realizar 0.8.4

Pure Rust ML inference engine built from scratch - model serving for GGUF and safetensors
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
323
324

/// Apply RMS norm in-place from `src` into `dst`, weighted by `gamma`.
fn rms_norm_weighted(src: &[f32], gamma: &[f32], dst: &mut [f32], eps: f32) {
    let n = src.len();
    let sq_sum: f32 = src.iter().map(|x| x * x).sum();
    let rms = (sq_sum / n as f32 + eps).sqrt();
    for i in 0..n {
        dst[i] = src[i] / rms * gamma[i];
    }
}

/// Apply RMS norm to a multi-token hidden state, returning normalized output.
fn rms_norm_batched(hidden: &[f32], gamma: &[f32], hidden_dim: usize, eps: f32) -> Vec<f32> {
    let seq_len = hidden.len() / hidden_dim;
    let mut out = Vec::with_capacity(hidden.len());
    for s in 0..seq_len {
        let slice = &hidden[s * hidden_dim..(s + 1) * hidden_dim];
        let sq_sum: f32 = slice.iter().map(|x| x * x).sum();
        let rms = (sq_sum / hidden_dim as f32 + eps).sqrt();
        for (i, &x) in slice.iter().enumerate() {
            out.push(x / rms * gamma[i]);
        }
    }
    out
}

/// Apply SwiGLU activation: silu(gate) * up, in-place on `up`.
fn apply_swiglu(up: &mut [f32], gate: &[f32]) {
    for i in 0..up.len() {
        let silu = gate[i] / (1.0 + (-gate[i]).exp());
        up[i] *= silu;
    }
}

/// Apply GELU activation in-place.
fn apply_gelu(data: &mut [f32]) {
    const SQRT_2_OVER_PI: f32 = 0.797_884_6;
    const GELU_COEFF: f32 = 0.044_715;
    for x in data.iter_mut() {
        let t = (SQRT_2_OVER_PI * (*x + GELU_COEFF * *x * *x * *x)).tanh();
        *x = 0.5 * *x * (1.0 + t);
    }
}

/// Embed token IDs into hidden states.
fn embed_tokens(token_ids: &[u32], embedding: &[f32], hidden_dim: usize) -> Vec<f32> {
    let mut hidden = Vec::with_capacity(token_ids.len() * hidden_dim);
    for &token_id in token_ids {
        let offset = (token_id as usize) * hidden_dim;
        if offset + hidden_dim <= embedding.len() {
            hidden.extend_from_slice(&embedding[offset..offset + hidden_dim]);
        } else {
            hidden.extend(std::iter::repeat_n(0.0, hidden_dim));
        }
    }
    hidden
}

/// Add `src` elementwise into `dst`.
fn residual_add(dst: &mut [f32], src: &[f32]) {
    for i in 0..dst.len() {
        dst[i] += src[i];
    }
}

impl QuantizedAprTransformerQ4 {
    /// Forward pass for a single token using scratch buffer (zero allocation).
    pub fn forward_single_with_scratch(
        &self,
        token_id: u32,
        scratch: &mut AprInferenceScratch,
    ) -> Result<Vec<f32>> {
        use crate::quantize::fused_q4_0_q8_0_parallel_matvec_into;

        let hidden_dim = self.config.hidden_dim;
        let num_heads = self.config.num_heads;
        let num_kv_heads = self.config.num_kv_heads;
        let head_dim = hidden_dim / num_heads;
        let eps = self.config.eps;

        // 1. Token embedding lookup
        let offset = (token_id as usize) * hidden_dim;
        if offset + hidden_dim <= self.token_embedding.len() {
            scratch.hidden[..hidden_dim]
                .copy_from_slice(&self.token_embedding[offset..offset + hidden_dim]);
        } else {
            scratch.hidden[..hidden_dim].fill(0.0);
        }

        // 2. Process through transformer layers
        for layer in &self.layers {
            rms_norm_weighted(&scratch.hidden, &layer.attn_norm_weight, &mut scratch.normed, eps);

            let qkv_dim = layer.qkv_weight.out_dim;
            fused_q4_0_q8_0_parallel_matvec_into(
                &layer.qkv_weight.data, &scratch.normed[..hidden_dim],
                hidden_dim, &mut scratch.qkv_out[..qkv_dim],
            )?;

            let q_dim = num_heads * head_dim;
            let kv_dim = num_kv_heads * head_dim;
            scratch.q[..q_dim].copy_from_slice(&scratch.qkv_out[..q_dim]);
            scratch.k[..kv_dim].copy_from_slice(&scratch.qkv_out[q_dim..q_dim + kv_dim]);
            scratch.v[..kv_dim]
                .copy_from_slice(&scratch.qkv_out[q_dim + kv_dim..q_dim + 2 * kv_dim]);

            self.apply_rope(&mut scratch.q[..q_dim], 0, num_heads);
            self.apply_rope(&mut scratch.k[..kv_dim], 0, num_kv_heads);

            // Single token: attention output = V (softmax of 1 element = 1.0)
            let group_size = num_heads / num_kv_heads;
            for head in 0..num_heads {
                let kv_head = head / group_size;
                scratch.attn_out[head * head_dim..(head + 1) * head_dim]
                    .copy_from_slice(&scratch.v[kv_head * head_dim..(kv_head + 1) * head_dim]);
            }

            fused_q4_0_q8_0_parallel_matvec_into(
                &layer.attn_output_weight.data, &scratch.attn_out[..hidden_dim],
                layer.attn_output_weight.in_dim,
                &mut scratch.ffn_out[..layer.attn_output_weight.out_dim],
            )?;
            residual_add(&mut scratch.hidden[..hidden_dim], &scratch.ffn_out[..hidden_dim]);

            // Pre-FFN norm
            if let Some(ffn_norm) = &layer.ffn_norm_weight {
                rms_norm_weighted(&scratch.hidden, ffn_norm, &mut scratch.ffn_input, eps);
            } else {
                scratch.ffn_input[..hidden_dim].copy_from_slice(&scratch.normed[..hidden_dim]);
            }

            // FFN
            let intermediate_dim = layer.ffn_up_weight.out_dim;
            fused_q4_0_q8_0_parallel_matvec_into(
                &layer.ffn_up_weight.data, &scratch.ffn_input[..hidden_dim],
                hidden_dim, &mut scratch.ffn_up[..intermediate_dim],
            )?;
            if let Some(gate) = &layer.ffn_gate_weight {
                fused_q4_0_q8_0_parallel_matvec_into(
                    &gate.data, &scratch.ffn_input[..hidden_dim],
                    hidden_dim, &mut scratch.ffn_gate[..intermediate_dim],
                )?;
                apply_swiglu(&mut scratch.ffn_up[..intermediate_dim], &scratch.ffn_gate[..intermediate_dim]);
            } else {
                apply_gelu(&mut scratch.ffn_up[..intermediate_dim]);
            }

            fused_q4_0_q8_0_parallel_matvec_into(
                &layer.ffn_down_weight.data, &scratch.ffn_up[..intermediate_dim],
                intermediate_dim, &mut scratch.ffn_out[..hidden_dim],
            )?;
            residual_add(&mut scratch.hidden[..hidden_dim], &scratch.ffn_out[..hidden_dim]);
        }

        // 3. Final RMS norm
        rms_norm_weighted(&scratch.hidden, &self.output_norm_weight, &mut scratch.normed, eps);

        // 4. LM head projection
        let vocab_size = self.config.vocab_size;
        let mut logits = vec![0.0f32; vocab_size];
        fused_q4_0_q8_0_parallel_matvec_into(
            &self.lm_head_weight.data, &scratch.normed[..hidden_dim],
            hidden_dim, &mut logits,
        )?;

        Ok(logits)
    }

    /// Forward pass with KV cache for efficient autoregressive generation.
    pub fn forward_with_cache(
        &self,
        token_ids: &[u32],
        cache: &mut AprKVCache,
    ) -> Result<Vec<f32>> {
        if token_ids.is_empty() {
            return Err(RealizarError::InvalidShape {
                reason: "Token sequence cannot be empty".to_string(),
            });
        }

        let hidden_dim = self.config.hidden_dim;
        let eps = self.config.eps;
        let cache_len = cache.len();
        let new_seq_len = token_ids.len();

        let mut hidden = embed_tokens(token_ids, &self.token_embedding, hidden_dim);

        for (layer_idx, layer) in self.layers.iter().enumerate() {
            let normed = rms_norm_batched(&hidden, &layer.attn_norm_weight, hidden_dim, eps);

            let attn_out = self.cached_layer_attention(
                &normed, layer, layer_idx, cache, cache_len, new_seq_len,
            )?;

            let proj_out = self.batched_matvec(&layer.attn_output_weight, &attn_out, new_seq_len)?;
            residual_add(&mut hidden, &proj_out);

            let ffn_input = match &layer.ffn_norm_weight {
                Some(ffn_norm) => rms_norm_batched(&hidden, ffn_norm, hidden_dim, eps),
                None => normed.clone(),
            };

            let ffn_out = self.cached_layer_ffn(layer, &ffn_input, new_seq_len)?;
            residual_add(&mut hidden, &ffn_out);
        }

        // Final norm + LM head (last token only)
        let last_start = (new_seq_len - 1) * hidden_dim;
        let last_hidden = &hidden[last_start..last_start + hidden_dim];
        let mut normed_final = vec![0.0f32; hidden_dim];
        rms_norm_weighted(last_hidden, &self.output_norm_weight, &mut normed_final, eps);

        use crate::quantize::fused_q4_0_q8_0_parallel_matvec;
        fused_q4_0_q8_0_parallel_matvec(
            &self.lm_head_weight.data, &normed_final, hidden_dim, self.config.vocab_size,
        )
    }

    /// Attention sublayer for cached forward pass.
    fn cached_layer_attention(
        &self,
        normed: &[f32],
        layer: &QuantizedAprLayerQ4,
        layer_idx: usize,
        cache: &mut AprKVCache,
        cache_len: usize,
        new_seq_len: usize,
    ) -> Result<Vec<f32>> {
        use crate::quantize::fused_q4_0_q8_0_parallel_matvec;

        let hidden_dim = self.config.hidden_dim;
        let num_heads = self.config.num_heads;
        let num_kv_heads = self.config.num_kv_heads;
        let head_dim = hidden_dim / num_heads;
        let qkv_dim = layer.qkv_weight.out_dim;
        let q_dim = num_heads * head_dim;
        let kv_dim = num_kv_heads * head_dim;

        let mut qkv_out = Vec::with_capacity(new_seq_len * qkv_dim);
        for s in 0..new_seq_len {
            let input = &normed[s * hidden_dim..(s + 1) * hidden_dim];
            let qkv = fused_q4_0_q8_0_parallel_matvec(
                &layer.qkv_weight.data, input, hidden_dim, qkv_dim,
            )?;
            qkv_out.extend(qkv);
        }

        let mut new_q = Vec::with_capacity(new_seq_len * q_dim);
        for s in 0..new_seq_len {
            let base = s * qkv_dim;
            let position = cache_len + s;
            let mut q = qkv_out[base..base + q_dim].to_vec();
            let mut k = qkv_out[base + q_dim..base + q_dim + kv_dim].to_vec();
            let v = qkv_out[base + q_dim + kv_dim..base + q_dim + 2 * kv_dim].to_vec();
            self.apply_rope(&mut q, position, num_heads);
            self.apply_rope(&mut k, position, num_kv_heads);
            new_q.extend_from_slice(&q);
            cache.append(layer_idx, &k, &v);
        }

        let (full_k, full_v) = cache.get(layer_idx);
        let total_seq_len = cache.len();
        Ok(self.causal_attention_cached(&new_q, full_k, full_v, new_seq_len, total_seq_len, cache_len))
    }

    /// FFN sublayer for cached forward pass (SwiGLU or GELU).
    fn cached_layer_ffn(
        &self,
        layer: &QuantizedAprLayerQ4,
        ffn_input: &[f32],
        new_seq_len: usize,
    ) -> Result<Vec<f32>> {
        let hidden_dim = self.config.hidden_dim;
        let intermediate_dim = layer.ffn_up_weight.out_dim;

        let mut up = self.batched_matvec(&layer.ffn_up_weight, ffn_input, new_seq_len)?;

        if let Some(gate) = &layer.ffn_gate_weight {
            let gate_out = self.batched_matvec(gate, ffn_input, new_seq_len)?;
            apply_swiglu(&mut up, &gate_out);
        } else {
            apply_gelu(&mut up);
        }

        self.batched_matvec_custom(&layer.ffn_down_weight, &up, new_seq_len, intermediate_dim, hidden_dim)
    }

    /// Batched matvec: multiply each token's hidden state by a quantized weight matrix.
    fn batched_matvec(
        &self,
        weight: &QuantizedAprTensorQ4,
        input: &[f32],
        seq_len: usize,
    ) -> Result<Vec<f32>> {
        use crate::quantize::fused_q4_0_q8_0_parallel_matvec;
        let in_dim = weight.in_dim;
        let out_dim = weight.out_dim;
        let mut out = Vec::with_capacity(seq_len * out_dim);
        for s in 0..seq_len {
            let slice = &input[s * in_dim..(s + 1) * in_dim];
            let row = fused_q4_0_q8_0_parallel_matvec(&weight.data, slice, in_dim, out_dim)?;
            out.extend(row);
        }
        Ok(out)
    }

    /// Batched matvec with explicit dimensions (for down projection where dims differ).
    fn batched_matvec_custom(
        &self,
        weight: &QuantizedAprTensorQ4,
        input: &[f32],
        seq_len: usize,
        in_dim: usize,
        out_dim: usize,
    ) -> Result<Vec<f32>> {
        use crate::quantize::fused_q4_0_q8_0_parallel_matvec;
        let mut out = Vec::with_capacity(seq_len * out_dim);
        for s in 0..seq_len {
            let slice = &input[s * in_dim..(s + 1) * in_dim];
            let row = fused_q4_0_q8_0_parallel_matvec(&weight.data, slice, in_dim, out_dim)?;
            out.extend(row);
        }
        Ok(out)
    }
}