Skip to content

docs: launch-rehearsal receipt — 6-box EU ring, 7 bugs, engine proven on sm_120#134

Merged
leyten merged 2 commits into
masterfrom
receipt/launch-rehearsal
Jul 21, 2026
Merged

docs: launch-rehearsal receipt — 6-box EU ring, 7 bugs, engine proven on sm_120#134
leyten merged 2 commits into
masterfrom
receipt/launch-rehearsal

Conversation

@leyten

@leyten leyten commented Jul 21, 2026

Copy link
Copy Markdown
Owner

Real-hardware rehearsal (~$8). 7 launch issues found (2 fatal manifest byte-integrity bugs that would fail every launch pull). Proven on real 5090s: self-provision, sm_120 cutlass engine load, 35GB verified pull, peer seeding, 6-box EU ring formation. NOT achieved: serve tokens (re-form storm starved it) — focused follow-up. Receipt in docs/receipts/.

leyten added 2 commits July 21, 2026 13:23
The engine computes mean_accept, toks_per_traversal (g), tok_s, prefill_s,
and the per-stage transport/compute split, then DROPPED them — SHARD_JOB_DONE
carried only the text + token count, so the daemon serve path was
unmeasurable (no way to see tok/s, acceptance, or where the time goes). Emit
an additive SHARD_JOB_METRICS line with whatever perf fields the run
produced; consumers that don't know it ignore it (the daemon logs unknown
contract lines). Found by the rehearsal-measurement design — turns a
black-box serve into an explainable one.
…proven on sm_120

Real-hardware rehearsal (~$8, 6-box EU ring serving MiniMax-M2.5 via the
stranger daemon path). Found 7 launch issues (4 pre-spend by a design panel,
3 on real hardware incl. TWO fatal manifest byte-integrity bugs that would
have failed EVERY launch pull). Proven green on real 5090s: self-provision,
sm_120 cutlass engine load, verified 35GB pull, peer seeding, and ring
formation across 6 EU boxes (the -announce fix). NOT achieved: end-to-end
serve tokens — the ring formed but a cold-cohort re-form storm starved the
serve (0 tokens), so no tok/s numbers yet; that's the focused follow-up.
@leyten
leyten merged commit dc9bd77 into master Jul 21, 2026
2 checks passed
@leyten
leyten deleted the receipt/launch-rehearsal branch July 21, 2026 16:08
leyten added a commit that referenced this pull request Jul 22, 2026
…120 (#134)

* feat(coordinate): surface serve perf metrics (tok/s, g, transport split)

The engine computes mean_accept, toks_per_traversal (g), tok_s, prefill_s,
and the per-stage transport/compute split, then DROPPED them — SHARD_JOB_DONE
carried only the text + token count, so the daemon serve path was
unmeasurable (no way to see tok/s, acceptance, or where the time goes). Emit
an additive SHARD_JOB_METRICS line with whatever perf fields the run
produced; consumers that don't know it ignore it (the daemon logs unknown
contract lines). Found by the rehearsal-measurement design — turns a
black-box serve into an explainable one.

* docs: launch-rehearsal receipt — 6-box EU ring, 7 bugs found, engine proven on sm_120

Real-hardware rehearsal (~$8, 6-box EU ring serving MiniMax-M2.5 via the
stranger daemon path). Found 7 launch issues (4 pre-spend by a design panel,
3 on real hardware incl. TWO fatal manifest byte-integrity bugs that would
have failed EVERY launch pull). Proven green on real 5090s: self-provision,
sm_120 cutlass engine load, verified 35GB pull, peer seeding, and ring
formation across 6 EU boxes (the -announce fix). NOT achieved: end-to-end
serve tokens — the ring formed but a cold-cohort re-form storm starved the
serve (0 tokens), so no tok/s numbers yet; that's the focused follow-up.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant