rlx-flow 0.2.4

Block assembly-line API for RLX model builders — fusion-first, config-driven
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
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98

// RLX — versatile ML compiler + runtime.
// Copyright (C) 2026 Eugene Hauptmann, Nataliya Kosmyna.
//
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU General Public License as published by
// the Free Software Foundation, version 3.
//
// This program is distributed in the hope that it will be useful,
// but WITHOUT ANY WARRANTY; without even the implied warranty of
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
// GNU General Public License for more details.
//
// You should have received a copy of the GNU General Public License
// along with this program. If not, see <https://www.gnu.org/licenses/>.

//! Block assembly-line API for RLX model builders.

pub mod blocks;
mod composite;
mod context;
mod dsl;
pub mod escape;
mod execution;
mod extension;
mod flow;
mod layer;
mod plugin;
mod profile;
mod recipe;
pub mod rope;
mod side;
mod stage;
mod stage_contract;
mod stage_interfaces;
pub mod stream;
mod value;
mod weight;

pub mod prelude;

pub use blocks::RopeTablesStage;
pub use blocks::{
    BertEncoderLayerSpec, BertEncoderLayerStage, BertQkvStyle, ClsTokenPoolStage,
    NomicEncoderLayerSpec, NomicEncoderLayerStage, Qwen3DecodeLayerSpec, Qwen3DecoderSpec,
    Qwen3DecoderStage, VitSelfAttnSpec, dinov2_layer_fused, nomic_vision_layer_fused,
};
pub use composite::LayerComposition;
pub use context::{DecodeBindings, FlowState, GdnInputSlots};
pub use escape::Emit;
pub use execution::{ExecutionPreset, ModelExecutionConfig};
pub use extension::FlowExtensionPlan;
pub use flow::{BuiltModel, ModelFlow};
pub use layer::LayerStack;
pub use plugin::{PluginStage, plugin, plugin_named};
pub use profile::{
    BackendOverrides, CompileProfile, CpuBackendProfile, FusionPolicyKind, FusionProfile,
    FusionTargetKind, MetalBackendProfile, MixedPrecisionKind, PassProfile, PrecisionKind,
    PrecisionProfile, ProfileMode,
};
pub use recipe::ModelRecipe;
pub use rope::{
    Llama3Scaling, YarnScaling, build_default_tables, build_mrope_text_tables, build_tables,
    default_inv_freq, inv_freq_with_factors, llama3_scaled_inv_freq, mrope_row_for_sections,
    mrope_section_for_pair, mrope_sections4, ntk_scaled_inv_freq, yarn_scaled_inv_freq,
};
pub use side::SideOutputs;
pub use stage::FlowStage;
pub use stage_contract::{BlockAsLayer, LayerStage, StageArtifacts};
pub use stage_interfaces::{AttentionStage, FfnStage, KvCacheContract, NormStage};
pub use stream::{
    DualStreamStage, LoadStreamStage, StoreStreamStage, dual_stream_stage, id as stream_id,
};
pub use value::FlowValue;
pub use weight::{MapWeights, WeightSource};

use std::collections::HashMap;

/// Compatibility shim: packed GGUF matmul weights (used by some model loaders).
#[derive(Debug, Clone, Default)]
pub struct GgufPackedParams {
    pub linears: HashMap<String, GgufPackedLinear>,
}

impl GgufPackedParams {
    pub fn get_linear(&self, key: &str) -> Option<&GgufPackedLinear> {
        self.linears.get(key)
    }
}

/// One packed linear weight: quantized bytes + bias.
#[derive(Debug, Clone)]
pub struct GgufPackedLinear {
    pub w_q: Vec<u8>,
    pub scheme: rlx_ir::quant::QuantScheme,
    pub in_dim: usize,
    pub out_dim: usize,
    pub bias: Vec<f32>,
}