fix: avoid SIGFPE in fwrite_weight_data when Mat is empty

[lx-99-lyt]

Problem

When running ncnn2int8 to quantize YOLOv8 models, the tool crashes with SIGFPE (Arithmetic exception) during the save() phase.

The crash occurs in ModelWriter::fwrite_weight_data() in tools/modelwriter.h:

ncnn::Mat data_flattened = data.reshape(data.w * data.h * data.d * data.c);

When data is an empty Mat (e.g. for certain layers in YOLOv8 detection head), data.w * data.h * data.d * data.c evaluates to 0, causing a division by zero inside Mat::reshape().

Fix

Add a null check before calling reshape:

if (!data.data || data.w * data.h * data.d * data.c == 0) return 0;

Reproduction

1. Convert YOLOv8n to ncnn format
2. Run ncnn2int8 with a calibration table
3. Tool crashes with SIGFPE after quantizing all convolution layers

[tencent-adm]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ lx-99-lyt
❌ nihui
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[Copilot]

Pull request overview

This PR updates the model serialization logic used by ncnn tools (e.g., ncnn2int8) to prevent crashes during save() when optional weight tensors (bias/affine/etc.) are empty, and adds writer support for the Dequantize layer type.

Changes:

• Guard several fwrite_weight_data() calls so optional parameters (e.g., bias_term, affine, bias_data_size) only write weights when enabled/present.
• Add Dequantize layer serialization (params + scale/bias weights).
• Adjust Scale serialization to skip writing weights for the special scale_data_size == -233 (multi-input) mode.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[nihui]

Thanks for tracking this down and sending the fix.

I’ve updated the patch to handle this at the ModelWriter call sites instead of making fwrite_weight_data() silently skip empty Mats. This keeps optional weights aligned with each layer’s load_model() logic, while still
preserving errors for required missing weights.

Thanks again for the contribution.