Skip to main content

Introducing Apache Mahout qumat 0.6.0

· 7 min read
Ryan Huang
Ryan Huang
Apache Mahout Committer
Mahout TeamMahout Team
Mahout Team
Apache Mahout

We're excited to announce the release of Apache Mahout qumat 0.6.0, rolling up 111 pull requests from contributors across the community since 0.5.0. This release brings AMD GPU support to first-class status, closes encoder parity gaps between CUDA and ROCm, adds new benchmarks for real-world datasets, and overhauls the documentation site.

Thank you to everyone who contributed to this release: Ryan Huang, Vic Wen, Jie-Kai Chang, Guan-Ming (Wesley) Chiu, Tim Hsiung, Kuan-Hao Huang, ChenChen Lai, Suyash Parmar, Yehfela, Shivam Mittal, Alisha, Hsien-Cheng Huang, Eddie Tsai, Andrew Musselman, Xin Hao, wdskuki, Trevor Grant, Karanjot Gaidu, Howardisme, and Han-Wen Tsao.

Key Highlights​

  • QDP encoding parity and new encodings — Phase, IQP, and IQP-Z encodings on both CUDA and AMD ROCm; float32 zero-copy batch paths and GPU-pointer encoding for the IQP family.
  • AMD GPU support — New Mahout-AMD framework with hand-written Triton kernels running natively on ROCm; AMD is now selectable from all QDP encoding and throughput benchmarks.
  • New benchmarks — SVHN Quantum Kernel SVM, MNIST amplitude encoding, and IQP latency/throughput benchmarks with PennyLane baselines.
  • Documentation overhaul — Frontmatter across all pages, self-hosted KaTeX, a new troubleshooting guide, and CONTRIBUTING merged into the README for easier onboarding.

Let's explore these in more detail.

QDP Encoding Parity and New Encodings​

Prior to 0.6.0, the CUDA and AMD ROCm backends had different encoder coverage — some encoding paths only existed on one backend, which made benchmarking apples-to-apples comparisons difficult and limited ROCm's usability in practice.

In 0.6.0, Phase, IQP, and IQP-Z encodings now ship on both NVIDIA CUDA and AMD ROCm backends, closing the parity gap with the existing angle and amplitude paths.

On the performance side, this release adds:

  • Float32 zero-copy batch paths for angle and basis encoders — both single-sample and batched, in both qdp-core (Rust) and the Python bindings — eliminating host-device copies in hot loops.
  • GPU-pointer encoding for the IQP family — pass a CUDA tensor directly via DLPack and skip the host round-trip entirely.
  • IQP kernel fusion and grid-stride optimizations — persistent kernels and fused encode passes reduce kernel launch overhead.
  • Async prefetching and native f32 dispatch pipelines — overlaps I/O with compute for throughput-bound workloads.

Data-to-state latency scaling (ms/vector, log scale) for amplitude, angle, and basis encoding. QDP (orange) achieves orders-of-magnitude lower latency than both CPU and GPU backends.

Figure: Data-to-state latency scaling (ms/vector, log scale) for amplitude, angle, and basis encoding. QDP (orange) achieves orders-of-magnitude lower latency than both CPU and GPU backends.

Install the QDP extra to try these:

pip install "qumat[qdp]==0.6.0"  # Linux x86_64 + NVIDIA CUDA

AMD GPU Support​

0.6.0 makes AMD ROCm a first-class backend via hand-written Triton kernels (TritonAmdEngine) that run natively on ROCm without going through PennyLane. The Triton path covers amplitude, angle, basis, phase, IQP, and IQP-Z encodings — the same set now available on CUDA — and is selectable from the QDP encoding and throughput benchmarks via --qdp-backend amd.

PennyLane-AMDGPU (lightning.amdgpu) appears in the throughput benchmark as a comparison baseline, not as Mahout's implementation path.

Docker images for the AMD environment are also included (Dockerfile.qdp-amd), making it easy to spin up a reproducible AMD test environment without a physical ROCm machine.

CUDA kernel build targets are now configurable for forward compatibility — specify the target compute capabilities at build time rather than hard-coding them, so the wheel stays valid on architectures released after the build.

New Benchmarks​

0.6.0 adds three new benchmarks that cover more realistic workloads than the synthetic microbenchmarks from earlier releases:

SVHN IQP variational classifier — trains a variational IQP classifier on the Street View House Numbers dataset (digit 1 vs 7, 200 samples, 200 iterations) on two RTX 3090 Ti GPUs. QDP offloads the IQP encoding step to the GPU in a single one-shot pass before training begins, keeping the training backend independent of encoding cost. QDP GPU consistently outperforms PennyLane GPU at every qubit count — 41 vs 33 samples/s at 4 qubits and 20 vs 15 samples/s at 10 qubits (~1.35×).

SVHN IQP training throughput: QDP GPU vs PennyLane GPU at 4–10 qubits (RTX 3090 Ti)

SVHN Quantum Kernel SVM — runs a precomputed squared inner-product quantum kernel SVM on SVHN amplitude-encoded features (12 qubits), measuring end-to-end pipeline time from raw pixels to SVM prediction. This is the first benchmark that exercises the full pipeline (feature encoding → kernel matrix → SVM fit/predict) rather than just the encoding step.

Data-to-State latency — isolates the full pipeline from CPU RAM to GPU-ready quantum state at 16 qubits (65,536-dimensional vectors). Mahout (QDP) delivers 0.160 ms/vector — 4.5× faster than PennyLane (0.716 ms), 56× faster than Qiskit Statevec (9.030 ms), and 477× faster than Qiskit Initialize (76.243 ms). PennyLane and Mahout both target GPU; Qiskit runs on CPU with Qiskit Init adding full circuit decomposition and transpilation overhead on top.

Data-to-state latency: Mahout 0.160 ms, PennyLane 0.716 ms, Qiskit Statevec 9.030 ms, Qiskit Init 76.243 ms (log scale)

DataLoader amplitude throughput — streams 12,800 amplitude-encoded vectors in batches of 64 at 16 qubits. Mahout (QDP) delivers 6,101 vectors/s — 3.8× faster than PennyLane (1,604 vectors/s).

Amplitude DataLoader throughput: Mahout 6,101 vs PennyLane 1,604 vectors/s

All benchmarks support AMD backend selection via --qdp-backend amd, so you can do a direct CUDA-vs-ROCm comparison on the same workload.

Documentation Overhaul​

The docs site received a significant refresh:

  • Frontmatter added to every docs/**/*.md page — fixes SEO metadata that was missing from most pages.
  • Self-hosted KaTeX — math rendering now works offline and no longer depends on the CDN, fixing the "KaTeX not found" error reported by several contributors.
  • Troubleshooting guide — a new guide covers the most common install and runtime issues, especially around GPU detection and CUDA/ROCm version mismatches.
  • CONTRIBUTING merged into the README — reduces the number of places a new contributor needs to read to get started.
  • Type hints added to qumat — initial pass adding Python type annotations, laying the groundwork for better IDE support and static analysis.
  • PR policy and review guidelines — documents the merge criteria and review expectations in one place.

Other Improvements​

  • Cloud storage support — the QDP data loader now supports S3 and GCS remote URLs in addition to local paths, enabling benchmarks and pipelines that read directly from object storage.
  • Encoding and Dtype enums — Encoding and Dtype are now proper Python enums rather than bare strings, with static dispatch in the encoder.
  • Pure-PyTorch reference implementations — added alongside the CUDA kernels for correctness comparison and CPU fallback.
  • Ruff rules expanded — E and most ANN rules are now enforced; document type-checking CI added.
  • pytest-xdist — parallel test execution cuts CI wall-clock time on multi-core machines.

Getting Started​

# Core package
pip install qumat==0.6.0

# With GPU-accelerated QDP extension (Linux x86_64 + NVIDIA CUDA)
pip install "qumat[qdp]==0.6.0"

We welcome feedback and contributions on the dev@mahout.apache.org mailing list and on GitHub.