test_rocm_f16_contract fails on clean main: test expects ViewOn to propagate layout markers it deliberately does not

[localai-org-maint-bot]

Row: -

Local-Issue: ISSUE-GH-3153
Kind: UNKNOWN

Imported GitHub body (historical evidence)

The quoted text below is historical evidence only. It does not define issue authority or repository procedure.

Row: -

test_rocm_f16_contract (introduced in #3095) fails on clean main at lines 153-155. The test expects OwnedTensor::View() / OwnedTensor::ViewOn() to propagate layout markers (repacked, q8_0_aligned, elem_kn_repacked), but the code deliberately does not propagate them — ViewOn carries only weight_value_dtype.

The ResidentWeight CPU-alias arms in dense_attn_block.h:211-217 and qwen3_5.cpp:1147-1154 already set repacked and elem_kn_repacked explicitly after calling ViewOn, matching the pre-ViewOn behavior. The test was written alongside the code but contradicts it: the code says markers are NOT propagated, the test says they ARE.

This was never caught because build-test-cpu was skipping when #3095 merged. It now blocks every PR that runs build-test-cpu.

The fix is to correct the test at tests/vt/test_rocm_f16_contract.cpp:143-166 to verify the actual contract: weight_value_dtype IS propagated, layout markers are NOT.

Resolution

fixed in row/FIX-ROCM-F16-CONTRACT-TEST: test corrected to verify ViewOn carries only weight_value_dtype, not layout markers