Pre-existing CUDA ctest failures with no issue: test_internlm2_paged_engine (and test_ops_gdn is only tentatively #305)

[localai-bot]

Measured as a BEFORE/AFTER pair on dgx.casa (GB10 sm_121a, CUDA 13.0.88) while gating #516: two clean builds from the same base, back to back, ctest -j 1, same box and lock. The BEFORE arm carries no part of that change.

test_internlm2_paged_engine fails identically in both arms, so it is pre-existing and unowned. A search over open and closed issues returns nothing for it, and #248 -- which groups four unowned CUDA failures including three paged-engine ones -- does not list it.

test	BEFORE (base, unmodified)	AFTER	tracked
test_internlm2_paged_engine	Failed 5.02 s	Failed 4.97 s	nothing
test_ops_gdn	Failed 2.52 s	Failed 2.40 s	#305, tentatively

Why test_ops_gdn is listed but not claimed

#305 is a conv_state initial-state/final-state race across blocks. The failure
seen here was recorded in .agents/specs/pool-device-key.md §10 as
:728 CHECK( bad == 0 ) -> 2609 == 0, a numeric check that may or may not be
that race. Someone who owns the GDN kernels should confirm or split it; this
issue exists so the tentative attribution is visible rather than assumed.

What is NOT established

Neither arm's LastTest.log survives: both 34 GB build trees were deleted to
keep a 98%-full filesystem under control, and the runs were plain ctest -j 1
without --output-on-failure. So the recorded evidence is the pass/fail, the
index and the duration -- not the assertion text. Re-running with
--output-on-failure is the first step for whoever picks this up.

Reproduce

Clean CUDA build on sm_121a (-DVLLM_CPP_CUDA=ON -DVLLM_CPP_CUDA_ARCHITECTURES=121a -DVLLM_CPP_CUTLASS_DIR=$HOME/cutlass-4.5.0 -DVLLM_CPP_TRITON=ON), then:

ctest -j 1 -R 'test_internlm2_paged_engine|test_ops_gdn' --output-on-failure

Use -j 1: GB10 unified memory stacks concurrent model gates into host RAM and -j 4 has OOM-rebooted this box.

Related

• SACRED gate red on main: test_glm4_moe_lite_paged_engine diverges in tokens (plus 4 more pre-existing ctest failures) #233 -- test_glm4_moe_lite_paged_engine, test_linear_method, test_serve_low_tools, test_gemma4_registry_e2e, test_minimax_h3
• Four pre-existing CUDA ctest failures on main, attributed but unowned: test_capi (SIGSEGV), test_qwen3_apc_e2e, test_minicpm3_paged_engine, test_llama_paged_engine #248 -- test_capi (SIGSEGV), test_qwen3_apc_e2e, test_minicpm3_paged_engine, test_llama_paged_engine
• test_minimax_h3 is RED on dgx (GB10): cudaFree invalid argument + SIGSEGV when two CUDA cases run in one process #486 / vllm::Pool() free list is keyed by size class with no DEVICE in the key: a cudaMalloc'd block can be handed to a CPU DBuf #516 -- test_minimax_h3 is the one test that CHANGES across this pair (SEGFAULT -> Passed); it is fixed by the device-keyed scratch pool and is not part of this issue.

Found while gating #516; not caused by it.