realizar 0.8.5

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


#[test]
fn test_batched_workspace_hidden_buffer() {
    use crate::cuda::executor::test_fixtures::{setup_executor_harness, HarnessConfig};
    let Some(mut exec) = create_executor() else {
        return;
    };
    let config = HarnessConfig::default();
    if setup_executor_harness(&mut exec, &config).is_err() {
        return;
    }

    // Initialize batched workspace
    let result = exec.init_batched_workspace(config.hidden_dim, config.intermediate_dim, 4);
    assert!(result.is_ok());

    // Verify workspace is initialized (fields vary by implementation)
    assert!(exec.workspace.hidden_buf1.is_some());
}

// ========================================================================
// Coverage Tests: RMSNorm and Batch Edge Cases (v1.36.0)
// ========================================================================

#[test]
fn test_rmsnorm_gpu_ptr_batched() {
    use crate::cuda::executor::test_fixtures::{setup_executor_harness, HarnessConfig};
    let Some(mut exec) = create_executor() else {
        return;
    };
    let config = HarnessConfig::default();
    if setup_executor_harness(&mut exec, &config).is_err() {
        return;
    }

    let layer_weights = &exec.indexed_layer_weights[0];
    let input = GpuBuffer::from_host(&exec.context, &vec![1.0f32; config.hidden_dim]).expect("input");

    // rmsnorm_gpu_ptr is defined in batched.rs as pub(crate)
    let result = exec.rmsnorm_gpu_ptr(
        &input,
        layer_weights.attn_norm_ptr,
        layer_weights.attn_norm_len,
        config.hidden_dim as u32,
        1e-5,
    );
    let _ = result;
}

#[test]
fn test_forward_batched_different_positions() {
    use crate::cuda::executor::test_fixtures::{setup_executor_harness, HarnessConfig};
    let Some(mut exec) = create_executor() else {
        return;
    };
    let config = HarnessConfig::default();
    if setup_executor_harness(&mut exec, &config).is_err() {
        return;
    }

    let _ = exec.init_batched_workspace(config.hidden_dim, config.intermediate_dim, 4);

    // Non-contiguous positions
    let positions: Vec<u32> = vec![0, 5, 10, 15];
    let inputs: Vec<f32> = vec![0.1; 4 * config.hidden_dim];

    let result = exec.forward_batched_to_token_ids(
        &inputs,
        &positions,
        config.num_layers,
        config.hidden_dim as u32,
        config.intermediate_dim as u32,
        config.vocab_size as u32,
        1e-5,
    );
    let _ = result;
}

#[test]
fn test_transformer_layer_batched_different_layers() {
    use crate::cuda::executor::test_fixtures::{setup_executor_harness, HarnessConfig};
    let Some(mut exec) = create_executor() else {
        return;
    };
    let mut config = HarnessConfig::default();
    config.num_layers = 4;
    if setup_executor_harness(&mut exec, &config).is_err() {
        return;
    }

    let _ = exec.init_batched_workspace(config.hidden_dim, config.intermediate_dim, 4);

    let inputs: Vec<f32> = vec![0.1; 4 * config.hidden_dim];
    let input_buf = GpuBuffer::from_host(&exec.context, &inputs).expect("input_buf");
    let positions: [u32; 4] = [0, 1, 2, 3];

    // Test each layer
    for layer_idx in 0..config.num_layers {
        if layer_idx >= exec.indexed_layer_weights.len() {
            break;
        }
        let layer_weights = exec.indexed_layer_weights[layer_idx].clone();

        let result = exec.transformer_layer_batched(
            &input_buf,
            layer_idx,
            &layer_weights,
            4,
            &positions,
            config.hidden_dim as u32,
            config.intermediate_dim as u32,
            1e-5,
        );
        let _ = result;
    }
}

#[test]
fn test_batched_forward_m1() {
    use crate::cuda::executor::test_fixtures::{setup_executor_harness, HarnessConfig};
    let Some(mut exec) = create_executor() else {
        return;
    };
    let config = HarnessConfig::default();
    if setup_executor_harness(&mut exec, &config).is_err() {
        return;
    }

    // Batch size 1 - edge case
    let _ = exec.init_batched_workspace(config.hidden_dim, config.intermediate_dim, 1);

    let positions: Vec<u32> = vec![0];
    let inputs: Vec<f32> = vec![0.1; config.hidden_dim];

    let result = exec.forward_batched_to_token_ids(
        &inputs,
        &positions,
        config.num_layers,
        config.hidden_dim as u32,
        config.intermediate_dim as u32,
        config.vocab_size as u32,
        1e-5,
    );
    let _ = result;
}

#[test]
fn test_batched_kv_cache_multiple_batch_sizes() {
    use crate::cuda::executor::test_fixtures::{setup_executor_harness, HarnessConfig};
    let Some(mut exec) = create_executor() else {
        return;
    };
    let config = HarnessConfig::default();
    if setup_executor_harness(&mut exec, &config).is_err() {
        return;
    }

    // Test different batch sizes for KV cache init
    for batch_size in [1, 4, 8, 16] {
        let result = exec.init_batched_kv_cache_gpu(config.num_layers, batch_size);
        let _ = result;
    }
}

#[test]
fn test_transformer_layer_batched_qtype_q6k() {
    use crate::cuda::executor::test_fixtures::{setup_executor_harness, HarnessConfig};
    let Some(mut exec) = create_executor() else {
        return;
    };
    let config = HarnessConfig::default();
    if setup_executor_harness(&mut exec, &config).is_err() {
        return;
    }

    let _ = exec.init_batched_workspace(config.hidden_dim, config.intermediate_dim, 4);

    let inputs: Vec<f32> = vec![0.1; 4 * config.hidden_dim];
    let input_buf = GpuBuffer::from_host(&exec.context, &inputs).expect("input_buf");

    if exec.indexed_layer_weights.is_empty() {
        return;
    }

    // Modify weights to use Q6K
    let mut layer_weights = exec.indexed_layer_weights[0].clone();
    layer_weights.inner_mut().attn_v_qtype = WeightQuantType::Q6K;

    let positions: [u32; 4] = [0, 1, 2, 3];
    let result = exec.transformer_layer_batched(
        &input_buf,
        0,
        &layer_weights,
        4,
        &positions,
        config.hidden_dim as u32,
        config.intermediate_dim as u32,
        1e-5,
    );
    let _ = result;
}

#[test]
fn test_batched_graph_token_count() {
    use crate::cuda::executor::test_fixtures::{setup_executor_harness, HarnessConfig};
    let Some(mut exec) = create_executor() else {
        return;
    };
    let config = HarnessConfig::default();
    if setup_executor_harness(&mut exec, &config).is_err() {
        return;
    }

    let initial_count = exec.decode_token_count;

    let _ = exec.init_batched_workspace(config.hidden_dim, config.intermediate_dim, 4);

    let positions: Vec<u32> = vec![0, 1, 2, 3];
    let inputs: Vec<f32> = vec![0.1; 4 * config.hidden_dim];

    // Execute batched graphed forward
    let _ = exec.forward_batched_to_token_ids_graphed(
        &inputs,
        &positions,
        config.num_layers,
        config.hidden_dim as u32,
        config.intermediate_dim as u32,
        config.vocab_size as u32,
        1e-5,
    );

    // Token count might change after graphed forward
    assert!(exec.decode_token_count >= initial_count);
}

#[test]
fn test_batched_forward_varying_input_values() {
    use crate::cuda::executor::test_fixtures::{setup_executor_harness, HarnessConfig};
    let Some(mut exec) = create_executor() else {
        return;
    };
    let config = HarnessConfig::default();
    if setup_executor_harness(&mut exec, &config).is_err() {
        return;
    }

    let _ = exec.init_batched_workspace(config.hidden_dim, config.intermediate_dim, 4);

    let positions: Vec<u32> = vec![0, 1, 2, 3];

    // Different input values per batch element
    let inputs: Vec<f32> = (0..4 * config.hidden_dim)
        .map(|i| (i as f32 / 1000.0).sin())
        .collect();

    let result = exec.forward_batched_to_token_ids(
        &inputs,
        &positions,
        config.num_layers,
        config.hidden_dim as u32,
        config.intermediate_dim as u32,
        config.vocab_size as u32,
        1e-5,
    );
    let _ = result;
}