burn_dragon_kernel 0.5.0

Fused GPU kernel crate for burn_dragon execution paths
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

#![recursion_limit = "256"]

//! Fused Dragon execution kernels and compiled-plan helpers.
//!
//! Preferred library-facing surface:
//! - [`api::recurrent`]
//! - [`api::spatial`]
//! - [`api::graph`]
//! - [`api::low_bit`]
//! - [`api::expert`] for lower-level kernel-plan access

mod dense_attention;
mod dense_causal_attention;
mod dense_scores;
mod fusion_compat;
mod local_grid_rho;
mod low_bit;
mod profiling;
mod recurrent;
mod relu_lowrank;
mod sequence;
mod sparse_graph_rho;
mod structured_pyramid_rho;
mod vision_rho;

pub mod api {
    //! Curated public surface for the fused execution layer.
    //!
    //! This mirrors the active kernel families instead of exposing the entire file/module layout.

    pub use crate::kernels::{attention, graph, low_bit, projection, recurrent, spatial};

    pub mod expert {
        //! Lower-level fused-kernel surface for advanced callers.

        pub use crate::kernels;
    }
}

/// Namespaced fused-kernel families exposed by this crate.
pub mod kernels {
    /// Sequence-kernel family namespace used by the language line.
    pub mod sequence {
        pub use crate::sequence::{linear, mamba, mamba2, mamba3, rwkv8};
    }

    /// Dense attention and score kernels used by language/vision recurrent executors.
    pub mod attention {
        pub use crate::dense_attention::{
            CompiledDenseAttentionPlan, supports_dense_attention_backend,
            try_fused_dense_row_l1_attention_wgpu, try_fused_dense_row_l1_attention_wgpu_with_plan,
        };
        pub use crate::dense_causal_attention::{
            CompiledDenseCausalAttentionPlan, supports_dense_causal_attention_backend,
            try_fused_dense_causal_attention_wgpu, try_fused_dense_causal_attention_wgpu_with_plan,
        };
        pub use crate::dense_scores::{
            CompiledDenseScoresPlan, supports_dense_scores_backend,
            try_fused_dense_row_l1_scores_wgpu, try_fused_dense_row_l1_scores_wgpu_with_plan,
        };
    }

    /// Sparse graph-routing kernels over recurrent `rho` state.
    pub mod graph {
        pub use crate::sparse_graph_rho::{
            SparseGraphCsr, SparseGraphRhoAttentionError, SparseGraphRhoAttentionOutput,
            SparseGraphRhoProfileSnapshot, fused_sparse_graph_rho_attention_wgpu,
            sparse_graph_rho_profile_reset, sparse_graph_rho_profile_snapshot,
            supports_sparse_graph_rho_backend, try_fused_sparse_graph_rho_attention_wgpu,
        };
    }

    /// Core recurrent attention kernels.
    pub mod recurrent {
        pub use crate::recurrent::{
            CompiledRecurrentAttentionPlan, RecurrentAttentionOutput, RecurrentProfileSnapshot,
            recurrent_profile_reset, recurrent_profile_snapshot,
            supports_backend as supports_recurrent_backend, try_fused_recurrent_attention_wgpu,
            try_fused_recurrent_attention_wgpu_with_plan,
        };
    }

    /// Fused low-rank projection kernels used in recurrent x/y projection paths.
    pub mod projection {
        pub use crate::relu_lowrank::{
            LowrankForwardRouteProfileSnapshot, LowrankGradInputExecutor,
            LowrankProjectionProfileSnapshot, relu_lowrank_forward_profile_reset,
            relu_lowrank_forward_profile_snapshot, relu_lowrank_forward_route_profile_reset,
            relu_lowrank_forward_route_profile_snapshot, relu_lowrank_grad_input_profile_reset,
            relu_lowrank_grad_input_profile_snapshot, relu_lowrank_grad_weight_profile_reset,
            relu_lowrank_grad_weight_profile_snapshot, supports_relu_lowrank_projection_backend,
            try_fused_relu_lowrank_projection_wgpu,
            try_fused_relu_lowrank_projection_wgpu_with_executor,
        };
    }

    /// Device-executed low-bit helpers for packed/static BitNet-style paths.
    pub mod low_bit {
        pub use crate::low_bit::{
            PackedRhoInt8BlockDeviceTensors, cached_wgpu_packed_dot_decoder_tail_support,
            cached_wgpu_packed_dot_lowrank_support, diagnose_wgpu_packed_dot_decoder_tail,
            diagnose_wgpu_packed_dot_lowrank_projection,
            diagnose_wgpu_quantize_pack_activation_i8x4, pack_decoder_input_codes_i8x4,
            pack_decoder_weight_codes_i8x4, pack_lowrank_input_codes_i8x4,
            pack_lowrank_weight_codes_i8x4, pack_rho_int8_block_device_reference,
            packed_decoder_tail_device_reference, packed_decoder_tail_grad_input_device_reference,
            packed_decoder_tail_grad_weight_device_reference,
            packed_lowrank_grad_input_device_reference,
            packed_lowrank_grad_weight_device_reference,
            packed_lowrank_projection_device_reference, supports_packed_low_bit_device_backend,
            supports_packed_rho_int8_block_device_backend, try_cube_fused_packed_decoder_tail_wgpu,
            try_cube_fused_packed_lowrank_projection_wgpu, try_fused_packed_decoder_tail,
            try_fused_packed_decoder_tail_grad_input, try_fused_packed_decoder_tail_grad_weight,
            try_fused_packed_decoder_tail_training_autodiff, try_fused_packed_lowrank_grad_input,
            try_fused_packed_lowrank_grad_weight, try_fused_packed_lowrank_projection,
            try_fused_packed_lowrank_training_autodiff,
            try_fused_packed_lowrank_training_autodiff_cuda_device_projection_scale,
            try_raw_cuda_packed_decoder_tail, try_raw_cuda_packed_decoder_tail_device_scale,
            try_raw_cuda_packed_decoder_tail_grad_input,
            try_raw_cuda_packed_decoder_tail_grad_weight,
            try_raw_cuda_packed_decoder_tail_prepacked_input,
            try_raw_cuda_packed_decoder_tail_prepacked_input_device_scale,
            try_raw_cuda_packed_lowrank_grad_input, try_raw_cuda_packed_lowrank_grad_weight,
            try_raw_cuda_packed_lowrank_projection,
            try_raw_cuda_packed_lowrank_projection_device_scale,
            try_raw_cuda_packed_lowrank_projection_prepacked_input,
            try_raw_cuda_packed_lowrank_projection_prepacked_input_device_scale,
            try_raw_cuda_quantize_pack_activation_i8x4, try_wgpu_packed_dot_decoder_tail,
            try_wgpu_packed_dot_decoder_tail_device_scale,
            try_wgpu_packed_dot_decoder_tail_prepacked_input_device_scale,
            try_wgpu_packed_dot_lowrank_projection,
            try_wgpu_packed_dot_lowrank_projection_device_scale,
            try_wgpu_packed_dot_lowrank_projection_from_f32_device_scale,
            try_wgpu_packed_dot_lowrank_projection_prepacked_input_device_scale,
            try_wgpu_quantize_activation_codes_i32, try_wgpu_quantize_pack_activation_i8x4,
            unpack_rho_int8_block_device_reference,
        };
        #[cfg(feature = "cuda")]
        pub use crate::low_bit::{
            packed_decoder_tail_grad_input_from_float_decoder_cuda,
            packed_lowrank_grad_input_from_float_weight_cuda,
            packed_lowrank_grad_input_from_transposed_float_weight_cuda,
        };
    }

    /// Spatial/topological recurrent kernels for local-grid, pyramid, and vision-rho paths.
    pub mod spatial {
        pub use crate::local_grid_rho::{
            CompiledLocalGridRhoPlan, LocalGridNeighborhood, LocalGridRhoAttentionOutput,
            LocalGridRhoPlanSpec, LocalGridRhoProfileSnapshot, LocalGridShape2d,
            local_grid_rho_profile_reset, local_grid_rho_profile_snapshot,
            supports_local_grid_rho_backend, try_fused_local_grid_rho_attention_wgpu,
            try_fused_local_grid_rho_attention_wgpu_head_decay,
            try_fused_local_grid_rho_attention_wgpu_head_decay_with_plan,
        };
        pub use crate::structured_pyramid_rho::{
            CompiledStructuredPyramidRhoPlan, CompiledStructuredPyramidSplitPlan,
            StructuredPyramidBankMode, StructuredPyramidCoarseOnlyNoPatchStepInput,
            StructuredPyramidCoarseOnlyStepInput, StructuredPyramidCoarseOnlyStepOutput,
            StructuredPyramidProfileSnapshot, StructuredPyramidRhoStepInput,
            StructuredPyramidRhoStepOutput, StructuredPyramidShape, StructuredPyramidSplitPlanSpec,
            StructuredPyramidSplitRhoStepInput, reference_structured_pyramid_rho_step,
            structured_pyramid_profile_reset, structured_pyramid_profile_snapshot,
            supports_structured_pyramid_rho_backend,
            try_fused_structured_pyramid_coarse_only_no_patch_step_wgpu_with_plan,
            try_fused_structured_pyramid_coarse_only_step_wgpu_with_plan,
            try_fused_structured_pyramid_rho_step_wgpu,
            try_fused_structured_pyramid_rho_step_wgpu_with_plan,
            try_fused_structured_pyramid_split_step_wgpu_with_plan,
        };
        pub use crate::vision_rho::{
            VisionRhoAttentionOutput, supports_vision_rho_backend,
            try_fused_vision_rho_attention_wgpu,
        };
    }
}