Row: UPSTREAM-SYNC-HEADPIN
Successor to #2593, which
closed when #2594 merged. That
wave measured the install-and-import half at
e126687a9a828d513c01a07cd69f025f27d63280 and left three things owed. They are
listed under ## Owed in .agents/specs/upstream-sync-headpin.md, which still
cites the now-closed #2593; this issue is the live owner.
Report: .agents/sync/2026-09-02-e126687.md.
1. The run half of gateability
The one that matters, because it is what keeps gateable = no. #2594 measured
RUNDEPS_RC=0 BUILD_RC=0 IMPORT_RC=0 and Qwen4ExpForCausalLM registered, and
said in as many words that this is not the run half: no weights were loaded, no
forward pass ran, no token was compared.
It is now reachable rather than blocked. thor:gpu0's container exposes the
device (CUDA_AVAIL=True, nvidia-smi answering NVIDIA Thor, 595.78), against
orin's CUDA_AVAIL=False, and nvidia-cuda-nvcc 13.3.73 installs as an
aarch64 wheel inside requirements/cuda.txt at
site-packages/nvidia/cu13/bin/nvcc. See §5.5 and §5.7.
Needs a source build: every #2594 job ran VLLM_USE_PRECOMPILED=1, which on
aarch64 yields an editable wheel and no vllm._C (EXT_RC=1). Locating the
compiler is not compiling with it.
2. The PORT-NOW queue for 5559679229..e126687a9a
290 entries, against #2589's 206 over its 1026-commit prefix. Estimated ~51 real
gaps by #2524's sample rate — an estimate from a sample, not a count. Nine of
#2524's eleven worked real gaps are inside this range, so its §13 is most of the
queue already. See §4 and §6.
This is not a licence to advance the pin. The queue has to be worked first,
and the run half has to be measured.
3. Whether upstream main installs under the same repair
Deliberately unmeasured; #2594 spent one lease, on the revision that serves the
row. Answered for RUNDEPS only, and by source identity rather than
measurement: requirements/cuda.txt is byte-identical at the target and at both
head revisions read, and the common.txt it includes differs by one mcp
version range. That argument does not transfer to BUILD or IMPORT. See §3.1.
Row:
UPSTREAM-SYNC-HEADPINSuccessor to #2593, which
closed when #2594 merged. That
wave measured the install-and-import half at
e126687a9a828d513c01a07cd69f025f27d63280and left three things owed. They arelisted under
## Owedin.agents/specs/upstream-sync-headpin.md, which stillcites the now-closed #2593; this issue is the live owner.
Report:
.agents/sync/2026-09-02-e126687.md.1. The run half of gateability
The one that matters, because it is what keeps
gateable = no. #2594 measuredRUNDEPS_RC=0 BUILD_RC=0 IMPORT_RC=0andQwen4ExpForCausalLMregistered, andsaid in as many words that this is not the run half: no weights were loaded, no
forward pass ran, no token was compared.
It is now reachable rather than blocked.
thor:gpu0's container exposes thedevice (
CUDA_AVAIL=True,nvidia-smiansweringNVIDIA Thor, 595.78), againstorin'sCUDA_AVAIL=False, andnvidia-cuda-nvcc 13.3.73installs as anaarch64 wheel inside
requirements/cuda.txtatsite-packages/nvidia/cu13/bin/nvcc. See §5.5 and §5.7.Needs a source build: every #2594 job ran
VLLM_USE_PRECOMPILED=1, which onaarch64 yields an editable wheel and no
vllm._C(EXT_RC=1). Locating thecompiler is not compiling with it.
2. The PORT-NOW queue for
5559679229..e126687a9a290 entries, against #2589's 206 over its 1026-commit prefix. Estimated ~51 real
gaps by #2524's sample rate — an estimate from a sample, not a count. Nine of
#2524's eleven worked real gaps are inside this range, so its §13 is most of the
queue already. See §4 and §6.
This is not a licence to advance the pin. The queue has to be worked first,
and the run half has to be measured.
3. Whether upstream
maininstalls under the same repairDeliberately unmeasured; #2594 spent one lease, on the revision that serves the
row. Answered for
RUNDEPSonly, and by source identity rather thanmeasurement:
requirements/cuda.txtis byte-identical at the target and at bothhead revisions read, and the
common.txtit includes differs by onemcpversion range. That argument does not transfer to
BUILDorIMPORT. See §3.1.