ruDNN 0.1.26

Neural network operators for Ruda.
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
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257

use super::{MoeError, RoutingPlan, elements, empty, float_tensor, kernels, same_device};
use ruda_core::tensor::{DType, Shape};
use ruda_kernel::{
    dsl::{Runtime, calculate_ruda_count_elemwise, prelude::{RudaDim, RudaCount}},
    tensor::{RudaTensor, contiguous::into_contiguous},
};

#[derive(Debug, Clone)]
pub struct DispatchedTokens<R: Runtime> {
    pub(super) routing: RoutingPlan<R>,
    pub(super) values: RudaTensor<R>,
    pub(super) row_experts: RudaTensor<R>,
    pub(super) offsets: RudaTensor<R>,
    pub(super) slot_rows: RudaTensor<R>,
    pub(super) sorted_slots: RudaTensor<R>,
}

impl<R: Runtime> RoutingPlan<R> {
    /// Compact device-side dispatch with no token dropping or host readback.
    /// Expert segments are contiguous; row order within each expert is unspecified.
    pub fn dispatch(self, input: RudaTensor<R>) -> Result<DispatchedTokens<R>, MoeError> {
        float_tensor(&input)?;
        same_device(&input, &self.indices)?;
        if input.meta.num_dims() != 2
            || input.meta.shape()[0] != self.tokens
            || input.meta.shape()[1] == 0
        {
            return Err(MoeError(
                "MoE dispatch requires [tokens, nonzero hidden] input",
            ));
        }
        let hidden = input.meta.shape()[1];
        let rows = elements(&[self.tokens, self.top_k])?;
        let size = elements(&[rows, hidden])?;
        let offset_count = self
            .experts
            .checked_add(1)
            .ok_or(MoeError("MoE offset size overflow"))?;
        elements(&[offset_count])?;
        let counts = empty(&input, [self.experts], DType::U32);
        let offsets = empty(&input, [offset_count], DType::U32);
        let ranks = empty(&input, [rows], DType::U32);
        let sorted_slots = empty(&input, [rows], DType::U32);
        let slot_rows = empty(&input, [rows], DType::U32);
        let row_experts = empty(&input, [rows], DType::U32);
        let values = empty(&input, [rows, hidden], input.dtype);
        let dim = RudaDim::new(input.client.properties(), self.experts);
        kernels::dispatch::clear_counts::launch::<R>(
            &input.client,
            calculate_ruda_count_elemwise(&input.client, self.experts, dim),
            dim,
            counts.clone().into_array_arg(),
        );
        if rows != 0 {
            let dim = RudaDim::new(input.client.properties(), rows);
            kernels::dispatch::count_routes::launch::<R>(
                &input.client,
                calculate_ruda_count_elemwise(&input.client, rows, dim),
                dim,
                self.indices.clone().into_array_arg(),
                counts.clone().into_array_arg(),
                ranks.clone().into_array_arg(),
            );
        }
        let dim = RudaDim::new(input.client.properties(), 1);
        kernels::dispatch::prefix::launch::<R>(
            &input.client,
            calculate_ruda_count_elemwise(&input.client, 1, dim),
            dim,
            counts.into_array_arg(),
            offsets.clone().into_array_arg(),
        );
        if rows != 0 {
            let dim = RudaDim::new(input.client.properties(), rows);
            kernels::dispatch::scatter_routes::launch::<R>(
                &input.client,
                calculate_ruda_count_elemwise(&input.client, rows, dim),
                dim,
                self.indices.clone().into_array_arg(),
                ranks.into_array_arg(),
                offsets.clone().into_array_arg(),
                sorted_slots.clone().into_array_arg(),
                slot_rows.clone().into_array_arg(),
                row_experts.clone().into_array_arg(),
            );
            let input = into_contiguous(input);
            let dim = RudaDim::new(input.client.properties(), size);
            kernels::dispatch::gather::launch::<R>(
                &input.client,
                calculate_ruda_count_elemwise(&input.client, size, dim),
                dim,
                input.clone().into_array_arg(),
                sorted_slots.clone().into_array_arg(),
                values.clone().into_array_arg(),
                hidden as u32,
                self.top_k as u32,
                input.dtype.into(),
            );
        }
        Ok(DispatchedTokens {
            routing: self,
            values,
            row_experts,
            offsets,
            slot_rows,
            sorted_slots,
        })
    }
}

/// Serial preserves the old reduction order; Plane is an explicit GPU tuning option.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum CombineGradientStrategy { Serial, Plane }

#[derive(Debug)]
pub struct CombineBackward<R:Runtime>{ pub dexpert:RudaTensor<R>, pub dweights:RudaTensor<R> }

impl<R: Runtime> DispatchedTokens<R> {
    pub fn values(&self) -> &RudaTensor<R> {
        &self.values
    }
    pub fn row_experts(&self) -> &RudaTensor<R> {
        &self.row_experts
    }
    /// Exclusive prefix offsets `[experts + 1]`, including empty expert segments.
    pub fn expert_offsets(&self) -> &RudaTensor<R> {
        &self.offsets
    }
    pub fn routing(&self) -> &RoutingPlan<R> {
        &self.routing
    }

    /// First-order backward of `combine`. Returns gradient for contiguous
    /// expert rows and FP32 gradients for the selected routing weights. It does
    /// not differentiate the discrete top-k indices.
    pub fn combine_backward(&self, expert_output:&RudaTensor<R>, grad_output:RudaTensor<R>)
        ->Result<CombineBackward<R>,MoeError> {
        self.combine_backward_with_strategy(expert_output, grad_output, CombineGradientStrategy::Serial)
    }

    pub fn combine_backward_with_strategy(&self, expert_output: &RudaTensor<R>,
        grad_output: RudaTensor<R>, strategy: CombineGradientStrategy)
        -> Result<CombineBackward<R>, MoeError>
    {
        float_tensor(expert_output)?; float_tensor(&grad_output)?;
        if !self.values.client.same_execution_queue(&expert_output.client)
            || !self.values.client.same_execution_queue(&grad_output.client) {
            return Err(MoeError("MoE backward operands must share an execution queue"));
        }
        same_device(&self.values,expert_output)?; same_device(&self.values,&grad_output)?;
        let rows=self.routing.tokens*self.routing.top_k;
        if expert_output.meta.num_dims()!=2 || expert_output.meta.shape()[0]!=rows
            || expert_output.meta.shape()[1] == 0
            || grad_output.meta.shape()[..]!=[self.routing.tokens,expert_output.meta.shape()[1]]
            || expert_output.dtype!=self.values.dtype || grad_output.dtype!=self.values.dtype
        { return Err(MoeError("MoE combine backward shape/dtype mismatch")); }
        let width=expert_output.meta.shape()[1]; let values=empty(expert_output,[rows,width],expert_output.dtype);
        let dweights=empty(expert_output,[self.routing.tokens,self.routing.top_k],DType::F32);
        let size=elements(&[rows,width])?;
        let lanes = values.client.properties().hardware.plane_size_max;
        if strategy == CombineGradientStrategy::Plane
            && (!matches!(lanes, 32 | 64) || lanes > values.client.properties().hardware.max_ruda_dim.0
                || rows > values.client.properties().hardware.max_ruda_count.0 as usize) {
            return Err(MoeError("parallel MoE weight gradient requires a legal full-plane grid"));
        }
        if size!=0 { let grad=into_contiguous(grad_output); let expert=into_contiguous(expert_output.clone());
            let dim=RudaDim::new(values.client.properties(),size);
            kernels::dispatch::combine_backward_values::launch::<R>(&values.client,
                calculate_ruda_count_elemwise(&values.client,size,dim),dim,
                grad.clone().into_array_arg(),self.routing.weights.clone().into_array_arg(),self.sorted_slots.clone().into_array_arg(),
                values.clone().into_array_arg(),width as u32,self.routing.top_k as u32,values.dtype.into(),self.routing.weights.dtype.into());
            let assignments=self.routing.tokens*self.routing.top_k;
            let dim=RudaDim::new(values.client.properties(),assignments);
            if strategy == CombineGradientStrategy::Plane {
                kernels::dispatch::combine_backward_weights_plane::launch::<R>(&values.client,
                    RudaCount::Static(assignments as u32, 1, 1), RudaDim::new_1d(lanes),
                    grad.into_array_arg(), expert.into_array_arg(), self.slot_rows.clone().into_array_arg(),
                    dweights.clone().into_array_arg(), width as u32, self.routing.top_k as u32,
                    lanes as usize, values.dtype.into());
            } else {
                kernels::dispatch::combine_backward_weights::launch::<R>(&values.client,
                    calculate_ruda_count_elemwise(&values.client,assignments,dim),dim,
                    grad.into_array_arg(),expert.into_array_arg(),self.slot_rows.clone().into_array_arg(),dweights.clone().into_array_arg(),
                    width as u32,self.routing.top_k as u32,values.dtype.into());
            }
        }
        Ok(CombineBackward{dexpert:values,dweights})
    }

    /// Backward of dispatch's COPY, not of combine: sum all selected expert-row
    /// input gradients for each source token. Never multiply by routing weights
    /// again. This closes the input-gradient chain after expert backward.
    pub fn dispatch_backward(&self, grad_values: RudaTensor<R>) -> Result<RudaTensor<R>, MoeError> {
        float_tensor(&grad_values)?;
        same_device(&self.values, &grad_values)?;
        if grad_values.meta.shape() != self.values.meta.shape() || grad_values.dtype != self.values.dtype
            || !grad_values.client.same_execution_queue(&self.values.client) {
            return Err(MoeError("dispatch backward requires matching expert-row gradients on the same queue"));
        }
        let width = self.values.meta.shape()[1];
        let size = elements(&[self.routing.tokens, width])?;
        let output = empty(&self.values, [self.routing.tokens, width], self.values.dtype);
        if size != 0 {
            let grad_values = into_contiguous(grad_values);
            let dim = RudaDim::new(output.client.properties(), size);
            kernels::dispatch::dispatch_backward::launch::<R>(&output.client,
                calculate_ruda_count_elemwise(&output.client, size, dim), dim,
                grad_values.into_array_arg(), self.slot_rows.clone().into_array_arg(),
                output.clone().into_array_arg(), width as u32, self.routing.top_k as u32, output.dtype.into());
        }
        Ok(output)
    }

    /// Undo dispatch and add weighted expert outputs in ascending expert-ID order.
    /// Each weighted contribution and each addition is rounded to the output dtype.
    pub fn combine(&self, expert_output: RudaTensor<R>) -> Result<RudaTensor<R>, MoeError> {
        float_tensor(&expert_output)?;
        same_device(&self.values, &expert_output)?;
        let rows = self.routing.tokens * self.routing.top_k;
        if expert_output.meta.num_dims() != 2
            || expert_output.meta.shape()[0] != rows
            || expert_output.meta.shape()[1] == 0
            || expert_output.dtype != self.values.dtype
        {
            return Err(MoeError(
                "MoE combine requires one output row per dispatched row with matching dtype",
            ));
        }
        let width = expert_output.meta.shape()[1];
        elements(&[rows, width])?;
        let size = elements(&[self.routing.tokens, width])?;
        let output = empty(
            &expert_output,
            Shape::from([self.routing.tokens, width]),
            expert_output.dtype,
        );
        if size != 0 {
            let expert_output = into_contiguous(expert_output);
            let dim = RudaDim::new(output.client.properties(), size);
            kernels::dispatch::combine::launch::<R>(
                &output.client,
                calculate_ruda_count_elemwise(&output.client, size, dim),
                dim,
                expert_output.into_array_arg(),
                self.routing.weights.clone().into_array_arg(),
                self.routing.indices.clone().into_array_arg(),
                self.slot_rows.clone().into_array_arg(),
                output.clone().into_array_arg(),
                width as u32,
                self.routing.top_k as u32,
                output.dtype.into(),
                self.routing.weights.dtype.into(),
            );
        }
        Ok(output)
    }
}