18#ifndef NV_INFER_IMPL_H
19#define NV_INFER_IMPL_H
55class IActivationLayer;
60class IConcatenationLayer;
63class IConvolutionLayer;
65class ICumulativeLayer;
66class IDeconvolutionLayer;
67class IDequantizeLayer;
69class IDynamicQuantizeLayer;
71class IElementWiseLayer;
72class IEngineInspector;
73class IExecutionContext;
76class IGridSampleLayer;
81class IIfConditionalInputLayer;
82class IIfConditionalOutputLayer;
84class IKVCacheUpdateLayer;
87class ILoopOutputLayer;
89class IMatrixMultiplyLayer;
91class IDistCollectiveLayer;
92class INetworkDefinition;
93class INormalizationLayer;
97class IOptimizationProfile;
99class IParametricReLULayer;
104class IPluginRegistry;
106class IRotaryEmbeddingLayer;
131class IRaggedSoftMaxLayer;
132class IRecurrenceLayer;
136class IReverseSequenceLayer;
138class ISafeSerializedNetworkArtifacts;
142class ISerializationConfig;
152struct TimingCacheKey;
153struct TimingCacheValue;
155using TimingCacheKey = v_1_0::TimingCacheKey;
156using TimingCacheValue = v_1_0::TimingCacheValue;
160class ITripLimitLayer;
162class IUnsqueezeLayer;
238 virtual ~VRoot() noexcept = default;
241class VHostMemory : public VRoot
244 virtual void* data() const noexcept = 0;
245 virtual std::
size_t size() const noexcept = 0;
246 virtual
DataType type() const noexcept = 0;
250class VDimensionExpr : public VRoot
253 virtual bool isConstant()
const = 0;
254 virtual int64_t getConstantValue()
const = 0;
255 virtual bool isSizeTensor()
const = 0;
258class VExprBuilder :
public VRoot
261 virtual IDimensionExpr
const* constant(int64_t value) = 0;
262 virtual IDimensionExpr
const* operation(
265 virtual IDimensionExpr
const* declareSizeTensor(
266 int32_t outputIndex, IDimensionExpr
const& opt, IDimensionExpr
const& upper)
270class VRuntime :
public VRoot
273 virtual IRuntime* getPImpl() noexcept = 0;
274 virtual
nvinfer1::ICudaEngine* deserializeCudaEngine(
void const* blob, std::
size_t size) noexcept = 0;
275 virtual
void setDLACore(int32_t dlaCore) noexcept = 0;
276 virtual int32_t getDLACore() const noexcept = 0;
277 virtual int32_t getNbDLACores() const noexcept = 0;
278 virtual
void setGpuAllocator(
IGpuAllocator* allocator) noexcept = 0;
279 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
282 virtual
bool setMaxThreads(int32_t maxThreads) noexcept = 0;
283 virtual int32_t getMaxThreads() const noexcept = 0;
284 virtual
void setTemporaryDirectory(
char const*) noexcept = 0;
285 virtual
char const* getTemporaryDirectory() const noexcept = 0;
289 virtual
void setPluginRegistryParent(IPluginRegistry* parent) noexcept = 0;
290 virtual IRuntime* loadRuntime(
char const* path) noexcept = 0;
291 virtual
void setEngineHostCodeAllowed(
bool allowed) noexcept = 0;
292 virtual
bool getEngineHostCodeAllowed() const noexcept = 0;
306class VRefitter : public VRoot
309 virtual IRefitter* getPImpl() noexcept = 0;
310 virtual
bool setWeights(
char const* layerName,
WeightsRole role, const Weights weights) noexcept = 0;
311 virtual
bool refitCudaEngine() noexcept = 0;
312 virtual int32_t getMissing(int32_t size,
char const** layerNames,
WeightsRole* roles) noexcept = 0;
313 virtual int32_t getAll(int32_t size,
char const** layerNames,
WeightsRole* roles) noexcept = 0;
314 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
316 virtual
bool setNamedWeights(
char const* name, Weights weights) noexcept = 0;
317 virtual int32_t getMissingWeights(int32_t size,
char const** weightsNames) noexcept = 0;
318 virtual int32_t getAllWeights(int32_t size,
char const** weightsNames) noexcept = 0;
320 virtual
bool setMaxThreads(int32_t maxThreads) noexcept = 0;
321 virtual int32_t getMaxThreads() const noexcept = 0;
322 virtual
bool setNamedWeightsWithLocation(
char const* name, Weights weights,
TensorLocation location) noexcept = 0;
323 virtual Weights getNamedWeights(
char const* weightsName) const noexcept = 0;
324 virtual
TensorLocation getWeightsLocation(
char const* weightsName) const noexcept = 0;
325 virtual
bool unsetNamedWeights(
char const* weightsName) noexcept = 0;
326 virtual
void setWeightsValidation(
bool weightsValidation) noexcept = 0;
327 virtual
bool getWeightsValidation() const noexcept = 0;
328 virtual
bool refitCudaEngineAsync(cudaStream_t stream) noexcept = 0;
329 virtual Weights getWeightsPrototype(
char const* weightsName) const noexcept = 0;
332class VOptimizationProfile : public VRoot
335 virtual bool setDimensions(
char const* inputName,
OptProfileSelector select,
Dims const& dims)
noexcept = 0;
337 virtual int32_t getNbShapeValues(
char const* inputName)
const noexcept = 0;
338 virtual bool setExtraMemoryTarget(
float target)
noexcept = 0;
339 virtual float getExtraMemoryTarget() const noexcept = 0;
340 virtual
bool isValid() const noexcept = 0;
343 char const* inputName,
OptProfileSelector select, int64_t const* values, int32_t nbValues) noexcept = 0;
348class VCudaEngine : public VRoot
351 virtual ICudaEngine* getPImpl() noexcept = 0;
352 virtual int32_t getNbLayers() const noexcept = 0;
353 virtual IHostMemory* serialize() const noexcept = 0;
355 virtual
bool isRefittable() const noexcept = 0;
356 virtual
char const* getName() const noexcept = 0;
357 virtual int32_t getNbOptimizationProfiles() const noexcept = 0;
359 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
363 virtual IEngineInspector* createEngineInspector() const noexcept = 0;
364 virtual
Dims getTensorShape(
char const* tensorName) const noexcept = 0;
365 virtual
DataType getTensorDataType(
char const* tensorName) const noexcept = 0;
366 virtual
TensorLocation getTensorLocation(
char const* tensorName) const noexcept = 0;
367 virtual
bool isShapeInferenceIO(
char const* tensorName) const noexcept = 0;
368 virtual
TensorIOMode getTensorIOMode(
char const* tensorName) const noexcept = 0;
369 virtual int32_t getTensorBytesPerComponent(
char const* tensorName) const noexcept = 0;
370 virtual int32_t getTensorComponentsPerElement(
char const* tensorName) const noexcept = 0;
371 virtual
TensorFormat getTensorFormat(
char const* tensorName) const noexcept = 0;
372 virtual
char const* getTensorFormatDesc(
char const* tensorName) const noexcept = 0;
373 virtual int32_t getTensorVectorizedDim(
char const* tensorName) const noexcept = 0;
374 virtual
Dims getProfileShape(
375 char const* tensorName, int32_t profileIndex,
OptProfileSelector select) const noexcept = 0;
376 virtual int32_t getNbIOTensors() const noexcept = 0;
377 virtual
char const* getIOTensorName(int32_t index) const noexcept = 0;
379 virtual int32_t getNbAuxStreams() const noexcept = 0;
381 virtual int32_t getTensorBytesPerComponentV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
382 virtual int32_t getTensorComponentsPerElementV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
383 virtual
TensorFormat getTensorFormatV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
384 virtual
char const* getTensorFormatDescV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
385 virtual int32_t getTensorVectorizedDimV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
387 virtual ISerializationConfig* createSerializationConfig() noexcept = 0;
388 virtual IHostMemory* serializeWithConfig(ISerializationConfig& config) const noexcept = 0;
390 virtual IRefitter* createRefitter(
ILogger& logger) noexcept = 0;
392 virtual int64_t getStreamableWeightsSize() const noexcept = 0;
394 virtual
bool isDebugTensor(
char const* name) const noexcept = 0;
397 virtual
bool setWeightStreamingBudgetV2(int64_t gpuMemoryBudget) noexcept = 0;
398 virtual int64_t getWeightStreamingBudgetV2() const noexcept = 0;
399 virtual int64_t getWeightStreamingAutomaticBudget() const noexcept = 0;
400 virtual int64_t getWeightStreamingScratchMemorySize() const noexcept = 0;
401 virtual int64_t getDeviceMemorySizeV2() const noexcept = 0;
402 virtual int64_t getDeviceMemorySizeForProfileV2(int32_t profileIndex) const noexcept = 0;
404 TRT_NODISCARD virtual int64_t const* getProfileTensorValuesV2(
405 char const* tensorName, int32_t profileIndex,
OptProfileSelector select) const noexcept = 0;
406 TRT_NODISCARD virtual IExecutionContext* createExecutionContextWithRuntimeConfig(
407 IRuntimeConfig* runtimeConfig) noexcept = 0;
408 TRT_NODISCARD virtual IRuntimeConfig* createRuntimeConfig() noexcept = 0;
411 TRT_NODISCARD virtual
char const* getAliasedInputTensor(
char const* tensorName) const noexcept = 0;
414class VExecutionContext : public VRoot
417 virtual IExecutionContext* getPImpl() noexcept = 0;
418 virtual
void setDebugSync(
bool sync) noexcept = 0;
419 virtual
bool getDebugSync() const noexcept = 0;
420 virtual
void setProfiler(
IProfiler*) noexcept = 0;
421 virtual
IProfiler* getProfiler() const noexcept = 0;
422 virtual ICudaEngine const& getEngine() const noexcept = 0;
423 virtual
void setName(
char const* name) noexcept = 0;
424 virtual
char const* getName() const noexcept = 0;
425 virtual
void setDeviceMemory(
void* memory) noexcept = 0;
426 virtual int32_t getOptimizationProfile() const noexcept = 0;
427 virtual
bool allInputDimensionsSpecified() const noexcept = 0;
428 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
430 virtual
bool executeV2(
void* const* bindings) noexcept = 0;
431 virtual
bool setOptimizationProfileAsync(int32_t profileIndex, cudaStream_t stream) noexcept = 0;
432 virtual
void setEnqueueEmitsProfile(
bool enqueueEmitsProfile) noexcept = 0;
433 virtual
bool getEnqueueEmitsProfile() const noexcept = 0;
434 virtual
bool reportToProfiler() const noexcept = 0;
435 virtual
bool setInputShape(
char const* tensorName,
Dims const& dims) noexcept = 0;
436 virtual
Dims getTensorShape(
char const* tensorName) const noexcept = 0;
437 virtual
Dims getTensorStrides(
char const* tensorName) const noexcept = 0;
438 virtual
bool setTensorAddress(
char const* tensorName,
void* data) noexcept = 0;
439 virtual
void const* getTensorAddress(
char const* tensorName) const noexcept = 0;
440 virtual
bool setInputTensorAddress(
char const* tensorName,
void const* data) noexcept = 0;
441 virtual
bool setOutputTensorAddress(
char const* tensorName,
void* data) noexcept = 0;
442 virtual int32_t inferShapes(int32_t nbMaxNames,
char const** tensorNames) noexcept = 0;
443 virtual
bool setInputConsumedEvent(cudaEvent_t event) noexcept = 0;
444 virtual cudaEvent_t getInputConsumedEvent() const noexcept = 0;
445 virtual
void* getOutputTensorAddress(
char const* tensorName) const noexcept = 0;
446 virtual
bool setOutputAllocator(
char const* tensorName,
IOutputAllocator* outputAllocator) noexcept = 0;
447 virtual
IOutputAllocator* getOutputAllocator(
char const* name) noexcept = 0;
448 virtual int64_t getMaxOutputSize(
char const* tensorName) const noexcept = 0;
449 virtual
bool setTemporaryStorageAllocator(
IGpuAllocator* allocator) noexcept = 0;
450 virtual
IGpuAllocator* getTemporaryStorageAllocator() const noexcept = 0;
451 virtual
bool enqueueV3(cudaStream_t stream) noexcept = 0;
452 virtual
void setPersistentCacheLimit(
size_t size) noexcept = 0;
453 virtual
size_t getPersistentCacheLimit() const noexcept = 0;
456 virtual
void setAuxStreams(cudaStream_t* auxStreams, int32_t nbStreams) noexcept = 0;
457 virtual
bool setDebugListener(
IDebugListener* listener) noexcept = 0;
459 virtual
bool setTensorDebugState(
char const* name,
bool flag) noexcept = 0;
460 virtual
bool getDebugState(
char const* name) const noexcept = 0;
461 virtual
bool setAllTensorsDebugState(
bool flag) noexcept = 0;
462 virtual
size_t updateDeviceMemorySizeForShapes() noexcept = 0;
463 virtual
void setDeviceMemoryV2(
void* memory, int64_t size) noexcept = 0;
464 TRT_NODISCARD virtual IRuntimeConfig* getRuntimeConfig() const noexcept = 0;
465 virtual
bool setUnfusedTensorsDebugState(
bool flag) noexcept = 0;
466 virtual
bool getUnfusedTensorsDebugState() const noexcept = 0;
467 virtual
bool setCommunicator(
void* communicator) noexcept = 0;
470class VEngineInspector : public VRoot
473 virtual IEngineInspector* getPImpl() noexcept = 0;
474 virtual
bool setExecutionContext(IExecutionContext const* context) noexcept = 0;
475 virtual IExecutionContext const* getExecutionContext() const noexcept = 0;
476 virtual
char const* getLayerInformation(int32_t layerIndex,
LayerInformationFormat format) const noexcept = 0;
478 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
482class VTensor : public VRoot
485 virtual void setName(
char const* name)
noexcept = 0;
486 virtual char const* getName() const noexcept = 0;
487 virtual
void setDimensions(
Dims const& dimensions) noexcept = 0;
488 virtual
Dims getDimensions() const noexcept = 0;
489 virtual
DataType getType() const noexcept = 0;
490 virtual
bool isNetworkInput() const noexcept = 0;
491 virtual
bool isNetworkOutput() const noexcept = 0;
492 virtual
void setBroadcastAcrossBatch(
bool broadcastAcrossBatch) noexcept = 0;
493 virtual
bool getBroadcastAcrossBatch() const noexcept = 0;
496 virtual
void setAllowedFormats(
TensorFormats formats) noexcept = 0;
497 virtual
TensorFormats getAllowedFormats() const noexcept = 0;
498 virtual
bool isShapeTensor() const noexcept = 0;
499 virtual
bool isExecutionTensor() const noexcept = 0;
500 virtual
void setDimensionName(int32_t index,
char const* name) noexcept = 0;
501 virtual
char const* getDimensionName(int32_t index) const noexcept = 0;
504class VLayer : public VRoot
507 virtual LayerType getType() const noexcept = 0;
508 virtual
void setName(
char const* name) noexcept = 0;
509 virtual
char const* getName() const noexcept = 0;
510 virtual int32_t getNbInputs() const noexcept = 0;
511 virtual ITensor* getInput(int32_t index) const noexcept = 0;
512 virtual int32_t getNbOutputs() const noexcept = 0;
513 virtual ITensor* getOutput(int32_t index) const noexcept = 0;
514 virtual
void setInput(int32_t index, ITensor& tensor) noexcept = 0;
515 virtual
DataType getOutputType(int32_t index) const noexcept = 0;
516 virtual
void setMetadata(
char const* docString) noexcept = 0;
517 virtual
char const* getMetadata() const noexcept = 0;
518 virtual
bool setNbRanks(int32_t nbRanks) noexcept = 0;
519 virtual int32_t getNbRanks() const noexcept = 0;
522class VConvolutionLayer : public VRoot
525 virtual void setNbOutputMaps(int64_t nbOutputMaps)
noexcept = 0;
526 virtual int64_t getNbOutputMaps() const noexcept = 0;
527 virtual
void setNbGroups(int64_t nbGroups) noexcept = 0;
528 virtual int64_t getNbGroups() const noexcept = 0;
529 virtual
void setKernelWeights(Weights weights) noexcept = 0;
530 virtual Weights getKernelWeights() const noexcept = 0;
531 virtual
void setBiasWeights(Weights weights) noexcept = 0;
532 virtual Weights getBiasWeights() const noexcept = 0;
533 virtual
void setPrePadding(
Dims const& padding) noexcept = 0;
534 virtual
Dims getPrePadding() const noexcept = 0;
535 virtual
void setPostPadding(
Dims const& padding) noexcept = 0;
536 virtual
Dims getPostPadding() const noexcept = 0;
537 virtual
void setPaddingMode(
PaddingMode paddingMode) noexcept = 0;
538 virtual
PaddingMode getPaddingMode() const noexcept = 0;
539 virtual
void setKernelSizeNd(
Dims const& kernelSize) noexcept = 0;
540 virtual
Dims getKernelSizeNd() const noexcept = 0;
541 virtual
void setStrideNd(
Dims const& stride) noexcept = 0;
542 virtual
Dims getStrideNd() const noexcept = 0;
543 virtual
void setPaddingNd(
Dims const& padding) noexcept = 0;
544 virtual
Dims getPaddingNd() const noexcept = 0;
545 virtual
void setDilationNd(
Dims const& dilation) noexcept = 0;
546 virtual
Dims getDilationNd() const noexcept = 0;
549class VActivationLayer : public VRoot
554 virtual
void setAlpha(
float alpha) noexcept = 0;
555 virtual
void setBeta(
float beta) noexcept = 0;
556 virtual
float getAlpha() const noexcept = 0;
557 virtual
float getBeta() const noexcept = 0;
560class VPoolingLayer : public VRoot
563 virtual void setPoolingType(
PoolingType type)
noexcept = 0;
564 virtual PoolingType getPoolingType() const noexcept = 0;
565 virtual
void setBlendFactor(
float blendFactor) noexcept = 0;
566 virtual
float getBlendFactor() const noexcept = 0;
567 virtual
void setAverageCountExcludesPadding(
bool exclusive) noexcept = 0;
568 virtual
bool getAverageCountExcludesPadding() const noexcept = 0;
569 virtual
void setPrePadding(
Dims const& padding) noexcept = 0;
570 virtual
Dims getPrePadding() const noexcept = 0;
571 virtual
void setPostPadding(
Dims const& padding) noexcept = 0;
572 virtual
Dims getPostPadding() const noexcept = 0;
573 virtual
void setPaddingMode(
PaddingMode paddingMode) noexcept = 0;
574 virtual
PaddingMode getPaddingMode() const noexcept = 0;
575 virtual
void setWindowSizeNd(
Dims const& windowSize) noexcept = 0;
576 virtual
Dims getWindowSizeNd() const noexcept = 0;
577 virtual
void setStrideNd(
Dims const& stride) noexcept = 0;
578 virtual
Dims getStrideNd() const noexcept = 0;
579 virtual
void setPaddingNd(
Dims const& padding) noexcept = 0;
580 virtual
Dims getPaddingNd() const noexcept = 0;
583class VLRNLayer : public VRoot
586 virtual void setWindowSize(int64_t windowSize)
noexcept = 0;
587 virtual int64_t getWindowSize() const noexcept = 0;
588 virtual
void setAlpha(
float alpha) noexcept = 0;
589 virtual
float getAlpha() const noexcept = 0;
590 virtual
void setBeta(
float beta) noexcept = 0;
591 virtual
float getBeta() const noexcept = 0;
592 virtual
void setK(
float k) noexcept = 0;
593 virtual
float getK() const noexcept = 0;
596class VScaleLayer : public VRoot
599 virtual void setMode(
ScaleMode mode)
noexcept = 0;
600 virtual ScaleMode getMode() const noexcept = 0;
601 virtual
void setShift(Weights shift) noexcept = 0;
602 virtual Weights getShift() const noexcept = 0;
603 virtual
void setScale(Weights scale) noexcept = 0;
604 virtual Weights getScale() const noexcept = 0;
605 virtual
void setPower(Weights power) noexcept = 0;
606 virtual Weights getPower() const noexcept = 0;
607 virtual int32_t getChannelAxis() const noexcept = 0;
608 virtual
void setChannelAxis(int32_t channelAxis) noexcept = 0;
611class VSoftMaxLayer : public VRoot
614 virtual void setAxes(uint32_t axes)
noexcept = 0;
615 virtual uint32_t getAxes() const noexcept = 0;
618class VConcatenationLayer : public VRoot
621 virtual void setAxis(int32_t axis)
noexcept = 0;
622 virtual int32_t getAxis() const noexcept = 0;
625class VDeconvolutionLayer : public VRoot
628 virtual void setNbOutputMaps(int64_t nbOutputMaps)
noexcept = 0;
629 virtual int64_t getNbOutputMaps() const noexcept = 0;
630 virtual
void setNbGroups(int64_t nbGroups) noexcept = 0;
631 virtual int64_t getNbGroups() const noexcept = 0;
632 virtual
void setKernelWeights(Weights weights) noexcept = 0;
633 virtual Weights getKernelWeights() const noexcept = 0;
634 virtual
void setBiasWeights(Weights weights) noexcept = 0;
635 virtual Weights getBiasWeights() const noexcept = 0;
636 virtual
void setPrePadding(
Dims const& padding) noexcept = 0;
637 virtual
Dims getPrePadding() const noexcept = 0;
638 virtual
void setPostPadding(
Dims const& padding) noexcept = 0;
639 virtual
Dims getPostPadding() const noexcept = 0;
640 virtual
void setPaddingMode(
PaddingMode paddingMode) noexcept = 0;
641 virtual
PaddingMode getPaddingMode() const noexcept = 0;
642 virtual
void setKernelSizeNd(
Dims const& kernelSize) noexcept = 0;
643 virtual
Dims getKernelSizeNd() const noexcept = 0;
644 virtual
void setStrideNd(
Dims const& stride) noexcept = 0;
645 virtual
Dims getStrideNd() const noexcept = 0;
646 virtual
void setPaddingNd(
Dims const& padding) noexcept = 0;
647 virtual
Dims getPaddingNd() const noexcept = 0;
648 virtual
void setDilationNd(
Dims const& dilation) noexcept = 0;
649 virtual
Dims getDilationNd() const noexcept = 0;
652class VElementWiseLayer : public VRoot
659class VGatherLayer : public VRoot
662 virtual void setGatherAxis(int32_t axis)
noexcept = 0;
663 virtual int32_t getGatherAxis() const noexcept = 0;
664 virtual
void setNbElementWiseDims(int32_t k) noexcept = 0;
665 virtual int32_t getNbElementWiseDims() const noexcept = 0;
666 virtual
void setMode(
GatherMode mode) noexcept = 0;
667 virtual
GatherMode getMode() const noexcept = 0;
670class VPluginLayer : public VRoot
673 virtual IPlugin& getPlugin() noexcept = 0;
676class VPluginV2Layer : public VRoot
679 virtual IPluginV2& getPlugin() noexcept = 0;
682class VPluginV3Layer : public VRoot
685 virtual IPluginV3& getPlugin() noexcept = 0;
688class VUnaryLayer : public VRoot
695class VReduceLayer : public VRoot
700 virtual
void setReduceAxes(uint32_t reduceAxes) noexcept = 0;
701 virtual uint32_t getReduceAxes() const noexcept = 0;
702 virtual
void setKeepDimensions(
bool keepDimensions) noexcept = 0;
703 virtual
bool getKeepDimensions() const noexcept = 0;
706class VPaddingLayer : public VRoot
709 virtual void setPrePaddingNd(
Dims const& padding)
noexcept = 0;
710 virtual Dims getPrePaddingNd() const noexcept = 0;
711 virtual
void setPostPaddingNd(
Dims const& padding) noexcept = 0;
712 virtual
Dims getPostPaddingNd() const noexcept = 0;
715class VShuffleLayer : public VRoot
718 virtual void setFirstTranspose(Permutation
const& permutation)
noexcept = 0;
719 virtual Permutation
const& getFirstTranspose() const noexcept = 0;
720 virtual
void setReshapeDimensions(
Dims const& dimensions) noexcept = 0;
721 virtual
Dims getReshapeDimensions() const noexcept = 0;
722 virtual
void setSecondTranspose(Permutation const& permutation) noexcept = 0;
723 virtual Permutation const& getSecondTranspose() const noexcept = 0;
724 virtual
void setZeroIsPlaceholder(
bool zeroIsPlaceholder) noexcept = 0;
725 virtual
bool getZeroIsPlaceholder() const noexcept = 0;
728class VSliceLayer : public VRoot
731 virtual void setStart(
Dims const& start)
noexcept = 0;
732 virtual Dims getStart() const noexcept = 0;
733 virtual
void setSize(
Dims const& size) noexcept = 0;
734 virtual
Dims getSize() const noexcept = 0;
735 virtual
void setStride(
Dims const& stride) noexcept = 0;
736 virtual
Dims getStride() const noexcept = 0;
737 virtual
void setMode(
SampleMode mode) noexcept = 0;
738 virtual
SampleMode getMode() const noexcept = 0;
739 virtual
void setAxes(
Dims const& axes) noexcept = 0;
740 virtual
Dims getAxes() const noexcept = 0;
743class VShapeLayer : public VRoot
748class VTopKLayer :
public VRoot
753 virtual
void setK(int32_t k) noexcept = 0;
754 virtual int32_t getK() const noexcept = 0;
755 virtual
void setReduceAxes(uint32_t reduceAxes) noexcept = 0;
756 virtual uint32_t getReduceAxes() const noexcept = 0;
757 virtual
bool setIndicesType(
DataType type) noexcept = 0;
758 virtual
DataType getIndicesType() const noexcept = 0;
761class VMatrixMultiplyLayer : public VRoot
764 virtual void setOperation(int32_t index,
MatrixOperation op)
noexcept = 0;
765 virtual MatrixOperation getOperation(int32_t index)
const noexcept = 0;
768class VNonZeroLayer :
public VRoot
771 virtual bool setIndicesType(
DataType type)
noexcept = 0;
772 virtual DataType getIndicesType() const noexcept = 0;
775class VRaggedSoftMaxLayer : public VRoot
780class VIdentityLayer :
public VRoot
785class VCastLayer :
public VRoot
788 virtual void setToType(
DataType toType)
noexcept = 0;
789 virtual DataType getToType() const noexcept = 0;
792class VConstantLayer : public VRoot
795 virtual void setWeights(Weights weights)
noexcept = 0;
796 virtual Weights getWeights() const noexcept = 0;
797 virtual
void setDimensions(
Dims const& dimensions) noexcept = 0;
798 virtual
Dims getDimensions() const noexcept = 0;
801class VParametricReLULayer : public VRoot
806class VResizeLayer :
public VRoot
809 virtual void setOutputDimensions(
Dims const& dimensions)
noexcept = 0;
810 virtual Dims getOutputDimensions() const noexcept = 0;
811 virtual
void setScales(
float const* scales, int32_t nbScales) noexcept = 0;
812 virtual int32_t getScales(int32_t size,
float* scales) const noexcept = 0;
817 virtual
void setSelectorForSinglePixel(
ResizeSelector selector) noexcept = 0;
818 virtual
ResizeSelector getSelectorForSinglePixel() const noexcept = 0;
821 virtual
void setCubicCoeff(
float value) noexcept = 0;
822 virtual
float getCubicCoeff() const noexcept = 0;
823 virtual
void setExcludeOutside(
bool value) noexcept = 0;
824 virtual
bool getExcludeOutside() const noexcept = 0;
827class VLoopBoundaryLayer : public VRoot
830 virtual ILoop* getLoop() const noexcept = 0;
833class VRecurrenceLayer : public VRoot
838class VLoopOutputLayer :
public VRoot
841 virtual LoopOutput getLoopOutput() const noexcept = 0;
842 virtual
void setAxis(int32_t axis) noexcept = 0;
843 virtual int32_t getAxis() const noexcept = 0;
846class VTripLimitLayer : public VRoot
849 virtual TripLimit getTripLimit() const noexcept = 0;
852class VIteratorLayer : public VRoot
855 virtual void setAxis(int32_t axis)
noexcept = 0;
856 virtual int32_t getAxis() const noexcept = 0;
857 virtual
void setReverse(
bool reverse) noexcept = 0;
858 virtual
bool getReverse() const noexcept = 0;
860class VLoop : public VRoot
863 virtual IRecurrenceLayer* addRecurrence(ITensor& initialValue)
noexcept = 0;
864 virtual ITripLimitLayer* addTripLimit(ITensor& tensor,
TripLimit limit)
noexcept = 0;
865 virtual IIteratorLayer* addIterator(ITensor& tensor, int32_t axis = 0,
bool reverse =
false) noexcept = 0;
866 virtual ILoopOutputLayer* addLoopOutput(ITensor& tensor,
LoopOutput outputKind, int32_t axis = 0) noexcept = 0;
867 virtual
void setName(
char const* name) noexcept = 0;
868 virtual
char const* getName() const noexcept = 0;
871class VConditionalBoundaryLayer : public VRoot
874 virtual IIfConditional* getConditional() const noexcept = 0;
877class VConditionLayer : public VRoot
882class VConditionalInputLayer :
public VRoot
887class VConditionalOutputLayer :
public VRoot
892class VIfConditional :
public VRoot
895 virtual IConditionLayer* setCondition(ITensor& tensor)
noexcept = 0;
896 virtual IIfConditionalInputLayer* addInput(ITensor& tensor)
noexcept = 0;
897 virtual IIfConditionalOutputLayer* addOutput(ITensor& trueTensor, ITensor& falseTensor)
noexcept = 0;
898 virtual void setName(
char const* name)
noexcept = 0;
899 virtual char const* getName() const noexcept = 0;
902class VAttentionBoundaryLayer : public VRoot
905 virtual IAttention* getAttention() const noexcept = 0;
908class VAttentionInputLayer : public VRoot
913class VAttentionOutputLayer :
public VRoot
918class VAttention :
public VRoot
921 TRT_NODISCARD virtual bool setInput(int32_t index, ITensor& input)
noexcept = 0;
922 TRT_NODISCARD virtual int32_t getNbInputs() const noexcept = 0;
923 TRT_NODISCARD virtual ITensor* getInput(int32_t index) const noexcept = 0;
924 TRT_NODISCARD virtual int32_t getNbOutputs() const noexcept = 0;
925 TRT_NODISCARD virtual ITensor* getOutput(int32_t index) const noexcept = 0;
926 TRT_NODISCARD virtual
bool setName(
char const* name) noexcept = 0;
927 TRT_NODISCARD virtual
char const* getName() const noexcept = 0;
930 TRT_DEPRECATED virtual
bool setCausal(
bool isCausal) noexcept = 0;
932 TRT_NODISCARD virtual
bool setMask(ITensor& mask) noexcept = 0;
934 TRT_NODISCARD virtual
bool setDecomposable(
bool decomposable) noexcept = 0;
935 TRT_NODISCARD virtual
bool getDecomposable() const noexcept = 0;
936 TRT_NODISCARD virtual
bool setNormalizationQuantizeScale(ITensor& tensor) noexcept = 0;
937 TRT_NODISCARD virtual ITensor* getNormalizationQuantizeScale() const noexcept = 0;
940 TRT_NODISCARD virtual
bool setMetadata(
char const* docString) noexcept = 0;
941 TRT_NODISCARD virtual
char const* getMetadata() const noexcept = 0;
942 TRT_NODISCARD virtual
bool setNbRanks(int32_t nbRanks) noexcept = 0;
943 TRT_NODISCARD virtual int32_t getNbRanks() const noexcept = 0;
950 TRT_NODISCARD virtual
bool setQueryLengths(ITensor* lengths) noexcept = 0;
951 TRT_NODISCARD virtual ITensor* getQueryLengths() const noexcept = 0;
952 TRT_NODISCARD virtual
bool setKeyValueLengths(ITensor* lengths) noexcept = 0;
953 TRT_NODISCARD virtual ITensor* getKeyValueLengths() const noexcept = 0;
956class VSelectLayer : public VRoot
960class VAssertionLayer :
public VRoot
963 virtual void setMessage(
char const* message)
noexcept = 0;
964 virtual char const* getMessage() const noexcept = 0;
967class VFillLayer : public VRoot
970 virtual void setDimensions(
Dims const& dimensions)
noexcept = 0;
971 virtual Dims getDimensions() const noexcept = 0;
974 virtual
void setAlpha(
double alpha) noexcept = 0;
975 virtual
double getAlpha() const noexcept = 0;
976 virtual
void setBeta(
double beta) noexcept = 0;
977 virtual
double getBeta() const noexcept = 0;
978 virtual
void setAlphaInt64(int64_t alpha) noexcept = 0;
979 virtual int64_t getAlphaInt64() const noexcept = 0;
980 virtual
void setBetaInt64(int64_t beta) noexcept = 0;
981 virtual int64_t getBetaInt64() const noexcept = 0;
982 virtual
bool isAlphaBetaInt64() const noexcept = 0;
983 virtual
DataType getToType() const noexcept = 0;
984 virtual
void setToType(
DataType toType) noexcept = 0;
987class VQuantizeLayer : public VRoot
990 virtual int32_t getAxis() const noexcept = 0;
991 virtual
void setAxis(int32_t axis) noexcept = 0;
992 virtual
DataType getToType() const noexcept = 0;
993 virtual
void setToType(
DataType toType) noexcept = 0;
994 virtual
Dims getBlockShape() const noexcept = 0;
995 virtual
bool setBlockShape(
Dims const& blockShape) noexcept = 0;
998class VDequantizeLayer : public VRoot
1001 virtual int32_t getAxis() const noexcept = 0;
1002 virtual
void setAxis(int32_t axis) noexcept = 0;
1003 virtual
DataType getToType() const noexcept = 0;
1004 virtual
void setToType(
DataType toType) noexcept = 0;
1005 virtual
Dims getBlockShape() const noexcept = 0;
1006 virtual
bool setBlockShape(
Dims const& blockShape) noexcept = 0;
1009class VDynamicQuantizeLayer : public VRoot
1014 TRT_DEPRECATED virtual int32_t getBlockSize() const noexcept = 0;
1015 TRT_DEPRECATED virtual
void setBlockSize(int32_t axis) noexcept = 0;
1016 virtual
DataType getScaleType() const noexcept = 0;
1017 virtual
void setScaleType(
DataType axis) noexcept = 0;
1018 virtual
DataType getToType() const noexcept = 0;
1019 virtual
void setToType(
DataType toType) noexcept = 0;
1020 virtual
Dims getBlockShape() const noexcept = 0;
1021 virtual
void setBlockShape(
Dims const& blockShape) noexcept = 0;
1024class VScatterLayer : public VRoot
1027 virtual void setMode(
ScatterMode mode)
noexcept = 0;
1029 virtual
void setAxis(int32_t axis) noexcept = 0;
1030 virtual int32_t getAxis() const noexcept = 0;
1033class VEinsumLayer : public VRoot
1036 virtual bool setEquation(
char const* equation)
noexcept = 0;
1037 virtual char const* getEquation() const noexcept = 0;
1040class VOneHotLayer : public VRoot
1043 virtual int32_t getAxis() const noexcept = 0;
1044 virtual
void setAxis(int32_t axis) noexcept = 0;
1047class VGridSampleLayer : public VRoot
1052 virtual
void setAlignCorners(
bool alignCorners) noexcept = 0;
1053 virtual
bool getAlignCorners() const noexcept = 0;
1054 virtual
bool setSampleMode(
SampleMode mode) noexcept = 0;
1055 virtual
SampleMode getSampleMode() const noexcept = 0;
1058class VNMSLayer : public VRoot
1063 virtual
void setTopKBoxLimit(int32_t limit) noexcept = 0;
1064 virtual int32_t getTopKBoxLimit() const noexcept = 0;
1065 virtual
bool setIndicesType(
DataType type) noexcept = 0;
1066 virtual
DataType getIndicesType() const noexcept = 0;
1069class VReverseSequenceLayer : public VRoot
1072 virtual void setBatchAxis(int32_t batchAxis)
noexcept = 0;
1073 virtual int32_t getBatchAxis() const noexcept = 0;
1075 virtual
void setSequenceAxis(int32_t sequenceAxis) noexcept = 0;
1076 virtual int32_t getSequenceAxis() const noexcept = 0;
1079class VNormalizationLayer : public VRoot
1082 virtual void setEpsilon(
float eps)
noexcept = 0;
1083 virtual float getEpsilon() const noexcept = 0;
1084 virtual
void setAxes(uint32_t axesMask) noexcept = 0;
1085 virtual uint32_t getAxes() const noexcept = 0;
1086 virtual
void setNbGroups(int64_t nbGroups) noexcept = 0;
1087 virtual int64_t getNbGroups() const noexcept = 0;
1088 virtual
bool isV2() const noexcept = 0;
1091class VSqueezeLayer : public VRoot
1095class VUnsqueezeLayer :
public VRoot
1099class VCumulativeLayer :
public VRoot
1104 virtual
void setExclusive(
bool exclusive) noexcept = 0;
1105 virtual
bool getExclusive() const noexcept = 0;
1106 virtual
void setReverse(
bool reverse) noexcept = 0;
1107 virtual
bool getReverse() const noexcept = 0;
1110class VRotaryEmbeddingLayer : public VRoot
1113 virtual void setInterleaved(
bool interleaved)
noexcept = 0;
1114 virtual bool getInterleaved() const noexcept = 0;
1115 virtual
bool setRotaryEmbeddingDim(int32_t rotaryEmbeddingDim) noexcept = 0;
1116 virtual int32_t getRotaryEmbeddingDim() const noexcept = 0;
1117 virtual
void setInput(int32_t index, ITensor& input) noexcept = 0;
1120class VKVCacheUpdateLayer : public VRoot
1127 TRT_NODISCARD virtual
bool setUpdateLengths(ITensor* lengths) noexcept = 0;
1128 TRT_NODISCARD virtual ITensor* getUpdateLengths() const noexcept = 0;
1131class VMoELayer : public VRoot
1134 virtual void setGatedWeights(ITensor& fcGateWeights, ITensor& fcUpWeights, ITensor& fcDownWeights,
MoEActType activationType)
noexcept = 0;
1135 virtual void setGatedBiases(ITensor& fcGateBiases, ITensor& fcUpBiases, ITensor& fcDownBiases)
noexcept = 0;
1136 virtual void setActivationType(
MoEActType activationType)
noexcept = 0;
1137 virtual MoEActType getActivationType() const noexcept = 0;
1138 virtual
void setQuantizationStatic(ITensor& fcDownActivationScale,
DataType dataType) noexcept = 0;
1139 virtual
void setQuantizationDynamicDblQ(ITensor& fcDownActivationDblQScale,
DataType dataType,
Dims const& blockShape,
DataType dynQOutputScaleType) noexcept = 0;
1140 virtual
void setQuantizationToType(
DataType type) noexcept = 0;
1141 virtual
DataType getQuantizationToType() const noexcept = 0;
1142 virtual
void setQuantizationBlockShape(
Dims const& blockShape) noexcept = 0;
1143 virtual
Dims getQuantizationBlockShape() const noexcept = 0;
1144 virtual
void setDynQOutputScaleType(
DataType type) noexcept = 0;
1145 virtual
DataType getDynQOutputScaleType() const noexcept = 0;
1146 virtual
void setSwigluParams(
float limit,
float alpha,
float beta) noexcept = 0;
1147 virtual
void setSwigluParamLimit(
float limit) noexcept = 0;
1148 virtual
float getSwigluParamLimit() const noexcept = 0;
1149 virtual
void setSwigluParamAlpha(
float alpha) noexcept = 0;
1150 virtual
float getSwigluParamAlpha() const noexcept = 0;
1151 virtual
void setSwigluParamBeta(
float beta) noexcept = 0;
1152 virtual
float getSwigluParamBeta() const noexcept = 0;
1153 virtual
void setInput(int32_t index, ITensor& tensor) noexcept = 0;
1157class VNetworkDefinition : public VRoot
1160 virtual ITensor* addInput(
char const* name,
DataType type,
Dims const& dimensions)
noexcept = 0;
1161 virtual void markOutput(ITensor& tensor)
noexcept = 0;
1162 virtual IActivationLayer* addActivation(ITensor& input,
ActivationType type)
noexcept = 0;
1163 virtual ILRNLayer* addLRN(ITensor& input, int64_t window,
float alpha,
float beta,
float k)
noexcept = 0;
1164 virtual IScaleLayer* addScale(
1165 ITensor& input,
ScaleMode mode, Weights shift, Weights scale, Weights power)
noexcept = 0;
1166 virtual ISoftMaxLayer* addSoftMax(ITensor& input)
noexcept = 0;
1167 virtual IConcatenationLayer* addConcatenation(ITensor*
const* inputs, int32_t nbInputs)
noexcept = 0;
1168 virtual IElementWiseLayer* addElementWise(ITensor& input1, ITensor& input2,
ElementWiseOperation op)
noexcept = 0;
1169 virtual IUnaryLayer* addUnary(ITensor& input,
UnaryOperation operation)
noexcept = 0;
1170 virtual IShuffleLayer* addShuffle(ITensor& input)
noexcept = 0;
1171 virtual int32_t getNbLayers() const noexcept = 0;
1172 virtual ILayer* getLayer(int32_t index) const noexcept = 0;
1173 virtual int32_t getNbInputs() const noexcept = 0;
1174 virtual ITensor* getInput(int32_t index) const noexcept = 0;
1175 virtual int32_t getNbOutputs() const noexcept = 0;
1176 virtual ITensor* getOutput(int32_t index) const noexcept = 0;
1177 virtual IReduceLayer* addReduce(
1178 ITensor& input,
ReduceOperation operation, uint32_t reduceAxes,
bool keepDimensions) noexcept
1180 virtual ITopKLayer* addTopK(ITensor& input,
TopKOperation op, int32_t k, uint32_t reduceAxes) noexcept = 0;
1181 virtual IGatherLayer* addGather(ITensor& data, ITensor& indices, int32_t axis) noexcept = 0;
1182 virtual IRaggedSoftMaxLayer* addRaggedSoftMax(ITensor& input, ITensor& bounds) noexcept = 0;
1183 virtual IMatrixMultiplyLayer* addMatrixMultiply(
1185 virtual IConstantLayer* addConstant(
Dims const& dimensions, Weights weights) noexcept = 0;
1186 virtual IIdentityLayer* addIdentity(ITensor& input) noexcept = 0;
1187 virtual
void removeTensor(ITensor& tensor) noexcept = 0;
1188 virtual
void unmarkOutput(ITensor& tensor) noexcept = 0;
1189 virtual IPluginV2Layer* addPluginV2(ITensor* const* inputs, int32_t nbInputs, IPluginV2& plugin) noexcept = 0;
1190 virtual IPluginV3Layer* addPluginV3(ITensor* const* inputs, int32_t nbInputs, ITensor* const* shapeInputs,
1191 int32_t nbShapeInputs,
IPluginV3& plugin) noexcept = 0;
1192 virtual ISliceLayer* addSlice(ITensor& input,
Dims const& start,
Dims const& size,
Dims const& stride) noexcept = 0;
1193 virtual
void setName(
char const* name) noexcept = 0;
1194 virtual
char const* getName() const noexcept = 0;
1195 virtual IShapeLayer* addShape(ITensor& input) noexcept = 0;
1196 virtual
bool hasImplicitBatchDimension() const noexcept = 0;
1197 virtual
bool markOutputForShapes(ITensor& tensor) noexcept = 0;
1198 virtual
bool unmarkOutputForShapes(ITensor& tensor) noexcept = 0;
1199 virtual IParametricReLULayer* addParametricReLU(ITensor& input, ITensor& slope) noexcept = 0;
1200 virtual IConvolutionLayer* addConvolutionNd(
1201 ITensor& input, int64_t nbOutputMaps,
Dims const& kernelSize, Weights kernelWeights, Weights biasWeights) noexcept
1203 virtual IPoolingLayer* addPoolingNd(ITensor& input,
PoolingType type,
Dims const& windowSize) noexcept = 0;
1204 virtual IDeconvolutionLayer* addDeconvolutionNd(
1205 ITensor& input, int64_t nbOutputMaps,
Dims const& kernelSize, Weights kernelWeights, Weights biasWeights) noexcept
1207 virtual IScaleLayer* addScaleNd(
1208 ITensor& input,
ScaleMode mode, Weights shift, Weights scale, Weights power, int32_t channelAxis) noexcept = 0;
1209 virtual IResizeLayer* addResize(ITensor& input) noexcept = 0;
1210 virtual ILoop* addLoop() noexcept = 0;
1211 virtual ISelectLayer* addSelect(ITensor& condition, ITensor& thenInput, ITensor& elseInput) noexcept = 0;
1212 virtual IPaddingLayer* addPaddingNd(ITensor& input,
Dims const& prePadding,
Dims const& postPadding) noexcept = 0;
1213 virtual
bool setWeightsName(Weights weights,
char const* name) noexcept = 0;
1214 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
1216 virtual IGatherLayer* addGatherV2(ITensor& data, ITensor& indices,
GatherMode mode) noexcept = 0;
1217 virtual IIfConditional* addIfConditional() noexcept = 0;
1218 virtual IScatterLayer* addScatter(ITensor& data, ITensor& indices, ITensor& updates,
ScatterMode mode) noexcept = 0;
1219 virtual IEinsumLayer* addEinsum(ITensor* const* inputs, int32_t nbInputs,
char const* equation) noexcept = 0;
1220 virtual IAssertionLayer* addAssertion(ITensor& condition,
char const* message) noexcept = 0;
1221 virtual IOneHotLayer* addOneHot(ITensor& indices, ITensor& values, ITensor& depth, int32_t axis) noexcept = 0;
1222 virtual INonZeroLayer* addNonZero(ITensor& input) noexcept = 0;
1223 virtual IGridSampleLayer* addGridSample(ITensor& input, ITensor& grid) noexcept = 0;
1224 virtual INMSLayer* addNMS(ITensor& boxes, ITensor& scores, ITensor& maxOutputBoxesPerClass) noexcept = 0;
1225 virtual IReverseSequenceLayer* addReverseSequence(ITensor& input, ITensor& sequenceLens) noexcept = 0;
1226 virtual INormalizationLayer* addNormalization(
1227 ITensor& input, ITensor& scale, ITensor& bias, uint32_t axesMask) noexcept = 0;
1228 virtual ICastLayer* addCast(ITensor& input,
DataType toType) noexcept = 0;
1229 virtual IBuilder& getBuilder() const noexcept = 0;
1232 virtual IQuantizeLayer* addQuantizeV2(ITensor& input, ITensor& scale,
DataType outputType) noexcept = 0;
1233 virtual IDequantizeLayer* addDequantizeV2(ITensor& input, ITensor& scale,
DataType outputType) noexcept = 0;
1235 virtual
bool markDebug(ITensor& tensor) noexcept = 0;
1236 virtual
bool unmarkDebug(ITensor& tensor) noexcept = 0;
1237 virtual
bool isDebugTensor(ITensor const& tensor) const noexcept = 0;
1238 virtual
bool markWeightsRefittable(
char const* name) noexcept = 0;
1239 virtual
bool unmarkWeightsRefittable(
char const* name) noexcept = 0;
1240 virtual
bool areWeightsMarkedRefittable(
char const* name) const noexcept = 0;
1241 virtual ISqueezeLayer* addSqueeze(ITensor& input, ITensor& axes) noexcept = 0;
1242 virtual IUnsqueezeLayer* addUnsqueeze(ITensor& input, ITensor& axes) noexcept = 0;
1243 virtual IDynamicQuantizeLayer* addDynamicQuantize(
1244 ITensor& input, int32_t axis, int32_t blockSize,
DataType toType,
DataType scaleType) noexcept = 0;
1245 virtual ICumulativeLayer* addCumulative(
1246 ITensor& input, ITensor& axis,
CumulativeOperation operation,
bool exclusive,
bool reverse) noexcept = 0;
1247 virtual
bool markUnfusedTensorsAsDebugTensors() noexcept = 0;
1248 virtual
bool unmarkUnfusedTensorsAsDebugTensors() noexcept = 0;
1249 virtual ITopKLayer* addTopKV2(
1250 ITensor& input,
TopKOperation op, int32_t k, uint32_t reduceAxes,
DataType indicesType) noexcept = 0;
1251 virtual INonZeroLayer* addNonZeroV2(ITensor& input,
DataType indicesType) noexcept = 0;
1252 virtual INMSLayer* addNMSV2(
1253 ITensor& boxes, ITensor& scores, ITensor& maxOutputBoxesPerClass,
DataType indicesType) noexcept = 0;
1256 virtual IRotaryEmbeddingLayer* addRotaryEmbedding(ITensor& input, ITensor& cosCache, ITensor& sinCache,
1257 bool interleaved, int32_t rotaryEmbeddingDim) noexcept = 0;
1258 virtual IDynamicQuantizeLayer* addDynamicQuantizeV2(
1260 virtual IKVCacheUpdateLayer* addKVCacheUpdate(
1261 ITensor& cache, ITensor& update, ITensor& writeIndices,
KVCacheMode cacheMode) noexcept = 0;
1262 virtual INormalizationLayer* addNormalizationV2(
1263 ITensor& input, ITensor& scale, ITensor& bias, uint32_t axesMask) noexcept = 0;
1264 virtual IMoELayer* addMoE(
1265 ITensor& hiddenStates, ITensor& selectedExpertsForTokens, ITensor& scoresForSelectedExperts) noexcept = 0;
1266 virtual IDistCollectiveLayer* addDistCollective(ITensor& input,
CollectiveOperation distCollectiveOp,
1267 ReduceOperation reduceOp, int64_t root, int64_t* groups, int64_t groupSize) noexcept = 0;
1272class VTimingCache : public VRoot
1276 virtual
bool combine(ITimingCache const& inputCache,
bool ignoreMismatch) noexcept = 0;
1277 virtual
bool reset() noexcept = 0;
1278 virtual int64_t queryKeys(TimingCacheKey* keyBuffer, int64_t capacity) const noexcept = 0;
1279 virtual TimingCacheValue query(TimingCacheKey const& key) const noexcept = 0;
1280 virtual
bool update(TimingCacheKey const& key, TimingCacheValue const& value) noexcept = 0;
1284class VBuilderConfig : public VRoot
1287 virtual void setAvgTimingIterations(int32_t avgTiming)
noexcept = 0;
1288 virtual int32_t getAvgTimingIterations() const noexcept = 0;
1289 virtual
void setEngineCapability(
EngineCapability capability) noexcept = 0;
1291 virtual
void setFlags(
BuilderFlags builderFlags) noexcept = 0;
1293 virtual
void clearFlag(
BuilderFlag builderFlag) noexcept = 0;
1294 virtual
void setFlag(
BuilderFlag builderFlag) noexcept = 0;
1295 virtual
bool getFlag(
BuilderFlag builderFlag) const noexcept = 0;
1296 virtual
void setDeviceType(ILayer const* layer,
DeviceType deviceType) noexcept = 0;
1297 virtual
DeviceType getDeviceType(ILayer const* layer) const noexcept = 0;
1298 virtual
bool isDeviceTypeSet(ILayer const* layer) const noexcept = 0;
1299 virtual
void resetDeviceType(ILayer const* layer) noexcept = 0;
1300 virtual
bool canRunOnDLA(ILayer const* layer) const noexcept = 0;
1301 virtual
void setDLACore(int32_t dlaCore) noexcept = 0;
1302 virtual int32_t getDLACore() const noexcept = 0;
1303 virtual
void setDefaultDeviceType(
DeviceType deviceType) noexcept = 0;
1304 virtual
DeviceType getDefaultDeviceType() const noexcept = 0;
1305 virtual
void reset() noexcept = 0;
1306 virtual
void setProfileStream(const cudaStream_t stream) noexcept = 0;
1307 virtual cudaStream_t getProfileStream() const noexcept = 0;
1308 virtual int32_t addOptimizationProfile(IOptimizationProfile const* profile) noexcept = 0;
1309 virtual int32_t getNbOptimizationProfiles() const noexcept = 0;
1312 virtual
bool setTacticSources(
TacticSources tacticSources) noexcept = 0;
1313 virtual
TacticSources getTacticSources() const noexcept = 0;
1314 virtual
nvinfer1::ITimingCache* createTimingCache(
void const* blob, std::
size_t size) const noexcept = 0;
1315 virtual
bool setTimingCache(ITimingCache const& cache,
bool ignoreMismatch) noexcept = 0;
1316 virtual
nvinfer1::ITimingCache const* getTimingCache() const noexcept = 0;
1317 virtual
void setMemoryPoolLimit(
MemoryPoolType pool, std::
size_t poolSize) noexcept = 0;
1318 virtual std::
size_t getMemoryPoolLimit(
MemoryPoolType pool) const noexcept = 0;
1319 virtual
void setPreviewFeature(
PreviewFeature feature,
bool enable) noexcept = 0;
1320 virtual
bool getPreviewFeature(
PreviewFeature feature) const noexcept = 0;
1321 virtual
void setBuilderOptimizationLevel(int32_t level) noexcept = 0;
1322 virtual int32_t getBuilderOptimizationLevel() const noexcept = 0;
1325 virtual
void setPluginsToSerialize(
char const* const* paths, int32_t nbPaths) noexcept = 0;
1326 virtual
char const* getPluginToSerialize(int32_t index) const noexcept = 0;
1327 virtual int32_t getNbPluginsToSerialize() const noexcept = 0;
1328 virtual
bool setMaxAuxStreams(int32_t nbStreams) noexcept = 0;
1329 virtual int32_t getMaxAuxStreams() const noexcept = 0;
1332 virtual
void setRuntimePlatform(
RuntimePlatform runtimePlatform) noexcept = 0;
1334 virtual
void setMaxNbTactics(int32_t maxTactics) noexcept = 0;
1335 virtual int32_t getMaxNbTactics() const noexcept = 0;
1338 virtual
bool setL2LimitForTiling(int64_t size) noexcept = 0;
1339 virtual int64_t getL2LimitForTiling() const noexcept = 0;
1340 virtual
bool setRemoteAutoTuningConfig(
char const* config) noexcept = 0;
1341 virtual
char const* getRemoteAutoTuningConfig() const noexcept = 0;
1342 virtual
bool setBuildRoute(
char const* buildRoute) noexcept = 0;
1343 virtual
char const* getBuildRoute() const noexcept = 0;
1344 virtual
char const* getAllBuildRoutes() const noexcept = 0;
1347class VSerializationConfig : public VRoot
1357class VBuilder : public VRoot
1360 virtual int32_t getMaxDLABatchSize() const noexcept = 0;
1361 virtual int32_t getNbDLACores() const noexcept = 0;
1362 virtual
void setGpuAllocator(
IGpuAllocator* allocator) noexcept = 0;
1363 virtual
nvinfer1::IBuilderConfig* createBuilderConfig() noexcept = 0;
1365 virtual
nvinfer1::IOptimizationProfile* createOptimizationProfile() noexcept = 0;
1366 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
1368 virtual
void reset() noexcept = 0;
1369 virtual
nvinfer1::IHostMemory* buildSerializedNetwork(
1370 INetworkDefinition& network, IBuilderConfig& config) noexcept = 0;
1371 virtual
bool isNetworkSupported(INetworkDefinition const& network, IBuilderConfig const& config) const noexcept = 0;
1373 virtual
bool setMaxThreads(int32_t maxThreads) noexcept = 0;
1374 virtual int32_t getMaxThreads() const noexcept = 0;
1376 virtual ICudaEngine* buildEngineWithConfig(INetworkDefinition& network, IBuilderConfig& config) noexcept = 0;
1377 virtual
bool buildSerializedNetworkToStream(
1378 INetworkDefinition& network, IBuilderConfig& config,
IStreamWriter& writer) noexcept = 0;
1379 virtual
nvinfer1::IHostMemory* buildSerializedNetworkWithKernelText(
1380 INetworkDefinition& network, IBuilderConfig& config, IHostMemory*& kernelText) noexcept
1384class VRuntimeConfig : public VRoot
1387 virtual IRuntimeConfig* getPImpl() noexcept = 0;
1393class 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: NvInferSafePlugin.h:33
uint32_t TacticSources
Represents a collection of one or more TacticSource values combine using bitwise-OR operations.
Definition: NvInferRuntime.h:2855
v_1_0::IOutputAllocator IOutputAllocator
Definition: NvInferRuntime.h:3979
ResizeSelector
The coordinate selector when resize to single pixel output.
Definition: NvInfer.h:3979
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:10299
AttentionIOForm
Enumerates the layout of the input/output tensors in an Attention layer.
Definition: NvInfer.h:6790
ScaleMode
Controls how shift, scale and power are applied in a Scale layer.
Definition: NvInfer.h:1649
RuntimePlatform
Describes the intended runtime platform (operating system and CPU architecture) for the execution of ...
Definition: NvInfer.h:9923
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:10407
CumulativeOperation
Enumerates the cumulative operations that may be performed by a Cumulative layer.
Definition: NvInfer.h:6598
SerializationFlag
List of valid flags that the engine can enable when serializing the bytes.
Definition: NvInferRuntime.h:2896
BoundingBoxFormat
Representation of bounding box data used for the Boxes input tensor in INMSLayer.
Definition: NvInfer.h:6165
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:2607
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
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:4982
EngineStat
The kind of engine statistics that queried from the ICudaEngine.
Definition: NvInferRuntime.h:3065
ResizeRoundMode
The rounding mode for nearest neighbor resize.
Definition: NvInfer.h:4006
v_1_0::ILogger ILogger
Definition: NvInferRuntimeBase.h:125
CausalMaskKind
Enumerates the causal mask alignment orientation for the attention.
Definition: NvInfer.h:6762
PaddingMode
Enumerates the modes of padding to perform in convolution, deconvolution and pooling layer,...
Definition: NvInfer.h:826
TripLimit
Enum that describes kinds of trip limits.
Definition: NvInfer.h:4364
uint32_t NetworkDefinitionCreationFlags
Represents one or more NetworkDefinitionCreationFlag flags using binary OR operations....
Definition: NvInfer.h:11463
PreviewFeature
Define preview features.
Definition: NvInfer.h:10373
TilingOptimizationLevel
Define the optimization levels for Tiling.
Definition: NvInfer.h:10457
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:9953
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:3109
GatherMode
Control form of IGatherLayer.
Definition: NvInfer.h:2348
v_1_0::IDebugListener IDebugListener
Definition: NvInferRuntime.h:4020
MoEActType
Enumerates the activation type for the MoE layer.
Definition: NvInfer.h:7679
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:2867
NetworkDefinitionCreationFlag
List of immutable network properties expressed at network creation time. NetworkDefinitionCreationFla...
Definition: NvInfer.h:11474
ElementWiseOperation
Enumerates the binary operations that may be performed by an ElementWise layer.
Definition: NvInfer.h:2259
CollectiveOperation
Enumerates the collective operations that may be performed by a DistCollective layer.
Definition: NvInfer.h:2737
uint32_t SerializationFlags
Represents one or more SerializationFlag values using binary OR operations, e.g., 1U << Serialization...
Definition: NvInferRuntime.h:2886
InterpolationMode
Enumerates various modes of interpolation.
Definition: NvInfer.h:3903
BuilderFlag
List of valid modes that the builder can enable when creating an engine from a network definition.
Definition: NvInfer.h:9963
TensorFormat
Format of the input/output tensors.
Definition: NvInferRuntime.h:1432
ExecutionContextAllocationStrategy
Different memory allocation behaviors for IExecutionContext.
Definition: NvInferRuntime.h:3003
TopKOperation
Enumerates the operations that may be performed by a TopK layer.
Definition: NvInfer.h:3394
ReduceOperation
Enumerates the reduce operations that may be performed by a Reduce layer.
Definition: NvInfer.h:2709
LayerInformationFormat
The format in which the IEngineInspector prints the layer information.
Definition: NvInferRuntime.h:4954
ScatterMode
Control form of IScatterLayer.
Definition: NvInfer.h:5892
MatrixOperation
Enumerates the operations that may be performed on a tensor by IMatrixMultiplyLayer before multiplica...
Definition: NvInfer.h:3549
ResizeCoordinateTransformation
The resize coordinate transformation function.
Definition: NvInfer.h:3928
LoopOutput
Enum that describes kinds of loop outputs.
Definition: NvInfer.h:4336
KVCacheMode
Enumerates the KVCache modes that may be performed by a KVCacheUpdate layer.
Definition: NvInfer.h:7526
v_1_0::IStreamReader IStreamReader
Definition: NvInferRuntime.h:700
PoolingType
The type of pooling to perform in a pooling layer.
Definition: NvInfer.h:1263
v_1_0::IProgressMonitor IProgressMonitor
Definition: NvInfer.h:10570
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:2614
AttentionNormalizationOp
Enumerates the operations that may be performed by the normalization in the attention subgraph.
Definition: NvInfer.h:6730