Fix incorrect output for QDQ Resize with non-nearest interpolation modes under ORT_ENABLE_ALL

[Copilot]

Description

• qdq_selectors.cc: In DropQDQNodeGroupSelector::Check, add a guard that skips the QDQ-drop optimization when the target node is Resize with a non-nearest mode attribute ("linear", "cubic", etc.). When mode is absent the ONNX default is "nearest", so the optimization remains allowed in that case.
• qdq_transformer_test.cc: Add Resize_Linear_No_QDQ_Drop test asserting that DQ/Q nodes are preserved around Resize for linear and cubic modes after Level2 optimization.

Motivation and Context

The DropQDQ optimizer incorrectly dropped the surrounding DQ/Q nodes for all Resize modes. Dropping these nodes bypasses float-space interpolation — Resize ends up operating directly on raw quantized integers, producing wrong results.

The optimization is only semantically valid for mode="nearest", which copies existing quantized values without arithmetic. For interpolating modes, the full DQ → Resize → Q chain must execute in float space:

# With ORT_DISABLE_ALL (correct):   DQ → float → Resize(linear) → Q  ✓
# With ORT_ENABLE_ALL (before fix): Resize(linear) on raw int8       ✗

Input [2.0, 7.0, 4.0] upsampled 1×1×1×3 → 1×1×1×6 with half_pixel:

• Expected / DISABLE_ALL: [2.0, 3.0, 6.0, 6.0, 5.0, 4.0]
• ENABLE_ALL before fix: [2.0, 3.0, 5.0, 6.0, 4.0, 4.0]

[titaiwangms]

Review — QDQ drop guard for non-nearest Resize

Verdict: Fix is correct and well-tested. One substantive caveat worth confirming; the rest are polish. The change correctly prevents DropQDQ from removing DQ/Q around interpolating Resize (linear/cubic), which must run in float space. Registration wiring, the inverse test, and .s() usage were verified against the repo.

🟠 Major (worth addressing / confirming scope)

Nearest is not a pure value copy under coordinate_transformation_mode="tf_crop_and_resize".

• upsample.cc:1110 passes extrapolation_value_ into UpsampleNearest; it fires when the coordinate mode is tf_crop_and_resize (upsamplebase.h:255). Out-of-crop positions get static_cast<T>(extrapolation_value) written raw into the quantized domain, but extrapolation_value is authored in the dequantized float domain. The correct stored value would be quantize(extrapolation_value). With the default extrapolation_value=0.0 and a nonzero zero_point, the dropped graph writes 0 where it should write zero_point → observable mismatch.
• This is pre-existing (the PR doesn't introduce it) and narrow, but it falsifies the new comment's claim that nearest "copies existing values." Recommendation: either also bail when coordinate_transformation_mode == "tf_crop_and_resize", or soften the comment so it doesn't assert an unconditional pure copy.
• For completeness: antialias and exclude_outside do not affect the nearest path, so the guard is complete with respect to those.

🟡 Minor

• Missing test for omitted mode. The fix relies on absent-mode ⇒ nearest ⇒ drop allowed, but no test builds Resize without a mode attribute asserting QDQ is dropped. Worth locking in the fallback semantics.
• Comment slightly misleading on the absence branch. After Graph::Resolve(), schema defaults are injected (graph.cc:3588-3598), so mode is always present at runtime — the != end clause does the real work and the absence branch is a defensive fallback for unresolved/hand-built graphs. Correct either way; just clarify the comment.
• Structural. This adds an op-specific OpType() == "Resize" guard inside the generic "data-preserving ops" DropQDQNodeGroupSelector. Acceptable as a targeted patch, but if per-op exceptions accumulate, consider a dedicated selector or a per-op predicate table near the DropQDQ registration.
• Naming. mode_attr_it breaks this file's _iter convention (e.g., block_size_iter, a_iter); mode_iter is the local idiom.

🔵 Nits

• Test name Resize_Linear_No_QDQ_Drop also covers cubic → consider Resize_NonNearest_No_QDQ_Drop.
• The inline comment largely duplicates the block comment above it; the only new info is the default-when-absent behavior.
• Magic "" for the nearest_mode argument (vs. the helper's declared default), and a minor sibling-test naming inconsistency (sizes_shape vs sizes_data).

✅ Praise

• Correct math distinction (value-copy vs. float interpolation), fails closed for non-nearest modes, and leaves the co-registered DepthToSpace selector path untouched.
• The new test is a genuine inverse of the existing Resize test and covers linear/cubic × contrib/non-contrib QDQ.

_{Reviewed with a multi-model agent team (readability, code, adversarial, and deep spec reviewers).}

[titaiwangms]

Non-blocking:

The new tf_crop_and_resize guard branch is untested (code + critical reviewers). The shared helper BuildQDQResizeTestCase always builds an empty {0} roi (qdq_test_utils.h:303), and a valid tf_crop_and_resize requires a 2*rank roi (upsample.cc:1091), so the new branch can't be reached through it. A regression deleting/misspelling that condition wouldn't be caught. Fix: a dedicated test builder supplying a valid roi + mode="nearest" + tf_crop_and_resize, asserting Q/DQ preserved. (The author honestly documented this gap in a test comment.)