b70-optimization-lab

Laguna M8 direct-XCCL runtime-graph preregistration

Date: 2026-07-24 America/Toronto

Scope

This authorizes one component-only probe of the runtime graph substrate. It does not load the Laguna target or DFlash draft, start an endpoint, generate a token, time endpoint throughput, stage a payload, use the network, or submit to LocalMaxxing.

The approved record remains unchanged:

33.89498511171744 tok/s;
LocalMaxxing cmrx6p5dv001bo4017hb7sixz;
vLLM 8936aac144929190c1e53f8b8624ca397ce16f5b; and
XPU kernels b6076ce1249ffee0e30bee528f4cd15c3bffb234.

The experimental target-only integration is separately committed at vLLM e09f34a008c31cb4c691697215a6eff3aa2eb5be. It is default-off, keeps the DFlash drafter eager, forces every target call except the true single-request DFlash-7 M=8 verifier eager before graph padding/dispatch, pins the graph output, and rejects any replay-time tensor pointer/offset/shape/stride/dtype or device drift. Its focused CPU gate passed 17/17 and an independent re-audit found no remaining dynamic-input blocker.

Question

Can the installed XPU graph and XCCL stack directly record and replay the incumbent Laguna collective pattern without stale data or changed BF16 bytes?

The probe records one graph containing exactly:

97 TP4 BF16 all_gather_into_tensor calls with per-rank shape [1, 8, 3072] and gathered shape [4, 8, 3072];
three literal rank-0, rank-1, rank-2, rank-3 BF16 additions after each gather; and
one final BF16 XCCL all-reduce.

It then performs 32 changing-input epochs with four unique replays per epoch. Every replay checks the raw uint8 view of all 97 gathered tensors, all 97 fixed-rank sums, and the final all-reduce. This is 24,960 raw comparisons per rank and 99,840 across the four-card process group. Source and output digests must change on all 127 transitions.

The final all-reduce fixture uses small integral BF16 values so its expected sum is order-independent. The 97 rank sums use the literal incumbent rank-ordered BF16 boundaries.

Frozen identity

tool commit: d9e88ff060c9b98dd169547d6aeb9ae31f55c629;
tool: experiments/laguna-s-2.1-xpu-b70/tools/probe_laguna_m8_xccl_graph.py;
tool SHA-256: 1797ae4df5f61f6583fa0ba8942cac7e212f4dfbdc36a28e971734d4f89f264f;
Python: /home/steve/.venvs/deepseek-v4-xpu/bin/python, SHA-256 202c17d1671602a4ef1d43e9b2fdbef0769443f37bf5e51f6b603e0b2c27d9d8;
oneCCL: /home/steve/.venvs/deepseek-v4-xpu/lib/libccl.so, SHA-256 ace144a390a53720b2743844decf127661c942b56f3b414900b9d8c11461acc3;
Level Zero adapter SHA-256: 26fa68779adb03b200a8c3001cf81e59fc9a3d63e0f38627ec0005ffce574e7a;
Level Zero loader SHA-256: 0fe232b18985ae078dd546b57bc6d11bacf1030834c0544f7e3feb53ed71c1d0;
boot: 0b7f98a5-e50a-46a5-81ea-15938b55317a;
kernel: 7.0.0-28-generic; and
output: /mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/laguna-m8-xccl-direct-graph-d9e88ff06-20260724T141124Z.

The tool rejects any output outside the internal-NVMe artifact subtree, requires a new root, and verifies its backing mount is ext4 on an NVMe block device. The external Corsair/NTFS volume is forbidden.

Frozen command

After a read-only four-card discovery/process check confirms no foreign XPU work:

env \
  PYTHONDONTWRITEBYTECODE=1 \
  PYTHONNOUSERSITE=1 \
  ONEAPI_DEVICE_SELECTOR=level_zero:0,1,2,3 \
  ZE_AFFINITY_MASK=0,1,2,3 \
  CCL_ATL_TRANSPORT=ofi \
  CCL_TOPO_P2P_ACCESS=1 \
  FI_TCP_IFACE=eno1 \
  CCL_KVS_IFACE=eno1 \
  LD_LIBRARY_PATH=/home/steve/.venvs/deepseek-v4-xpu/lib:/opt/intel/oneapi/umf/1.1/lib:/opt/intel/oneapi/compiler/2026.0/lib:/opt/intel/oneapi/compiler/2026.0/opt/compiler/lib \
  timeout --signal=TERM --kill-after=15s 240s \
  /home/steve/.venvs/deepseek-v4-xpu/bin/python \
  -m torch.distributed.run \
  --standalone \
  --nproc_per_node=4 \
  /home/steve/llm-optimizations/experiments/laguna-s-2.1-xpu-b70/tools/probe_laguna_m8_xccl_graph.py \
  --output-root /mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/laguna-m8-xccl-direct-graph-d9e88ff06-20260724T141124Z \
  --epochs 32 \
  --replays-per-epoch 4

Decision rule

Continue the direct-capture branch only if all four rank results and the aggregate report pass, all 99,840 raw comparisons match, every freshness transition changes, and there is no timeout, fallback, device loss, or runtime error.
A capture/runtime failure or raw mismatch closes direct collective capture. It does not authorize a retry of unchanged code; the next permissible design is a separately reviewed segmented graph that keeps all 98 collectives eager with caller-owned persistent buffers.
A tooling or infrastructure failure before graph construction may be fixed only under a new tool identity and new preregistration. It is not a performance or correctness result.
Even a pass authorizes only an actual target-M8 component/trace gate. It does not authorize an endpoint. Model/revision/quantization hashes and the canonical q1 teacher contract must be bound separately before any model execution.

This site is open source. Improve this page.