ruda-nn 0.21.56

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
164
165
166
167
168
169
170
171
172
173
174
175
176
177

use super::*;
use ruda_model::{module::ModuleDisplay, tensor::{Bool, IntegerTensorCollective, VariableTensorCollective, MoeDispatchOps, MoeReceivedOps}};
use crate::expert_parallel::ExpertParallelReceived;
use crate::transformer::{ExpertParallelMhcFeedForward, ExpertParallelMhcError, MixedMhcFeedForward, MixedMhcBranchError,
    MhcResidualBranchShape, MhcResidualBranch, TransformerProjection};
use crate::attention::{CompressedAttentionProjection, CompressedAttentionOutput, PackedCompressedAttentionOutput, PackedSequenceLayout};

/// Original expert-owned mHC branch and explicitly present shared FFN, with all
/// persistent native weights kept as local data-axis slices.
#[derive(Module, Debug)]
pub struct FullyShardedExpertParallelMhcFeedForward<B: Backend, P: Module<B>, E: Module<B>> {
    pub routed: FullyShardedExpertParallelMoeLayer<B, P, E>,
    pub shared: Option<FullyShardedProjectedFeedForward<B, P>>,
}
#[derive(Module, Debug)]
pub enum FullyShardedMixedMhcFeedForward<B: Backend, L: Module<B>, P: Module<B>, E: Module<B>> {
    Local(L),
    Parallel(FullyShardedExpertParallelMhcFeedForward<B, P, E>),
}

impl<B: Backend> ShardingContext<B> {
    pub fn expert_parallel_mhc<P: ShardTransformerProjection<B>, E: ShardOwnedExperts<B>>(&mut self,
        source: ExpertParallelMhcFeedForward<B, P, E>) -> FullyShardedExpertParallelMhcFeedForward<B, P::Sharded, E::Sharded> {
        source.validate_branch();
        FullyShardedExpertParallelMhcFeedForward { routed: self.expert_parallel_layer(source.routed),
            shared: source.shared.map(|value| self.awq_feed_forward(value)) }
    }
}
impl<B: Backend, P: ShardTransformerProjection<B>, E: ShardOwnedExperts<B>> ShardMhcResidualBranch<B> for ExpertParallelMhcFeedForward<B, P, E> {
    type Sharded = FullyShardedExpertParallelMhcFeedForward<B, P::Sharded, E::Sharded>;
    fn shard_branch(self, context: &mut ShardingContext<B>) -> Self::Sharded { context.expert_parallel_mhc(self) }
}
impl<B: Backend, L: ShardMhcResidualBranch<B>, P: ShardTransformerProjection<B>, E: ShardOwnedExperts<B>> ShardMhcResidualBranch<B>
    for MixedMhcFeedForward<B, L, P, E> {
    type Sharded = FullyShardedMixedMhcFeedForward<B, L::Sharded, P::Sharded, E::Sharded>;
    fn shard_branch(self, context: &mut ShardingContext<B>) -> Self::Sharded {
        match self { Self::Local(value) => FullyShardedMixedMhcFeedForward::Local(value.shard_branch(context)),
            Self::Parallel(value) => FullyShardedMixedMhcFeedForward::Parallel(value.shard_branch(context)) }
    }
}

macro_rules! gather_mhc_expert_parallel {
    ($backend:ty, [$($generics:tt)*], $gather:ident) => {
        impl<$($generics)*, P: GatherTransformerProjection<$backend, B>, E: GatherOwnedExperts<$backend, B>> GatherMhcResidualBranch<$backend, B>
            for FullyShardedExpertParallelMhcFeedForward<$backend, P, E> {
            type Gathered = ExpertParallelMhcFeedForward<$backend, P::Gathered, E::Gathered>;
            fn gather_branch<C: IntegerTensorCollective<B>>(&self, data: C) -> Result<Self::Gathered, C::Error> {
                Ok(ExpertParallelMhcFeedForward::from_parts(self.routed.$gather(data.clone())?,
                    self.shared.as_ref().map(|value| value.$gather(data)).transpose()?))
            }
        }
        impl<$($generics)*, L: GatherMhcResidualBranch<$backend, B>, P: GatherTransformerProjection<$backend, B>, E: GatherOwnedExperts<$backend, B>>
            GatherMhcResidualBranch<$backend, B> for FullyShardedMixedMhcFeedForward<$backend, L, P, E> {
            type Gathered = MixedMhcFeedForward<$backend, L::Gathered, P::Gathered, E::Gathered>;
            fn gather_branch<C: IntegerTensorCollective<B>>(&self, data: C) -> Result<Self::Gathered, C::Error> {
                match self { Self::Local(value) => value.gather_branch(data).map(MixedMhcFeedForward::Local),
                    Self::Parallel(value) => value.gather_branch(data).map(MixedMhcFeedForward::Parallel) }
            }
        }
    };
}
gather_mhc_expert_parallel!(B, [B: Backend], gather_inference);
gather_mhc_expert_parallel!(Autodiff<B, S>, [B: Backend, S: CheckpointStrategy], gather);

#[derive(Debug)]
pub enum FullyShardedExpertParallelMhcError<D: core::fmt::Debug, C: core::fmt::Debug, P: core::fmt::Debug, E: core::fmt::Debug> {
    Data(D),
    Branch(ExpertParallelMhcError<C, P, E>),
}
impl<D: core::fmt::Debug, C: core::fmt::Debug, P: core::fmt::Debug, E: core::fmt::Debug> core::fmt::Display for FullyShardedExpertParallelMhcError<D, C, P, E> {
    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
        match self { Self::Data(error) => write!(f, "mHC owned branch data gather: {error:?}"), Self::Branch(error) => write!(f, "{error}") }
    }
}
impl<D: core::fmt::Debug, C: core::fmt::Debug, P: core::fmt::Debug, E: core::fmt::Debug> core::error::Error for FullyShardedExpertParallelMhcError<D, C, P, E> {}

macro_rules! execute_sharded_mhc_experts {
    ($backend:ty, [$($generics:tt)*], $native:ident, $forward:ident) => {
        impl<$($generics)*, P: GatherTransformerProjection<$backend, B>, E: GatherOwnedExperts<$backend, B>>
            FullyShardedExpertParallelMhcFeedForward<$backend, P, E>
        where P::Gathered: TransformerProjection<$backend>, E::Gathered: ExpertParallelReceived<$backend> {
            /// Separate caller-owned groups/scopes retain their own transport
            /// derivatives. Objective weighting/completion remains caller-selected.
            pub fn $forward<D: IntegerTensorCollective<B>, C: VariableTensorCollective<B>, const N: usize>(&self,
                input: Tensor<$backend, N>, data: D, expert: C)
                -> Result<Tensor<$backend, N>, FullyShardedExpertParallelMhcError<D::Error, C::Error,
                    <P::Gathered as TransformerProjection<$backend>>::Error, <E::Gathered as ExpertParallelReceived<$backend>>::Error>> {
                self.gather_branch(data).map_err(FullyShardedExpertParallelMhcError::Data)?
                    .$native(input, expert).map_err(FullyShardedExpertParallelMhcError::Branch)
            }
        }
    };
}
execute_sharded_mhc_experts!(B, [B: MoeDispatchOps + MoeReceivedOps], forward_inference, forward_inference);
execute_sharded_mhc_experts!(Autodiff<B, S>, [B: MoeDispatchOps + MoeReceivedOps, S: CheckpointStrategy], forward, forward);

/// Native branch errors retain local, router, owned-expert and expert-transport
/// categories, separately from data gathers and the actual task-head error.
pub type FullyShardedMixedMhcModelError<D, C, L, P, E, H> = FullyShardedMhcModelError<D, MixedMhcBranchError<L, C, P, E>, H>;

macro_rules! execute_mixed_mhc_model {
    ($backend:ty, [$($generics:tt)*], $native:ident, $with:ident, $hidden_with:ident, $aux_with:ident, $packed_with:ident, $packed_aux_with:ident,
        $forward:ident, $hidden:ident, $aux:ident, $packed:ident, $packed_aux:ident) => {
        impl<$($generics)*, A: GatherTransformerProjection<$backend, B>, L: GatherMhcResidualBranch<$backend, B>,
            P: GatherTransformerProjection<$backend, B>, E: GatherOwnedExperts<$backend, B>, H: GatherTransformerProjection<$backend, B>>
            FullyShardedMhcResidualModel<$backend, A, FullyShardedMixedMhcFeedForward<$backend, L, P, E>, H>
        where A::Gathered: CompressedAttentionProjection<$backend>, L::Gathered: MhcResidualBranch<$backend>,
            P::Gathered: TransformerProjection<$backend>, E::Gathered: ExpertParallelReceived<$backend>, H::Gathered: TransformerProjection<$backend> {
            /// The expert-group factory is evaluated only on actual parallel
            /// branches. It may return separately scope-bound per-layer groups;
            /// data gather order and original native branch graphs are unchanged.
            pub fn $forward<D, C, G>(&self, tokens: Tensor<$backend, 2, Int>, valid: Option<Tensor<$backend, 2, Bool>>, data: D, mut experts: G)
                -> Result<Tensor<$backend, 3>, FullyShardedMixedMhcModelError<D::Error, C::Error,
                    <L::Gathered as MhcResidualBranch<$backend>>::Error, <P::Gathered as TransformerProjection<$backend>>::Error,
                    <E::Gathered as ExpertParallelReceived<$backend>>::Error, <H::Gathered as TransformerProjection<$backend>>::Error>>
            where D: IntegerTensorCollective<B>, C: VariableTensorCollective<B>, G: FnMut(usize) -> C {
                self.$with(tokens, valid, data, |index, feed, input, _| feed.$native(input, || experts(index)))
            }
            pub fn $hidden<D, C, G>(&self, tokens: Tensor<$backend, 2, Int>, valid: Option<Tensor<$backend, 2, Bool>>, data: D, mut experts: G)
                -> Result<Tensor<$backend, 3>, FullyShardedMixedMhcModelError<D::Error, C::Error,
                    <L::Gathered as MhcResidualBranch<$backend>>::Error, <P::Gathered as TransformerProjection<$backend>>::Error,
                    <E::Gathered as ExpertParallelReceived<$backend>>::Error, <H::Gathered as TransformerProjection<$backend>>::Error>>
            where D: IntegerTensorCollective<B>, C: VariableTensorCollective<B>, G: FnMut(usize) -> C {
                self.$hidden_with(tokens, valid, data, |index, feed, input, _| feed.$native(input, || experts(index)))
            }
            pub fn $aux<D, C, G>(&self, tokens: Tensor<$backend, 2, Int>, valid: Option<Tensor<$backend, 2, Bool>>,
                indexer_warmup: bool, data: D, mut experts: G)
                -> Result<CompressedAttentionOutput<$backend>, FullyShardedMixedMhcModelError<D::Error, C::Error,
                    <L::Gathered as MhcResidualBranch<$backend>>::Error, <P::Gathered as TransformerProjection<$backend>>::Error,
                    <E::Gathered as ExpertParallelReceived<$backend>>::Error, <H::Gathered as TransformerProjection<$backend>>::Error>>
            where D: IntegerTensorCollective<B>, C: VariableTensorCollective<B>, G: FnMut(usize) -> C {
                self.$aux_with(tokens, valid, indexer_warmup, data, |index, feed, input, _| feed.$native(input, || experts(index)))
            }
            pub fn $packed<D, C, G>(&self, tokens: Tensor<$backend, 1, Int>, layout: &PackedSequenceLayout,
                valid: Option<Tensor<$backend, 1, Bool>>, data: D, mut experts: G)
                -> Result<Tensor<$backend, 2>, FullyShardedMixedMhcModelError<D::Error, C::Error,
                    <L::Gathered as MhcResidualBranch<$backend>>::Error, <P::Gathered as TransformerProjection<$backend>>::Error,
                    <E::Gathered as ExpertParallelReceived<$backend>>::Error, <H::Gathered as TransformerProjection<$backend>>::Error>>
            where D: IntegerTensorCollective<B>, C: VariableTensorCollective<B>, G: FnMut(usize) -> C {
                self.$packed_with(tokens, layout, valid, data, |index, feed, input, _| feed.$native(input, || experts(index)))
            }
            pub fn $packed_aux<D, C, G>(&self, tokens: Tensor<$backend, 1, Int>, layout: &PackedSequenceLayout,
                valid: Option<Tensor<$backend, 1, Bool>>, indexer_warmup: bool, data: D, mut experts: G)
                -> Result<PackedCompressedAttentionOutput<$backend>, FullyShardedMixedMhcModelError<D::Error, C::Error,
                    <L::Gathered as MhcResidualBranch<$backend>>::Error, <P::Gathered as TransformerProjection<$backend>>::Error,
                    <E::Gathered as ExpertParallelReceived<$backend>>::Error, <H::Gathered as TransformerProjection<$backend>>::Error>>
            where D: IntegerTensorCollective<B>, C: VariableTensorCollective<B>, G: FnMut(usize) -> C {
                self.$packed_aux_with(tokens, layout, valid, indexer_warmup, data, |index, feed, input, _| feed.$native(input, || experts(index)))
            }
        }
    };
}
execute_mixed_mhc_model!(B, [B: MoeDispatchOps + MoeReceivedOps], forward_inference,
    try_forward_with_inference, try_forward_hidden_with_inference, try_forward_with_aux_inference, try_forward_packed_with_inference, try_forward_packed_with_aux_inference,
    forward_with_experts_inference, forward_hidden_with_experts_inference, forward_with_experts_aux_inference, forward_packed_with_experts_inference, forward_packed_with_experts_aux_inference);
execute_mixed_mhc_model!(Autodiff<B, S>, [B: MoeDispatchOps + MoeReceivedOps, S: CheckpointStrategy], forward,
    try_forward_with, try_forward_hidden_with, try_forward_with_aux, try_forward_packed_with, try_forward_packed_with_aux,
    forward_with_experts, forward_hidden_with_experts, forward_with_experts_aux, forward_packed_with_experts, forward_packed_with_experts_aux);

impl<B: Backend, P: FullyShardedModule<B> + ModuleDisplay, E: FullyShardedModule<B> + ModuleDisplay> FullyShardedModule<B>
    for FullyShardedExpertParallelMhcFeedForward<B, P, E> {
    fn visit_shards<F: FnMut(&ShardedParameter<B>)>(&self, visitor: &mut F) { self.routed.visit_shards(visitor); self.shared.visit_shards(visitor); }
    fn visit_packed_shards<F: FnMut(&ShardedPackedParameter<B>)>(&self, visitor: &mut F) { self.routed.visit_packed_shards(visitor); self.shared.visit_packed_shards(visitor); }
}
impl<B: Backend, P: FullyShardedAdapterModule<B> + ModuleDisplay, E: FullyShardedAdapterModule<B> + ModuleDisplay> FullyShardedAdapterModule<B>
    for FullyShardedExpertParallelMhcFeedForward<B, P, E> {
    fn visit_adapter_shards<F: FnMut(&ShardedParameter<B>)>(&self, visitor: &mut F) { self.routed.visit_adapter_shards(visitor); self.shared.visit_adapter_shards(visitor); }
}
impl<B: Backend, L: FullyShardedModule<B> + ModuleDisplay, P: FullyShardedModule<B> + ModuleDisplay, E: FullyShardedModule<B> + ModuleDisplay>
    FullyShardedModule<B> for FullyShardedMixedMhcFeedForward<B, L, P, E> {
    fn visit_shards<F: FnMut(&ShardedParameter<B>)>(&self, visitor: &mut F) { match self { Self::Local(value) => value.visit_shards(visitor), Self::Parallel(value) => value.visit_shards(visitor) } }
    fn visit_packed_shards<F: FnMut(&ShardedPackedParameter<B>)>(&self, visitor: &mut F) { match self { Self::Local(value) => value.visit_packed_shards(visitor), Self::Parallel(value) => value.visit_packed_shards(visitor) } }
}
impl<B: Backend, L: FullyShardedAdapterModule<B> + ModuleDisplay, P: FullyShardedAdapterModule<B> + ModuleDisplay, E: FullyShardedAdapterModule<B> + ModuleDisplay>
    FullyShardedAdapterModule<B> for FullyShardedMixedMhcFeedForward<B, L, P, E> {
    fn visit_adapter_shards<F: FnMut(&ShardedParameter<B>)>(&self, visitor: &mut F) { match self { Self::Local(value) => value.visit_adapter_shards(visitor), Self::Parallel(value) => value.visit_adapter_shards(visitor) } }
}