C++ GPT Runtime — TensorRT-LLM (original) (raw)

TensorRT-LLM includes a C++ component to execute TensorRT engines built with the Python API as described in the TensorRT-LLM Architecture section. That component is called the C++ runtime.

The API of the C++ runtime is composed of the classes declared incpp/include/tensorrt_llm/runtime and implemented in cpp/tensorrt_llm/runtime.

Even if the different components described in that document mention GPT in their name, they are not restricted to this specific model. Those classes can be used to implement auto-regressive models like BLOOM, GPT-J, GPT-NeoX or LLaMA, for example.

Complete support of encoder-decoder models, like T5, will be added to TensorRT-LLM in a future release. An experimental version, only in Python for now, can be found in the examples/models/core/enc_dec folder.

Overview#

Runtime models are described by an instance of theModelConfigclass and a pointer to the TensorRT engine that must be executed to perform the inference. The environment is configured through theWorldConfig(that name comes fromMPI and its “famous”MPI_COMM_WORLD default communicator). The SamplingConfigclass encapsulates parameters that control thegeneration of new tokens.

Model Configuration#

The model configuration is an instance of theModelConfig class. That class encapsulates the following parameters (they are declared as private member variables and exposed through getters and setters):

• vocabSize, the size of the vocabulary,
• numLayers, the number of layers in the model,
• numHeads, the number of heads in the attention block,
• numKvHeads, the number of heads for K and V in the attention component. When the number of K/V heads is the same as the number of (Q) heads, the model uses multi-head attention. When the number of K/V heads is 1, it uses multi-query attention. Otherwise, it uses group-query attention. Refer to Multi-Head, Multi-Query, and Group-Query Attention for more information,
• hiddenSize, the size of the hidden dimension,
• dataType, the datatype that was used to build the TensorRT engine and that must be used to run the model during inference,
• useGptAttentionPlugin, indicates if the Multi-Head, Multi-Query, and Group-Query Attention operator was compiled using theGPT Attention plugin,
• inputPacked, indicates that the input must be packed (or padded when set to false). For performance reasons, it is recommended to always use packed, even if its default is set to false (will be changed in a future release). Refer to Multi-Head, Multi-Query, and Group-Query Attention for more information,
• pagedKvCache, indicates if the K/V cache uses paging. Refer to Multi-Head, Multi-Query, and Group-Query Attention for more information,
• tokensPerBlock, is the number of tokens in each block of the K/V cache. It’s relevant when the paged K/V cache is enabled. By default, the value is 64. Refer to Multi-Head, Multi-Query, and Group-Query Attention for more information,
• quantMode, controls the quantization method. Refer to Numerical Precision for more information.
• maxBatchSize, indicates the maximum batch size that the TensorRT engine was built for,
• maxInputLen, the maximum size of the input sequences,
• maxSequenceLen, the maximum total size (input+output) of the sequences.

World Configuration#

Familiarity withMPI, is not required to utilize the TensorRT-LMM C++ runtime. There are two main things you need to know:

• The C++ Runtime in TensorRT-LLM usesprocesses to execute TensorRT engines on the different GPUs. Those GPUs can be located on a single node as well as on different nodes in a cluster. Each process is called a_rank_ in MPI.
• The ranks are grouped in communication groups. The TensorRT-LLM C++ Runtime calls that group the world.

The world configuration is an instance of theWorldConfigclass, which encapsulates the following parameters:

• tensorParallelism, the number of ranks that collaborate together to implement Tensor Parallelism (TP). With TP, each GPU performs computations for all the layers of the model. Some of those computations are distributed across the GPU. TP is more balanced than Pipeline Parallelism (PP), in most cases, but requires higher bandwidth between the GPUs. It is the recommended setting in the presence of NVLINK between GPUs,
• pipelineParallelism, the number of ranks that collaborate together to implement Pipeline Parallelism (PP). With PP, each GPU works on a subset of consecutive layers. Communications between the GPUs happen only at the boundaries of the subsets of layers. It is harder to guarantee the full utilization of the GPUs with PP but it requires less memory bandwidth. It is the recommended setting in the absence of NVLINK between GPUs,
• rank, the unique identifier of the rank,
• gpusPerNode, indicates the number of GPUs on each node. Having that information allows the C++ runtime to optimize communications between GPUs in a node (like taking advantage of theNVLINKinterconnect between GPUs of an A100DGXnode).

Sampling Parameters#

The SamplingConfigclass encapsulates parameters that control thegeneration of new tokens. A comparison of selecting decoding method is listed as the table below (X means it is not supported yet). Except for the beamWidth parameter, all the fields are optional and the runtime will use a default value if no values are provided by the user. For vector fields, the TensorRT-LLM runtime supports one value per sequence (that is, the vector contains batchSize values). If all the sequences use the same value for a given parameter, the vector can be limited to a single element (that is, size() == 1).

General

• The tokens of input prompt are included during adopting repetitionPenalty, presencePenalty, and frequencyPenalty onto logits.
• The parameters repetitionPenalty, presencePenalty, and frequencyPenalty are not mutually exclusive.

Sampling

• If setting topK = 0 and topP = 0.0f, greedy search is performed.
• If setting topK > 0 and topP = 0.0f, topK tokens of highest probabilities will become the candidates of sampling (named TopK sampling in TRT-LLM).
• If setting topK = 0 and topP > 0.0f, tokens will be sorted with probability descendly, then the tokens with highest probabilities which the accumulated probability larger than topP will become the candidates of sampling (named TopP sampling in TRT-LLM).
• If setting topK > 0 and topP > 0.0f, topK tokens of highest probabilities will be selected, then those selected tokens will be sorted with probability descendly and their probability will be normalized, then the tokens with highest normalized probabilities which the accumulated probability larger than topP will become the candidates of sampling (named TopKTopP sampling in TRT-LLM)
• If different topK values are provided for the different sequences in the batch, the performance of the implementation will depend on the largest value. For efficiency reasons, we recommend to batch requests with similar topK values together.
• topPDecay, topPMin and topPResetIds are explained inFactuality Enhanced Language Models for Open-Ended Text Generation.topPDecay is the decay, topPMin is the lower-bound and topPResetIds indicates where to reset the decay.
• minP is explained in Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs.

Beam-search

• Beam-search algorithm: beam search.
• Parameter diversity_penalty in HF is only used for diverse beam-search decoding (or named Group-Beam-Search), which is not supported by TRT-LLM yet.
• If setting earlyStopping = 1, decoding will stop once beamWidth finished sentences are generated.
• If setting earlyStopping = 0, decoding will keep going until no better sentences (with better score) can be generated.
• If setting earlyStopping to other values, decoding will stop only depending on lengthlengthPenalty.
• beamWidthArray is a list of beam width for each step. Using beamWidthArray = [20,40,80] as an example, beam width will be 20 for the first step, 40 for second step, 80 for the later all steps.
• The beamWidth parameter is a scalar value. It means that in this release of TensorRT-LLM, it is not possible to specify a different width for each input sequence. This limitation is likely to be removed in a future release.

Internal Components#

The TllmRuntime is in charge of the execution of the TensorRT engine. The TllmRuntime class is an internal component and you are not expected to use that class directly. The GptDecoder generates tokens from the logits. The GptDecoder can be used directly to implement a custom generation loop and for use cases that cannot be satisfied by the TRT-LLM implementation.