baracuda-kernels 0.0.1-alpha.68

Unified ML op facade for the baracuda CUDA ecosystem. Exposes every primitive an ML framework would expect (union of PyTorch torch.* + nn.functional and JAX lax.* / numpy ops) through a single Plan-based Rust surface, internally dispatching to baracuda-cutlass, the baracuda-* NVIDIA-library wrappers, or bespoke baracuda-kernels-sys kernels.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51

//! Random / sampling op family — Phase 4.5 (Category Q).
//!
//! Plan-per-op-family because the random ops have heterogeneous arg
//! shapes:
//!
//! - [`RandomPlan`] — pure-generator ops with no input tensor. Today
//!   wires `Uniform`, `Normal` (f32 / f64 via cuRAND) and `Bernoulli`
//!   (Bool output via cuRAND uniform + custom threshold kernel).
//!
//! - [`DropoutPlan`] / [`DropoutBackwardPlan`] — dropout takes an input
//!   tensor and returns both the output and the saved mask; backward
//!   replays the mask against `dy`. f32 / f64 only.
//!
//! Multinomial / Randint / exponential / gamma / quasi-random / stateful
//! RNG replay are reserved for future milestones (see the
//! `~/.claude/plans/warm-prancing-comet.md` Phase 4 deferral list).
//!
//! ## Generator lifetime
//!
//! cuRAND generators are stateful objects that bind to a CUDA stream.
//! Each `*Plan` creates one lazily on first `run` (the generator API
//! requires a live CUDA context, which `select()` cannot rely on having
//! at construction time) and rebinds it via `curandSetStream` on every
//! call so the plan is reusable across streams. The handle is destroyed
//! on `Drop`. cuRAND generators are *not* thread-safe — the plan is
//! `!Sync` by virtue of the `Cell<curandGenerator_t>` it holds.

pub mod dropout;
pub mod plan;
// Phase 46 — FlashInfer sort-free top-K/top-P/min-P sampling.
pub mod topk_topp_sampling;
// Phase 66 Tier 2 — per-row sampling + speculative-decode verification.
pub mod perrow_spec_sampling;
// Phase 66 Tier 2 — bespoke token-penalty logit transform.
pub mod token_penalty;

pub use dropout::{
    DropoutArgs, DropoutBackwardArgs, DropoutBackwardDescriptor, DropoutBackwardPlan,
    DropoutDescriptor, DropoutPlan,
};
pub use plan::{RandomArgs, RandomBoolArgs, RandomDescriptor, RandomPlan};
// Phase 46 — sort-free sampling re-exports.
pub use topk_topp_sampling::{
    SamplerKind, TopKTopPSamplingArgs, TopKTopPSamplingDescriptor, TopKTopPSamplingPlan,
};
// Phase 66 Tier 2 re-exports.
pub use perrow_spec_sampling::{
    PerRowSampler, PerRowSamplingArgs, PerRowSamplingDescriptor, PerRowSamplingPlan,
    SpeculativeSamplingArgs, SpeculativeSamplingDescriptor, SpeculativeSamplingPlan,
};
pub use token_penalty::{TokenPenaltyArgs, TokenPenaltyDescriptor, TokenPenaltyPlan};