support compressed-tensors refactor

[xin3he]

Description

compressed_tensors PR vllm-project/compressed-tensors#610 (commit 927f6d5, Mar 17 2026) removed CompressedLinear as a functional class. The class is now a stub that raises ValueError from from_linear().

Models loaded with the new compressed_tensors produce regular torch.nn.Linear modules with quantization_scheme/quantization_status attributes instead of CompressedLinear wrappers.

Type of Change

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

Related Issues

Fixes or relates to #1578

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

Updates AutoRound’s compressed_tensors integration to handle the upstream removal of functional CompressedLinear wrappers by detecting/processing quantized torch.nn.Linear modules that carry quantization_scheme / quantization_status metadata.

Changes:

• Updated weight-type handlers to detect quantized modules via quantization_scheme (new compressed_tensors behavior) in addition to legacy CompressedLinear/compressor logic.
• Added decompress_module(...) handling for new-style quantized Linear modules in FP8 and NVFP4 conversions.
• Re-enabled/adjusted CPU tests to validate FP8/MXFP4 models without relying on CompressedLinear type checks (NVFP4 still skipped due to upstream issue).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 5 comments.

File	Description
auto_round/utils/weight_handler.py	Extends detection/conversion to support new compressed_tensors quantized Linear modules.
test/test_cpu/advanced/test_low_precision_input_model.py	Updates assertions to match new model module types/attributes and unskips FP8/MXFP4 coverage.

[yiliu30]

Code review

Found 1 issue:

1. FP8Handler.detect_layer uses .data_type (lines 497, 499) to access quantization scheme fields, but compressed_tensors QuantizationArgs only has a .type field. The other three handlers (MXFP4, MXFP8, NVFP4) in this same PR correctly use .type. This will raise AttributeError at runtime when detecting FP8-BLOCK models via the new quantization_scheme path.

auto-round/auto_round/utils/weight_handler.py

Lines 496 to 500 in e99111f

	q_scheme.weights.num_bits == 8
	and "float" in q_scheme.weights.data_type
	and q_scheme.input_activations.num_bits == 8
	and "float" in q_scheme.input_activations.data_type
	):

Also, is the new quantization_scheme-based detection compatible with the older compressed_tensors versions that still use CompressedLinear? Want to make sure we don't break backward compatibility.

Otherwise the approach looks good -- approve once the above is addressed.

[yiliu30]

Approach looks good. One bug to fix (.data_type -> .type in FP8Handler) and a question about backward compat with older compressed_tensors versions -- see comment.

[chensuyue]

Remove the version limit here to test your code,

auto-round/test/test_cpu/requirements.txt

Line 13 in 88824a1

compressed-tensors==0.14.1a20260313 # temporary pin for llmcompressor

[xin3he]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[xin3he]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[xin3he]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).