trnrand 0.1.0__py3-none-any.whl

trnrand/__init__.py

@@ -0,0 +1,37 @@
+"""
+trnrand — Random number generation for AWS Trainium via NKI.
+
+Pseudo-random and quasi-random sequences for scientific computing.
+Reproducible seeding, standard distributions, and low-discrepancy
+sequences for quasi-Monte Carlo integration.
+Part of the trnsci scientific computing suite.
+"""
+
+__version__ = "0.1.0"
+
+# Generator
+from .generator import Generator, manual_seed, get_default_generator
+
+# Distributions
+from .distributions import (
+    uniform, normal, standard_normal, exponential,
+    bernoulli, randint, randperm, truncated_normal,
+)
+
+# Quasi-random sequences
+from .quasi import sobol, halton, latin_hypercube
+
+# Backend control
+from .nki import HAS_NKI, set_backend, get_backend
+
+__all__ = [
+    # Generator
+    "Generator", "manual_seed", "get_default_generator",
+    # Distributions
+    "uniform", "normal", "standard_normal", "exponential",
+    "bernoulli", "randint", "randperm", "truncated_normal",
+    # Quasi-random
+    "sobol", "halton", "latin_hypercube",
+    # Backend
+    "HAS_NKI", "set_backend", "get_backend",
+]

trnrand/distributions.py

@@ -0,0 +1,135 @@
+"""
+Random number distributions for Trainium.
+
+Standard distributions with reproducible seeding via Generator.
+All functions accept an optional Generator for deterministic streams.
+
+For scientific computing:
+- uniform/normal: Monte Carlo sampling, weight initialization
+- exponential: Poisson process simulation, radioactive decay
+- multinomial: importance sampling
+- truncated_normal: bounded parameter initialization
+"""
+
+from __future__ import annotations
+
+import math
+import torch
+from typing import Optional, Tuple, Union
+
+from .generator import Generator, get_default_generator
+
+
+def uniform(
+    *size: int,
+    low: float = 0.0,
+    high: float = 1.0,
+    dtype: torch.dtype = torch.float32,
+    generator: Optional[Generator] = None,
+) -> torch.Tensor:
+    """Uniform distribution on [low, high)."""
+    gen = (generator or get_default_generator()).torch_generator
+    return torch.empty(*size, dtype=dtype).uniform_(low, high, generator=gen)
+
+
+def normal(
+    *size: int,
+    mean: float = 0.0,
+    std: float = 1.0,
+    dtype: torch.dtype = torch.float32,
+    generator: Optional[Generator] = None,
+) -> torch.Tensor:
+    """Normal (Gaussian) distribution."""
+    gen = (generator or get_default_generator()).torch_generator
+    return torch.empty(*size, dtype=dtype).normal_(mean, std, generator=gen)
+
+
+def standard_normal(
+    *size: int,
+    dtype: torch.dtype = torch.float32,
+    generator: Optional[Generator] = None,
+) -> torch.Tensor:
+    """Standard normal N(0, 1)."""
+    return normal(*size, mean=0.0, std=1.0, dtype=dtype, generator=generator)
+
+
+def exponential(
+    *size: int,
+    rate: float = 1.0,
+    dtype: torch.dtype = torch.float32,
+    generator: Optional[Generator] = None,
+) -> torch.Tensor:
+    """Exponential distribution with given rate (λ).
+
+    PDF: f(x) = λ exp(-λx)
+    Mean: 1/λ
+    """
+    gen = (generator or get_default_generator()).torch_generator
+    u = torch.empty(*size, dtype=dtype).uniform_(0, 1, generator=gen)
+    # Clamp to avoid log(0)
+    u = torch.clamp(u, min=1e-10)
+    return -torch.log(u) / rate
+
+
+def bernoulli(
+    *size: int,
+    p: float = 0.5,
+    dtype: torch.dtype = torch.float32,
+    generator: Optional[Generator] = None,
+) -> torch.Tensor:
+    """Bernoulli distribution: 1 with probability p, 0 otherwise."""
+    gen = (generator or get_default_generator()).torch_generator
+    return torch.empty(*size, dtype=dtype).bernoulli_(p, generator=gen)
+
+
+def randint(
+    *size: int,
+    low: int = 0,
+    high: int = 1,
+    dtype: torch.dtype = torch.long,
+    generator: Optional[Generator] = None,
+) -> torch.Tensor:
+    """Random integers from [low, high)."""
+    gen = (generator or get_default_generator()).torch_generator
+    return torch.randint(low, high, size, dtype=dtype, generator=gen)
+
+
+def randperm(
+    n: int,
+    dtype: torch.dtype = torch.long,
+    generator: Optional[Generator] = None,
+) -> torch.Tensor:
+    """Random permutation of 0..n-1."""
+    gen = (generator or get_default_generator()).torch_generator
+    return torch.randperm(n, dtype=dtype, generator=gen)
+
+
+def truncated_normal(
+    *size: int,
+    mean: float = 0.0,
+    std: float = 1.0,
+    low: float = -2.0,
+    high: float = 2.0,
+    dtype: torch.dtype = torch.float32,
+    generator: Optional[Generator] = None,
+) -> torch.Tensor:
+    """Truncated normal distribution clipped to [low, high] (in std units).
+
+    Uses rejection sampling. Useful for bounded weight initialization.
+    """
+    result = torch.empty(*size, dtype=dtype)
+    gen = (generator or get_default_generator()).torch_generator
+
+    # Rejection sampling
+    remaining = result.reshape(-1)
+    n_remaining = remaining.numel()
+    idx = 0
+    while idx < n_remaining:
+        batch_size = min(n_remaining - idx, n_remaining * 2)
+        samples = torch.empty(batch_size, dtype=dtype).normal_(mean, std, generator=gen)
+        valid = samples[(samples >= mean + low * std) & (samples <= mean + high * std)]
+        take = min(len(valid), n_remaining - idx)
+        remaining[idx:idx + take] = valid[:take]
+        idx += take
+
+    return result

trnrand/generator.py

@@ -0,0 +1,71 @@
+"""
+Random number generator state management for Trainium.
+
+Provides a Generator class that wraps torch.Generator for CPU/CUDA
+and will wrap NKI-based on-device RNG for Trainium. Reproducible
+seeding is critical for scientific computing — same seed, same stream,
+regardless of backend.
+
+The Philox counter-based RNG is the natural NKI target because it's
+stateless (counter + key → output), embarrassingly parallel across
+tiles, and used by both PyTorch and JAX as their default engine.
+"""
+
+from __future__ import annotations
+
+import torch
+from typing import Optional
+
+
+class Generator:
+    """Seeded random number generator with backend dispatch.
+
+    Wraps torch.Generator for CPU/CUDA. On NKI, would use a Philox
+    counter-based RNG running on the GpSimd engine.
+    """
+
+    def __init__(self, seed: Optional[int] = None, device: str = "cpu"):
+        self._device = device
+        self._torch_gen = torch.Generator(device=device)
+        if seed is not None:
+            self._torch_gen.manual_seed(seed)
+        self._seed = seed
+
+    @property
+    def seed(self) -> Optional[int]:
+        return self._seed
+
+    @property
+    def device(self) -> str:
+        return self._device
+
+    def manual_seed(self, seed: int) -> Generator:
+        self._seed = seed
+        self._torch_gen.manual_seed(seed)
+        return self
+
+    def get_state(self) -> torch.Tensor:
+        return self._torch_gen.get_state()
+
+    def set_state(self, state: torch.Tensor):
+        self._torch_gen.set_state(state)
+
+    @property
+    def torch_generator(self) -> torch.Generator:
+        """Access underlying torch.Generator for use with torch ops."""
+        return self._torch_gen
+
+
+# Module-level default generator
+_default_generator = Generator(seed=None)
+
+
+def manual_seed(seed: int) -> Generator:
+    """Seed the default generator."""
+    global _default_generator
+    _default_generator = Generator(seed=seed)
+    return _default_generator
+
+
+def get_default_generator() -> Generator:
+    return _default_generator

trnrand/nki/__init__.py

@@ -0,0 +1,5 @@
+"""NKI kernel dispatch for Trainium RNG acceleration."""
+
+from .dispatch import HAS_NKI, set_backend, get_backend
+
+__all__ = ["HAS_NKI", "set_backend", "get_backend"]

trnrand/nki/dispatch.py

@@ -0,0 +1,224 @@
+"""
+NKI dispatch for random number generation.
+
+The Philox 4×32 counter-based RNG is the primary NKI target:
+- Stateless: (counter, key) → deterministic output
+- Parallel: each tile generates independently with disjoint counters
+- Used by PyTorch (cuRAND) and JAX as default engine
+
+On Trainium, Philox runs on the GpSimd engine (general-purpose SIMD)
+since it's integer arithmetic, not matmul. The Tensor Engine isn't
+useful here — RNG is compute-light, bandwidth-heavy.
+
+For scientific workloads, on-device RNG avoids the host→device transfer
+of random tensors, which can be significant for large Monte Carlo runs.
+"""
+
+from __future__ import annotations
+
+import math
+
+import torch
+
+try:
+    import neuronxcc.nki as nki
+    import neuronxcc.nki.language as nl
+    HAS_NKI = True
+except ImportError:
+    HAS_NKI = False
+
+TWO_PI = 2.0 * math.pi
+
+_backend = "auto"
+
+
+def set_backend(backend: str):
+    global _backend
+    assert backend in ("auto", "pytorch", "nki")
+    if backend == "nki" and not HAS_NKI:
+        raise RuntimeError("NKI backend requires neuronxcc")
+    _backend = backend
+
+
+def get_backend() -> str:
+    return _backend
+
+
+def _use_nki() -> bool:
+    if _backend == "nki":
+        return True
+    if _backend == "pytorch":
+        return False
+    return HAS_NKI
+
+
+# Philox 4×32-10 constants (Salmon et al., SC'11; matches cuRAND, JAX).
+PHILOX_M0 = 0xD2511F53
+PHILOX_M1 = 0xCD9E8D57
+PHILOX_W0 = 0x9E3779B9
+PHILOX_W1 = 0xBB67AE85
+PHILOX_ROUNDS = 10
+UINT32_MASK = 0xFFFFFFFF
+
+
+def philox4x32_reference(counter: torch.Tensor, key: torch.Tensor) -> torch.Tensor:
+    """CPU reference implementation of Philox 4×32-10.
+
+    Used as the conformance oracle for the NKI kernel.
+
+    Args:
+        counter: int64 tensor of shape (..., 4) — the 4 counter words per stream.
+        key:     int64 tensor of shape (..., 2) — the 2 key words per stream.
+
+    Returns:
+        int64 tensor of shape (..., 4) holding the 4 output uint32 words
+        (stored in the low 32 bits of int64 for arithmetic safety).
+    """
+    assert counter.shape[-1] == 4 and key.shape[-1] == 2
+    c0 = counter[..., 0] & UINT32_MASK
+    c1 = counter[..., 1] & UINT32_MASK
+    c2 = counter[..., 2] & UINT32_MASK
+    c3 = counter[..., 3] & UINT32_MASK
+    k0 = key[..., 0] & UINT32_MASK
+    k1 = key[..., 1] & UINT32_MASK
+
+    for _ in range(PHILOX_ROUNDS):
+        prod0 = c0 * PHILOX_M0
+        prod1 = c2 * PHILOX_M1
+        hi0 = (prod0 >> 32) & UINT32_MASK
+        lo0 = prod0 & UINT32_MASK
+        hi1 = (prod1 >> 32) & UINT32_MASK
+        lo1 = prod1 & UINT32_MASK
+        c0, c1, c2, c3 = (
+            (hi1 ^ c1 ^ k0) & UINT32_MASK,
+            lo1 & UINT32_MASK,
+            (hi0 ^ c3 ^ k1) & UINT32_MASK,
+            lo0 & UINT32_MASK,
+        )
+        k0 = (k0 + PHILOX_W0) & UINT32_MASK
+        k1 = (k1 + PHILOX_W1) & UINT32_MASK
+
+    return torch.stack([c0, c1, c2, c3], dim=-1)
+
+
+def box_muller_cpu(uniforms: torch.Tensor) -> torch.Tensor:
+    """Box-Muller transform: pairs of uniforms → standard-normal pairs.
+
+    Same algorithm as the Vector Engine kernel — kept on CPU so the
+    `normal_nki` path can be tested for distributional correctness without
+    Trainium hardware.
+
+    Args:
+        uniforms: even-length 1-D tensor of U(0, 1) samples.
+
+    Returns:
+        Tensor of the same shape, holding standard-normal samples.
+    """
+    assert uniforms.numel() % 2 == 0, "Box-Muller needs pairs"
+    u = uniforms.view(-1, 2)
+    # Clamp u1 away from 0 to avoid log(0) = -inf.
+    u1 = torch.clamp(u[:, 0], min=1e-10)
+    u2 = u[:, 1]
+    r = torch.sqrt(-2.0 * torch.log(u1))
+    theta = TWO_PI * u2
+    z = torch.stack([r * torch.cos(theta), r * torch.sin(theta)], dim=-1)
+    return z.reshape(uniforms.shape)
+
+
+def philox_uniform_cpu(
+    n_elements: int,
+    seed: int,
+    counter_offset: int = 0,
+    dtype: torch.dtype = torch.float32,
+) -> torch.Tensor:
+    """CPU Philox uniform stream — used as fallback and as the oracle for tests.
+
+    Generates `n_elements` floats in [0, 1) by emitting `ceil(n/4)` Philox
+    blocks with sequential counters starting at `counter_offset`.
+    """
+    n_blocks = (n_elements + 3) // 4
+    counters = torch.zeros(n_blocks, 4, dtype=torch.int64)
+    counters[:, 0] = torch.arange(counter_offset, counter_offset + n_blocks, dtype=torch.int64)
+    key = torch.tensor([seed & UINT32_MASK, (seed >> 32) & UINT32_MASK], dtype=torch.int64)
+    key = key.expand(n_blocks, 2)
+    out_u32 = philox4x32_reference(counters, key).reshape(-1)[:n_elements]
+    # uint32 → float32 in [0, 1) via mantissa-only conversion (cuRAND convention).
+    return (out_u32.to(torch.float64) * (1.0 / 2**32)).to(dtype)
+
+
+if HAS_NKI:
+
+    @nki.jit
+    def philox4x32_kernel(counter_lo_ref, key_lo_ref, key_hi_ref, out_ref):
+        """Philox 4×32-10 NKI kernel running on GpSimd.
+
+        Each lane processes one independent Philox stream:
+        - `counter_lo` (per-lane) is the 32-bit counter increment; the
+          remaining counter words start at zero. This is enough to give
+          each tile a disjoint 2^96-element stream.
+        - `key_lo`, `key_hi` are the two key words (broadcast across lanes).
+
+        Output: 4 uint32 words per lane, written contiguously to out_ref.
+
+        Status: scaffolded against the Philox 4×32-10 spec but not yet
+        validated on trn1/trn2. The CPU reference in `philox4x32_reference`
+        is the conformance oracle (see tests/test_nki_philox.py).
+        """
+        # Load one tile of counter_lo values (one per lane); other counter
+        # words start at 0 — they get written by the round function.
+        c0 = nl.load(counter_lo_ref)
+        c1 = nl.zeros_like(c0)
+        c2 = nl.zeros_like(c0)
+        c3 = nl.zeros_like(c0)
+        k0 = nl.load(key_lo_ref)
+        k1 = nl.load(key_hi_ref)
+
+        # 10 rounds of the Philox round function.
+        # NKI mul-hi-lo: result is i64; high 32 / low 32 isolated by shift+mask.
+        for _ in nl.static_range(PHILOX_ROUNDS):
+            prod0 = nl.multiply(c0, PHILOX_M0)  # i64
+            prod1 = nl.multiply(c2, PHILOX_M1)  # i64
+            hi0 = nl.bitwise_and(nl.right_shift(prod0, 32), UINT32_MASK)
+            lo0 = nl.bitwise_and(prod0, UINT32_MASK)
+            hi1 = nl.bitwise_and(nl.right_shift(prod1, 32), UINT32_MASK)
+            lo1 = nl.bitwise_and(prod1, UINT32_MASK)
+
+            new_c0 = nl.bitwise_xor(nl.bitwise_xor(hi1, c1), k0)
+            new_c1 = lo1
+            new_c2 = nl.bitwise_xor(nl.bitwise_xor(hi0, c3), k1)
+            new_c3 = lo0
+            c0, c1, c2, c3 = new_c0, new_c1, new_c2, new_c3
+
+            k0 = nl.bitwise_and(nl.add(k0, PHILOX_W0), UINT32_MASK)
+            k1 = nl.bitwise_and(nl.add(k1, PHILOX_W1), UINT32_MASK)
+
+        # Write 4 output words per lane interleaved.
+        nl.store(out_ref[0::4], c0)
+        nl.store(out_ref[1::4], c1)
+        nl.store(out_ref[2::4], c2)
+        nl.store(out_ref[3::4], c3)
+
+    @nki.jit
+    def box_muller_kernel(uniforms_ref, out_ref):
+        """Box-Muller transform on the Vector Engine: U(0,1) pairs → N(0,1) pairs.
+
+        Pairs of uniform inputs (u1, u2) → (z1, z2) with
+            r = sqrt(-2 ln u1),  θ = 2π u2,
+            z1 = r cos θ,  z2 = r sin θ.
+
+        Chosen over Marsaglia polar because the Vector Engine has hardware
+        cos/sin/log/sqrt; Marsaglia's rejection step kills SIMD throughput.
+
+        Status: scaffolded but awaits on-hardware validation per #2.
+        Iterate after #1 (Philox kernel) lands cleanly on trn1.
+        """
+        u1 = nl.load(uniforms_ref[0::2])
+        u2 = nl.load(uniforms_ref[1::2])
+        # Clamp u1 away from 0 to avoid log(0) = -inf.
+        u1_safe = nl.maximum(u1, 1e-10)
+        r = nl.sqrt(nl.multiply(nl.log(u1_safe), -2.0))
+        theta = nl.multiply(u2, TWO_PI)
+        z1 = nl.multiply(r, nl.cos(theta))
+        z2 = nl.multiply(r, nl.sin(theta))
+        nl.store(out_ref[0::2], z1)
+        nl.store(out_ref[1::2], z2)

trnrand/quasi.py

@@ -0,0 +1,115 @@
+"""
+Quasi-random (low-discrepancy) sequences for Trainium.
+
+Sobol and Halton sequences for quasi-Monte Carlo integration.
+These provide better convergence than pseudo-random sampling for
+high-dimensional integrals — O(1/N) vs O(1/√N).
+
+Use cases:
+- Numerical integration in quantum chemistry (electron integrals)
+- Stochastic trace estimation (Hutchinson with QMC)
+- Hyperparameter search (Sobol beats random for small budgets)
+- Importance sampling with low-discrepancy proposals
+"""
+
+from __future__ import annotations
+
+import math
+import torch
+from typing import Optional
+
+
+def sobol(
+    n_points: int,
+    n_dims: int,
+    seed: Optional[int] = None,
+    scramble: bool = True,
+    dtype: torch.dtype = torch.float32,
+) -> torch.Tensor:
+    """Sobol quasi-random sequence.
+
+    Args:
+        n_points: Number of points to generate
+        n_dims: Dimensionality
+        seed: Random seed for scrambling
+        scramble: Apply Owen scrambling for better equidistribution
+
+    Returns:
+        Tensor of shape (n_points, n_dims) with values in [0, 1)
+    """
+    engine = torch.quasirandom.SobolEngine(
+        dimension=n_dims,
+        scramble=scramble,
+        seed=seed,
+    )
+    return engine.draw(n_points).to(dtype)
+
+
+def halton(
+    n_points: int,
+    n_dims: int,
+    dtype: torch.dtype = torch.float32,
+) -> torch.Tensor:
+    """Halton quasi-random sequence.
+
+    Uses the first n_dims primes as bases.
+    Simpler than Sobol but degrades in high dimensions (>~20).
+
+    Args:
+        n_points: Number of points
+        n_dims: Dimensionality (recommend ≤ 20)
+
+    Returns:
+        Tensor of shape (n_points, n_dims) with values in (0, 1)
+    """
+    primes = _first_n_primes(n_dims)
+    result = torch.zeros(n_points, n_dims, dtype=dtype)
+    for d in range(n_dims):
+        result[:, d] = _halton_sequence(n_points, primes[d], dtype)
+    return result
+
+
+def _halton_sequence(n: int, base: int, dtype: torch.dtype) -> torch.Tensor:
+    """Generate Halton sequence for a single base."""
+    result = torch.zeros(n, dtype=dtype)
+    for i in range(n):
+        f = 1.0
+        r = 0.0
+        idx = i + 1  # Skip 0
+        while idx > 0:
+            f /= base
+            r += f * (idx % base)
+            idx //= base
+        result[i] = r
+    return result
+
+
+def _first_n_primes(n: int) -> list[int]:
+    """Return first n prime numbers."""
+    primes = []
+    candidate = 2
+    while len(primes) < n:
+        if all(candidate % p != 0 for p in primes):
+            primes.append(candidate)
+        candidate += 1
+    return primes
+
+
+def latin_hypercube(
+    n_points: int,
+    n_dims: int,
+    dtype: torch.dtype = torch.float32,
+    generator: Optional[torch.Generator] = None,
+) -> torch.Tensor:
+    """Latin Hypercube Sampling.
+
+    Stratified sampling that ensures each row and column of the
+    hypercube contains exactly one sample. Better space-filling
+    than pure random for small sample sizes.
+    """
+    result = torch.zeros(n_points, n_dims, dtype=dtype)
+    for d in range(n_dims):
+        perm = torch.randperm(n_points, generator=generator)
+        u = torch.rand(n_points, generator=generator, dtype=dtype)
+        result[:, d] = (perm.to(dtype) + u) / n_points
+    return result

trnrand-0.1.0.dist-info/METADATA

@@ -0,0 +1,122 @@
+Metadata-Version: 2.4
+Name: trnrand
+Version: 0.1.0
+Summary: Random number generation for AWS Trainium via NKI
+Author-email: Scott Friedman <scttfrdmn@gmail.com>
+License-Expression: Apache-2.0
+Project-URL: Homepage, https://github.com/trnsci/trnrand
+Project-URL: Documentation, https://trnsci.dev/trnrand/
+Project-URL: Repository, https://github.com/trnsci/trnrand
+Project-URL: Issues, https://github.com/trnsci/trnrand/issues
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering
+Classifier: Topic :: Scientific/Engineering :: Mathematics
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: torch>=2.1
+Requires-Dist: numpy>=1.24
+Provides-Extra: neuron
+Requires-Dist: neuronxcc>=2.24; extra == "neuron"
+Requires-Dist: torch-neuronx>=2.9; extra == "neuron"
+Provides-Extra: dev
+Requires-Dist: pytest>=7.0; extra == "dev"
+Requires-Dist: pytest-benchmark>=4.0; extra == "dev"
+Requires-Dist: scipy>=1.11; extra == "dev"
+Requires-Dist: mkdocs>=1.5; extra == "dev"
+Requires-Dist: mkdocs-material>=9.0; extra == "dev"
+Dynamic: license-file
+
+# trnrand
+
+[![CI](https://github.com/trnsci/trnrand/actions/workflows/ci.yml/badge.svg)](https://github.com/trnsci/trnrand/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/trnrand)](https://pypi.org/project/trnrand/)
+[![Python](https://img.shields.io/pypi/pyversions/trnrand)](https://pypi.org/project/trnrand/)
+[![License](https://img.shields.io/github/license/trnsci/trnrand)](LICENSE)
+[![Docs](https://img.shields.io/badge/docs-mkdocs-blue)](https://trnsci.github.io/trnrand/)
+
+Random number generation for AWS Trainium via NKI.
+
+Seeded pseudo-random distributions, quasi-random sequences for quasi-Monte Carlo, and on-device Philox RNG targeting the GpSimd engine.
+
+Part of the trnsci scientific computing suite ([github.com/trnsci](https://github.com/trnsci)).
+
+## Install
+
+```bash
+pip install trnrand
+
+# With Neuron hardware support
+pip install trnrand[neuron]
+```
+
+## Usage
+
+```python
+import trnrand
+
+# Seeded, reproducible generation
+g = trnrand.Generator(seed=42)
+
+# Standard distributions
+x = trnrand.normal(1000, mean=0.0, std=1.0, generator=g)
+u = trnrand.uniform(1000, low=-1.0, high=1.0, generator=g)
+e = trnrand.exponential(1000, rate=2.0, generator=g)
+
+# Quasi-random sequences (better convergence for MC integration)
+sobol_pts = trnrand.sobol(1024, n_dims=5, seed=42)
+halton_pts = trnrand.halton(1024, n_dims=3)
+lhs_pts = trnrand.latin_hypercube(100, n_dims=4)
+
+# Module-level seeding
+trnrand.manual_seed(42)
+x = trnrand.standard_normal(256)
+```
+
+## Operations
+
+| Category | Function | Description |
+|----------|----------|-------------|
+| Distributions | `uniform` | U[low, high) |
+| | `normal` | N(μ, σ²) |
+| | `standard_normal` | N(0, 1) |
+| | `exponential` | Exp(λ) |
+| | `bernoulli` | Bernoulli(p) |
+| | `randint` | Uniform integers [low, high) |
+| | `randperm` | Random permutation |
+| | `truncated_normal` | Bounded normal (rejection sampling) |
+| Quasi-random | `sobol` | Sobol sequence (scrambled) |
+| | `halton` | Halton sequence |
+| | `latin_hypercube` | Latin Hypercube Sampling |
+
+## MC vs QMC Example
+
+```bash
+python examples/mc_integration.py
+```
+
+Compares pseudo-random vs Sobol quasi-random for estimating the volume of a 5-D hypersphere. QMC converges O(1/N) vs O(1/√N).
+
+## Status
+
+- [x] Seeded Generator with state management
+- [x] Standard distributions (uniform, normal, exponential, Bernoulli, etc.)
+- [x] Sobol, Halton, Latin Hypercube sequences
+- [x] MC vs QMC integration example
+- [ ] NKI Philox kernel on GpSimd
+- [ ] On-device Box-Muller (uniform → normal)
+- [ ] Benchmarks vs cuRAND
+
+## Related Projects
+
+| Project | What |
+|---------|------|
+| [trnfft](https://github.com/trnsci/trnfft) | FFT + complex ops |
+| [trnblas](https://github.com/trnsci/trnblas) | BLAS operations |
+| [trnsolver](https://github.com/trnsci/trnsolver) | Linear solvers |
+
+## License
+
+Apache 2.0 — Copyright 2026 Scott Friedman

trnrand-0.1.0.dist-info/RECORD

@@ -0,0 +1,11 @@
+trnrand/__init__.py,sha256=aDhG3auvvzoTbf54WE7LrmQlmVnR--btUGO2DfG7uTk,1046
+trnrand/distributions.py,sha256=5y94mDXGY_BiPpOXoWt-UIBLpZahddkg1PtYP0LE4eA,4072
+trnrand/generator.py,sha256=ghxtHSrmQoECA-1lntO5UDiyD-Y28IKLVU37jYWq4lY,2012
+trnrand/quasi.py,sha256=V80fkZBgEeMEUIo39tyKOwHeDPXKHNcwA3AZ9aD4YQY,3220
+trnrand/nki/__init__.py,sha256=xjAwT_HOYC1_NgCwmeAMQEMFVOSKNxcLDq-5cjUQ39U,167
+trnrand/nki/dispatch.py,sha256=6_YLNffct8cDrQNUCPKeYP1dmk4lLqQIlJ4x7y_Re-A,7941
+trnrand-0.1.0.dist-info/licenses/LICENSE,sha256=hTt_varGm4HTIVm8GQ1pGR7jah4vW5IrddFIu94FPAE,11175
+trnrand-0.1.0.dist-info/METADATA,sha256=hOfRucwRYjWM9jmY9VsQCt9RfJAvwBq9AnfggU0XrQ8,4089
+trnrand-0.1.0.dist-info/WHEEL,sha256=aeYiig01lYGDzBgS8HxWXOg3uV61G9ijOsup-k9o1sk,91
+trnrand-0.1.0.dist-info/top_level.txt,sha256=M-7c9pzN73CKWdnPxKWLuzzLSGZmQ2XuI6NKGdVM37E,8
+trnrand-0.1.0.dist-info/RECORD,,

trnrand-0.1.0.dist-info/WHEEL

@@ -0,0 +1,5 @@
+Wheel-Version: 1.0
+Generator: setuptools (82.0.1)
+Root-Is-Purelib: true
+Tag: py3-none-any
+

trnrand-0.1.0.dist-info/licenses/LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   Copyright 2026 Scott Friedman
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for describing the origin of the Work and
+      reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may accept support,
+      obligations, or other liability only on behalf of Yourself and on
+      Your sole responsibility, not on behalf of any other Contributor,
+      and only if You agree to indemnify, defend, and hold each Contributor
+      harmless for any liability incurred by, or claims asserted against,
+      such Contributor by reason of your accepting any such warranty or
+      additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Scott Friedman
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

trnrand-0.1.0.dist-info/top_level.txt

@@ -0,0 +1 @@
+trnrand