cuda/README.md

RANGE CUDA preview

This preview publishes the retained Phase60 CUDA pipeline, previously installed and measured in Phase61. Its CUDA binary and private exact CPU fallback are byte-identical to that installation. The current integration changes package imports and connects the pipeline to the maintained RANGE API.

Public Python package release is coming soon. These are preview wheels, not a PyPI release. The hosted app runs on CPU; downloading its CUDA preview gives you a separate local GPU execution path.

Install the preview

Use 64-bit Windows, CPython 3.10, and an NVIDIA sm75 GPU with a compatible CUDA 12 driver/runtime. The original qualification used an RTX 2070 Max-Q. Other GPU architectures and Linux CUDA execution are not supported by this preview. CPU-only users can install the CPU wheel without the CUDA companion.

Download both matching files from the Space's downloads folder:

• RANGE CPU wheel
• RANGE CUDA companion

From the folder containing both downloads, create a Python 3.10 environment and install both files together:

py -3.10 -m venv .venv
.\.venv\Scripts\python.exe -m pip install .\range_curate-0.1.0a3-cp310-cp310-win_amd64.whl .\range_curate_cuda-0.1.0a3-cp310-cp310-win_amd64.whl

Pip installs the companion's pinned CuPy, NumPy, PyArrow and CUDA-pathfinder dependencies. No RANGE Rust or CUDA compilation is needed to install these wheels. A compatible NVIDIA driver/runtime is still required. The port was checked with existing installed dependencies; a fresh dependency download on another machine has not been qualified.

Run bulk Arrow screening

from datetime import datetime, timezone
import pyarrow as pa
from range_curate.examples.general_economy import RelevanceEngine

engine = RelevanceEngine(backend="cuda", workers=10)
data = pa.table({
    "text": ["The Federal Reserve cut interest rates."],
    "timestamp": pa.array(
        [datetime(2025, 1, 15, tzinfo=timezone.utc)],
        type=pa.timestamp("us", tz="UTC"),
    ),
})
facts = engine.screen_arrow(data).to_pyarrow()

CUDA applies to bulk screen_arrow. Rich predictions, explanations, list screening, and the domain-neutral CurationEngine continue to run on CPU. A tiny example demonstrates the API, not GPU throughput. The CUDA input is a materialized Arrow Table or RecordBatch with non-null text and timezone-aware timestamps.

You can pass a supported custom Taxonomy using taxonomy=.... The pipeline uses exact CPU fallback for unsupported rules and rows that exceed its GPU bounds. Missing or incompatible CUDA raises a clear error; it does not silently substitute a CPU run for a failed GPU runtime. Choose backend="cpu" explicitly for CPU use.

What the speed record means

Phase60 measured 1,833,059 articles/s and the Phase61 installation measured 1,809,133 articles/s on their recorded General Economy workload. This preview contains that retained pipeline. Six finite installed correctness checks passed for the current API port. No performance experiment was rerun, at the maintainer's request; this is not a new timing claim for the renamed wheel.

The kernel SHA-256 is dd943d9adac702e41c2bb0a2b58f78698498edd8a3dfab16cf9d224c8c6a83fa. The private CPU fallback SHA-256 is 8b1dddd69ccc13380cced1f2df4977f2ecb750805c9d127b0c8046fef812563f.

See the integration record, source provenance, and benchmark conditions. The source tree includes the retained orchestration and CUDA kernels. build_wheel.py assembles the companion from the hash-bound Phase61 reference wheel and the matching current CPU wheel; it preserves third-party notices and records all packaged file hashes.