The Batch Manager in TensorRT-LLM

TensorRT-LLM relies on a component, called the Batch Manager, to support in-flight batching of requests (also known in the community as continuous batching or iteration-level batching). That technique that aims at reducing wait times in queues, eliminating the need for padding requests and allowing for higher GPU utilization.

In more details, this feature allows for the inclusion of newly arrived requests and the return of newly completed requests at each iteration of the token generation loop. In-flight batching is accessed via a TensorRT-LLM component called the Batch Manager. That batch manager exposes hooks for the user to register function pointers to define how TensorRT-LLM reads in new requests and how it returns completed requests to the user.

The Batch Manager API

A software component (called the client in the text that follows) can interact with the batch manager using two mandatory, and several optional callbacks. Their signatures are defined in the callbacks.h file.

These callbacks are invoked in the generation loop at regular intervals and serve a variety of functions described below.

Get and Send Callbacks

The entry point to pass new requests to the batch manager is a callback of type GetInferenceRequestsCallback. An implementation of that callback must return a list of requests (std::list<std::shared_ptr<InferenceRequest>) to be processed by the batch manager. It takes a parameter indicating the maximum number of requests that can be accepted (a negative value indicates that an unbounded number of requests can be accepted). The complete signature of that callback is:

using GetInferenceRequestsCallback = std::function<std::list<std::shared_ptr<InferenceRequest>>(int32_t)>;

For each new request, the client must provide the batch manager with its input tensors and a 64-bit unsigned number (uint64_t) that will uniquely identify the request. That identifier is called the request ID in the text that follows (and in the code of the batch manager). The input tensors are collected in a map (std::map<std::string, Tensor>) that associates input names to tensor. Refer to InferenceRequest.h for more information.

Responses are delivered to the client through a callback of type SendResponseCallback. A conforming callback must accept the 64-bit request ID that uniquely identifies the request, the list of output tensors, a boolean (identifying the last response for the request when set to true) and a potentially non-empty error message. A non-empty error message indicates that an error has been encountered. In that case, the boolean indicating that this is the last response will be set to true, and the callback must properly handle the error. Its signature is:

using SendResponseCallback = std::function<void(uint64_t, std::list<std::shared_ptr<Tensor>> const&, bool, const std::string&)>;

The batch manager will reject any request sent using the GetInferenceRequestsCallback callback if the request ID passed by the client corresponds to the request ID of a request that is being processed by the batch manager. A request ID can be reused after it appears in a call to the SendResponseCallback callback marked as final (third argument set to true).

Request Interruption

The batch manager allows users to stop the execution of requests currently in-flight. The set of request IDs to be stopped can be passed to the batch manager through the callback:

using PollStopSignalCallback = std::function<std::unordered_set<uint64_t>()>;

When an active request appears in the set of requests to be interrupted, the batch manager will ensure that it is properly stopped.

Statistics

The batch manager can report execution statistics when provided with the following callback:

using ReturnBatchManagerStatsCallback = std::function<void(const std::string&)>;

The statistics are packaged as a JSON string. That string contains the following fields:

  • Timestamp, the timestamp of the request (obtained using std::put_time(&tm, "%m-%d-%Y %H:%M:%S")),

  • Iteration Counter, a global step counter value that increases monotonically over time

  • Active Request Count, the number of active requests in batch manager

  • Max Request Count, the max number of requests batch manager can support at a time

When using paged KV cache, following statistics are reported:

  • Max KV cache blocks, the maximum number of KV cache blocks per GPU

  • Free KV cache blocks, number of free KV cache blocks per GPU

  • Used KV cache blocks, number of used KV cache blocks per GPU

  • Tokens per KV cache block, number of tokens per KV cache block

  • Scheduled Requests, number of requests scheduled this iteration

When using in-flight batching, the following additional statistics are reported per step/iteration:

  • Scheduled Requests, number of total requests scheduled

  • Context Requests, number of requests in Context phase

  • Generation Requests, number of requests in Generation phase

  • Total Context Tokens, total number of tokens across requests in context phase

  • MicroBatch ID, micro batch ID

When using V1 batching, the following additional statistics are reported per V1 iteration:

  • Scheduled Requests, number of total requests scheduled

  • Context Requests, number of requests in Context phase

  • Total Generation Tokens, Total number of tokens generated

  • Total Context Tokens, total number of tokens across requests in context phase

  • Empty Generation Slots, total number of padded Slots during generation phase

Logits Post-Processor (optional)

Users can alter the logits produced the network, with a callback attached to an InferenceRequest:

  using LogitsPostProcessor = std::function<TensorPtr(RequestIdType, TensorPtr&, BeamTokens const&, TStream)>;

The first argument is the request id, second is the logits tensor, third are the tokens produced by the request so far, and last one is the operation stream used by the logits tensor.

Users must use the stream to access the logits tensor. For example, performing a addition with a bias tensor should be enqueued on that stream. Alternatively, users may call stream->synchronize(), however, that will slow down the entire execution pipeline.

Note: this feature isn’t supported with the V1 batching scheme for the moment.

Other mandatory GptManager parameters

  • trtEnginePath, path to the directory containing the TRT-LLM engine that GptManager wraps

  • modelType, batching scheme - V1, InflightBatching or InflightFusedBatching.

    • V1 refers to the traditional batching scheme with a batch of requests running in lockstep until the full generation for all of them is complete. Requests in a batch are all padded up to the maximum input and output sequence length of any member of the batch.

    • InflightBatching refers to a scheme where newly arrived requests are dynamically incorporated into the batch under execution, and requests are returned as soon as the end condition is met without any padding.

    • InflightFusedBatching is an improvement on InflightBatching, leveraging additional operation fusion opportunities and is expected to be strictly superior to it.

  • maxBeamWidth, the maximum beam width GptManager will allow for any request.

  • schedulerPolicy, policy used to select the subset available requests in each iteration of the InflightBatching generation loop.

    • MAX_UTILIZATION packs as many requests as the underlying TRT engine can support in any iteration of the InflightBatching generation loop. While this is expected to maximize GPU throughput, it might require that some requests be paused and restarted depending on peak KV cache memory availability.

    • GUARANTEED_NO_EVICT uses KV cache more conservatively guaranteeing that a request, once started, will run to completion without eviction.

Optional GptManager parameters

  • TrtGptModelOptionalParams class encapsulates the following fields:

    • kvCacheConfig

      • maxTokens (default: unspecified) refers to the maximum number of tokens reserved for KV cache across all requests. If specified, the final allocated KV cache considers this parameter as well as freeGpuMemoryFraction below.

      • maxAttentionWindow (default: unspecified) refers to the maximum number of tokens attended to in the model when using features like sliding window attention or StreamingLLM. If unspecified, each generated tokens attends to all previous tokens like traditional MHA or MQA.

      • freeGpuMemoryFraction (default: 0.9) a number between 0 and 1 to indicate the maximum fraction of GPU memory (after loading the model) that may be used for KV cache. If maxTokens is specified, allocated KV cache is the minimum of maxTokens and the value inferred from freeGpuMemoryFraction.

      • enableBlockReuse (default: false) allow reuse of previously computed KV cache blocks across requests. This is expected to optimize memory use and computation.

    • enableTrtOverlap (default: false) when true, GptManager partitions available requests into 2 ‘microbatches’ that can be run concurrently to hide exposed CPU runtime. However, it may not give performance benefits when the size of the model is not big enough to overlap the host overhead, or when the number of requests is too small.

    • enableChunkedContext (default: false) Whether to enable context chunking. Context chunking increases the possibility of batching the context and generation phases, which in turn improves performance. When set to false, it indicates that the context chunk is disabled.

    • peftCacheManagerConfig (currently only supports LoRA, and requires --use_lora_plugin during engine build)

      • numHostModuleLayer (default: 0) number of adapter_size 1 single module single layer LoRA weight rows the host cache can hold. Overrides hostCacheSize if non-zero.

      • numDeviceModuleLayer (default: 0) number of adapter_size 1 single module single layer LoRA weight rows the device cache can hold. Overrides deviceCachePercent if non-zero.

      • optimalAdapterSize (default: 8) Used to size cache pages. Typically optimally sized adapters will fix exactly into 1 cache page.

      • maxAdapterSize (default: 64) Used to set the minimum size of a cache page. Pages must be at least large enough to fit a single module, single later adapter_size maxAdapterSize row of weights.

      • numPutWorkers (default: 1) Number of CPU workers used to put weights into host cache.

      • numEnsureWorkers (default: 1) Number of CPU workers used to ensure all weights needed for the next forward pass are in the GPU cache.

      • numCopyStreams (default: 1) Number of CUDA streams used for H2D copies of cache pages

      • maxPagesPerBlockHost (default: 24) Number of cache pages per host memory allocation

      • maxPagesPerBlockDevice (default: 24) Number of cache pages per device memory allocation

      • deviceCachePercent (default: 0.05) percent of device memory used for PEFT cache after engine load and KV cache allocation

      • hostCacheSize (default: 1G) size in bytes of the host PEFT cache

Responses content

The responses from SendResponseCallback are stored in a std::shared_ptr<Tensor> list, which contains the following tensors of a specific request:

  • output Ids: a CPU tensor that contains the output token IDs. Its shape is [1, beamWidth, maxSeqLength].

  • sequence length: a CPU tensor that indicates the length of inputID + outputID. Its shape is [1, 1].

  • context logits: a CPU tensor that contains context logits. Its shape is [1, promptLength, vocabSizePadded] if the engine is built with gather_context_logits or gather_all_token_logits. Otherwise, it is a dummy tensor with shape [1, 1, 1].

  • generation logits: a CPU tensor that contains generation logits. Its shape is [1, beamWidth, outputLength, vocabSizePadded]. if the engine is built with gather_generation_logits or gather_all_token_logits. Otherwise, it is a dummy tensor with shape [1, 1, 1, 1]. If you are using gptManagerBenchmark.cpp, please remember to pass corresponding parameters --return-context-logits and/or --return-generation-logits to obtain these logits. Note that enabling return logits will require more device memory for converting and storing logits. To reduce redundant memory buffer allocation as much as possible, we recommend that the max_batch_size, max_beam_width, max_input_len, max_output_len, and other parameters set when building the engine are close to the values required during actual inference.

  • logProb: a CPU tensor that stores the log-prob of the generated tokens. Its shape is [1, beamWidth, outputLength]

  • cumLogProb: a CPU tensor that stores the cumLogProb. Its shape is [1, beamWidth]

GptManager Design

Batch Manager is designed to integrate into an inference server that’s executing a pool of active work items populated by a stream of requests actively received by the server. GptManager assumes a GPT-style autoregressive model architecture. GptManager spawns a worker thread in its constructor that then persistently runs the token generation loop. The worker thread invokes GetInferenceRequestsCallback at the start of each loop iteration, which is intended to read new requests. It invokes SendResponseCallback at the end of each iteration when one or more requests have generated a response to send back to the user. This response can be a single token in the case of requests that have streaming mode enabled or the full response when streaming mode is disabled. PollStopSignalCallback and ReturnBatchManagerStatsCallback, if provided, are both invoked at the end of each iteration loop. ReturnBatchManagerStatsCallback is not called when the system has no active requests. The server can safely retire requests from its pool of work items when notified of completion (via the final_response boolean argument) by the batch manager in SendResponseCallback. All TensorRT-LLM internal state related to that request will have been freed before this point. An instance of the batch manager to serve an auto-regressive model like GPT can be created as follows:

#include <tensorrt_llm/batch_manager/GptManager.h>

using namespace tensorrt_llm::batch_manager;

GptManager batchManager(pathToTrtEngine,                   // Path to the TensorRT engine of the model,
                        TrtGptModelType::InflightFusedBatching, // Use in-flight batching,
                        maxBeamWidth,                      // Maximum beam width (must be >= 1),
                        schedulerPolicy,                   // Scheduling policy (see below),
                        getInferenceRequestsCb,            // The Get callback (see above),
                        sendResponseCb,                    // The Send callback (see above),
                        pollStopSignalCb,                  // The Stop signals callback (see above),
                        returnBatchManagerStatsCb);        // The Return stats callback (see above),

The scheduler policy helps the batch manager adjust how requests are scheduled for execution. The batch manager can try to maximize the utilization of the GPUs by aggressively scheduling requests (schedulerPolicy set to MAX_UTILIZATION) at the risk of having to pause requests if it runs short on memory for KV caches. Note that any paused request will be automatically resumed and the only user-visible effect may be increased latency. It can also adopt a more conservative approach and schedule requests only when it knows that the memory allocation will be sufficient to process all active requests even in the worst case of KV cache consumption. That mode corresponds to a schedulerPolicy set to GUARANTEED_NO_EVICT.

The GptManager’s worker thread terminates when the GptManager destructor is called and there are no more active requests.

Multi-GPU execution

When running on multiple GPUs using either tensor or pipeline parallelism, it is assumed that the server launches as many processes as GPU ranks, and each process runs its own instance of GptManager. The number of GPUs visible on a given node can be controlled using the CUDA_VISIBLE_DEVICES environment variable.

Care must be taken to ensure all ranks see the same inputs at each iteration of the generation loop. In TensorRT-LLM Triton backend, an MPI broadcast is performed in GetInferenceRequestsCallback to ensure the same set of requests is seen by each of the MPI ranks. ReturnBatchManagerStatsCallback need only be called from a single rank; all ranks hold identical copies of the final results.

In-flight Batching with the Triton Inference Server

A Triton Inference Server C++ backend is provided with TensorRT-LLM that includes the mechanisms needed to serve models using in-flight batching. That backend is also a good starting example how to implement in-flight batching using the TensorRT-LLM batch manager.