boostr 0.1.0

ML framework built on numr - attention, quantization, model architectures
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

//! MoE token permutation and unpermutation.
//!
//! THE algorithm — same for all backends.
//! Composes numr primitives: argsort, index_select, bincount, cumsum, scatter, cat.

use crate::error::{Error, Result};
use numr::dtype::DType;
use numr::ops::{
    BinaryOps, IndexingOps, ReduceOps, ScalarOps, ShapeOps, SortingOps, TensorOps,
    TypeConversionOps,
};
use numr::runtime::{Runtime, RuntimeClient};
use numr::tensor::Tensor;

/// Permute tokens into expert-grouped order.
///
/// Creates a flat token list where all tokens for expert 0 come first,
/// then expert 1, etc. Tokens assigned to multiple experts (top-k > 1)
/// are duplicated.
///
/// # Returns
/// `(permuted [N*k, hidden], expert_offsets [E+1] I32, sort_indices [N*k] I32)`
pub fn moe_permute_tokens_impl<R, C>(
    client: &C,
    tokens: &Tensor<R>,
    indices: &Tensor<R>,
    num_experts: usize,
) -> Result<(Tensor<R>, Tensor<R>, Tensor<R>)>
where
    R: Runtime<DType = DType>,
    C: RuntimeClient<R>
        + BinaryOps<R>
        + IndexingOps<R>
        + ScalarOps<R>
        + SortingOps<R>
        + ShapeOps<R>
        + TensorOps<R>
        + TypeConversionOps<R>,
{
    let tok_shape = tokens.shape();
    let idx_shape = indices.shape();
    if tok_shape.len() != 2 {
        return Err(Error::InvalidArgument {
            arg: "tokens",
            reason: format!("expected 2D [num_tokens, hidden], got {}D", tok_shape.len()),
        });
    }
    if idx_shape.len() != 2 {
        return Err(Error::InvalidArgument {
            arg: "indices",
            reason: format!("expected 2D [num_tokens, k], got {}D", idx_shape.len()),
        });
    }

    let num_tokens = tok_shape[0];
    let _hidden_dim = tok_shape[1];
    let k = idx_shape[1];
    let total = num_tokens * k;
    let device = tokens.device();

    // Flatten indices to [N*k] — already I32 from routing
    let flat_indices = indices.reshape(&[total]).map_err(Error::Numr)?.contiguous();

    // Sort by expert index to group tokens by expert
    // argsort gives us the permutation that sorts flat_indices (stable sort)
    let sort_perm = client
        .argsort(&flat_indices, 0, false)
        .map_err(Error::Numr)?;

    // Build token indices: for each (token_i, expert_j), the source token is token_i
    // token_source[i*k + j] = i → we need to expand tokens by k
    // Create token source indices [N*k] where entry i*k+j = i
    let token_source_data: Vec<i32> = (0..num_tokens)
        .flat_map(|t| std::iter::repeat_n(t as i32, k))
        .collect();
    let token_source = Tensor::<R>::from_slice(&token_source_data, &[total], device);

    // Reorder token_source by sort_perm to get the token index for each position in sorted order
    let sorted_token_indices = client
        .index_select(&token_source, 0, &sort_perm)
        .map_err(Error::Numr)?;

    // Gather tokens in sorted order: permuted[i] = tokens[sorted_token_indices[i]]
    let permuted = client
        .index_select(tokens, 0, &sorted_token_indices)
        .map_err(Error::Numr)?;

    // Compute expert_offsets via bincount
    let sorted_expert_ids = client
        .index_select(&flat_indices, 0, &sort_perm)
        .map_err(Error::Numr)?;
    let counts = client
        .bincount(&sorted_expert_ids, None, num_experts)
        .map_err(Error::Numr)?;

    // Convert counts to CSR offsets via GPU ops: [0, c0, c0+c1, ..., total]
    // bincount dtype varies by backend (I64 on CPU, I32 on WebGPU) — cast on GPU first.
    let counts_i32 = if counts.dtype() == DType::I32 {
        counts
    } else {
        client.cast(&counts, DType::I32).map_err(Error::Numr)?
    };
    let cumsum = client.cumsum(&counts_i32, 0).map_err(Error::Numr)?;
    // Prepend 0 to form CSR offsets [0, c0, c0+c1, ..., total]
    let zero = Tensor::<R>::zeros(&[1], DType::I32, device);
    let expert_offsets = client.cat(&[&zero, &cumsum], 0).map_err(Error::Numr)?;

    // Cast sort_perm to I32 (argsort returns I64 on CPU, I32 on WebGPU)
    let sort_perm_i32 = if sort_perm.dtype() != DType::I32 {
        client.cast(&sort_perm, DType::I32).map_err(Error::Numr)?
    } else {
        sort_perm
    };

    Ok((permuted, expert_offsets, sort_perm_i32))
}

/// Unpermute expert outputs back to original token order with weighted combination.
///
/// For each original token, gathers its k expert outputs and computes
/// a weighted sum using the routing weights.
pub fn moe_unpermute_tokens_impl<R, C>(
    client: &C,
    expert_output: &Tensor<R>,
    sort_indices: &Tensor<R>,
    weights: &Tensor<R>,
    num_tokens: usize,
) -> Result<Tensor<R>>
where
    R: Runtime<DType = DType>,
    C: RuntimeClient<R> + IndexingOps<R> + ShapeOps<R> + ScalarOps<R> + ReduceOps<R> + TensorOps<R>,
{
    let out_shape = expert_output.shape();
    if out_shape.len() != 2 {
        return Err(Error::InvalidArgument {
            arg: "expert_output",
            reason: format!("expected 2D [N*k, hidden], got {}D", out_shape.len()),
        });
    }

    let total = out_shape[0];
    let hidden_dim = out_shape[1];
    let w_shape = weights.shape();
    if w_shape.len() != 2 {
        return Err(Error::InvalidArgument {
            arg: "weights",
            reason: format!("expected 2D [num_tokens, k], got {}D", w_shape.len()),
        });
    }
    let k = w_shape[1];

    if num_tokens * k != total {
        return Err(Error::InvalidArgument {
            arg: "num_tokens",
            reason: format!(
                "num_tokens({}) * k({}) = {} != expert_output rows({})",
                num_tokens,
                k,
                num_tokens * k,
                total
            ),
        });
    }

    // Compute inverse permutation on GPU: inv_perm[sort_indices[i]] = i
    // Use scatter: dst[index[i]] = src[i] with src = arange(total).
    // arange is a static CPU-generated sequence of known size — not a GPU data download.
    let device = expert_output.device();
    let arange_data: Vec<i32> = (0..total as i32).collect();
    let values = Tensor::<R>::from_slice(&arange_data, &[total], device);
    let inv_perm_base = Tensor::<R>::zeros(&[total], DType::I32, device);
    let inv_perm = client
        .scatter(&inv_perm_base, 0, sort_indices, &values)
        .map_err(Error::Numr)?;

    // Gather back to original order: unsorted[i] = expert_output[inv_perm[i]]
    let unsorted = client
        .index_select(expert_output, 0, &inv_perm)
        .map_err(Error::Numr)?;

    // unsorted is now [N*k, hidden] in original (token, expert) order
    // Reshape to [N, k, hidden]
    let reshaped = unsorted
        .reshape(&[num_tokens, k, hidden_dim])
        .map_err(Error::Numr)?;

    // weights: [N, k] → [N, k, 1] for broadcasting
    let w_expanded = weights.reshape(&[num_tokens, k, 1]).map_err(Error::Numr)?;

    // Weighted sum: sum(reshaped * weights, dim=1) → [N, hidden]
    let weighted = client.mul(&reshaped, &w_expanded).map_err(Error::Numr)?;
    let output = client.sum(&weighted, &[1], false).map_err(Error::Numr)?;

    Ok(output)
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::test_utils::cpu_setup;
    use numr::runtime::cpu::CpuRuntime;

    #[test]
    fn test_permute_unpermute_roundtrip() {
        let (client, device) = cpu_setup();
        let num_tokens = 4;
        let hidden_dim = 8;
        let num_experts = 3;
        let k = 2;

        let tokens_data: Vec<f32> = (0..num_tokens * hidden_dim)
            .map(|i| i as f32 * 0.1)
            .collect();
        let tokens =
            Tensor::<CpuRuntime>::from_slice(&tokens_data, &[num_tokens, hidden_dim], &device);

        // Create expert indices [num_tokens, k]
        let indices_data: Vec<i32> = vec![0, 1, 2, 0, 1, 2, 0, 1];
        let indices = Tensor::<CpuRuntime>::from_slice(&indices_data, &[num_tokens, k], &device);

        // Equal weights
        let weights_data: Vec<f32> = vec![0.5; num_tokens * k];
        let weights = Tensor::<CpuRuntime>::from_slice(&weights_data, &[num_tokens, k], &device);

        let (permuted, offsets, sort_indices) =
            moe_permute_tokens_impl(&client, &tokens, &indices, num_experts).unwrap();

        assert_eq!(permuted.shape(), &[num_tokens * k, hidden_dim]);
        assert_eq!(offsets.shape(), &[num_experts + 1]);

        // Unpermute with identity-like weights should recover tokens
        // (each token appears k times with weight 1/k, so weighted sum = original)
        let result =
            moe_unpermute_tokens_impl(&client, &permuted, &sort_indices, &weights, num_tokens)
                .unwrap();

        assert_eq!(result.shape(), &[num_tokens, hidden_dim]);

        let result_vec = result.to_vec::<f32>();
        for (i, (&got, &expected)) in result_vec.iter().zip(tokens_data.iter()).enumerate() {
            assert!(
                (got - expected).abs() < 1e-5,
                "roundtrip mismatch at {}: got {}, expected {}",
                i,
                got,
                expected
            );
        }
    }
}