b70-optimization-lab

Laguna S 2.1 exact Breakable graph record

Date: 2026-07-24 America/Toronto

Status: VERIFIED, SUBMITTED, APPROVED RECORD. LocalMaxxing ID cmrzjb7i906x4o401egrnm05m. The conservative lower graph start is 92.16352215694299 tok/s, compared with the prior approved 33.89498511171744 tok/s record.

Result

The preregistered fresh-service A1/B1/B2/A2 campaign passed every quality, honesty, causal, reproducibility, and record gate:

Leg Treatment Median tok/s, tokens 1-100 after TTFT Canonical exact Cached tokens
A1 eager 34.49116381990456 13/13 0 on 13/13
B1 Breakable graph 92.76071675369911 13/13 0 on 13/13
B2 Breakable graph 92.16352215694299 13/13 0 on 13/13
A2 eager 34.591122672009526 13/13 0 on 13/13

The submitted value is the lower B start. It improves the prior approved record by 58.26853704522555 tok/s (+171.90902091614203%, or 2.71909020916142x).

Both adjacent comparisons passed the preregistered causal gates:

Pair Graph row wins Median paired gain Decode-cycle saving Acceptance drift
B1 vs A1 13/13 +169.42082090615466% 55.0490377685078 ms 0.0003070338847266929
B2 vs A2 12/13 +169.36519273519548% 54.22048225686592 ms 0.0003066772832345799

All four legs are mutually bitwise identical and match the canonical q1 greedy teacher. Long-then-next passed on every leg, as did the 863-token rollover row. Metrics started at zero, each leg recorded exactly 13 decode requests, and every service used a private cache and RPC path.

What changed

The candidate retains the approved exact DFlash7, fused W1-route-W2, route-interleaved expert GEMM, shared-elementwise, QKNorm/RoPE, and W1 N64 stack. The only treatment difference from eager is the validated production Breakable graph runtime:

mode=NONE
cudagraph_mode=PIECEWISE
cudagraph_capture_sizes=[8]
max_cudagraph_capture_size=8
VLLM_XPU_LAGUNA_M8_BREAKABLE_GRAPH=1
VLLM_USE_BREAKABLE_CUDAGRAPH=1
XPU_GRAPH=1
VLLM_XPU_ENABLE_XPU_GRAPH=1

Each graph start performed its first lazy capture inside the first measured cold request. Each logged exactly one audited capture and one replay on ranks 0 through 3, with 146 graph segments and 145 eager breaks. The eager controls logged no graph capture or replay.

No warm-up generation, retry, prefix cache, history/ngram acceleration, response reuse, context checkpoint, qualification timing, or prior-run output was used.

Frozen identity

Models, caches, RPC state, and evidence all used internal NVMe paths.

Sealed evidence

/mnt/fast-ai/llm-optimization-artifacts/laguna-s-2.1/runs/laguna-m8-formal-graph-crossover-4b4b5dd9c-0ce373a31-20260724T220518Z

The root and leg directories are mode 0500; evidence files are mode 0400.

Key SHA256:

f8f2697ce6abcde7a78d9af2a16f89628027ef55a3a60e8d60b025e31c135378  full-analysis.json
9467b3d105a918180fb3e92e6b951f4217a5fc285df0a939d14ecfad1a456852  all-vs-teacher.json
f2ea7649923f47e8e0e3c4230c56fc7a7926157f24f981a91718eeed65a6248f  cross-leg.json
616445a0996ec2795bed130d3dc9c79261ae4b9fd3c36d5583c040df22efdebc  B2-graph/bench.json
393d2ea3c29328cae249df4b2a101ee30689d9b7d3ca4eddf06eef49a146c3d9  B2-graph/exactness-vs-q1.json
3342886132977e73782046fcc9b1fb78efcfb8b4e12cd2f448e564447074792a  B2-graph/server.log

Two independent read-only audits recomputed raw token equality, cache-zero status, all performance statistics, metrics deltas, topology, environment, 26 idle snapshots per leg, cleanup, sealing, and provenance. Both reported no discrepancy.

Decision

Promote the exact Breakable graph stack. The independently audited payload submitted only 92.16352215694299 tok/s, the lower graph start. LocalMaxxing returned HTTP 201 and approved it as cmrzjb7i906x4o401egrnm05m. Preserve both graph starts and both controls as the reproducibility record.

This site is open source. Improve this page.