[WebGPU] fp16 nanochat produces NaNs (CPU works fine)

[xenova]

Describe the issue

When running the fp16 or q4f16 variants of onnx-community/nanochat-d32-ONNX, the model fails to produce valid output. This issue is present on both JSEP and Native WebGPU EPs; WASM/CPU works fine. This probably points to some overflow issue, but it would be nice to get to the bottom of it.

To reproduce

1. Load model from https://huggingface.co/onnx-community/nanochat-d32-ONNX/blob/main/onnx/model_q4f16.onnx (weights stored at https://huggingface.co/onnx-community/nanochat-d32-ONNX/blob/main/onnx/model_q4f16.onnx_data)
2. Run with Transformers.js or onnxruntime-web
3. Observe invalid output

Urgency

Blocks nanochat from running on q4f16. q4 works fine, so I'll use that in the meantime.

ONNX Runtime Installation

Built from Source

ONNX Runtime Version or Commit ID

latest

Execution Provider

'webgpu' (WebGPU)