b70-optimization-lab

W1 N128 NVMe recovery: XCCL log-framing preflight abort

The fourth local-NVMe recovery root is closed as an infrastructure-only abort:

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/w1-n128-nvme-postreboot-recovery-20260723T140038Z

It ran on clean boot 0b7f98a5-e50a-46a5-81ea-15938b55317a, kernel 7.0.0-28-generic, with taint 0. The gate verified all 118 local model files, four-card discovery/mapping and strict idle, four oneAPI Level Zero devices, and the exact four-device peer read.

The first XCCL command also exited zero. All four sequential device checks reported device_count 1 and ok 2097152.0. Each rank emitted exactly one init marker, one barrier marker, and one exact allreduce ok 4.0 marker. However, torchrun multiplexed multiple rank writes onto the same physical lines:

rank 0 init okrank 1 init okrank 3 init okrank 2 init ok

The frozen parser required each marker to occupy a complete physical line, so it rejected the valid zero-exit pass and failed closed with status 1. The second XCCL pass, both N64 gates, N128, a model service, and model generation did not run. Kernel delta contained only -- No entries --; the Xe reject file was empty and kernel taint remained zero.

The evidence manifest verifies completely. Stable hashes are:

evidence.sha256          7907353bc59879be89cfd719cfda537e011accb12a7365b439b042b29520866e
final-status.txt         be629cb74eb5804b8933e0e710f2374a0a20155c2725b56ac36fba089b767b9c
sycl-ls-verbose.log      a6c004432418cea7c58b444c44ccfc1d4c9d4ae9acbf39f6bf50ce2ee875c617
peer-read.log            8e4d7110a7999cc510c69d352aa6a69553d6f22d1f20f116eb0e8cf8a6b35752
xccl-pass-1.log          1fb7f3522b8b85c982563dc07dd35b213ec302203c767dfce02aef4f59f68fcd
kernel-delta.txt         19c4e28b2f54ea8f217db4e622b0a5c758efa7b77ce2dc0394b6d159568c0b2e
kernel-reject-events.txt e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855

The correction changes only log framing: it still requires every rank’s exact init, barrier, and numeric all-reduce marker exactly once, but counts fixed substrings instead of complete physical lines. The XCCL commands, ranks, numeric expectations, timeouts, and two-pass requirement are unchanged.

The new recovery root is registered as:

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/w1-n128-nvme-postreboot-recovery-20260723T140536Z

This retry is an infrastructure parser correction on the same clean boot, not a performance-conditioned restart. A1 remains the first model generation after a complete recovery pass.