Engine Capabilities and Version Boundaries¶
简体中文 · Previous: Import and migration · Next: CLI and build
This chapter is for Engine operators and advanced users checking exact capability boundaries. Studio does not calculate these quantities itself; it negotiates explicit Engine capability and strictly validates returned artifacts.
Current route matrix¶
| Route | User capability | Project | Execution | Current status/boundary |
|---|---|---|---|---|
| v0.1 | 4D Twiss optics | v0.1 | Interactive one-shot | Available, SCREENING |
| v0.1 | Generic CPU tracking | v0.1 | Durable | Available, at most 50M particle-turns |
| v0.2 | Supplied-slice radiation–RF equilibrium | v0.2 | One-shot | Available, result/0.1 |
| v0.3 | Lattice-derived slices + equilibrium | v0.2 | One-shot | Available, 8/32/128 convergence, result/0.2 |
| v0.4 | Matched distribution | v0.2 + qualified equilibrium/0.2 | One-shot | 128–8192 particles; linear/planar/uncoupled |
| v0.5 | Local momentum acceptance | v0.2 + qualified equilibrium/0.2 | One-shot, ≤1800 s | Sparse aperture, not lifetime |
| v0.6 | Touschek | v0.2 | Historical durable | Formula bug; no new submit, recovery only |
| v0.7 | Corrected Touschek | v0.2 + beam/equilibrium gates | Durable | Screening lifetime, not total lifetime |
| v0.8 | Exact-lattice DA/FMA | v0.2 | Durable | CPU fixed-δ 4D finite-turn |
| v0.9 | Orbit error/correction | v0.2 + reviewed profile | Durable | Planar reference-δ ideal model |
| v0.10 | v0.8 Exact-lattice DA/FMA | v0.2 | Durable Slurm | Deployment-bound CPU scheduling; unchanged v0.8 physics |
| v0.11 | Native Xsuite load gate | Attached native v0.2 package | One-shot | Exact-source reconstruction only; not a physics result |
| v0.12 | Native Xsuite equilibrium | Attached native v0.2 package | One-shot | Radiation–RF screening on original Line |
| v0.13 | Native matched distribution | Attached native v0.2 package | One-shot | Recomputes v0.12; linear/planar/uncoupled |
| v0.14 | Native 4D Twiss | Attached native v0.2 package | One-shot | Original Line; rejects c_minus >= 1e-6 |
| v0.15 | Native radiation slices | Attached native v0.2 package | One-shot | Independent 8/32/128 convergence diagnostic |
| v0.16 | Native local momentum acceptance | Attached native v0.2 package | One-shot | Sparse finite-turn screening at 12 locations |
| v0.17 | Native tracking/nonlinear/orbit | Attached native v0.2 package | One-shot CPU | Synthetic ensemble and virtual devices |
| v0.18 | v0.8 Exact-lattice DA/FMA | v0.2 | Durable Slurm/CUDA | ContextCupy Xsuite tracking; no CPU fallback; unchanged physics boundary |
One-shot means no durable reconnect/cancel contract after client interruption. Durable work has a stable remote identity, state, and reconnect path.
Accepted Engine sources¶
These identities are deployment evidence, not project inputs:
| Route | Accepted/required Engine source |
|---|---|
| v0.1 | e50a6e6952cbb85fd1e3a831e4b5b2010b9b6bd9 |
| v0.2 | dc9860b43cc9a2d929e0568df814f28d733270bc |
| v0.3 | fdab5539893d66bf37bf05fb4ab246ecab1476b7 |
| v0.4 / v0.5 | f4189eba95469be1e286d6169e2781376c4e58c8 |
| v0.6 historical | 5063daaa2ac6501645742c6997fc72d46679aabb |
| v0.7 | 79fad5d1bb3d2f3b3b9be417d6eb23a6bb1332a9 |
| v0.8 current gate | 5253d6758042381bab15ed9c385aef78bd0331f6 |
| v0.9 current gate | 3b14dbcc5e525235a44db32ad96e20be5c33e3ef |
| v0.10 Slurm scheduling gate | b05e977b77f28bfbf98c95b6947ca77b3daea819 |
| v0.11–v0.17 current unified deployment | ecd919a6b0ffd5edb16386ac6a2cfeaffefe1719 |
| v0.18 Slurm/CUDA gate | e9677a5a50415ef35b062c116bc0474e2bab4891 |
Do not assume a newer commit satisfies an older exact identity. Deploy accepted releases and use separate Studio targets for incompatible gates.
v0.8 fixed limits and exclusions¶
- CPU-only turns 512/1024/2048;
- up to 7 deltas, 2,048 turns, 8,192 final launches, and 65,536 convergence trials;
- radial/angle intervals 8/16/32 and magnet slices 2/4/8;
- three artifacts: lattice projection, nonlinear screening, job summary;
- no RF kick, radiation, quantum excitation, collective effects, errors/correction, strong coupling, or GPU;
- source
5253…runs directly through frozen v0.8; sourceb05e…may wrap the same CPU physics in v0.10 deployment-bound Slurm scheduling without changing its qualification boundary.
v0.10 Slurm scheduling boundary¶
- Engine installation declares Slurm
requiredoroff; Studio does not infer scheduler policy from SSH reachability. - For a
requireddeployment, Studio requests a fresh catalog immediately before Submit, shows only Engine-authorized partitions/offers, and returns the selected opaque signed offer. - Partition, account, QoS, and raw resource flags remain Engine-owned. Studio never constructs an
sbatchcommand from user text. - This route is CPU-only and supports only the frozen nonlinear operation/profile above. It is not a generic Slurm or GPU execution interface.
v0.18 GPU scheduling boundary¶
- The accepted Engine source is exactly
e9677a5a50415ef35b062c116bc0474e2bab4891. - Studio accepts only Engine-signed
cuda_gpu,xsuite-cuda/0.1, one-GPU-per-node offers. - Engine discovery and worker Submit both include
--gpus-per-node=1. - Xsuite tracking must bind ContextCupy, device, CUDA driver/runtime, CuPy, and
trackingExecutedOnCuda=true; Studio verifies the evidence before retention. - Projection, optics, post-processing, and independent PyAT replay stay on CPU. CUDA failures are terminal and never cause CPU fallback.
- Real 107 A100 acceptance passed with detached Submit, fresh-process reconnect, continuous phase recovery, and strict Fetch of the three expected artifacts in 1459.44 seconds.
- This remains the frozen finite-turn screening operation, not a generic GPU/Slurm RPC or machine dynamic-aperture result.
v0.9 fixed limits and exclusions¶
- one reviewed portable-FODO profile;
- maximum 512 element definitions, 128 occurrences/BPMs/correctors, 512 seeds, 32 scales;
- formal profile uses 8 BPMs, 16 correctors, 8 seeds, 3 scales;
- four artifacts: error model, correction screening, tolerance summary, job summary;
- no coupling/skew, dispersion/chromaticity, LOCO, survey/measured response, hysteresis, power-supply dynamics, GPU, or SLURM.
Explicitly unavailable system capabilities¶
Beyond exact v0.18 nonlinear Xsuite tracking, there is no generic GPU or Slurm RPC execution, arbitrary workflow daemon/dashboard, normal-mode coupling, strong-coupling correction, complete collective/IBS/wake/space-charge model, DA/FMA/ lifetime under realistic errors, measured-machine loop, or AI-agent automatic execution interface. Read-only OFF switches cannot be bypassed by editing JSON.
Compatibility checklist¶
- Confirm SSH principal and runtime user.
- Point the executable to the accepted build.
- Check expected protocol/namespace/build/source in Hello/Capabilities.
- Check CPU, permissions, durability/reconnect/fetch semantics.
- Check request/project/artifact limits.
- Keep v0.8/v0.9 state roots, namespaces, credentials, and job prefixes isolated.
- In Studio start the intended calculation, choose its source, and accept Check source; the SSH/Engine checks run first and the specialized runtime gate must still pass before execution.
- Label fixtures as reference/acceptance; never promote them into a user's formal result.
Frozen protocol and acceptance records remain under protocol/ and docs/round_*; ordinary users
do not need to read protobuf definitions.