ruCCL 0.21.27

Ruda collective communication algorithms and orchestration.
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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322

use super::{TensorDevice, TensorDeviceError, TensorElement, storage};
use crate::ReduceOperation;
use crate::rank::{ReductionOperation, communicator::RankCommunicator};
use ruda_core::tensor::collective::CollectiveShape;
use ruda_tensor::{Backend, DType, Shape, TensorMetadata, bf16, f16};

impl<B: Backend> ruda_tensor::collective::TensorCollective<B> for RankCommunicator<TensorDevice<B>> {
    type Error = TensorDeviceError;

    fn world_size(&self) -> u32 {
        RankCommunicator::world_size(self)
    }

    fn all_gather_float(
        &self,
        value: B::FloatTensorPrimitive,
    ) -> Result<B::FloatTensorPrimitive, Self::Error> {
        RankCommunicator::all_gather_float(self, value)
    }

    fn reduce_scatter_sum(
        &self,
        value: B::FloatTensorPrimitive,
    ) -> Result<B::FloatTensorPrimitive, Self::Error> {
        RankCommunicator::reduce_scatter_float(self, value, ReduceOperation::Sum)
    }
}

impl<B: Backend> ruda_tensor::collective::ReplicatedTensorCollective<B>
    for RankCommunicator<TensorDevice<B>>
{
    fn all_reduce_sum(
        &self,
        value: B::FloatTensorPrimitive,
    ) -> Result<B::FloatTensorPrimitive, Self::Error> {
        RankCommunicator::all_reduce_float(self, value, ReduceOperation::Sum)
    }
}

impl<B: Backend> ruda_tensor::collective::BroadcastTensorCollective<B>
    for RankCommunicator<TensorDevice<B>>
{
    fn rank(&self) -> u32 {
        RankCommunicator::rank(self)
    }

    fn broadcast_float(
        &self,
        value: B::FloatTensorPrimitive,
        root: u32,
    ) -> Result<B::FloatTensorPrimitive, Self::Error> {
        RankCommunicator::broadcast_float(self, value, root)
    }
}

impl<B: Backend> RankCommunicator<TensorDevice<B>> {
    /// Gather equal-size floating tensors along axis zero in rank order.
    /// Retains dtype, remaining axes and input snapshots; transfers are host-staged.
    pub fn all_gather_float(
        &self,
        value: B::FloatTensorPrimitive,
    ) -> Result<B::FloatTensorPrimitive, TensorDeviceError> {
        match value.dtype() {
            DType::F32 => self.sharded_float::<f32>(value, None),
            DType::F16 => self.sharded_float::<f16>(value, None),
            DType::BF16 => self.sharded_float::<bf16>(value, None),
            dtype => Err(TensorDeviceError::UnsupportedDType(dtype)),
        }
    }

    /// Reduce floating tensors and return this rank's equal axis-zero shard.
    /// The leading axis must divide by world size. All ranks use matching contracts.
    pub fn reduce_scatter_float(
        &self,
        value: B::FloatTensorPrimitive,
        operation: ReduceOperation,
    ) -> Result<B::FloatTensorPrimitive, TensorDeviceError> {
        match value.dtype() {
            DType::F32 => self.sharded_float::<f32>(value, Some(operation)),
            DType::F16 => self.sharded_float::<f16>(value, Some(operation)),
            DType::BF16 => self.sharded_float::<bf16>(value, Some(operation)),
            dtype => Err(TensorDeviceError::UnsupportedDType(dtype)),
        }
    }

    /// Gather I32/I64 tensors in rank order without floating-point conversion.
    pub fn all_gather_int(
        &self,
        value: B::IntTensorPrimitive,
    ) -> Result<B::IntTensorPrimitive, TensorDeviceError> {
        match value.dtype() {
            DType::I32 => self.sharded_int::<i32>(value, None),
            DType::I64 => self.sharded_int::<i64>(value, None),
            dtype => Err(TensorDeviceError::UnsupportedDType(dtype)),
        }
    }

    /// Reduce I32/I64 tensors using backend arithmetic and return an axis-zero shard.
    pub fn reduce_scatter_int(
        &self,
        value: B::IntTensorPrimitive,
        operation: ReductionOperation,
    ) -> Result<B::IntTensorPrimitive, TensorDeviceError> {
        match value.dtype() {
            DType::I32 => self.sharded_int::<i32>(value, Some(operation)),
            DType::I64 => self.sharded_int::<i64>(value, Some(operation)),
            dtype => Err(TensorDeviceError::UnsupportedDType(dtype)),
        }
    }

    fn sharded_float<T: TensorElement>(
        &self,
        value: B::FloatTensorPrimitive,
        operation: Option<ReduceOperation>,
    ) -> Result<B::FloatTensorPrimitive, TensorDeviceError> {
        let execution = self.execution();
        execution.validate_type::<T>()?;
        if &B::float_device(&value) != execution.device() {
            return Err(TensorDeviceError::DeviceMismatch);
        }
        let plan = sharded_shape(value.shape(), self.world_size(), operation.is_some())?;
        storage::checked_length::<T>(plan.input_elements)?;
        storage::checked_length::<T>(plan.output_elements)?;
        let buffer = execution
            .import_float::<T>(B::float_reshape(value, Shape::new([plan.input_elements])))?;
        let collective = self.tensor_collective::<T>();
        let (output, _) = if operation.is_some() {
            collective.reduce_scatter(&buffer, ReductionOperation::Sum, &ReductionOperation::Sum)?
        } else {
            collective.all_gather(&buffer)?
        };
        let mut output = output.float_tensor()?;
        if operation == Some(ReduceOperation::Mean) {
            output = B::float_div_scalar(output, (self.world_size() as f32).into());
        }
        Ok(B::float_reshape(output, plan.output))
    }

    fn sharded_int<T: TensorElement>(
        &self,
        value: B::IntTensorPrimitive,
        operation: Option<ReductionOperation>,
    ) -> Result<B::IntTensorPrimitive, TensorDeviceError> {
        let execution = self.execution();
        execution.validate_type::<T>()?;
        if &B::int_device(&value) != execution.device() {
            return Err(TensorDeviceError::DeviceMismatch);
        }
        let plan = sharded_shape(value.shape(), self.world_size(), operation.is_some())?;
        storage::checked_length::<T>(plan.input_elements)?;
        storage::checked_length::<T>(plan.output_elements)?;
        let buffer =
            execution.import_int::<T>(B::int_reshape(value, Shape::new([plan.input_elements])))?;
        let collective = self.tensor_collective::<T>();
        let (output, _) = if let Some(operation) = operation {
            let kernel = execution.reduction_kernel::<T>(operation)?;
            collective.reduce_scatter(&buffer, operation, &kernel)?
        } else {
            collective.all_gather(&buffer)?
        };
        Ok(B::int_reshape(output.int_tensor()?, plan.output))
    }

    /// Broadcast an I32 or I64 tensor without converting integer storage to floating point.
    /// Shape, dtype and each rank's local device are retained; transfers are host-staged.
    pub fn broadcast_int(
        &self,
        value: B::IntTensorPrimitive,
        root: u32,
    ) -> Result<B::IntTensorPrimitive, TensorDeviceError> {
        match value.dtype() {
            DType::I32 => self.integer_collective::<i32>(value, None, root),
            DType::I64 => self.integer_collective::<i64>(value, None, root),
            dtype => Err(TensorDeviceError::UnsupportedDType(dtype)),
        }
    }

    /// Reduce an I32 or I64 tensor using the selected backend's integer arithmetic.
    /// Supports sum, product, minimum, maximum and integer bitwise operations.
    /// All ranks must use matching shapes, dtypes and operations in the same order.
    pub fn all_reduce_int(
        &self,
        value: B::IntTensorPrimitive,
        operation: ReductionOperation,
    ) -> Result<B::IntTensorPrimitive, TensorDeviceError> {
        match value.dtype() {
            DType::I32 => self.integer_collective::<i32>(value, Some(operation), 0),
            DType::I64 => self.integer_collective::<i64>(value, Some(operation), 0),
            dtype => Err(TensorDeviceError::UnsupportedDType(dtype)),
        }
    }

    fn integer_collective<T: TensorElement>(
        &self,
        value: B::IntTensorPrimitive,
        operation: Option<ReductionOperation>,
        root: u32,
    ) -> Result<B::IntTensorPrimitive, TensorDeviceError> {
        let execution = self.execution();
        execution.validate_type::<T>()?;
        if &B::int_device(&value) != execution.device() {
            return Err(TensorDeviceError::DeviceMismatch);
        }
        let shape = value.shape();
        let length = shape.iter().try_fold(1_usize, |length, dim| {
            length
                .checked_mul(*dim)
                .ok_or(TensorDeviceError::InvalidBuffer(
                    "collective tensor element count overflow",
                ))
        })?;
        storage::checked_length::<T>(length)?;
        let buffer = execution.import_int::<T>(B::int_reshape(value, Shape::new([length])))?;
        if let Some(operation) = operation {
            let kernel = execution.reduction_kernel::<T>(operation)?;
            self.tensor_collective::<T>()
                .all_reduce(&buffer, operation, &kernel)?;
        } else {
            self.tensor_collective::<T>().broadcast(&buffer, root)?;
        }
        Ok(B::int_reshape(buffer.int_tensor()?, shape))
    }

    /// Broadcast a floating tensor, retaining its shape, dtype and local device.
    pub fn broadcast_float(
        &self,
        value: B::FloatTensorPrimitive,
        root: u32,
    ) -> Result<B::FloatTensorPrimitive, TensorDeviceError> {
        match value.dtype() {
            DType::F32 => self.broadcast_typed::<f32>(value, root),
            DType::F16 => self.broadcast_typed::<f16>(value, root),
            DType::BF16 => self.broadcast_typed::<bf16>(value, root),
            dtype => Err(TensorDeviceError::UnsupportedDType(dtype)),
        }
    }

    fn broadcast_typed<T: TensorElement>(
        &self,
        value: B::FloatTensorPrimitive,
        root: u32,
    ) -> Result<B::FloatTensorPrimitive, TensorDeviceError> {
        let execution = self.execution();
        execution.validate_type::<T>()?;
        if &B::float_device(&value) != execution.device() {
            return Err(TensorDeviceError::DeviceMismatch);
        }
        let shape = value.shape();
        let length = shape.iter().try_fold(1_usize, |length, dim| {
            length.checked_mul(*dim).ok_or(TensorDeviceError::InvalidBuffer(
                "collective tensor element count overflow",
            ))
        })?;
        storage::checked_length::<T>(length)?;
        let buffer = execution.import_float::<T>(B::float_reshape(value, Shape::new([length])))?;
        self.tensor_collective::<T>().broadcast(&buffer, root)?;
        Ok(B::float_reshape(buffer.float_tensor()?, shape))
    }

    /// Reduce a floating tensor through this rank's configured transport.
    ///
    /// All ranks must submit the same tensor shapes, dtypes and operations in
    /// the same order. The backend reshapes non-contiguous inputs according to
    /// its normal tensor semantics; output shape, dtype and device are retained.
    /// Transfers are host-staged, while arithmetic uses the tensor backend.
    pub fn all_reduce_float(
        &self,
        value: B::FloatTensorPrimitive,
        operation: ReduceOperation,
    ) -> Result<B::FloatTensorPrimitive, TensorDeviceError> {
        match value.dtype() {
            DType::F32 => self.reduce_float::<f32>(value, operation),
            DType::F16 => self.reduce_float::<f16>(value, operation),
            DType::BF16 => self.reduce_float::<bf16>(value, operation),
            dtype => Err(TensorDeviceError::UnsupportedDType(dtype)),
        }
    }

    fn reduce_float<T: TensorElement>(
        &self,
        value: B::FloatTensorPrimitive,
        operation: ReduceOperation,
    ) -> Result<B::FloatTensorPrimitive, TensorDeviceError> {
        let execution = self.execution();
        execution.validate_type::<T>()?;
        if &B::float_device(&value) != execution.device() {
            return Err(TensorDeviceError::DeviceMismatch);
        }
        let shape = value.shape();
        let length = shape.iter().try_fold(1_usize, |length, dim| {
            length.checked_mul(*dim).ok_or(TensorDeviceError::InvalidBuffer(
                "collective tensor element count overflow",
            ))
        })?;
        storage::checked_length::<T>(length)?;
        let value = B::float_reshape(value, Shape::new([length]));
        let buffer = execution.import_float::<T>(value)?;
        self.tensor_collective::<T>().all_reduce(
            &buffer,
            ReductionOperation::Sum,
            &ReductionOperation::Sum,
        )?;
        let mut value = buffer.float_tensor()?;
        if operation == ReduceOperation::Mean {
            value = B::float_div_scalar(value, (self.world_size() as f32).into());
        }
        Ok(B::float_reshape(value, shape))
    }
}

fn sharded_shape(
    shape: Shape,
    world_size: u32,
    scatter: bool,
) -> Result<CollectiveShape, TensorDeviceError> {
    let result = if scatter {
        CollectiveShape::reduce_scatter(shape, world_size as usize)
    } else {
        CollectiveShape::all_gather(shape, world_size as usize)
    };
    result.map_err(|error| TensorDeviceError::Data(format!("{error:?}")))
}