18#ifndef NV_INFER_IMPL_H
19#define NV_INFER_IMPL_H
57class IActivationLayer;
62class IConcatenationLayer;
65class IConvolutionLayer;
67class ICumulativeLayer;
68class IDeconvolutionLayer;
69class IDequantizeLayer;
71class IDynamicQuantizeLayer;
73class IElementWiseLayer;
74class IEngineInspector;
75class IExecutionContext;
78class IGridSampleLayer;
83class IIfConditionalInputLayer;
84class IIfConditionalOutputLayer;
86class IKVCacheUpdateLayer;
89class ILoopOutputLayer;
91class IMatrixMultiplyLayer;
93class IDistCollectiveLayer;
94class INetworkDefinition;
95class INormalizationLayer;
99class IOptimizationProfile;
101class IParametricReLULayer;
106class IPluginRegistry;
108class IRotaryEmbeddingLayer;
134class IRaggedSoftMaxLayer;
135class IRecurrenceLayer;
139class IReverseSequenceLayer;
141class ISafeSerializedNetworkArtifacts;
145class ISerializationConfig;
155struct TimingCacheKey;
156struct TimingCacheValue;
158using TimingCacheKey = v_1_0::TimingCacheKey;
159using TimingCacheValue = v_1_0::TimingCacheValue;
163class ITripLimitLayer;
165class IUnsqueezeLayer;
245 virtual ~VRoot() noexcept = default;
248class VHostMemory : public VRoot
251 virtual void* data() const noexcept = 0;
252 virtual std::
size_t size() const noexcept = 0;
253 virtual
DataType type() const noexcept = 0;
257class VDimensionExpr : public VRoot
260 virtual bool isConstant()
const = 0;
261 virtual int64_t getConstantValue()
const = 0;
262 virtual bool isSizeTensor()
const = 0;
265class VExprBuilder :
public VRoot
268 virtual IDimensionExpr
const* constant(int64_t value) = 0;
269 virtual IDimensionExpr
const* operation(
272 virtual IDimensionExpr
const* declareSizeTensor(
273 int32_t outputIndex, IDimensionExpr
const& opt, IDimensionExpr
const& upper)
277class VRuntime :
public VRoot
280 virtual IRuntime* getPImpl() noexcept = 0;
281 virtual
nvinfer1::ICudaEngine* deserializeCudaEngine(
void const* blob, std::
size_t size) noexcept = 0;
282 virtual
void setDLACore(int32_t dlaCore) noexcept = 0;
283 virtual int32_t getDLACore() const noexcept = 0;
284 virtual int32_t getNbDLACores() const noexcept = 0;
285 virtual
void setGpuAllocator(
IGpuAllocator* allocator) noexcept = 0;
286 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
289 virtual
bool setMaxThreads(int32_t maxThreads) noexcept = 0;
290 virtual int32_t getMaxThreads() const noexcept = 0;
291 virtual
void setTemporaryDirectory(
char const*) noexcept = 0;
292 virtual
char const* getTemporaryDirectory() const noexcept = 0;
296 virtual
void setPluginRegistryParent(IPluginRegistry* parent) noexcept = 0;
297 virtual IRuntime* loadRuntime(
char const* path) noexcept = 0;
298 virtual
void setEngineHostCodeAllowed(
bool allowed) noexcept = 0;
299 virtual
bool getEngineHostCodeAllowed() const noexcept = 0;
302 virtual int64_t getEngineHeaderSize() const noexcept = 0;
303 virtual
EngineValidity getEngineValidity(
void const* blob, int64_t blobSize, uint64_t* diagnostics) const noexcept
305 virtual
void setDeferredWeightsLoading(
bool defer) noexcept = 0;
306 virtual
bool getDeferredWeightsLoading() const noexcept = 0;
318class VWeightsManager : public VRoot
321 TRT_NODISCARD virtual IWeightsManager* getPImpl() noexcept = 0;
324 void* dst, int64_t dstSize,
TensorLocation location, cudaStream_t stream) noexcept = 0;
326 TRT_NODISCARD virtual
bool unload(cudaStream_t stream) noexcept = 0;
328 void const* src, int64_t srcSize,
TensorLocation location, cudaStream_t stream) noexcept = 0;
330 uint64_t allocationHandle, int64_t offset, int64_t size, cudaStream_t stream) noexcept = 0;
333class VRefitter : public VRoot
336 virtual IRefitter* getPImpl() noexcept = 0;
337 virtual
bool setWeights(
char const* layerName,
WeightsRole role, const Weights weights) noexcept = 0;
338 virtual
bool refitCudaEngine() noexcept = 0;
339 virtual int32_t getMissing(int32_t size,
char const** layerNames,
WeightsRole* roles) noexcept = 0;
340 virtual int32_t getAll(int32_t size,
char const** layerNames,
WeightsRole* roles) noexcept = 0;
341 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
343 virtual
bool setNamedWeights(
char const* name, Weights weights) noexcept = 0;
344 virtual int32_t getMissingWeights(int32_t size,
char const** weightsNames) noexcept = 0;
345 virtual int32_t getAllWeights(int32_t size,
char const** weightsNames) noexcept = 0;
347 virtual
bool setMaxThreads(int32_t maxThreads) noexcept = 0;
348 virtual int32_t getMaxThreads() const noexcept = 0;
349 virtual
bool setNamedWeightsWithLocation(
char const* name, Weights weights,
TensorLocation location) noexcept = 0;
350 virtual Weights getNamedWeights(
char const* weightsName) const noexcept = 0;
351 virtual
TensorLocation getWeightsLocation(
char const* weightsName) const noexcept = 0;
352 virtual
bool unsetNamedWeights(
char const* weightsName) noexcept = 0;
353 virtual
void setWeightsValidation(
bool weightsValidation) noexcept = 0;
354 virtual
bool getWeightsValidation() const noexcept = 0;
355 virtual
bool refitCudaEngineAsync(cudaStream_t stream) noexcept = 0;
356 virtual Weights getWeightsPrototype(
char const* weightsName) const noexcept = 0;
358 virtual
bool releaseRefitResources() noexcept = 0;
361class VOptimizationProfile : public VRoot
364 virtual bool setDimensions(
char const* inputName,
OptProfileSelector select,
Dims const& dims)
noexcept = 0;
366 virtual int32_t getNbShapeValues(
char const* inputName)
const noexcept = 0;
367 virtual bool setExtraMemoryTarget(
float target)
noexcept = 0;
368 virtual float getExtraMemoryTarget() const noexcept = 0;
369 virtual
bool isValid() const noexcept = 0;
372 char const* inputName,
OptProfileSelector select, int64_t const* values, int32_t nbValues) noexcept
378 virtual
void setProfileStream(cudaStream_t stream) noexcept = 0;
379 TRT_NODISCARD virtual cudaStream_t getProfileStream() const noexcept = 0;
382class VCudaEngine : public VRoot
385 virtual ICudaEngine* getPImpl() noexcept = 0;
386 virtual int32_t getNbLayers() const noexcept = 0;
387 virtual IHostMemory* serialize() const noexcept = 0;
389 virtual
bool isRefittable() const noexcept = 0;
390 virtual
char const* getName() const noexcept = 0;
391 virtual int32_t getNbOptimizationProfiles() const noexcept = 0;
393 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
397 virtual IEngineInspector* createEngineInspector() const noexcept = 0;
398 virtual
Dims getTensorShape(
char const* tensorName) const noexcept = 0;
399 virtual
DataType getTensorDataType(
char const* tensorName) const noexcept = 0;
400 virtual
TensorLocation getTensorLocation(
char const* tensorName) const noexcept = 0;
401 virtual
bool isShapeInferenceIO(
char const* tensorName) const noexcept = 0;
402 virtual
TensorIOMode getTensorIOMode(
char const* tensorName) const noexcept = 0;
403 virtual int32_t getTensorBytesPerComponent(
char const* tensorName) const noexcept = 0;
404 virtual int32_t getTensorComponentsPerElement(
char const* tensorName) const noexcept = 0;
405 virtual
TensorFormat getTensorFormat(
char const* tensorName) const noexcept = 0;
406 virtual
char const* getTensorFormatDesc(
char const* tensorName) const noexcept = 0;
407 virtual int32_t getTensorVectorizedDim(
char const* tensorName) const noexcept = 0;
408 virtual
Dims getProfileShape(
409 char const* tensorName, int32_t profileIndex,
OptProfileSelector select) const noexcept = 0;
410 virtual int32_t getNbIOTensors() const noexcept = 0;
411 virtual
char const* getIOTensorName(int32_t index) const noexcept = 0;
413 virtual int32_t getNbAuxStreams() const noexcept = 0;
415 virtual int32_t getTensorBytesPerComponentV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
416 virtual int32_t getTensorComponentsPerElementV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
417 virtual
TensorFormat getTensorFormatV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
418 virtual
char const* getTensorFormatDescV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
419 virtual int32_t getTensorVectorizedDimV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
421 virtual ISerializationConfig* createSerializationConfig() noexcept = 0;
422 virtual IHostMemory* serializeWithConfig(ISerializationConfig& config) const noexcept = 0;
424 virtual IRefitter* createRefitter(
ILogger& logger) noexcept = 0;
426 virtual int64_t getStreamableWeightsSize() const noexcept = 0;
428 virtual
bool isDebugTensor(
char const* name) const noexcept = 0;
431 virtual
bool setWeightStreamingBudgetV2(int64_t gpuMemoryBudget) noexcept = 0;
432 virtual int64_t getWeightStreamingBudgetV2() const noexcept = 0;
433 virtual int64_t getWeightStreamingAutomaticBudget() const noexcept = 0;
434 virtual int64_t getWeightStreamingScratchMemorySize() const noexcept = 0;
435 virtual int64_t getDeviceMemorySizeV2() const noexcept = 0;
436 virtual int64_t getDeviceMemorySizeForProfileV2(int32_t profileIndex) const noexcept = 0;
438 TRT_NODISCARD virtual int64_t const* getProfileTensorValuesV2(
439 char const* tensorName, int32_t profileIndex,
OptProfileSelector select) const noexcept = 0;
440 TRT_NODISCARD virtual IExecutionContext* createExecutionContextWithRuntimeConfig(
441 IRuntimeConfig* runtimeConfig) noexcept = 0;
442 TRT_NODISCARD virtual IRuntimeConfig* createRuntimeConfig() noexcept = 0;
445 TRT_NODISCARD virtual
char const* getAliasedInputTensor(
char const* tensorName) const noexcept = 0;
446 TRT_NODISCARD virtual
bool areWeightsLoaded() const noexcept = 0;
447 TRT_NODISCARD virtual
bool loadWeights(
void const* blob, int64_t size) noexcept = 0;
451 TRT_NODISCARD virtual IWeightsManager* createWeightsManager() noexcept = 0;
454class VExecutionContext : public VRoot
457 virtual IExecutionContext* getPImpl() noexcept = 0;
458 virtual
void setDebugSync(
bool sync) noexcept = 0;
459 virtual
bool getDebugSync() const noexcept = 0;
460 virtual
void setProfiler(
IProfiler*) noexcept = 0;
461 virtual
IProfiler* getProfiler() const noexcept = 0;
462 virtual ICudaEngine const& getEngine() const noexcept = 0;
463 virtual
void setName(
char const* name) noexcept = 0;
464 virtual
char const* getName() const noexcept = 0;
465 virtual
void setDeviceMemory(
void* memory) noexcept = 0;
466 virtual int32_t getOptimizationProfile() const noexcept = 0;
467 virtual
bool allInputDimensionsSpecified() const noexcept = 0;
468 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
470 virtual
bool executeV2(
void* const* bindings) noexcept = 0;
471 virtual
bool setOptimizationProfileAsync(int32_t profileIndex, cudaStream_t stream) noexcept = 0;
472 virtual
void setEnqueueEmitsProfile(
bool enqueueEmitsProfile) noexcept = 0;
473 virtual
bool getEnqueueEmitsProfile() const noexcept = 0;
474 virtual
bool reportToProfiler() const noexcept = 0;
475 virtual
bool setInputShape(
char const* tensorName,
Dims const& dims) noexcept = 0;
476 virtual
Dims getTensorShape(
char const* tensorName) const noexcept = 0;
477 virtual
Dims getTensorStrides(
char const* tensorName) const noexcept = 0;
478 virtual
bool setTensorAddress(
char const* tensorName,
void* data) noexcept = 0;
479 virtual
void const* getTensorAddress(
char const* tensorName) const noexcept = 0;
480 virtual
bool setInputTensorAddress(
char const* tensorName,
void const* data) noexcept = 0;
481 virtual
bool setOutputTensorAddress(
char const* tensorName,
void* data) noexcept = 0;
482 virtual int32_t inferShapes(int32_t nbMaxNames,
char const** tensorNames) noexcept = 0;
483 virtual
bool setInputConsumedEvent(cudaEvent_t event) noexcept = 0;
484 virtual cudaEvent_t getInputConsumedEvent() const noexcept = 0;
485 virtual
void* getOutputTensorAddress(
char const* tensorName) const noexcept = 0;
486 virtual
bool setOutputAllocator(
char const* tensorName,
IOutputAllocator* outputAllocator) noexcept = 0;
487 virtual
IOutputAllocator* getOutputAllocator(
char const* name) noexcept = 0;
488 virtual int64_t getMaxOutputSize(
char const* tensorName) const noexcept = 0;
489 virtual
bool setTemporaryStorageAllocator(
IGpuAllocator* allocator) noexcept = 0;
490 virtual
IGpuAllocator* getTemporaryStorageAllocator() const noexcept = 0;
491 virtual
bool enqueueV3(cudaStream_t stream) noexcept = 0;
492 virtual
void setPersistentCacheLimit(
size_t size) noexcept = 0;
493 virtual
size_t getPersistentCacheLimit() const noexcept = 0;
496 virtual
void setAuxStreams(cudaStream_t* auxStreams, int32_t nbStreams) noexcept = 0;
497 virtual
bool setDebugListener(
IDebugListener* listener) noexcept = 0;
499 virtual
bool setTensorDebugState(
char const* name,
bool flag) noexcept = 0;
500 virtual
bool getDebugState(
char const* name) const noexcept = 0;
501 virtual
bool setAllTensorsDebugState(
bool flag) noexcept = 0;
502 virtual
size_t updateDeviceMemorySizeForShapes() noexcept = 0;
503 virtual
void setDeviceMemoryV2(
void* memory, int64_t size) noexcept = 0;
504 TRT_NODISCARD virtual IRuntimeConfig* getRuntimeConfig() const noexcept = 0;
505 virtual
bool setUnfusedTensorsDebugState(
bool flag) noexcept = 0;
506 virtual
bool getUnfusedTensorsDebugState() const noexcept = 0;
507 virtual
bool isStreamCapturable(cudaStream_t stream) const noexcept = 0;
508 virtual
bool setCommunicator(
void* communicator) noexcept = 0;
511class VEngineInspector : public VRoot
514 virtual IEngineInspector* getPImpl() noexcept = 0;
515 virtual
bool setExecutionContext(IExecutionContext const* context) noexcept = 0;
516 virtual IExecutionContext const* getExecutionContext() const noexcept = 0;
517 virtual
char const* getLayerInformation(int32_t layerIndex,
LayerInformationFormat format) const noexcept = 0;
519 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
523class VTensor : public VRoot
526 virtual void setName(
char const* name)
noexcept = 0;
527 virtual char const* getName() const noexcept = 0;
528 virtual
void setDimensions(
Dims const& dimensions) noexcept = 0;
529 virtual
Dims getDimensions() const noexcept = 0;
530 virtual
DataType getType() const noexcept = 0;
531 virtual
bool isNetworkInput() const noexcept = 0;
532 virtual
bool isNetworkOutput() const noexcept = 0;
533 virtual
void setAllowedFormats(
TensorFormats formats) noexcept = 0;
534 virtual
TensorFormats getAllowedFormats() const noexcept = 0;
535 virtual
bool isShapeTensor() const noexcept = 0;
536 virtual
bool isExecutionTensor() const noexcept = 0;
537 virtual
void setDimensionName(int32_t index,
char const* name) noexcept = 0;
538 virtual
char const* getDimensionName(int32_t index) const noexcept = 0;
541class VLayer : public VRoot
544 virtual LayerType getType() const noexcept = 0;
545 virtual
void setName(
char const* name) noexcept = 0;
546 virtual
char const* getName() const noexcept = 0;
547 virtual int32_t getNbInputs() const noexcept = 0;
548 virtual ITensor* getInput(int32_t index) const noexcept = 0;
549 virtual int32_t getNbOutputs() const noexcept = 0;
550 virtual ITensor* getOutput(int32_t index) const noexcept = 0;
551 virtual
void setInput(int32_t index, ITensor& tensor) noexcept = 0;
552 virtual
DataType getOutputType(int32_t index) const noexcept = 0;
553 virtual
void setMetadata(
char const* docString) noexcept = 0;
554 virtual
char const* getMetadata() const noexcept = 0;
555 virtual
bool setNbRanks(int32_t nbRanks) noexcept = 0;
556 virtual int32_t getNbRanks() const noexcept = 0;
559class VConvolutionLayer : public VRoot
562 virtual void setNbOutputMaps(int64_t nbOutputMaps)
noexcept = 0;
563 virtual int64_t getNbOutputMaps() const noexcept = 0;
564 virtual
void setNbGroups(int64_t nbGroups) noexcept = 0;
565 virtual int64_t getNbGroups() const noexcept = 0;
566 virtual
void setKernelWeights(Weights weights) noexcept = 0;
567 virtual Weights getKernelWeights() const noexcept = 0;
568 virtual
void setBiasWeights(Weights weights) noexcept = 0;
569 virtual Weights getBiasWeights() const noexcept = 0;
570 virtual
void setPrePadding(
Dims const& padding) noexcept = 0;
571 virtual
Dims getPrePadding() const noexcept = 0;
572 virtual
void setPostPadding(
Dims const& padding) noexcept = 0;
573 virtual
Dims getPostPadding() const noexcept = 0;
574 virtual
void setPaddingMode(
PaddingMode paddingMode) noexcept = 0;
575 virtual
PaddingMode getPaddingMode() const noexcept = 0;
576 virtual
void setKernelSizeNd(
Dims const& kernelSize) noexcept = 0;
577 virtual
Dims getKernelSizeNd() const noexcept = 0;
578 virtual
void setStrideNd(
Dims const& stride) noexcept = 0;
579 virtual
Dims getStrideNd() const noexcept = 0;
580 virtual
void setPaddingNd(
Dims const& padding) noexcept = 0;
581 virtual
Dims getPaddingNd() const noexcept = 0;
582 virtual
void setDilationNd(
Dims const& dilation) noexcept = 0;
583 virtual
Dims getDilationNd() const noexcept = 0;
586class VActivationLayer : public VRoot
591 virtual
void setAlpha(
float alpha) noexcept = 0;
592 virtual
void setBeta(
float beta) noexcept = 0;
593 virtual
float getAlpha() const noexcept = 0;
594 virtual
float getBeta() const noexcept = 0;
597class VPoolingLayer : public VRoot
600 virtual void setPoolingType(
PoolingType type)
noexcept = 0;
601 virtual PoolingType getPoolingType() const noexcept = 0;
602 virtual
void setBlendFactor(
float blendFactor) noexcept = 0;
603 virtual
float getBlendFactor() const noexcept = 0;
604 virtual
void setAverageCountExcludesPadding(
bool exclusive) noexcept = 0;
605 virtual
bool getAverageCountExcludesPadding() const noexcept = 0;
606 virtual
void setPrePadding(
Dims const& padding) noexcept = 0;
607 virtual
Dims getPrePadding() const noexcept = 0;
608 virtual
void setPostPadding(
Dims const& padding) noexcept = 0;
609 virtual
Dims getPostPadding() const noexcept = 0;
610 virtual
void setPaddingMode(
PaddingMode paddingMode) noexcept = 0;
611 virtual
PaddingMode getPaddingMode() const noexcept = 0;
612 virtual
void setWindowSizeNd(
Dims const& windowSize) noexcept = 0;
613 virtual
Dims getWindowSizeNd() const noexcept = 0;
614 virtual
void setStrideNd(
Dims const& stride) noexcept = 0;
615 virtual
Dims getStrideNd() const noexcept = 0;
616 virtual
void setPaddingNd(
Dims const& padding) noexcept = 0;
617 virtual
Dims getPaddingNd() const noexcept = 0;
620class VLRNLayer : public VRoot
623 virtual void setWindowSize(int64_t windowSize)
noexcept = 0;
624 virtual int64_t getWindowSize() const noexcept = 0;
625 virtual
void setAlpha(
float alpha) noexcept = 0;
626 virtual
float getAlpha() const noexcept = 0;
627 virtual
void setBeta(
float beta) noexcept = 0;
628 virtual
float getBeta() const noexcept = 0;
629 virtual
void setK(
float k) noexcept = 0;
630 virtual
float getK() const noexcept = 0;
633class VScaleLayer : public VRoot
636 virtual void setMode(
ScaleMode mode)
noexcept = 0;
637 virtual ScaleMode getMode() const noexcept = 0;
638 virtual
void setShift(Weights shift) noexcept = 0;
639 virtual Weights getShift() const noexcept = 0;
640 virtual
void setScale(Weights scale) noexcept = 0;
641 virtual Weights getScale() const noexcept = 0;
642 virtual
void setPower(Weights power) noexcept = 0;
643 virtual Weights getPower() const noexcept = 0;
644 virtual int32_t getChannelAxis() const noexcept = 0;
645 virtual
void setChannelAxis(int32_t channelAxis) noexcept = 0;
648class VSoftMaxLayer : public VRoot
651 virtual void setAxes(uint32_t axes)
noexcept = 0;
652 virtual uint32_t getAxes() const noexcept = 0;
655class VConcatenationLayer : public VRoot
658 virtual void setAxis(int32_t axis)
noexcept = 0;
659 virtual int32_t getAxis() const noexcept = 0;
662class VDeconvolutionLayer : public VRoot
665 virtual void setNbOutputMaps(int64_t nbOutputMaps)
noexcept = 0;
666 virtual int64_t getNbOutputMaps() const noexcept = 0;
667 virtual
void setNbGroups(int64_t nbGroups) noexcept = 0;
668 virtual int64_t getNbGroups() const noexcept = 0;
669 virtual
void setKernelWeights(Weights weights) noexcept = 0;
670 virtual Weights getKernelWeights() const noexcept = 0;
671 virtual
void setBiasWeights(Weights weights) noexcept = 0;
672 virtual Weights getBiasWeights() const noexcept = 0;
673 virtual
void setPrePadding(
Dims const& padding) noexcept = 0;
674 virtual
Dims getPrePadding() const noexcept = 0;
675 virtual
void setPostPadding(
Dims const& padding) noexcept = 0;
676 virtual
Dims getPostPadding() const noexcept = 0;
677 virtual
void setPaddingMode(
PaddingMode paddingMode) noexcept = 0;
678 virtual
PaddingMode getPaddingMode() const noexcept = 0;
679 virtual
void setKernelSizeNd(
Dims const& kernelSize) noexcept = 0;
680 virtual
Dims getKernelSizeNd() const noexcept = 0;
681 virtual
void setStrideNd(
Dims const& stride) noexcept = 0;
682 virtual
Dims getStrideNd() const noexcept = 0;
683 virtual
void setPaddingNd(
Dims const& padding) noexcept = 0;
684 virtual
Dims getPaddingNd() const noexcept = 0;
685 virtual
void setDilationNd(
Dims const& dilation) noexcept = 0;
686 virtual
Dims getDilationNd() const noexcept = 0;
689class VElementWiseLayer : public VRoot
696class VGatherLayer : public VRoot
699 virtual void setGatherAxis(int32_t axis)
noexcept = 0;
700 virtual int32_t getGatherAxis() const noexcept = 0;
701 virtual
void setNbElementWiseDims(int32_t k) noexcept = 0;
702 virtual int32_t getNbElementWiseDims() const noexcept = 0;
703 virtual
void setMode(
GatherMode mode) noexcept = 0;
704 virtual
GatherMode getMode() const noexcept = 0;
707class VPluginLayer : public VRoot
710 virtual IPlugin& getPlugin() noexcept = 0;
713class VPluginV2Layer : public VRoot
716 virtual IPluginV2& getPlugin() noexcept = 0;
719class VPluginV3Layer : public VRoot
722 virtual IPluginV3& getPlugin() noexcept = 0;
725class VUnaryLayer : public VRoot
732class VReduceLayer : public VRoot
737 virtual
void setReduceAxes(uint32_t reduceAxes) noexcept = 0;
738 virtual uint32_t getReduceAxes() const noexcept = 0;
739 virtual
void setKeepDimensions(
bool keepDimensions) noexcept = 0;
740 virtual
bool getKeepDimensions() const noexcept = 0;
743class VPaddingLayer : public VRoot
746 virtual void setPrePaddingNd(
Dims const& padding)
noexcept = 0;
747 virtual Dims getPrePaddingNd() const noexcept = 0;
748 virtual
void setPostPaddingNd(
Dims const& padding) noexcept = 0;
749 virtual
Dims getPostPaddingNd() const noexcept = 0;
752class VShuffleLayer : public VRoot
755 virtual void setFirstTranspose(Permutation
const& permutation)
noexcept = 0;
756 virtual Permutation
const& getFirstTranspose() const noexcept = 0;
757 virtual
void setReshapeDimensions(
Dims const& dimensions) noexcept = 0;
758 virtual
Dims getReshapeDimensions() const noexcept = 0;
759 virtual
void setSecondTranspose(Permutation const& permutation) noexcept = 0;
760 virtual Permutation const& getSecondTranspose() const noexcept = 0;
761 virtual
void setZeroIsPlaceholder(
bool zeroIsPlaceholder) noexcept = 0;
762 virtual
bool getZeroIsPlaceholder() const noexcept = 0;
765class VSliceLayer : public VRoot
768 virtual void setStart(
Dims const& start)
noexcept = 0;
769 virtual Dims getStart() const noexcept = 0;
770 virtual
void setSize(
Dims const& size) noexcept = 0;
771 virtual
Dims getSize() const noexcept = 0;
772 virtual
void setStride(
Dims const& stride) noexcept = 0;
773 virtual
Dims getStride() const noexcept = 0;
774 virtual
void setMode(
SampleMode mode) noexcept = 0;
775 virtual
SampleMode getMode() const noexcept = 0;
776 virtual
void setAxes(
Dims const& axes) noexcept = 0;
777 virtual
Dims getAxes() const noexcept = 0;
780class VShapeLayer : public VRoot
785class VTopKLayer :
public VRoot
790 virtual
void setK(int32_t k) noexcept = 0;
791 virtual int32_t getK() const noexcept = 0;
792 virtual
void setReduceAxes(uint32_t reduceAxes) noexcept = 0;
793 virtual uint32_t getReduceAxes() const noexcept = 0;
794 virtual
bool setIndicesType(
DataType type) noexcept = 0;
795 virtual
DataType getIndicesType() const noexcept = 0;
798class VMatrixMultiplyLayer : public VRoot
801 virtual void setOperation(int32_t index,
MatrixOperation op)
noexcept = 0;
802 virtual MatrixOperation getOperation(int32_t index)
const noexcept = 0;
805class VNonZeroLayer :
public VRoot
808 virtual bool setIndicesType(
DataType type)
noexcept = 0;
809 virtual DataType getIndicesType() const noexcept = 0;
812class VRaggedSoftMaxLayer : public VRoot
817class VIdentityLayer :
public VRoot
822class VCastLayer :
public VRoot
825 virtual void setToType(
DataType toType)
noexcept = 0;
826 virtual DataType getToType() const noexcept = 0;
829class VConstantLayer : public VRoot
832 virtual void setWeights(Weights weights)
noexcept = 0;
833 virtual Weights getWeights() const noexcept = 0;
834 virtual
void setDimensions(
Dims const& dimensions) noexcept = 0;
835 virtual
Dims getDimensions() const noexcept = 0;
838class VParametricReLULayer : public VRoot
843class VResizeLayer :
public VRoot
846 virtual void setOutputDimensions(
Dims const& dimensions)
noexcept = 0;
847 virtual Dims getOutputDimensions() const noexcept = 0;
848 virtual
void setScales(
float const* scales, int32_t nbScales) noexcept = 0;
849 virtual int32_t getScales(int32_t size,
float* scales) const noexcept = 0;
854 virtual
void setSelectorForSinglePixel(
ResizeSelector selector) noexcept = 0;
855 virtual
ResizeSelector getSelectorForSinglePixel() const noexcept = 0;
858 virtual
void setCubicCoeff(
float value) noexcept = 0;
859 virtual
float getCubicCoeff() const noexcept = 0;
860 virtual
void setExcludeOutside(
bool value) noexcept = 0;
861 virtual
bool getExcludeOutside() const noexcept = 0;
864class VLoopBoundaryLayer : public VRoot
867 virtual ILoop* getLoop() const noexcept = 0;
870class VRecurrenceLayer : public VRoot
875class VLoopOutputLayer :
public VRoot
878 virtual LoopOutput getLoopOutput() const noexcept = 0;
879 virtual
void setAxis(int32_t axis) noexcept = 0;
880 virtual int32_t getAxis() const noexcept = 0;
883class VTripLimitLayer : public VRoot
886 virtual TripLimit getTripLimit() const noexcept = 0;
889class VIteratorLayer : public VRoot
892 virtual void setAxis(int32_t axis)
noexcept = 0;
893 virtual int32_t getAxis() const noexcept = 0;
894 virtual
void setReverse(
bool reverse) noexcept = 0;
895 virtual
bool getReverse() const noexcept = 0;
897class VLoop : public VRoot
900 virtual IRecurrenceLayer* addRecurrence(ITensor& initialValue)
noexcept = 0;
901 virtual ITripLimitLayer* addTripLimit(ITensor& tensor,
TripLimit limit)
noexcept = 0;
902 virtual IIteratorLayer* addIterator(ITensor& tensor, int32_t axis = 0,
bool reverse =
false) noexcept = 0;
903 virtual ILoopOutputLayer* addLoopOutput(ITensor& tensor,
LoopOutput outputKind, int32_t axis = 0) noexcept = 0;
904 virtual
void setName(
char const* name) noexcept = 0;
905 virtual
char const* getName() const noexcept = 0;
908class VConditionalBoundaryLayer : public VRoot
911 virtual IIfConditional* getConditional() const noexcept = 0;
914class VConditionLayer : public VRoot
919class VConditionalInputLayer :
public VRoot
924class VConditionalOutputLayer :
public VRoot
929class VIfConditional :
public VRoot
932 virtual IConditionLayer* setCondition(ITensor& tensor)
noexcept = 0;
933 virtual IIfConditionalInputLayer* addInput(ITensor& tensor)
noexcept = 0;
934 virtual IIfConditionalOutputLayer* addOutput(ITensor& trueTensor, ITensor& falseTensor)
noexcept = 0;
935 virtual void setName(
char const* name)
noexcept = 0;
936 virtual char const* getName() const noexcept = 0;
939class VAttentionBoundaryLayer : public VRoot
942 virtual IAttention* getAttention() const noexcept = 0;
945class VAttentionInputLayer : public VRoot
950class VAttentionOutputLayer :
public VRoot
955class VAttention :
public VRoot
958 TRT_NODISCARD virtual bool setInput(int32_t index, ITensor& input)
noexcept = 0;
959 TRT_NODISCARD virtual int32_t getNbInputs() const noexcept = 0;
960 TRT_NODISCARD virtual ITensor* getInput(int32_t index) const noexcept = 0;
961 TRT_NODISCARD virtual int32_t getNbOutputs() const noexcept = 0;
962 TRT_NODISCARD virtual ITensor* getOutput(int32_t index) const noexcept = 0;
963 TRT_NODISCARD virtual
bool setName(
char const* name) noexcept = 0;
964 TRT_NODISCARD virtual
char const* getName() const noexcept = 0;
967 TRT_DEPRECATED virtual
bool setCausal(
bool isCausal) noexcept = 0;
969 TRT_NODISCARD virtual
bool setMask(ITensor& mask) noexcept = 0;
971 TRT_NODISCARD virtual
bool setDecomposable(
bool decomposable) noexcept = 0;
972 TRT_NODISCARD virtual
bool getDecomposable() const noexcept = 0;
973 TRT_NODISCARD virtual
bool setNormalizationQuantizeScale(ITensor& tensor) noexcept = 0;
974 TRT_NODISCARD virtual ITensor* getNormalizationQuantizeScale() const noexcept = 0;
977 TRT_NODISCARD virtual
bool setMetadata(
char const* docString) noexcept = 0;
978 TRT_NODISCARD virtual
char const* getMetadata() const noexcept = 0;
979 TRT_NODISCARD virtual
bool setNbRanks(int32_t nbRanks) noexcept = 0;
980 TRT_NODISCARD virtual int32_t getNbRanks() const noexcept = 0;
987 TRT_NODISCARD virtual
bool setQueryLengths(ITensor* lengths) noexcept = 0;
988 TRT_NODISCARD virtual ITensor* getQueryLengths() const noexcept = 0;
989 TRT_NODISCARD virtual
bool setKeyValueLengths(ITensor* lengths) noexcept = 0;
990 TRT_NODISCARD virtual ITensor* getKeyValueLengths() const noexcept = 0;
993class VSelectLayer : public VRoot
997class VAssertionLayer :
public VRoot
1000 virtual void setMessage(
char const* message)
noexcept = 0;
1001 virtual char const* getMessage() const noexcept = 0;
1004class VFillLayer : public VRoot
1007 virtual void setDimensions(
Dims const& dimensions)
noexcept = 0;
1008 virtual Dims getDimensions() const noexcept = 0;
1011 virtual
void setAlpha(
double alpha) noexcept = 0;
1012 virtual
double getAlpha() const noexcept = 0;
1013 virtual
void setBeta(
double beta) noexcept = 0;
1014 virtual
double getBeta() const noexcept = 0;
1015 virtual
void setAlphaInt64(int64_t alpha) noexcept = 0;
1016 virtual int64_t getAlphaInt64() const noexcept = 0;
1017 virtual
void setBetaInt64(int64_t beta) noexcept = 0;
1018 virtual int64_t getBetaInt64() const noexcept = 0;
1019 virtual
bool isAlphaBetaInt64() const noexcept = 0;
1020 virtual
DataType getToType() const noexcept = 0;
1021 virtual
void setToType(
DataType toType) noexcept = 0;
1024class VQuantizeLayer : public VRoot
1027 virtual int32_t getAxis() const noexcept = 0;
1028 virtual
void setAxis(int32_t axis) noexcept = 0;
1029 virtual
DataType getToType() const noexcept = 0;
1030 virtual
void setToType(
DataType toType) noexcept = 0;
1031 virtual
Dims getBlockShape() const noexcept = 0;
1032 virtual
bool setBlockShape(
Dims const& blockShape) noexcept = 0;
1035class VDequantizeLayer : public VRoot
1038 virtual int32_t getAxis() const noexcept = 0;
1039 virtual
void setAxis(int32_t axis) noexcept = 0;
1040 virtual
DataType getToType() const noexcept = 0;
1041 virtual
void setToType(
DataType toType) noexcept = 0;
1042 virtual
Dims getBlockShape() const noexcept = 0;
1043 virtual
bool setBlockShape(
Dims const& blockShape) noexcept = 0;
1046class VDynamicQuantizeLayer : public VRoot
1051 TRT_DEPRECATED virtual int32_t getBlockSize() const noexcept = 0;
1052 TRT_DEPRECATED virtual
void setBlockSize(int32_t axis) noexcept = 0;
1053 virtual
DataType getScaleType() const noexcept = 0;
1054 virtual
void setScaleType(
DataType axis) noexcept = 0;
1055 virtual
DataType getToType() const noexcept = 0;
1056 virtual
void setToType(
DataType toType) noexcept = 0;
1057 virtual
Dims getBlockShape() const noexcept = 0;
1058 virtual
void setBlockShape(
Dims const& blockShape) noexcept = 0;
1061class VScatterLayer : public VRoot
1064 virtual void setMode(
ScatterMode mode)
noexcept = 0;
1066 virtual
void setAxis(int32_t axis) noexcept = 0;
1067 virtual int32_t getAxis() const noexcept = 0;
1070class VEinsumLayer : public VRoot
1073 virtual bool setEquation(
char const* equation)
noexcept = 0;
1074 virtual char const* getEquation() const noexcept = 0;
1077class VOneHotLayer : public VRoot
1080 virtual int32_t getAxis() const noexcept = 0;
1081 virtual
void setAxis(int32_t axis) noexcept = 0;
1084class VGridSampleLayer : public VRoot
1089 virtual
void setAlignCorners(
bool alignCorners) noexcept = 0;
1090 virtual
bool getAlignCorners() const noexcept = 0;
1091 virtual
bool setSampleMode(
SampleMode mode) noexcept = 0;
1092 virtual
SampleMode getSampleMode() const noexcept = 0;
1095class VNMSLayer : public VRoot
1100 virtual
void setTopKBoxLimit(int32_t limit) noexcept = 0;
1101 virtual int32_t getTopKBoxLimit() const noexcept = 0;
1102 virtual
bool setIndicesType(
DataType type) noexcept = 0;
1103 virtual
DataType getIndicesType() const noexcept = 0;
1106class VReverseSequenceLayer : public VRoot
1109 virtual void setBatchAxis(int32_t batchAxis)
noexcept = 0;
1110 virtual int32_t getBatchAxis() const noexcept = 0;
1112 virtual
void setSequenceAxis(int32_t sequenceAxis) noexcept = 0;
1113 virtual int32_t getSequenceAxis() const noexcept = 0;
1116class VNormalizationLayer : public VRoot
1119 virtual void setEpsilon(
float eps)
noexcept = 0;
1120 virtual float getEpsilon() const noexcept = 0;
1121 virtual
void setAxes(uint32_t axesMask) noexcept = 0;
1122 virtual uint32_t getAxes() const noexcept = 0;
1123 virtual
void setNbGroups(int64_t nbGroups) noexcept = 0;
1124 virtual int64_t getNbGroups() const noexcept = 0;
1125 virtual
bool isV2() const noexcept = 0;
1128class VSqueezeLayer : public VRoot
1132class VUnsqueezeLayer :
public VRoot
1136class VCumulativeLayer :
public VRoot
1141 virtual
void setExclusive(
bool exclusive) noexcept = 0;
1142 virtual
bool getExclusive() const noexcept = 0;
1143 virtual
void setReverse(
bool reverse) noexcept = 0;
1144 virtual
bool getReverse() const noexcept = 0;
1147class VRotaryEmbeddingLayer : public VRoot
1150 virtual void setInterleaved(
bool interleaved)
noexcept = 0;
1151 virtual bool getInterleaved() const noexcept = 0;
1152 virtual
bool setRotaryEmbeddingDim(int32_t rotaryEmbeddingDim) noexcept = 0;
1153 virtual int32_t getRotaryEmbeddingDim() const noexcept = 0;
1154 virtual
void setInput(int32_t index, ITensor& input) noexcept = 0;
1157class VKVCacheUpdateLayer : public VRoot
1164 TRT_NODISCARD virtual
bool setUpdateLengths(ITensor* lengths) noexcept = 0;
1165 TRT_NODISCARD virtual ITensor* getUpdateLengths() const noexcept = 0;
1168class VMoELayer : public VRoot
1171 virtual void setGatedWeights(ITensor& fcGateWeights, ITensor& fcUpWeights, ITensor& fcDownWeights,
MoEActType activationType)
noexcept = 0;
1172 virtual void setGatedBiases(ITensor& fcGateBiases, ITensor& fcUpBiases, ITensor& fcDownBiases)
noexcept = 0;
1173 virtual void setActivationType(
MoEActType activationType)
noexcept = 0;
1174 virtual MoEActType getActivationType() const noexcept = 0;
1175 virtual
void setQuantizationStatic(ITensor& fcDownActivationScale,
DataType dataType) noexcept = 0;
1176 virtual
void setQuantizationDynamicDblQ(ITensor& fcDownActivationDblQScale,
DataType dataType,
Dims const& blockShape,
DataType dynQOutputScaleType) noexcept = 0;
1177 virtual
void setQuantizationToType(
DataType type) noexcept = 0;
1178 virtual
DataType getQuantizationToType() const noexcept = 0;
1179 virtual
void setQuantizationBlockShape(
Dims const& blockShape) noexcept = 0;
1180 virtual
Dims getQuantizationBlockShape() const noexcept = 0;
1181 virtual
void setDynQOutputScaleType(
DataType type) noexcept = 0;
1182 virtual
DataType getDynQOutputScaleType() const noexcept = 0;
1183 virtual
void setSwigluParams(
float limit,
float alpha,
float beta) noexcept = 0;
1184 virtual
void setSwigluParamLimit(
float limit) noexcept = 0;
1185 virtual
float getSwigluParamLimit() const noexcept = 0;
1186 virtual
void setSwigluParamAlpha(
float alpha) noexcept = 0;
1187 virtual
float getSwigluParamAlpha() const noexcept = 0;
1188 virtual
void setSwigluParamBeta(
float beta) noexcept = 0;
1189 virtual
float getSwigluParamBeta() const noexcept = 0;
1190 virtual
void setInput(int32_t index, ITensor& tensor) noexcept = 0;
1194class VNetworkDefinition : public VRoot
1197 virtual ITensor* addInput(
char const* name,
DataType type,
Dims const& dimensions)
noexcept = 0;
1198 virtual void markOutput(ITensor& tensor)
noexcept = 0;
1199 virtual IActivationLayer* addActivation(ITensor& input,
ActivationType type)
noexcept = 0;
1200 virtual ILRNLayer* addLRN(ITensor& input, int64_t window,
float alpha,
float beta,
float k)
noexcept = 0;
1201 virtual IScaleLayer* addScale(
1202 ITensor& input,
ScaleMode mode, Weights shift, Weights scale, Weights power)
noexcept = 0;
1203 virtual ISoftMaxLayer* addSoftMax(ITensor& input)
noexcept = 0;
1204 virtual IConcatenationLayer* addConcatenation(ITensor*
const* inputs, int32_t nbInputs)
noexcept = 0;
1205 virtual IElementWiseLayer* addElementWise(ITensor& input1, ITensor& input2,
ElementWiseOperation op)
noexcept = 0;
1206 virtual IUnaryLayer* addUnary(ITensor& input,
UnaryOperation operation)
noexcept = 0;
1207 virtual IShuffleLayer* addShuffle(ITensor& input)
noexcept = 0;
1208 virtual int32_t getNbLayers() const noexcept = 0;
1209 virtual ILayer* getLayer(int32_t index) const noexcept = 0;
1210 virtual int32_t getNbInputs() const noexcept = 0;
1211 virtual ITensor* getInput(int32_t index) const noexcept = 0;
1212 virtual int32_t getNbOutputs() const noexcept = 0;
1213 virtual ITensor* getOutput(int32_t index) const noexcept = 0;
1214 virtual IReduceLayer* addReduce(
1215 ITensor& input,
ReduceOperation operation, uint32_t reduceAxes,
bool keepDimensions) noexcept
1217 virtual ITopKLayer* addTopK(ITensor& input,
TopKOperation op, int32_t k, uint32_t reduceAxes) noexcept = 0;
1218 virtual IGatherLayer* addGather(ITensor& data, ITensor& indices, int32_t axis) noexcept = 0;
1219 virtual IRaggedSoftMaxLayer* addRaggedSoftMax(ITensor& input, ITensor& bounds) noexcept = 0;
1220 virtual IMatrixMultiplyLayer* addMatrixMultiply(
1222 virtual IConstantLayer* addConstant(
Dims const& dimensions, Weights weights) noexcept = 0;
1223 virtual IIdentityLayer* addIdentity(ITensor& input) noexcept = 0;
1224 virtual
void removeTensor(ITensor& tensor) noexcept = 0;
1225 virtual
void unmarkOutput(ITensor& tensor) noexcept = 0;
1226 virtual ISliceLayer* addSlice(ITensor& input,
Dims const& start,
Dims const& size,
Dims const& stride) noexcept = 0;
1227 virtual
void setName(
char const* name) noexcept = 0;
1228 virtual
char const* getName() const noexcept = 0;
1229 virtual IShapeLayer* addShape(ITensor& input) noexcept = 0;
1230 virtual
bool hasImplicitBatchDimension() const noexcept = 0;
1231 virtual
bool markOutputForShapes(ITensor& tensor) noexcept = 0;
1232 virtual
bool unmarkOutputForShapes(ITensor& tensor) noexcept = 0;
1233 virtual IParametricReLULayer* addParametricReLU(ITensor& input, ITensor& slope) noexcept = 0;
1234 virtual IConvolutionLayer* addConvolutionNd(
1235 ITensor& input, int64_t nbOutputMaps,
Dims const& kernelSize, Weights kernelWeights, Weights biasWeights) noexcept
1237 virtual IPoolingLayer* addPoolingNd(ITensor& input,
PoolingType type,
Dims const& windowSize) noexcept = 0;
1238 virtual IDeconvolutionLayer* addDeconvolutionNd(
1239 ITensor& input, int64_t nbOutputMaps,
Dims const& kernelSize, Weights kernelWeights, Weights biasWeights) noexcept
1241 virtual IScaleLayer* addScaleNd(
1242 ITensor& input,
ScaleMode mode, Weights shift, Weights scale, Weights power, int32_t channelAxis) noexcept = 0;
1243 virtual IResizeLayer* addResize(ITensor& input) noexcept = 0;
1244 virtual ILoop* addLoop() noexcept = 0;
1245 virtual ISelectLayer* addSelect(ITensor& condition, ITensor& thenInput, ITensor& elseInput) noexcept = 0;
1246 virtual IPaddingLayer* addPaddingNd(ITensor& input,
Dims const& prePadding,
Dims const& postPadding) noexcept = 0;
1247 virtual
bool setWeightsName(Weights weights,
char const* name) noexcept = 0;
1248 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
1250 virtual IGatherLayer* addGatherV2(ITensor& data, ITensor& indices,
GatherMode mode) noexcept = 0;
1251 virtual IIfConditional* addIfConditional() noexcept = 0;
1252 virtual IScatterLayer* addScatter(ITensor& data, ITensor& indices, ITensor& updates,
ScatterMode mode) noexcept = 0;
1253 virtual IEinsumLayer* addEinsum(ITensor* const* inputs, int32_t nbInputs,
char const* equation) noexcept = 0;
1254 virtual IAssertionLayer* addAssertion(ITensor& condition,
char const* message) noexcept = 0;
1255 virtual IOneHotLayer* addOneHot(ITensor& indices, ITensor& values, ITensor& depth, int32_t axis) noexcept = 0;
1256 virtual INonZeroLayer* addNonZero(ITensor& input) noexcept = 0;
1257 virtual IGridSampleLayer* addGridSample(ITensor& input, ITensor& grid) noexcept = 0;
1258 virtual INMSLayer* addNMS(ITensor& boxes, ITensor& scores, ITensor& maxOutputBoxesPerClass) noexcept = 0;
1259 virtual IReverseSequenceLayer* addReverseSequence(ITensor& input, ITensor& sequenceLens) noexcept = 0;
1260 virtual INormalizationLayer* addNormalization(
1261 ITensor& input, ITensor& scale, ITensor& bias, uint32_t axesMask) noexcept = 0;
1262 virtual ICastLayer* addCast(ITensor& input,
DataType toType) noexcept = 0;
1263 virtual IBuilder& getBuilder() const noexcept = 0;
1266 virtual IQuantizeLayer* addQuantizeV2(ITensor& input, ITensor& scale,
DataType outputType) noexcept = 0;
1267 virtual IDequantizeLayer* addDequantizeV2(ITensor& input, ITensor& scale,
DataType outputType) noexcept = 0;
1269 virtual
bool markDebug(ITensor& tensor) noexcept = 0;
1270 virtual
bool unmarkDebug(ITensor& tensor) noexcept = 0;
1271 virtual
bool isDebugTensor(ITensor const& tensor) const noexcept = 0;
1272 virtual
bool markWeightsRefittable(
char const* name) noexcept = 0;
1273 virtual
bool unmarkWeightsRefittable(
char const* name) noexcept = 0;
1274 virtual
bool areWeightsMarkedRefittable(
char const* name) const noexcept = 0;
1275 virtual ISqueezeLayer* addSqueeze(ITensor& input, ITensor& axes) noexcept = 0;
1276 virtual IUnsqueezeLayer* addUnsqueeze(ITensor& input, ITensor& axes) noexcept = 0;
1277 virtual IDynamicQuantizeLayer* addDynamicQuantize(
1278 ITensor& input, int32_t axis, int32_t blockSize,
DataType toType,
DataType scaleType) noexcept = 0;
1279 virtual ICumulativeLayer* addCumulative(
1280 ITensor& input, ITensor& axis,
CumulativeOperation operation,
bool exclusive,
bool reverse) noexcept = 0;
1281 virtual
bool markUnfusedTensorsAsDebugTensors() noexcept = 0;
1282 virtual
bool unmarkUnfusedTensorsAsDebugTensors() noexcept = 0;
1283 virtual ITopKLayer* addTopKV2(
1284 ITensor& input,
TopKOperation op, int32_t k, uint32_t reduceAxes,
DataType indicesType) noexcept = 0;
1285 virtual INonZeroLayer* addNonZeroV2(ITensor& input,
DataType indicesType) noexcept = 0;
1286 virtual INMSLayer* addNMSV2(
1287 ITensor& boxes, ITensor& scores, ITensor& maxOutputBoxesPerClass,
DataType indicesType) noexcept = 0;
1290 virtual IRotaryEmbeddingLayer* addRotaryEmbedding(ITensor& input, ITensor& cosCache, ITensor& sinCache,
1291 bool interleaved, int32_t rotaryEmbeddingDim) noexcept = 0;
1292 virtual IDynamicQuantizeLayer* addDynamicQuantizeV2(
1294 virtual IKVCacheUpdateLayer* addKVCacheUpdate(
1295 ITensor& cache, ITensor& update, ITensor& writeIndices,
KVCacheMode cacheMode) noexcept = 0;
1296 virtual INormalizationLayer* addNormalizationV2(
1297 ITensor& input, ITensor& scale, ITensor& bias, uint32_t axesMask) noexcept = 0;
1298 virtual IMoELayer* addMoE(
1299 ITensor& hiddenStates, ITensor& selectedExpertsForTokens, ITensor& scoresForSelectedExperts) noexcept = 0;
1300 virtual IDistCollectiveLayer* addDistCollective(ITensor& input,
CollectiveOperation distCollectiveOp,
1301 ReduceOperation reduceOp, int64_t root, int64_t* groups, int64_t groupSize) noexcept = 0;
1304 virtual ISqueezeLayer* addSqueezeV2(ITensor& input, ITensor* axes) noexcept = 0;
1307class VTimingCache : public VRoot
1311 virtual
bool combine(ITimingCache const& inputCache,
bool ignoreMismatch) noexcept = 0;
1312 virtual
bool reset() noexcept = 0;
1313 virtual int64_t queryKeys(TimingCacheKey* keyBuffer, int64_t capacity) const noexcept = 0;
1314 virtual TimingCacheValue query(TimingCacheKey const& key) const noexcept = 0;
1315 virtual
bool update(TimingCacheKey const& key, TimingCacheValue const& value) noexcept = 0;
1319class VBuilderConfig : public VRoot
1322 virtual void setAvgTimingIterations(int32_t avgTiming)
noexcept = 0;
1323 virtual int32_t getAvgTimingIterations() const noexcept = 0;
1324 virtual
void setEngineCapability(
EngineCapability capability) noexcept = 0;
1326 virtual
void setFlags(
BuilderFlags builderFlags) noexcept = 0;
1328 virtual
void clearFlag(
BuilderFlag builderFlag) noexcept = 0;
1329 virtual
void setFlag(
BuilderFlag builderFlag) noexcept = 0;
1330 virtual
bool getFlag(
BuilderFlag builderFlag) const noexcept = 0;
1331 virtual
void setDeviceType(ILayer const* layer,
DeviceType deviceType) noexcept = 0;
1332 virtual
DeviceType getDeviceType(ILayer const* layer) const noexcept = 0;
1333 virtual
bool isDeviceTypeSet(ILayer const* layer) const noexcept = 0;
1334 virtual
void resetDeviceType(ILayer const* layer) noexcept = 0;
1335 virtual
bool canRunOnDLA(ILayer const* layer) const noexcept = 0;
1336 virtual
void setDLACore(int32_t dlaCore) noexcept = 0;
1337 virtual int32_t getDLACore() const noexcept = 0;
1338 virtual
void setDefaultDeviceType(
DeviceType deviceType) noexcept = 0;
1339 virtual
DeviceType getDefaultDeviceType() const noexcept = 0;
1340 virtual
void reset() noexcept = 0;
1341 virtual
void setProfileStream(const cudaStream_t stream) noexcept = 0;
1342 virtual cudaStream_t getProfileStream() const noexcept = 0;
1343 virtual int32_t addOptimizationProfile(IOptimizationProfile const* profile) noexcept = 0;
1344 virtual int32_t getNbOptimizationProfiles() const noexcept = 0;
1347 virtual
bool setTacticSources(
TacticSources tacticSources) noexcept = 0;
1348 virtual
TacticSources getTacticSources() const noexcept = 0;
1349 virtual
nvinfer1::ITimingCache* createTimingCache(
void const* blob, std::
size_t size) const noexcept = 0;
1350 virtual
bool setTimingCache(ITimingCache const& cache,
bool ignoreMismatch) noexcept = 0;
1351 virtual
nvinfer1::ITimingCache const* getTimingCache() const noexcept = 0;
1352 virtual
void setMemoryPoolLimit(
MemoryPoolType pool, std::
size_t poolSize) noexcept = 0;
1353 virtual std::
size_t getMemoryPoolLimit(
MemoryPoolType pool) const noexcept = 0;
1354 virtual
void setPreviewFeature(
PreviewFeature feature,
bool enable) noexcept = 0;
1355 virtual
bool getPreviewFeature(
PreviewFeature feature) const noexcept = 0;
1356 virtual
void setBuilderOptimizationLevel(int32_t level) noexcept = 0;
1357 virtual int32_t getBuilderOptimizationLevel() const noexcept = 0;
1360 virtual
void setPluginsToSerialize(
char const* const* paths, int32_t nbPaths) noexcept = 0;
1361 virtual
char const* getPluginToSerialize(int32_t index) const noexcept = 0;
1362 virtual int32_t getNbPluginsToSerialize() const noexcept = 0;
1363 virtual
bool setMaxAuxStreams(int32_t nbStreams) noexcept = 0;
1364 virtual int32_t getMaxAuxStreams() const noexcept = 0;
1367 virtual
void setRuntimePlatform(
RuntimePlatform runtimePlatform) noexcept = 0;
1369 virtual
void setMaxNbTactics(int32_t maxTactics) noexcept = 0;
1370 virtual int32_t getMaxNbTactics() const noexcept = 0;
1373 virtual
bool setL2LimitForTiling(int64_t size) noexcept = 0;
1374 virtual int64_t getL2LimitForTiling() const noexcept = 0;
1375 TRT_NODISCARD virtual
bool setNbComputeCapabilities(int32_t maxNbComputeCapabilities) noexcept = 0;
1376 TRT_NODISCARD virtual int32_t getNbComputeCapabilities() const noexcept = 0;
1379 virtual
bool setBuildRoute(
char const* buildRoute) noexcept = 0;
1380 virtual
char const* getBuildRoute() const noexcept = 0;
1381 virtual
char const* getAllBuildRoutes() const noexcept = 0;
1384class VSerializationConfig : public VRoot
1394class VBuilder : public VRoot
1397 virtual int32_t getMaxDLABatchSize() const noexcept = 0;
1398 virtual int32_t getNbDLACores() const noexcept = 0;
1399 virtual
void setGpuAllocator(
IGpuAllocator* allocator) noexcept = 0;
1400 virtual
nvinfer1::IBuilderConfig* createBuilderConfig() noexcept = 0;
1402 virtual
nvinfer1::IOptimizationProfile* createOptimizationProfile() noexcept = 0;
1403 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
1405 virtual
void reset() noexcept = 0;
1406 virtual
nvinfer1::IHostMemory* buildSerializedNetwork(
1407 INetworkDefinition& network, IBuilderConfig& config) noexcept = 0;
1408 virtual
bool isNetworkSupported(INetworkDefinition const& network, IBuilderConfig const& config) const noexcept = 0;
1410 virtual
bool setMaxThreads(int32_t maxThreads) noexcept = 0;
1411 virtual int32_t getMaxThreads() const noexcept = 0;
1413 virtual
bool buildSerializedNetworkToStream(
1414 INetworkDefinition& network, IBuilderConfig& config,
IStreamWriter& writer) noexcept = 0;
1417class VRuntimeConfig : public VRoot
1420 virtual IRuntimeConfig* getPImpl() noexcept = 0;
1424 virtual
bool setRuntimeCache(
nvinfer1::IRuntimeCache const& cache) noexcept = 0;
1426 virtual
void setDynamicShapesKernelSpecializationStrategy(
1433class VRuntimeCache : public VRoot
1437 virtual
bool deserialize(
void const* blob,
size_t size) noexcept = 0;
1438 virtual
bool reset() noexcept = 0;
1441class VDistCollectiveLayer : public VRoot
TENSORRTAPI nvinfer1::IPluginRegistry * getPluginRegistry() noexcept
Return the plugin registry.
TENSORRTAPI nvinfer1::ILogger * getLogger() noexcept
Return the logger object.
#define TRT_NODISCARD
A stand-in for [[nodiscard]] and [[nodiscard(REASON)]] that works with older compilers.
Definition: NvInferRuntimeBase.h:57
#define TRT_DEPRECATED
Definition: NvInferRuntimeBase.h:42
Structure to define the dimensions of a tensor.
User-implemented callback for notification when value of a debug tensor is updated.
Reference counted application-implemented error reporting interface for TensorRT objects.
Application-implemented class for controlling allocation on the GPU.
Callback from ExecutionContext::enqueueV3()
Plugin class for the V3 generation of user-implemented layers.
Application-implemented interface for profiling.
Application-implemented progress reporting interface for TensorRT.
Application-implemented class for reading data in a stream-based manner.
Application-implemented class for reading data in a stream-based manner asynchronously....
Application-implemented class for writing data in a stream-based manner.
Class to handle library allocated memory that is accessible to the user.
Definition: NvInferRuntime.h:150
The TensorRT API version 1 namespace.
Definition: NvInferPluginBase.h:29
uint32_t TacticSources
Represents a collection of one or more TacticSource values combine using bitwise-OR operations.
Definition: NvInferRuntime.h:3035
v_1_0::IOutputAllocator IOutputAllocator
Definition: NvInferRuntime.h:4617
ResizeSelector
The coordinate selector when resize to single pixel output.
Definition: NvInfer.h:3916
EngineCapability
List of supported engine capability flows.
Definition: NvInferRuntime.h:76
DimensionOperation
An operation on two IDimensionExpr, which represent integer expressions used in dimension computation...
Definition: NvInferRuntime.h:189
MemoryPoolType
The type for memory pools used by TensorRT.
Definition: NvInfer.h:10230
AttentionIOForm
Enumerates the layout of the input/output tensors in an Attention layer.
Definition: NvInfer.h:6733
ScaleMode
Controls how shift, scale and power are applied in a Scale layer.
Definition: NvInfer.h:1586
RuntimePlatform
Describes the intended runtime platform (operating system and CPU architecture) for the execution of ...
Definition: NvInfer.h:9841
TensorIOMode
Definition of tensor IO Mode.
Definition: NvInferRuntimeBase.h:664
HardwareCompatibilityLevel
Describes requirements of compatibility with GPU architectures other than that of the GPU on which th...
Definition: NvInfer.h:10343
CumulativeOperation
Enumerates the cumulative operations that may be performed by a Cumulative layer.
Definition: NvInfer.h:6541
SerializationFlag
List of valid flags that the engine can enable when serializing the bytes.
Definition: NvInferRuntime.h:3076
BoundingBoxFormat
Representation of bounding box data used for the Boxes input tensor in INMSLayer.
Definition: NvInfer.h:6102
ComputeCapability
Describes compute capability that an engine will be built for.
Definition: NvInfer.h:10392
v_1_0::IPluginV3 IPluginV3
Definition: NvInferPluginBase.h:273
UnaryOperation
Enumerates the unary operations that may be performed by a Unary layer.
Definition: NvInfer.h:2544
DLAWorkspaceAllocationStrategy
Describes how DLA workspace memory is allocated.
Definition: NvInferRuntime.h:1576
v_1_0::IStreamWriter IStreamWriter
Definition: NvInferRuntime.h:710
v_1_0::IProfiler IProfiler
Definition: NvInferRuntime.h:1320
DynamicShapesKernelSpecializationStrategy
Different kernel specialization strategies for dynamic shapes.
Definition: NvInferRuntime.h:3268
v_1_0::IStreamReaderV2 IStreamReaderV2
Definition: NvInferRuntime.h:787
ActivationType
Enumerates the types of activation to perform in an activation layer.
Definition: NvInfer.h:143
uint32_t TempfileControlFlags
Represents a collection of one or more TempfileControlFlag values combined using bitwise-OR operation...
Definition: NvInferRuntime.h:1398
FillOperation
Enumerates the tensor fill operations that may performed by a fill layer.
Definition: NvInfer.h:4919
EngineStat
The kind of engine statistics that queried from the ICudaEngine.
Definition: NvInferRuntime.h:3598
CudaGraphStrategy
Strategies available for CUDA graphs optimizations for JIT (Just-In-Time) inference.
Definition: NvInferRuntime.h:3304
ResizeRoundMode
The rounding mode for nearest neighbor resize.
Definition: NvInfer.h:3943
EngineValidity
Whether a TensorRT-RTX engine is likely to be valid on the current system.
Definition: NvInferRuntime.h:1898
v_1_0::ILogger ILogger
Definition: NvInferRuntimeBase.h:125
CausalMaskKind
Enumerates the causal mask alignment orientation for the attention.
Definition: NvInfer.h:6705
PaddingMode
Enumerates the modes of padding to perform in convolution, deconvolution and pooling layer,...
Definition: NvInfer.h:763
TripLimit
Enum that describes kinds of trip limits.
Definition: NvInfer.h:4301
uint32_t NetworkDefinitionCreationFlags
Represents one or more NetworkDefinitionCreationFlag flags using binary OR operations....
Definition: NvInfer.h:11497
PreviewFeature
Define preview features.
Definition: NvInfer.h:10309
TilingOptimizationLevel
Define the optimization levels for Tiling.
Definition: NvInfer.h:10420
DataType
The type of weights and tensors. The datatypes other than kBOOL, kINT32, and kINT64 are "activation d...
Definition: NvInferRuntimeBase.h:151
uint32_t BuilderFlags
Represents one or more BuilderFlag values using binary OR operations, e.g., 1U << BuilderFlag::kDEBUG...
Definition: NvInfer.h:9871
DeviceType
The device that this layer/network will execute on.
Definition: NvInferRuntime.h:1352
LayerType
The type values of layer classes.
Definition: NvInfer.h:58
SampleMode
Controls how ISliceLayer and IGridSample handle out-of-bounds coordinates.
Definition: NvInfer.h:3046
GatherMode
Control form of IGatherLayer.
Definition: NvInfer.h:2285
v_1_0::IDebugListener IDebugListener
Definition: NvInferRuntime.h:4658
MoEActType
Enumerates the activation type for the MoE layer.
Definition: NvInfer.h:7622
WeightsRole
How a layer uses particular Weights.
Definition: NvInferRuntime.h:1330
uint32_t TensorFormats
It is capable of representing one or more TensorFormat by binary OR operations, e....
Definition: NvInfer.h:135
ProfilingVerbosity
List of verbosity levels of layer information exposed in NVTX annotations and in IEngineInspector.
Definition: NvInferRuntime.h:3047
NetworkDefinitionCreationFlag
List of immutable network properties expressed at network creation time. NetworkDefinitionCreationFla...
Definition: NvInfer.h:11508
ElementWiseOperation
Enumerates the binary operations that may be performed by an ElementWise layer.
Definition: NvInfer.h:2196
CollectiveOperation
Enumerates the collective operations that may be performed by a DistCollective layer.
Definition: NvInfer.h:2674
uint32_t SerializationFlags
Represents one or more SerializationFlag values using binary OR operations, e.g., 1U << Serialization...
Definition: NvInferRuntime.h:3066
InterpolationMode
Enumerates various modes of interpolation.
Definition: NvInfer.h:3840
BuilderFlag
List of valid modes that the builder can enable when creating an engine from a network definition.
Definition: NvInfer.h:9881
TensorFormat
Format of the input/output tensors.
Definition: NvInferRuntime.h:1432
ExecutionContextAllocationStrategy
Different memory allocation behaviors for IExecutionContext.
Definition: NvInferRuntime.h:3183
TopKOperation
Enumerates the operations that may be performed by a TopK layer.
Definition: NvInfer.h:3331
ReduceOperation
Enumerates the reduce operations that may be performed by a Reduce layer.
Definition: NvInfer.h:2646
LayerInformationFormat
The format in which the IEngineInspector prints the layer information.
Definition: NvInferRuntime.h:5643
ScatterMode
Control form of IScatterLayer.
Definition: NvInfer.h:5829
MatrixOperation
Enumerates the operations that may be performed on a tensor by IMatrixMultiplyLayer before multiplica...
Definition: NvInfer.h:3486
ResizeCoordinateTransformation
The resize coordinate transformation function.
Definition: NvInfer.h:3865
LoopOutput
Enum that describes kinds of loop outputs.
Definition: NvInfer.h:4273
KVCacheMode
Enumerates the KVCache modes that may be performed by a KVCacheUpdate layer.
Definition: NvInfer.h:7469
v_1_0::IStreamReader IStreamReader
Definition: NvInferRuntime.h:700
PoolingType
The type of pooling to perform in a pooling layer.
Definition: NvInfer.h:1200
v_1_0::IProgressMonitor IProgressMonitor
Definition: NvInfer.h:10533
TensorLocation
The location for tensor data storage, device or host.
Definition: NvInferRuntime.h:214
OptProfileSelector
When setting or querying optimization profile parameters (such as shape tensor inputs or dynamic dime...
Definition: NvInferRuntime.h:2756
AttentionNormalizationOp
Enumerates the operations that may be performed by the normalization in the attention subgraph.
Definition: NvInfer.h:6673