A single F16 weight overflow broke Qwen3-tts in llama.cpp

The Qwen3-tts code predictor produced NaN output in llama.cpp, and the cause traced to a single layer's numeric range. That's a hazard for anyone quantizing models to save memory. One layer running past F16's range corrupts every output downstream, with nothing louder than NaN to show for it.

The ffn_down weight in that predictor sees intermediate activations peaking around 1.5e5, past F16's 65504 ceiling. Since mul_mat casts its input to match the weight's type, storing ffn_down in F16 turned that peak into infinity. The residual connection carried the infinity forward, and the next rms_norm turned it into NaN.

The b10760 release fixes it by keeping that one layer in F32 while the rest of the model stays quantized. A reference forward pass run in float32 puts the real value at 145109. The graph's F16 path put it at 145396. That's close enough to confirm the layer's activations run that high, and that F16 was never going to hold them.

The fix is narrow by design. Nobody moved the whole model to F32. They moved one layer. Precision needs are decided per layer, not per model. A weight's own magnitude doesn't say whether it's safe to shrink. The activations flowing through it do, and those aren't visible until something breaks.