This document is relevant for: Inf1, Inf2, Trn1, Trn2, Trn3
nrt_experimental.h#
Neuron Runtime Experimental API - Features under development and subject to change.
Source: src/libnrt/include/nrt/nrt_experimental.h
Note
Experimental APIs are provided for testing and feedback and may not be appropriate for production environments.
Enumerations#
nrt_tensor_usage_t#
typedef enum nrt_tensor_usage {
NRT_TENSOR_USAGE_INPUT = 0,
NRT_TENSOR_USAGE_OUTPUT,
} nrt_tensor_usage_t;
Usage of a Tensor in the NEFF.
Source: nrt_experimental.h:18
Structures#
nrt_tensor_info_t#
typedef struct nrt_tensor_info {
char name[NRT_TENSOR_NAME_MAX];
nrt_tensor_usage_t usage;
size_t size;
nrt_dtype_t dtype;
uint32_t *shape;
uint32_t ndim;
} nrt_tensor_info_t;
Tensor information including name, usage, size, data type, and shape.
Source: nrt_experimental.h:25
nrt_tensor_info_array_t#
typedef struct nrt_tensor_info_array {
uint64_t tensor_count;
nrt_tensor_info_t tensor_array[];
} nrt_tensor_info_array_t;
Array of tensor information.
Source: nrt_experimental.h:34
nrt_model_info_t#
typedef struct nrt_model_info {
uint32_t vnc;
} nrt_model_info_t;
Model information structure.
Source: nrt_experimental.h:139
Functions#
nrt_get_model_tensor_info#
NRT_STATUS nrt_get_model_tensor_info(nrt_model_t *model, nrt_tensor_info_array_t **tensor_info);
Return input/output tensor information for a given model.
Parameters:
model[in] - Model for which tensor information needs to be extracted.tensor_info[out] - Pointer to store the result.
Returns: NRT_STATUS_SUCCESS on success.
Source: nrt_experimental.h:48
nrt_trace_start#
NRT_STATUS nrt_trace_start(bool trace_mem);
Enable tracing for all VNCs visible to the app.
Parameters:
trace_mem[in] - collect memory allocation info
Returns: NRT_SUCCESS on success.
Source: nrt_experimental.h:68
nrt_trace_stop#
NRT_STATUS nrt_trace_stop(const char *filename);
Serialize all data and disable tracing.
Parameters:
filename[in] - filename to write to
Returns: NRT_SUCCESS on success.
Source: nrt_experimental.h:75
nrt_barrier#
NRT_STATUS nrt_barrier(int32_t vnc, uint32_t g_device_id, uint32_t g_device_count);
Implements a barrier by running a small all-reduce over all workers.
Parameters:
vnc[in] - local VNC (within the instance)global_device_id[in] - global worker IDglobal_device_count[in] - total number of workers
Returns: NRT_STATUS_SUCCESS on success.
Source: nrt_experimental.h:115
This document is relevant for: Inf1, Inf2, Trn1, Trn2, Trn3