triton-csprng

triton-csprng is a small PyTorch/Triton package for counter-based random streams on CPU and NVIDIA GPUs. It provides ChaCha20-backed tensor generation and sampling primitives for cryptography-adjacent, simulation, and FHE-style workloads.

The package exposes low-level stream and sampling building blocks without depending on any downstream library's RNG API.

What is implemented

• ChaCha20 block generation with vectorized PyTorch CPU and Triton CUDA paths.
• Explicit key / nonce / counter stream state.
• Raw uint32(...) and bytes(...) APIs returning ordinary PyTorch tensors.
• Bounded integer sampling with scalar or per-channel bounds.
• Centered integer discrete Gaussian sampling from a 128-bit half-plane CDT.
• Stochastic rounding for CPU and CUDA floating tensors.
• RnsRandomStreams, a convenience manager for RNS-like layouts with:
  • independent streams per device for non-repeated channels;
  • repeated channels that reproduce the same values across devices;
  • state-dict roundtrip for deterministic continuation.
• No C++/CUDA extension or torch.ops registration step.

Installation

For ordinary package use after a PyPI release:

python -m pip install triton-csprng

For development:

git clone git@github.com:visualDust/triton-csprng.git
cd triton-csprng
python -m pip install -e ".[dev]"

Runtime dependencies are PyTorch, NumPy for vectorized CPU sampling, and mpmath for high-precision CDT construction. CPU execution does not require Triton. CUDA execution additionally requires Triton; install triton-csprng[cuda] after installing the PyTorch build that matches the target CUDA stack.

Quick start

import torch
from triton_csprng import ChaCha20Rng

rng = ChaCha20Rng(
    key=list(range(8)),      # 8 little-endian uint32 words = 256 bits
    nonce=[123, 456],        # 2 little-endian uint32 words = 64 bits
    counter=0,
    device="cuda:0",
)

words = rng.uint32((1024,))
raw = rng.bytes((4096,))
mod_q = rng.randint([17, 257], (2, 1024))
gauss = rng.discrete_gaussian((4, 1024), sigma=3.2)
rounded = rng.stochastic_round(torch.randn(1024, device="cuda:0"))

Every result above is a normal PyTorch tensor. Use device="cpu" to select the CPU implementation without changing the stream or sampling APIs.

Stream semantics

ChaCha20Rng is a deterministic counter-based stream:

from triton_csprng import ChaCha20Rng

rng1 = ChaCha20Rng(key=list(range(8)), nonce=[1, 2], counter=9)
rng2 = ChaCha20Rng(key=list(range(8)), nonce=[1, 2], counter=9)

assert torch.equal(rng1.uint32((3, 7)), rng2.uint32((3, 7)))

The stream buffers unused bytes internally, so chunked reads are equivalent to a single larger read:

one_shot = ChaCha20Rng(key=list(range(8)), nonce=[5, 6])
chunked = ChaCha20Rng(key=list(range(8)), nonce=[5, 6])

expected = one_shot.uint32(40)
got = torch.cat([chunked.uint32(17), chunked.uint32(23)])
assert torch.equal(got, expected)

Stream state can be checkpointed:

state = rng1.state_dict()
restored = ChaCha20Rng.from_state_dict(state)

Sampling APIs

Bounded integers

x = rng.randint(17, (4, 1024))
y = rng.randint([17, 257, 65537], (3, 1024))

The output dtype is torch.int64; bounds therefore must fit in signed int64. Multiple bounds are interpreted as leading channels. Internally the sampler consumes a 128-bit ChaCha word U and returns floor(bound * U / 2**128). This has no retry or fallback branch. Over the finite 128-bit source domain, output bucket counts differ by at most one, so the statistical distance from the ideal uniform distribution is bounded by roughly bound / 2**128.

Discrete Gaussian

e = rng.discrete_gaussian((8, 32768), sigma=3.2)

The sampler builds a 128-bit half-plane cumulative distribution table using mpmath precision 2 * security_bits, chooses num_sampling_points = 2**ceil(log2(6*sigma)), reserves one random bit for the sign, and folds the remaining truncated tail into the last bucket. This keeps the CUDA path compact and constant-shape while making the finite table construction explicit and testable.

Stochastic rounding

rounded = rng.stochastic_round(values)

values must be a CUDA floating tensor on the same device as the RNG stream. The result is torch.int64 with Bernoulli rounding by the fractional part of abs(values), then sign restoration.

RNS-style stream manager

RnsRandomStreams helps express layouts where each configured CPU or CUDA device gets independent non-repeated channels, while repeated channels are generated from matching streams on every device.

from triton_csprng import RnsRandomStreams

streams = RnsRandomStreams(
    num_coeffs=32768,
    channel_counts=[8, 8],
    repeated_channels=2,
    devices=["cuda:0", "cuda:1"],
    key=list(range(8)),
    nonce=[1, 2],
)

u32 = streams.uint32_channels()
gauss = streams.discrete_gaussian_channels(sigma=3.2)
ints = streams.randint_channels([
    [17] * 8 + [257] * 2,
    [19] * 8 + [257] * 2,
])

For each returned list item:

shape = [non_repeated_channels + repeated_channels, num_coeffs]

The repeated tail channels are reproducible across devices when their bounds and distribution parameters match.

Why there is no torch op wrapper

A Triton kernel can be launched directly with PyTorch CUDA tensors:

out = torch.empty_like(x)
_kernel[grid](x, out, ...)

CUDA paths launch Triton kernels this way; CPU paths use ordinary PyTorch tensor operations. A torch.ops.* custom op is unnecessary for normal Python/PyTorch integration and would reintroduce dispatcher/wrapper maintenance. A torch.library wrapper can still be added later if a downstream project needs formal fake-tensor or torch.compile dispatcher integration.

Developer checks

Install the optional development tools and pre-commit hooks:

python -m pip install -e ".[dev]"
pre-commit install

Run the same checks manually:

python -m ruff check .
python -m ruff format --check .
python -m pytest tests -q

Validation

Current local validation:

python -m ruff check .
python -m ruff format --check .
python -m pre_commit run --all-files
python -m pytest tests -q

The tests cover:

• ChaCha20 CPU and Triton output against a Python reference implementation;
• non-multiple block counts;
• stream determinism and chunking behavior;
• state-dict restore;
• bounded integer range and multiply-high mapping checks;
• difficult bounds and distribution sanity;
• half-plane CDT table shape/range checks;
• rough discrete Gaussian moments and symmetry;
• stochastic-rounding determinism and integer cases;
• exact seeded CPU/CUDA stream and sampling parity;
• CPU RNS state restore and repeated-channel equality across two GPUs.