TensorRT-LLMs/docs/source/gpt_runtime.md

C++ GPT Runtime

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

The API of the C++ runtime is composed of the classes declared in cpp/include/tensorrt_llm/runtime and implemented in cpp/tensorrt_llm/runtime. An example of how to use the C++ runtime for a GPT-like auto-regressive model can be found in cpp/tests/runtime/gptSessionTest.cpp.

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/enc_dec folder.

The Session

The main component of the C++ runtime is the session. For GPT-like auto-regressive models, it is the GptSession class.

Creation

The constructors of that class allow users to specify the model and the environment to execute it. The model is described by an instance of the GptModelConfig class and a pointer to the TensorRT engine that must be executed to perform the inference. The environment is configured through the WorldConfig (that name comes from MPI and its "famous" MPI_COMM_WORLD default communicator). The constructor also accepts an optional object to log information, warnings and errors:

#include <tensorrt_llm/runtime/gptSession.h>

using namespace tensorrt_llm::runtime;

GptSession session(sessionConfig, // Configuration of the session,
                   modelConfig,   // Description of the model,
                   worldConfig,   // Description of the environment,
                   engineBuffer,  // The compiled TensorRT engine (const void*),
                   engineSize,    // The size in bytes of the TensorRT engine (size_t),
                   logger);       // The optional logger.

The above constructor accepts a const void* pointer to the engine and the associated size (in bytes) of that buffer. There exist other overloaded versions that take std::vector<uint8_t> or std::string arguments to encapsulate the engine.

Session Configuration

The session configuration is an instance of the GptSession::Config class. The constructor of this class requires three arguments:

• maxBatchSize, the maximum number of sequences in a batch,
• maxBeamWidth, the maximum width of the beams in beam-search,
• maxSequenceLength, the length of the longest input sequence,

Additionally, the class encapsulates the following optional parameters (they are declared as public member variables and can be accessed directly):

• decoderPerRequest, whether the session will use a different decoder per request. It must be set to true when running in-flight batching,
• cudaGraphMode, whether the session will use CUDA graphs for the engine execution in generation phase,
• kvCacheConfig encapsulates parameters to configure paged KV cache, when the paged KV cache is enabled in the engine:
  • maxTokens, the maximum number of tokens that will have to be stored in the paged KV cache,
  • freeGpuMemoryFraction, the fraction of free GPU memory that will be reserved for paged KV cache,
• ctxMicroBatchSize, the micro batch size to be used in context phase. Batches entered in GptSession::generation will be split into smaller micro batches of this size,
• genMicroBatchSize, the micro batch size to be used in generation phase, Batches entered in GptSession::generation will be split into smaller micro batches of this size.

Model Configuration

The model configuration is an instance of the GptModelConfig 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, is 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. See GPT Attention,
• 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 GPT Attention operator was compiled using the GPT 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). See GPT Attention,
• pagedKvCache, indicates if the K/V cache uses paging. See GPT Attention,
• 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. See GPT Attention,
• quantMode, controls the quantization method. See Numerical Precision.
• maxBatchSize, indicates the maximum batch size that the TensorRT engine was built for,
• maxInputLen, is the maximum size of the input sequences,
• maxSequenceLen, is the maximum total size (input+output) of the sequences.

World Configuration

Familiarity with MPI, is not required to utilize the TensorRT-LMM C++ runtime. There are two main things you need to know: (1) The C++ Runtime in TensorRT-LLM uses processes 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. (2) The ranks are grouped in communication groups. The TensorRT-LLM C++ Runtime calls that group the world.

The world configuration is an instance of the WorldConfig class. In this release, that class encapsulates the following parameters:

• tensorParallelism, is 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 PP (see below), in most cases, but requires higher bandwidth between the GPUs. It is the recommended setting in the presence of NVLINK between GPUs,
• pipelineParallelism, is the number of ranks that collaborate together to implement Pipeline Parallelism (PP). With PP, each GPU works on a subset of consecutive layers and 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 recommended in the absence of NVLINK between GPUs,
• rank, is the unique identifier of the rank (see below),
• 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 the NVLINK interconnect between GPUs of an A100 DGX node).

For a multi-GPU configuration (single or multi-node), each rank must create its own instance of GptSession using its own WorldConfig. A typical example is:

#include "tensorrt_llm/common/mpiUtils.h"

// Get the unique identifier for each rank.
auto const rank = COMM_SESSION.getRank();

// Create the TensorRT-LLM Runtime WorldConfig.
tensorrt_llm::runtime::WorldConfig worldConfig(tensorParallelism, pipelineParallelism, rank);

// Create the GPT session (as shown above).
tensorrt_llm::runtime::GptSession session(sessionConfig, modelConfig, worldConfig, ...);

For simplicity, TensorRT-LLM provides users with the following simplified API:

auto worldConfig = tensorrt_llm::runtime::WorldConfig::mpi();

Once compiled, that C++ code must be executed using the mpirun command installed on the system (talk to your system administrator if needed):

# Launch the program using two processes (worldSize == 2 and ranks == {0, 1}).
mpirun -n 2 ...

Generation

The GptSession::generate member function performs the generation loop. Given input tensors to read from, output tensors to populate, that member function will run the generation loop until it reaches the maximum number of tokens that can be produced or each sequence has reached completion (due to the production of "end-of-sequence" or a word in the list of "stop words"). The pseudo-code of that function looks like (member function names were changed to keep the presentation simple):

// Have all the sequences in the batch reached completion?
bool allFinished = false;

// Until all sequences are finished or the number of steps reaches the limit...
for (int step = 0; !allFinished && step < maxNewTokens; ++step) {

  // Trigger the computation of the logits...
  computeLogits(...);

  // Run the sampling to produce a token (for each active sequence) from the logits.
  allFinished = generateTokensFromLogits(...);

  // Callback to stream the output tokens while the generation loop continues.
  onTokenGenerated(...);
}

Inputs and Outputs

The generate member function takes an instance of the GenerationInput class and populates an instance of the GenerationOutput class.

Mandatory inputs

• endId, is the token ID that marks the end of the input sequence (aka EOS or end-of-sequence). It's 50,256 for the GPT2 model which has a vocabulary of 50,257 tokens, for example,
• padId, is the token ID that is used for padding (i.e. fills in the slots that are at an index greater-or-equal to the input length for padded sequences). It can be set to the same value as endId,
• ids, is the tensor of input IDs. That tensor must be allocated on the GPU. When the input tensor is padded, the shape of ids is [batchSize, maxInputLength], where batchSize and maxInputLength must respect the maximum sizes in sessionConfig passed to the GptSession constructor. When the input is packed, the shape of ids is [numTokens], where numTokens is the sum of the lengths of the different sequences in the batch,
• lengths, is the tensor of input sequence lengths. That tensor must be allocated on the GPU and contain batchSize values,
• packed, indicates if the ids tensor is packed or padded. In this release, that flag must match the value passed to the constructor through the instance of the ModelConfig class. In a future release, the session may be made more flexible and automatically pad or pack the input,

Optional inputs

• embeddingBiasOpt, is a tensor of floating-point values on the GPU that contains the bias to add to the logits during sampling (after the projection from hidden states to logits as the last step of the model). This tensor must have vocabSize elements (as defined in the modelConfig argument passed to the constructor),
• badWordsList, is a tensor of integers on the GPU that encodes the list of words that have to be banned from generated sequences. Its shape is [2, badWordsLength], as explained below, or [batchSize, 2, badWordsLength] when there is a different list for each sequence in the batch,
• stopWordsList, is a tensor of integers on the GPU that encodes the list of words that trigger the end of the generation for a sequence. Its shape is [2, stopWordsLength], as explained below, or [batchSize, 2, stopWordsLength] when there is a different list for each sequence in the batch,
• maxNewTokens, is the maximum number of tokens to generate.

The badWordsList and stopWordsList tensors have the same shape [2, length]. Let's consider an example with three words to describe the representation of those lists. The first word contains tokens [5, 7, 3], the second one contains [9, 2] and the third one is composed of tokens [6, 2, 4, 1]. In total, there are 9 tokens. That's the length. The shape of the tensor is [2, 9]. The first row of the tensor must contain the 9 token IDs and the second row must store the inclusive prefix-sum of the word lengths as shown on the following diagram:

   0           3       5              9
   |           |       |              |
   V           V       V              V
[  5,  7,  3,  9,  2,  6,  2,  4,  1]
[  3,  5,  9, -1, -1, -1, -1, -1, -1]

In case all the words are made of a single token, the inner-most dimension of the tensor must be increased by 1 (i.e. the length for 4 words, each made of a single token, must be 5 instead of 4 -- the shape is [2, 5]).

Mandatory outputs

• ids, is a tensor that contains the output token IDs. Its shape is [batchSize, beamWidth, maxSeqLength] where maxSeqLength is the sum of maxInputLength and maxNewTokens. After generation, it contains, for each sequence, a copy of the input tokens followed by the output tokens. When a sequence is shorter than maxSeqLength, padding tokens are added at the end of the sequence.

Note that the shape of that tensor is different in this version of TensorRT-LLM from its shape in previous versions where it was [maxSeqLength, batchSize, beamWidth].

Optional outputs

• logProbs, is a tensor of floating-point values on the GPU to store the log-prob of the generated tokens. Its shape is [maxNewTokens, batchSize, beamWidth]. Its shape will likely change in a future release to match the shape of the output ids tensor.

• contextLogits, is a tensor of values on the GPU (same datatype as the computation type) to store the logits for the context. Its shape is [batchSize, maxSequenceLength, vocabSizePadded]. If use remove_input_padding, its shape is [packedSize, vocabSizePadded]. This buffer will only be filled in if the TensorRT engine was built with the gather_context_logits or gather_all_token_logits parameter enabled.

  After inference is complete, you can get the context logits in GenerationOutput.contextLogits, these are variables on the GPU. For specific acquisition methods, please refer to the example of gptSessionBenchmark.cpp.

  It is important to point out that enabling the computation may have an impact on performance (the language modeling head (LM head) has to perform a matrix multiplication on all the context tokens instead of a just the last one).

• generationLogits, is a tensor of values on the GPU (same datatype as the computation type) to store the logits for the generation. Its shape is [batchSize, beamWidth, maxOutputLen, vocabSizePadded]. This buffer will only be filled in if the TensorRT engine was built with the gather_generation_logits or gather_all_token_logits parameter enabled.

  Generation logits can also be obtained through GenerationOutput.generationLogits after inference is completed.

• onTokenGenerated, is a callback function invoked in the generation loop to pass newly generated tokens to the caller while the loop continues to execute. An implementation of that callback must accept the output ids tensor, the generation step and a boolean flag that indicates if the generation is complete.

Sampling Parameters

The SamplingConfig class encapsulates parameters that control the generation of new tokens. 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 (i.e. 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 (i.e. size() == 1).

General

• temperature, a vector of floating-point numbers to control the modulation of logits when sampling new tokens. The default value is 1.0f,
• minLength, a vector of integers to set a lower-bound on the number of tokens generated. The default value is 0,
• repetitionPenalty, a vector of float-point numbers to penalize tokens based on how often they appear in the sequence. The default value is 0.f,
• presencePenalty, a vector of float-point numbers to penalize tokens already present in the sequence (irrespective of the number of appearances). The default value is 0.f,
• frequencyPenalty, a vector of float-point numbers to penalize tokens already present in the sequence (dependent on the number of appearances). The default value is 0.f,

The parameters repetitionPenalty, presencePenalty, and frequencyPenalty are not mutually exclusive.

Sampling

• randomSeed, a vector of 64-bit integers to control the random seed used by the random number generator in sampling. Its default value is 0,
• topK, a vector of integers to control the number of logits to sample from. Its default value is 0. Note that if different 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,
• topP, a vector of floating-point values to control the top-P probability to sample from. Its default value is 0.f,
• topPDecay, topPMin and topPResetIds, vectors to control the decay in the top-P algorithm. The top-P values are modulated by a decay that exponentially depends on the length of the sequence as explained in Factuality 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. Defaults are 1.f, 1.0e-6,f and -1,

If both topK and topP fields are set, the top-K method will be run for sequences with a topK value greater than 0.f. In that case, the topP value for that sequence also influences the result. If the topK values for some sequences are 0.f, the top-P method will be used for those remaining sequences. If both topK and topP are zero, greedy search is performed.

Beam-search

• beamWidth, is the width used for the beam search sampling algorithm. There is no explicit upper-bound on the beam width but increasing the beam width will likely increase the latency. Use 1 to disable beam-search,
• beamSearchDiversityRate, a floating-point value that controls the diversity in beam-search. Its default value is 0.f,
• lengthPenalty, a floating-point value that controls how to penalize the longer sequences in beam-search (the log-probability of a sequence will be penalized by a factor that depends on 1.f / (length ^ lengthPenalty)). The default is value 0.f. The parameter lengthPenalty may be renamed to beamSearchLengthPenalty in a future release,
• earlyStopping, a integer value that controls whether the generation process finishes once beamWidth sentences are generated (end up with end_token). Default value 1 means earlyStopping is enabled, value 0 means earlyStopping is disable, other values means the generation process is depended on length_penalty. 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 GptSession class encapsulates two main components. The TllmRuntime is in charge of the execution of the TensorRT engine. The GptDecoder does the generation of the tokens from the logits. The TllmRuntime class is an internal component and users are not expected to use that class directly. The GptDecoder can be used directly to implement very custom generation loop and for use cases that cannot be satisfied by the implementation in GptSession.

In-flight Batching Support

In this release, in-flight batching is supported using separate decoders per request. The biggest difference compared to using a single decoder is in how the token generation from logits is managed. A batch is split into batchSize individual requests and kernels are issued using separated CUDA streams. This behavior may be revisited in a future release to maintain the structure of the batch and improve efficiency.

Know Issues and Future Changes

• In the current release of TensorRT-LLM, the C++ and Python runtimes are two separate software components and the C++ runtime is being more actively developed (with features like in-flight batching). An objective, for a future release, could be to rebuild the Python runtime on top of the C++ one.