + FP16/BF16/FP32 LoRA​
    • Weight-Only Group-wise + FP16/BF16/FP32 LoRA
  • Added LoRA support to Qwen2, see “Run models with LoRA” section in examples/qwen/README.md.
  • Added support for Phi-3-mini/small FP8 base + FP16/BF16 LoRA, see “Run Phi-3 with LoRA” section in examples/phi/README.md.
  • Added support for starcoder-v2 FP8 base + FP16/BF16 LoRA, see “Run StarCoder2 with LoRA” section in examples/gpt/README.md.
• Encoder-decoder models C++ runtime enhancements
  • Supported paged KV cache and inflight batching. (#800)
  • Supported tensor parallelism.
• Supported INT8 quantization with embedding layer excluded.
• Updated default model for Whisper to distil-whisper/distil-large-v3, thanks to the contribution from @IbrahimAmin1 in #1337.
• Supported HuggingFace model automatically download for the Python high level API.
• Supported explicit draft tokens for in-flight batching.
• Supported local custom calibration datasets, thanks to the contribution from @DreamGenX in #1762.
• Added batched logits post processor.
• Added Hopper qgmma kernel to XQA JIT codepath.
• Supported tensor parallelism and expert parallelism enabled together for MoE.
• Supported the pipeline parallelism cases when the number of layers cannot be divided by PP size.
• Added numQueuedRequests to the iteration stats log of the executor API.
• Added iterLatencyMilliSec to the iteration stats log of the executor API.
• Add HuggingFace model zoo from the community, thanks to the contribution from @matichon-vultureprime in #1674.

API Changes

• [BREAKING CHANGE] trtllm-build command
  • Migrated Whisper to unified workflow (trtllm-build command), see documents: examples/whisper/README.md.
  • max_batch_size in trtllm-build command is switched to 256 by default.
  • max_num_tokens in trtllm-build command is switched to 8192 by default.
  • Deprecated max_output_len and added max_seq_len.
  • Removed unnecessary --weight_only_precision argument from trtllm-build command.
  • Removed attention_qk_half_accumulation argument from trtllm-build command.
  • Removed use_context_fmha_for_generation argument from trtllm-build command.
  • Removed strongly_typed argument from trtllm-build command.
  • The default value of max_seq_len reads from the HuggingFace mode config now.
• C++ runtime
  • [BREAKING CHANGE] Renamed free_gpu_memory_fraction in ModelRunnerCpp to kv_cache_free_gpu_memory_fraction.
  • [BREAKING CHANGE] Refactored GptManager API
    • Moved maxBeamWidth into TrtGptModelOptionalParams.
    • Moved schedulerConfig into TrtGptModelOptionalParams.
  • Added some more options to ModelRunnerCpp, including max_tokens_in_paged_kv_cache, kv_cache_enable_block_reuse and enable_chunked_context.
• [BREAKING CHANGE] Python high-level API
  • Removed the ModelConfig class, and all the options are moved to LLM class.
  • Refactored the LLM class, please refer to examples/high-level-api/README.md
    • Moved the most commonly used options in the explicit arg-list, and hidden the expert options in the kwargs.
    • Exposed model to accept either HuggingFace model name or local HuggingFace model/TensorRT-LLM checkpoint/TensorRT-LLM engine.
    • Support downloading model from HuggingFace model hub, currently only Llama variants are supported.
    • Support build cache to reuse the built TensorRT-LLM engines by setting environment variable TLLM_HLAPI_BUILD_CACHE=1 or passing enable_build_cache=True to LLM class.
    • Exposed low-level options including BuildConfig, SchedulerConfig and so on in the kwargs, ideally you should be able to configure details about the build and runtime phase.
  • Refactored LLM.generate() and LLM.generate_async() API.
    • Removed SamplingConfig.
    • Added SamplingParams with more extensive parameters, see tensorrt_llm/hlapi/utils.py.
      • The new SamplingParams contains and manages fields from Python bindings of SamplingConfig, OutputConfig, and so on.
    • Refactored LLM.generate() output as RequestOutput, see tensorrt_llm/hlapi/llm.py.
  • Updated the apps examples, specially by rewriting both chat.py and fastapi_server.py using the LLM APIs, please refer to the examples/apps/README.md for details.
    • Updated the chat.py to support multi-turn conversation, allowing users to chat with a model in the terminal.
    • Fixed the fastapi_server.py and eliminate the need for mpirun in multi-GPU scenarios.
• [BREAKING CHANGE] Speculative decoding configurations unification
  • Introduction of SpeculativeDecodingMode.h to choose between different speculative decoding techniques.
  • Introduction of SpeculativeDecodingModule.h base class for speculative decoding techniques.
  • Removed decodingMode.h.
• gptManagerBenchmark
  • [BREAKING CHANGE] api in gptManagerBenchmark command is executor by default now.
  • Added a runtime max_batch_size.
  • Added a runtime max_num_tokens.
• [BREAKING CHANGE] Added a bias argument to the LayerNorm module, and supports non-bias layer normalization.
• [BREAKING CHANGE] Removed GptSession Python bindings.

Model Updates

• Supported Jais, see examples/jais/README.md.
• Supported DiT, see examples/dit/README.md.
• Supported VILA 1.5.
• Supported Video NeVA, see Video NeVAsection in examples/multimodal/README.md.
• Supported Grok-1, see examples/grok/README.md.
• Supported Qwen1.5-110B with FP8 PTQ.
• Supported Phi-3 small model with block sparse attention.
• Supported InternLM2 7B/20B, thanks to the contribution from @RunningLeon in #1392.
• Supported Phi-3-medium models, see examples/phi/README.md.
• Supported Qwen1.5 MoE A2.7B.
• Supported phi 3 vision multimodal.

Fixed Issues

• Fixed brokens outputs for the cases when batch size is larger than 1. (#1539)
• Fixed top_k type in executor.py, thanks to the contribution from @vonjackustc in #1329.
• Fixed stop and bad word list pointer offset in Python runtime, thanks to the contribution from @fjosw in #1486.
• Fixed some typos for Whisper model, thanks to the contribution from @Pzzzzz5142 in #1328.
• Fixed export failure with CUDA driver < 526 and pynvml >= 11.5.0, thanks to the contribution from @CoderHam in #1537.
• Fixed an issue in NMT weight conversion, thanks to the contribution from @Pzzzzz5142 in #1660.
• Fixed LLaMA Smooth Quant conversion, thanks to the contribution from @lopuhin in #1650.
• Fixed qkv_bias shape issue for Qwen1.5-32B (#1589), thanks to the contribution from @Tlntin in #1637.
• Fixed the error of Ada traits for fpA_intB, thanks to the contribution from @JamesTheZ in #1583.
• Update examples/qwenvl/requirements.txt, thanks to the contribution from @ngoanpv in #1248.
• Fixed rsLoRA scaling in lora_manager, thanks to the contribution from @TheCodeWrangler in #1669.
• Fixed Qwen1.5 checkpoint convert failure #1675.
• Fixed Medusa safetensors and AWQ conversion, thanks to the contribution from @Tushar-ml in #1535.
• Fixed convert_hf_mpt_legacy call failure when the function is called in other than global scope, thanks to the contribution from @bloodeagle40234 in #1534.
• Fixed use_fp8_context_fmha broken outputs (#1539).
• Fixed pre-norm weight conversion for NMT models, thanks to the contribution from @Pzzzzz5142 in #1723.
• Fixed random seed initialization issue, thanks to the contribution from @pathorn in #1742.
• Fixed stop words and bad words in python bindings. (#1642)
• Fixed the issue that when converting checkpoint for Mistral 7B v0.3, thanks to the contribution from @Ace-RR: #1732.
• Fixed broken inflight batching for fp8 Llama and Mixtral, thanks to the contribution from @bprus: #1738
• Fixed the failure when quantize.py is export data to config.json, thanks to the contribution from @janpetrov: #1676
• Raise error when autopp detects unsupported quant plugin #1626.
• Fixed the issue that shared_embedding_table is not being set when loading Gemma #1799, thanks to the contribution from @mfuntowicz.
• Fixed stop and bad words list contiguous for ModelRunner #1815, thanks to the contribution from @Marks101.
• Fixed missing comment for FAST_BUILD, thanks to the support from @lkm2835 in #1851.
• Fixed the issues that Top-P sampling occasionally produces invalid tokens. #1590
• Fixed #1424.
• Fixed #1529.
• Fixed benchmarks/cpp/README.md for #1562 and #1552.
• Fixed dead link, thanks to the help from @DefTruth, @buvnswrn and @sunjiabin17 in: triton-inference-server/tensorrtllm_backend#478, triton-inference-server/tensorrtllm_backend#482 and triton-inference-server/tensorrtllm_backend#449.

Infrastructure Changes

• Base Docker image for TensorRT-LLM is updated to nvcr.io/nvidia/pytorch:24.05-py3.
• Base Docker image for TensorRT-LLM backend is updated to `nvcr.io/nvidia/...