Implementation:Ggml org Llama cpp Debug Header

Knowledge Sources	Ggml_org_Llama_cpp
Domains	Debugging, Tensor
Last Updated	2026-02-15 00:00 GMT

Overview

Declares the tensor debugging interface for printing tensor data and registering evaluation callbacks during computation graph processing.

Description

This header declares the template function `common_debug_print_tensor` for formatted tensor data output, parameterized on `abort_on_nan` behavior, and `common_debug_cb_eval` as a ggml backend evaluation callback. The `base_callback_data` struct holds a data buffer for tensor content and a vector of compiled regex filters for selective tensor inspection. Its constructor takes `common_params` and filter patterns, compiling them into anchored regex objects and registering the callback on the params struct.

Usage

Include this header when building debugging tools that need to inspect specific tensors during model execution. Use the `base_callback_data` struct to configure which tensors to inspect by providing regex patterns that match tensor names.

Code Reference

Source Location

• Repository: Ggml_org_Llama_cpp
• File: common/debug.h
• Lines: 1-43

Signature

template <bool abort_on_nan>
void common_debug_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n);

template <bool abort_on_nan>
bool common_debug_cb_eval(struct ggml_tensor * t, bool ask, void * user_data);

struct base_callback_data {
    std::vector<uint8_t>    data;
    std::vector<std::regex> tensor_filters;

    base_callback_data() = default;
    base_callback_data(common_params & params, const std::vector<std::string> & filter_patterns);
};

Import

#include "debug.h"

I/O Contract

Inputs

Name	Type	Required	Description
params	common_params &	Yes	Common parameters struct; callback fields are set by the constructor
filter_patterns	const std::vector<std::string> &	Yes	Regex patterns to match tensor names for selective inspection
data	uint8_t *	Yes	Raw tensor data bytes for print_tensor
type	ggml_type	Yes	Tensor quantization/data type
ne	const int64_t *	Yes	Tensor dimensions array
nb	const size_t *	Yes	Tensor strides array
n	int64_t	Yes	Number of elements to print before truncating

Outputs

Name	Type	Description
base_callback_data instance	base_callback_data	Configured callback data with compiled regex filters and registered callback
cb_eval return	bool	Whether to continue processing the tensor

Usage Examples

#include "debug.h"

// Create callback data with filter patterns
common_params params;
std::vector<std::string> filters = {"attn_q", "attn_k"};
base_callback_data cb_data(params, filters);

// params.cb_eval and params.cb_eval_user_data are now set
// Use params normally with model initialization

Related Pages

• Principle:Ggml_org_Llama_cpp_Tensor_Debugging