wai-quantum: jobs placed on certified GPU members or split over several, byte for byte the CPU's; Clifford+T as sums of stabilizer states #557

Merged

dcharlot merged 7 commits from wai/quantum-scale-out into main 2026-10-31 14:53:38 -04:00

Conversation 0 Commits 7 Files changed 49 +10009 -116

2026-10-31

Transaction Science

wai-quantum-py: the getting-started notebook's method list names stabilizer_rank ...

The notebook prints every method with its accuracy, and the Python
workflow holds each cell to its stored output. The stabilizer-rank
method is new in this branch, so its line goes into the stored output,
in the engine's order. Nothing else in any notebook's outputs changed.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

bda065b52f

• View at this point in history

2026-10-31 13:38:54 -04:00

Transaction Science

wai-quantum: the agent docs blessed and the views page rebuilt with the fabric, the stabilizer-rank method and everything before them ...

The views page was rebuilt with wai/wai-quantum-views/build.sh in catthehacker/ubuntu:act-22.04
with Rust 1.98.0, as the views workflow rebuilds it.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

6c4f6d5154

• View at this point in history

2026-10-31 09:54:34 -04:00

Transaction Science

quantum-studio: the stabilizer-rank sum by its plain name, and the views page rebuilt ...

The Studio's method names gain stabilizer_rank, "Stabilizer-rank sum" in
English and its Simplified Chinese, so a plan or result made by it shows a
name rather than the method's key. The Studio's words feed the bundled views
page, so the page is rebuilt in the CI image (act-22.04, Rust 1.98.0,
wasm32-unknown-unknown), where build-views --check holds it to its 48
sources; it carries this branch's engine too.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

a11492f913

• View at this point in history

2026-10-31 09:54:34 -04:00

Transaction Science

wai-quantum: auto keeps the state vector on the CPU until a member is measured faster ...

With one launch per gate and the state copied to the device and back, a
whole job measured about as fast, or slower on the GPU, between 18 and 22
qubits on the fleet's GPUs (22 qubits on the laptop GPU: 159 ms on the CPU,
202 ms placed). A planner must not route by default to a path that is not
better, so `auto` now places on the CPU on every host and looks for no
member: a job that does not ask runs as it always did, its result unnamed.

A member runs a job only when the job asks: `accelerator` takes the first
certified member that holds the state, and one that fails during the run
hands the job to the next, the receipt listing each that failed
(`fell_back_from`, now a list, each with how it failed); when every one
fails the job is refused (J013). A member named by id or key is alone.
Discovery, certification, the receipt's fabric block, the split and
`wai-quantum fabric` are as they were; `fabric_members` states `auto`
instead of a threshold. A test with stand-in devices holds the fall-back to
the CPU's bits and to the schema.

The hardware legs ask for `accelerator` and check that auto names nothing;
both replayed on jetson-hub and framework-128gb. Docs, help, both MCP
servers' descriptions, the guide and the changelog say so; agent docs
re-blessed.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

780b0526ce

• View at this point in history

2026-10-31 09:54:34 -04:00

Transaction Science

wai-quantum: a build without the fabric honours only auto placement, and says so ...

The engine without quantum_fabric (the desktop app, the browser engine, the
contract's build) reads every `fabric` and refuses all but auto (J013), with a
test of its own; the placement tests that need the fabric are gated on it, and
the helpers only the fabric uses are compiled only with it, so those builds
carry no dead code.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

21b9369471

• View at this point in history

2026-10-31 09:54:34 -04:00

Transaction Science

wai-quantum: Clifford+T circuits as sums of stabilizer states (stabilizer_rank) ...

A T gate is (e^{iπ/8} I + e^{−iπ/8} S) / (2 cos π/8), and T† its conjugate,
so a circuit with t of them is a sum of 2^t Clifford circuits of equal
weight. quantum_stabrank walks every branch on the phased tableau, whose new
PhasedTableau::amplitudes gives every amplitude of a stabilizer state exactly
from one reduction of the stabilizers (d = x ⊕ r in the X span, the phase
read off the product of pivot rows). Every term is a power of 1/√2 at a whole
number of sixteenths of a turn, counted as a whole number (ExactSum), so the
sum does not depend on how the branches are split between threads and is
evaluated once in double precision. An expectation value is the amplitude
⟨0|U†PU|0⟩ of one circuit with twice the T gates.

Exact mode: terms grow as 2^t (4^t for an expectation value). Sampled mode
(terms, or epsilon, at confidence): branches drawn uniformly, seeded, each
from a stream of its own; the estimate is unbiased with mean squared distance
(sec²(π/8)^t − 1)/terms, so each value is within
sqrt((sec²(π/8)^t − 1)/(terms·(1 − confidence))) with the stated probability,
computed from t, terms and confidence and stated beside each value.

Program::clifford_t lowers t, tdg, p/rz/rx/ry at odd multiples of π/4, cp at
odd multiples of π/2, ccz and ccx, with the global phase kept exactly.
Ground truth first: every amplitude of random Clifford+T circuits (up to 7
qubits, 10 T gates) and of random programs through the job engine equals the
dense state vector's, phase and all, within 1e-12, by fidelity; expectation
values likewise; the sampled bound holds in at least its stated fraction of
seeds, measured on the whole state. The planner names stabilizer_rank and
auto chooses it by T count (amplitudes or expectation values past the dense
state vector, its branches within one job's work, calibrated at about 4e10
tableau word operations a second on a 16-core host). Optional t_count, terms,
bound and confidence in the result schema; the methods list, both MCP
servers, the service and the docs name it.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

57638f48a5

• View at this point in history

2026-10-31 09:54:34 -04:00

Transaction Science

wai-quantum: the state vector placed on the fabric's members, and split over several ...

A job's new `fabric` says where the statevector method's state runs: auto
(the default), cpu, accelerator, a member's id or key, or a list of members
(a power of two), the state split by its top qubits into one partition on
each. The CUDA, HIP and OpenCL members' f64 state vectors are discovered
once per process and certified against the CPU on the pinned corpus; only a
member certified bit-exact is placed, so a placed job's answers are the
CPU's byte for byte. A placement the host cannot honour is refused (J013),
never made elsewhere. With no member asked for and none placed, the plan and
the result are what they always were: the existing pins hold.

The partitioned run (sv64::run_split) runs the launches the CPU plans on each
partition's amplitudes; a gate on a partition qubit first has that qubit
exchanged with a local one through the host's copy. Ground truth first: CPU
partitions equal the CPU's state vector bit for bit over the pinned corpus at
every partition count and on wide states with exchanges. On the fleet, one
state split four ways over a laptop GPU (by two backends), the CPU by OpenCL
and the CPU reference gives the CPU's answers byte for byte, as does the
integrated GPU by HIP beside OpenCL.

The result carries `fabric` (member, parity tier, certificate digest, corpus,
energy unmetered with the reason), sealed by the job's receipt, and its
provenance names the members: optional members of the job request, job result
and provenance schemas, and fabric_members in the tools schema. Surfaces:
`wai-quantum run --fabric` and `wai-quantum fabric`, the CLI's cuda, hip and
opencl features, `fabric` in both MCP servers' job tools and `list_fabric` on
the command line's, the service reading it. Members gain run_launches and
max_qubits. Hardware legs in the CUDA and HIP workflows, absent legs and the
CPU-partition tests by name in containers.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

a0bd65d6af

• View at this point in history

2026-10-31 09:54:34 -04:00