Current origin/main (8b00f79f2) does not compile the parity suite.
tests/parity/hf_snapshot.h:52 — Nemotron35LightningSnapshot() calls HfSnapshot(...), but HfSnapshot is declared 11 lines below it, at :58.
hs.h: In function ‘std::string parity::Nemotron35LightningSnapshot()’:
hs.h:52:10: error: ‘HfSnapshot’ was not declared in this scope
52 | return HfSnapshot("models--nvidia--NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4",
| ^~~~~~~~~~
Reproduced in isolation against the pristine blob — this is the header itself, not any local merge or build configuration:
git show origin/main:tests/parity/hf_snapshot.h > hs.h
printf #include "hs.h"\nint main(){return 0;}\n > t.cpp
g++ -std=c++20 -fsyntax-only t.cpp # fails as above, no vllm.cpp include path involvedIntroduced by af8170154 ("evidence(MODEL-TEXT-nemotron-h): the pinned oracle LOADS AND RUNS Nemotron-3.5-Lightning on GB10 (#517)").
Blast radius: 14 translation units fail to build, every one with this identical error — test_qwen36_weights, test_op_parity, test_qwen36_paged_engine, test_qwen36_async_serving, test_qwen36_spec_decode, the test_qwen27*/test_qwen27n* suites, test_qwen3_dflash_draft_parity, test_qwen3_dflash_kvprep_parity, and test_hf_snapshot_pinning. A serial ctest reports 97% tests passed, 14 tests failed out of 401, where all 14 are ***Not Run (missing executables) rather than assertion failures.
Why this matters beyond the build break: those are the checkpoint-gated parity suites — the SACRED 27B/35B/Coder token gates. Any row currently owing GPU evidence cannot obtain it against main until this is fixed.
Why it was not caught: these TUs are checkpoint-gated, so on a host without the snapshots they skip at runtime; but they still have to compile. A CPU-only or snapshot-less CI leg that builds only the default target would not surface it, and the failure mode on a full build is 14 identical errors that look like an environment problem.
Fix is a one-line ordering change: move Nemotron35LightningSnapshot() below HfSnapshot’s definition (or forward-declare HfSnapshot). Found while merging origin/main into row/MODEL-QWEN38-TEXT-ONLY (#490); not repaired there, since a reviewer/implementer must not silently fix an unrelated defect inside another row.
Current
origin/main(8b00f79f2) does not compile the parity suite.tests/parity/hf_snapshot.h:52—Nemotron35LightningSnapshot()callsHfSnapshot(...), butHfSnapshotis declared 11 lines below it, at:58.Reproduced in isolation against the pristine blob — this is the header itself, not any local merge or build configuration:
Introduced by
af8170154("evidence(MODEL-TEXT-nemotron-h): the pinned oracle LOADS AND RUNS Nemotron-3.5-Lightning on GB10 (#517)").Blast radius: 14 translation units fail to build, every one with this identical error —
test_qwen36_weights,test_op_parity,test_qwen36_paged_engine,test_qwen36_async_serving,test_qwen36_spec_decode, thetest_qwen27*/test_qwen27n*suites,test_qwen3_dflash_draft_parity,test_qwen3_dflash_kvprep_parity, andtest_hf_snapshot_pinning. A serialctestreports97% tests passed, 14 tests failed out of 401, where all 14 are***Not Run(missing executables) rather than assertion failures.Why this matters beyond the build break: those are the checkpoint-gated parity suites — the SACRED 27B/35B/Coder token gates. Any row currently owing GPU evidence cannot obtain it against
mainuntil this is fixed.Why it was not caught: these TUs are checkpoint-gated, so on a host without the snapshots they skip at runtime; but they still have to compile. A CPU-only or snapshot-less CI leg that builds only the default target would not surface it, and the failure mode on a full build is 14 identical errors that look like an environment problem.
Fix is a one-line ordering change: move
Nemotron35LightningSnapshot()belowHfSnapshot’s definition (or forward-declareHfSnapshot). Found while mergingorigin/mainintorow/MODEL-QWEN38-TEXT-ONLY(#490); not repaired there, since a reviewer/implementer must not silently fix an unrelated defect inside another row.