Massively parallel execution work tracker¶
Parent: MPP-00 (#818). Status: architecture and issue plan only, 2026-11-02. No distributed implementation, scientific acceptance or performance result is claimed by this increment. The programme is unscheduled in the existing production MPI and large-system scaling milestone; it adds no #557 gate.
Read the architecture and
measurement/acceptance contract. The configured
library collection massive-parallel-vibe-qc-2026-11-02 supplies full reading
and acquisition provenance; public DOI/section anchors are in the architecture.
Missing DBCSR/Bussy/GauXC/DMRG/HIP-DFPT sources remain explicit uncertainties.
The #804 descriptor/completion ADR specifies portable identities, bounded leases, publication and synthetic complex fixtures. The bounded CPU binding/completion slice now has portable byte envelopes and real 1/2/3/empty-rank checks through the opt-in #820 transfer adapter. GPU and physical distributed consumers remain open; no scaling result is claimed.
Ownership reconciliation¶
GitLab issue descriptions and recent ownership/evidence notes were refreshed on 2026-11-02, including #793-#814, #472/#594, #580/#581, #617/#639, #613, #96 and #144. These were open at inspection. Their titles, assignees, existing work and scientific authority were preserved. #798/#799 have landed bounded CPU source slices, not a complete distributed implementation. #594 has scoped native evaluated-state evidence; consume only the admitted route. #96 GDF scaling and #144 experimental chi output farming retain their own acceptance.
Create no second tensor/runtime/Fock/integral framework. #804 owns common placement, remote leases and completion; #803 owns GPU event/transport mechanics. The new MPP issues own missing integration/capacity work. #805-#812 retain method registration and scientific migration. Their linked distribution addenda identify the next capability rather than copying a second backlog. No owner is reassigned and no worker chat is launched by this plan.
New bounded work¶
Issue |
Deliverable |
First increment |
|---|---|---|
MPP-01 (#819) |
Baseline inventory, cost records and multi-node benchmark pilots |
Implement the portable record and two fixture manifests, then capture/reuse qualified baseline evidence; leave unmeasured hardware explicit. |
MPP-02 (#820) |
MPI runtime ownership, collective errors and rank/thread policy |
Write the phase/thread ownership ADR and implement one bounded admission/error protocol with real uneven/empty-rank coverage. |
MPP-03 (#821) |
Complete multi-k SCF pools with global occupations and coherent output |
Pin one route and implement global occupation/reduction fixtures plus its k-pool SCF seam; retain the serial reference path. |
MPP-04 (#822) |
q/pair-owned periodic GDF factors without full-cache allgather |
Define the q/pair provider and ownership census, then run a bounded fixed-density complex exchange slice without allgathering factors. |
MPP-05 (#823) |
Molecular distributed HF/DFT through the existing native builder seam |
Audit the coarse native RI-J seam and implement one fixed-density distributed operation using existing tensor and builder contracts. |
MPP-06 (#824) |
Distributed Gamma GPW grid and FFT-Poisson capacity slice |
Write grid-layout/normalization and workspace ADR, compare one pinned optional FFT prototype to serial without changing the production dependency set. |
MPP-07 (#825) |
Optional distributed generalized eigensolver and matrix-state capacity |
Pin the required solver operations, compare API/build/license coverage and memory models, then one synthetic generalized solve through the existing operator contract. |
MPP-08 (#826) |
Partition-independent checkpoint and relaunch restart for one distributed route |
Select one supported safe boundary and write the versioned manifest/global-key contract, then interruption and repartition tests without broad format changes. |
Reused capability and method owners¶
Issue |
Distributed increment |
Capability prerequisites |
|---|---|---|
#804 (TENS-11) |
Specify distribution and completion fields, then one bounded distributed complex contraction over existing tensor providers. |
MPP-02 (#820) protocol; specific #795/#796/#797/#800 capabilities only. ADR and synthetic fixture design can start now; consume landed #798/#799 pieces without waiting for umbrella closure. |
#803 (TENS-10) |
Extend one existing event/lease operation with bounded staged transport before a direct device-buffer experiment. |
#804 completion contract, MPP-02 (#820) owner/progress policy, existing #797/#800 and qualified G1 operation. Local GPU work does not wait for distributed consumers. |
#805 (TENS-12) |
Audit and document one RI-J native builder registration used by MPP-05 (#823), implementing only a missing bounded seam. |
Existing #798/#797 for local registration; MPP-05 (#823)/#804 only for distributed integration. Preserve current local migration priorities. |
#806 (TENS-13) |
Implement one fixed-source DF-MP2 pair reduction with full preparation accounting, then its complete-method validation. |
#804 ownership/provider/reduction and MPP-02 (#820); #798 local contraction and #611 reference; MPP-04 (#822) only for consumers actually requiring distributed periodic factors. |
#807 (TENS-14) |
Choose a small admitted pair graph with at least one cross-owner coupling and match the serial fixed-amplitude residual. |
#804 remote leases/reduction, MPP-02 (#820); #799/#801 needed ragged capabilities and existing #580/#581. Reuse #585/#586 transport/solver scientific work rather than replacing it. |
#808 (TENS-15) |
Fix a small determinant space and implement bounded owner-based sigma contribution aggregation against exact serial action. |
#804 ownership/credits and MPP-02 (#820), required #801/#802 sparse/provider capabilities; existing admitted local operation. Inventory fixed-space action now, then implementation as capabilities land. |
#809 (TENS-16) |
Distribute a fixed TDA block action and global dots against the same-source serial operator before integrating the full root solve. |
#804/#797 operation/lease/reduction and MPP-02 (#820); migrated or registered admitted operator, #612 state contract. Does not require MPP-07 (#825) dense solver. |
#810 (TENS-17) |
Select one supported analytic derivative with admitted primal state and distribute a bounded contraction plus its full force comparison. |
Accepted primal route and #804 derivative provider/reduction; MPP-02 (#820) and existing #636/#637 capabilities only where needed. No unrelated family closure. |
#811 (TENS-18) |
Bind one existing complex periodic provider to stable mesh/source keys and reconstruction checks used by MPP-04 (#822). |
#799/#801 source descriptors and #804 transport where distributed; MPP-03 (#821)/04/06 own their bounded execution. #144 existing chi farming evidence remains separately open; #577/#607/#617 scientific owners unchanged. |
#812 (TENS-19) |
Select an admitted SCC route, enumerate charge/energy contributions and prototype one bounded reduction without altering its nonlinear iteration. |
#804 reduction/ownership, MPP-02 (#820) and existing #795-#797 capabilities; #609/#610 admitted route. Periodic overlay #811 only where relevant; local migration can proceed independently. |
Dependency and start order¶
flowchart TD
B["MPP-01 (#819) baselines"] --> K["MPP-03 (#821) complete k SCF"]
R["MPP-02 (#820) runtime protocol"] --> T["804 distributed ownership"]
C["797 and 800 needed resource/planner capabilities"] --> T
T --> K
T --> M["MPP-05 (#823) molecular HF/DFT"]
T --> F["MPP-04 (#822) q/pair GDF storage"]
K --> F
T --> G["MPP-06 (#824) GPW grids/FFT"]
T --> E["MPP-07 (#825) distributed solver"]
T --> U["803 GPU transport"]
T --> X["806-812 method-specific distribution"]
K --> S["MPP-08 (#826) checkpoint/relaunch"]
M --> S
Arrows mean the particular interface/capability used by an increment, not completion of the preceding issue. In particular fixed-density MPP-04 (#822) work can start before complete MPP-03 (#821) acceptance; MPP-08 (#826) needs one admitted route, not both. CPU method work does not wait for GPU transport. Layout/solver/FFT prototypes do not require full tensor or all-method closure. Reference physics and numerical-library licensing/API/build decisions precede only affected integration, not baseline or protocol design.
Ready now: MPP-01 (#819) baseline/record work and MPP-02 (#820) runtime policy/tests. #804 can start its descriptor/completion ADR and synthetic fixtures using landed CPU contracts, specifying missing async capabilities explicitly. MPP-05 (#823) can inventory its native builder seam; MPP-06 (#824)/07 can compare layouts and pinned optional backends; MPP-08 (#826) can define its schema boundary. These are bounded preparatory increments, not authorization to bypass implementation prerequisites or claim a working multi-node backend.
Route coverage and state¶
The MPP-01 (#819) inventory records each supported method/source/spin/k/derivative variant, entry point, current replica/communication behavior, owner, first slice, source landing, numerical acceptance and measured scaling. Expand it as source inspection finds variants; do not infer implementation from a roadmap family name. Coverage initially maps as follows:
Route |
Distribution owner |
Scientific/registration owner |
|---|---|---|
Molecular direct/DF/seminumerical HF/DFT and AO transforms |
MPP-05 (#823), #804 |
#805, #472/#594, #595/#596/#597, #617 |
Periodic GDF k-SCF and q/pair factors |
MPP-03 (#821)/MPP-04 (#822) |
#601/#600, #811, #581/#592 |
GPW grids and optional FFT |
MPP-06 (#824) |
#603/#605/#811; augmentation separately #604 |
Generalized Fock solvers and distributed AO state |
MPP-07 (#825) |
#472/#594/#615 |
Canonical MPn/CC/triples |
#806 |
#611 plus existing method owners |
Coupled local correlation |
#807 |
#580/#581/#585/#586 and #508 scope |
CI/CAS/MR/DMRG |
#808 |
Existing admitted method/solver owners |
Response, EOM, properties |
#809 |
#612 and producer method |
Nuclear/cell derivatives, Hessians and stress |
#810 |
Primal/response owner, #636/#637 |
Ewald/COSX/GAPW/BIPOLE and experimental CCM/chi overlays |
#811 |
#577/#607/#617 and route-specific owners, #144 only in its scope |
Supported semiempirical and experimental SECCM |
#812 |
#609/#610; #811 for periodic overlays |
Device transport |
#803 |
Existing GPU stages and #639 MANTA qualification |
Distributed output/checkpoint/relaunch |
MPP-08 (#826) |
#613 and |
Each route has separate source landed, scientifically accepted, and scaling measured states. Record exact source/native identity, applicable reference, measured observables, limitations and next action. A route may have landed source while runtime delivery or numerical/scaling acceptance remains pending. Unsupported physics stays explicit. Parent closure requires the inventory’s declared distributed scope to be reconciled; merely registering an adapter is insufficient. Optional compiler/compression/ULFM research remains separate and cannot keep otherwise completed exact execution open.
Pickup and completion¶
Each linked issue contains the concrete deliverable/exclusions, source paths, ownership/lifetime/error contract, dependencies, primary literature locators, focused tests, complete-method reference checks, memory/communication/performance acceptance and a self-contained development-chat prompt. Read its latest notes before claiming a bounded increment and files. Preserve existing active work.
Dev runs focused checks, lands source on main with CI skipped and completes own
cleanup. Validator owns applicable complete vq and reference acceptance and
may be the same worker; pending downstream work does not extend a finished Dev
assignment. Do not close scientific/scaling acceptance on source tests alone.
Set numerical tolerances from method evidence and performance targets from
baseline/pilot measurements before optimized runs. Keep all private operational
records outside the checkout; retain portable summaries and citations here.