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
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

//! Tensor-parallel LLaMA block, attention, and MLP sub-modules.

use crate::distributed::tensor_parallel::{ColumnParallelLinear, RowParallelLinear};
use crate::error::{Error, Result};
use crate::inference::kv_cache::KvCache;
use crate::model::traits::ModelClient;
use crate::nn::{RmsNorm, RoPE};
use crate::ops::impl_generic::attention::multi_head_attention_impl;
use crate::ops::impl_generic::attention::rope::apply_rope_interleaved_impl;
use numr::autograd::{Var, var_add, var_mul, var_narrow, var_reshape, var_silu};
use numr::dtype::DType;
use numr::ops::{
    ActivationOps, BinaryOps, CompareOps, ConditionalOps, IndexingOps, ReduceOps, ScalarOps,
    ShapeOps, TensorOps, UnaryOps,
};
use numr::runtime::Runtime;

pub(super) struct LlamaBlockTp<R: Runtime> {
    pub(super) input_layernorm: RmsNorm<R>,
    pub(super) self_attn: LlamaAttentionTp<R>,
    pub(super) post_attention_layernorm: RmsNorm<R>,
    pub(super) mlp: LlamaMlpTp<R>,
}

pub(super) struct LlamaAttentionTp<R: Runtime> {
    pub(super) q_proj: ColumnParallelLinear<R>,
    pub(super) k_proj: ColumnParallelLinear<R>,
    pub(super) v_proj: ColumnParallelLinear<R>,
    pub(super) o_proj: RowParallelLinear<R>,
    pub(super) num_heads: usize,    // LOCAL heads (total / world_size)
    pub(super) num_kv_heads: usize, // LOCAL kv heads (total / world_size)
    pub(super) head_dim: usize,
    /// Optional Q/K layer norms (Command-R, Cohere)
    pub(super) q_norm: Option<RmsNorm<R>>,
    pub(super) k_norm: Option<RmsNorm<R>>,
}

pub(super) struct LlamaMlpTp<R: Runtime> {
    pub(super) gate_proj: ColumnParallelLinear<R>,
    pub(super) up_proj: ColumnParallelLinear<R>,
    pub(super) down_proj: RowParallelLinear<R>,
}

// ── LlamaBlockTp ────────────────────────────────────────────────────

impl<R: Runtime<DType = DType>> LlamaBlockTp<R> {
    pub(super) fn forward<C>(&self, client: &C, x: &Var<R>, rope: &RoPE<R>) -> Result<Var<R>>
    where
        C: ModelClient<R>,
        R::Client: TensorOps<R>
            + ScalarOps<R>
            + ReduceOps<R>
            + IndexingOps<R>
            + ShapeOps<R>
            + ActivationOps<R>
            + BinaryOps<R>
            + UnaryOps<R>
            + CompareOps<R>
            + ConditionalOps<R>,
    {
        let normed = self.input_layernorm.forward(client, x)?;
        let attn_out = self.self_attn.forward(client, &normed, rope)?;
        let h = var_add(x, &attn_out, client).map_err(Error::Numr)?;

        let normed = self.post_attention_layernorm.forward(client, &h)?;
        let mlp_out = self.mlp.forward(client, &normed)?;
        var_add(&h, &mlp_out, client).map_err(Error::Numr)
    }

    /// Forward pass with KV cache for autoregressive inference.
    pub(super) fn forward_with_kv_cache<C>(
        &self,
        client: &C,
        x: &Var<R>,
        rope: &RoPE<R>,
        kv_cache: &mut KvCache<R>,
        position: usize,
    ) -> Result<Var<R>>
    where
        C: ModelClient<R>,
        R::Client: TensorOps<R>
            + ScalarOps<R>
            + ReduceOps<R>
            + IndexingOps<R>
            + ShapeOps<R>
            + ActivationOps<R>
            + BinaryOps<R>
            + UnaryOps<R>
            + CompareOps<R>
            + ConditionalOps<R>,
    {
        let normed = self.input_layernorm.forward(client, x)?;
        let attn_out = self
            .self_attn
            .forward_with_kv_cache(client, &normed, rope, kv_cache, position)?;
        let h = var_add(x, &attn_out, client).map_err(Error::Numr)?;

        let normed = self.post_attention_layernorm.forward(client, &h)?;
        let mlp_out = self.mlp.forward(client, &normed)?;
        var_add(&h, &mlp_out, client).map_err(Error::Numr)
    }
}

// ── LlamaAttentionTp ────────────────────────────────────────────────

impl<R: Runtime<DType = DType>> LlamaAttentionTp<R> {
    pub(super) fn forward<C>(&self, client: &C, x: &Var<R>, rope: &RoPE<R>) -> Result<Var<R>>
    where
        C: ModelClient<R>,
        R::Client: TensorOps<R>
            + ScalarOps<R>
            + ReduceOps<R>
            + IndexingOps<R>
            + ShapeOps<R>
            + ActivationOps<R>
            + BinaryOps<R>
            + UnaryOps<R>
            + CompareOps<R>
            + ConditionalOps<R>,
    {
        let shape = x.shape().to_vec();
        let batch = shape[0];
        let seq_len = shape[1];

        // Q/K/V: column-parallel, each rank gets local heads
        let q = self.q_proj.forward(client, x)?;
        let k = self.k_proj.forward(client, x)?;
        let v = self.v_proj.forward(client, x)?;

        // Reshape to [B, S, local_H, D] then permute to [B, local_H, S, D]
        let q = numr::autograd::var_reshape(&q, &[batch, seq_len, self.num_heads, self.head_dim])
            .map_err(Error::Numr)?;
        let k =
            numr::autograd::var_reshape(&k, &[batch, seq_len, self.num_kv_heads, self.head_dim])
                .map_err(Error::Numr)?;
        let v =
            numr::autograd::var_reshape(&v, &[batch, seq_len, self.num_kv_heads, self.head_dim])
                .map_err(Error::Numr)?;

        let q = numr::autograd::var_permute(&q, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let q = var_contiguous(&q);
        let k = numr::autograd::var_permute(&k, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let k = var_contiguous(&k);
        let v = numr::autograd::var_permute(&v, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let v = var_contiguous(&v);

        // Optional Q/K layer norms (Command-R, Cohere) — applied before RoPE
        let q = match &self.q_norm {
            Some(norm) => norm.forward(client, &q)?,
            None => q,
        };
        let k = match &self.k_norm {
            Some(norm) => norm.forward(client, &k)?,
            None => k,
        };

        // RoPE on Q and K
        let q = apply_rope_interleaved_impl(client, &q, rope.cos_cache(), rope.sin_cache())?;
        let k = apply_rope_interleaved_impl(client, &k, rope.cos_cache(), rope.sin_cache())?;

        // GQA repeat KV if needed (local heads)
        let (k, v) = if self.num_kv_heads < self.num_heads {
            let repeat = self.num_heads / self.num_kv_heads;
            let k = repeat_kv(&k, repeat).map_err(Error::Numr)?;
            let v = repeat_kv(&v, repeat).map_err(Error::Numr)?;
            (k, v)
        } else {
            (k, v)
        };

        // Local attention (no communication)
        let attn_out = multi_head_attention_impl(client, &q, &k, &v, None, self.num_heads)?;

        // [B, H, S, D] -> [B, S, H, D] -> [B, S, H*D]
        let attn_out =
            numr::autograd::var_permute(&attn_out, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let attn_out = var_contiguous(&attn_out);
        let attn_out = numr::autograd::var_reshape(
            &attn_out,
            &[batch, seq_len, self.num_heads * self.head_dim],
        )
        .map_err(Error::Numr)?;

        // O projection (row-parallel → all-reduce)
        self.o_proj.forward(client, &attn_out)
    }

    /// Forward with KV cache for autoregressive inference.
    pub(super) fn forward_with_kv_cache<C>(
        &self,
        client: &C,
        x: &Var<R>,
        rope: &RoPE<R>,
        kv_cache: &mut KvCache<R>,
        position: usize,
    ) -> Result<Var<R>>
    where
        C: ModelClient<R>,
        R::Client: TensorOps<R>
            + ScalarOps<R>
            + ReduceOps<R>
            + IndexingOps<R>
            + ShapeOps<R>
            + ActivationOps<R>
            + BinaryOps<R>
            + UnaryOps<R>
            + CompareOps<R>
            + ConditionalOps<R>,
    {
        let shape = x.shape().to_vec();
        let batch = shape[0];
        let seq_len = shape[1];

        // Q/K/V: column-parallel, each rank gets local heads
        let q = self.q_proj.forward(client, x)?;
        let k = self.k_proj.forward(client, x)?;
        let v = self.v_proj.forward(client, x)?;

        // Reshape to [B, S, local_H, D] then permute to [B, local_H, S, D]
        let q = var_reshape(&q, &[batch, seq_len, self.num_heads, self.head_dim])
            .map_err(Error::Numr)?;
        let k = var_reshape(&k, &[batch, seq_len, self.num_kv_heads, self.head_dim])
            .map_err(Error::Numr)?;
        let v = var_reshape(&v, &[batch, seq_len, self.num_kv_heads, self.head_dim])
            .map_err(Error::Numr)?;

        let q = numr::autograd::var_permute(&q, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let q = var_contiguous(&q);
        let k = numr::autograd::var_permute(&k, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let k = var_contiguous(&k);
        let v = numr::autograd::var_permute(&v, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let v = var_contiguous(&v);

        // Optional Q/K layer norms (Command-R, Cohere) — applied before RoPE
        let q = match &self.q_norm {
            Some(norm) => norm.forward(client, &q)?,
            None => q,
        };
        let k = match &self.k_norm {
            Some(norm) => norm.forward(client, &k)?,
            None => k,
        };

        // Apply RoPE with position offset
        let cos_offset = var_narrow(rope.cos_cache(), 0, position, seq_len).map_err(Error::Numr)?;
        let sin_offset = var_narrow(rope.sin_cache(), 0, position, seq_len).map_err(Error::Numr)?;
        let q = apply_rope_interleaved_impl(client, &q, &cos_offset, &sin_offset)?;
        let k = apply_rope_interleaved_impl(client, &k, &cos_offset, &sin_offset)?;

        // Update KV cache
        kv_cache.update_fused(k.tensor(), v.tensor(), client)?;

        // Attention using full KV cache
        let kv_seq_len = kv_cache.seq_len();
        let is_prefill = seq_len > 1;
        let (attn_out, _lse) = client.flash_attention_fwd(
            q.tensor(),
            kv_cache.k_cache_raw(),
            kv_cache.v_cache_raw(),
            self.num_heads,
            self.num_kv_heads,
            self.head_dim,
            is_prefill,
            0,
            Some(kv_seq_len),
        )?;

        // [B, H, S, D] -> [B, S, H, D] -> [B, S, H*D]
        let attn_out = Var::new(attn_out, false);
        let attn_out =
            numr::autograd::var_permute(&attn_out, &[0, 2, 1, 3]).map_err(Error::Numr)?;
        let attn_out = var_contiguous(&attn_out);
        let attn_out = var_reshape(&attn_out, &[batch, seq_len, self.num_heads * self.head_dim])
            .map_err(Error::Numr)?;

        // O projection (row-parallel → all-reduce)
        self.o_proj.forward(client, &attn_out)
    }
}

fn var_contiguous<R: Runtime>(v: &Var<R>) -> Var<R> {
    Var::new(v.tensor().contiguous(), v.requires_grad())
}

fn repeat_kv<R: Runtime>(x: &Var<R>, repeat: usize) -> numr::error::Result<Var<R>> {
    if repeat == 1 {
        return Ok(x.clone());
    }
    let shape = x.shape();
    let [b, h_kv, s, d] = [shape[0], shape[1], shape[2], shape[3]];
    let expanded = x.tensor().reshape(&[b, h_kv, 1, s, d])?;
    let expanded = expanded.broadcast_to(&[b, h_kv, repeat, s, d])?;
    let result = expanded.contiguous().reshape(&[b, h_kv * repeat, s, d])?;
    Ok(Var::new(result, x.requires_grad()))
}

// ── LlamaMlpTp ──────────────────────────────────────────────────────

impl<R: Runtime<DType = DType>> LlamaMlpTp<R> {
    pub(super) fn forward<C>(&self, client: &C, x: &Var<R>) -> Result<Var<R>>
    where
        C: ModelClient<R>,
        R::Client: TensorOps<R>
            + ScalarOps<R>
            + ReduceOps<R>
            + IndexingOps<R>
            + ShapeOps<R>
            + ActivationOps<R>
            + BinaryOps<R>
            + UnaryOps<R>
            + CompareOps<R>
            + ConditionalOps<R>,
    {
        let gate = self.gate_proj.forward(client, x)?;
        let up = self.up_proj.forward(client, x)?;
        let gate_silu = var_silu(&gate, client).map_err(Error::Numr)?;
        let hidden = var_mul(&gate_silu, &up, client).map_err(Error::Numr)?;
        // down_proj is row-parallel → all-reduce
        self.down_proj.forward(client, &hidden)
    }
}