wai-quantum 0.4.4: the fabric's CUDA member #243

Merged

dcharlot merged 1 commit from wai/quantum-every-processor into main 2026-10-14 08:55:18 -04:00

Conversation 0 Commits 1 Files changed 17 +1600 -37

dcharlot commented 2026-10-14 01:14:36 -04:00

Owner

Copy link

What

A second accelerator member for the fabric, beside the WebGPU one: the gate kernel in CUDA C. The CUDA runtime compiler builds it at run time for each device's own architecture, and each device is certified bit-exact against the CPU reference's pinned corpus digests. It is behind the feature quantum_fabric_cuda, which is opt-in and native-only. This is wai-quantum 0.4.4.

Why it is bit-exact

src/quantum_fabric/lanes64.cu transcribes lanes::exact_row and apply_words into 64-bit integers, which every CUDA device has:

• each fx(a, b) = (a·b + 2^29) >> 30 is exact in 64 bits, and so is each row's sum;
• the lane keeps the sum's low 32 bits;
• a sum outside 32 bits records the gate's op-log index, and the lowest index wins;
• each thread handles one amplitude pair and each launch is one gate, so no thread reads what another writes within a launch.

The member's probe checks the device's arithmetic before it joins. Its certificates check whole runs.

Shape

The member presents the same surface as the WebGPU member:

• discovery, with typed refusals;
• the arithmetic probe, whose pair generator and mismatch check now live in lanes and are shared by both members;
• run_lanes and run_program; simulate_from and simulate_metered; certify, member and member_metered;
• a Runner for the Fabric.

Its id is local/gpu-cuda-<12 hex>, binding the device, its UUID, PCI location, driver, compiler and memory. Its CI key is gpu-cuda-sm<major><minor>.

The driver and runtime compiler are loaded on first use, so a build with the feature runs on a host without CUDA. There it finds no member and says why (Discovery::unavailable).

Verified

On the RTX 4050 (sm_89), tests/cuda_member.rs passes 9 of 9:

Check	Result
Arithmetic probe	agrees on 65,536 pairs
Kernel vs run_words	equal on 2,000 random states and gate words, out-of-range runs included
Pinned certificates	all four reproduced: two gatesets × random-v1 and fabric-v1
Random circuits, up to 22 qubits, and QFTs	equal to native
Refusals	typed
Kernel mutants	all 6 caught
Receipt, placement, Fabric run	as expected
Repeated runs	identical

On a Mac without CUDA, the same tests pass over no member.

The mutation testing caught one weak mutant of my own. A sign flip on the imaginary part of the top-left entry changes nothing, because that entry is real for every gate in the set, so it was replaced with a flip on an entry Y does have.

CI

wai-quantum-cuda.yml adds two jobs:

• on the GPU host: all four certificates and every mutant caught are required;
• on a runner without CUDA: the member must report that no CUDA device is present.

Lockfiles

Every lockfile that includes wai-quantum moves to 0.4.4 by its version line alone. check-lockfiles passes on all seven.

🤖 Generated with Claude Code

## What A second accelerator member for the fabric, beside the WebGPU one: **the gate kernel in CUDA C**. The CUDA runtime compiler builds it at run time for each device's own architecture, and each device is certified bit-exact against the CPU reference's pinned corpus digests. It is behind the feature `quantum_fabric_cuda`, which is opt-in and native-only. This is wai-quantum **0.4.4**. ## Why it is bit-exact `src/quantum_fabric/lanes64.cu` transcribes `lanes::exact_row` and `apply_words` into 64-bit integers, which every CUDA device has: - each `fx(a, b) = (a·b + 2^29) >> 30` is exact in 64 bits, and so is each row's sum; - the lane keeps the sum's low 32 bits; - a sum outside 32 bits records the gate's op-log index, and the lowest index wins; - each thread handles one amplitude pair and each launch is one gate, so no thread reads what another writes within a launch. The member's probe checks the device's arithmetic before it joins. Its certificates check whole runs. ## Shape The member presents the same surface as the WebGPU member: - discovery, with typed refusals; - the arithmetic probe, whose pair generator and mismatch check now live in `lanes` and are shared by both members; - `run_lanes` and `run_program`; `simulate_from` and `simulate_metered`; `certify`, `member` and `member_metered`; - a `Runner` for the `Fabric`. Its id is `local/gpu-cuda-<12 hex>`, binding the device, its UUID, PCI location, driver, compiler and memory. Its CI key is `gpu-cuda-sm<major><minor>`. The driver and runtime compiler are loaded on first use, so a build with the feature runs on a host without CUDA. There it finds no member and says why (`Discovery::unavailable`). ## Verified On the RTX 4050 (sm_89), `tests/cuda_member.rs` passes 9 of 9: | Check | Result | |---|---| | Arithmetic probe | agrees on 65,536 pairs | | Kernel vs `run_words` | equal on 2,000 random states and gate words, out-of-range runs included | | Pinned certificates | all four reproduced: two gatesets × random-v1 and fabric-v1 | | Random circuits, up to 22 qubits, and QFTs | equal to native | | Refusals | typed | | Kernel mutants | all 6 caught | | Receipt, placement, `Fabric` run | as expected | | Repeated runs | identical | On a Mac without CUDA, the same tests pass over no member. The mutation testing caught one weak mutant of my own. A sign flip on the imaginary part of the top-left entry changes nothing, because that entry is real for every gate in the set, so it was replaced with a flip on an entry `Y` does have. ## CI `wai-quantum-cuda.yml` adds two jobs: - **on the GPU host:** all four certificates and every mutant caught are required; - **on a runner without CUDA:** the member must report that no CUDA device is present. ## Lockfiles Every lockfile that includes wai-quantum moves to 0.4.4 by its version line alone. `check-lockfiles` passes on all seven. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

dcharlot added 1 commit 2026-10-14 01:14:36 -04:00

wai-quantum 0.4.4: the fabric's CUDA member ... 7443ce1992

A second accelerator member beside the WebGPU one: the gate kernel in CUDA
C, compiled at run time by the CUDA runtime compiler for each device's own
architecture, certified bit-exact against the CPU reference's pinned
corpus digests. Feature `quantum_fabric_cuda`, opt-in, native targets
only.

The kernel (src/quantum_fabric/lanes64.cu) is a transcription of
lanes::exact_row and apply_words in 64-bit integers, which every CUDA
device has: each fx(a, b) = (a·b + 2^29) >> 30 is exact in 64 bits, each
row's sum too, the lane keeps the low 32 bits, and a sum outside 32 bits
records the gate's op-log index, the lowest winning. One thread per
amplitude pair, one launch per gate, so no thread reads what another
writes within a launch.

The driver and the runtime compiler are loaded when first used, so a
build with the feature runs on a host with no CUDA and finds no member,
with the reason (Discovery::unavailable).

The member mirrors the WebGPU member's surface: discovery with typed
refusals, the arithmetic probe (shared now as lanes::probe_pairs and
first_mismatch), run_lanes / run_program, simulate_from and
simulate_metered, certify, member and member_metered, a Runner for the
Fabric, with_max_qubits and with_source for tests. Its id binds the
device, UUID, PCI location, driver, compiler and memory:
local/gpu-cuda-<12 hex>, keyed in CI as gpu-cuda-sm<major><minor>.

Device tests (tests/cuda_member.rs), on the RTX 4050 (sm_89): the probe
agrees on 65,536 pairs; the kernel matches run_words on 2,000 random
states and gate words, out-of-range runs included; all four pinned
certificates reproduced; random circuits up to 22 qubits and QFTs equal
to native; typed refusals; all six kernel mutants caught; a receipt, a
placement and a Fabric run; repeated runs identical. On a host with no
CUDA the same tests pass over no member.

CI: wai-quantum-cuda.yml, on the GPU host (requires the four
certificates and every mutant caught) and on a runner with no CUDA
(requires the member to say so).

Every lockfile that includes wai-quantum moves to 0.4.4 by its version
line alone.

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

dcharlot merged commit 8869612c27 into main 2026-10-14 08:55:18 -04:00

dcharlot referenced this pull request from a commit 2026-10-14 08:55:19 -04:00

Merge pull request 'wai-quantum 0.4.4: the fabric's CUDA member' (#243) from wai/quantum-every-processor into main

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

Transaction-Science/open-standards!243

No description provided.