Skip to content

Add OLMoE model - #1444

Merged
merrymercy merged 1 commit into
sgl-project:mainfrom
janimo:olmoe
Sep 17, 2024
Merged

merrymercy merged 1 commit into
sgl-project:mainfrom
janimo:olmoe

Conversation

@janimo

@janimo janimo commented Sep 17, 2024

Copy link
Copy Markdown
Contributor

Based on VLLM WIP at vllm-project/vllm#7922

@merrymercy
merrymercy merged commit 1f4b5f7 into sgl-project:main Sep 17, 2024
@janimo
janimo deleted the olmoe branch September 17, 2024 12:28
@Muennighoff Muennighoff mentioned this pull request Sep 19, 2024
timethink pushed a commit to timethink/sglang that referenced this pull request Mar 9, 2025
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 16, 2026
… previous boot's MEASURED record, not the xsn14 constant

DC_MEASURED_D_XCHG_MIB = (2588, 3084, 2588) was read at boot weg2xsn14 while
the weights_draft tag was still RESIDENT on the sleeping cards.  The tag has
travelled with the exchange since; the front measured 1616-1686 MiB (5090)
and 1256-1356 MiB (3080) at boot weg2xsn203.  P's per-card budget paid the
constant anyway: 1000-1150 MiB per card for bytes that are not there.

* front: measures the device residue BEFORE sampling the dormant image and
  stamps it (vram_residue_mib per uuid) plus the weight form (--weight-form,
  set by the launcher) into the dormant-image record it already appends.
* host_ledger.dormant_image_sample: carries the two fields (empty when the
  caller measured none).
* launcher.dc_residue_from_record: ONE selector -- newest group-D record of
  the SAME form with a value for EVERY card of this boot, + 256 MiB margin
  (2x the 100 MiB flip-to-flip spread); otherwise the constant with a printed
  reason.  sgl-project#1444 DC-RESIDUE log line names source and per-card reserve.
  SGLANG_WEG2_DC_D_RECORD=0 prices the constant again.

Tests: test_weg2_dc_record_1444 (10: same form prices measured+margin and
undercuts the constant on every card, other form / missing card / bogus
value / kill switch refuse, record fields typed, AST wiring pins) +
w19_form_residue_1273 + xchg_reserve_1273 + launch_replay_1329: 67 passed
(hermetic, CUDA_VISIBLE_DEVICES="").
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 16, 2026
… constant was priced

Boot weg2xsn205 priced group D's residue from the record (sgl-project#1444 DC-RESIDUE
source=RECORD) and still printed 'dormant residue RESERVE ... of boot
weg2xsn14 (2588/3084/2588 MiB ...)' right after it -- two provenances for one
number.  The old line is now under 'if _dc_from_record is None'.  Only the
one log call moved; parse_user_reserve and the B4i WEG2-XCHG-RESERVE line
stay unconditional.

Tests: dc_record_1444 + launch_replay_1329 + xchg_reserve_1273: 52 passed
(hermetic).
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 16, 2026
… residue attributed on the rank at every release

The front measures the dormant residue as ONE number per card (WEG2-DC,
1616-1686 MiB on the 5090 at weg2xsn205) and sgl-project#1444 prices it.  Section [1y]
attributed 1462 of rg6's 1820 MiB from dumps; with the graph pool and the
float workspace now tagged (458 MiB released) the measured residue still
sits ~650 MiB above the [1y] rows.  This is the live attribution:

  nvml_proc = torch_reserved (allocated) + tms_resident {tag: MiB}
              + other (context + driver + communicator + non-torch)
  tms_paused {tag: MiB} beside it: unmapped, NOT in nvml_proc.

* weg2_memory_saver.dc_breakdown / format_dc_breakdown: pure, MiB, a reading
  that could not be taken prints n/a, never 0.
* weight_updater._weg2_nvml_self_bytes (NVML per-process bytes for this pid
  = the front's own quantity) and _weg2_log_dc_breakdown, called at the end
  of every Weg-2 release RPC (kv stage and each weights leg), fail-soft.

Tests: test_weg2_dc_breakdown_1446 (5) + seam_async_1437: passed (hermetic).
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 24, 2026
…gl-project#1444-Record (fnFL2x111 W19)

fnFL2x111 (P-Chunk 8192) fiel vor dem ersten Flip am W19-Riegel: D-TP0 hielt beim
ersten Sleep 2000 MiB (torch_reserved 26248 statt 26198 MiB), die Front gradete
gegen 1960 MiB. Die 1960 kamen aus der Austausch-Census (Boot fnFL2x86: 1896 + 64
Slack), obwohl der sgl-project#1444-Record des vorigen Boots (x110) 1950 gemessen und
1950+256+64 = 2270 gepreist hatte -- die census-korrigierte Schleife ueberschrieb
den neueren Record mit der aelteren Census. Jetzt gewinnt je Karte die groessere
der beiden gepreisten Reserven, die Logzeile nennt den Sieger je Karte. P's
Budget auf der 5090 sinkt damit um die 310 MiB, die der Record ohnehin vorsah.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 24, 2026
Nutzer-Order 24.09.: "dadurch koennen auch mehr experten im P layout gehalten
werden."

WAS IST
- weg2/draft_post.py (rein, aus den safetensors-KOEPFEN gerechnet, kein
  Boot-Zahlwert): P-Draft-Posten der letzten P-Stufe = Checkpoint ohne
  lm_head (teilt die letzte Stufe mit dem Ziel) + Runner-Puffer 64,1 MiB
  (x137 COVER) + Produzenten-Transiente 622,8 MiB (fnFL2x83); in ganze
  Experten-Zeilen je Layer dieser Stufe umgerechnet (layers x row_mib),
  Deckel H5 (E-2)/E. Keine Reserve, kein Handpin.
- launcher.apply_p_draft_post: in solve_p_cut VOR dem Pool-Term, nur ohne
  Draft auf P; die gehobene Fraktion geht an ALLE DREI Stellen, an denen FR_P
  lebt (--pp-cut-expert-device-fraction, --extra-p
  --rank-moe-resident-fraction, --env-p SGLANG_MOE_RESIDENT_EXPERT_FRACTION).
  Zeile 'PP-CUT draft post (H25) card=nvml2 stage=2 freed_mib=3422 ... ->
  experts +176 rows ... FR_P[2] 0.3900 -> 0.733887' im Dry-Run.
- ns._expert_row_mib wird jetzt gesetzt (der Leser in der
  PLATZTAUSCH-KARTE-Zeile hatte keinen Schreiber; "Store ~GiB" fehlte).
- WEG2-HOST DRAFT-SWAP (Praezisierung 08:30Z "systemram gegen vram
  tausch"): Store vor/nach aus der Platztausch-Karte (build_nested mit der
  Arm-FR_P und der gehobenen), d_draft_host, net. x137-Zahlen: Store 378 ->
  378 Plaetze (43872 MiB), d_draft_host 1587 MiB, net +1587 MiB. Das
  Nullsummen-Praemisse haelt in dieser Form NICHT: der Store ist
  total - |Schnittmenge aller Stufen|, PP0 (0.26, 134 Ids) bindet ihn, und
  die neuen PP2-Zeilen sind P-Extras, die D in der D-Phase nicht haelt --
  ihr Byte muss dann im Host liegen. Der Store folgt nur der bindenden Stufe
  (Test: PP0 >= 138 -> 374 Plaetze = D's Boden).
- 5090: kein Gewinn gegen heute -- D's Draft verliess die Karte schon ueber
  den Tausch (x137 D TP0 SLEEP-TAG-TIME weights_draft pause 1556 MiB); der
  sgl-project#1444-Record (1698 MiB) enthaelt ihn nicht.
- Tests: test_weg2_draft_post_h25b.py (8), x137-Zahlen, Checkpoint als
  duennbesetzte safetensors-Attrappe mit den gemessenen Tensorklassen.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 24, 2026
…READY gate replay (desk, no GPU)

python <tree>/python/sglang/srt/weg2/form_matrix_gate.py [--form 27B-DFLASH|NF-MTP]
[--launcher-args "<arm argv>"] [--base-tree <tree>] [--explain KEY=why] [--reuse]

Per form: (1) launcher --dry-run of THIS tree with the arm's argv/env
(canonical: arm_xsn418 step-6 'deviation' argv / arm_fnFL2 LARGS defaults;
--launcher-args hands the arm's exact argv over); (2) the WEG2-FORM line must
name the form's axes; (3) group P/D argv + WEG2-GROUP-ENV diffed against the
NEWEST boot of this form that reached front READY, every row classified
SOLVED / PER-BOOT / FORM / EXPLAINED / UNEXPLAINED (UNEXPLAINED fails for the
27B; NF reference is information because its canonical argv is the arm
defaults); (4) argv vs the last GOOD boot (27B: weg2xsn411), informational;
(5) THIS tree's post-READY gates (W7/W10 P+D, W9, W45 census, W10, W11/W11b)
replayed on the reference boot's own P/D logs; (6) --base-tree: the same dry
run on a second tree, every difference must be the form's own or come from a
MOVED calibration source (named per source: sgl-project#1444 record, P-cut log, run
sample). The parent runs by path, torch-free (import sglang = 612 MiB), all
dry runs before the replay imports the launcher.

Live desk runs (incg cgroup, CUDA_VISIBLE_DEVICES=""):
- 27B-DFLASH on this tree: PASS. Dry run rc 0; WEG2-FORM dense/none/dflash/
  cold/paged_dcp/family/transient; vs weg2xsn418: SOLVED P --max-total-tokens
  279632->272668, P --rank-gpu-memory-mib 26408,16264,15992->26272,15856,15576
  (reserve now 27B's own D residue 1404+256 on the 3080s instead of the
  census-corrected 1314 that W19 stopped), D budget (dry expectation);
  FORM: sgl-project#114 transient gone, SGLANG_WEG2_FORM added; rest PER-BOOT.
  Replay on xsn418: W7/W10 P+D, W9, W45 (bound 648806), W10 (SKIPPED by
  name, produce off), W11/W11b PASS. vs weg2xsn411: only SOLVED flags.
- NF-MTP with --base-tree d37b793: argv/env differ ONLY by CALIBRATION --
  the base tree priced weg2xsn418's (27B) D residue for the NF form, this tree
  NF's own fnFL2x144 (P/D --rank-gpu-memory-mib, P --max-total-tokens) -- plus
  SGLANG_WEG2_FORM and per-run paths. Replay on fnFL2x144: W7/W10, W9, W45 PASS,
  W10/W11 SKIPPED (that boot ran --draft-kv-on-p off).
Tests: test_weg2_form_matrix_gate.py 10 passed, test_weg2_form_axes.py 38 passed.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 24, 2026
sgl-project#1444/P-cut/depth/host record keyed on checkpoint+form)

User order 24.09.: "da muessen schalter rein, in den code, dense, moe,
vollstaendig im vram, mit offload ... sonst knallts doch an jeder stelle".
Inventory /root/.claude/jobs/1ab4cd30/tmp/27b-up-c/FORM_AXES_INVENTORY.md.

weg2/form.py (new, torch-free): axes arch{dense,moe} experts{none,resident,
offload} draft{dflash,mtp,none} p_draft{compute,cold,none} kv{paged_dcp,
qsa_forma} + flip/vision, each DERIVED from the flags that build the argv
(arch from the checkpoint config). --profile = expectations, --form-* flags
override them (--form-draft/--form-p-draft also drive --spec-form /
--draft-kv-on-p / --dflash-produce-on-p when those are not given);
contradictions refused by name: W140 Weg2FormContradiction. ONE line
(WEG2-FORM ...), ONE env (SGLANG_WEG2_FORM, build_env writes it from the
resolved form and pops an inherited one), ONE accessor (current_form).

Calibration identity (the xsn417/xsn418 class):
- sgl-project#114 P prefill transient published only for the checkpoint it was measured
  on (Next-Flash fnFL2x113/116/118); 27B P no longer gets 918,815,623 MiB
  (xsn417 PP0 pool 246155 < 262144).
- sgl-project#1444 DC record, host-ledger record (run peak/image): only samples of this
  checkpoint and residue form (arch/experts/draft/kv/flip; pre-form boots by
  their drafter). xsn418 died W19 on fnFL2x144's (NF) 3080 residue 768 vs
  measured 1404; the 27B now reads weg2xsn418's own record.
- P-cut design prefix and P-depth bubble: newest P log OF THIS CHECKPOINT
  (xsn417 cut on fnFL2x142's census; now weg2xsn411's).

MoE-only gates declared (sgl-project#106/sgl-project#107/sgl-project#134/sgl-project#140/sgl-project#145 fraction/H14): a dense form
logs '<gate> SKIPPED (form ...)' instead of ENTFAELLT/none; MoE forms unchanged.

Desk evidence (no GPU): 27B-DFLASH dry run (arm_xsn418 argv) base d37b793
vs this tree: group P/D argv byte-identical; P env loses
SGLANG_KV_BUDGET_PREFILL_TRANSIENT_MIB, both gain SGLANG_WEG2_FORM; design
prefix 18905 (fnFL2x144 NF) -> 42971 (weg2xsn411 27B); DC record weg2xsn418.
Tests: test_weg2_form_axes.py 38 passed; 47 affected weg2 files run one by
one: failures only the pre-existing ones, identical counts on the base
(anon_baseline 2, fix8_image 1, lane_cut 2, ring_absent 1, store_priced 2,
lanes_concurrent 2, xchg_shadow 1).

(cherry picked from commit 2d4d4cf)
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 26, 2026
…rofile that measured it (NF), the 27B keeps weg2xsn14

NF cb1575e (22.09., user "diesen wert nicht doppelt oder gar nicht
nehmen") replaced the constant DC_MEASURED_D_XCHG_* by the census reading in
xchg_form_dormant_reserve AND in dc_expect_d (P budget, W19 riegel). In the
unified tree that ran for every profile. On a 27B exchange boot WITH a sgl-project#1444
record the record wins (max), so the 27B dry runs stayed byte-identical to
RC9; WITHOUT one (no record for this checkpoint/form under the S9 calibration
identity, SGLANG_WEG2_DC_RECORD=0) the 27B priced D at its census --
weg2xsn246 1622/1250/1250 MiB -- instead of 3084/2588/2588, below every 27B D
residue the records hold (2074-2150 MiB on the 5090): P ~1.4 GiB/card too
large, W19 at D's first sleep.

Registry row decides: ModelProfile.d_residue_census (nextflash True, qwen27b
False); launcher.xchg_census_is_reserve(profile) is read by the report and by
main's dc_expect_d override (one decision). 27B: the bfc6bd8 line form,
byte-identical; NF (--profile nextflash): unchanged.

Red before: test_weg2_xchg_reserve_1273 x3 + test_weg2_w19_form_residue_1273
(27B pins, red since the unified base a3b9479, red on NF d1c7094, green
on 27B bfc6bd8) and the new test_weg2_xchg_reserve_profile_redfix (6/6 red
before, green after).
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 26, 2026
…faces (798/951 retired with #1015c, slice-A, fix3, 1290, 1233, 1153)

Pre-existing reds on desk/27b-unified-0926 (b3180fb), all but 1153 also
red on the 27B line (bfc6bd8) and the NF line (d1c7094). Test files only.

- pp_void_slot_advance_798 x7 -> successor pins. sgl-project#1015 (01a391f) deleted
  the admission-decision arc the whole file drove (_pp_send/_recv_admission_
  decision, _pp_drain_voided_proxy, ...) and made the proxy channel
  unconditional; #1015c (206661a) retired the downstream refusal
  (pp_upstream_void_pending False; boot_pp3solo_01a391fa03: every downstream
  pass refused, 0 served). The file now pins both halves of that retirement
  (sgl-project#631's restored _pp_reconcile_incoming_admission excluded by name).
- pp_upstream_void_before_formation_951 x3 (found in the same family): the
  three refusal cases rewritten to the #1015c form (pass reaches the
  formation, predicate False on every shape, no void -> streak 0, slot kept).
- scheduling_slice_a_0907: t1 sgl-project#1459 34b6d14 (P takes p_concurrency +
  SGLANG_WEG2_P_QUEUE_AHEAD, default 1); t4/g3 sgl-project#1356 bfc252c (argv_p
  keyword-only); f4c sgl-project#1444 dc15687 (front_argv_for reads
  ns.weg2_weight_source -> the launcher's own parser namespace); f1a: the
  case set awake='P' mid-flip, which since sgl-project#1443/xsn347 (4336934) is the
  P->D flip D legitimately admits into; it now models the D->P flip as flip()
  runs it (awake stays 'D' until the end). Mutant (post-acquire re-check
  removed) -> red. t3 is the front defect, next commit.
- sched_fix3_0907 b1d-b1g: H42b 7370b10 hoisted the pass-level burst
  verdict into a local tested in the loop (b1d: locals assigned from a
  call/attribute count as state; b1e-g: the slice gets its stock values
  _burst_hold=None, _carried_n=0) and xsn326 a3489cc reads module `time`
  (real clock). Round-5 mutant (`if req is not None: continue` above the gate)
  still red on all four.
- no_route_1290 x2: Front.__new__ harness + vision (sgl-project#58 0aeb52e), the sgl-project#49
  epoch keyword of price_remainder (196f6a8), H84 x_start_tokens
  (b895928).
- s3s4_refusals_1233 w25: #1474 c7cd8ae wrapped handle_generate_request
  around _handle_generate_request_impl; the seam is the impl's first
  statement, the wrapper is pinned to clock-read + delegation.
- pp_forwarded_refusal_stop_1153 x2 (unified only): H93 RU 6cda2aa plants
  the usable-match floor on tree_cache; harness tree_cache=None.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 26, 2026
…aseline take the launcher's own namespace

Test files only. Pre-existing on c51591f (admin_key also on b3180fb).

- sched_fix3_0907 b1e-b1g: H91b (bf708a1) hoisted a second pass-level
  gate into a local tested in the waiting-queue loop
  (`_d_park_gate = self._weg2_d_park_admission(...)`, None = stock loop), the
  same shape as the H42b burst verdict; the slice gets its stock None. The
  round-5 mutant (`if req is not None: continue` above the X gate) still turns
  b1d-b1g red.
- admin_key_1275 x2, anon_baseline_1269 x2: the hand-built SimpleNamespace
  lacked weg2_weight_source, which front_argv_for reads since sgl-project#1444
  (dc15687); both use launcher.build_parser() like slice-A f4c.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 26, 2026
…faces (798/951 retired with #1015c, slice-A, fix3, 1290, 1233, 1153)

Pre-existing reds on desk/27b-unified-0926 (b3180fb), all but 1153 also
red on the 27B line (bfc6bd8) and the NF line (d1c7094). Test files only.

- pp_void_slot_advance_798 x7 -> successor pins. sgl-project#1015 (01a391f) deleted
  the admission-decision arc the whole file drove (_pp_send/_recv_admission_
  decision, _pp_drain_voided_proxy, ...) and made the proxy channel
  unconditional; #1015c (206661a) retired the downstream refusal
  (pp_upstream_void_pending False; boot_pp3solo_01a391fa03: every downstream
  pass refused, 0 served). The file now pins both halves of that retirement
  (sgl-project#631's restored _pp_reconcile_incoming_admission excluded by name).
- pp_upstream_void_before_formation_951 x3 (found in the same family): the
  three refusal cases rewritten to the #1015c form (pass reaches the
  formation, predicate False on every shape, no void -> streak 0, slot kept).
- scheduling_slice_a_0907: t1 sgl-project#1459 34b6d14 (P takes p_concurrency +
  SGLANG_WEG2_P_QUEUE_AHEAD, default 1); t4/g3 sgl-project#1356 bfc252c (argv_p
  keyword-only); f4c sgl-project#1444 dc15687 (front_argv_for reads
  ns.weg2_weight_source -> the launcher's own parser namespace); f1a: the
  case set awake='P' mid-flip, which since sgl-project#1443/xsn347 (4336934) is the
  P->D flip D legitimately admits into; it now models the D->P flip as flip()
  runs it (awake stays 'D' until the end). Mutant (post-acquire re-check
  removed) -> red. t3 is the front defect, next commit.
- sched_fix3_0907 b1d-b1g: H42b 7370b10 hoisted the pass-level burst
  verdict into a local tested in the loop (b1d: locals assigned from a
  call/attribute count as state; b1e-g: the slice gets its stock values
  _burst_hold=None, _carried_n=0) and xsn326 a3489cc reads module `time`
  (real clock). Round-5 mutant (`if req is not None: continue` above the gate)
  still red on all four.
- no_route_1290 x2: Front.__new__ harness + vision (sgl-project#58 0aeb52e), the sgl-project#49
  epoch keyword of price_remainder (196f6a8), H84 x_start_tokens
  (b895928).
- s3s4_refusals_1233 w25: #1474 c7cd8ae wrapped handle_generate_request
  around _handle_generate_request_impl; the seam is the impl's first
  statement, the wrapper is pinned to clock-read + delegation.
- pp_forwarded_refusal_stop_1153 x2 (unified only): H93 RU 6cda2aa plants
  the usable-match floor on tree_cache; harness tree_cache=None.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 26, 2026
…aseline take the launcher's own namespace

Test files only. Pre-existing on c51591f (admin_key also on b3180fb).

- sched_fix3_0907 b1e-b1g: H91b (bf708a1) hoisted a second pass-level
  gate into a local tested in the waiting-queue loop
  (`_d_park_gate = self._weg2_d_park_admission(...)`, None = stock loop), the
  same shape as the H42b burst verdict; the slice gets its stock None. The
  round-5 mutant (`if req is not None: continue` above the X gate) still turns
  b1d-b1g red.
- admin_key_1275 x2, anon_baseline_1269 x2: the hand-built SimpleNamespace
  lacked weg2_weight_source, which front_argv_for reads since sgl-project#1444
  (dc15687); both use launcher.build_parser() like slice-A f4c.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 29, 2026
…s statt aus der alten Census

dc_expect_d ist dormant_other in P's Budget, der --dc-reserve der Front
und der W19-Riegel. Auf Next Flash galt je Karte max(neuester sgl-project#1444-Record,
xchg-Census). Die Census stammt aus fnFL2x86 (andere Form) und preiste in
rc12z30r3 1140/1538 MiB (nvml0/2), obwohl jeder D-Record dieser Form seit
28.09. 612-680 MiB mass. Damit lagen je 3080 ~460/860 MiB in P's Budget
brach -- Platz fuer P's residente Experten.

- dormant_max_from_records(recs, cards, form, n, group): der Max-Leser
  von D-EXPECT, jetzt gruppenneutral. p_dormant_from_records ist ein
  Wrapper, das Verhalten ist unveraendert.
- d_reserve_from_records_or_census(): Gibt es D-Records dieser Identitaet
  (_dc_accept) und Form, wird das Maximum der neuesten 8 +
  DC_RECORD_MARGIN_MIB + Slack gepreist; die Census wird dann nur noch
  genannt. Ohne Records bleibt die alte max(Record, Census)-Regel
  byte-gleich (fnFL2x111: eine aeltere Census unterbietet keine neuere
  Messung). Die Lehre gilt weiter: ein einzelner niedriger Wert preist
  nie.
- Mit SGLANG_WEG2_D_RESERVE_RECORD_MAX=0 gilt die alte Regel.
- Neue Zeile 'D-RESERVE RECORD-MAX: <Quelle>'. Nur im NF-Zweig
  (xchg_census_is_reserve), die 27B-Linie ist unberuehrt.

Gerechnet mit z30r3-Records: 2012/998/1000 statt 2006/1140/1538 (nvml1/0/2),
also -142/-538 MiB auf den 3080ern und +6 auf der 5090.

Test test_weg2_d_reserve_record_max_0929: 6/6 rot auf b1ed70a, gruen.
D-EXPECT-Wiring-Test angepasst (der einzige direkte Record-Lesezugriff in
main ist der fuer Gruppe D). 34 Nachbardateien: 823 passed, 5 Altlasten
identisch, 0 neue Rote.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 29, 2026
…s statt aus der alten Census

dc_expect_d ist dormant_other in P's Budget, der --dc-reserve der Front
und der W19-Riegel. Auf Next Flash galt je Karte max(neuester sgl-project#1444-Record,
xchg-Census). Die Census stammt aus fnFL2x86 (andere Form) und preiste in
rc12z30r3 1140/1538 MiB (nvml0/2), obwohl jeder D-Record dieser Form seit
28.09. 612-680 MiB mass. Damit lagen je 3080 ~460/860 MiB in P's Budget
brach -- Platz fuer P's residente Experten.

- dormant_max_from_records(recs, cards, form, n, group): der Max-Leser
  von D-EXPECT, jetzt gruppenneutral. p_dormant_from_records ist ein
  Wrapper, das Verhalten ist unveraendert.
- d_reserve_from_records_or_census(): Gibt es D-Records dieser Identitaet
  (_dc_accept) und Form, wird das Maximum der neuesten 8 +
  DC_RECORD_MARGIN_MIB + Slack gepreist; die Census wird dann nur noch
  genannt. Ohne Records bleibt die alte max(Record, Census)-Regel
  byte-gleich (fnFL2x111: eine aeltere Census unterbietet keine neuere
  Messung). Die Lehre gilt weiter: ein einzelner niedriger Wert preist
  nie.
- Mit SGLANG_WEG2_D_RESERVE_RECORD_MAX=0 gilt die alte Regel.
- Neue Zeile 'D-RESERVE RECORD-MAX: <Quelle>'. Nur im NF-Zweig
  (xchg_census_is_reserve), die 27B-Linie ist unberuehrt.

Gerechnet mit z30r3-Records: 2012/998/1000 statt 2006/1140/1538 (nvml1/0/2),
also -142/-538 MiB auf den 3080ern und +6 auf der 5090.

Test test_weg2_d_reserve_record_max_0929: 6/6 rot auf b1ed70a, gruen.
D-EXPECT-Wiring-Test angepasst (der einzige direkte Record-Lesezugriff in
main ist der fuer Gruppe D). 34 Nachbardateien: 823 passed, 5 Altlasten
identisch, 0 neue Rote.
efschu pushed a commit to efschu/htsglang that referenced this pull request Sep 29, 2026
…ITY UNRESOLVABLE), kein stilles CONSTANT

29.09. x27ra: Ein Trockenlauf auf einem git-archive-Baum ohne .git konnte fuer
keinen Record belegen, dass sein Boot-Commit Vorfahr des Baum-HEAD ist. Belegen
liess sich nur ein Boot auf der deklarierten Linie 103712c. Dadurch fiel
beim 27B jeder Record still auf CONSTANT:
- sgl-project#1444 DC-RESIDUE 3148/2652/2652 statt RECORD 2106/1676/1676
- P lag 1040 MiB unter dem echten Boot z30j

CalibrationIdentity.head_unresolvable_line() liefert eine laute Zeile, sobald
der line-Term den HEAD des Baums nicht lesen kann. Die Zeile nennt den Baum,
die deklarierten Linienkoepfe und die Folge ("falls to CONSTANT"). main
druckt sie direkt unter der Zeile 'WEG2-PROFILE calibration identity'.
Verhalten sonst unveraendert: Es gibt keinen neuen Riegel, und Records werden
wie bisher angenommen oder verworfen.

Test test_weg2_record_identity_head_0929 (4):
- Baum ohne git wird benannt.
- Lesbarer HEAD schweigt.
- Ohne line-Term schweigt die Identitaet.
- Aufrufkante in main vorhanden.
Auf bd1bc7a rot (Methode fehlt).

7 Nachbardateien (CalibrationIdentity/is_line_ancestor): 135 passed auf
Zweig und Basis, 0 neue Rote.
efschu pushed a commit to efschu/htsglang that referenced this pull request Oct 6, 2026
… product they stand in for (no product code)

grid_claim_1298 / leg2_store_probe_1290: the cache stub binds the real _weg2_extent_topup (fnFL2x36) and the _reduce double answers the prefetch_cut_rank MIN (sgl-project#249) -- 13 red -> green.
tail_fold_store_h63d: _Tree takes the real _weg2_mark_park_end, _P0 the real _weg2_note_lost_anchors (5 -> 0).
xchg_seam_wiring_1273: the carrier decision reads weights_cpu_backup_armed() since sgl-project#1369, so the fake flag now drives the real predicate through its on/off env knob (5 -> 0).
front_lan_1321: the namespace carries weg2_weight_source (sgl-project#1444) (2 -> 0).
region_form_lane_cap_1464: xchg_max_tag_bytes answered by a stand-in, it is not the subject (1 -> 0).
xchg_bounce_phases_1330: the adapter test drives a real manager with the sequential transport doubled (run_sequential_units replaced run_bounce_leg in sgl-project#1378 xsn45); the second-expression sweep is the lane family, not park_l3.py:164 (2 -> 0).
xchg_solo_draft_xsn389: the draft-region fixture is one manifest per (rank, region) as the writer produces them; a shadow goes empty (sgl-project#105), a non-empty rank that moves nothing is still refused; a tensor held whole by a subset joins REPLICATED with widths (whole,0,0) (sgl-project#76) (4 -> 0).
unify_standard_form_profile: the kv resume selector reads "kv_cache" in tags (FLIPCYCLE H6) (1 -> 0).

Rest of the triage: /spinning/gpu-arb/deskq/done/1533-bericht.md.
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.

2 participants