Summary
The ONNX PTQ entry points expose NVFP4 for transformer-family models but not for convolutional
models. A CNN detector cannot request NVFP4 through the ONNX PTQ API at all — this is an
unreachable code path rather than a quality limitation, so it fails as "unsupported" rather than as
"poor accuracy".
Originally raised as a comment on #362, which is closed; re-filing standalone so it can be tracked.
Why it matters
We benchmark CNN detectors (YOLOv8/11/26-family) and a transformer detector (RT-DETRv2) through the
same export and quantization pipeline, and compare precisions across them. NVFP4 is reachable for
the transformer and not for the CNNs, so the comparison has a hole that is a property of the API
surface rather than of the models.
Environment
ModelOpt 0.46.0, ONNX PTQ path, explicit-QDQ (strongly typed) export.
What would help
Either an NVFP4 config selection path for convolutional graphs, or — if that is out of scope — an
explicit, early error naming NVFP4-on-CNN as unsupported, so callers do not have to discover it by
inspecting realized precisions afterwards.
I am happy to supply a representative CNN detector ONNX and the exact config that fails to select
NVFP4.
Summary
The ONNX PTQ entry points expose NVFP4 for transformer-family models but not for convolutional
models. A CNN detector cannot request NVFP4 through the ONNX PTQ API at all — this is an
unreachable code path rather than a quality limitation, so it fails as "unsupported" rather than as
"poor accuracy".
Originally raised as a comment on #362, which is closed; re-filing standalone so it can be tracked.
Why it matters
We benchmark CNN detectors (YOLOv8/11/26-family) and a transformer detector (RT-DETRv2) through the
same export and quantization pipeline, and compare precisions across them. NVFP4 is reachable for
the transformer and not for the CNNs, so the comparison has a hole that is a property of the API
surface rather than of the models.
Environment
ModelOpt 0.46.0, ONNX PTQ path, explicit-QDQ (strongly typed) export.
What would help
Either an NVFP4 config selection path for convolutional graphs, or — if that is out of scope — an
explicit, early error naming NVFP4-on-CNN as unsupported, so callers do not have to discover it by
inspecting realized precisions afterwards.
I am happy to supply a representative CNN detector ONNX and the exact config that fails to select
NVFP4.