Add autoround xpu ut

[Kaihui-intel]

User description

Type of Change

UT

Description

for BMG
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu
If encountering ipex error: pip uninstall intel_extension_for_pytorch
or use nightly
pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu
or
pip3 install torch==2.8 torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu

Expected Behavior & Potential Risk

the expected behavior that triggered by this PR

How has this PR been tested?

how to reproduce the test (including hardware information)

Dependency Change?

any library dependency introduced or removed

---

PR Type

Tests

---

Description

• Add XPU unit tests for autoround

• Test various schemes and formats

• Include VLM and LM head quantization tests

---

Diagram Walkthrough

flowchart LR
  A["Add XPU UT"] -- "Test schemes" --> B["Test formats"]
  B -- "Test VLM models" --> C["Test LM head quantization"]

Loading

File Walkthrough

	Relevant files
Tests	test_autoround.py Add XPU autoround unit tests                                                          test/3x/torch/quantization/weight_only/test_autoround.py Added XPU unit tests for autoround Included tests for different schemes and formats Added tests for VLM models and LM head quantization +161/-0

---

[PRAgent4INC]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 PR contains tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Possible Issue Ensure that all test cases cover edge cases and invalid inputs to prevent unexpected behavior. import copy import shutil import pytest import torch import transformers from packaging.version import Version, parse import os from functools import lru_cache @lru_cache(None) def is_habana_framework_installed(): """Check if Habana framework is installed. Only check for the habana_frameworks package without importing it to avoid initializing lazy-mode-related components. """ from importlib.util import find_spec package_spec = find_spec("habana_frameworks") return package_spec is not None def set_hpu_torch_compile_envs(): if not is_habana_framework_installed(): return None import torch._dynamo.config as dynamo_config import torch._inductor.config as inductor_config os.environ["PT_HPU_LAZY_MODE"] = "0" os.environ["PT_ENABLE_INT64_SUPPORT"] = "1" inductor_config.force_disable_caches = True dynamo_config.inline_inbuilt_nn_modules = True # The `TestAutoRoundHPU` is expected to be run with `compile` mode, # so set the HPU environment variables before importing INC. if is_habana_framework_installed(): set_hpu_torch_compile_envs() from neural_compressor.torch.quantization import ( AutoRoundConfig, convert, get_default_AutoRound_config, prepare, quantize, ) from neural_compressor.torch.utils import logger torch.backends.__allow_nonbracketed_mutation_flag = True try: import auto_round from auto_round.export.export_to_itrex.model_wrapper import WeightOnlyLinear auto_round_installed = True except ImportError: auto_round_installed = False try: import compressed_tensors ct_installed = True except ImportError: ct_installed = False @torch.no_grad() def run_fn(model, dataloader): for data in dataloader: if isinstance(data, tuple) or isinstance(data, list): model(*data) elif isinstance(data, dict): model(**data) else: model(data) @pytest.mark.skipif(is_habana_framework_installed(), reason="These tests are not supported on HPU for now.") @pytest.mark.skipif(not auto_round_installed, reason="auto_round module is not installed") class TestAutoRoundCPU: @classmethod def setup_class(self): self.gptj = transformers.AutoModelForCausalLM.from_pretrained( "hf-internal-testing/tiny-random-GPTJForCausalLM", torchscript=True, ) self.inp = torch.ones([1, 10], dtype=torch.long) self.tokenizer = transformers.AutoTokenizer.from_pretrained( "hf-internal-testing/tiny-random-GPTJForCausalLM", trust_remote_code=True ) from neural_compressor.torch.algorithms.weight_only.autoround import get_dataloader self.dataloader = get_dataloader(self.tokenizer, 32, dataset_name="NeelNanda/pile-10k", seed=42, bs=8, nsamples=10) self.label = self.gptj(self.inp)[0] @classmethod def teardown_class(self): shutil.rmtree("saved_results", ignore_errors=True) def setup_method(self, method):

[PRAgent4INC]

PR Code Suggestions ✨

[chensuyue]

only skip when no xpu detected.