[Shreya Boyane (Sonica-B)]

Describe your changes

OnnxStaticQuantization silently ignores reduce_range and nodes_to_quantize on its default path (quant_format="QDQ", prepare_qdq_config=True, non-QNN EP, onnxruntime >= 1.21).

On that path, get_static_run_config calls onnxruntime's get_qdq_config(...) and then uses the members of the returned StaticQuantConfig as the quantize_static arguments. Two user settings get lost on the way:

• reduce_range is a parameter of get_qdq_config, but Olive never passes it, so StaticQuantConfig keeps its default reduce_range=False.
• Neither get_qdq_config nor get_qnn_qdq_config has a nodes_to_quantize parameter, so StaticQuantConfig.nodes_to_quantize stays None and every node of the selected op types is quantized.

With prepare_qdq_config=False both settings work, so the pass does mean to support them. The pass-configuration docs example (docs/source/how-to/configure-workflows/pass-configuration.md) searches over reduce_range for OnnxQuantization; on the default path both search points currently produce the same model.

flowchart LR
    subgraph Before
        A1[pass config<br/>reduce_range, nodes_to_quantize] --> B1["get_qdq_config(...)<br/>no reduce_range passed"]
        B1 --> C1["StaticQuantConfig<br/>reduce_range=False<br/>nodes_to_quantize=None"]
        C1 --> D1[quantize_static]
    end
    subgraph After
        A2[pass config<br/>reduce_range, nodes_to_quantize] --> B2["get_qdq_config(..., reduce_range)"]
        B2 --> C2[StaticQuantConfig members]
        A2 -- "nodes_to_quantize copied" --> C2
        C2 --> D2[quantize_static]
    end

Loading

Changes in olive/passes/onnx/quantization.py:

• pass reduce_range to get_qdq_config;
• on the QNN path, log a warning when reduce_range is set, since get_qnn_qdq_config has no such parameter (QNN behavior is otherwise unchanged);
• copy a user-set nodes_to_quantize onto the final run config, for both the QNN and non-QNN helpers.

Reproduction: a two-MatMul ONNX model (x -> MatMul(W1) -> Relu -> MatMul(W2)), OnnxStaticQuantization with precision=int8, weight_symmetric=True, everything else default. The script prints the largest absolute value in each quantized weight (127 is the full symmetric int8 range, 64 is the 7-bit reduce_range range).

Before (main @ 1fce2a6):

{}                                                           -> W1: int8 max|q|=127, W2: int8 max|q|=127
{'reduce_range': True}                                       -> W1: int8 max|q|=127, W2: int8 max|q|=127
{'reduce_range': True, 'prepare_qdq_config': False}          -> W1: int8 max|q|=64, W2: int8 max|q|=64
{'nodes_to_quantize': ['mm1']}                               -> W1: int8 max|q|=127, W2: int8 max|q|=127
{'nodes_to_quantize': ['mm1'], 'prepare_qdq_config': False}  -> W1: int8 max|q|=127, W2: not quantized (float)

After:

{}                                                           -> W1: int8 max|q|=127, W2: int8 max|q|=127
{'reduce_range': True}                                       -> W1: int8 max|q|=64, W2: int8 max|q|=64
{'reduce_range': True, 'prepare_qdq_config': False}          -> W1: int8 max|q|=64, W2: int8 max|q|=64
{'nodes_to_quantize': ['mm1']}                               -> W1: int8 max|q|=127, W2: not quantized (float)
{'nodes_to_quantize': ['mm1'], 'prepare_qdq_config': False}  -> W1: int8 max|q|=127, W2: not quantized (float)

Tests (test/passes/onnx/test_quantization.py):

• test_nodes_and_ops gains a nodes_to_quantize case for both QNN and non-QNN and checks that nodes_to_quantize reaches quantize_static;
• new test_static_quantization_applies_reduce_range_when_prepare_qdq_config runs real static quantization and checks that the quantized weights span 127 without reduce_range and 64 with it.

On main the three new cases fail (assert set() == {'node_linear'} twice, assert 127 == 64); with the fix the file passes (16 passed). Validated on Windows 11 with CPU torch (Python 3.11.9, onnxruntime 1.31.0, onnx 1.20.1, torch 2.14.1+cpu); test_add_metadata.py, test_split_model.py, test_quantize_embedding.py, test/engine/test_engine.py and test/cli/test_cli.py also pass there. CI will confirm on the other platforms. The QNN EP itself was not run; the QNN path is covered by the mocked test_nodes_and_ops[True-*] cases.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary. (No doc change needed; the documented options now take effect.)
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.

Release note: OnnxStaticQuantization / OnnxQuantization now apply reduce_range and nodes_to_quantize when prepare_qdq_config=True (the default for QDQ). Before this change both were ignored on that path. Configs that set reduce_range: true will now produce 7-bit weights, as documented.

(Optional) Issue link

None; found while reading the static quantization path.

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.