Found on the windows-msvc-cpu job of PR #1821, whose own change touches four
files and none of them is this one. The break is inherited from main.
What is red
dots3_note.cpp(606,31): error C2220: the following warning is treated as an error
dots3_note.cpp(606,31): warning C4646: function declared with 'noreturn' has non-void return type
src/vllm/model_executor/models/dots3_note.h:282 declares
[[noreturn]] static ForwardLogits ForwardDevice(
[[noreturn]] on a function whose return type is not void. GCC and Clang
accept it silently, MSVC emits C4646, and this build treats warnings as errors,
so the whole vllm project fails to compile on Windows. build-test-cpu,
build-test-cpu-arm64, build-newest-gcc and verify (cpu) are all green on
the same commit, which is why nothing on the POSIX side saw it.
Why it landed unseen
windows-msvc-cpu and windows-msvc-vulkan are PULL-REQUEST-ONLY jobs with no
main baseline, so main never receives a Windows verdict. The declaration
arrived with 849a7dd73
(#1805, row
MODEL-MM-dots3-note, lane #699), and that commit is an ancestor of
af320abb2. Every pull request branched from main after it now carries the
same red, and each one has to prove the red is not its own before it can
proceed.
This is NOT #584. #584 is the runtime exit -1073740791
(STATUS_STACK_BUFFER_OVERRUN) in test_openai_api_server.exe; this is a
compile error, and a reader who stops at the job name will wave a real break
through as the known one.
Owner
Row MODEL-MM-dots3-note, lane #699. Not repaired by #1821: the fix is a
semantic decision about a declaration in an actively-developed file that
another row owns -- ForwardDevice overrides into a registry hook and cannot
simply become void, so the choice is between dropping the attribute and
restructuring the refusal, and it needs the row that wrote it. #1821 also has
no MSVC to verify a fix against.
tests/vllm/models/test_kimi_k3_*.cpp's KimiK3Model::ForwardDevice is named
in dots3_note.h:276 as the pattern this mirrors; whoever fixes this should
check whether that one carries the same declaration.
Found on the
windows-msvc-cpujob of PR #1821, whose own change touches fourfiles and none of them is this one. The break is inherited from
main.What is red
src/vllm/model_executor/models/dots3_note.h:282declares[[noreturn]]on a function whose return type is notvoid. GCC and Clangaccept it silently, MSVC emits C4646, and this build treats warnings as errors,
so the whole
vllmproject fails to compile on Windows.build-test-cpu,build-test-cpu-arm64,build-newest-gccandverify (cpu)are all green onthe same commit, which is why nothing on the POSIX side saw it.
Why it landed unseen
windows-msvc-cpuandwindows-msvc-vulkanare PULL-REQUEST-ONLY jobs with nomainbaseline, somainnever receives a Windows verdict. The declarationarrived with
849a7dd73(#1805, row
MODEL-MM-dots3-note, lane #699), and that commit is an ancestor ofaf320abb2. Every pull request branched frommainafter it now carries thesame red, and each one has to prove the red is not its own before it can
proceed.
This is NOT #584. #584 is the runtime
exit -1073740791(
STATUS_STACK_BUFFER_OVERRUN) intest_openai_api_server.exe; this is acompile error, and a reader who stops at the job name will wave a real break
through as the known one.
Owner
Row
MODEL-MM-dots3-note, lane #699. Not repaired by #1821: the fix is asemantic decision about a declaration in an actively-developed file that
another row owns --
ForwardDeviceoverrides into a registry hook and cannotsimply become
void, so the choice is between dropping the attribute andrestructuring the refusal, and it needs the row that wrote it. #1821 also has
no MSVC to verify a fix against.
tests/vllm/models/test_kimi_k3_*.cpp'sKimiK3Model::ForwardDeviceis namedin
dots3_note.h:276as the pattern this mirrors; whoever fixes this shouldcheck whether that one carries the same declaration.