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


// =============================================================================
// Q8K-optimized path (requires hidden_dim % 256 == 0)
// =============================================================================

/// Create a llama model with hidden_dim=256 to trigger Q8K paths
fn create_llama_256_model() -> crate::gguf::OwnedQuantizedModel {
    use crate::gguf::{OwnedQKVWeights, OwnedQuantizedLayer, OwnedQuantizedModel};

    let config = GGUFConfig {
        architecture: "llama".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("llama"),
        hidden_dim: 256,
        intermediate_dim: 512,
        num_heads: 4,
        num_kv_heads: 4,
        num_layers: 1,
        vocab_size: 100,
        context_length: 512,
        rope_theta: 10000.0,
        eps: 1e-5,
        rope_type: 0,
            explicit_head_dim: None,
        bos_token_id: None,
            eos_token_id: None,
    };
    let hidden_dim = config.hidden_dim;
    let intermediate_dim = config.intermediate_dim;
    let kv_dim = config.num_kv_heads * (hidden_dim / config.num_heads);
    let qkv_out_dim = hidden_dim + 2 * kv_dim;

    let layer = OwnedQuantizedLayer {
        attn_norm_weight: vec![1.0f32; hidden_dim],
        attn_norm_bias: None,
        qkv_weight: OwnedQKVWeights::Fused(create_q4k_test_data(hidden_dim, qkv_out_dim)),
        qkv_bias: None,
        attn_output_weight: create_q4k_test_data(hidden_dim, hidden_dim),
        attn_output_bias: None,
        ffn_up_weight: create_q4k_test_data(hidden_dim, intermediate_dim),
        ffn_up_bias: None,
        ffn_down_weight: create_q4k_test_data(intermediate_dim, hidden_dim),
        ffn_down_bias: None,
        ffn_gate_weight: Some(create_q4k_test_data(hidden_dim, intermediate_dim)),
        ffn_gate_bias: None,
        ffn_norm_weight: Some(vec![1.0f32; hidden_dim]),
        ffn_norm_bias: None,
        attn_q_norm_weight: None,
        attn_k_norm_weight: None,
    };

    OwnedQuantizedModel {
        config: config.clone(),
        token_embedding: vec![0.1f32; config.vocab_size * hidden_dim],
        position_embedding: None,
        layers: vec![layer],
        encoder_layers: vec![],
        encoder_output_norm_weight: None,
        encoder_output_norm_bias: None,
        output_norm_weight: vec![1.0f32; hidden_dim],
        output_norm_bias: None,
        lm_head_weight: create_q4k_test_data(hidden_dim, config.vocab_size),
        lm_head_bias: None,
        #[cfg(feature = "cuda")]
        cuda_executor: None,
        #[cfg(feature = "cuda")]
        cuda_kernel_count: std::sync::atomic::AtomicU64::new(0),
        #[cfg(feature = "cuda")]
        cached_weight_names: std::sync::Mutex::new(std::collections::HashSet::new()),
    }
}

#[test]
fn test_forward_single_with_scratch_q8k_path() {
    // hidden_dim=256 => Q8K-optimized path enabled
    let model = create_llama_256_model();
    let config = &model.config;
    let mut cache = OwnedQuantizedKVCache::from_config(config, 128);
    let mut scratch = InferenceScratchBuffer::from_config(config);

    let r1 = model.forward_single_with_scratch(5, &mut cache, 0, &mut scratch);
    assert!(r1.is_ok(), "Q8K scratch: {}", r1.unwrap_err());
    assert!(scratch.logits.len() == config.vocab_size);

    let r2 = model.forward_single_with_scratch(10, &mut cache, 1, &mut scratch);
    assert!(r2.is_ok(), "Q8K scratch 2nd: {}", r2.unwrap_err());
}

#[test]
fn test_forward_single_with_cache_q8k_path() {
    // Q8K path also used in forward_single_with_cache when hidden_dim % 256 == 0
    let model = create_llama_256_model();
    let config = &model.config;
    let mut cache = OwnedQuantizedKVCache::from_config(config, 128);

    let r1 = model.forward_single_with_cache(5, &mut cache, 0);
    assert!(r1.is_ok(), "Q8K cache: {}", r1.unwrap_err());
    assert_eq!(r1.expect("r1").len(), config.vocab_size);

    let r2 = model.forward_single_with_cache(10, &mut cache, 1);
    assert!(r2.is_ok(), "Q8K cache 2nd: {}", r2.unwrap_err());
}

/// Create a phi-style model with hidden_dim=256 to trigger Q8K GELU paths
fn create_phi_256_model() -> crate::gguf::OwnedQuantizedModel {
    use crate::gguf::{OwnedQKVWeights, OwnedQuantizedLayer, OwnedQuantizedModel};

    let config = GGUFConfig {
        architecture: "phi2".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("phi2"),
        hidden_dim: 256,
        intermediate_dim: 512,
        num_heads: 4,
        num_kv_heads: 4,
        num_layers: 1,
        vocab_size: 100,
        context_length: 512,
        rope_theta: 10000.0,
        eps: 1e-5,
        rope_type: 0,
            explicit_head_dim: None,
        bos_token_id: None,
            eos_token_id: None,
    };
    let hidden_dim = config.hidden_dim;
    let intermediate_dim = config.intermediate_dim;
    let kv_dim = config.num_kv_heads * (hidden_dim / config.num_heads);
    let qkv_out_dim = hidden_dim + 2 * kv_dim;

    let layer = OwnedQuantizedLayer {
        attn_norm_weight: vec![1.0f32; hidden_dim],
        attn_norm_bias: Some(vec![0.0f32; hidden_dim]), // LayerNorm
        qkv_weight: OwnedQKVWeights::Fused(create_q4k_test_data(hidden_dim, qkv_out_dim)),
        qkv_bias: Some(vec![0.0f32; qkv_out_dim]),
        attn_output_weight: create_q4k_test_data(hidden_dim, hidden_dim),
        attn_output_bias: Some(vec![0.0f32; hidden_dim]),
        ffn_up_weight: create_q4k_test_data(hidden_dim, intermediate_dim),
        ffn_up_bias: Some(vec![0.0f32; intermediate_dim]),
        ffn_down_weight: create_q4k_test_data(intermediate_dim, hidden_dim),
        ffn_down_bias: Some(vec![0.0f32; hidden_dim]),
        ffn_gate_weight: None, // GELU, no gate
        ffn_gate_bias: None,
        ffn_norm_weight: Some(vec![1.0f32; hidden_dim]),
        ffn_norm_bias: Some(vec![0.0f32; hidden_dim]),
        attn_q_norm_weight: None,
        attn_k_norm_weight: None,
    };

    OwnedQuantizedModel {
        config: config.clone(),
        token_embedding: vec![0.1f32; config.vocab_size * hidden_dim],
        position_embedding: None,
        layers: vec![layer],
        encoder_layers: vec![],
        encoder_output_norm_weight: None,
        encoder_output_norm_bias: None,
        output_norm_weight: vec![1.0f32; hidden_dim],
        output_norm_bias: Some(vec![0.0f32; hidden_dim]),
        lm_head_weight: create_q4k_test_data(hidden_dim, config.vocab_size),
        lm_head_bias: Some(vec![0.0f32; config.vocab_size]),
        #[cfg(feature = "cuda")]
        cuda_executor: None,
        #[cfg(feature = "cuda")]
        cuda_kernel_count: std::sync::atomic::AtomicU64::new(0),
        #[cfg(feature = "cuda")]
        cached_weight_names: std::sync::Mutex::new(std::collections::HashSet::new()),
    }
}

/// Create a GQA llama model with hidden_dim=256 (num_kv_heads < num_heads)
fn create_llama_256_gqa_model() -> crate::gguf::OwnedQuantizedModel {
    use crate::gguf::{OwnedQKVWeights, OwnedQuantizedLayer, OwnedQuantizedModel};

    let config = GGUFConfig {
        architecture: "llama".to_string(),
        constraints: crate::gguf::ArchConstraints::from_architecture("llama"),
        hidden_dim: 256,
        intermediate_dim: 512,
        num_heads: 8,    // 8 query heads
        num_kv_heads: 2, // 2 KV heads => GQA group_size=4
        num_layers: 1,
        vocab_size: 100,
        context_length: 512,
        rope_theta: 10000.0,
        eps: 1e-5,
        rope_type: 0,
            explicit_head_dim: None,
        bos_token_id: None,
            eos_token_id: None,
    };
    let hidden_dim = config.hidden_dim;
    let intermediate_dim = config.intermediate_dim;
    let head_dim = hidden_dim / config.num_heads; // 32
    let kv_dim = config.num_kv_heads * head_dim; // 64
    let qkv_out_dim = hidden_dim + 2 * kv_dim; // 256 + 128 = 384

    let layer = OwnedQuantizedLayer {
        attn_norm_weight: vec![1.0f32; hidden_dim],
        attn_norm_bias: None,
        qkv_weight: OwnedQKVWeights::Fused(create_q4k_test_data(hidden_dim, qkv_out_dim)),
        qkv_bias: None,
        attn_output_weight: create_q4k_test_data(hidden_dim, hidden_dim),
        attn_output_bias: None,
        ffn_up_weight: create_q4k_test_data(hidden_dim, intermediate_dim),
        ffn_up_bias: None,
        ffn_down_weight: create_q4k_test_data(intermediate_dim, hidden_dim),
        ffn_down_bias: None,
        ffn_gate_weight: Some(create_q4k_test_data(hidden_dim, intermediate_dim)),
        ffn_gate_bias: None,
        ffn_norm_weight: Some(vec![1.0f32; hidden_dim]),
        ffn_norm_bias: None,
        attn_q_norm_weight: None,
        attn_k_norm_weight: None,
    };

    OwnedQuantizedModel {
        config: config.clone(),
        token_embedding: vec![0.1f32; config.vocab_size * hidden_dim],
        position_embedding: None,
        layers: vec![layer],
        encoder_layers: vec![],
        encoder_output_norm_weight: None,
        encoder_output_norm_bias: None,
        output_norm_weight: vec![1.0f32; hidden_dim],
        output_norm_bias: None,
        lm_head_weight: create_q4k_test_data(hidden_dim, config.vocab_size),
        lm_head_bias: None,
        #[cfg(feature = "cuda")]
        cuda_executor: None,
        #[cfg(feature = "cuda")]
        cuda_kernel_count: std::sync::atomic::AtomicU64::new(0),
        #[cfg(feature = "cuda")]
        cached_weight_names: std::sync::Mutex::new(std::collections::HashSet::new()),
    }
}

#[test]
fn test_forward_single_with_scratch_phi_256_q8k_gelu() {
    // Phi model with hidden_dim=256: covers Q8K GELU up/down paths in scratch
    let model = create_phi_256_model();
    let config = &model.config;
    let mut cache = OwnedQuantizedKVCache::from_config(config, 128);
    let mut scratch = InferenceScratchBuffer::from_config(config);

    let r1 = model.forward_single_with_scratch(5, &mut cache, 0, &mut scratch);
    assert!(r1.is_ok(), "Phi-256 scratch first: {}", r1.unwrap_err());

    // Second token covers attention_with_cache path
    let r2 = model.forward_single_with_scratch(10, &mut cache, 1, &mut scratch);
    assert!(r2.is_ok(), "Phi-256 scratch second: {}", r2.unwrap_err());
}

#[test]
fn test_forward_single_with_cache_phi_256_q8k() {
    // Phi 256 through the cache path for Q8K GELU coverage
    let model = create_phi_256_model();
    let config = &model.config;
    let mut cache = OwnedQuantizedKVCache::from_config(config, 128);

    let r1 = model.forward_single_with_cache(5, &mut cache, 0);
    assert!(r1.is_ok(), "Phi-256 cache first: {}", r1.unwrap_err());

    let r2 = model.forward_single_with_cache(10, &mut cache, 1);
    assert!(r2.is_ok(), "Phi-256 cache second: {}", r2.unwrap_err());
}

#[test]
fn test_forward_single_with_scratch_gqa_256() {
    // GQA model (num_kv_heads=2, num_heads=8) with hidden_dim=256
    // Covers: GQA V-expansion on first token, attention_with_cache_gqa on subsequent
    let model = create_llama_256_gqa_model();
    let config = &model.config;
    let mut cache = OwnedQuantizedKVCache::from_config(config, 128);
    let mut scratch = InferenceScratchBuffer::from_config(config);

    // First token: covers GQA V-expansion branch in scratch
    let r1 = model.forward_single_with_scratch(5, &mut cache, 0, &mut scratch);
    assert!(r1.is_ok(), "GQA-256 scratch first: {}", r1.unwrap_err());

    // Second token: covers attention_with_cache_gqa_into branch
    let r2 = model.forward_single_with_scratch(10, &mut cache, 1, &mut scratch);
    assert!(r2.is_ok(), "GQA-256 scratch second: {}", r2.unwrap_err());

    // Third token: further exercises cached attention
    let r3 = model.forward_single_with_scratch(15, &mut cache, 2, &mut scratch);
    assert!(r3.is_ok(), "GQA-256 scratch third: {}", r3.unwrap_err());
}

#[test]
fn test_forward_single_with_cache_gqa_256() {
    // GQA model through cache path
    let model = create_llama_256_gqa_model();
    let config = &model.config;
    let mut cache = OwnedQuantizedKVCache::from_config(config, 128);

    for i in 0..5 {
        let r = model.forward_single_with_cache(i as u32, &mut cache, i);
        assert!(r.is_ok(), "GQA-256 cache pos {}: {}", i, r.unwrap_err());
    }
}

#[test]
fn test_forward_single_with_scratch_q8k_multi_position() {
    // Multiple positions through Q8K path to exercise cached attention
    let model = create_llama_256_model();
    let config = &model.config;
    let mut cache = OwnedQuantizedKVCache::from_config(config, 128);
    let mut scratch = InferenceScratchBuffer::from_config(config);

    for i in 0..10 {
        let r = model.forward_single_with_scratch(i as u32, &mut cache, i, &mut scratch);
        assert!(r.is_ok(), "Q8K multi pos {}: {}", i, r.unwrap_err());
    }
}