burn_dragon_core 0.5.0

burn dragon core model and utilities
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
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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206

#![recursion_limit = "256"]

//! Shared Dragon/Hatchling core math and state contracts.
//!
//! Preferred library-facing surface:
//! - [`api::config`]
//! - [`api::state`]
//! - [`api::recurrent`]
//! - [`api::mhc`]
//! - [`api::experimental`] for research-only connector and sequence-kernel surfaces
//! - [`api::expert`] for lower-level implementation details

pub mod constants;
pub mod experimental;
pub mod kernel;
pub mod model;
pub mod positional;

pub use model::LanguageModuleLrScaleTarget;

pub mod api {
    //! Curated public surface for the Dragon core.
    //!
    //! These modules group the primary stable-ish concepts rather than exposing the full internal
    //! implementation layout.

    pub mod config {
        pub use crate::model::LanguageModuleLrScaleTarget;
        pub use crate::{
            AttentionResidualConfig, BDHConfig, BdhFiringTargetConfig, BdhFiringTargetKind,
            BdhInitializationConfig, BdhInitializationKind, BdhNeuronGainConfig, BdhNeuronGainKind,
            BdhResidualScalingConfig, BdhResidualScalingKind, BdhTopologyPriorConfig,
            BdhTopologyPriorKind, BitNetLowBitProtocol, BlockAttentionResidualConfig,
            BlockAttentionResidualSummaryMode, ClockedSlowMemoryConfig, DragonNormConfig,
            DragonNormKind, FusedAttentionExecutor, FusedKernelConfig, FusedProjectionExecutor,
            LanguageHeadConfig, LatentFanoutScheduleConfig, LowBitActivationFormat,
            LowBitActivationGrouping, LowBitInferenceMode, LowBitMemoryBucketEstimate,
            LowBitMemoryEstimateInput, LowBitNativeProjectionProfileSnapshot,
            LowBitQuantizationConfig, LowBitRhoConfig, LowBitSavedActivationConfig,
            LowBitSavedActivationInventory, LowBitSavedActivationMode,
            LowBitSavedActivationRecomputePolicy, LowBitSavedActivationTensorInventoryEntry,
            LowBitTargetModule, LowBitTrainingMode, LowBitWeightFormat, LowBitWeightGrouping,
            MambaSequenceConfig, ManifoldHyperConnectionCoefficientPolicy,
            ManifoldHyperConnectionsConfig, ResidualConnectorKind, RhoCompressionConfig,
            RhoCompressionInterval, RhoPrecisionConfig, SequenceKernelConfig, SequenceMemorySystem,
            SequenceTrainingExecutor, SummaryMemoryConfig, YNeuronRecurrenceConfig,
        };
        pub use burn_dragon_kernel::api::projection::LowrankGradInputExecutor;
    }

    pub mod state {
        pub use crate::{
            BankedRhoState, ModelState, StructuredGridState, StructuredRoutingSpec,
            StructuredStepMode, StructuredTopologyState,
        };
    }

    pub mod recurrent {
        pub use crate::{
            BDH, DragonNorm, HaltHead, LanguageMhcLayerDiagnostics, LanguagePipelineState,
            LogitsProjectionProfileSnapshot, LowBitTrainingProjectionMemoryProfileSnapshot,
            LowBitTrainingProjectionMemoryStageSnapshot, LowBitTrainingQuantizeProfileSnapshot,
            LowRankResidualMemoryProfileSnapshot, LowRankResidualMemoryStageSnapshot,
            LowRankResidualOutput, LowRankResidualProfileSnapshot, StructuredDenseUpdateOutput,
            logits_projection_profile_reset, logits_projection_profile_snapshot,
            low_bit_training_lowrank_memory_profile_snapshot,
            low_bit_training_quantize_profile_snapshot, lowrank_residual_memory_profile_reset,
            lowrank_residual_memory_profile_snapshot, lowrank_residual_profile_reset,
            lowrank_residual_profile_snapshot, lowrank_residual_step, lowrank_residual_step_next,
            structured_dense_update_tokens,
        };
        #[cfg(any(feature = "probe", test))]
        pub use crate::{
            HeadTensorComparisonDiagnostics, HeadTensorGeometryDiagnostics,
            LanguageBdhInitLayerDiagnostics, LanguageLayerStateDeltaDiagnostics,
            LanguageLayerStateSummaryDiagnostics, LanguageLowRankLayerComparisonDiagnostics,
            LanguageLowRankLayerGeometryDiagnostics, TensorComparisonDiagnostics,
            TensorDistributionDiagnostics, TensorStateDeltaDiagnostics,
            TensorStateSummaryDiagnostics, compare_model_states, summarize_model_state,
        };
    }

    pub mod mhc {
        pub use crate::{
            ManifoldHyperConnectionCoefficients, ManifoldHyperConnectionStreamCoefficients,
            ManifoldHyperConnectionStreamOutput, ManifoldHyperConnectionWidthOutput,
            ManifoldHyperConnections, mhc_merge, mhc_merge_with_coefficients, mhc_passthrough,
            mhc_passthrough_with_coefficients, mhc_split, mhc_split_with_coefficients,
        };
    }

    pub mod experimental {
        //! Experimental BDH surfaces that are still expected to evolve quickly.
        //!
        //! These are available for research and ablation work, but they are intentionally grouped
        //! away from the smaller stable-ish `config` / `state` / `recurrent` surface.

        pub mod connectors {
            pub use crate::{
                AttentionResidual, AttentionResidualConfig, BlockAttentionResidual,
                BlockAttentionResidualConfig, BlockAttentionResidualSummaryMode,
                ResidualConnectorKind,
            };
        }

        pub mod sequence {
            pub use crate::{
                MambaSequenceConfig, SequenceKernelConfig, SequenceMemorySystem,
                SequenceTrainingExecutor,
            };
        }

        pub mod bitnet_reference {
            pub use crate::experimental::bitnet_reference::{
                BdhBitNetStaticArtifacts, BitLinearReferenceSpec, BitNetReferenceActivationMode,
                BitNetReferenceWeightMode, PackedTernaryBuffer, PackedWeightArtifact,
                PackedWeightEncoding, QuantizedBuffer, bitlinear_reference_forward,
                dequantize_activation_i8, dequantize_weight_codes, pack_binary_1bit,
                pack_ternary_2bit, pack_weight_artifact_from_format,
                quantize_activation_symmetric_i8, quantize_binary_sign, quantize_ternary_absmean,
                ste_passthrough, unpack_binary_1bit, unpack_ternary_2bit,
                unpack_weight_artifact_to_f32,
            };
        }
    }

    pub mod expert {
        //! Lower-level implementation surface for advanced callers.
        //!
        //! Prefer the smaller `config` / `state` / `recurrent` / `mhc` modules unless you
        //! explicitly need internal kernel/layout utilities.

        pub use crate::constants;
        pub use crate::kernel;
        pub use crate::model;
        pub use crate::positional;
    }
}

pub use burn_dragon_kernel::api::projection::LowrankGradInputExecutor;
pub use kernel::{BlockPattern1d, BlockPattern2d, BlockSparseConfig};
#[cfg(any(feature = "probe", test))]
pub use model::LanguageBdhInitLayerDiagnostics;
#[cfg(any(feature = "viz", feature = "probe"))]
pub use model::LayerVizState;
pub use model::{
    AttentionResidual, AttentionResidualConfig, BDH, BDHConfig, BankedRhoState,
    BdhActivationThresholds, BdhBitNetDeployScaffold, BdhFiringTargetConfig, BdhFiringTargetKind,
    BdhInitializationConfig, BdhInitializationKind, BdhInitializer, BdhNeuronGainConfig,
    BdhNeuronGainKind, BdhProjectionRole, BdhResidualScalingConfig, BdhResidualScalingKind,
    BdhTopologyPriorConfig, BdhTopologyPriorKind, BitNetLowBitProtocol, BlockAttentionResidual,
    BlockAttentionResidualConfig, BlockAttentionResidualSummaryMode, ClockedSlowMemoryConfig,
    DragonNorm, DragonNormConfig, DragonNormKind, FusedAttentionExecutor, FusedKernelConfig,
    FusedProjectionExecutor, HaltHead, LanguageHeadConfig, LanguageMhcLayerDiagnostics,
    LanguagePipelineState, LatentFanoutScheduleConfig, LogitsProjectionProfileSnapshot,
    LowBitActivationFormat, LowBitActivationGrouping, LowBitInferenceMode,
    LowBitKernelCapabilities, LowBitKernelFallbackReason, LowBitKernelPlan,
    LowBitKernelRuntimeKind, LowBitMemoryBucketEstimate, LowBitMemoryEstimateInput,
    LowBitNativeProjectionProfileSnapshot, LowBitProjectionPlan, LowBitQuantizationConfig,
    LowBitRhoConfig, LowBitSavedActivationCache, LowBitSavedActivationConfig,
    LowBitSavedActivationInventory, LowBitSavedActivationMode,
    LowBitSavedActivationRecomputePolicy, LowBitSavedActivationTensorInventoryEntry,
    LowBitTargetModule, LowBitTrainingMode, LowBitTrainingProjectionMemoryProfileSnapshot,
    LowBitTrainingProjectionMemoryStageSnapshot, LowBitTrainingQuantizeProfileSnapshot,
    LowBitWeightFormat, LowBitWeightGrouping, LowRankResidualMemoryProfileSnapshot,
    LowRankResidualMemoryStageSnapshot, LowRankResidualOutput, LowRankResidualProfileSnapshot,
    MambaSequenceConfig, ManifoldHyperConnectionCoefficientPolicy,
    ManifoldHyperConnectionCoefficients, ManifoldHyperConnectionStreamCoefficients,
    ManifoldHyperConnectionStreamOutput, ManifoldHyperConnectionWidthOutput,
    ManifoldHyperConnections, ManifoldHyperConnectionsConfig, MicroTransformerBlock, ModelState,
    PackedLowBitProjectionArtifacts, PackedRhoBlockEncoding, PackedRhoBlockState,
    PackedRhoInt8DeviceState, PackedSavedActivationBuffer, PackedSavedActivationState,
    ResidualConnectorKind, RhoCompressionConfig, RhoCompressionInterval, RhoCompressionQualityGate,
    RhoCompressionStatsSnapshot, RhoPrecisionConfig, SequenceKernelConfig, SequenceMemorySystem,
    SequenceTrainingExecutor, SharedLowrankActivationBatchStats,
    SharedLowrankContinualBackpropRuntime, SharedLowrankFeatureMetrics, SharedLowrankParamIds,
    StructuredBankRole, StructuredDenseUpdateOutput, StructuredGridState, StructuredRouteOperation,
    StructuredRoutePattern, StructuredRouteSpec, StructuredRoutingSpec, StructuredStepMode,
    StructuredTopologyState, SummaryMemoryConfig, YNeuronRecurrenceConfig,
    build_low_bit_saved_activation_inventory, estimate_low_bit_memory_buckets,
    logits_projection_profile_reset, logits_projection_profile_snapshot,
    low_bit_kernel_capabilities, low_bit_kernel_capabilities_for_backend_name,
    low_bit_native_decoder_tail_profile_snapshot, low_bit_native_lowrank_profile_snapshot,
    low_bit_native_projection_profile_reset, low_bit_training_lowrank_memory_profile_snapshot,
    low_bit_training_quantize_profile_snapshot, lowrank_residual_memory_profile_reset,
    lowrank_residual_memory_profile_snapshot, lowrank_residual_profile_reset,
    lowrank_residual_profile_snapshot, lowrank_residual_step, lowrank_residual_step_next,
    mhc_merge, mhc_merge_with_coefficients, mhc_passthrough, mhc_passthrough_with_coefficients,
    mhc_split, mhc_split_with_coefficients, near_critical_embedding_initializer,
    near_critical_embedding_std, near_critical_projection_std, near_critical_residual_output_std,
    pack_saved_activation_state, resolve_low_bit_kernel_plan,
    resolve_low_bit_kernel_plan_for_backend_name, rho_compression_profile_reset,
    rho_compression_profile_snapshot, rho_compression_snapshot_passes_gate,
    structured_dense_update_tokens, structured_predict_decay, target_major_apply_decay,
    target_major_decay_add, target_major_identity_read, target_major_identity_write,
    target_major_outer_product, unpack_saved_activation_state,
};
#[cfg(any(feature = "probe", test))]
pub use model::{
    HeadTensorComparisonDiagnostics, HeadTensorGeometryDiagnostics,
    LanguageLayerStateDeltaDiagnostics, LanguageLayerStateSummaryDiagnostics,
    LanguageLowRankLayerComparisonDiagnostics, LanguageLowRankLayerGeometryDiagnostics,
    TensorComparisonDiagnostics, TensorDistributionDiagnostics, TensorStateDeltaDiagnostics,
    TensorStateSummaryDiagnostics, compare_model_states, summarize_model_state,
};
pub use positional::RotaryEmbedding;