API documentation

Core API

Core functionality of the API, available directly via qai_hub.

upload_dataset	Upload a dataset that expires in 30 days.
upload_model	Uploads a model.
get_dataset	Returns a dataset for a given id.
get_datasets	Returns a list of datasets visible to you.
get_devices	Returns a list of available devices.
get_device_attributes	Returns the super set of available device attributes.
get_job	Returns a job for a given id.
get_job_summaries	Returns summary information for jobs matching the specified filters.
get_model	Returns a model for a given id.
get_models	Returns a list of models.
set_verbose	If true, API calls may print progress to standard output.
submit_compile_job	Submits a compile job.
submit_quantize_job	Submits a quantize job.
submit_profile_job	Submits a profile job.
submit_inference_job	Submits an inference job.
submit_link_job	Submits a link job.
submit_compile_and_profile_jobs	Submits a compilation job and a profile job.
submit_compile_and_quantize_jobs	Compiles a model to onnx and runs a quantize job on the produced onnx model.
get_jobs	Deprecated.

Managed objects

Classes returned by the API. These objects are managed by the API and should not be instantiated directly. For example, to construct a Model instance, call upload_model() or get_model().

Dataset	A dataset should not be constructed directly.
Device	Create a target device representation.
Model	Neural network model object.
Job	abstract Job base class.
JobSummary	Summary information about a job and its current status.
CompileJobSummary	Summary information about a compile job and its current status.
QuantizeJobSummary	Summary information about a quantize job and its current status.
ProfileJobSummary	Summary information about a profile job and its current status.
InferenceJobSummary	Summary information about an inference job and its current status.
LinkJobSummary	Summary information about a link job and its current status.
CompileJob	Compile job for a model, a set of input specs, and a set of device.
ProfileJob	Profile job for a model, a set of input specs, and a device.
QuantizeJob	Quantize job for a model, a set of input specs, and a set of device.
InferenceJob	Inference job for a model, user provided inputs, and a device.
LinkJob	Link job for a collection of models.
CompileJobResult	Compile Job result structure.
ProfileJobResult	Profile Job result structure.
QuantizeJobResult	Quantize Job result structure.
InferenceJobResult	Inference Job result structure.
LinkJobResult	Link Job result structure.
QuantizeDtype	Supported data types when submitting quantize jobs.
JobStatus	Status of a job.
JobType	The type of a job (compile, profile, etc.)
SourceModelType	Set of supported input model types.
ModelMetadataKey	Model metadata key.

Exceptions

Exceptions thrown by the API:

Error	Base class for all exceptions explicitly thrown by the API.
UserError	Something in the user input caused a failure; you may need to adjust your input.
InternalError	Internal API failure; please contact ai-hub-support@qti.qualcomm.com for assistance.

Common Options

Options for submit_compile_job(), submit_profile_job(), submit_inference_job():

Option

Description

--compute_unit <units>

Specifies the target compute unit(s). When used in a compile job, this can optimize the code generation of various layers. When used in a profile or inference job, this targets the specified compute unit(s). Implicit fall back to CPU will be determined by the target_runtime; TfLite and ONNX Runtime always include CPU fallback. Options for <units> (comma-separated values are unordered): all Target all available compute units with priority NPU, then GPU, then CPU. npu Target NPU. gpu Target GPU. cpu Target CPU. Default: npu when the target runtime is qnn_lib_aarch64_android all for any other target runtime Example: --compute_unit npu,cpu

Compile Options

Options for submit_compile_job():

Compile option	Description
--target_runtime <runtime>	Overrides the default target runtime. The default target runtime is derived from the source model and the target device. Options for <runtime>: tflite: TensorFlow Lite (.tflite) qnn_lib_aarch64_android: Qualcomm® AI Engine Direct model library (.so) targeting AArch64 Android qnn_context_binary: Qualcomm® AI Engine Direct context binary (.bin) targeting the hardware specified in the compile job. onnx: ONNX Runtime (.onnx) precompiled_qnn_onnx: ONNX Runtime model with a pre-compiled QNN context binary. The default target is decided as follows: .aimet (TorchScript) models are compiled to qnn_lib_aarch64_android Default target is set to tflite for any Android device Default target is set to onnx for any Windows device Examples: --target_runtime tflite --target_runtime qnn_lib_aarch64_android
--output_names "<name>[,name]..."	Overrides the default output names. By default, the output names of a model are output_0, output_1 …. Options: comma separated list of output names Requirements: When used, a name must be specified for each model output. Default: Output names of a model are output_0, output_1 …. Examples: --output_names "new_name_0,new_name_1"
--truncate_64bit_tensors [true|false]	instructs the compiler to treat intermediate int64 tensors as int32. Options: true (or no argument) false Requirements: Must be used for the TensorFlow Lite target runtime when the source model has int64 intermediate tensors ignored for other target runtimes Default: Intermediate tensors are not truncated Examples: --truncate_64bit_tensors
--truncate_64bit_io [true|false]	instructs the compiler to treat int64 input and output tensors as int32. Options: true (or no argument) false Requirements: Must be used for TensorFlow Lite and QNN target runtime when the source model has int64 input or output tensors ignored for other target runtimes Default: Inputs and output are not truncated Examples: --truncate_64bit_io
--force_channel_last_input "<name>[,name]..." --force_channel_last_output "<name>[,name]..."	Overrides the channel layout. By default, the compiler maintains the layout of the model’s input and output tensors. This results in additional transpose layers in the final model, most commonly at the beginning and end. These options instruct the compiler to generate code in such a way that the generated model can consume inputs in channel-last layout and/or produce outputs in channel-last layout. Options: comma separated list of input tensor names Requirements: This option cannot be used if the target runtime is ONNX. Note that --force_channel_last_output takes either the default names or the names from the --output_names option. Default: The compiler will maintain the layout of the model’s input and output tensors. Examples: --force_channel_last_input "image,indices" --force_channel_last_output "output_0,output_1"
--quantize_full_type <type>	Quantizes an unquantized model to the specified type. It quantizes both activations and weights using a representative dataset. If not such dataset is provided, a randomly generated one is used. In that case, the generated model can be used as a proxy for the achievable performance only, as the model will not be able to produce accurate results. Options: int8: quantize activations and weights to int8 int16: quantize activations and weights to int16 w8a16: quantize weights to int8 and activations to int16 (recommended over int16) w4a8: quantize weights to int4 and activations to int8 w4a16: quantize weights to int4 and activations to int16 Requirements: This option cannot be used if the input is an AIMET model, or if the target runtime is ONNX. Default: No fake quantization. Examples: --quantize_full_type int8
--quantize_io [true|false]	Quantize the input and outputs when quantizing a model. Options: true (or no argument) false Default: Inputs and output are not quantized. Examples: --quantize_io
--quantize_io_type <type>	Specify the input/output type if --quantize_io is specified. This is currently only available for the following runtime and quantization type (--quantization_type_full): TensorFlow Lite and int8: int8 (default), uint8 Requirements: This option can only be used if the target runtime is TensorFlow Lite. Default: Type is determined based on the runtime and quantization type. Examples: --quantize_io --quantize_io_type uint8
--qnn_graph_name <name>	Defines the graph name for QNN targets. This is particularly consequential for linked models that contain multiple graphs and need dinstinguishing names. Requirements: The <name> must be a valid variable name in C++. Examples: --qnn_graph_name large_input
--qnn_context_binary_vtcm <num>	Specify the amount in MB of VTCM memory to reserve and utilize. Default: 0 (Use the maximum amount of VTCM for a specific device) This option can only be used along with --target_runtime qnn_context_binary or --target_runtime precompiled_qnn_onnx Examples: --qnn_context_binary_vtcm 2
--qnn_context_binary_optimization_level <num>	Only applicable when compiling to QNN context binary. Specify to set the graph optimization value in range 1 to 3. Default: 3 1: Fast preparation time, less optimal graph 2: Longer preparation time, more optimal graph 3: Longest preparation time, most likely even more optimal graph This option can only be used along with --target_runtime qnn_context_binary or --target_runtime precompiled_qnn_onnx Examples: --qnn_context_binary_optimization_level 3

Quantize Options

Options for submit_quantize_job():

Quantize option

Description

--range_scheme <scheme>

Which algorithm to use to determine the quantization range of each tensor in the network. Default is auto. Options: auto Defers to the server to choose an algorithm. Currently this always uses mse_minimizer. mse_minimizer Collects histograms for each tensor and sets the range to include most of the data, while clipping outliers. min_max Sets the range as the absolute minimum and maximum found for each tensor during calibration. Useful when outlier values are important to the network.

Link Options

There are currently no options for submit_link_job().

Profile & Inference Options

Options for submit_profile_job(), submit_inference_job():

Profile option	Description
--dequantize_outputs [true|false]	Dequantize the output. Default is true. Options: true false Default: Output is dequantized Examples: --dequantize_outputs
--tflite_delegates <delegate>[,<delegate>]...	Specifies the TensorFlow Lite delegates to load and target. Multiple delegates can be specified. They will be used in the order specified. Specify the best delegate(s) for the device to run on the desired compute units. Options: qnn qnn-gpu nnapi nnapi-gpu gpu xnnpack Default: The selection of delegates is informed by the compute units specified. Examples: --tflite_delegates qnn,qnn-gpu
--tflite_options <option>[;<option>]...	Specify behavior for the TensorFlow Lite delegates. This option is specified as a semicolon separated list of sub-options. The full set of sub-options are in the TensorFlow Lite Options, TensorFlow Lite Delegate Options for Qualcomm® AI Engine Direct, TensorFlow Lite Delegate Options for GPUv2, and TensorFlow Lite Delegate Options for NNAPI tables below. Examples: --tflite_options number_of_threads=4;qnn_htp_precision=kHtpQuantized
--qnn_options <option>[;<option>]...	Specify behavior when targeting a Qualcomm® AI Engine Direct model library. This option is specified as a semicolon separated list of sub-options. The full set of sub-options are in the Qualcomm® AI Engine Direct Options table below. Requirements: --qnn_options is only applicable when used with --target_runtime qnn_lib_aarch64_android. Examples: --target_runtime qnn_lib_aarch64_android --qnn_options context_priority=HIGH;context_htp_performance_mode=BALANCED
--onnx_options <option>[;<option>]...	Specify behavior for the ONNX Runtime and ONNX Runtime execution providers. This option is specified as a semicolon separated list of sub-options. The full set of sub-options are in the ONNX Runtime Options and ONNX Runtime QNN Execution Provider options tables below. Examples: --onnx_options qnn_htp_performance_mode=burst
--onnx_execution_providers <ep>[,<ep>]...	Specifies the ONNX Runtime execution providers to load and target. This option can be used with --target_runtime onnx. Multiple execution providers can be specified. Options: qnn directml Default: The selection of execution providers is informed by the compute units specified. Examples: --onnx_execution_providers qnn

Profile Options

Options for submit_profile_job():

Profile option	Description
--max_profiler_iterations <num>	Specifies the maximum number of profile iterations. Fewer iterations may be run if their cumulative runtime is predicted to exceed the execution timeout defined by --max_profiler_time. Options: <num> The maximum number of iterations to run. Default: 100 – 100 iterations Examples: --max_profiler_iterations 50
--max_profiler_time <seconds>	Specifies the maximum amount of time to spend performing inference. At runtime, the number of iterations adapts to complete approximately within this limit, never exceeding the value specified by --max_profiler_iterations. For example, let’s say a profile job is submitted with --max_profiler_iterations 100 --max_profiler_time 10. This indicates that no more than 10 seconds should be spent in a loop profiling inference, allowing for up to 100 iterations if they can be completed within that time. The loop is terminated if the running average of inference times predicts that another iteration would cause the timeout to be exceeded. <seconds> The soft limit of how long to spend performing inference on-device. Requirements: This value must be strictly positive and not exceed 600. Default: 600 – 10 minutes Examples: --max_profiler_time 30

ONNX Runtime Options

Sub-options for --onnx_options that are applicable to all execution providers, although some execution providers require certain values.

ONNX Runtime option	Description
execution_mode=<mode>	Sets the execution mode for the session. Default is SEQUENTIAL. Options: SEQUENTIAL PARALLEL
intra_op_num_threads=<num>	Set the number of threads used to parallelize the execution within nodes. Default is 0 – default number of threads.
inter_op_num_threads=<num>	Set the number of threads used to parallelize the execution of the graph. If nodes can be run in parallel, this sets the maximum number of threads to use to run them in parallel. If sequential execution is enabled this option is ignored. Default is 0 – default number of threads.
enable_memory_pattern=[true|false]	Enable a memory pattern optimization. If the input shapes are the same, generate a memory pattern for a future request so that just one allocation can handle all internal memory allocation. Default is false.
enable_cpu_memory_arena=[true|false]	Enable a memory arena on the CPU for future usage. Default is false.
graph_optimization_level=<level>	Set the optimization level to apply when loading a graph. See <https://onnxruntime.ai/docs/performance/model-optimizations/graph-optimizations.htmlOptons> for details. Default is ENABLE_ALL. Options: DISABLE_ALL ENABLE_BASIC ENABLE_EXTENDED ENABLE_ALL

ONNX Runtime QNN Execution Provider options

Sub-options for --onnx_options specific to the QNN execution provider. For additional information on the QNN execution provider options see <https://onnxruntime.ai/docs/execution-providers/QNN-ExecutionProvider.html#configuration-options>.

ONNX Runtime option (QNN)	Description
qnn_htp_performance_mode=<option>	HTP performance mode. Default is burst. Options: default low_power_saver power_saver high_power_saver low_balanced balanced high_performance sustained_high_performance burst
qnn_htp_graph_optimization_mode=[1|2|3]	Set the graph optimization level. Default is 3. Options: 1 Faster preparation time, less optimal graph. 2 Longer preparation time, more optimal graph. 3 Longest preparation time, most likely even more optimal graph.
qnn_enable_htp_fp16_precision=[0|1]	Enable a fp32 model to be inferenced with fp16 precision. This option greatly improves performance of models with any fp32 inputs. Default is 1 (enable).

ONNX Runtime DirectML Execution Provider options

There are no user options for the DirectML execution provider. However, the following options will be automatically set when using DirectML: execution_mode=SEQUENTIAL;enable_memory_pattern=false.

TensorFlow Lite Options

Sub-options for --tflite_options that are applicable to all delegates.

TensorFlow Lite option	Description
enable_fallback=[true|false]	Fallback to lower priority delegates when one fails to prepare. When this is false, any delegate failure will fail the entire job. Default is true.
invoke_interpreter_on_cold_load=[true|false]	Run the interpreter during cold load. This validates that delegates can execute successfully. If an error occurs, failed delegates will be eliminated and the model will be reloaded until a working configuration can be found. Subsequent loads will not re-run the interpreter. Default is false.
allow_fp32_as_fp16=[true|false]	Allow delegates to use reduced fp16 precision for fp32 operations. Delegate options may override this. Default is true.
force_opengl=[true|false]	Force the use of OpenGL when possible. This exists for debugging purposes. Default is false.
number_of_threads=<num>	Number of threads to use. Default: is -1, which then uses std::thread::hardware_concurrency() / 2.
release_dynamic_tensors=[true|false]	Force all intermediate dynamic tensors to be released once they are not used by the model. Please use this configuration with caution, since it might reduce the peak memory usage of the model at the cost of a slower inference speed. Default is false.

TensorFlow Lite Delegate Options for Qualcomm® AI Engine Direct

Sub-options for --tflite_options specific to the QNN delegate. For additional information on the QNN options see the QNN GPU backend and QNN HTP backend pages in the Qualcomm® AI Engine Direct Documentation. Note that not all options are available for all backends.

TfLite option (QNN)	Description
qnn_log_level=<option>	Set the log level for qnn. Default: kLogLevelWarn. Options: kLogOff kLogLevelError kLogLevelWarn kLogLevelInfo kLogLevelVerbose kLogLevelDebug
qnn_graph_priority=<option>	Set the underlying graph priority. Default: kQnnPriorityDefault. Options: kQnnPriorityDefault kQnnPriorityLow kQnnPriorityNormal kQnnPriorityNormalHigh kQnnPriorityHigh kQnnPriorityUndefined
qnn_gpu_precision=<option>	Set the precision for the GPU backend that defines the optimization levels of the graph tensors that are not input nor output tensors. Default is kGpuFp16. Options: kGpuUserProvided kGpuFp32 kGpuFp16 kGpuHybrid
qnn_gpu_performance_mode=<option>	Performance mode for the GPU. Higher performance consumes more power. Default is kGpuHigh. Options: kGpuDefault kGpuHigh kGpuNormal kGpuLow
qnn_dsp_performance_mode=<option>	DSP performance mode. Default is kDspBurst. Options: kDspLowPowerSaver kDspPowerSaver kDspHighPowerSaver kDspLowBalanced kDspBalanced kDspHighPerformance kDspSustainedHighPerformance kDspBurst
qnn_dsp_encoding=<option>	Specify the origin of the quanitization parameters. Default is kDspStatic. Options: kDspStatic: Use quantization parameters from the model. kDspDynamic: Use quantization parameters from the model inputs.
qnn_htp_performance_mode=<option>	HTP performance mode. Default is kHtpBurst. Options: kHtpLowPowerSaver kHtpPowerSaver kHtpHighPowerSaver kHtpLowBalanced kHtpBalanced kHtpHighPerformance kHtpSustainedHighPerformance kHtpBurst
qnn_htp_precision=<option>	HTP precision mode. Only applicable on 8gen1 and newer. Default: is kHtpFp16, when supported. Options: kHtpQuantized kHtpFp16
qnn_htp_optimization_strategy=<option>	Optimize for load time or inference time. Default is kHtpOptimizeForInference. kHtpOptimizeForInference kHtpOptimizeForPrepare
qnn_htp_use_conv_hmx=[true|false]	Using short conv hmx may lead to better performance. However, convolutions that have short depth and/or weights that are not symmetric could exhibit inaccurate results. Default is true.
qnn_htp_use_fold_relu=[true|false]	Using fold relu may lead to better performance. Quantization ranges for convolution should be equal or a subset of the Relu operation for correct results. Default is false.
qnn_htp_vtcm_size=<size>	Set VTCM size in MB. If VTCM size is not set, the default VTCM size will be used. VTCM size must be bigger than 0, and if VTCM size is set to bigger than VTCM size available for this device, it will be set to the VTCM size available for this device.
qnn_htp_num_hvx_threads=<num>	Set the number of HVX threads. If num exceeds the max number of HVX threads, the number will be clipped to the maximum number of threads supported. The num must be greater than 0.

TensorFlow Lite Delegate Options for GPUv2

Sub-options for --tflite_options specific to the GPUv2 delegate. The GPUv2 delegate is chosen with --tflite_delegates gpu. For additional information on the TensorFlow Lite GPU options see delegate_options.h.

TfLite option (GPUv2)	Description
gpu_inference_preference=<option>	Specify the compilation/runtime trade-off. Default is TFLITE_GPU_INFERENCE_PREFERENCE_SUSTAINED_SPEED. Options: TFLITE_GPU_INFERENCE_PREFERENCE_FAST_SINGLE_ANSWER: The delegate will be used only once, therefore, bootstrap/init time should be taken into account. TFLITE_GPU_INFERENCE_PREFERENCE_SUSTAINED_SPEED: Prefer maximizing the throughput. The same delegate will be used repeatedly on multiple inputs. TFLITE_GPU_INFERENCE_PREFERENCE_BALANCED: Balance init latency and throughput. This option will result in slightly higher init latency than TFLITE_GPU_INFERENCE_PREFERENCE_FAST_SINGLE_ANSWER but should have inference latency closer to TFLITE_GPU_INFERENCE_PREFERENCE_SUSTAINED_SPEED.
gpu_inference_priority1=<option> gpu_inference_priority2=<option> gpu_inference_priority3=<option>	Set the top ordered priorities. Default for priority1 is TFLITE_GPU_INFERENCE_PRIORITY_MIN_LATENCY. Default for priority2 is TFLITE_GPU_INFERENCE_PRIORITY_MIN_MEMORY_USAGE. Default for priority3 is TFLITE_GPU_INFERENCE_PRIORITY_MAX_PRECISION. To specify maximum precision, use the following set for priority1-priority3: TFLITE_GPU_INFERENCE_PRIORITY_MAX_PRECISION, TFLITE_GPU_INFERENCE_PRIORITY_AUTO, TFLITE_GPU_INFERENCE_PRIORITY_AUTO. Options: TFLITE_GPU_INFERENCE_PRIORITY_AUTO TFLITE_GPU_INFERENCE_PRIORITY_MAX_PRECISION TFLITE_GPU_INFERENCE_PRIORITY_MIN_LATENCY TFLITE_GPU_INFERENCE_PRIORITY_MIN_MEMORY_USAGE
gpu_max_delegated_partitions=<num>	A graph could have multiple partitions that can be delegated. This limits the maximum number of partitions to be delegated. Default is 1.

TensorFlow Lite Delegate Options for NNAPI

Sub-options for --tflite_options specific to the NNAPI delegate. For additional information on the NNAPI specific options see nnapi_delegate_c_api.h.

TensorFlow Lite option (NNAPI)	Description
nnapi_execution_preference=<option>	Power/performance trade-off. Default is kSustainedSpeed. Options: kLowPower kFastSingleAnswer kSustainedSpeed
nnapi_max_number_delegated_partitions=<max>	Set the maximum number of partitions. A value less than or equal to zero means no limit. If the delegation of the full set of supported nodes would generate a number of partitions greater than this parameter, only <max> of them will be actually accelerated. The selection is currently done sorting partitions in decreasing order of number of nodes and selecting them until the limit is reached. Default is 3.
nnapi_allow_fp16=[true|false]	Allow fp32 to run as fp16. Default is true.

Qualcomm® AI Engine Direct Options

Sub-options for --qnn_options. These options are applicable when the target runtime is qnn_lib_aarch64_android. For additional information on the Qualcomm® AI Engine Direct options see the QNN GPU backend and QNN HTP backend pages in the Qualcomm® AI Engine Direct Documentation. Note that not all options are available for all backends.

Note that the Qualcomm® AI Engine Direct SDK does not guarantee model libraries will be ABI compatible with all versions of the SDK. Using a model library with a newer SDK may result in undefined behavior. If you experience an error using an older model library with AI Hub, you may need to recompile the model library. The version of the QNN SDK used is visible on the AI Hub Compile Job details page.

Qualcomm® AI Engine Direct option	Description
context_async_execution_queue_depth_numeric=<num>	Specify asynchronous execution queue depth.
context_enable_graphs=<name>	The name of the graph to deserialize from a context binary. This argument is necessary if the context binary contains more than one graph. An error is generated if an invalid graph name is provided or multiple graphs are used.
context_error_reporting_options_level=<option>	Sets the error reporting level. Options: BRIEF DETAILED
context_error_reporting_options_storage_limit=<kb>	Amount of memory to be reserved for error information. Specified in KB.
context_memory_limit_hint=<mb>	Sets the peak memory limit hint of a deserialized context in MB.
context_priority=<option>	Sets the default priority for graphs in this context. Options: LOW NORMAL NORMAL_HIGH HIGH
context_gpu_performance_hint=<option>	Set the GPU performance hint. Default is HIGH. Options: HIGH: Maximize GPU clock frequencies. NORMAL: Balance performance dependent upon power management. LOW: Use lowest power consumption at the expense of inference latency.
context_gpu_use_gl_buffers=[true|false]	OpenGL buffers will be used if set to true.
context_htp_performance_mode=<option>	Set the HTP performance mode. Default is BURST. Options: EXTREME_POWER_SAVER LOW_POWER_SAVER POWER_SAVER HIGH_POWER_SAVER LOW_BALANCED BALANCED HIGH_PERFORMANCE SUSTAINED_HIGH_PERFORMANCE BURST
default_graph_priority=<option>	Options: LOW NORMAL NORMAL_HIGH HIGH
default_graph_gpu_precision=<option>	Specify the precision mode to use. Default is USER_PROVIDED. Options: FLOAT32: Convert tensor data types to FP32 and select kernels that use an FP32 accumulator. FLOAT16: Convert tensor data types to FP16 and select kernels that use an FP16 accumulator where possible. HYBRID: Convert tensor data types to FP16 and select kernels that use an FP32 accumulator. USER_PROVIDED: Do not optimize tensor data types.
default_graph_gpu_disable_memory_optimizations=[true|false]	When true, each tensor in the model will be allocated unique memory and sharing is disabled.
default_graph_gpu_disable_node_optimizations=[true|false]	When true, operations will not be fused and will be kept separate.
default_graph_gpu_disable_queue_recording=[true|false]	The QNN GPU backend will use queue recording to improve performance. When true, queue recording is disabled.
default_graph_htp_disable_fold_relu_activation_into_conv=[true|false]	For any graph where a convolution or convolution-like operation is followed by Relu or ReluMinMax, the Relu is folded into the convolution operation. Default is false.
default_graph_htp_num_hvx_threads=<num>	Define number of HVX threads to reserve and utilize for a particular graph. Default is 4.
default_graph_htp_optimization_type=<option>	Set an optimization level to balance prepare and execution. Options: FINALIZE_OPTIMIZATION_FLAG
default_graph_htp_optimization_value=<num>	Specify in combination with default_graph_htp_optimization_type. Options for <num>: 1: Faster preparation time, less optimal graph. 2: Longer preparation time, more optimal graph 3: Longest preparation time, most likely even more optimal graph.
default_graph_htp_precision=<option>	If no precision value is set, the QNN HTP backend assumes that the client expects to run a quantized network. When the precision value is set to FLOAT16, the QNN HTP backend will convert user provided float32 inputs to float16 and execute the graph with float16 math. Options: FLOAT16
default_graph_htp_disable_short_depth_conv_on_hmx=[true|false]	Run all Convolution operations using HMX instructions. Default is false.
default_graph_htp_vtcm_size=<mb>	Set the amount of VTCM memory (in MB) to reserve and utilize. Specify 0 to use the maximum amount. Default is 4.