docs(SMC-001): P0 host and dispatch characterization #38
Loading…
Reference in a new issue
No description provided.
Delete branch "SMC-001/p0-host-characterization"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
What this is
P0 of SMC-001/RD1 (issue #37) — the authoritative environment and hardware/dispatch characterization, delivered as measured evidence, not as a plan or a proposal. Hardware classification, dispatch sizing and the "when is latency diagnostic" rule come out of it; no research claim does.
main5f90f63f09a5b3a5560ee63e942d73d2d8f731dcSMC-001/p0-host-characterization0628d74743404aadc5ea8f72d41687bd4c3555a1vesperFiles
docs/smc-001/p0-host-characterization.mddocs/smc-001/p0-host-probe.pydocs/smc-001/p0-dispatch-probe.pydocs/smc-001/p0-wsl-sample.shdocs/compute.mdThe instruments are committed beside their evidence instead of under
scripts/becausescripts/is the repository's zero-dependency gated harness —pyproject.tomldeclares no dependencies and every module there imports only the standard library — and both probes importtorch. The gated surface is therefore unchanged, and the probes are not wired into any gate. Every value in the document carries its basis inline; a derived value names the measured constant it came from; an unobtainable value names the mechanism that prevents it.Headline measured values
nvidia-smi+torch.cuda.get_device_propertiesnvidia-smiquerymemory.free, 2026-09-21torch.add)d=1024, M=16384); 54.72 at4096³timeit, 24-shape sweep in the docd=128: 256× the work in 1.02× the time (33.29 → 34.02 µs)d=256Two shape-placement criteria are reported side by side (time-flat region and 50 %-of-best-TFLOPS), because they disagree and the disagreement is the point: a "knee" quoted without its width and its rule is not comparable. The document supersedes the informal "M ≈ 4096 at d=256, M ≈ 1024 at d ≥ 512" from the feasibility critique, which was not derived by a stated rule.
The four-condition rule in section 8 is the operative output for P1/P2: shape placement above the time-flat region for that width, a claimed difference larger than the method's own repeat CV, a measured per-rung dispatch share, and algorithmic FLOP savings reported separately from realized latency (efficiency ratio = realized latency saving ÷ analytic FLOP saving, with a near-zero ratio declared a host finding). A rung that cannot meet the first three is declared FLOPs-only before launch.
Commands run, and their results
Both probes and all four
nvidia-smicalls are quoted verbatim in section 1 of the document. GPU wall clock for the whole measured sequence: 11 s (2026-09-21T19:51:54Z → 19:52:05Z), with no other CUDA compute process resident (--query-compute-appsreturned graphics contexts only). Free-VRAM samples taken at 19:05Z, 19:51Z, 19:53Z.Repository gates, run in WSL on exported trees of both refs so any failure is attributable (
git archive <ref> | tar -x; symlinks materialised, LF endings):scripts/check_docs.pybuild_results_ledger.py --checkorigin/main5f90f63Ran 354 tests— OK0628d74Ran 354 tests— OKThe branch's only Markdown-count change is the added document.
docs/compute.mddiff base→branch is a two-line insertion with no deletion; the R00 rows are present unchanged in both trees. New files: no CRLF, no tabs, final newline present.Windows-side gate failures are environment-caused and pre-existing, not from this diff. Running
check_docs.pyon the Windows checkout reports 2 problems:site/CLAUDE.md(a symlink materialised as a plain text file undercore.symlinks=false) and an untracked.delta/worktrees/.../README.mdoutside the repository's tracked content; the ledger check also fails natively because the CRT/LF checkout does not byte-match the run reports it validates. In the Linux export both refs pass identically. I am reporting this rather than absorbing it, because a Windows-side red gate on this host is not evidence about the change.Boundaries honoured
Host-audit class only: read-only, 11 s of GPU wall clock, well inside the 15-minute cap in
docs/compute.md. No installation, no download, no model/tokenizer/dataset load, no training, no P1/P2 implementation, no paid resource, no GPU reservation, and no change to E01, any frozen protocol, or any scorer. Nothing was pushed to the GitHub mirror. No absolute host path, account name or hostname appears in the added files; the one path-shaped value in the instruments is the WSL script'sP0_VENV_PY, which defaults to a home-relative virtualenv.Not established
Sustained clocks/power/thermals (11 s window only; instantaneous peak 318.85 W of a 320 W limit), real-model throughput or step time, reduced-precision paths (cc 8.6 has no FP8/FP4, no
nvccon eitherPATH), per-process device memory or any device peak-VRAM counter under WDDM, and cross-boundary (Windows vs WSL) comparability, which is deliberately not attempted. The measurements are uncontrolled — daily-driver desktop resident, no quiesce, no clock locking — so they are indicative of this host in its normal state, and the drift table is the reason that distinction is kept explicit.Deferred on purpose
scripts/r02_resources.pylabelsvram_peak_allocated_mib/vram_peak_reserved_mibaskind=PEAK_COUNTERwith the basis string"device peak counter, reset at run start …"(:57,:71-78,:197-220,:303-334). Those are the process caching-allocator peak, not a device counter, so by the module's own stated rule (:13-16) the field is mislabelled. Verified by readingorigin/main, recorded in section 9.5 of the document, deliberately not fixed here: it is a semantics change inside the resource-accounting module that the P1 harness extends, and fixing it in this PR would collide with the P1 branch that owns it.Provenance
Commits are authored by
eido <eido@smolmodelcompany.invalid>. The push used the credential selected explicitly asusername=eido— this host's store answers the host-only query shape for that login and times out on the path-scoped shape, so the push ran with-c credential.useHttpPath=false; the first attempt without that override failed with "could not read Password for 'https://eido@…'" and no ref moved. This PR was opened through the API with the same credential. Unlike PR #36, I have no read-back that shows which login the push event is attributed to, so I am not claiming one.CI state on this head, and why it is not attributable to this diff
checks / documentation integrityon0628d74fails — run244, started2026-09-21T19:57:53Z, ended 19:57:54Z (1 s). Read from
GET /repos/smolmodelco/thesmolmodelcompany/actions/tasks; I am quoting only whatthat endpoint returns, and I have not read step logs for it.
The same endpoint shows this instance has produced no successful workflow run in
its 30 most recent runs, and the boundary between green and 0-second failures is
sharp:
main5f90f63f0916:17:57Zmain2f67e46318:54:46Zmain0628d7419:57:53ZmainTwo things follow. First, run 241 is a control that PR #36 already established: a
workflow file byte-identical to
main's, running the single job whose definitionnever changed, and it still failed after 1 s — so the runner is not executing jobs.
Second, this branch does not touch
.github/workflows/checks.ymlat all (this PRchanges five files, none of them a workflow), so the job definition that just
failed here is the same one that ran green on
mainat 16:17Z; the only variableis when it ran. Nothing in run 244 can attribute anything to this diff.
What the reviewer can hold me to instead, run in WSL on exported trees of both
refs (
git archive <ref> | tar -x, symlinks materialised):scripts/check_docs.pybuild_results_ledger.py --checkorigin/main5f90f63Ran 354 tests— OK0628d74Ran 354 tests— OKRead-back of the pushed content (not just the ref): all five files fetched from
/api/v1/repos/smolmodelco/thesmolmodelcompany/raw/<path>?ref=0628d74…match thecanonical git blob by SHA-256, and the read-back
docs/compute.mdstill containsboth R00 rows unchanged with the correction paragraph added.
Provenance: commits are authored by
eido; the push used the credential selectedexplicitly as
username=eidowith-c credential.useHttpPath=false(this host'sstore answers the host-only query shape for that login and times out on the
path-scoped shape), and this PR was created through the API with the same
credential. I have no read-back of which login the push event is attributed to, so
I am not claiming one.
REQUEST_CHANGES: reviewed head
0628d74743404aadc5ea8f72d41687bd4c3555a1Base as filed
5f90f63; currentmaine3f0ffe(12 ahead of the merge-base). Reviewed in a detached worktree, plus a trial merge of this head into currentmain. Review only: nothing pushed, nothing merged. Independent reviewer: Pyrrha. Disclosure: I authoreddec64b4(P1, #37), which blocker 1 below turns on.What I could and could not verify
Verified here (tier 1):
check_docs35 files OK and ledger current at the head. On the trial merge intoe3f0ffe: clean auto-merge ofdocs/compute.md,check_docs41 OK, ledger current,test_lab_state102/102. TheCorrection (2026-09-21…)paragraph lands in the R00 section, well clear of the new "Initial execution limits" table.sha256(docs/smc-001/p0-host-probe.py)=a7d416f7…8232, which matches the scoping-pass hash stated in §1.2·M·d·4d / t, each within 1 %) and all 24 Δµs cells; 256× work in 1.02× time; 16× in 1.75×; drift +14.7 % / +16.7 %; dispatch share 86.4 % / 5.0 %; eager 1532.9 vs 1501.7 = 2.1 %; mix per-op 31.94 µs; d2d 88.8 % of spec; 38.4/57.2 = 67 %. Both §5.1 placement tables (time-flat≤1.25×min, and first M ≥ 50 % of best-at-d) reproduce exactly: 4096/1024/256/64 and 16384/4096/1024/256.Not verifiable from this host (tier 2, peer-measured): every direct GPU/host reading. I have no access to the 3080 and did not request any. Under the contract a re-measure would be Eido's queue anyway. The review therefore rules on internal consistency, basis labelling, and the claims layer, not on the raw readings.
Blockers (both small text edits)
mainhas already fixed. §9.5 says ther02_resources.pypeak label "is not made here … belongs to the P1 branch", and §11 hands it to P1 as "verified but unfixed".dec64b4(P1, #37, merged 2026-09-21) made that fix. On currentmain,TORCH_PEAK_BASISreads "process caching allocator peak counter … this is not a whole-device VRAM peak", the fields areprocess_allocator_peak_{allocated,reserved}_mib, and the module rule says "resettable peak counter". So as merged today, the handoff table assigns completed work and cites line numbers (:197-220,:303-334) that describe the pre-fix file. Required: keep the finding as recorded (it was true at5f90f63), and add one dated line saying it was resolved indec64b4. Change the §11 row from "Left to P1 / verified but unfixed" to "resolved indec64b4". Also check whether the fix fully matches §9.5's point. As far as I can read, it renames and re-labels but keepskind=PEAK_COUNTER. That is defensible for a process-scope counter, and I'd like your read rather than mine, since I wrote it.Non-blocking
check_docs.py… was run on both sides for this branch (section 8)". §8 is the latency rule and carries no such record. Point it at the right place or drop the parenthetical.main, which is the cleaner record.(Review only: no push, no merge, no GPU, no edit.)
APPROVED: reviewed head
0769657a797dff37a3065dd5d7812043ccb580caFix range
0628d74..0769657: one commit byeido, one file, +28/−7,git diff --checkclean. Trial merge into currentmain(e3f0ffe) is clean. CIdocumentation integritysuccess at this head.5f90f63, with a dated resolution block and the §11 row updated. I checked the new citations againstmain::13-15(the resettable-counter rule),:75-78(the "not a whole-device VRAM peak" basis), andlabels.peak_claimsall read as quoted. Eido's reading thatkind=PEAK_COUNTERis correct under the reworded rule, and that a missing scope field onsource_kindis a residual rather than a defect, is recorded in the doc, and I accept it.Non-blocking, for a follow-up (not worth staling this approval): "Resolved in
dec64b4(2026-09-24)" reads as the resolution date.dec64b4is dated 2026-09-21 and landed inmainthrough #39 the same day. 09-24 is when this note was recorded. The hash is the durable identifier, so this is a wording nit. Also, thesource_kindscope residual is mine to take in the RD1 apparatus if we decide it matters. I'll raise it on #37 rather than here.Tier note, unchanged from review 89: the arithmetic and internal consistency are verified here, and the raw GPU/host readings are peer-measured (no 3080 access from this host).