Add MLX format export support for Apple Silicon and support vlm in AutoScheme

[wenhuach21]

Description

test qwen3.5-4b, qwen3-0.6b. As there is no devices to test, so the oob support may not enough

Type of Change

• Bug fix
• New feature
• Documentation update
• Performance improvement
• Code refactoring
• Other (please specify):

Related Issues

Fixes or relates to #

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds MLX export + Apple Silicon inference support by introducing an MLX backend, an MLX-format exporter, and MLX-specific quantized linear layers, plus tests covering export/inference behavior (including mixed-bit and VLM config validation).

Changes:

• Introduces mlx output format and MLX exporter (export_to_mlx) that writes MLX-compatible config.json quantization blocks.
• Adds MLX inference backend and QuantLinearMLX (including GPTQ→MLX post-init repacking on macOS).
• Adds pytest coverage for MLX/native + auto_round flows and helper paths for Qwen models.

Reviewed changes

Copilot reviewed 13 out of 14 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
auto_round/export/export_to_mlx/export.py	Implements MLX packing + config.json generation (mixed-bit + VLM handling).
auto_round/formats.py	Registers mlx / auto_round:mlx formats and routes saving through MLX exporter.
auto_round/inference/backend.py	Adds mlx backend and OS-based backend filtering; adjusts backend requirements.
auto_round/inference/convert_model.py	Adds MPS device support and macOS GPTQ→MLX post-init conversion.
auto_round_extension/mlx/qlinear_mlx.py	Adds QuantLinearMLX with MLX-kernel forward + GPTQ→MLX repacking logic.
auto_round_extension/torch/qlinear_mlx.py	Adds backward-compat shim to the new MLX module location.
auto_round/utils/common.py	Extends supported format list with mlx and auto_round:mlx.
auto_round/schemes.py	Adds W5A16 and W6A16 preset schemes.
test/test_mlx/test_mlx_format.py	Adds comprehensive MLX export/inference pytest suite (incl. mixed-bit + VLM config assertions).
test/helpers.py	Adds a new helper path variable for a Qwen3 VL 9B model.
test_mlx_export.py	Adds a standalone MLX export test script.

Comments suppressed due to low confidence (1)

auto_round/inference/backend.py:1

• The auto_awq:gemm backend no longer declares its dependency requirement, but dynamic_import_inference_linear(...) still imports awq.modules.linear.WQLinear_GEMM for AWQ backends. Without requirements=[\"autoawq\"] (or the correct package name used in your environment), backend selection may succeed and then fail at runtime with an ImportError. Re-add an explicit requirement for the AWQ package so compatibility checks prevent selecting this backend when the dependency is missing.

# Copyright (c) 2024 Intel Corporation

[wenhuach21]

visual module's grad is 0 in autoscheme