rlx_runtime/
lib.rs

1// RLX — versatile ML compiler + runtime.
2// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
3//
4// This program is free software: you can redistribute it and/or modify
5// it under the terms of the GNU General Public License as published by
6// the Free Software Foundation, version 3.
7//
8// This program is distributed in the hope that it will be useful,
9// but WITHOUT ANY WARRANTY; without even the implied warranty of
10// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11// GNU General Public License for more details.
12//
13// You should have received a copy of the GNU General Public License
14// along with this program. If not, see <https://www.gnu.org/licenses/>.
15
16//! RLX Runtime — the user-facing API.
17//!
18//! Provides a unified [`Session`] that compiles and executes IR graphs
19//! on the selected backend. Backend selection is via Cargo features:
20//!
21//! ```toml
22//! [dependencies]
23//! rlx-runtime = { version = "0.1", features = ["cpu"] }                # CPU (default)
24//! rlx-runtime = { version = "0.1", features = ["blas-accelerate"] }    # CPU + Apple Accelerate
25//! rlx-runtime = { version = "0.1", features = ["blas-mkl"] }           # CPU + Intel MKL
26//! # rlx-runtime = { version = "0.1", features = ["gpu"] }             # GPU via wgpu
27//! # rlx-runtime = { version = "0.1", features = ["cuda"] }            # GPU via CUDA
28//! ```
29//!
30//! # Example
31//! ```rust,no_run
32//! use rlx_runtime::*;
33//! use rlx_ir::*;
34//!
35//! // Build a graph
36//! let mut g = Graph::new("example");
37//! let x = g.input("x", Shape::new(&[2, 4], DType::F32));
38//! let w = g.param("w", Shape::new(&[4, 3], DType::F32));
39//! let b = g.param("b", Shape::new(&[3], DType::F32));
40//! let mm = g.matmul(x, w, Shape::new(&[2, 3], DType::F32));
41//! let out = g.binary(op::BinaryOp::Add, mm, b, Shape::new(&[2, 3], DType::F32));
42//! g.set_outputs(vec![out]);
43//!
44//! // Compile and execute
45//! let session = Session::new(Device::Cpu);
46//! let mut compiled = session.compile(g);
47//! compiled.set_param("w", &[1.0f32; 12]);
48//! compiled.set_param("b", &[0.0f32; 3]);
49//! let result = compiled.run(&[("x", &[1.0f32; 8])]);
50//! ```
51
52// Driver-layer concerns (device, arena, handle, stream, buffer)
53// live in rlx-driver as of plan #58; re-exported below so
54// existing callers compile unchanged.
55pub mod aot_cache;
56pub mod backend;
57pub mod compile_cache;
58pub mod compiled;
59pub mod cost;
60pub mod device_ext;
61pub mod expert_pool;
62pub mod jacfwd;
63pub mod kernel_trace;
64pub mod lora_scheduler;
65pub mod memory_estimate;
66pub mod model_pipeline;
67pub mod moe_expert_store;
68pub mod op_registry;
69pub mod options;
70pub mod paged_kv;
71pub mod precision;
72pub mod record_replay;
73pub mod reflect;
74pub mod registry;
75pub mod router;
76pub mod session;
77pub mod stages;
78pub mod subgraph;
79pub mod trace;
80pub mod weight_registry;
81pub mod weights;
82pub mod worker_pool;
83/// PLAN L3 — Perfetto / chrome-trace JSON tracing. Lives in `rlx-ir`
84/// (alongside the `Tick` cycle counter it depends on) so every backend
85/// can instrument per-thunk without crate-dep gymnastics. Re-exported
86/// here so callers see one consistent `rlx_runtime::perfetto::TraceSpan`.
87pub use rlx_ir::perfetto;
88pub mod custom_ops;
89pub mod hwinfo;
90pub mod logit_verify;
91pub mod nan_check;
92pub mod phase;
93pub mod spec_decode;
94pub mod telemetry;
95pub mod validators;
96
97// Always-available now that serde is a non-optional dep + #32
98// router consumes the OpenAI-shaped structs unconditionally.
99pub mod mock_requests;
100
101// Driver-layer types — re-exported from rlx-driver (plan #58).
102pub use rlx_driver::{Buffer, BufferHandle, CommandStream, Device, DeviceArena, SyncStream};
103// Symmetric-memory primitives (plan #49) — foundation for #12.
104pub use rlx_driver::{
105    CollectiveError, LocalTransport, Rank, SymmetricBuffer, SymmetricHeap, SymmetricTransport,
106};
107// Collective ops (plan #12).
108pub use aot_cache::{AotCache, AotCacheError};
109pub use backend::{Backend, ExecutableGraph, compile_hir, compile_module};
110pub use compile_cache::{
111    BucketedCompileCache, CompileCache, DynamicDimCompileCache, pad_rows, slice_rows,
112};
113pub use compiled::CompiledGraph;
114#[cfg(feature = "apple")]
115pub use device_ext::available_apple_devices;
116pub use device_ext::{
117    available_devices, dispatch_report_for_device, dispatch_report_for_device_with_options,
118    first_unsupported_op, first_unsupported_op_with_options, full_name, is_available,
119    legalize_graph_for_device, legalize_graph_for_device_with_options,
120    legalize_graph_for_device_with_report, supports, supports_graph, supports_graph_with_options,
121};
122pub use expert_pool::{
123    ExpertPool, ExpertPoolConfig, ExpertPoolStats, ExpertRefreshPolicy, ExpertRefreshResult,
124    MoEExecMode, gpu_expert_budget_from_vram,
125};
126pub use memory_estimate::{MoeOffloadEstimate, estimate_moe_offload};
127pub use model_pipeline::ModelCompilePipeline;
128pub use options::CompileOptions;
129pub use precision::Precision;
130pub use reflect::{ModelReflection, load_hir_template_with_extensions, specialize_entry};
131pub use registry::{BackendFactory, backend_for, register_backend, registered_devices};
132#[cfg(feature = "cpu")]
133pub use rlx_cpu::moe_residency::MoeResidencyStats;
134#[cfg(feature = "cpu")]
135pub use rlx_cpu::moe_topk_capture::MoeTopkCapture;
136pub use rlx_driver::{ReduceKind, all_gather, all_reduce, reduce_scatter};
137pub use rlx_ir::env::{self, RlxEnv, RuntimeOverrides};
138pub use session::Session;
139pub use stages::{
140    compile_graph_stages, compile_graph_stages_for_backend, compile_hir_stages,
141    compile_module_stages, fusion_target_for, graph_from_lir, maybe_log_fusion,
142    options_with_supported_ops, pipeline_for,
143};
144pub use subgraph::{SubgraphCache, run_if, run_while};
145
146pub use expert_pool::{merged_resident_mask, per_layer_resident_masks};
147pub use moe_expert_store::{ExpertStackF32, LayerMoeWeights, MoeExpertStore};
148pub use weight_registry::{WeightEntry, WeightHandle, WeightKind, WeightRegistry};
149pub use weights::{BytesWeightLoader, WeightLoader};
150
151// Cycle-accurate timing primitive lives in rlx-ir (lowest crate); re-export
152// here so `rlx_runtime::Tick` works without forcing callers to add the IR
153// crate as a direct dep.
154pub use rlx_ir::{AsyncCopy, BarrierToken, DoubleBuffer, SyncCopy};
155pub use rlx_ir::{CacheBuster, Tick, time_ns};
156
157// Re-export precision policy from rlx-opt for convenience
158pub use rlx_ir::{
159    inspect_graph, inspect_hir, inspect_hir_stats, inspect_lir, inspect_mir, inspect_mir_stats,
160};
161pub use rlx_opt::{OpKind, PrecisionPolicy};
162pub use rlx_opt::{PipelineInspect, inspect_pipeline};
163
164// Re-export IR types for convenience
165pub use rlx_ir::logical_kernel::{KernelDispatchConfig, KernelDispatchPolicy};
166pub use rlx_ir::op;
167pub use rlx_ir::{
168    BindingManifest, CompilationMode, DType, Graph, HirExtensionFn, HirReflection, IoBindingEntry,
169    ManifestDiff, ModelComponent, ModelPhase, ModelVariant, Node, NodeId, Op, Shape, WeightBlock,
170    apply_hir_extensions, register_hir_extension, registered_hir_extensions,
171};
172
173// Re-export proc macro
174pub use rlx_macros::pipeline_schedule;
175pub use rlx_macros::rlx_model;
rlx_runtime/lib.rs

rlx_runtime/
lib.rs
