Cannot run unsloth finetuned `Gemma-3-4b-it` model on webgpu

[azizbekabdul]

System Info

transformers.js: 3.7.5

Environment/Platform

• Website/web-app
• Browser extension
• Server-side (e.g., Node.js, Deno, Bun)
• Desktop app (e.g., Electron)
• Other (e.g., VSCode extension)

Description

Description
We fine-tuned unsloth/gemma-3-4b-it and merged LoRA adapters into the base model in bf16 to prepare for ONNX export.

• HF sources

  • Base: https://huggingface.co/unsloth/gemma-3-4b-it
  • Our merged model: https://huggingface.co/azizbekabdullaev/gemma-3-4b-it

What we tried

• Transformers.js conversion script

  • python -m scripts.convert --model_id azizbekabdullaev/gemma-3-4b-it --quantize

  • Result: conversion fails during ONNX export for Gemma-3.

  • python -m onnxruntime_genai.models.builder -i "C:\models\Gemma3-4B-SFT-merged-bf16" -o "C:\models\gemma3-4B-webgpu" -e webgpu -p int4

  • This produced a WebGPU Q4 package but it does not run in Transformers.js.

Expected
A path to convert Gemma-3-4B-IT to a Transformers.js-compatible ONNX package for WebGPU text-generation.

Reproduction

# Try official conversion
python -m scripts.convert --model_id azizbekabdullaev/gemma-3-4b-it --quantize

# Then attempt to load in JS
import { pipeline } from "@huggingface/transformers";
const pipe = await pipeline("text-generation", "./models/gemma-3-4b-it", {
  device: "webgpu",
  dtype: "q4",
});