add aitk webgpu

[xieofxie]

need patch utf-8 ?

        import builtins
        import os
        import olive
        import olive.workflows

        # PYTHONUTF8=1 is not honoured by the Python-WebGPU launcher because
        # the env var must be read by the actual CPython binary at startup, and
        # the WebGPU runtime delegates to a uv-managed CPython that never sees it.
        # Patch builtins.open for the duration of the Olive call so any text file
        # opened without an explicit encoding (e.g. chat_template.jinja in
        # transformers) uses UTF-8 instead of the Windows default (cp1252).
        _original_open = builtins.open

        def _utf8_open(file, mode="r", *args, **kwargs):
            if "b" not in mode and "encoding" not in kwargs:
                kwargs["encoding"] = "utf-8"
            return _original_open(file, mode, *args, **kwargs)

        builtins.open = _utf8_open
        try:
            output = olive.workflows.run(oliveJson)
        finally:
            builtins.open = _original_open

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(path)

will fail for ds

need patch model ?

will fail for ds, llama, qwen

model = og.Model(model_folder)
      6 tokenizer = og.Tokenizer(model)
      7 tokenizer_stream = tokenizer.create_stream()
      8 

RuntimeError: Node (/model/layers.2/attn/o_proj/MatMulNBits) Op (MatMulNBits) [ShapeInferenceError] Incompatible dimensions for matrix multiplication