Status of testing Apache Mahout Qumat 0.6.0rc1

[ryankert01]

Testing Apache Mahout Qumat 0.6.0rc1

We kindly request all contributors to help test the Release Candidate for qumat 0.6.0rc1 and qumat-qdp 0.2.0rc1.

Requirements

• Python 3.10, 3.11, or 3.12
• For QDP extension: Linux with NVIDIA GPU and CUDA

Installation

# Core package
pip install --pre "qumat==0.6.0rc1"

# With QDP extension
pip install --pre "qumat[qdp]==0.6.0rc1"

Changes to Test

• docs: fix broken blog link to the quantum computing primer (#1300) (#1306): @bearomorphism
• [QDP] Close AMD-vs-CUDA encoder parity: add iqp, iqp-z, phase (#1292): @ryankert01
• [QDP] Add Mahout-AMD and PennyLane-AMDGPU frameworks to throughput benchmark (#1291): @ryankert01
• docs(qdp): sync getting-started guide with current engine and encoding behavior (#1290): @ryankert01
• [QDP] Add AMD backend selection to QDP encoding benchmarks (#1289): @400Ping
• docs(qdp): sync python-api page with current qumat_qdp (#1287): @ryankert01
• docs(qdp): document phase and iqp-family encodings (#1286): @ryankert01
• [Feature] Make QDP CUDA kernel build targets configurable and future-compatible (#1283): @viiccwen
• [Feature] Add batch float32 zero-copy angle encoding in qdp-python bindings (#1278): @viiccwen
• feat(qdp): Encoding + Dtype enums, static encoder dispatch (#1276): @rich7420
• [Feature][QDP] F32 support for angle/basis encoders, fidelity metrics and pipeline improvements (#1275): @rich7420
• feat: Add batch float32 zero-copy angle encoding in qdp-core (#1268): @viiccwen
• feat: Add float32 zero-copy angle encoding in Python bindings (single sample) (#1264): @viiccwen
• fix: remove hardcoded IQP qubit limit and reuse shared validation (#1261): @viiccwen
• feat: add IQP / IQP-Z latency and throughput benchmark (#1257): @viiccwen
• [Docs] Rename community meeting link to Subscribe (#1255): @400Ping
• [Feature][QDP] IQP kernel fusion and persistent kernel / grid-stride optimizations (#1253): @400Ping
• [QDP] Add AMD Dockerfile without startup smoke test (#1251): @400Ping
• feat: Add float32 zero-copy angle encoding in qdp-core (single sample) (#1249): @viiccwen
• [Fix] Add AMD docker context (#1247): @400Ping
• [QDP] Add Dockerfile.qdp-amd for AMD environment testing (#1246): @400Ping
• [Website] Fix CSP violation by self-hosting cubes.png texture (#1244): @karanjot-gaidu
• perf(qdp): Implement async prefetching and native f32 dispatch pipelines (#1242): @rich7420
• fix(website): katex and google font not found (#1139) (#1240): @grorge123
• [Fix] Website main page image (#1223): @400Ping
• Add phase encoding (#1199): @vvvdwbvvv
• Hoist loop-invariant norm computation in iqp_encode_batch_kernel_naive (#1198): @wdskuki
• Initial effort on adding type hints to qumat (#1190): @bearomorphism
• Add pure-PyTorch reference implementations (#1189): @ryankert01
• refactor: iqp_encode_kernel and iqp_encode_batch_kernel (#1188): @ryankert01
• [QDP] Add SVHN IQP encoding benchmark with PennyLane baseline and QDP pipeline (#1186): @ryankert01
• [Website] [Deploy] Add check to ensure deploy on apache/mahout repo instead of my local fork (#1185): @ryankert01
• [QDP] Add angle and basis encoding tests for numpy input (#1179): @yehfelareborn
• [merge it when release] Add pypi badge (#1176): @ryankert01
• feat: add SVHN Quantum Kernel SVM benchmark (#1175): @rich7420
• [QDP] Add GCS remote URL support (#1174): @400Ping
• docs: fix link in CONTRIBUTING.md (#1165): @bearomorphism
• feat: add MNIST amplitude encoding benchmarks for PennyLane and QDP pipelines (#1161): @ryankert01
• [Feature][QDP] Add AMD GPU support via Triton backend (#1158): @400Ping
• [Docs] Fix website getting-started url (#1157): @0lai0
• [QDP] Enhance benchmark's installation journey (#1154): @0lai0
• docs: self host katex.min.css to fix not css loading issue (#1140): @bearomorphism
• docs: fix doc version inconsistency in navbar (#1138): @bearomorphism
• docs: use relative path to avoid link to docs with different version (#1137): @bearomorphism
• [QDP] Optimize run_throughput_pipeline to avoid per-iteration Vec allocations (#1136): @0lai0
• [QDP] Remove circular qumat dependency from qumat-qdp (#1135): @guan404ming
• docs: merge CONTRIBUTING.md into README.md and add troubleshooting guide, update stale information (#1134): @bearomorphism
• docs: add frontmatter to every docs/**/*.md (#1133): @bearomorphism
• docs: better hot reload (#1132): @bearomorphism
• docs: improve website math format (#1131): @bearomorphism
• [QDP] Add remote URL support for cloud object storage (S3) (#1130): @guan404ming
• docs(katex): remove custom .katex style to address math expression duplication issue (#1129): @bearomorphism
• docs(README): update basic-gates link to point to the documentation site (#1128): @bearomorphism
• docs(basic-gates.md): fix broken LaTeX math in GitHub Markdown (#1127): @bearomorphism
• docs(README): update broken internal docs link (#1126): @bearomorphism
• [QDP]Remove duplicate cuda_error_to_string in gpu/memory.rs and reuse crate error mapping (#1124): @0lai0
• [Doc] PR policy and Review Guidelines (#1122): @400Ping
• [QDP] Unified venv Environment documentation (#1121): @400Ping
• [Website][Docs] Redirect website & unified download Page (#1119): @400Ping
• feat: add CUDA float32 batch amplitude encoding support in Python bindings (#1116): @viiccwen
• feat: support CUDA tensor GPU-pointer encoding for iqp and iqp-z (#1115): @viiccwen
• Improve test coverage for qiskit_backend.py to 99% (#1112): @Howardisme
• Add default null handling to run_throughput_pipeline_py function (#1110): @ryankert01
• fix: misaligned vector loads for odd-length samples in batch amplitude and norm kernels (#1108): @viiccwen
• Add unified and configurable null handling (#1101): @ryankert01
• Fix/dynamic cuda encoding method list (#1100): @SuyashParmar
• Added the run_throughput_pipeline_py PyO3 binding function so make te… (#1098): @ryankert01
• [QDP] feat: improve iris benchmark example (#1095): @rich7420
• Enable CUDA tensor validation for iqp and iqp-z encodings (#1093): @SuyashParmar
• Update README with QDP installation instructions (#1092): @ryankert01
• [Docs] Fix Biweekly Community Meeting Section (#1090): @400Ping
• Align Iris amplitude encoding benchmark between PennyLane and QDP (#1088): @rich7420
• [Docs] Update how-to-contribute page to use GitHub Issues #1077 (#1087): @alisha-1000
• [Examples] [QDP] re-org files & add simple example for qdp (#1085): @ryankert01
• [Docs] [QDP] add qdp example (#1079): @ryankert01
• avoid mutating user backend config (#1076): @shiavm006
• Fix Cirq backend circuit mutation when adding measurements (#1072): @shiavm006
• added PyPI classifiers for frameworks and environment (#1051): @SuyashParmar
• [Bug] Fix website Redirect (#1046): @400Ping
• Update the docs versioning (#1040): @SuyashParmar
• feat: Add classifiers to pyproject.toml (#1034): @ryankert01
• MAHOUT-901: Fix test_rust by excluding qdp-python from cargo test (#1031): @viiccwen
• [QDP] Add zero-copy amplitude batch encoding from float32 GPU tensors (#1029): @viiccwen
• [QDP] I/O errors: expose underlying cause via Error::source() (#1026): @rich7420
• [QDP] Support float32 CUDA amplitude encoding in Python bindings (#1025): @viiccwen
• [Docs] Update Angle Encoding Core Concepts Doc (#1022): @400Ping
• chore(qdp-python): remove deprecated PyO3 extension-module feature (#1021): @rich7420
• [QDP] DLPack Tensor unsafe Cleanup Refactoring (#1011): @0lai0
• [QDP] File and streaming Quantum Data Loader (#1010): @rich7420
• [QDP] Delegate GPU pointer encoding to QuantumEncoder trait (#1006): @viiccwen
• [QDP] Add a Quantum Data Loader and API refactor (#1000): @rich7420
• [QDP] Add zero-copy amplitude encoding from float32 GPU tensors (#999): @viiccwen
• [QDP] Extend GpuStateVector to support Float32 precision (#995): @viiccwen
• Re organized docs/ structure (#992): @ryankert01
• [Docs] update qumat getting started! (#991): @ryankert01
• [QDP] Baseline and Data Collection (#972): @rich7420
• [QDP] Add f32 GPU inverse-norm helper for amplitude encoder (#971): @viiccwen
• [QDP] Fix invalid CUDA kernel launch when num_samples exceeds grid dimension limit (#968): @viiccwen
• Update release docs and scripts (#964): @guan404ming
• docs: add instructions for creating versioned documentation (#961): @ryankert01
• [QDP] Replace panic with Result error for Arrow IPC capacity overflow (#960): @shiavm006
• [Website] Fix typo (#955): @400Ping
• [QDP] Add observability tools to diagnose pipeline performance (#945): @rich7420
• MAHOUT-900 [IQP] Implement FWT optimization for IQP encoding (#938): @ryankert01
• [QDP] DLPack input extraction (#935): @400Ping
• [QDP] basis GPU‑pointer support (#934): @400Ping
• [QDP] Extend CUDA tensor direct GPU encode to angle (#932): @400Ping
• [QDP] PyTorch CUDA stream‑aware encode for GPU tensors (#930): @400Ping
• [QDP] PyTorch GPU Pointer Validation (#929): @400Ping
• MAHOUT-802: Add float32 L2 norm reduction kernel for batch processing (#918): @viiccwen
• MAHOUT-800: Unified CONTRIBUTING.md (#910): @viiccwen

Compatibility

• Python 3.10
• Python 3.11
• Python 3.12

Thanks to all who contributed to the release: @0lai0 @400Ping @alisha-1000 @bearomorphism @grorge123 @guan404ming @Howardisme @karanjot-gaidu @rich7420 @ryankert01 @shiavm006 @SuyashParmar @viiccwen @vvvdwbvvv @wdskuki @yehfelareborn

Committer

• I acknowledge that I am a maintainer/committer of the Apache Mahout project.

[ryankert01]

+1, find a small bug that doesn't affact release.

[andrewmusselman]

Verified RC1 builds and tests cleanly on Linux/CUDA 12.4 with the
following environment notes:

• GPU: NVIDIA GeForce GTX 1060 (Pascal, sm_61)
• CUDA Toolkit: 12.4 (apt nvidia-cuda-toolkit)
• Host compiler: required CXX=g++-13 (gcc 14 default is unsupported
  by CUDA 12.4 headers)
• Pascal arch: required QDP_CUDA_ARCH_LIST=61+PTX
• PyTorch: required CUDA_VISIBLE_DEVICES= to mask the sm_61 GPU
  (current wheel ships sm_70+ only)

With those, make tests and uv run pytest -v pass end-to-end and
cargo llvm-cov reports 67.21% region coverage on qdp-core.

Three developer-experience hiccups hit along the way which I'll file
as separate non-blocking issues for 0.6.1 with patches attached:

1. qdp-core declares CUDA Runtime API extern symbols but doesn't
   emit its own -lcudart directive; relies on qdp-kernels'
   build.rs, which only emits when nvcc is found. Driver-only
   hosts get a confusing 12-symbol linker error instead of a
   clean build via the stub path.

2. Makefile's HAS_NVIDIA check gates test_rust on nvidia-smi, which
   succeeds with just the driver. test_rust actually requires the
   CUDA Toolkit (nvcc + libcudart). Should gate on HAS_NVCC and
   print a clear "install the toolkit" message otherwise.

3. qumat_qdp.loader and qumat_qdp.api pick cuda:N based on
   torch.cuda.is_available() alone, without intersecting the
   device's compute capability against torch.cuda.get_arch_list().
   Result: on a Pascal GPU with a recent PyTorch wheel, 12 tests
   error with cudaErrorNoKernelImageForDevice instead of either
   passing on CPU fallback or skipping with a clear reason.

None of these are release-blockers — the library itself works correctly
on supported hardware (Volta+) and the tests pass once the environment
matches the project's implicit assumptions. Filing as 0.6.1 followups.

[rich7420]

+1, find a little bug and it has been merged #1324

[viiccwen]

+1

[guan404ming]

+1

[400Ping]

btw has anyone tested on AMD GPUs, this release has introduced triton backend for AMD ecosystems

[0lai0]

+1