ruda-nn 0.21.38

Ruda neural network layers, activation modules and losses.
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
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163

use alloc::{format,string::String};
use ruda_model::{
    module::{Module,ModuleDTypeRecord},
    record::{PrecisionSettings,Record,Recorder,RecorderError},
    tensor::{DType,backend::Backend},
};
use crate::{Linear,LoRALinear};

/// Caller-identified frozen base and the actual adapter continuation contract.
#[derive(Clone,Debug,PartialEq)]
pub struct LoRAAdapterSchema {
    /// Adapter-only format version, independent of the full module record.
    pub version: u32,
    /// Caller-provided identity of the exact frozen weights/configuration.
    pub base_id: String,
    /// Actual logical [input,output] dimensions of the base projection.
    pub base_shape: [usize;2],
    /// Original frozen weight storage, including its precision.
    pub base_dtype: DType,
    /// Original optional frozen bias storage; its width equals base_shape[1].
    pub base_bias_dtype: Option<DType>,
    /// Actual intermediate adapter rank.
    pub rank: usize,
    /// Actual forward multiplier, including the selected LoRA/rsLoRA convention.
    pub scale: f64,
    /// Actual adapter-input dropout probability.
    pub dropout: f64,
    /// Actual A/B trainable settings, retained independently from recorder precision.
    pub trainable: [bool;2],
}

impl<B: Backend> Record<B> for LoRAAdapterSchema {
    type Item<S: PrecisionSettings> = (u32,String,([usize;2],DType,Option<DType>),(usize,f64,f64),[bool;2]);
    fn into_item<S: PrecisionSettings>(self) -> Self::Item<S> {
        (self.version,self.base_id,(self.base_shape,self.base_dtype,self.base_bias_dtype),
            (self.rank,self.scale,self.dropout),self.trainable)
    }
    fn from_item<S: PrecisionSettings>(item: Self::Item<S>,_device: &B::Device) -> Self {
        Self {version:item.0,base_id:item.1,base_shape:item.2.0,base_dtype:item.2.1,base_bias_dtype:item.2.2,
            rank:item.3.0,scale:item.3.1,dropout:item.3.2,trainable:item.4}
    }
}

/// Native A/B-only record; no base tensors, optimizer state or input-source state.
/// Use a full-precision recorder when exact optimizer continuation is required.
pub struct LoRAAdapterRecord<B: Backend> {
    /// Actual immutable frozen-base/configuration contract.
    pub schema: LoRAAdapterSchema,
    adapter_a: <Linear<B> as Module<B>>::Record,
    adapter_b: <Linear<B> as Module<B>>::Record,
    dtypes: ModuleDTypeRecord,
}

impl<B: Backend> Record<B> for LoRAAdapterRecord<B> {
    type Item<S: PrecisionSettings> = (
        <LoRAAdapterSchema as Record<B>>::Item<S>,
        <<Linear<B> as Module<B>>::Record as Record<B>>::Item<S>,
        <<Linear<B> as Module<B>>::Record as Record<B>>::Item<S>,
        <ModuleDTypeRecord as Record<B>>::Item<S>,
    );
    fn into_item<S: PrecisionSettings>(self) -> Self::Item<S> {
        (<LoRAAdapterSchema as Record<B>>::into_item::<S>(self.schema),
            self.adapter_a.into_item::<S>(),self.adapter_b.into_item::<S>(),
            <ModuleDTypeRecord as Record<B>>::into_item::<S>(self.dtypes))
    }
    fn from_item<S: PrecisionSettings>(item: Self::Item<S>,device: &B::Device) -> Self {
        Self {schema:<LoRAAdapterSchema as Record<B>>::from_item::<S>(item.0,device),
            adapter_a:<<Linear<B> as Module<B>>::Record as Record<B>>::from_item::<S>(item.1,device),
            adapter_b:<<Linear<B> as Module<B>>::Record as Record<B>>::from_item::<S>(item.2,device),
            dtypes:<ModuleDTypeRecord as Record<B>>::from_item::<S>(item.3,device)}
    }
}

fn invalid(reason: &str) -> RecorderError { RecorderError::Unknown(format!("Invalid native adapter record: {reason}")) }

impl LoRAAdapterSchema {
    /// Read actual projection metadata, without downloading frozen base values.
    pub fn capture<B: Backend>(layer: &LoRALinear<B>,base_id: &str) -> Result<Self,RecorderError> {
        if base_id.is_empty() { return Err(invalid("base identity must be supplied explicitly")); }
        let base = layer.base.weight.val();
        let [input,output] = base.dims();
        if input == 0 || output == 0 || base.is_require_grad() || !base.dtype().is_float() {
            return Err(invalid("the declared dense base must be nonempty, floating and frozen"));
        }
        let base_bias_dtype = if let Some(bias) = &layer.base.bias {
            let bias = bias.val();
            if bias.dims() != [output] || bias.device() != base.device() || bias.is_require_grad() {
                return Err(invalid("base bias geometry/device or frozen setting differs"));
            }
            Some(bias.dtype())
        } else { None };
        if layer.adapter_a.bias.is_some() || layer.adapter_b.bias.is_some() {
            return Err(invalid("native adapter A/B projections must be bias-free"));
        }
        let a = layer.adapter_a.weight.val();
        let b = layer.adapter_b.weight.val();
        let [a_input,rank] = a.dims();
        if rank == 0 || a_input != input || b.dims() != [rank,output]
            || a.device() != base.device() || b.device() != base.device()
            || !a.dtype().is_float() || !b.dtype().is_float() {
            return Err(invalid("actual adapter A/B geometry, storage or device differs"));
        }
        if !layer.scale.is_finite() || !layer.dropout.prob.is_finite() || !(0.0..1.0).contains(&layer.dropout.prob) {
            return Err(invalid("invalid adapter multiplier/dropout"));
        }
        Ok(Self {version:1,base_id:base_id.into(),base_shape:[input,output],base_dtype:base.dtype(),base_bias_dtype,
            rank,scale:layer.scale,dropout:layer.dropout.prob,trainable:[a.is_require_grad(),b.is_require_grad()]})
    }

    /// Match an already-prepared layer before replacing any adapter parameters.
    /// Base identity is caller-owned; this method does not invent a content hash.
    pub fn validate_for<B: Backend>(&self,layer: &LoRALinear<B>,base_id: &str) -> Result<(),RecorderError> {
        if self.version != 1 { return Err(invalid("unsupported format version")); }
        let actual = Self::capture(layer,base_id)?;
        if self.base_id != actual.base_id || self.base_shape != actual.base_shape || self.base_dtype != actual.base_dtype
            || self.base_bias_dtype != actual.base_bias_dtype || self.rank != actual.rank || self.trainable != actual.trainable
            || self.scale.to_bits() != actual.scale.to_bits() || self.dropout.to_bits() != actual.dropout.to_bits() {
            return Err(invalid("frozen base, rank, precision, training settings or forward configuration differs"));
        }
        Ok(())
    }
}

impl<B: Backend> LoRAAdapterRecord<B> {
    /// Pin only actual A/B values and per-parameter dtypes at a caller-owned boundary.
    /// Frozen base handles/values are never retained in this record.
    pub fn capture(layer: &LoRALinear<B>,base_id: &str) -> Result<Self,RecorderError> {
        let schema = LoRAAdapterSchema::capture(layer,base_id)?;
        let adapters = (layer.adapter_a.clone(),layer.adapter_b.clone());
        let dtypes = ModuleDTypeRecord::capture(&adapters)?;
        Ok(Self {schema,adapter_a:adapters.0.into_record(),adapter_b:adapters.1.into_record(),dtypes})
    }

    /// Save A/B and continuation metadata through an existing recorder.
    pub fn save<R: Recorder<B>>(self,recorder: &R,args: R::RecordArgs) -> Result<R::RecordOutput,RecorderError> {
        recorder.record(self,args)
    }

    /// Read A/B onto the caller's device, without loading or creating a base model.
    pub fn load<R: Recorder<B>>(recorder: &R,args: R::LoadArgs,device: &B::Device) -> Result<Self,RecorderError> {
        recorder.load(args,device)
    }

    /// Restore original adapter IDs/dtypes on a prepared layer, leaving the base unchanged.
    /// Restore optimizer/pending-gradient records only after obtaining this new layer.
    pub fn restore_into(self,layer: LoRALinear<B>,base_id: &str) -> Result<LoRALinear<B>,RecorderError> {
        self.schema.validate_for(&layer,base_id)?;
        let device = layer.base.weight.val().device();
        let a = layer.adapter_a.load_record(self.adapter_a).fork(&device);
        let b = layer.adapter_b.load_record(self.adapter_b).fork(&device);
        let (a,b) = self.dtypes.apply((a,b))?;
        let restored = LoRALinear {base:layer.base,adapter_a:a,adapter_b:b,dropout:layer.dropout,scale:layer.scale};
        self.schema.validate_for(&restored,base_id)?;
        Ok(restored)
    }
}

impl<B: Backend> LoRALinear<B> {
    /// Export only this layer's actual A/B state, with explicit frozen-base identity.
    pub fn adapter_record(&self,base_id: &str) -> Result<LoRAAdapterRecord<B>,RecorderError> {
        LoRAAdapterRecord::capture(self,base_id)
    }
}