Skip to content

[None][fix] using arrival time in llmapi when creating LlmRequest in pytorch workflow #7553

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 5 commits into from
Sep 15, 2025
Merged

[None][fix] using arrival time in llmapi when creating LlmRequest in pytorch workflow #7553

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8e3c208
provide arrivalTime when creating LlmRequest
zhengd-nv Sep 5, 2025
52c3b20
llmRequest update
zhengd-nv Sep 8, 2025
55bdd3b
Merge branch 'main' into arrival-time
zhengd-nv Sep 9, 2025
5c4a221
Merge branch 'main' into arrival-time
zhengd-nv Sep 10, 2025
13b1588
Merge remote-tracking branch 'origin/main' into arrival-time
zhengd-nv Sep 15, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
103 changes: 12 additions & 91 deletions cpp/include/tensorrt_llm/batch_manager/llmRequest.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -100,6 +100,7 @@ class GenericLlmRequest
RequestIdType, TensorPtr&, BeamTokens const&, TStream const&, std::optional<RequestIdType>)>;
using RequestPtr = std::shared_ptr<GenericLlmRequest>;
using MillisecondsType = std::chrono::milliseconds;
using TimePoint = std::chrono::time_point<std::chrono::steady_clock>;
using CacheSaltIDType = runtime::CacheSaltIDType;

GenericLlmRequest(RequestIdType requestId, SizeType32 maxNewTokens, std::shared_ptr<VecTokens> const& inputTokens,
Expand Down Expand Up @@ -138,7 +139,7 @@ class GenericLlmRequest
std::optional<SizeType32> languageAdapterUid = std::nullopt,
std::optional<MillisecondsType> allottedTimeMs = std::nullopt,
std::optional<executor::ContextPhaseParams> const& contextPhaseParams = std::nullopt,
std::optional<CacheSaltIDType> cacheSaltID = std::nullopt)
std::optional<CacheSaltIDType> cacheSaltID = std::nullopt, std::optional<TimePoint> arrivalTime = std::nullopt)
: mRequestId(requestId)
, mPromptLen(inputTokens->size())
, mMaxNewTokens(maxNewTokens)
Expand Down Expand Up @@ -202,7 +203,7 @@ class GenericLlmRequest
mState = LlmRequestState::kENCODER_INIT;
}

initialize(*inputTokens, returnLogProbs);
initialize(*inputTokens, returnLogProbs, arrivalTime);
}

GenericLlmRequest(RequestIdType requestId, SizeType32 maxNewTokens, VecTokens const& inputTokens,
Expand Down Expand Up @@ -2054,7 +2055,8 @@ class GenericLlmRequest
std::optional<CacheSaltIDType> mCacheSaltID{std::nullopt};

private:
void initialize(VecTokens const& inputTokens, bool outputLogProbs)
void initialize(
VecTokens const& inputTokens, bool outputLogProbs, std::optional<TimePoint> arrivalTime = std::nullopt)
{
if (mLlmRequestType == LlmRequestType::LLMREQUEST_TYPE_GENERATION_ONLY)
{
Expand Down Expand Up @@ -2148,7 +2150,7 @@ class GenericLlmRequest

if (mReturnPerfMetrics)
{
mPerfMetrics.timingMetrics.arrivalTime = std::chrono::steady_clock::now();
mPerfMetrics.timingMetrics.arrivalTime = arrivalTime.value_or(std::chrono::steady_clock::now());
}
mStartTime = std::chrono::steady_clock::now();
}
Expand Down Expand Up @@ -2197,61 +2199,9 @@ class LlmRequest : public GenericLlmRequest<runtime::ITensor::SharedPtr>
using TokenExtraIdType = Base::TokenExtraIdType;
using VecTokenExtraIds = Base::VecTokenExtraIds;

// 49 parameters, 49 parameters in Base class constructor
LlmRequest(RequestIdType requestId, SizeType32 maxNewTokens, std::shared_ptr<VecTokens> inputTokens,
runtime::SamplingConfig const& samplingConfig, bool isStreaming, std::optional<SizeType32> endId = std::nullopt,
std::optional<SizeType32> padId = std::nullopt, std::optional<TensorPtr> embeddingBias = std::nullopt,
std::optional<TensorPtr> badWordsList = std::nullopt, std::optional<TensorPtr> stopWordsList = std::nullopt,
std::optional<std::shared_ptr<std::vector<SizeType32>>> positionIds = std::nullopt,
std::optional<TensorPtr> promptEmbeddingTable = std::nullopt,
std::optional<SizeType32> promptVocabSize = std::nullopt,
std::optional<std::shared_ptr<std::vector<std::vector<SizeType32>>>> multimodalHashes = std::nullopt,
std::optional<std::shared_ptr<std::vector<SizeType32>>> multimodalPositions = std::nullopt,
std::optional<std::shared_ptr<std::vector<SizeType32>>> multimodalLengths = std::nullopt,
std::optional<TensorPtr> multimodalEmbedding = std::nullopt,
std::optional<TensorPtr> mropeRotaryCosSin = std::nullopt,
std::optional<SizeType32> mropePositionDeltas = std::nullopt,
std::optional<LoraTaskIdType> loraTaskId = std::nullopt, std::optional<TensorPtr> loraWeights = std::nullopt,
std::optional<TensorPtr> loraConfig = std::nullopt,
std::optional<executor::LookaheadDecodingConfig> lookaheadConfig = std::nullopt,
std::optional<executor::KvCacheRetentionConfig> kvCacheRetentionConfig = std::nullopt,
bool returnLogProbs = false, bool returnContextLogits = false, bool returnGenerationLogits = false,
std::optional<std::shared_ptr<VecTokens>> const& draftTokens = std::nullopt,
std::optional<TensorPtr> draftLogits = std::nullopt, bool excludeInputFromOutput = false,
std::optional<LogitsPostProcessor> logitsPostProcessor = std::nullopt,
bool applyLogitsPostProcessorBatched = false,
std::optional<std::shared_ptr<VecTokens>> encoderInputTokens = std::nullopt, bool returnEncoderOutput = false,
std::optional<RequestIdType> clientId = std::nullopt,
executor::PriorityType priority = executor::Request::kDefaultPriority,
std::optional<TensorPtr> encoderInputFeatures = std::nullopt,
std::optional<SizeType32> encoderOutputLength = std::nullopt,
std::optional<TensorPtr> crossAttentionMask = std::nullopt,
LlmRequestType llmRequestType = LlmRequestType::LLMREQUEST_TYPE_CONTEXT_AND_GENERATION,
std::optional<std::shared_ptr<VecTokenExtraIds>> inputTokenExtraIds = std::nullopt,
SizeType32 numReturnSequences = 1, std::optional<executor::EagleConfig> eagleConfig = std::nullopt,
std::optional<TensorPtr> skipCrossAttnBlocks = std::nullopt, bool returnPerfMetrics = false,
std::optional<executor::GuidedDecodingParams> guidedDecodingParams = std::nullopt,
std::optional<SizeType32> languageAdapterUid = std::nullopt,
std::optional<MillisecondsType> allottedTimeMs = std::nullopt,
std::optional<executor::ContextPhaseParams> const& contextPhaseParams = std::nullopt,
std::optional<CacheSaltIDType> cacheSaltID = std::nullopt)
: Base(requestId, maxNewTokens, std::move(inputTokens), samplingConfig, isStreaming, endId, padId,
std::move(embeddingBias), std::move(badWordsList), std::move(stopWordsList), std::move(positionIds),
std::move(promptEmbeddingTable), promptVocabSize, std::move(multimodalHashes),
std::move(multimodalPositions), std::move(multimodalLengths), std::move(multimodalEmbedding),
std::move(mropeRotaryCosSin), mropePositionDeltas, loraTaskId, std::move(loraWeights),
std::move(loraConfig), std::move(lookaheadConfig), std::move(kvCacheRetentionConfig), returnLogProbs,
returnContextLogits, returnGenerationLogits, std::move(draftTokens), std::move(draftLogits),
excludeInputFromOutput, std::move(logitsPostProcessor), applyLogitsPostProcessorBatched,
std::move(encoderInputTokens), returnEncoderOutput, clientId, priority, std::move(encoderInputFeatures),
std::move(encoderOutputLength), std::move(crossAttentionMask), llmRequestType,
std::move(inputTokenExtraIds), numReturnSequences, std::move(eagleConfig), std::move(skipCrossAttnBlocks),
returnPerfMetrics, std::move(guidedDecodingParams), languageAdapterUid, allottedTimeMs, contextPhaseParams,
cacheSaltID)
{
}

// 49 parameters, 49 parameters in Base class constructor
// inherit constructors
using Base::Base;

LlmRequest(RequestIdType requestId, SizeType32 maxNewTokens, std::vector<TokenIdType> inputTokens,
runtime::SamplingConfig const& samplingConfig, bool isStreaming, std::optional<SizeType32> endId = std::nullopt,
std::optional<SizeType32> padId = std::nullopt, std::optional<TensorPtr> embeddingBias = std::nullopt,
Expand Down Expand Up @@ -2286,7 +2236,7 @@ class LlmRequest : public GenericLlmRequest<runtime::ITensor::SharedPtr>
std::optional<SizeType32> languageAdapterUid = std::nullopt,
std::optional<MillisecondsType> allottedTimeMs = std::nullopt,
std::optional<executor::ContextPhaseParams> const& contextPhaseParams = std::nullopt,
std::optional<CacheSaltIDType> cacheSaltID = std::nullopt)
std::optional<CacheSaltIDType> cacheSaltID = std::nullopt, std::optional<TimePoint> arrivalTime = std::nullopt)
: Base(requestId, maxNewTokens, std::make_shared<std::vector<TokenIdType>>(std::move(inputTokens)),
samplingConfig, isStreaming, endId, padId, std::move(embeddingBias), std::move(badWordsList),
std::move(stopWordsList),
Expand Down Expand Up @@ -2316,37 +2266,8 @@ class LlmRequest : public GenericLlmRequest<runtime::ITensor::SharedPtr>
inputTokenExtraIds ? std::make_optional(std::make_shared<VecTokenExtraIds>(std::move(*inputTokenExtraIds)))
: std::optional<std::shared_ptr<VecTokenExtraIds>>(std::nullopt),
numReturnSequences, std::move(eagleConfig), skipCrossAttnBlocks, returnPerfMetrics,
std::move(guidedDecodingParams), languageAdapterUid, allottedTimeMs, contextPhaseParams, cacheSaltID)
{
}

// 32 parameters, 32 parameters in Base class constructor
LlmRequest(RequestIdType requestId, SizeType32 maxNewTokens, VecTokens const& inputTokens,
runtime::SamplingConfig const& samplingConfig, bool isStreaming, std::optional<SizeType32> endId = std::nullopt,
std::optional<SizeType32> padId = std::nullopt, std::optional<TensorPtr> embeddingBias = std::nullopt,
std::optional<TensorPtr> badWordsList = std::nullopt, std::optional<TensorPtr> stopWordsList = std::nullopt,
std::optional<std::shared_ptr<std::vector<SizeType32>>> positionIds = std::nullopt,
std::optional<TensorPtr> promptEmbeddingTable = std::nullopt,
std::optional<SizeType32> promptVocabSize = std::nullopt,
std::optional<LoraTaskIdType> loraTaskId = std::nullopt, std::optional<TensorPtr> loraWeights = std::nullopt,
std::optional<TensorPtr> loraConfig = std::nullopt,
std::optional<executor::LookaheadDecodingConfig> lookaheadConfig = std::nullopt, bool returnLogProbs = false,
bool returnContextLogits = false, bool returnGenerationLogits = false,
std::optional<VecTokens> draftTokens = std::nullopt, std::optional<TensorPtr> draftLogits = std::nullopt,
bool excludeInputFromOutput = false, std::optional<LogitsPostProcessor> logitsPostProcessor = std::nullopt,
bool applyLogitsPostProcessorBatched = false, std::optional<VecTokens> encoderInputTokens = std::nullopt,
bool returnEncoderOutput = false, std::optional<RequestIdType> clientId = std::nullopt,
executor::PriorityType priority = executor::Request::kDefaultPriority, SizeType32 numReturnSequences = 1,
std::optional<SizeType32> languageAdapterUid = std::nullopt,
std::optional<executor::ContextPhaseParams> const& contextPhaseParams = std::nullopt,
std::optional<CacheSaltIDType> cacheSaltID = std::nullopt)
: Base(requestId, maxNewTokens, inputTokens, samplingConfig, isStreaming, endId, padId,
std::move(embeddingBias), std::move(badWordsList), std::move(stopWordsList), std::move(positionIds),
std::move(promptEmbeddingTable), promptVocabSize, loraTaskId, std::move(loraWeights), std::move(loraConfig),
lookaheadConfig, returnLogProbs, returnContextLogits, returnGenerationLogits, std::move(draftTokens),
std::move(draftLogits), excludeInputFromOutput, std::move(logitsPostProcessor),
applyLogitsPostProcessorBatched, std::move(encoderInputTokens), returnEncoderOutput, clientId, priority,
numReturnSequences, languageAdapterUid, contextPhaseParams, cacheSaltID)
std::move(guidedDecodingParams), languageAdapterUid, allottedTimeMs, contextPhaseParams, cacheSaltID,
arrivalTime)
{
}

Expand Down
11 changes: 7 additions & 4 deletions cpp/tensorrt_llm/nanobind/batch_manager/bindings.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,7 @@

#include <ATen/ATen.h>
#include <nanobind/nanobind.h>
#include <nanobind/stl/chrono.h>
#include <nanobind/stl/optional.h>
#include <nanobind/stl/shared_ptr.h>
#include <nanobind/stl/tuple.h>
Expand Down Expand Up @@ -289,7 +290,8 @@ void initBindings(nb::module_& m)
std::optional<tb::LlmRequest::SizeType32> language_adapter_uid,
std::optional<tb::LlmRequest::MillisecondsType> allotted_time_ms,
std::optional<executor::ContextPhaseParams> context_phase_params,
std::optional<tb::LlmRequest::CacheSaltIDType> cache_salt_id)
std::optional<tb::LlmRequest::CacheSaltIDType> cache_salt_id,
std::optional<tb::LlmRequest::TimePoint> arrival_time)
{
auto makeOptionalTensor = [](std::optional<at::Tensor> const& atTensor, bool unsqueeze = false)
{
Expand Down Expand Up @@ -329,8 +331,8 @@ void initBindings(nb::module_& m)
encoder_input_tokens, return_encoder_output, client_id, priority, encoder_input_features_tensor_ptr,
encoder_output_length, cross_attention_mask_tensor_ptr, llm_request_type, input_token_extra_ids,
num_return_sequences, eagle_config, skip_cross_attn_blocks_tensor_ptr, return_perf_metrics,
guided_decoding_params, language_adapter_uid, allotted_time_ms, context_phase_params,
cache_salt_id};
guided_decoding_params, language_adapter_uid, allotted_time_ms, context_phase_params, cache_salt_id,
arrival_time};
},
nb::arg("request_id"), nb::arg("max_new_tokens"), nb::arg("input_tokens"), nb::arg("sampling_config"),
nb::arg("is_streaming"), nb::arg("end_id") = std::nullopt, nb::arg("pad_id") = std::nullopt,
Expand All @@ -355,7 +357,8 @@ void initBindings(nb::module_& m)
nb::arg("eagle_config") = std::nullopt, nb::arg("skip_cross_attn_blocks") = std::nullopt,
nb::arg("return_perf_metrics") = false, nb::arg("guided_decoding_params") = std::nullopt,
nb::arg("language_adapter_uid") = std::nullopt, nb::arg("allotted_time_ms") = std::nullopt,
nb::arg("context_phase_params") = std::nullopt, nb::arg("cache_salt_id") = std::nullopt)
nb::arg("context_phase_params") = std::nullopt, nb::arg("cache_salt_id") = std::nullopt,
nb::arg("arrival_time") = std::nullopt)
.def("check_token_id_range", &tb::LlmRequest::checkTokenIdRange, nb::arg("vocab_size"))
.def(nb::init<tb::LlmRequest const&>())
.def("validate", &tb::LlmRequest::validate, nb::arg("max_input_len"), nb::arg("max_seq_len"),
Expand Down
3 changes: 2 additions & 1 deletion cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -126,6 +126,7 @@ std::shared_ptr<tb::LlmRequest> LlmRequest::toTrtLlm() const
mLanguageAdapterUid, //
mAllottedTimeMs, //
mContextPhaseParams, //
mCacheSaltID //
mCacheSaltID, //
mPerfMetrics.timingMetrics.arrivalTime //
);
}
6 changes: 4 additions & 2 deletions cpp/tensorrt_llm/nanobind/batch_manager/llmRequest.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -85,7 +85,8 @@ class LlmRequest : public tb::GenericLlmRequest<at::Tensor, c10::Stream>
std::optional<SizeType32> languageAdapterUid = std::nullopt,
std::optional<MillisecondsType> allottedTimeMs = std::nullopt,
std::optional<executor::ContextPhaseParams> const& contextPhaseParams = std::nullopt,
std::optional<CacheSaltIDType> cacheSaltID = std::nullopt)
std::optional<CacheSaltIDType> cacheSaltID = std::nullopt,
std::optional<TimePoint> arrivalTime = std::nullopt)
: Base(requestId, //
maxNewTokens, //
std::make_shared<std::vector<TokenIdType>>(std::move(inputTokens)), //
Expand Down Expand Up @@ -147,7 +148,8 @@ class LlmRequest : public tb::GenericLlmRequest<at::Tensor, c10::Stream>
languageAdapterUid, //
allottedTimeMs, //
contextPhaseParams, //
cacheSaltID //
cacheSaltID, //
arrivalTime //
)
{
}
Expand Down
3 changes: 3 additions & 0 deletions cpp/tensorrt_llm/nanobind/bindings.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@
#include <nanobind/nanobind.h>
#include <nanobind/operators.h>
#include <nanobind/stl/bind_vector.h>
#include <nanobind/stl/chrono.h>
#include <nanobind/stl/filesystem.h>
#include <nanobind/stl/optional.h>
#include <nanobind/stl/shared_ptr.h>
Expand Down Expand Up @@ -511,4 +512,6 @@ NB_MODULE(TRTLLM_NB_MODULE, m)
m.def("ipc_nvls_allocate", &tr::ipcNvlsAllocate, nb::rv_policy::reference);
m.def("ipc_nvls_free", &tr::ipcNvlsFree);
m.def("ipc_nvls_supported", &tr::ipcNvlsSupported);

m.def("steady_clock_now", []() { return std::chrono::steady_clock::now(); });
}
8 changes: 5 additions & 3 deletions cpp/tensorrt_llm/pybind/batch_manager/bindings.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -295,7 +295,8 @@ void initBindings(pybind11::module_& m)
std::optional<tb::LlmRequest::SizeType32> language_adapter_uid,
std::optional<tb::LlmRequest::MillisecondsType> allotted_time_ms,
std::optional<executor::ContextPhaseParams> context_phase_params,
std::optional<tb::LlmRequest::CacheSaltIDType> cache_salt_id)
std::optional<tb::LlmRequest::CacheSaltIDType> cache_salt_id,
std::optional<tb::LlmRequest::TimePoint> arrival_time)
{
auto makeOptionalTensor = [](std::optional<at::Tensor> const& atTensor, bool unsqueeze = false)
{
Expand Down Expand Up @@ -336,7 +337,7 @@ void initBindings(pybind11::module_& m)
encoder_input_features_tensor_ptr, encoder_output_length, cross_attention_mask_tensor_ptr,
llm_request_type, input_token_extra_ids, num_return_sequences, eagle_config,
skip_cross_attn_blocks_tensor_ptr, return_perf_metrics, guided_decoding_params,
language_adapter_uid, allotted_time_ms, context_phase_params, cache_salt_id};
language_adapter_uid, allotted_time_ms, context_phase_params, cache_salt_id, arrival_time};
}),
py::arg("request_id"), py::arg("max_new_tokens"), py::arg("input_tokens"), py::arg("sampling_config"),
py::arg("is_streaming"), py::arg("end_id") = std::nullopt, py::arg("pad_id") = std::nullopt,
Expand All @@ -362,7 +363,8 @@ void initBindings(pybind11::module_& m)
py::arg("eagle_config") = std::nullopt, py::arg("skip_cross_attn_blocks") = std::nullopt,
py::arg("return_perf_metrics") = false, py::arg("guided_decoding_params") = std::nullopt,
py::arg("language_adapter_uid") = std::nullopt, py::arg("allotted_time_ms") = std::nullopt,
py::arg("context_phase_params") = std::nullopt, py::arg("cache_salt_id") = std::nullopt)
py::arg("context_phase_params") = std::nullopt, py::arg("cache_salt_id") = std::nullopt,
py::arg("arrival_time") = std::nullopt)
.def("check_token_id_range", &tb::LlmRequest::checkTokenIdRange, py::arg("vocab_size"))
.def(py::init<tb::LlmRequest const&>())
.def("validate", &tb::LlmRequest::validate, py::arg("max_input_len"), py::arg("max_seq_len"),
Expand Down
3 changes: 2 additions & 1 deletion cpp/tensorrt_llm/pybind/batch_manager/llmRequest.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -125,6 +125,7 @@ std::shared_ptr<tb::LlmRequest> LlmRequest::toTrtLlm() const
mLanguageAdapterUid, //
mAllottedTimeMs, //
mContextPhaseParams, //
mCacheSaltID //
mCacheSaltID, //
mPerfMetrics.timingMetrics.arrivalTime //
);
}
Loading