neuwirth/levineuwirth.org
neuwirth
/
levineuwirth.org
1
1
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

auto: 2026-08-30T17:54:47Z [skip ci]

This commit is contained in:
Levi Neuwirth 2026-08-30 19:54:47 +02:00
parent 52a7482e80
commit 9fdf0e202f
No known key found for this signature in database
2 changed files with 21 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
content/essays/verified-inference/index.md
View File

@ -220,18 +220,27 @@ A profiler that only ever confirms its author is not a measurement instrument.
Predictions go out before the measurement, and get corrected in public when the Predictions go out before the measurement, and get corrected in public when the
measurement disagrees — otherwise the numbers above would be worth very little. measurement disagrees — otherwise the numbers above would be worth very little.
The correction is now upstream rather than only narrated here. The original
scorecard reported roughly 950 GB per sweep of cross-shard traffic and returned
a binding verdict on interconnect; with the traffic model corrected — settlement
vectors reduce to per-shard scalars rather than shipping whole — the same
configuration moves 2.4 to 2.8 GB per sweep, and communication is negligible at
every bandwidth swept. Nearly all of the original figure was my own modeling
error. The retraction is part of the merged record, not a footnote to it.
### What is upstream, and what is not ### What is upstream, and what is not
**Merged code** — the dry-run profiler and its parts: manifest contract, **Merged code** — the dry-run profiler and its parts: manifest contract,
extractor, cost model, execution DAG, partition scorecard; and the RMSNorm extractor, cost model, execution DAG, partition scorecard; the RMSNorm
cost-model correction. cost-model correction; and, as of 30 August 2026, the calibration suite, the
machine-profile tooling, and the corrected traffic model.
**Publicly reported validation** — the B200 tape extraction and the measurements **Publicly inspectable evidence** — the B200 session archive, the extracted
that came out of it, recorded in the profiler pull request itself. manifests, the calibration logs, and the measured machine profile are all in the
repository. An earlier version of this section asked you to treat the figures
**Pending upstream** — the calibration suite, the machine-profile tooling, and above as reported by me rather than independently checkable. That caveat has
the remaining B200 artifacts. Until those land, treat the figures above as been discharged: the inputs that produced them can now be read, and the
reported by me rather than independently inspectable. cross-checks rerun.
**On agent assistance.** Implementation was agent-assisted. I owned the research **On agent assistance.** Implementation was agent-assisted. I owned the research
direction, the cost-model derivation and checking, the experimental design, the direction, the cost-model derivation and checking, the experimental design, the

6
content/work.md
View File

@ -52,8 +52,10 @@ work is on proving multi-GPU, which is the ceiling on model scale, context
length, and mixture-of-experts breadth. length, and mixture-of-experts breadth.
::: {.work-limit} ::: {.work-limit}
**Ongoing.** The multi-GPU work is unmerged and its results are preliminary. **Ongoing.** The profiler and calibration tooling are merged upstream; the
Technical write-up expected Q4 2026, for review and publication. multi-GPU figures are projections from a validated cost model rather than
measurements at that scale. Technical write-up expected Q4 2026, for review and
publication.
::: :::
::: {.work-entry-links} ::: {.work-entry-links}