VectorBlox vnnx_tflite.py Fix: Problem 5 - Constant Tensor Buffer Structure Compatibility Issue Resolution

Problem 5: Constant Tensor Buffer Structure Compatibility

Compilation Failure Log

struct.pack error: required argument is not an integer

File "vnnx_tflite.py", line XXXX

# Constant tensor serialization failed

Root Cause

TFLite models store weights or constant values as constant tensors. These tensors have buffer fields, and VectorBlox’s Tensor structure defines this as a 2-element integer array int32_t buffer[2]:

// VectorBlox structure
typedef struct {
    int32_t buffer[2];  // [buffer_id, offset]
    // ...
} Tensor;

However, in the original vnnx_tflite.py, there were cases where constant tensor buffers were set as a single integer:

tn.buffer = buffer_id  # Single integer

When serializing this with struct.pack(), an error occurred saying “expected 2 integers but got 1”.

Solution Process

All tensor buffers needed to be set consistently in the [buffer_id, offset] format.

Before modification, different formats were used depending on the case:

# General tensor
tn.buffer = [buffer_id, 0]  # Array

# Constant tensor
tn.buffer = buffer_id  # Single integer

After modification, all cases were unified to arrays:

# All tensors
if isinstance(buffer_info, int):
    # Convert single integer to array
    tn.buffer = [buffer_info, 0]
elif isinstance(buffer_info, dict):
    # Convert dict to array
    tn.buffer = [buffer_info.get('id', 0), buffer_info.get('offset', 0)]
else:
    # Keep as is if already an array
    tn.buffer = buffer_info

# Always guarantee [buffer_id, offset] format

The Python code was modified to maintain the same structure as the Tensor structure definition.

Solution Result

After modification, models including constant tensors also compile normally:

[Constant tensor processing]
Unified buffer to [buffer_id, offset] array
struct.pack successful
Compilation successful

Modification Summary

Conclusion

The constant tensor buffer structure compatibility problem occurred due to a mismatch between VectorBlox C structure’s int32_t buffer[2] definition and the Python code.

The solution unified all tensor buffers to the [buffer_id, offset] array format to ensure compatibility with C structures.

Final result: SPNv2’s constant tensors can now be serialized and executed normally on VectorBlox hardware.

Key Improvements

• Guaranteed buffer format consistency
• Perfect compatibility with C structures
• struct.pack serialization stability
• Support for various constant tensor types

Series Posts

• Previous: Problem 4 - INT8 Constant ADD/SUB Operation

Language: 한국어 (Korean)