Row: ENG-HYBRID-PLACEMENT
What
The MoE placement seam's entire contract is "placement must never change a
value", and nothing in the tree enforces it byte-for-byte.
The short-circuit that makes an unplaced run identical is at
include/vllm/model_executor/moe_placement_seam.h:103-111 (twin in
RunMoePlacedPair at :187):
if (placed_on == engine_device) {
// THE UNPLACED PATH, and it is the existing call with nothing around it: no
// copy, no allocation, and the value the architecture already produced.
dense_attn::DBuf out = body(engine, dh);Every architecture now routes through it — deepseek_v2.cpp:531,
kimi_linear_device.cpp:936,1231,1616, nemotron_h_device.cpp:1177,1865,
dots3_note_device.cpp:1369, glm_moe_dsa_forward.cpp:473,
qwen3_5.cpp:7398,7683, qwen3_moe.cpp:110, qwen4_exp_forward.cpp:692.
The gate that covered this was deleted
Confirmed from history rather than a failed grep:
tests/vllm/model_executor/test_placed_moe_roundtrip.cpp was deleted at
9164b6cb7 (merge 6416aab85). The commit body is candid — it called
vllm::RunMoeBlockPlaced, which 866075b2f had removed, and main would not
build without deleting it.
Nothing replaced it. The two surviving placement tests are
test_device_placement.cpp and test_placement_dump_dtype.cpp, and the closest
case asserts only a negative (test_device_placement.cpp:433-437):
// (It resolves to kCUDA, which has no backend registered in a CPU build, so
// only assert it is not OUR refusal that fires.)
The placed body never executes on a CPU tier, so no bytes are ever compared.
Why a same-device test cannot substitute
The short-circuit means placed_on == engine_device returns the original value
by construction. A test that does not genuinely cross devices exercises the
branch that cannot differ, and proves nothing about the one that can. This needs
a real cross-device run — a CUDA/ROCm box, or possibly Vulkan via lavapipe
(unconfirmed; nobody has checked whether that target builds here).
Class
Gate. No defect is claimed — the point is that a divergence would currently be
invisible, on a seam every MoE architecture in the tree now depends on.
Row:
ENG-HYBRID-PLACEMENTWhat
The MoE placement seam's entire contract is "placement must never change a
value", and nothing in the tree enforces it byte-for-byte.
The short-circuit that makes an unplaced run identical is at
include/vllm/model_executor/moe_placement_seam.h:103-111(twin inRunMoePlacedPairat:187):Every architecture now routes through it —
deepseek_v2.cpp:531,kimi_linear_device.cpp:936,1231,1616,nemotron_h_device.cpp:1177,1865,dots3_note_device.cpp:1369,glm_moe_dsa_forward.cpp:473,qwen3_5.cpp:7398,7683,qwen3_moe.cpp:110,qwen4_exp_forward.cpp:692.The gate that covered this was deleted
Confirmed from history rather than a failed grep:
tests/vllm/model_executor/test_placed_moe_roundtrip.cppwas deleted at9164b6cb7(merge6416aab85). The commit body is candid — it calledvllm::RunMoeBlockPlaced, which866075b2fhad removed, andmainwould notbuild without deleting it.
Nothing replaced it. The two surviving placement tests are
test_device_placement.cppandtest_placement_dump_dtype.cpp, and the closestcase asserts only a negative (
test_device_placement.cpp:433-437):The placed body never executes on a CPU tier, so no bytes are ever compared.
Why a same-device test cannot substitute
The short-circuit means
placed_on == engine_devicereturns the original valueby construction. A test that does not genuinely cross devices exercises the
branch that cannot differ, and proves nothing about the one that can. This needs
a real cross-device run — a CUDA/ROCm box, or possibly Vulkan via lavapipe
(unconfirmed; nobody has checked whether that target builds here).
Class
Gate. No defect is claimed — the point is that a divergence would currently be
invisible, on a seam every MoE architecture in the tree now depends on.