> ## Documentation Index
> Fetch the complete documentation index at: https://docs.thrindex.com/llms.txt
> Use this file to discover all available pages before exploring further.

# E0403 — NonFiniteWeights

> A weight tensor in the artifact contains NaN or Inf values that cannot be quantized.

<Warning>E0403: non-finite weights cannot be quantized for akida-akd1500</Warning>

## What happened

A `.thx` artifact was validated against the `akida-akd1500` backend and one of its weight tensors contains one or more `NaN` or `Inf` values. AKD1500 requires integer-quantizable weights; quantization of non-finite floats is undefined.

## Why

The AKD1500 compile path maps `f32` weights to 4-bit signed integers in the range `[-7, 7]` using the scale `max|W| / 7`. If any weight is `NaN` or `Inf`, this scale computation produces `NaN`, and the resulting quantized tensor is garbage. Rather than produce silently incorrect inference output, the backend rejects the artifact at validation time.

Non-finite weights almost always indicate a numerical problem in the training pipeline: exploding gradients, a missing gradient clip, or an unstable loss function.

## How to fix

1. **Identify the source.** Check where in training the weights diverge. Add gradient clipping:

   ```python theme={null}
   torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
   ```

2. **Check your loss function** for instability — a `log(0)` or division by zero in the loss will propagate `Inf` through the backward pass.

3. **Re-export the model** after fixing the training issue and verify with:

   ```python theme={null}
   for name, param in model.named_parameters():
       assert torch.isfinite(param).all(), f"{name} contains non-finite values"
   ```

4. **Recompile** the corrected model for `akida-akd1500`.

## Example

```
E0403: layer[2] weights_b64 contains non-finite values.
Observed: 14 NaN or Inf values found in decoded weight tensor.
Why: AKD1500 requires integer-quantizable weights. Quantization of non-finite
     floats is undefined and would produce silent garbage output.
What to do: Check the training pipeline for numerical instability.
            Re-export the model with finite weights.
Docs: https://docs.thrindex.com/errors/E0403
```