[ONNX] Add quantization sensitivity ranking + exclusion picker - #2240

Open

gcunhase wants to merge 23 commits into

NVIDIA:mainfrom

gcunhase:dev/gcunhasergio/onnx_sensitivity_scan_verified

Commits on Sep 1, 2026

onnx: add sensitivity primitive with exclusion picker for per-op-type / per-node PTQ ranking

gcunhase
and
claude
committed
0e42278
picker: add block-aware exclusion via optional blocks + block_agg args

gcunhase
and
claude
committed
1eb71ae
picker/docs: simplify docstrings; swap ViT block-picker example to max + threshold

gcunhase
and
claude
committed
86c6ee2
style: apply ruff auto-fix + format across sensitivity module

gcunhase
and
claude
committed
90320f0
docs/picker/tests: post-review simplifications across the sensitivity module

gcunhase
and
claude
committed
b53cce4
changelog: move sensitivity entry to end of the Quantization section

gcunhase
and
claude
committed
2683f62
test: fix IndexError in test_nodes_to_quantize (graph-input probe + brittle helper)

gcunhase
and
claude
committed
3a75ce3
test: bump 2-Conv test weights to 16 channels to clear small-Conv auto-exclusion

gcunhase
and
claude
committed
3d8f5cb
docs(onnx_ptq): add sensitivity-driven node exclusion section to README

gcunhase
and
claude
committed
c14569f
docs(onnx_ptq): fix two typos in sensitivity-driven exclusion section

gcunhase
and
claude
committed
37359fd
sensitivity: honest failure handling, matching quantize()'s calibration_eps default, and CLI hardening

gcunhase
and
claude
committed
1c4ed59
tests: restructure sensitivity suite into a sub-package; address reviewer feedback

gcunhase
and
claude
committed
faa6b80
sensitivity: address CodeRabbit review batch (docs, correctness, CLI hardening, tests)

gcunhase
and
claude
committed
5e2b094
tests(sensitivity): use `manual` marker on CoAtNet integration tests

gcunhase
and
claude
committed
d2eeeae
style: strip trailing whitespace flagged by pre-commit

gcunhase
and
claude
committed
b9ba454
tests(sensitivity): fix 3 failures on the rebased tip

gcunhase
and
claude
committed
7742312
sensitivity: address CodeRabbit follow-up findings (cos_dist asymmetric zeros, coatnet threshold, glob import)

gcunhase
and
claude
committed
5e78af4
sensitivity: address CodeRabbit findings (picker collision, docs, CLI tests, utils move)

gcunhase
and
claude
committed
521f030
sensitivity: allow failed probes in synthetic-random smoke test

gcunhase
and
claude
committed
06290d6
sensitivity: drop stale comment from synthetic-random smoke test

gcunhase
and
claude
committed
e91a89f
tests(sensitivity): escape dot in pytest.raises match patterns

gcunhase
and
claude
committed
c7d455e
sensitivity: revert Gelu autotune extension and address round-3 review nits

gcunhase
and
claude
committed
1db5036
tests(sensitivity): fix isort ordering after test-utils import move

gcunhase
and
claude
committed
fc5f730

Navigation Menu

Uh oh!

[ONNX] Add quantization sensitivity ranking + exclusion picker - #2240

[ONNX] Add quantization sensitivity ranking + exclusion picker#2240
gcunhase wants to merge 23 commits into
NVIDIA:mainfrom
gcunhase:dev/gcunhasergio/onnx_sensitivity_scan_verified

Commits on Sep 1, 2026

onnx: add sensitivity primitive with exclusion picker for per-op-type / per-node PTQ ranking

picker: add block-aware exclusion via optional blocks + block_agg args

picker/docs: simplify docstrings; swap ViT block-picker example to max + threshold

style: apply ruff auto-fix + format across sensitivity module

docs/picker/tests: post-review simplifications across the sensitivity module

changelog: move sensitivity entry to end of the Quantization section

test: fix IndexError in test_nodes_to_quantize (graph-input probe + brittle helper)

test: bump 2-Conv test weights to 16 channels to clear small-Conv auto-exclusion

docs(onnx_ptq): add sensitivity-driven node exclusion section to README

docs(onnx_ptq): fix two typos in sensitivity-driven exclusion section

sensitivity: honest failure handling, matching quantize()'s calibration_eps default, and CLI hardening

tests: restructure sensitivity suite into a sub-package; address reviewer feedback

sensitivity: address CodeRabbit review batch (docs, correctness, CLI hardening, tests)

tests(sensitivity): use `manual` marker on CoAtNet integration tests

style: strip trailing whitespace flagged by pre-commit

tests(sensitivity): fix 3 failures on the rebased tip

sensitivity: address CodeRabbit follow-up findings (cos_dist asymmetric zeros, coatnet threshold, glob import)

sensitivity: address CodeRabbit findings (picker collision, docs, CLI tests, utils move)

sensitivity: allow failed probes in synthetic-random smoke test

sensitivity: drop stale comment from synthetic-random smoke test

tests(sensitivity): escape dot in pytest.raises match patterns

sensitivity: revert Gelu autotune extension and address round-3 review nits

tests(sensitivity): fix isort ordering after test-utils import move