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
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

//! T-COV-95 Coverage Bridge (Part 14 - GgufToAprQ4KConverter::convert full pipeline)
//!
//! Targets uncovered lines in convert/mod.rs:
//!   - GgufToAprQ4KConverter::convert() (lines 525-831) - GGUF->APR Q4K file pipeline
//!   - Metadata extraction, tensor iteration, binary index, alignment, CRC32
//!   - Error paths: missing file, truncated tensor, bounds check

use super::*;
use crate::apr::{HEADER_SIZE, MAGIC};
use crate::gguf::test_factory::build_minimal_llama_gguf;
use std::io::Read;

fn write_gguf_to_tempfile(gguf_data: &[u8]) -> (tempfile::NamedTempFile, std::path::PathBuf) {
    use std::io::Write;
    let mut tmp = tempfile::NamedTempFile::new().expect("create temp GGUF");
    tmp.write_all(gguf_data).expect("write GGUF");
    tmp.flush().expect("flush");
    let path = tmp.path().to_path_buf();
    (tmp, path)
}

// ============================================================================
// Happy path: full GGUF -> APR Q4K conversion
// ============================================================================

#[test]
fn test_q4k_convert_llama_produces_valid_apr() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    let result = GgufToAprQ4KConverter::convert(&gguf_path, &apr_path);
    assert!(result.is_ok(), "convert() failed: {:?}", result.err());

    let stats = result.expect("stats");
    assert_eq!(stats.architecture, "llama");
    assert_eq!(stats.num_layers, 1);
    assert_eq!(stats.hidden_size, 64);
    assert!(stats.tensor_count > 0, "should have tensors");
    assert!(stats.q4k_tensor_count > 0, "should have Q4K tensors");
    assert!(stats.total_bytes > 0, "should have written bytes");
}

#[test]
fn test_q4k_convert_output_has_apr_magic() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    GgufToAprQ4KConverter::convert(&gguf_path, &apr_path).expect("should convert");

    let apr_data = std::fs::read(&apr_path).expect("read APR");
    assert!(apr_data.len() >= HEADER_SIZE, "APR file too small");
    assert_eq!(&apr_data[0..4], &MAGIC, "APR magic mismatch");
    assert_eq!(apr_data[4], 2, "version major should be 2");
    assert_eq!(apr_data[5], 0, "version minor should be 0");
}

#[test]
fn test_q4k_convert_output_header_offsets_valid() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    GgufToAprQ4KConverter::convert(&gguf_path, &apr_path).expect("should convert");

    let apr_data = std::fs::read(&apr_path).expect("read APR");

    // Parse header offsets
    let tensor_count = u32::from_le_bytes(apr_data[8..12].try_into().expect("type conversion"));
    let metadata_offset = u64::from_le_bytes(apr_data[12..20].try_into().expect("type conversion"));
    let metadata_len = u32::from_le_bytes(apr_data[20..24].try_into().expect("type conversion"));
    let tensor_index_offset = u64::from_le_bytes(apr_data[24..32].try_into().expect("type conversion"));
    let data_offset = u64::from_le_bytes(apr_data[32..40].try_into().expect("type conversion"));

    assert!(tensor_count > 0, "should have tensors in header");
    assert_eq!(
        metadata_offset, HEADER_SIZE as u64,
        "metadata starts after header"
    );
    assert!(metadata_len > 0, "metadata should not be empty");
    assert!(
        tensor_index_offset > metadata_offset,
        "tensor index after metadata"
    );
    assert!(data_offset > tensor_index_offset, "data after tensor index");
    assert_eq!(data_offset % 64, 0, "data offset must be 64-byte aligned");
    assert!(
        (apr_data.len() as u64) >= data_offset,
        "file must be at least as large as data offset"
    );
}

#[test]
fn test_q4k_convert_metadata_contains_architecture() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    GgufToAprQ4KConverter::convert(&gguf_path, &apr_path).expect("should convert");

    let apr_data = std::fs::read(&apr_path).expect("read APR");

    // Extract metadata JSON
    let metadata_offset = u64::from_le_bytes(apr_data[12..20].try_into().expect("type conversion")) as usize;
    let metadata_len = u32::from_le_bytes(apr_data[20..24].try_into().expect("type conversion")) as usize;
    let metadata_bytes = &apr_data[metadata_offset..metadata_offset + metadata_len];
    let metadata: serde_json::Value =
        serde_json::from_slice(metadata_bytes).expect("parse metadata JSON");

    assert_eq!(metadata["architecture"], "llama");
    assert_eq!(metadata["hidden_size"], 64);
    assert_eq!(metadata["num_hidden_layers"], 1);
    assert_eq!(metadata["num_attention_heads"], 4);
    assert_eq!(metadata["num_key_value_heads"], 4);
    assert_eq!(metadata["vocab_size"], 32);
    assert_eq!(metadata["intermediate_size"], 128);
    assert_eq!(metadata["model_type"], "transformer_lm_q4k");
    assert!(metadata["quantization"]["quant_type"] == "Q4_K");
}

#[test]
fn test_q4k_convert_tensor_count_matches_stats() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    let stats = GgufToAprQ4KConverter::convert(&gguf_path, &apr_path).expect("should convert");

    let apr_data = std::fs::read(&apr_path).expect("read APR");
    let header_tensor_count = u32::from_le_bytes(apr_data[8..12].try_into().expect("type conversion")) as usize;
    assert_eq!(header_tensor_count, stats.tensor_count);
}

#[test]
fn test_q4k_convert_crc32_checksum_valid() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    GgufToAprQ4KConverter::convert(&gguf_path, &apr_path).expect("should convert");

    let apr_data = std::fs::read(&apr_path).expect("read APR");
    let stored_checksum = u32::from_le_bytes(apr_data[40..44].try_into().expect("type conversion"));

    // Recompute checksum over header bytes (excluding checksum field at [40..44])
    let header = &apr_data[0..HEADER_SIZE];
    let recomputed = compute_apr_header_checksum(header);
    assert_eq!(stored_checksum, recomputed, "CRC32 checksum mismatch");
}

// ============================================================================
// GQA model (different num_kv_heads)
// ============================================================================

#[test]
fn test_q4k_convert_gqa_model() {
    // GQA: 8 query heads, 2 KV heads (ratio 4)
    let gguf_data = build_minimal_llama_gguf(32, 128, 256, 8, 2);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    let stats = GgufToAprQ4KConverter::convert(&gguf_path, &apr_path).expect("should convert");
    assert_eq!(stats.hidden_size, 128);

    let apr_data = std::fs::read(&apr_path).expect("read APR");
    let metadata_offset = u64::from_le_bytes(apr_data[12..20].try_into().expect("type conversion")) as usize;
    let metadata_len = u32::from_le_bytes(apr_data[20..24].try_into().expect("type conversion")) as usize;
    let metadata: serde_json::Value =
        serde_json::from_slice(&apr_data[metadata_offset..metadata_offset + metadata_len])
            .expect("parse metadata");

    assert_eq!(metadata["num_attention_heads"], 8);
    assert_eq!(metadata["num_key_value_heads"], 2);
}

// ============================================================================
// Error paths
// ============================================================================

#[test]
fn test_q4k_convert_missing_input_file() {
    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let result = GgufToAprQ4KConverter::convert(
        std::path::Path::new("/nonexistent/model.gguf"),
        tmp_out.path(),
    );
    assert!(result.is_err(), "should fail for missing input");
}

#[test]
fn test_q4k_convert_invalid_gguf_data() {
    let bad_data = vec![0xFF; 200];
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&bad_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let result = GgufToAprQ4KConverter::convert(&gguf_path, tmp_out.path());
    assert!(result.is_err(), "should fail for invalid GGUF");
}

#[test]
fn test_q4k_convert_empty_gguf() {
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&[]);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let result = GgufToAprQ4KConverter::convert(&gguf_path, tmp_out.path());
    assert!(result.is_err(), "should fail for empty GGUF");
}

// ============================================================================
// Roundtrip: convert GGUF -> APR Q4K -> load APR back
// ============================================================================

#[test]
fn test_q4k_convert_roundtrip_loadable() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    GgufToAprQ4KConverter::convert(&gguf_path, &apr_path).expect("should convert");

    // Verify APR file can be loaded back as AprTransformer
    let apr_data = std::fs::read(&apr_path).expect("read APR");
    let result = crate::apr_transformer::AprTransformer::from_apr_bytes(&apr_data);
    assert!(
        result.is_ok(),
        "APR file should be loadable: {:?}",
        result.err()
    );

    let transformer = result.expect("transformer");
    assert_eq!(transformer.config.architecture, "llama");
    assert_eq!(transformer.config.hidden_dim, 64);
    assert_eq!(transformer.config.num_layers, 1);
}

// ============================================================================
// Dtype coverage: verify Q4K tensors are correctly typed in output
// ============================================================================

#[test]
fn test_q4k_convert_has_quantized_flag() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    GgufToAprQ4KConverter::convert(&gguf_path, &apr_path).expect("should convert");

    let apr_data = std::fs::read(&apr_path).expect("read APR");
    // Flags at offset [6..8]
    let flags = u16::from_le_bytes(apr_data[6..8].try_into().expect("type conversion"));
    assert_eq!(flags & 0x0020, 0x0020, "QUANTIZED flag should be set");
}

#[test]
fn test_q4k_convert_q4k_tensor_count_correct() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    let stats = GgufToAprQ4KConverter::convert(&gguf_path, &apr_path).expect("should convert");

    // build_minimal_llama_gguf has 7 Q4K tensors:
    // attn_q, attn_k, attn_v, attn_output, ffn_up, ffn_down, ffn_gate
    assert_eq!(stats.q4k_tensor_count, 7, "should have 7 Q4K tensors");
}

// ============================================================================
// Rope type inference
// ============================================================================

#[test]
fn test_q4k_convert_rope_type_in_metadata() {
    let gguf_data = build_minimal_llama_gguf(32, 64, 128, 4, 4);
    let (_tmp_in, gguf_path) = write_gguf_to_tempfile(&gguf_data);

    let tmp_out = tempfile::NamedTempFile::new().expect("create temp APR");
    let apr_path = tmp_out.path().to_path_buf();

    GgufToAprQ4KConverter::convert(&gguf_path, &apr_path).expect("should convert");

    let apr_data = std::fs::read(&apr_path).expect("read APR");
    let metadata_offset = u64::from_le_bytes(apr_data[12..20].try_into().expect("type conversion")) as usize;
    let metadata_len = u32::from_le_bytes(apr_data[20..24].try_into().expect("type conversion")) as usize;
    let metadata: serde_json::Value =
        serde_json::from_slice(&apr_data[metadata_offset..metadata_offset + metadata_len])
            .expect("parse metadata");

    // LLaMA uses NEOX style (type 2) or NORM (type 0) depending on metadata
    assert!(
        metadata.get("rope_type").is_some(),
        "rope_type should be in metadata"
    );
}