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;
132class IRaggedSoftMaxLayer;
133class IRecurrenceLayer;
137class IReverseSequenceLayer;
142class ISerializationConfig;
152struct TimingCacheKey;
153struct TimingCacheValue;
155using TimingCacheKey = v_1_0::TimingCacheKey;
156using TimingCacheValue = v_1_0::TimingCacheValue;
160class ITripLimitLayer;
162class IUnsqueezeLayer;
240 virtual ~VRoot() noexcept = default;
243class VHostMemory : public VRoot
246 virtual void* data() const noexcept = 0;
247 virtual std::
size_t size() const noexcept = 0;
248 virtual
DataType type() const noexcept = 0;
251class VDimensionExpr : public VRoot
254 virtual bool isConstant()
const = 0;
255 virtual int64_t getConstantValue()
const = 0;
256 virtual bool isSizeTensor()
const = 0;
259class VExprBuilder :
public VRoot
262 virtual IDimensionExpr
const* constant(int64_t value) = 0;
263 virtual IDimensionExpr
const* operation(
266 virtual IDimensionExpr
const* declareSizeTensor(
267 int32_t outputIndex, IDimensionExpr
const& opt, IDimensionExpr
const& upper)
271class VRuntime :
public VRoot
274 virtual IRuntime* getPImpl() noexcept = 0;
275 virtual
nvinfer1::ICudaEngine* deserializeCudaEngine(
void const* blob, std::
size_t size) noexcept = 0;
276 virtual
void setDLACore(int32_t dlaCore) noexcept = 0;
277 virtual int32_t getDLACore() const noexcept = 0;
278 virtual int32_t getNbDLACores() const noexcept = 0;
279 virtual
void setGpuAllocator(
IGpuAllocator* allocator) noexcept = 0;
280 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
283 virtual
bool setMaxThreads(int32_t maxThreads) noexcept = 0;
284 virtual int32_t getMaxThreads() const noexcept = 0;
285 virtual
void setTemporaryDirectory(
char const*) noexcept = 0;
286 virtual
char const* getTemporaryDirectory() const noexcept = 0;
290 virtual
void setPluginRegistryParent(IPluginRegistry* parent) noexcept = 0;
291 virtual IRuntime* loadRuntime(
char const* path) noexcept = 0;
292 virtual
void setEngineHostCodeAllowed(
bool allowed) noexcept = 0;
293 virtual
bool getEngineHostCodeAllowed() const noexcept = 0;
296 virtual int64_t getEngineHeaderSize() const noexcept = 0;
297 virtual
EngineValidity getEngineValidity(
void const* blob, int64_t blobSize, uint64_t* diagnostics) const noexcept = 0;
300class VRefitter : public VRoot
303 virtual IRefitter* getPImpl() noexcept = 0;
304 virtual
bool setWeights(
char const* layerName,
WeightsRole role, const Weights weights) noexcept = 0;
305 virtual
bool refitCudaEngine() noexcept = 0;
306 virtual int32_t getMissing(int32_t size,
char const** layerNames,
WeightsRole* roles) noexcept = 0;
307 virtual int32_t getAll(int32_t size,
char const** layerNames,
WeightsRole* roles) noexcept = 0;
308 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
310 virtual
bool setNamedWeights(
char const* name, Weights weights) noexcept = 0;
311 virtual int32_t getMissingWeights(int32_t size,
char const** weightsNames) noexcept = 0;
312 virtual int32_t getAllWeights(int32_t size,
char const** weightsNames) noexcept = 0;
314 virtual
bool setMaxThreads(int32_t maxThreads) noexcept = 0;
315 virtual int32_t getMaxThreads() const noexcept = 0;
316 virtual
bool setNamedWeightsWithLocation(
char const* name, Weights weights,
TensorLocation location) noexcept = 0;
317 virtual Weights getNamedWeights(
char const* weightsName) const noexcept = 0;
318 virtual
TensorLocation getWeightsLocation(
char const* weightsName) const noexcept = 0;
319 virtual
bool unsetNamedWeights(
char const* weightsName) noexcept = 0;
320 virtual
void setWeightsValidation(
bool weightsValidation) noexcept = 0;
321 virtual
bool getWeightsValidation() const noexcept = 0;
322 virtual
bool refitCudaEngineAsync(cudaStream_t stream) noexcept = 0;
323 virtual Weights getWeightsPrototype(
char const* weightsName) const noexcept = 0;
326class VOptimizationProfile : public VRoot
329 virtual bool setDimensions(
char const* inputName,
OptProfileSelector select,
Dims const& dims)
noexcept = 0;
331 virtual int32_t getNbShapeValues(
char const* inputName)
const noexcept = 0;
332 virtual bool setExtraMemoryTarget(
float target)
noexcept = 0;
333 virtual float getExtraMemoryTarget() const noexcept = 0;
334 virtual
bool isValid() const noexcept = 0;
337 char const* inputName,
OptProfileSelector select, int64_t const* values, int32_t nbValues) noexcept = 0;
342class VCudaEngine : public VRoot
345 virtual ICudaEngine* getPImpl() noexcept = 0;
346 virtual int32_t getNbLayers() const noexcept = 0;
347 virtual IHostMemory* serialize() const noexcept = 0;
349 virtual
bool isRefittable() const noexcept = 0;
350 virtual
char const* getName() const noexcept = 0;
351 virtual int32_t getNbOptimizationProfiles() const noexcept = 0;
353 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
357 virtual IEngineInspector* createEngineInspector() const noexcept = 0;
358 virtual
Dims getTensorShape(
char const* tensorName) const noexcept = 0;
359 virtual
DataType getTensorDataType(
char const* tensorName) const noexcept = 0;
360 virtual
TensorLocation getTensorLocation(
char const* tensorName) const noexcept = 0;
361 virtual
bool isShapeInferenceIO(
char const* tensorName) const noexcept = 0;
362 virtual
TensorIOMode getTensorIOMode(
char const* tensorName) const noexcept = 0;
363 virtual int32_t getTensorBytesPerComponent(
char const* tensorName) const noexcept = 0;
364 virtual int32_t getTensorComponentsPerElement(
char const* tensorName) const noexcept = 0;
365 virtual
TensorFormat getTensorFormat(
char const* tensorName) const noexcept = 0;
366 virtual
char const* getTensorFormatDesc(
char const* tensorName) const noexcept = 0;
367 virtual int32_t getTensorVectorizedDim(
char const* tensorName) const noexcept = 0;
368 virtual
Dims getProfileShape(
369 char const* tensorName, int32_t profileIndex,
OptProfileSelector select) const noexcept = 0;
370 virtual int32_t getNbIOTensors() const noexcept = 0;
371 virtual
char const* getIOTensorName(int32_t index) const noexcept = 0;
373 virtual int32_t getNbAuxStreams() const noexcept = 0;
375 virtual int32_t getTensorBytesPerComponentV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
376 virtual int32_t getTensorComponentsPerElementV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
377 virtual
TensorFormat getTensorFormatV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
378 virtual
char const* getTensorFormatDescV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
379 virtual int32_t getTensorVectorizedDimV2(
char const* tensorName, int32_t profileIndex) const noexcept = 0;
381 virtual ISerializationConfig* createSerializationConfig() noexcept = 0;
382 virtual IHostMemory* serializeWithConfig(ISerializationConfig& config) const noexcept = 0;
384 virtual IRefitter* createRefitter(
ILogger& logger) noexcept = 0;
386 virtual int64_t getStreamableWeightsSize() const noexcept = 0;
388 virtual
bool isDebugTensor(
char const* name) const noexcept = 0;
391 virtual
bool setWeightStreamingBudgetV2(int64_t gpuMemoryBudget) noexcept = 0;
392 virtual int64_t getWeightStreamingBudgetV2() const noexcept = 0;
393 virtual int64_t getWeightStreamingAutomaticBudget() const noexcept = 0;
394 virtual int64_t getWeightStreamingScratchMemorySize() const noexcept = 0;
395 virtual int64_t getDeviceMemorySizeV2() const noexcept = 0;
396 virtual int64_t getDeviceMemorySizeForProfileV2(int32_t profileIndex) const noexcept = 0;
398 TRT_NODISCARD virtual int64_t const* getProfileTensorValuesV2(
399 char const* tensorName, int32_t profileIndex,
OptProfileSelector select) const noexcept = 0;
400 TRT_NODISCARD virtual IExecutionContext* createExecutionContextWithRuntimeConfig(
401 IRuntimeConfig* runtimeConfig) noexcept = 0;
402 TRT_NODISCARD virtual IRuntimeConfig* createRuntimeConfig() noexcept = 0;
405 TRT_NODISCARD virtual
char const* getAliasedInputTensor(
char const* tensorName) const noexcept = 0;
408class VExecutionContext : public VRoot
411 virtual IExecutionContext* getPImpl() noexcept = 0;
412 virtual
void setDebugSync(
bool sync) noexcept = 0;
413 virtual
bool getDebugSync() const noexcept = 0;
414 virtual
void setProfiler(
IProfiler*) noexcept = 0;
415 virtual
IProfiler* getProfiler() const noexcept = 0;
416 virtual ICudaEngine const& getEngine() const noexcept = 0;
417 virtual
void setName(
char const* name) noexcept = 0;
418 virtual
char const* getName() const noexcept = 0;
419 virtual
void setDeviceMemory(
void* memory) noexcept = 0;
420 virtual int32_t getOptimizationProfile() const noexcept = 0;
421 virtual
bool allInputDimensionsSpecified() const noexcept = 0;
422 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
424 virtual
bool executeV2(
void* const* bindings) noexcept = 0;
425 virtual
bool setOptimizationProfileAsync(int32_t profileIndex, cudaStream_t stream) noexcept = 0;
426 virtual
void setEnqueueEmitsProfile(
bool enqueueEmitsProfile) noexcept = 0;
427 virtual
bool getEnqueueEmitsProfile() const noexcept = 0;
428 virtual
bool reportToProfiler() const noexcept = 0;
429 virtual
bool setInputShape(
char const* tensorName,
Dims const& dims) noexcept = 0;
430 virtual
Dims getTensorShape(
char const* tensorName) const noexcept = 0;
431 virtual
Dims getTensorStrides(
char const* tensorName) const noexcept = 0;
432 virtual
bool setTensorAddress(
char const* tensorName,
void* data) noexcept = 0;
433 virtual
void const* getTensorAddress(
char const* tensorName) const noexcept = 0;
434 virtual
bool setInputTensorAddress(
char const* tensorName,
void const* data) noexcept = 0;
435 virtual
bool setOutputTensorAddress(
char const* tensorName,
void* data) noexcept = 0;
436 virtual int32_t inferShapes(int32_t nbMaxNames,
char const** tensorNames) noexcept = 0;
437 virtual
bool setInputConsumedEvent(cudaEvent_t event) noexcept = 0;
438 virtual cudaEvent_t getInputConsumedEvent() const noexcept = 0;
439 virtual
void* getOutputTensorAddress(
char const* tensorName) const noexcept = 0;
440 virtual
bool setOutputAllocator(
char const* tensorName,
IOutputAllocator* outputAllocator) noexcept = 0;
441 virtual
IOutputAllocator* getOutputAllocator(
char const* name) noexcept = 0;
442 virtual int64_t getMaxOutputSize(
char const* tensorName) const noexcept = 0;
443 virtual
bool setTemporaryStorageAllocator(
IGpuAllocator* allocator) noexcept = 0;
444 virtual
IGpuAllocator* getTemporaryStorageAllocator() const noexcept = 0;
445 virtual
bool enqueueV3(cudaStream_t stream) noexcept = 0;
446 virtual
void setPersistentCacheLimit(
size_t size) noexcept = 0;
447 virtual
size_t getPersistentCacheLimit() const noexcept = 0;
450 virtual
void setAuxStreams(cudaStream_t* auxStreams, int32_t nbStreams) noexcept = 0;
451 virtual
bool setDebugListener(
IDebugListener* listener) noexcept = 0;
453 virtual
bool setTensorDebugState(
char const* name,
bool flag) noexcept = 0;
454 virtual
bool getDebugState(
char const* name) const noexcept = 0;
455 virtual
bool setAllTensorsDebugState(
bool flag) noexcept = 0;
456 virtual
size_t updateDeviceMemorySizeForShapes() noexcept = 0;
457 virtual
void setDeviceMemoryV2(
void* memory, int64_t size) noexcept = 0;
458 TRT_NODISCARD virtual IRuntimeConfig* getRuntimeConfig() const noexcept = 0;
459 virtual
bool setUnfusedTensorsDebugState(
bool flag) noexcept = 0;
460 virtual
bool getUnfusedTensorsDebugState() const noexcept = 0;
461 virtual
bool isStreamCapturable(cudaStream_t stream) const noexcept = 0;
462 virtual
bool setCommunicator(
void* communicator) noexcept = 0;
465class VEngineInspector : public VRoot
468 virtual IEngineInspector* getPImpl() noexcept = 0;
469 virtual
bool setExecutionContext(IExecutionContext const* context) noexcept = 0;
470 virtual IExecutionContext const* getExecutionContext() const noexcept = 0;
471 virtual
char const* getLayerInformation(int32_t layerIndex,
LayerInformationFormat format) const noexcept = 0;
473 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
477class VTensor : public VRoot
480 virtual void setName(
char const* name)
noexcept = 0;
481 virtual char const* getName() const noexcept = 0;
482 virtual
void setDimensions(
Dims const& dimensions) noexcept = 0;
483 virtual
Dims getDimensions() const noexcept = 0;
484 virtual
DataType getType() const noexcept = 0;
485 virtual
bool isNetworkInput() const noexcept = 0;
486 virtual
bool isNetworkOutput() const noexcept = 0;
487 virtual
void setAllowedFormats(
TensorFormats formats) noexcept = 0;
488 virtual
TensorFormats getAllowedFormats() const noexcept = 0;
489 virtual
bool isShapeTensor() const noexcept = 0;
490 virtual
bool isExecutionTensor() const noexcept = 0;
491 virtual
void setDimensionName(int32_t index,
char const* name) noexcept = 0;
492 virtual
char const* getDimensionName(int32_t index) const noexcept = 0;
495class VLayer : public VRoot
498 virtual LayerType getType() const noexcept = 0;
499 virtual
void setName(
char const* name) noexcept = 0;
500 virtual
char const* getName() const noexcept = 0;
501 virtual int32_t getNbInputs() const noexcept = 0;
502 virtual ITensor* getInput(int32_t index) const noexcept = 0;
503 virtual int32_t getNbOutputs() const noexcept = 0;
504 virtual ITensor* getOutput(int32_t index) const noexcept = 0;
505 virtual
void setInput(int32_t index, ITensor& tensor) noexcept = 0;
506 virtual
DataType getOutputType(int32_t index) const noexcept = 0;
507 virtual
void setMetadata(
char const* docString) noexcept = 0;
508 virtual
char const* getMetadata() const noexcept = 0;
509 virtual
bool setNbRanks(int32_t nbRanks) noexcept = 0;
510 virtual int32_t getNbRanks() const noexcept = 0;
513class VConvolutionLayer : public VRoot
516 virtual void setNbOutputMaps(int64_t nbOutputMaps)
noexcept = 0;
517 virtual int64_t getNbOutputMaps() const noexcept = 0;
518 virtual
void setNbGroups(int64_t nbGroups) noexcept = 0;
519 virtual int64_t getNbGroups() const noexcept = 0;
520 virtual
void setKernelWeights(Weights weights) noexcept = 0;
521 virtual Weights getKernelWeights() const noexcept = 0;
522 virtual
void setBiasWeights(Weights weights) noexcept = 0;
523 virtual Weights getBiasWeights() const noexcept = 0;
524 virtual
void setPrePadding(
Dims const& padding) noexcept = 0;
525 virtual
Dims getPrePadding() const noexcept = 0;
526 virtual
void setPostPadding(
Dims const& padding) noexcept = 0;
527 virtual
Dims getPostPadding() const noexcept = 0;
528 virtual
void setPaddingMode(
PaddingMode paddingMode) noexcept = 0;
529 virtual
PaddingMode getPaddingMode() const noexcept = 0;
530 virtual
void setKernelSizeNd(
Dims const& kernelSize) noexcept = 0;
531 virtual
Dims getKernelSizeNd() const noexcept = 0;
532 virtual
void setStrideNd(
Dims const& stride) noexcept = 0;
533 virtual
Dims getStrideNd() const noexcept = 0;
534 virtual
void setPaddingNd(
Dims const& padding) noexcept = 0;
535 virtual
Dims getPaddingNd() const noexcept = 0;
536 virtual
void setDilationNd(
Dims const& dilation) noexcept = 0;
537 virtual
Dims getDilationNd() const noexcept = 0;
540class VActivationLayer : public VRoot
545 virtual
void setAlpha(
float alpha) noexcept = 0;
546 virtual
void setBeta(
float beta) noexcept = 0;
547 virtual
float getAlpha() const noexcept = 0;
548 virtual
float getBeta() const noexcept = 0;
551class VPoolingLayer : public VRoot
554 virtual void setPoolingType(
PoolingType type)
noexcept = 0;
555 virtual PoolingType getPoolingType() const noexcept = 0;
556 virtual
void setBlendFactor(
float blendFactor) noexcept = 0;
557 virtual
float getBlendFactor() const noexcept = 0;
558 virtual
void setAverageCountExcludesPadding(
bool exclusive) noexcept = 0;
559 virtual
bool getAverageCountExcludesPadding() const noexcept = 0;
560 virtual
void setPrePadding(
Dims const& padding) noexcept = 0;
561 virtual
Dims getPrePadding() const noexcept = 0;
562 virtual
void setPostPadding(
Dims const& padding) noexcept = 0;
563 virtual
Dims getPostPadding() const noexcept = 0;
564 virtual
void setPaddingMode(
PaddingMode paddingMode) noexcept = 0;
565 virtual
PaddingMode getPaddingMode() const noexcept = 0;
566 virtual
void setWindowSizeNd(
Dims const& windowSize) noexcept = 0;
567 virtual
Dims getWindowSizeNd() const noexcept = 0;
568 virtual
void setStrideNd(
Dims const& stride) noexcept = 0;
569 virtual
Dims getStrideNd() const noexcept = 0;
570 virtual
void setPaddingNd(
Dims const& padding) noexcept = 0;
571 virtual
Dims getPaddingNd() const noexcept = 0;
574class VLRNLayer : public VRoot
577 virtual void setWindowSize(int64_t windowSize)
noexcept = 0;
578 virtual int64_t getWindowSize() const noexcept = 0;
579 virtual
void setAlpha(
float alpha) noexcept = 0;
580 virtual
float getAlpha() const noexcept = 0;
581 virtual
void setBeta(
float beta) noexcept = 0;
582 virtual
float getBeta() const noexcept = 0;
583 virtual
void setK(
float k) noexcept = 0;
584 virtual
float getK() const noexcept = 0;
587class VScaleLayer : public VRoot
590 virtual void setMode(
ScaleMode mode)
noexcept = 0;
591 virtual ScaleMode getMode() const noexcept = 0;
592 virtual
void setShift(Weights shift) noexcept = 0;
593 virtual Weights getShift() const noexcept = 0;
594 virtual
void setScale(Weights scale) noexcept = 0;
595 virtual Weights getScale() const noexcept = 0;
596 virtual
void setPower(Weights power) noexcept = 0;
597 virtual Weights getPower() const noexcept = 0;
598 virtual int32_t getChannelAxis() const noexcept = 0;
599 virtual
void setChannelAxis(int32_t channelAxis) noexcept = 0;
602class VSoftMaxLayer : public VRoot
605 virtual void setAxes(uint32_t axes)
noexcept = 0;
606 virtual uint32_t getAxes() const noexcept = 0;
609class VConcatenationLayer : public VRoot
612 virtual void setAxis(int32_t axis)
noexcept = 0;
613 virtual int32_t getAxis() const noexcept = 0;
616class VDeconvolutionLayer : public VRoot
619 virtual void setNbOutputMaps(int64_t nbOutputMaps)
noexcept = 0;
620 virtual int64_t getNbOutputMaps() const noexcept = 0;
621 virtual
void setNbGroups(int64_t nbGroups) noexcept = 0;
622 virtual int64_t getNbGroups() const noexcept = 0;
623 virtual
void setKernelWeights(Weights weights) noexcept = 0;
624 virtual Weights getKernelWeights() const noexcept = 0;
625 virtual
void setBiasWeights(Weights weights) noexcept = 0;
626 virtual Weights getBiasWeights() const noexcept = 0;
627 virtual
void setPrePadding(
Dims const& padding) noexcept = 0;
628 virtual
Dims getPrePadding() const noexcept = 0;
629 virtual
void setPostPadding(
Dims const& padding) noexcept = 0;
630 virtual
Dims getPostPadding() const noexcept = 0;
631 virtual
void setPaddingMode(
PaddingMode paddingMode) noexcept = 0;
632 virtual
PaddingMode getPaddingMode() const noexcept = 0;
633 virtual
void setKernelSizeNd(
Dims const& kernelSize) noexcept = 0;
634 virtual
Dims getKernelSizeNd() const noexcept = 0;
635 virtual
void setStrideNd(
Dims const& stride) noexcept = 0;
636 virtual
Dims getStrideNd() const noexcept = 0;
637 virtual
void setPaddingNd(
Dims const& padding) noexcept = 0;
638 virtual
Dims getPaddingNd() const noexcept = 0;
639 virtual
void setDilationNd(
Dims const& dilation) noexcept = 0;
640 virtual
Dims getDilationNd() const noexcept = 0;
643class VElementWiseLayer : public VRoot
650class VGatherLayer : public VRoot
653 virtual void setGatherAxis(int32_t axis)
noexcept = 0;
654 virtual int32_t getGatherAxis() const noexcept = 0;
655 virtual
void setNbElementWiseDims(int32_t k) noexcept = 0;
656 virtual int32_t getNbElementWiseDims() const noexcept = 0;
657 virtual
void setMode(
GatherMode mode) noexcept = 0;
658 virtual
GatherMode getMode() const noexcept = 0;
661class VPluginLayer : public VRoot
664 virtual IPlugin& getPlugin() noexcept = 0;
667class VPluginV2Layer : public VRoot
670 virtual IPluginV2& getPlugin() noexcept = 0;
673class VPluginV3Layer : public VRoot
676 virtual IPluginV3& getPlugin() noexcept = 0;
679class VUnaryLayer : public VRoot
686class VReduceLayer : public VRoot
691 virtual
void setReduceAxes(uint32_t reduceAxes) noexcept = 0;
692 virtual uint32_t getReduceAxes() const noexcept = 0;
693 virtual
void setKeepDimensions(
bool keepDimensions) noexcept = 0;
694 virtual
bool getKeepDimensions() const noexcept = 0;
697class VPaddingLayer : public VRoot
700 virtual void setPrePaddingNd(
Dims const& padding)
noexcept = 0;
701 virtual Dims getPrePaddingNd() const noexcept = 0;
702 virtual
void setPostPaddingNd(
Dims const& padding) noexcept = 0;
703 virtual
Dims getPostPaddingNd() const noexcept = 0;
706class VShuffleLayer : public VRoot
709 virtual void setFirstTranspose(Permutation
const& permutation)
noexcept = 0;
710 virtual Permutation
const& getFirstTranspose() const noexcept = 0;
711 virtual
void setReshapeDimensions(
Dims const& dimensions) noexcept = 0;
712 virtual
Dims getReshapeDimensions() const noexcept = 0;
713 virtual
void setSecondTranspose(Permutation const& permutation) noexcept = 0;
714 virtual Permutation const& getSecondTranspose() const noexcept = 0;
715 virtual
void setZeroIsPlaceholder(
bool zeroIsPlaceholder) noexcept = 0;
716 virtual
bool getZeroIsPlaceholder() const noexcept = 0;
719class VSliceLayer : public VRoot
722 virtual void setStart(
Dims const& start)
noexcept = 0;
723 virtual Dims getStart() const noexcept = 0;
724 virtual
void setSize(
Dims const& size) noexcept = 0;
725 virtual
Dims getSize() const noexcept = 0;
726 virtual
void setStride(
Dims const& stride) noexcept = 0;
727 virtual
Dims getStride() const noexcept = 0;
728 virtual
void setMode(
SampleMode mode) noexcept = 0;
729 virtual
SampleMode getMode() const noexcept = 0;
730 virtual
void setAxes(
Dims const& axes) noexcept = 0;
731 virtual
Dims getAxes() const noexcept = 0;
734class VShapeLayer : public VRoot
739class VTopKLayer :
public VRoot
744 virtual
void setK(int32_t k) noexcept = 0;
745 virtual int32_t getK() const noexcept = 0;
746 virtual
void setReduceAxes(uint32_t reduceAxes) noexcept = 0;
747 virtual uint32_t getReduceAxes() const noexcept = 0;
748 virtual
bool setIndicesType(
DataType type) noexcept = 0;
749 virtual
DataType getIndicesType() const noexcept = 0;
752class VMatrixMultiplyLayer : public VRoot
755 virtual void setOperation(int32_t index,
MatrixOperation op)
noexcept = 0;
756 virtual MatrixOperation getOperation(int32_t index)
const noexcept = 0;
759class VNonZeroLayer :
public VRoot
762 virtual bool setIndicesType(
DataType type)
noexcept = 0;
763 virtual DataType getIndicesType() const noexcept = 0;
766class VRaggedSoftMaxLayer : public VRoot
771class VIdentityLayer :
public VRoot
776class VCastLayer :
public VRoot
779 virtual void setToType(
DataType toType)
noexcept = 0;
780 virtual DataType getToType() const noexcept = 0;
783class VConstantLayer : public VRoot
786 virtual void setWeights(Weights weights)
noexcept = 0;
787 virtual Weights getWeights() const noexcept = 0;
788 virtual
void setDimensions(
Dims const& dimensions) noexcept = 0;
789 virtual
Dims getDimensions() const noexcept = 0;
792class VParametricReLULayer : public VRoot
797class VResizeLayer :
public VRoot
800 virtual void setOutputDimensions(
Dims const& dimensions)
noexcept = 0;
801 virtual Dims getOutputDimensions() const noexcept = 0;
802 virtual
void setScales(
float const* scales, int32_t nbScales) noexcept = 0;
803 virtual int32_t getScales(int32_t size,
float* scales) const noexcept = 0;
808 virtual
void setSelectorForSinglePixel(
ResizeSelector selector) noexcept = 0;
809 virtual
ResizeSelector getSelectorForSinglePixel() const noexcept = 0;
812 virtual
void setCubicCoeff(
float value) noexcept = 0;
813 virtual
float getCubicCoeff() const noexcept = 0;
814 virtual
void setExcludeOutside(
bool value) noexcept = 0;
815 virtual
bool getExcludeOutside() const noexcept = 0;
818class VLoopBoundaryLayer : public VRoot
821 virtual ILoop* getLoop() const noexcept = 0;
824class VRecurrenceLayer : public VRoot
829class VLoopOutputLayer :
public VRoot
832 virtual LoopOutput getLoopOutput() const noexcept = 0;
833 virtual
void setAxis(int32_t axis) noexcept = 0;
834 virtual int32_t getAxis() const noexcept = 0;
837class VTripLimitLayer : public VRoot
840 virtual TripLimit getTripLimit() const noexcept = 0;
843class VIteratorLayer : public VRoot
846 virtual void setAxis(int32_t axis)
noexcept = 0;
847 virtual int32_t getAxis() const noexcept = 0;
848 virtual
void setReverse(
bool reverse) noexcept = 0;
849 virtual
bool getReverse() const noexcept = 0;
851class VLoop : public VRoot
854 virtual IRecurrenceLayer* addRecurrence(ITensor& initialValue)
noexcept = 0;
855 virtual ITripLimitLayer* addTripLimit(ITensor& tensor,
TripLimit limit)
noexcept = 0;
856 virtual IIteratorLayer* addIterator(ITensor& tensor, int32_t axis = 0,
bool reverse =
false) noexcept = 0;
857 virtual ILoopOutputLayer* addLoopOutput(ITensor& tensor,
LoopOutput outputKind, int32_t axis = 0) noexcept = 0;
858 virtual
void setName(
char const* name) noexcept = 0;
859 virtual
char const* getName() const noexcept = 0;
862class VConditionalBoundaryLayer : public VRoot
865 virtual IIfConditional* getConditional() const noexcept = 0;
868class VConditionLayer : public VRoot
873class VConditionalInputLayer :
public VRoot
878class VConditionalOutputLayer :
public VRoot
883class VIfConditional :
public VRoot
886 virtual IConditionLayer* setCondition(ITensor& tensor)
noexcept = 0;
887 virtual IIfConditionalInputLayer* addInput(ITensor& tensor)
noexcept = 0;
888 virtual IIfConditionalOutputLayer* addOutput(ITensor& trueTensor, ITensor& falseTensor)
noexcept = 0;
889 virtual void setName(
char const* name)
noexcept = 0;
890 virtual char const* getName() const noexcept = 0;
893class VAttentionBoundaryLayer : public VRoot
896 virtual IAttention* getAttention() const noexcept = 0;
899class VAttentionInputLayer : public VRoot
904class VAttentionOutputLayer :
public VRoot
909class VAttention :
public VRoot
912 TRT_NODISCARD virtual bool setInput(int32_t index, ITensor& input)
noexcept = 0;
913 TRT_NODISCARD virtual int32_t getNbInputs() const noexcept = 0;
914 TRT_NODISCARD virtual ITensor* getInput(int32_t index) const noexcept = 0;
915 TRT_NODISCARD virtual int32_t getNbOutputs() const noexcept = 0;
916 TRT_NODISCARD virtual ITensor* getOutput(int32_t index) const noexcept = 0;
917 TRT_NODISCARD virtual
bool setName(
char const* name) noexcept = 0;
918 TRT_NODISCARD virtual
char const* getName() const noexcept = 0;
921 TRT_DEPRECATED virtual
bool setCausal(
bool isCausal) noexcept = 0;
923 TRT_NODISCARD virtual
bool setMask(ITensor& mask) noexcept = 0;
925 TRT_NODISCARD virtual
bool setDecomposable(
bool decomposable) noexcept = 0;
926 TRT_NODISCARD virtual
bool getDecomposable() const noexcept = 0;
927 TRT_NODISCARD virtual
bool setNormalizationQuantizeScale(ITensor& tensor) noexcept = 0;
928 TRT_NODISCARD virtual ITensor* getNormalizationQuantizeScale() const noexcept = 0;
931 TRT_NODISCARD virtual
bool setMetadata(
char const* docString) noexcept = 0;
932 TRT_NODISCARD virtual
char const* getMetadata() const noexcept = 0;
933 TRT_NODISCARD virtual
bool setNbRanks(int32_t nbRanks) noexcept = 0;
934 TRT_NODISCARD virtual int32_t getNbRanks() const noexcept = 0;
941 TRT_NODISCARD virtual
bool setQueryLengths(ITensor* lengths) noexcept = 0;
942 TRT_NODISCARD virtual ITensor* getQueryLengths() const noexcept = 0;
943 TRT_NODISCARD virtual
bool setKeyValueLengths(ITensor* lengths) noexcept = 0;
944 TRT_NODISCARD virtual ITensor* getKeyValueLengths() const noexcept = 0;
947class VSelectLayer : public VRoot
951class VAssertionLayer :
public VRoot
954 virtual void setMessage(
char const* message)
noexcept = 0;
955 virtual char const* getMessage() const noexcept = 0;
958class VFillLayer : public VRoot
961 virtual void setDimensions(
Dims const& dimensions)
noexcept = 0;
962 virtual Dims getDimensions() const noexcept = 0;
965 virtual
void setAlpha(
double alpha) noexcept = 0;
966 virtual
double getAlpha() const noexcept = 0;
967 virtual
void setBeta(
double beta) noexcept = 0;
968 virtual
double getBeta() const noexcept = 0;
969 virtual
void setAlphaInt64(int64_t alpha) noexcept = 0;
970 virtual int64_t getAlphaInt64() const noexcept = 0;
971 virtual
void setBetaInt64(int64_t beta) noexcept = 0;
972 virtual int64_t getBetaInt64() const noexcept = 0;
973 virtual
bool isAlphaBetaInt64() const noexcept = 0;
974 virtual
DataType getToType() const noexcept = 0;
975 virtual
void setToType(
DataType toType) noexcept = 0;
978class VQuantizeLayer : public VRoot
981 virtual int32_t getAxis() const noexcept = 0;
982 virtual
void setAxis(int32_t axis) noexcept = 0;
983 virtual
DataType getToType() const noexcept = 0;
984 virtual
void setToType(
DataType toType) noexcept = 0;
985 virtual
Dims getBlockShape() const noexcept = 0;
986 virtual
bool setBlockShape(
Dims const& blockShape) noexcept = 0;
989class VDequantizeLayer : public VRoot
992 virtual int32_t getAxis() const noexcept = 0;
993 virtual
void setAxis(int32_t axis) noexcept = 0;
994 virtual
DataType getToType() const noexcept = 0;
995 virtual
void setToType(
DataType toType) noexcept = 0;
996 virtual
Dims getBlockShape() const noexcept = 0;
997 virtual
bool setBlockShape(
Dims const& blockShape) noexcept = 0;
1000class VDynamicQuantizeLayer : public VRoot
1005 TRT_DEPRECATED virtual int32_t getBlockSize() const noexcept = 0;
1006 TRT_DEPRECATED virtual
void setBlockSize(int32_t axis) noexcept = 0;
1007 virtual
DataType getScaleType() const noexcept = 0;
1008 virtual
void setScaleType(
DataType axis) noexcept = 0;
1009 virtual
DataType getToType() const noexcept = 0;
1010 virtual
void setToType(
DataType toType) noexcept = 0;
1011 virtual
Dims getBlockShape() const noexcept = 0;
1012 virtual
void setBlockShape(
Dims const& blockShape) noexcept = 0;
1015class VScatterLayer : public VRoot
1018 virtual void setMode(
ScatterMode mode)
noexcept = 0;
1020 virtual
void setAxis(int32_t axis) noexcept = 0;
1021 virtual int32_t getAxis() const noexcept = 0;
1024class VEinsumLayer : public VRoot
1027 virtual bool setEquation(
char const* equation)
noexcept = 0;
1028 virtual char const* getEquation() const noexcept = 0;
1031class VOneHotLayer : public VRoot
1034 virtual int32_t getAxis() const noexcept = 0;
1035 virtual
void setAxis(int32_t axis) noexcept = 0;
1038class VGridSampleLayer : public VRoot
1043 virtual
void setAlignCorners(
bool alignCorners) noexcept = 0;
1044 virtual
bool getAlignCorners() const noexcept = 0;
1045 virtual
bool setSampleMode(
SampleMode mode) noexcept = 0;
1046 virtual
SampleMode getSampleMode() const noexcept = 0;
1049class VNMSLayer : public VRoot
1054 virtual
void setTopKBoxLimit(int32_t limit) noexcept = 0;
1055 virtual int32_t getTopKBoxLimit() const noexcept = 0;
1056 virtual
bool setIndicesType(
DataType type) noexcept = 0;
1057 virtual
DataType getIndicesType() const noexcept = 0;
1060class VReverseSequenceLayer : public VRoot
1063 virtual void setBatchAxis(int32_t batchAxis)
noexcept = 0;
1064 virtual int32_t getBatchAxis() const noexcept = 0;
1066 virtual
void setSequenceAxis(int32_t sequenceAxis) noexcept = 0;
1067 virtual int32_t getSequenceAxis() const noexcept = 0;
1070class VNormalizationLayer : public VRoot
1073 virtual void setEpsilon(
float eps)
noexcept = 0;
1074 virtual float getEpsilon() const noexcept = 0;
1075 virtual
void setAxes(uint32_t axesMask) noexcept = 0;
1076 virtual uint32_t getAxes() const noexcept = 0;
1077 virtual
void setNbGroups(int64_t nbGroups) noexcept = 0;
1078 virtual int64_t getNbGroups() const noexcept = 0;
1079 virtual
bool isV2() const noexcept = 0;
1082class VSqueezeLayer : public VRoot
1086class VUnsqueezeLayer :
public VRoot
1090class VCumulativeLayer :
public VRoot
1095 virtual
void setExclusive(
bool exclusive) noexcept = 0;
1096 virtual
bool getExclusive() const noexcept = 0;
1097 virtual
void setReverse(
bool reverse) noexcept = 0;
1098 virtual
bool getReverse() const noexcept = 0;
1101class VRotaryEmbeddingLayer : public VRoot
1104 virtual void setInterleaved(
bool interleaved)
noexcept = 0;
1105 virtual bool getInterleaved() const noexcept = 0;
1106 virtual
bool setRotaryEmbeddingDim(int32_t rotaryEmbeddingDim) noexcept = 0;
1107 virtual int32_t getRotaryEmbeddingDim() const noexcept = 0;
1108 virtual
void setInput(int32_t index, ITensor& input) noexcept = 0;
1111class VKVCacheUpdateLayer : public VRoot
1118 TRT_NODISCARD virtual
bool setUpdateLengths(ITensor* lengths) noexcept = 0;
1119 TRT_NODISCARD virtual ITensor* getUpdateLengths() const noexcept = 0;
1122class VMoELayer : public VRoot
1125 virtual void setGatedWeights(ITensor& fcGateWeights, ITensor& fcUpWeights, ITensor& fcDownWeights,
MoEActType activationType)
noexcept = 0;
1126 virtual void setGatedBiases(ITensor& fcGateBiases, ITensor& fcUpBiases, ITensor& fcDownBiases)
noexcept = 0;
1127 virtual void setActivationType(
MoEActType activationType)
noexcept = 0;
1128 virtual MoEActType getActivationType() const noexcept = 0;
1129 virtual
void setQuantizationStatic(ITensor& fcDownActivationScale,
DataType dataType) noexcept = 0;
1130 virtual
void setQuantizationDynamicDblQ(ITensor& fcDownActivationDblQScale,
DataType dataType,
Dims const& blockShape,
DataType dynQOutputScaleType) noexcept = 0;
1131 virtual
void setQuantizationToType(
DataType type) noexcept = 0;
1132 virtual
DataType getQuantizationToType() const noexcept = 0;
1133 virtual
void setQuantizationBlockShape(
Dims const& blockShape) noexcept = 0;
1134 virtual
Dims getQuantizationBlockShape() const noexcept = 0;
1135 virtual
void setDynQOutputScaleType(
DataType type) noexcept = 0;
1136 virtual
DataType getDynQOutputScaleType() const noexcept = 0;
1137 virtual
void setSwigluParams(
float limit,
float alpha,
float beta) noexcept = 0;
1138 virtual
void setSwigluParamLimit(
float limit) noexcept = 0;
1139 virtual
float getSwigluParamLimit() const noexcept = 0;
1140 virtual
void setSwigluParamAlpha(
float alpha) noexcept = 0;
1141 virtual
float getSwigluParamAlpha() const noexcept = 0;
1142 virtual
void setSwigluParamBeta(
float beta) noexcept = 0;
1143 virtual
float getSwigluParamBeta() const noexcept = 0;
1144 virtual
void setInput(int32_t index, ITensor& tensor) noexcept = 0;
1148class VNetworkDefinition : public VRoot
1151 virtual ITensor* addInput(
char const* name,
DataType type,
Dims const& dimensions)
noexcept = 0;
1152 virtual void markOutput(ITensor& tensor)
noexcept = 0;
1153 virtual IActivationLayer* addActivation(ITensor& input,
ActivationType type)
noexcept = 0;
1154 virtual ILRNLayer* addLRN(ITensor& input, int64_t window,
float alpha,
float beta,
float k)
noexcept = 0;
1155 virtual IScaleLayer* addScale(
1156 ITensor& input,
ScaleMode mode, Weights shift, Weights scale, Weights power)
noexcept = 0;
1157 virtual ISoftMaxLayer* addSoftMax(ITensor& input)
noexcept = 0;
1158 virtual IConcatenationLayer* addConcatenation(ITensor*
const* inputs, int32_t nbInputs)
noexcept = 0;
1159 virtual IElementWiseLayer* addElementWise(ITensor& input1, ITensor& input2,
ElementWiseOperation op)
noexcept = 0;
1160 virtual IUnaryLayer* addUnary(ITensor& input,
UnaryOperation operation)
noexcept = 0;
1161 virtual IShuffleLayer* addShuffle(ITensor& input)
noexcept = 0;
1162 virtual int32_t getNbLayers() const noexcept = 0;
1163 virtual ILayer* getLayer(int32_t index) const noexcept = 0;
1164 virtual int32_t getNbInputs() const noexcept = 0;
1165 virtual ITensor* getInput(int32_t index) const noexcept = 0;
1166 virtual int32_t getNbOutputs() const noexcept = 0;
1167 virtual ITensor* getOutput(int32_t index) const noexcept = 0;
1168 virtual IReduceLayer* addReduce(
1169 ITensor& input,
ReduceOperation operation, uint32_t reduceAxes,
bool keepDimensions) noexcept
1171 virtual ITopKLayer* addTopK(ITensor& input,
TopKOperation op, int32_t k, uint32_t reduceAxes) noexcept = 0;
1172 virtual IGatherLayer* addGather(ITensor& data, ITensor& indices, int32_t axis) noexcept = 0;
1173 virtual IRaggedSoftMaxLayer* addRaggedSoftMax(ITensor& input, ITensor& bounds) noexcept = 0;
1174 virtual IMatrixMultiplyLayer* addMatrixMultiply(
1176 virtual IConstantLayer* addConstant(
Dims const& dimensions, Weights weights) noexcept = 0;
1177 virtual IIdentityLayer* addIdentity(ITensor& input) noexcept = 0;
1178 virtual
void removeTensor(ITensor& tensor) noexcept = 0;
1179 virtual
void unmarkOutput(ITensor& tensor) noexcept = 0;
1180 virtual ISliceLayer* addSlice(ITensor& input,
Dims const& start,
Dims const& size,
Dims const& stride) noexcept = 0;
1181 virtual
void setName(
char const* name) noexcept = 0;
1182 virtual
char const* getName() const noexcept = 0;
1183 virtual IShapeLayer* addShape(ITensor& input) noexcept = 0;
1184 virtual
bool hasImplicitBatchDimension() const noexcept = 0;
1185 virtual
bool markOutputForShapes(ITensor& tensor) noexcept = 0;
1186 virtual
bool unmarkOutputForShapes(ITensor& tensor) noexcept = 0;
1187 virtual IParametricReLULayer* addParametricReLU(ITensor& input, ITensor& slope) noexcept = 0;
1188 virtual IConvolutionLayer* addConvolutionNd(
1189 ITensor& input, int64_t nbOutputMaps,
Dims const& kernelSize, Weights kernelWeights, Weights biasWeights) noexcept
1191 virtual IPoolingLayer* addPoolingNd(ITensor& input,
PoolingType type,
Dims const& windowSize) noexcept = 0;
1192 virtual IDeconvolutionLayer* addDeconvolutionNd(
1193 ITensor& input, int64_t nbOutputMaps,
Dims const& kernelSize, Weights kernelWeights, Weights biasWeights) noexcept
1195 virtual IScaleLayer* addScaleNd(
1196 ITensor& input,
ScaleMode mode, Weights shift, Weights scale, Weights power, int32_t channelAxis) noexcept = 0;
1197 virtual IResizeLayer* addResize(ITensor& input) noexcept = 0;
1198 virtual ILoop* addLoop() noexcept = 0;
1199 virtual ISelectLayer* addSelect(ITensor& condition, ITensor& thenInput, ITensor& elseInput) noexcept = 0;
1200 virtual IPaddingLayer* addPaddingNd(ITensor& input,
Dims const& prePadding,
Dims const& postPadding) noexcept = 0;
1201 virtual
bool setWeightsName(Weights weights,
char const* name) noexcept = 0;
1202 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
1204 virtual IGatherLayer* addGatherV2(ITensor& data, ITensor& indices,
GatherMode mode) noexcept = 0;
1205 virtual IIfConditional* addIfConditional() noexcept = 0;
1206 virtual IScatterLayer* addScatter(ITensor& data, ITensor& indices, ITensor& updates,
ScatterMode mode) noexcept = 0;
1207 virtual IEinsumLayer* addEinsum(ITensor* const* inputs, int32_t nbInputs,
char const* equation) noexcept = 0;
1208 virtual IAssertionLayer* addAssertion(ITensor& condition,
char const* message) noexcept = 0;
1209 virtual IOneHotLayer* addOneHot(ITensor& indices, ITensor& values, ITensor& depth, int32_t axis) noexcept = 0;
1210 virtual INonZeroLayer* addNonZero(ITensor& input) noexcept = 0;
1211 virtual IGridSampleLayer* addGridSample(ITensor& input, ITensor& grid) noexcept = 0;
1212 virtual INMSLayer* addNMS(ITensor& boxes, ITensor& scores, ITensor& maxOutputBoxesPerClass) noexcept = 0;
1213 virtual IReverseSequenceLayer* addReverseSequence(ITensor& input, ITensor& sequenceLens) noexcept = 0;
1214 virtual INormalizationLayer* addNormalization(
1215 ITensor& input, ITensor& scale, ITensor& bias, uint32_t axesMask) noexcept = 0;
1216 virtual ICastLayer* addCast(ITensor& input,
DataType toType) noexcept = 0;
1217 virtual IBuilder& getBuilder() const noexcept = 0;
1220 virtual IQuantizeLayer* addQuantizeV2(ITensor& input, ITensor& scale,
DataType outputType) noexcept = 0;
1221 virtual IDequantizeLayer* addDequantizeV2(ITensor& input, ITensor& scale,
DataType outputType) noexcept = 0;
1223 virtual
bool markDebug(ITensor& tensor) noexcept = 0;
1224 virtual
bool unmarkDebug(ITensor& tensor) noexcept = 0;
1225 virtual
bool isDebugTensor(ITensor const& tensor) const noexcept = 0;
1226 virtual
bool markWeightsRefittable(
char const* name) noexcept = 0;
1227 virtual
bool unmarkWeightsRefittable(
char const* name) noexcept = 0;
1228 virtual
bool areWeightsMarkedRefittable(
char const* name) const noexcept = 0;
1229 virtual ISqueezeLayer* addSqueeze(ITensor& input, ITensor& axes) noexcept = 0;
1230 virtual IUnsqueezeLayer* addUnsqueeze(ITensor& input, ITensor& axes) noexcept = 0;
1231 virtual IDynamicQuantizeLayer* addDynamicQuantize(
1232 ITensor& input, int32_t axis, int32_t blockSize,
DataType toType,
DataType scaleType) noexcept = 0;
1233 virtual ICumulativeLayer* addCumulative(
1234 ITensor& input, ITensor& axis,
CumulativeOperation operation,
bool exclusive,
bool reverse) noexcept = 0;
1235 virtual
bool markUnfusedTensorsAsDebugTensors() noexcept = 0;
1236 virtual
bool unmarkUnfusedTensorsAsDebugTensors() noexcept = 0;
1237 virtual ITopKLayer* addTopKV2(
1238 ITensor& input,
TopKOperation op, int32_t k, uint32_t reduceAxes,
DataType indicesType) noexcept = 0;
1239 virtual INonZeroLayer* addNonZeroV2(ITensor& input,
DataType indicesType) noexcept = 0;
1240 virtual INMSLayer* addNMSV2(
1241 ITensor& boxes, ITensor& scores, ITensor& maxOutputBoxesPerClass,
DataType indicesType) noexcept = 0;
1244 virtual IRotaryEmbeddingLayer* addRotaryEmbedding(ITensor& input, ITensor& cosCache, ITensor& sinCache,
1245 bool interleaved, int32_t rotaryEmbeddingDim) noexcept = 0;
1246 virtual IDynamicQuantizeLayer* addDynamicQuantizeV2(
1248 virtual IKVCacheUpdateLayer* addKVCacheUpdate(
1249 ITensor& cache, ITensor& update, ITensor& writeIndices,
KVCacheMode cacheMode) noexcept = 0;
1250 virtual INormalizationLayer* addNormalizationV2(
1251 ITensor& input, ITensor& scale, ITensor& bias, uint32_t axesMask) noexcept = 0;
1252 virtual IMoELayer* addMoE(
1253 ITensor& hiddenStates, ITensor& selectedExpertsForTokens, ITensor& scoresForSelectedExperts) noexcept = 0;
1254 virtual IDistCollectiveLayer* addDistCollective(ITensor& input,
CollectiveOperation distCollectiveOp,
1255 ReduceOperation reduceOp, int64_t root, int64_t* groups, int64_t groupSize) noexcept = 0;
1260class VTimingCache : public VRoot
1264 virtual
bool combine(ITimingCache const& inputCache,
bool ignoreMismatch) noexcept = 0;
1265 virtual
bool reset() noexcept = 0;
1266 virtual int64_t queryKeys(TimingCacheKey* keyBuffer, int64_t capacity) const noexcept = 0;
1267 virtual TimingCacheValue query(TimingCacheKey const& key) const noexcept = 0;
1268 virtual
bool update(TimingCacheKey const& key, TimingCacheValue const& value) noexcept = 0;
1271class VBuilderConfig : public VRoot
1274 virtual void setAvgTimingIterations(int32_t avgTiming)
noexcept = 0;
1275 virtual int32_t getAvgTimingIterations() const noexcept = 0;
1276 virtual
void setEngineCapability(
EngineCapability capability) noexcept = 0;
1278 virtual
void setFlags(
BuilderFlags builderFlags) noexcept = 0;
1280 virtual
void clearFlag(
BuilderFlag builderFlag) noexcept = 0;
1281 virtual
void setFlag(
BuilderFlag builderFlag) noexcept = 0;
1282 virtual
bool getFlag(
BuilderFlag builderFlag) const noexcept = 0;
1283 virtual
void setDeviceType(ILayer const* layer,
DeviceType deviceType) noexcept = 0;
1284 virtual
DeviceType getDeviceType(ILayer const* layer) const noexcept = 0;
1285 virtual
bool isDeviceTypeSet(ILayer const* layer) const noexcept = 0;
1286 virtual
void resetDeviceType(ILayer const* layer) noexcept = 0;
1287 virtual
bool canRunOnDLA(ILayer const* layer) const noexcept = 0;
1288 virtual
void setDLACore(int32_t dlaCore) noexcept = 0;
1289 virtual int32_t getDLACore() const noexcept = 0;
1290 virtual
void setDefaultDeviceType(
DeviceType deviceType) noexcept = 0;
1291 virtual
DeviceType getDefaultDeviceType() const noexcept = 0;
1292 virtual
void reset() noexcept = 0;
1293 virtual
void setProfileStream(const cudaStream_t stream) noexcept = 0;
1294 virtual cudaStream_t getProfileStream() const noexcept = 0;
1295 virtual int32_t addOptimizationProfile(IOptimizationProfile const* profile) noexcept = 0;
1296 virtual int32_t getNbOptimizationProfiles() const noexcept = 0;
1299 virtual
bool setTacticSources(
TacticSources tacticSources) noexcept = 0;
1300 virtual
TacticSources getTacticSources() const noexcept = 0;
1301 virtual
nvinfer1::ITimingCache* createTimingCache(
void const* blob, std::
size_t size) const noexcept = 0;
1302 virtual
bool setTimingCache(ITimingCache const& cache,
bool ignoreMismatch) noexcept = 0;
1303 virtual
nvinfer1::ITimingCache const* getTimingCache() const noexcept = 0;
1304 virtual
void setMemoryPoolLimit(
MemoryPoolType pool, std::
size_t poolSize) noexcept = 0;
1305 virtual std::
size_t getMemoryPoolLimit(
MemoryPoolType pool) const noexcept = 0;
1306 virtual
void setPreviewFeature(
PreviewFeature feature,
bool enable) noexcept = 0;
1307 virtual
bool getPreviewFeature(
PreviewFeature feature) const noexcept = 0;
1308 virtual
void setBuilderOptimizationLevel(int32_t level) noexcept = 0;
1309 virtual int32_t getBuilderOptimizationLevel() const noexcept = 0;
1312 virtual
void setPluginsToSerialize(
char const* const* paths, int32_t nbPaths) noexcept = 0;
1313 virtual
char const* getPluginToSerialize(int32_t index) const noexcept = 0;
1314 virtual int32_t getNbPluginsToSerialize() const noexcept = 0;
1315 virtual
bool setMaxAuxStreams(int32_t nbStreams) noexcept = 0;
1316 virtual int32_t getMaxAuxStreams() const noexcept = 0;
1319 virtual
void setRuntimePlatform(
RuntimePlatform runtimePlatform) noexcept = 0;
1321 virtual
void setMaxNbTactics(int32_t maxTactics) noexcept = 0;
1322 virtual int32_t getMaxNbTactics() const noexcept = 0;
1325 virtual
bool setL2LimitForTiling(int64_t size) noexcept = 0;
1326 virtual int64_t getL2LimitForTiling() const noexcept = 0;
1327 TRT_NODISCARD virtual
bool setNbComputeCapabilities(int32_t maxNbComputeCapabilities) noexcept = 0;
1328 TRT_NODISCARD virtual int32_t getNbComputeCapabilities() const noexcept = 0;
1333class VSerializationConfig : public VRoot
1343class VBuilder : public VRoot
1346 virtual int32_t getMaxDLABatchSize() const noexcept = 0;
1347 virtual int32_t getNbDLACores() const noexcept = 0;
1348 virtual
void setGpuAllocator(
IGpuAllocator* allocator) noexcept = 0;
1349 virtual
nvinfer1::IBuilderConfig* createBuilderConfig() noexcept = 0;
1351 virtual
nvinfer1::IOptimizationProfile* createOptimizationProfile() noexcept = 0;
1352 virtual
void setErrorRecorder(
IErrorRecorder* recorder) noexcept = 0;
1354 virtual
void reset() noexcept = 0;
1355 virtual
nvinfer1::IHostMemory* buildSerializedNetwork(
1356 INetworkDefinition& network, IBuilderConfig& config) noexcept = 0;
1357 virtual
bool isNetworkSupported(INetworkDefinition const& network, IBuilderConfig const& config) const noexcept = 0;
1359 virtual
bool setMaxThreads(int32_t maxThreads) noexcept = 0;
1360 virtual int32_t getMaxThreads() const noexcept = 0;
1362 virtual
bool buildSerializedNetworkToStream(
1363 INetworkDefinition& network, IBuilderConfig& config,
IStreamWriter& writer) noexcept = 0;
1366class VRuntimeConfig : public VRoot
1369 virtual IRuntimeConfig* getPImpl() noexcept = 0;
1373 virtual
bool setRuntimeCache(
nvinfer1::IRuntimeCache const& cache) noexcept = 0;
1375 virtual
void setDynamicShapesKernelSpecializationStrategy(
1382class VRuntimeCache : public VRoot
1386 virtual
bool deserialize(
void const* blob,
size_t size) noexcept = 0;
1387 virtual
bool reset() noexcept = 0;
1390class 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:139
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:2870
v_1_0::IOutputAllocator IOutputAllocator
Definition: NvInferRuntime.h:4170
ResizeSelector
The coordinate selector when resize to single pixel output.
Definition: NvInfer.h:3917
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:178
MemoryPoolType
The type for memory pools used by TensorRT.
Definition: NvInfer.h:10185
AttentionIOForm
Enumerates the layout of the input/output tensors in an Attention layer.
Definition: NvInfer.h:6728
ScaleMode
Controls how shift, scale and power are applied in a Scale layer.
Definition: NvInfer.h:1587
RuntimePlatform
Describes the intended runtime platform (operating system and CPU architecture) for the execution of ...
Definition: NvInfer.h:9803
TensorIOMode
Definition of tensor IO Mode.
Definition: NvInferRuntimeBase.h:662
HardwareCompatibilityLevel
Describes requirements of compatibility with GPU architectures other than that of the GPU on which th...
Definition: NvInfer.h:10293
CumulativeOperation
Enumerates the cumulative operations that may be performed by a Cumulative layer.
Definition: NvInfer.h:6536
SerializationFlag
List of valid flags that the engine can enable when serializing the bytes.
Definition: NvInferRuntime.h:2911
BoundingBoxFormat
Representation of bounding box data used for the Boxes input tensor in INMSLayer.
Definition: NvInfer.h:6103
ComputeCapability
Describes compute capability that an engine will be built for.
Definition: NvInfer.h:10339
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:2545
v_1_0::IStreamWriter IStreamWriter
Definition: NvInferRuntime.h:699
v_1_0::IProfiler IProfiler
Definition: NvInferRuntime.h:1309
DynamicShapesKernelSpecializationStrategy
Different kernel specialization strategies for dynamic shapes.
Definition: NvInferRuntime.h:3091
v_1_0::IStreamReaderV2 IStreamReaderV2
Definition: NvInferRuntime.h:776
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:1387
FillOperation
Enumerates the tensor fill operations that may performed by a fill layer.
Definition: NvInfer.h:4920
EngineStat
The kind of engine statistics that queried from the ICudaEngine.
Definition: NvInferRuntime.h:3282
CudaGraphStrategy
Strategies available for CUDA graphs optimizations for JIT (Just-In-Time) inference.
Definition: NvInferRuntime.h:3127
ResizeRoundMode
The rounding mode for nearest neighbor resize.
Definition: NvInfer.h:3944
EngineValidity
Whether a TensorRT-RTX engine is likely to be valid on the current system.
Definition: NvInferRuntime.h:1859
v_1_0::ILogger ILogger
Definition: NvInferRuntimeBase.h:125
CausalMaskKind
Enumerates the causal mask alignment orientation for the attention.
Definition: NvInfer.h:6700
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:4302
uint32_t NetworkDefinitionCreationFlags
Represents one or more NetworkDefinitionCreationFlag flags using binary OR operations....
Definition: NvInfer.h:11345
PreviewFeature
Define preview features.
Definition: NvInfer.h:10259
TilingOptimizationLevel
Define the optimization levels for Tiling.
Definition: NvInfer.h:10367
DataType
The type of weights and tensors. The datatypes other than kBOOL, kINT32, and kINT64 are "activation d...
Definition: NvInferRuntimeBase.h:149
uint32_t BuilderFlags
Represents one or more BuilderFlag values using binary OR operations, e.g., 1U << BuilderFlag::kDEBUG...
Definition: NvInfer.h:9832
DeviceType
The device that this layer/network will execute on.
Definition: NvInferRuntime.h:1341
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:3047
GatherMode
Control form of IGatherLayer.
Definition: NvInfer.h:2286
v_1_0::IDebugListener IDebugListener
Definition: NvInferRuntime.h:4211
MoEActType
Enumerates the activation type for the MoE layer.
Definition: NvInfer.h:7617
WeightsRole
How a layer uses particular Weights.
Definition: NvInferRuntime.h:1319
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:2882
NetworkDefinitionCreationFlag
List of immutable network properties expressed at network creation time. NetworkDefinitionCreationFla...
Definition: NvInfer.h:11356
ElementWiseOperation
Enumerates the binary operations that may be performed by an ElementWise layer.
Definition: NvInfer.h:2197
CollectiveOperation
Enumerates the collective operations that may be performed by a DistCollective layer.
Definition: NvInfer.h:2675
uint32_t SerializationFlags
Represents one or more SerializationFlag values using binary OR operations, e.g., 1U << Serialization...
Definition: NvInferRuntime.h:2901
InterpolationMode
Enumerates various modes of interpolation.
Definition: NvInfer.h:3841
BuilderFlag
List of valid modes that the builder can enable when creating an engine from a network definition.
Definition: NvInfer.h:9842
TensorFormat
Format of the input/output tensors.
Definition: NvInferRuntime.h:1421
ExecutionContextAllocationStrategy
Different memory allocation behaviors for IExecutionContext.
Definition: NvInferRuntime.h:3018
TopKOperation
Enumerates the operations that may be performed by a TopK layer.
Definition: NvInfer.h:3332
ReduceOperation
Enumerates the reduce operations that may be performed by a Reduce layer.
Definition: NvInfer.h:2647
LayerInformationFormat
The format in which the IEngineInspector prints the layer information.
Definition: NvInferRuntime.h:5166
ScatterMode
Control form of IScatterLayer.
Definition: NvInfer.h:5830
MatrixOperation
Enumerates the operations that may be performed on a tensor by IMatrixMultiplyLayer before multiplica...
Definition: NvInfer.h:3487
ResizeCoordinateTransformation
The resize coordinate transformation function.
Definition: NvInfer.h:3866
LoopOutput
Enum that describes kinds of loop outputs.
Definition: NvInfer.h:4274
KVCacheMode
Enumerates the KVCache modes that may be performed by a KVCacheUpdate layer.
Definition: NvInfer.h:7464
v_1_0::IStreamReader IStreamReader
Definition: NvInferRuntime.h:689
PoolingType
The type of pooling to perform in a pooling layer.
Definition: NvInfer.h:1200
v_1_0::IProgressMonitor IProgressMonitor
Definition: NvInfer.h:10480
TensorLocation
The location for tensor data storage, device or host.
Definition: NvInferRuntime.h:203
OptProfileSelector
When setting or querying optimization profile parameters (such as shape tensor inputs or dynamic dime...
Definition: NvInferRuntime.h:2629
AttentionNormalizationOp
Enumerates the operations that may be performed by the normalization in the attention subgraph.
Definition: NvInfer.h:6668