`smoke_fp16_v0` — CI fp16 I/O round-trip probe¶

Status — CI internal. This is not a quality model and must not be used for video quality assessment. It exists solely as a fp16 cast-path probe for the libvmaf DNN integration CI gate.

smoke_fp16_v0 is a minimal ONNX graph (Identity op, no initializer tensors, opset 17) with fp16 input and output tensors. It exercises the fp16 I/O cast path in the C-side DNN loader — specifically the path that casts float32 activations to fp16 before inference and back to float32 after. No trained weights are involved; the Identity op passes values through unchanged.

Checkpoint facts¶

Field	Value
Model id	`smoke_fp16_v0`
Location	`model/tiny/smoke_fp16_v0.onnx`
Architecture	Identity (fp16 I/O) — intentional CI fp16 cast probe
Trainable parameters	0 (no initializer tensors)
Input	`x` — `[1, 1, 2, 2]` (cast from float32 to fp16 by libvmaf on the host)
Output	`y` — `[1, 1, 2, 2]` (cast back to float32 by libvmaf)
Training	None: an Identity graph with no weights
ONNX opset	17
License	BSD-2-Clause-Patent
Registry entry	`smoke_fp16_v0` in `model/tiny/registry.json` (`"smoke": true`)
SHA-256	`6cbf16be5d2cfb858f1eb60bfdcc9e674c15f17b3ff365afa475bbe9be76258b`

Purpose¶

The libvmaf DNN integration supports models that declare fp16 input/output tensors, with the loader performing cast from float32 → fp16 on input and fp16 → float32 on output. smoke_fp16_v0 verifies that the full round-trip cast path functions correctly:

The loader detects fp16 I/O via the ONNX graph element type.
The cast-to-fp16 step runs without error.
ORT executes the Identity op in fp16.
The cast-from-fp16 step runs and the output is numerically close to the input (within fp16 precision, ~3 ULP for values in the VMAF feature range).

The model is registered with "smoke": true in registry.json so validation scripts know to skip quality-metric assertions.

Evaluation metrics¶

Intentional smoke probe — not evaluated on any quality corpus. PLCC / SROCC / RMSE figures are not applicable. The only gate the model must pass is ORT load-and-run without error, plus round-trip numerical proximity within fp16 precision.

Runnable usage example¶

# Verify the fp16 I/O round-trip via the C unit test suite:
python3 "$(git rev-parse --show-toplevel)/scripts/ci/run_meson_test.py" -- \
  -C build --suite=dnn test_ep_fp16

# Or inspect the model via Python:
python3 -c "import onnxruntime as ort; sess = ort.InferenceSession('model/tiny/smoke_fp16_v0.onnx'); print('Inputs:', [i.name for i in sess.get_inputs()])"

Known limitations / when NOT to use¶

Do not use for video quality assessment of any kind.
fp16 I/O incurs precision loss relative to fp32. For production VMAF fusion models the fork ships fp32-only checkpoints; the fp16 path is reserved for future optimized inference variants.
The "smoke": true registry flag means model-registry validation scripts skip quality assertions.

smoke_fp16_v0 — CI fp16 I/O round-trip probe¶