NVIDIA/cudaqx 0.8.0

0.8.0

Repository: NVIDIA/cudaqx

Tag: 0.8.0

Published: 2026-09-12T21:59:10Z

Prerelease: no

Release notes:

CUDA-Q QEC 0.8.0

This is a release of CUDA-Q QEC version 0.8.0. (CUDA-Q Solvers is not included in this release; see CUDA-Q Algorithms for a replacement for CUDA-Q Solvers.)

CUDA-Q QEC 0.8.0 is a decoder-focused release centered on new decoder options. The most notable change is:

• A new NV Fusion Decoder. `nv_fusion_decoder` is a multi-threaded minimum-weight perfect matching decoder that combines *fusion blossom* with PyMatching's *sparse blossom*: the matching graph is partitioned into temporal blocks solved independently and then fused across their boundaries, dispatched to a worker pool as syndrome data arrives, so it is designed for low-latency streaming/realtime decoding as well as offline batch decoding (contributed by @tlshannon). This is currently closed source, so source code for this decoder is not available in GitHub at this time.

Additionally, nv-qldpc-decoder (closed source) gains two new features. The `osd_init_method="min_llr"` option makes BP+OSD far more accurate on full circuit-level (joint XYZ) decoding problems, cutting the logical error rate by up to 117x while running significantly fewer BP iterations. The `relay_solutions` helper records every Relay BP convergence during a decode, so the whole RelayBP-N stop_nconv trade-off curve can be computed offline from a single run instead of re-decoding once per N value (61 runs in the documented example).

Please check out the docs and examples for how to get started using CUDA-Q QEC.

> Dependency note: CUDA-Q QEC 0.8.0 depends on CUDA-Q 0.16.

---

Features and Enhancements (QEC) 🎉

Decoders

• Add NV Fusion Decoder QEC plugin - a new multi-threaded fusion-blossom/sparse-blossom MWPM decoder for streaming and batch decoding, with decoder stats, latency speedups, and auto-selected block_leaf_size by @tlshannon.
• Allow a realtime decoder to be constructed from a raw Stim DEM (stim_dem_path in YAML) by @tlshannon in https://github.com/NVIDIA/cudaqx/pull/801 — a third way (alongside flat-matrix and dem_chunks forms) to describe a decoder's error model on the realtime path, enabling DEM-native decoders such as Chromobius there. In particular, this lets a realtime color-code predecoder pipeline chain Chromobius in as the global decoder (stim_dem_path reaches the nested global decoder via global_decoder_params["stim_dem"]), since Chromobius rejects a parity-check-matrix representation outright and previously could not be configured on the realtime path at all. See NVIDIA/Ising-Decoding for the color-code predecoder this is designed to pair with.
• Add common decoder-stats functions (decoder_stats) for consistent latency/replay logging across decoders by @tlshannon in https://github.com/NVIDIA/cudaqx/pull/802.
• Add [Core] rvalue insert(key, T&&) overload to heterogeneous_map, avoiding unnecessary deep copies of large payloads (e.g. syndrome/LLR history) by @bmhowe23 in https://github.com/NVIDIA/cudaqx/pull/781.

Bindings / examples

• Replace msm execution contexts with cudaq::dem_from_kernel in app examples by @bmhowe23 in https://github.com/NVIDIA/cudaqx/pull/696.
• Accept both cc.StdvecType and cc.SequenceType in py_code, tracking a CUDA-Q internal rename by @bmhowe23 in https://github.com/NVIDIA/cudaqx/pull/786.

nv-qldpc-decoder Updates (Closed Source)

New features and options

• `osd_init_method="min_llr"` (@bmhowe23) — a new OSD seeding mode that initializes the OSD solver with the minimum-LLR-magnitude scoreboard rather than the existing default. This makes BP+OSD far more accurate on full circuit-level (joint XYZ) decoding problems while running significantly fewer BP iterations.
• `relay_solutions` post-processing helper (@bmhowe23) — a pure-NumPy module that reconstructs, offline, what relay-BP would have returned for *any* stop_nconv setting from a single uncapped recording run, plus batch_opt_results plumbing to carry relay_solutions records through the decoder API. It records every relay-BP convergence (not just the accepted one).
• Speed up BP LLR-history collection on the sparse-GPU path (@bmhowe23).

Bug Fixes (QEC) 🐛

• Fix PyMatching parallel-edge merge mapping and realtime observable output (native decode_to_obs output instead of an error-frame-to-observable projection) — cuts realtime PyMatching decode latency by ~30-35% on GB200 by @vedika-saravanan in https://github.com/NVIDIA/cudaqx/pull/799.
• Reduce realtime decode latency by resolving detectors as measurements arrive and reusing PyMatching scratch buffers (steady-state decode() now allocates nothing) by @tlshannon in https://github.com/NVIDIA/cudaqx/pull/812.
• Fix decoder lifetime during async decoding by @kaiqiy-nv in https://github.com/NVIDIA/cudaqx/pull/711.
• Validate AI decoder RPC slot capacity to prevent gateway output from overwriting adjacent slots by @kaiqiy-nv in https://github.com/NVIDIA/cudaqx/pull/764.

Breaking Changes & Deprecations ⚠️

• CUDA-Q Solvers removed from this repository (#796, #797): Solvers is deprecated in favor of the new CUDA-Q Algorithms library; future development in this repo targets QEC only, and Algorithms development continues at https://github.com/NVIDIA/cudaq-algorithms.

Documentation ✏️

• Add a gamma-ensemble Relay-BP performance-tuning guide by @eliotheinrich in https://github.com/NVIDIA/cudaqx/pull/755, plus the accompanying standalone benchmarking scripts by @eliotheinrich in https://github.com/NVIDIA/cudaqx/pull/766.
• Correct realtime decoding example commands (surface_code_1.py filename, --emulate opt-in, Quantinuum payload provider args) by @vedika-saravanan in https://github.com/NVIDIA/cudaqx/pull/749.
• Add versioning to the docs site (gh-pages version switcher) by @anjbur in https://github.com/NVIDIA/cudaqx/pull/761.
• Update doc layout to match the 0.7.0 release branch's layout while keeping main's content by @melody-ren in...