ruda-nn 0.21.52

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

use super::*;
use super::super::super::{VocabParallelGreedySelection,VocabParallelTopKSelection};
use crate::{pool::SequencePooling,transformer::SequenceHeadOutput};
use ruda_model::tensor::Bool;

impl<B:Backend> TensorParallelEncoderDecoderModel<B> {
    /// Complete native paired hidden output, before the original target head.
    pub fn forward_hidden_inference_with<C,T,E,F>(&self,source:TensorParallelTransformerInput<B>,target:TensorParallelTransformerInput<B>,
        source_group:C,source_layout:&VocabParallelLossLayout,encoder:E,target_group:T,target_layout:&VocabParallelLossLayout,decoder:F)
        -> Result<Tensor<B,3>,C::Error>
        where C:BroadcastTensorCollective<B>,T:BroadcastTensorCollective<B,Error=C::Error>,
            E:FnMut(usize,&TensorParallelAdaptedStackLayer<B>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,3>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error> {
        let memory = self.encode_inference_with(source,source_group,source_layout,encoder)?;
        self.decode_hidden_inference_with(target,memory,target_group,target_layout,decoder)
    }

    /// Complete packed paired hidden inference, retaining independent actual document lengths.
    pub fn forward_packed_hidden_inference_with<C,T,E,F>(&self,source:TensorParallelTransformerInput<B,1>,target:TensorParallelTransformerInput<B,1>,
        source_packed:&PackedSequenceLayout,target_packed:&PackedSequenceLayout,source_group:C,source_layout:&VocabParallelLossLayout,encoder:E,
        target_group:T,target_layout:&VocabParallelLossLayout,decoder:F) -> Result<Tensor<B,2>,C::Error>
        where C:BroadcastTensorCollective<B>,T:BroadcastTensorCollective<B,Error=C::Error>,
            E:FnMut(usize,&TensorParallelAdaptedStackLayer<B>,Tensor<B,2>)->Result<Tensor<B,2>,C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,2>,Tensor<B,2>)->Result<Tensor<B,2>,C::Error> {
        assert_eq!(source_packed.documents(),target_packed.documents(),"native packed paired hidden document counts differ");
        let memory = self.encode_packed_inference_with(source,source_packed,source_group,source_layout,encoder)?;
        self.decode_packed_hidden_inference_with(target,memory,source_packed,target_packed,target_group,target_layout,decoder)
    }

    /// Paired target-sequence classification using the original native real-token pooling policy.
    pub fn forward_sequence_inference_with<C,T,O,E,F>(&self,source:TensorParallelTransformerInput<B>,target:TensorParallelTransformerInput<B>,
        visible:Tensor<B,2,Bool>,pooling:SequencePooling,source_group:C,source_layout:&VocabParallelLossLayout,encoder:E,
        target_group:T,target_layout:&VocabParallelLossLayout,decoder:F,output_group:O,output_layout:&VocabParallelLossLayout,gather_output:bool)
        -> Result<SequenceHeadOutput<B>,C::Error>
        where C:BroadcastTensorCollective<B>,T:BroadcastTensorCollective<B,Error=C::Error>,O:BroadcastTensorCollective<B,Error=C::Error>,
            E:FnMut(usize,&TensorParallelAdaptedStackLayer<B>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,3>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error> {
        assert_eq!(visible.dims(),target.tokens.dims(),"native paired sequence visibility/decoder rows differ");
        let hidden = self.forward_hidden_inference_with(source,target,source_group,source_layout,encoder,target_group,target_layout,decoder)?;
        self.head.forward_sequence_inference(hidden,visible,pooling,output_group,output_layout,gather_output)
    }

    /// Native packed paired document classification, preserving empty-document visibility and counts.
    pub fn forward_packed_sequences_inference_with<C,T,O,E,F>(&self,source:TensorParallelTransformerInput<B,1>,target:TensorParallelTransformerInput<B,1>,
        source_packed:&PackedSequenceLayout,target_packed:&PackedSequenceLayout,visible:Option<Tensor<B,1,Bool>>,pooling:SequencePooling,
        source_group:C,source_layout:&VocabParallelLossLayout,encoder:E,target_group:T,target_layout:&VocabParallelLossLayout,decoder:F,
        output_group:O,output_layout:&VocabParallelLossLayout,gather_output:bool) -> Result<SequenceHeadOutput<B>,C::Error>
        where C:BroadcastTensorCollective<B>,T:BroadcastTensorCollective<B,Error=C::Error>,O:BroadcastTensorCollective<B,Error=C::Error>,
            E:FnMut(usize,&TensorParallelAdaptedStackLayer<B>,Tensor<B,2>)->Result<Tensor<B,2>,C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,2>,Tensor<B,2>)->Result<Tensor<B,2>,C::Error> {
        let hidden = self.forward_packed_hidden_inference_with(source,target,source_packed,target_packed,source_group,source_layout,encoder,target_group,target_layout,decoder)?;
        self.head.forward_packed_sequences_inference(hidden,target_packed,visible,pooling,output_group,output_layout,gather_output)
    }

    /// Complete paired prompt-to-global greedy IDs, projecting only the actual last target row.
    /// No source/target tokenization, EOS termination or sampling policy is selected.
    pub fn forward_greedy_last_inference_with<C,T,O,E,F>(&self,source:TensorParallelTransformerInput<B>,target:TensorParallelTransformerInput<B>,
        source_group:C,source_layout:&VocabParallelLossLayout,encoder:E,target_group:T,target_layout:&VocabParallelLossLayout,decoder:F,
        output_group:O,output_layout:&VocabParallelLossLayout,visible:Option<Tensor<B,1,Bool>>) -> Result<VocabParallelGreedySelection<B>,C::Error>
        where C:BroadcastTensorCollective<B>,T:BroadcastTensorCollective<B,Error=C::Error>,O:BroadcastTensorCollective<B,Error=C::Error>,
            E:FnMut(usize,&TensorParallelAdaptedStackLayer<B>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,3>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error> {
        assert!(target.tokens.dims()[1] > 0,"native paired prompt greedy requires an actual target token");
        let hidden = self.forward_hidden_inference_with(source,target,source_group,source_layout,encoder,target_group,target_layout,decoder)?;
        self.head.forward_greedy_last_inference(hidden,output_group,output_layout,visible)
    }

    /// Complete paired prompt-to-global top-k candidates, without projecting/gathering full prompt logits.
    pub fn forward_topk_last_inference_with<C,T,O,E,F>(&self,source:TensorParallelTransformerInput<B>,target:TensorParallelTransformerInput<B>,
        source_group:C,source_layout:&VocabParallelLossLayout,encoder:E,target_group:T,target_layout:&VocabParallelLossLayout,decoder:F,
        output_group:O,output_layout:&VocabParallelLossLayout,k:usize,visible:Option<Tensor<B,1,Bool>>) -> Result<VocabParallelTopKSelection<B>,C::Error>
        where C:BroadcastTensorCollective<B>,T:BroadcastTensorCollective<B,Error=C::Error>,O:BroadcastTensorCollective<B,Error=C::Error>,
            E:FnMut(usize,&TensorParallelAdaptedStackLayer<B>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,3>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error> {
        assert!(target.tokens.dims()[1] > 0,"native paired prompt top-k requires an actual target token");
        let hidden = self.forward_hidden_inference_with(source,target,source_group,source_layout,encoder,target_group,target_layout,decoder)?;
        self.head.forward_topk_last_inference(hidden,output_group,output_layout,k,visible)
    }

    /// Native complete source/target inference with original independent per-layer policies.
    pub fn forward_inference_with<C,T,O,E,F>(&self,source:TensorParallelTransformerInput<B>,target:TensorParallelTransformerInput<B>,
        source_group:C,source_layout:&VocabParallelLossLayout,encoder:E,target_group:T,target_layout:&VocabParallelLossLayout,decoder:F,
        output_group:O,output_layout:&VocabParallelLossLayout,gather_output:bool) -> Result<Tensor<B,3>,C::Error>
        where C:BroadcastTensorCollective<B>,T:BroadcastTensorCollective<B,Error=C::Error>,O:BroadcastTensorCollective<B,Error=C::Error>,
            E:FnMut(usize,&TensorParallelAdaptedStackLayer<B>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,3>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error> {
        let memory = self.encode_inference_with(source,source_group,source_layout,encoder)?;
        let hidden = self.decode_hidden_inference_with(target,memory,target_group,target_layout,decoder)?;
        self.head.forward_inference(hidden,output_group,output_layout,gather_output)
    }

    /// Native paired flat-document inference with real independent source/target boundaries.
    pub fn forward_packed_inference_with<C,T,O,E,F>(&self,source:TensorParallelTransformerInput<B,1>,target:TensorParallelTransformerInput<B,1>,
        source_packed:&PackedSequenceLayout,target_packed:&PackedSequenceLayout,source_group:C,source_layout:&VocabParallelLossLayout,encoder:E,
        target_group:T,target_layout:&VocabParallelLossLayout,decoder:F,output_group:O,output_layout:&VocabParallelLossLayout,gather_output:bool)
        -> Result<Tensor<B,2>,C::Error>
        where C:BroadcastTensorCollective<B>,T:BroadcastTensorCollective<B,Error=C::Error>,O:BroadcastTensorCollective<B,Error=C::Error>,
            E:FnMut(usize,&TensorParallelAdaptedStackLayer<B>,Tensor<B,2>)->Result<Tensor<B,2>,C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,2>,Tensor<B,2>)->Result<Tensor<B,2>,C::Error> {
        assert_eq!(source_packed.documents(),target_packed.documents(),"native packed paired model document counts differ");
        let memory = self.encode_packed_inference_with(source,source_packed,source_group,source_layout,encoder)?;
        let hidden = self.decode_packed_hidden_inference_with(target,memory,source_packed,target_packed,target_group,target_layout,decoder)?;
        self.head.forward_inference(hidden,output_group,output_layout,gather_output)
    }

    /// Encode the actual source once, then prepare each real decoder layer's positioned source K/V.
    /// The callback explicitly supplies persistent visibility and original cross-memory position policy.
    pub fn prepare_kv_cache_inference_with<C,E,F>(&self,source:TensorParallelTransformerInput<B>,source_group:C,
        source_layout:&VocabParallelLossLayout,encoder:E,initial_capacity:usize,prepare:F) -> Result<EncoderDecoderKvCache<B>,C::Error>
        where C:BroadcastTensorCollective<B>,E:FnMut(usize,&TensorParallelAdaptedStackLayer<B>,Tensor<B,3>)->Result<Tensor<B,3>,C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedDecoderCrossAttention<B>,Tensor<B,3>)->Result<ProjectedKvCache<B>,C::Error> {
        let memory = self.encode_inference_with(source,source_group,source_layout,encoder)?;
        self.decoder.prepare_kv_cache_inference_with(memory,initial_capacity,prepare)
    }

    /// Actual new target rows over original prepared source/decoder history, without rerunning the encoder.
    pub fn decode_cached_hidden_inference_with<C,F>(&self,target:TensorParallelTransformerInput<B>,cache:&mut EncoderDecoderKvCache<B>,
        target_group:C,target_layout:&VocabParallelLossLayout,decoder:F) -> Result<Tensor<B,3>,C::Error>
        where C:BroadcastTensorCollective<B>,F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,3>,&mut ProjectedKvCache<B>,&ProjectedKvCache<B>)
            ->Result<Tensor<B,3>,C::Error> {
        let hidden = target.embed_inference(&self.target_embeddings,target_group,target_layout)?;
        self.decoder.forward_cached_inference_with(hidden,cache,decoder).map(|hidden|self.target_finish(hidden))
    }

    /// Complete native cached target decoding and actual local output projection.
    /// Existing cache completion follows the decoder contract; head transport is a subsequent stage.
    pub fn forward_cached_inference_with<C,O,F>(&self,target:TensorParallelTransformerInput<B>,cache:&mut EncoderDecoderKvCache<B>,
        target_group:C,target_layout:&VocabParallelLossLayout,decoder:F,output_group:O,output_layout:&VocabParallelLossLayout,gather_output:bool)
        -> Result<Tensor<B,3>,C::Error>
        where C:BroadcastTensorCollective<B>,O:BroadcastTensorCollective<B,Error=C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,3>,&mut ProjectedKvCache<B>,&ProjectedKvCache<B>)->Result<Tensor<B,3>,C::Error> {
        let hidden = self.decode_cached_hidden_inference_with(target,cache,target_group,target_layout,decoder)?;
        self.head.forward_inference(hidden,output_group,output_layout,gather_output)
    }

    /// Native cached target-to-global-greedy output, projecting only the actual last new decoder row.
    pub fn forward_cached_greedy_inference_with<C,O,F>(&self,target:TensorParallelTransformerInput<B>,cache:&mut EncoderDecoderKvCache<B>,
        target_group:C,target_layout:&VocabParallelLossLayout,decoder:F,output_group:O,output_layout:&VocabParallelLossLayout,
        visible:Option<Tensor<B,1,ruda_model::tensor::Bool>>) -> Result<VocabParallelGreedySelection<B>,C::Error>
        where C:BroadcastTensorCollective<B>,O:BroadcastTensorCollective<B,Error=C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,3>,&mut ProjectedKvCache<B>,&ProjectedKvCache<B>)->Result<Tensor<B,3>,C::Error> {
        assert!(target.tokens.dims()[1] > 0,"cached paired greedy model requires an actual last target token");
        let hidden = self.decode_cached_hidden_inference_with(target,cache,target_group,target_layout,decoder)?;
        self.head.forward_greedy_last_inference(hidden,output_group,output_layout,visible)
    }

    /// Native paired cached decoder-to-global-top-k candidates over already prepared actual source memory.
    pub fn forward_cached_topk_inference_with<C,O,F>(&self,target:TensorParallelTransformerInput<B>,cache:&mut EncoderDecoderKvCache<B>,
        target_group:C,target_layout:&VocabParallelLossLayout,decoder:F,output_group:O,output_layout:&VocabParallelLossLayout,
        k:usize,visible:Option<Tensor<B,1,ruda_model::tensor::Bool>>) -> Result<VocabParallelTopKSelection<B>,C::Error>
        where C:BroadcastTensorCollective<B>,O:BroadcastTensorCollective<B,Error=C::Error>,
            F:FnMut(usize,&TensorParallelAdaptedEncoderDecoderLayer<B>,Tensor<B,3>,&mut ProjectedKvCache<B>,&ProjectedKvCache<B>)->Result<Tensor<B,3>,C::Error> {
        assert!(target.tokens.dims()[1] > 0,"cached paired top-k model requires an actual last target token");
        let hidden = self.decode_cached_hidden_inference_with(target,cache,target_group,target_layout,decoder)?;
        self.head.forward_topk_last_inference(hidden,output_group,output_layout,k,visible)
    }
}