lattice-embed 0.13.0

SIMD-accelerated vector operations and embedding generation
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
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
use super::*;
use lattice_inference::model::qwen35_config::{LayerType, RopeParams, VisionModelConfig};
use lattice_inference::vision::checkpoint::{
    VisualBlockWeights, VisualMergerWeights, resolve_qwen35_single_decoder_safetensors,
};
use lattice_inference::weights::f16_weights::{
    F16AttentionWeights, F16CommonLayerWeights, F16FeedForwardWeights,
    F16FullAttentionLayerWeights, f32_to_f16_slice,
};

/// Deterministic pseudo-random f32 fill (xorshift LCG), matching the
/// fixture builder in `lattice_inference::vision::pooled_embed`'s own
/// unit tests, so this crate's wrapper is exercised against
/// non-trivial weights without needing a real checkpoint.
fn pseudo_random_fill(seed: u32, n: usize) -> Vec<f32> {
    let mut state = seed | 1;
    let mut next = move || {
        state ^= state << 13;
        state ^= state >> 17;
        state ^= state << 5;
        (state as f32 / u32::MAX as f32) * 0.2 - 0.1
    };
    (0..n).map(|_| next()).collect()
}

fn tiny_vision_cfg() -> VisionModelConfig {
    VisionModelConfig {
        depth: 1,
        hidden_size: 8,
        num_heads: 2,
        patch_size: 2,
        spatial_merge_size: 2,
        out_hidden_size: 8, // must equal decoder hidden_size below
        temporal_patch_size: 1,
        num_position_embeddings: 16,
        in_channels: 1,
        deepstack_visual_indexes: vec![],
        intermediate_size: None,
    }
}

fn tiny_vision_weights(vision_cfg: &VisionModelConfig, seed: u32) -> Qwen35VisionWeights {
    let hidden = vision_cfg.hidden_size;
    let patch_len = vision_cfg.in_channels
        * vision_cfg.temporal_patch_size
        * vision_cfg.patch_size
        * vision_cfg.patch_size;
    let mlp_dim = 2 * hidden;
    let merge_in = vision_cfg.spatial_merge_size * vision_cfg.spatial_merge_size * hidden;

    let block = VisualBlockWeights {
        qkv_weight: pseudo_random_fill(seed, 3 * hidden * hidden),
        qkv_bias: pseudo_random_fill(seed.wrapping_add(1), 3 * hidden),
        proj_weight: pseudo_random_fill(seed.wrapping_add(2), hidden * hidden),
        proj_bias: pseudo_random_fill(seed.wrapping_add(3), hidden),
        fc1_weight: pseudo_random_fill(seed.wrapping_add(4), mlp_dim * hidden),
        fc1_bias: pseudo_random_fill(seed.wrapping_add(5), mlp_dim),
        fc2_weight: pseudo_random_fill(seed.wrapping_add(6), hidden * mlp_dim),
        fc2_bias: pseudo_random_fill(seed.wrapping_add(7), hidden),
        norm1_weight: vec![1.0; hidden],
        norm1_bias: vec![0.0; hidden],
        norm2_weight: vec![1.0; hidden],
        norm2_bias: vec![0.0; hidden],
    };

    Qwen35VisionWeights {
        patch_embed_weight: pseudo_random_fill(seed.wrapping_add(8), hidden * patch_len),
        patch_embed_weight_shape: vec![
            hidden,
            vision_cfg.in_channels,
            vision_cfg.temporal_patch_size,
            vision_cfg.patch_size,
            vision_cfg.patch_size,
        ],
        patch_embed_bias: pseudo_random_fill(seed.wrapping_add(9), hidden),
        pos_embed: pseudo_random_fill(
            seed.wrapping_add(10),
            vision_cfg.num_position_embeddings * hidden,
        ),
        blocks: vec![block],
        merger: VisualMergerWeights {
            fc1_weight: pseudo_random_fill(seed.wrapping_add(11), merge_in * merge_in),
            fc1_bias: pseudo_random_fill(seed.wrapping_add(12), merge_in),
            fc2_weight: pseudo_random_fill(
                seed.wrapping_add(13),
                vision_cfg.out_hidden_size * merge_in,
            ),
            fc2_bias: pseudo_random_fill(seed.wrapping_add(14), vision_cfg.out_hidden_size),
            norm_weight: vec![1.0; hidden],
            norm_bias: vec![0.0; hidden],
        },
    }
}

/// A minimal one-layer full-attention decoder + vision config wired
/// together: small enough to hand-construct, non-trivial (pseudo-random)
/// projections so the pipeline is actually exercised end to end.
fn tiny_vlm_fixture() -> (Qwen35Config, F16ModelWeights, Qwen35VisionWeights) {
    let hidden = 8usize;
    let vocab = 16usize;
    let vision_cfg = tiny_vision_cfg();

    let cfg = Qwen35Config {
        hidden_size: hidden,
        num_hidden_layers: 1,
        vocab_size: vocab,
        intermediate_size: 4,
        rms_norm_eps: 1e-6,
        num_attention_heads: 1,
        num_key_value_heads: 1,
        head_dim: hidden,
        rope_theta: 1.0e7,
        partial_rotary_factor: 1.0,
        rope_parameters: Some(RopeParams {
            rope_theta: 1.0e7,
            partial_rotary_factor: Some(1.0),
            mrope_section: Some(vec![2, 1, 1]),
            mrope_interleaved: Some(true),
        }),
        linear_num_key_heads: 2,
        linear_num_value_heads: Some(2),
        linear_key_head_dim: 32,
        linear_value_head_dim: 32,
        linear_conv_kernel_dim: 4,
        num_experts: None,
        num_experts_per_tok: None,
        moe_intermediate_size: None,
        shared_expert_intermediate_size: None,
        output_router_logits: false,
        router_aux_loss_coef: None,
        tie_word_embeddings: true,
        full_attention_interval: 1,
        layer_types: vec![LayerType::FullAttention],
        layer_mask: vec![true],
        eos_token_id: 999,
        max_position_embeddings: 512,
        mtp_num_hidden_layers: 0,
        mtp_use_dedicated_embeddings: false,
        quarot_rotation_seed: None,
        vision_config: Some(vision_cfg.clone()),
        image_token_id: Some(9),
        video_token_id: None,
        vision_start_token_id: Some(10),
        vision_end_token_id: Some(11),
    };

    let to_f16 = |src: &[f32]| -> Vec<u16> {
        let mut dst = vec![0u16; src.len()];
        f32_to_f16_slice(src, &mut dst);
        dst
    };

    let embed_tokens_f32 = pseudo_random_fill(777, vocab * hidden);
    let q_dim = cfg.full_q_dim();
    let kv_dim = cfg.full_kv_dim();
    let full_weights = F16FullAttentionLayerWeights {
        q_proj: to_f16(&pseudo_random_fill(101, 2 * q_dim * hidden)),
        k_proj: to_f16(&pseudo_random_fill(102, kv_dim * hidden)),
        v_proj: to_f16(&pseudo_random_fill(103, kv_dim * hidden)),
        o_proj: to_f16(&pseudo_random_fill(104, hidden * q_dim)),
        q_norm: vec![0.0f32; hidden],
        k_norm: vec![0.0f32; hidden],
    };
    let common = F16CommonLayerWeights {
        input_layernorm: vec![0.0f32; hidden],
        post_attention_layernorm: vec![0.0f32; hidden],
        ffn: F16FeedForwardWeights::Dense {
            gate_proj: to_f16(&vec![0.0f32; 4 * hidden]),
            up_proj: to_f16(&vec![0.0f32; 4 * hidden]),
            down_proj: to_f16(&vec![0.0f32; hidden * 4]),
        },
    };
    let weights = F16ModelWeights {
        embed_tokens: to_f16(&embed_tokens_f32),
        final_norm: vec![0.0f32; hidden],
        layers: vec![(F16AttentionWeights::Full(full_weights), common)],
    };

    let vision_weights = tiny_vision_weights(&vision_cfg, 555);
    (cfg, weights, vision_weights)
}

fn make_test_png(w: u32, h: u32, seed: u8) -> Vec<u8> {
    use image::RgbImage;
    let mut img = RgbImage::new(w, h);
    for y in 0..h {
        for x in 0..w {
            let v = ((x + y + seed as u32) % 256) as u8;
            img.put_pixel(x, y, image::Rgb([v, v, v]));
        }
    }
    let mut buf = Vec::new();
    img.write_to(&mut std::io::Cursor::new(&mut buf), image::ImageFormat::Png)
        .unwrap();
    buf
}

fn tiny_tokenizer() -> BpeTokenizer {
    let mut vocab_map = std::collections::HashMap::new();
    for (i, c) in ["describe", "this", "image"].iter().enumerate() {
        vocab_map.insert((*c).to_string(), i as u32);
    }
    BpeTokenizer::from_vocab_and_merges(vocab_map, vec![]).expect("tokenizer constructs")
}

/// Single-character vocab: with no merges, a byte-level BPE tokenizer
/// falls back to per-character tokens, so (unlike `tiny_tokenizer`'s
/// whole-word entries) this actually produces non-empty `real_length`
/// output — required by `embed_text_vlm_f16`'s empty-prompt guard.
/// Mirrors the tokenizer `cpu_f16.rs`'s own `embed_text_vlm_f16` tests use.
fn single_char_tokenizer() -> BpeTokenizer {
    let mut vocab_map = std::collections::HashMap::new();
    for (i, c) in ["a", "b", "c"].iter().enumerate() {
        vocab_map.insert((*c).to_string(), i as u32);
    }
    BpeTokenizer::from_vocab_and_merges(vocab_map, vec![]).expect("tokenizer constructs")
}

fn tiny_vlm_checkpoint_shapes() -> Vec<(String, Vec<usize>)> {
    let hidden = 8usize;
    let mut shapes = vec![
        (
            "model.language_model.embed_tokens.weight".to_string(),
            vec![16, hidden],
        ),
        ("model.language_model.norm.weight".to_string(), vec![hidden]),
        (
            "model.language_model.layers.0.input_layernorm.weight".to_string(),
            vec![hidden],
        ),
        (
            "model.language_model.layers.0.post_attention_layernorm.weight".to_string(),
            vec![hidden],
        ),
        (
            "model.language_model.layers.0.mlp.gate_proj.weight".to_string(),
            vec![4, hidden],
        ),
        (
            "model.language_model.layers.0.mlp.up_proj.weight".to_string(),
            vec![4, hidden],
        ),
        (
            "model.language_model.layers.0.mlp.down_proj.weight".to_string(),
            vec![hidden, 4],
        ),
        (
            "model.language_model.layers.0.self_attn.q_proj.weight".to_string(),
            vec![16, hidden],
        ),
        (
            "model.language_model.layers.0.self_attn.k_proj.weight".to_string(),
            vec![hidden, hidden],
        ),
        (
            "model.language_model.layers.0.self_attn.v_proj.weight".to_string(),
            vec![hidden, hidden],
        ),
        (
            "model.language_model.layers.0.self_attn.o_proj.weight".to_string(),
            vec![hidden, hidden],
        ),
        (
            "model.language_model.layers.0.self_attn.q_norm.weight".to_string(),
            vec![hidden],
        ),
        (
            "model.language_model.layers.0.self_attn.k_norm.weight".to_string(),
            vec![hidden],
        ),
        (
            "model.visual.patch_embed.proj.weight".to_string(),
            vec![hidden, 3, 1, 2, 2],
        ),
        (
            "model.visual.patch_embed.proj.bias".to_string(),
            vec![hidden],
        ),
        (
            "model.visual.pos_embed.weight".to_string(),
            vec![16, hidden],
        ),
        (
            "model.visual.merger.linear_fc1.weight".to_string(),
            vec![32, 32],
        ),
        ("model.visual.merger.linear_fc1.bias".to_string(), vec![32]),
        (
            "model.visual.merger.linear_fc2.weight".to_string(),
            vec![hidden, 32],
        ),
        (
            "model.visual.merger.linear_fc2.bias".to_string(),
            vec![hidden],
        ),
        ("model.visual.merger.norm.weight".to_string(), vec![hidden]),
        ("model.visual.merger.norm.bias".to_string(), vec![hidden]),
    ];
    for (suffix, shape) in [
        ("attn.qkv.weight", vec![24, hidden]),
        ("attn.qkv.bias", vec![24]),
        ("attn.proj.weight", vec![hidden, hidden]),
        ("attn.proj.bias", vec![hidden]),
        ("mlp.linear_fc1.weight", vec![32, hidden]),
        ("mlp.linear_fc1.bias", vec![32]),
        ("mlp.linear_fc2.weight", vec![hidden, 32]),
        ("mlp.linear_fc2.bias", vec![hidden]),
        ("norm1.weight", vec![hidden]),
        ("norm1.bias", vec![hidden]),
        ("norm2.weight", vec![hidden]),
        ("norm2.bias", vec![hidden]),
    ] {
        shapes.push((format!("model.visual.blocks.0.{suffix}"), shape));
    }
    shapes
}

fn write_f32_safetensors(path: &Path, shapes: &[(String, Vec<usize>)]) {
    write_f32_safetensors_with_offset(path, shapes, 0.0);
}

fn write_f32_safetensors_with_offset(path: &Path, shapes: &[(String, Vec<usize>)], offset: f32) {
    let mut header_parts = Vec::with_capacity(shapes.len());
    let mut data = Vec::new();
    for (i, (name, shape)) in shapes.iter().enumerate() {
        let start = data.len();
        let numel: usize = shape.iter().product();
        for _ in 0..numel {
            data.extend_from_slice(&(offset + (i + 1) as f32 / 100.0).to_le_bytes());
        }
        let end = data.len();
        let shape = shape
            .iter()
            .map(usize::to_string)
            .collect::<Vec<_>>()
            .join(",");
        header_parts.push(format!(
            r#""{name}":{{"dtype":"F32","shape":[{shape}],"data_offsets":[{start},{end}]}}"#
        ));
    }
    let header = format!("{{{}}}", header_parts.join(","));
    let mut bytes = Vec::with_capacity(8 + header.len() + data.len());
    bytes.extend_from_slice(&(header.len() as u64).to_le_bytes());
    bytes.extend_from_slice(header.as_bytes());
    bytes.extend_from_slice(&data);
    std::fs::write(path, bytes).expect("write safetensors fixture");
}

fn write_tiny_tokenizer_json(dir: &Path) {
    let tokenizer = r#"{
            "model": {
                "type": "BPE",
                "vocab": {
                    "a": 0, "b": 1, "c": 2, "d": 3,
                    "e": 4, "f": 5, "g": 6, "h": 7,
                    "i": 8, "j": 9, "k": 10, "l": 11,
                    "m": 12, "n": 13, "o": 14, "p": 15
                },
                "merges": []
            }
        }"#;
    std::fs::write(dir.join("tokenizer.json"), tokenizer).expect("write tokenizer.json");
}

fn write_tiny_vlm_checkpoint(dir: &Path, indexed: bool) {
    write_tiny_vlm_checkpoint_with_max_position_embeddings(dir, indexed, 512);
}

fn write_tiny_vlm_checkpoint_with_max_position_embeddings(
    dir: &Path,
    indexed: bool,
    max_position_embeddings: usize,
) {
    let config = format!(
        r#"{{
            "text_config": {{
                "hidden_size": 8,
                "num_hidden_layers": 1,
                "vocab_size": 16,
                "intermediate_size": 4,
                "rms_norm_eps": 0.000001,
                "num_attention_heads": 1,
                "num_key_value_heads": 1,
                "head_dim": 8,
                "rope_theta": 10000000.0,
                "partial_rotary_factor": 1.0,
                "rope_parameters": {{
                    "rope_theta": 10000000.0,
                    "partial_rotary_factor": 1.0,
                    "mrope_section": [2, 1, 1],
                    "mrope_interleaved": true
                }},
                "linear_num_key_heads": 2,
                "linear_num_value_heads": 2,
                "linear_key_head_dim": 32,
                "linear_value_head_dim": 32,
                "linear_conv_kernel_dim": 4,
                "tie_word_embeddings": true,
                "full_attention_interval": 1,
                "layer_types": ["full_attention"],
                "layer_mask": [true],
                "eos_token_id": 15,
                "max_position_embeddings": {max_position_embeddings}
            }},
            "vision_config": {{
                "depth": 1,
                "hidden_size": 8,
                "num_heads": 2,
                "patch_size": 2,
                "spatial_merge_size": 2,
                "out_hidden_size": 8,
                "temporal_patch_size": 1,
                "num_position_embeddings": 16,
                "in_channels": 3,
                "deepstack_visual_indexes": []
            }},
            "image_token_id": 9,
            "vision_start_token_id": 10,
            "vision_end_token_id": 11,
            "tie_word_embeddings": true
        }}"#
    );
    std::fs::write(dir.join("config.json"), &config).expect("write config.json");
    write_tiny_tokenizer_json(dir);

    let shapes = tiny_vlm_checkpoint_shapes();
    let shard_name = if indexed {
        "model-00001-of-00001.safetensors"
    } else {
        "model.safetensors"
    };
    write_f32_safetensors(&dir.join(shard_name), &shapes);
    if indexed {
        let weight_map = shapes
            .iter()
            .map(|(name, _)| format!(r#""{name}":"{shard_name}""#))
            .collect::<Vec<_>>()
            .join(",");
        std::fs::write(
            dir.join("model.safetensors.index.json"),
            format!(r#"{{"weight_map":{{{weight_map}}}}}"#),
        )
        .expect("write one-shard index");
    }
}

/// The embed-crate wrapper must return the exact same vector as calling
/// the raw inference-crate primitive directly: wiring adds no numerical
/// difference. This is the core claim of this module (a wire-through,
/// not a reimplementation).
#[test]
fn embed_image_matches_raw_inference_primitive() {
    let (cfg, weights, vision_weights) = tiny_vlm_fixture();
    let tokenizer = tiny_tokenizer();
    let png = make_test_png(8, 8, 0);

    let model = VisionEmbeddingModel::new(
        weights.clone(),
        cfg.clone(),
        vision_weights.clone(),
        tokenizer.clone(),
    );
    let via_wrapper = model
        .embed_image(
            &png,
            "describe this image",
            PoolingStrategy::MeanVisualTokens,
        )
        .expect("wrapper embed_image succeeds");

    let via_raw = embed_image_from_bytes_f16(
        &weights,
        &cfg,
        &vision_weights,
        &tokenizer,
        &png,
        "describe this image",
        PoolingStrategy::MeanVisualTokens,
    )
    .expect("raw primitive succeeds");

    assert_eq!(
        via_wrapper, via_raw,
        "embed-crate wrapper must return the identical vector to the raw primitive"
    );
}

/// Same wire-through claim as
/// `embed_image_matches_raw_inference_primitive`, for the Metal entry
/// point: the crate wrapper must add no numerical difference relative to
/// calling `embed_image_from_bytes_f16_metal` directly.
#[cfg(all(target_os = "macos", feature = "metal-gpu"))]
#[test]
fn embed_image_metal_matches_raw_inference_primitive() {
    use lattice_inference::vision::embed_image_from_bytes_f16_metal;

    let (cfg, weights, vision_weights) = tiny_vlm_fixture();
    let tokenizer = tiny_tokenizer();
    let png = make_test_png(8, 8, 0);

    let model = VisionEmbeddingModel::new(
        weights.clone(),
        cfg.clone(),
        vision_weights.clone(),
        tokenizer.clone(),
    );
    let via_wrapper = model
        .embed_image_metal(
            &png,
            "describe this image",
            PoolingStrategy::MeanVisualTokens,
        )
        .expect("wrapper embed_image_metal succeeds");

    let via_raw = embed_image_from_bytes_f16_metal(
        &weights,
        &cfg,
        &vision_weights,
        &tokenizer,
        &png,
        "describe this image",
        PoolingStrategy::MeanVisualTokens,
    )
    .expect("raw metal primitive succeeds");

    assert_eq!(
        via_wrapper, via_raw,
        "embed-crate Metal wrapper must return the identical vector to the raw primitive"
    );
}

/// Off this cfg gate, the wrapper must surface the CPU/Metal
/// distinction the inference crate makes (`InferenceError::
/// UnsupportedModel`) as `EmbedError::InferenceFailed`, never silently
/// substituting `embed_image`'s CPU output.
#[cfg(not(all(target_os = "macos", feature = "metal-gpu")))]
#[test]
fn embed_image_metal_fails_closed_without_metal_gpu() {
    let (cfg, weights, vision_weights) = tiny_vlm_fixture();
    let tokenizer = tiny_tokenizer();
    let png = make_test_png(8, 8, 0);
    let model = VisionEmbeddingModel::new(weights, cfg, vision_weights, tokenizer);

    let err = model
        .embed_image_metal(
            &png,
            "describe this image",
            PoolingStrategy::MeanVisualTokens,
        )
        .expect_err("Metal wrapper must fail without the metal-gpu feature");
    assert!(matches!(err, EmbedError::InferenceFailed(_)));
}

#[test]
fn embed_image_is_deterministic_and_normalized() {
    let (cfg, weights, vision_weights) = tiny_vlm_fixture();
    let tokenizer = tiny_tokenizer();
    let png = make_test_png(8, 8, 0);
    let model = VisionEmbeddingModel::new(weights, cfg.clone(), vision_weights, tokenizer);

    let v1 = model
        .embed_image(
            &png,
            "describe this image",
            PoolingStrategy::MeanVisualTokens,
        )
        .expect("embed succeeds");
    let v2 = model
        .embed_image(
            &png,
            "describe this image",
            PoolingStrategy::MeanVisualTokens,
        )
        .expect("embed succeeds");

    assert_eq!(
        v1, v2,
        "same image + prompt must produce an identical vector"
    );
    assert_eq!(v1.len(), model.dimensions());
    assert!(v1.iter().all(|x| x.is_finite()));
    let norm: f32 = v1.iter().map(|x| x * x).sum::<f32>().sqrt();
    assert!((norm - 1.0).abs() < 1e-4, "expected unit norm, got {norm}");
}

#[test]
fn embed_image_rejects_non_vlm_checkpoint() {
    let (mut cfg, weights, vision_weights) = tiny_vlm_fixture();
    cfg.vision_config = None;
    let tokenizer = tiny_tokenizer();
    let png = make_test_png(8, 8, 0);
    let model = VisionEmbeddingModel::new(weights, cfg, vision_weights, tokenizer);

    let err = model
        .embed_image(
            &png,
            "describe this image",
            PoolingStrategy::MeanVisualTokens,
        )
        .expect_err("a checkpoint with no vision_config must be rejected");
    let msg = err.to_string();
    assert!(matches!(err, EmbedError::InvalidInput(_)));
    assert!(
        msg.contains("vision_config"),
        "error must name the missing field, got: {msg}"
    );
}

#[test]
fn embed_image_rejects_misaligned_image() {
    let (cfg, weights, vision_weights) = tiny_vlm_fixture();
    let tokenizer = tiny_tokenizer();
    // factor = patch_size(2) * merge(2) = 4; 6 is not a multiple of 4.
    let png = make_test_png(6, 4, 0);
    let model = VisionEmbeddingModel::new(weights, cfg, vision_weights, tokenizer);

    let err = model
        .embed_image(
            &png,
            "describe this image",
            PoolingStrategy::MeanVisualTokens,
        )
        .expect_err("a misaligned image must be rejected, not panic");
    assert!(matches!(err, EmbedError::InvalidInput(_)));
}

/// Issue #1408 sibling hole (`embed_image`): unlike the sibling
/// `lattice_inference::serve::embeddings::EmbeddingModel::
/// image_scaffold_token_count`, whose one production caller always
/// passes an empty `prompt` (so its own doc comment notes the
/// truncation-vs-window mismatch is not live there), `embed_image` is a
/// general public entry point where a caller can pass any `prompt`. An
/// 8x8 test image against this fixture's tiny vision config produces 4
/// merged patch pads, so its scaffold (`vision_start` + 4 pads +
/// `vision_end`) is 6 tokens -- already over a 5-token window with an
/// empty prompt, so this exercises the scaffold-only side of the
/// admission check without needing any text tokens at all.
#[test]
fn embed_image_rejects_when_scaffold_exceeds_context_window() {
    let (mut cfg, weights, vision_weights) = tiny_vlm_fixture();
    cfg.max_position_embeddings = 5;
    let tokenizer = tiny_tokenizer();
    let model = VisionEmbeddingModel::new(weights, cfg, vision_weights, tokenizer);
    let png = make_test_png(8, 8, 0);

    let err = model
        .embed_image(&png, "", PoolingStrategy::MeanVisualTokens)
        .expect_err("scaffold tokens alone exceeding the context window must be rejected");
    assert!(matches!(err, EmbedError::InvalidInput(_)), "got: {err:?}");
    let msg = err.to_string();
    assert!(
        msg.contains('6'),
        "error must name the scaffold token total (6), got: {msg}"
    );
    assert!(
        msg.contains('5'),
        "error must name the context window (5), got: {msg}"
    );
}

/// Sibling boundary case: scaffold tokens landing exactly at the context
/// window (empty prompt, so no text tokens add to the total) must still
/// embed successfully.
#[test]
fn embed_image_accepts_when_scaffold_exactly_fits_context_window() {
    let (mut cfg, weights, vision_weights) = tiny_vlm_fixture();
    cfg.max_position_embeddings = 6;
    let tokenizer = tiny_tokenizer();
    let model = VisionEmbeddingModel::new(weights, cfg, vision_weights, tokenizer);
    let png = make_test_png(8, 8, 0);

    model
        .embed_image(&png, "", PoolingStrategy::MeanVisualTokens)
        .expect("scaffold tokens exactly at the context window boundary must still embed");
}

/// Same property as `embed_image_rejects_when_scaffold_exceeds_context_window`,
/// for the Metal-dispatching sibling: the admission check must run (and
/// reject) before `embed_image_metal` ever attempts a Metal dispatch, so
/// this must reject identically on a build/platform with no Metal
/// support at all -- no GPU is touched by this test.
#[test]
fn embed_image_metal_rejects_when_scaffold_exceeds_context_window() {
    let (mut cfg, weights, vision_weights) = tiny_vlm_fixture();
    cfg.max_position_embeddings = 5;
    let tokenizer = tiny_tokenizer();
    let model = VisionEmbeddingModel::new(weights, cfg, vision_weights, tokenizer);
    let png = make_test_png(8, 8, 0);

    let err = model
        .embed_image_metal(&png, "", PoolingStrategy::MeanVisualTokens)
        .expect_err(
            "scaffold tokens alone exceeding the context window must be rejected \
                         before any Metal dispatch is attempted",
        );
    assert!(matches!(err, EmbedError::InvalidInput(_)), "got: {err:?}");
}

#[test]
fn embed_text_matches_raw_inference_primitive() {
    let (cfg, weights, vision_weights) = tiny_vlm_fixture();
    let tokenizer = single_char_tokenizer();
    let model = VisionEmbeddingModel::new(
        weights.clone(),
        cfg.clone(),
        vision_weights,
        tokenizer.clone(),
    );

    let via_wrapper = model
        .embed_text("abc", PoolingStrategy::LastToken)
        .expect("wrapper embed_text succeeds");
    let via_raw = lattice_inference::forward::cpu_f16::embed_text_vlm_f16(
        &weights,
        &cfg,
        &tokenizer,
        "abc",
        PoolingStrategy::LastToken,
    )
    .expect("raw primitive succeeds");

    assert_eq!(via_wrapper, via_raw);
}

/// Issue #1408 (library-path admission guard, `embed_text`): a prompt
/// whose PRE-truncation token count exceeds the checkpoint's context
/// window must be rejected as caller input, before the forward pass ever
/// runs. This fixture's tokenizer cap (4096, `capped_tokenizer` never
/// lowers it) sits well above the 1-token window, so before this guard
/// existed "abc" tokenized to its full 3 real ids and only failed
/// downstream, as `EmbedError::InferenceFailed`, once the assembled
/// request's length was compared against `max_position_embeddings` deep
/// inside the forward pass -- a caller-input problem misreported as a
/// runtime one. A checkpoint whose window instead sits between the
/// tokenizer's default and a wider `max_position_embeddings` (the
/// scenario `capped_tokenizer` exists for) would raise the tokenizer's
/// own cap to match, so the same over-window prompt would truncate
/// silently there and never reach any error at all. Reverting the
/// `check_prompt_admission` call in `embed_text` reddens this.
#[test]
fn embed_text_rejects_prompt_exceeding_context_window_before_truncation() {
    let (mut cfg, weights, vision_weights) = tiny_vlm_fixture();
    cfg.max_position_embeddings = 1;
    let tokenizer = single_char_tokenizer();
    let model = VisionEmbeddingModel::new(weights, cfg, vision_weights, tokenizer);

    let err = model
        .embed_text("abc", PoolingStrategy::LastToken)
        .expect_err(
            "a prompt whose pre-truncation length exceeds the context window must be rejected",
        );
    assert!(
        matches!(err, EmbedError::InvalidInput(_)),
        "over-window prompt must be rejected as caller input, not embedded from a \
             truncated prefix, got: {err:?}"
    );
    let msg = err.to_string();
    assert!(
        msg.contains('3'),
        "error must name the pre-truncation token count (3), got: {msg}"
    );
    assert!(
        msg.contains('1'),
        "error must name the context window (1), got: {msg}"
    );
}

/// Sibling boundary case: a prompt whose pre-truncation length lands
/// exactly at the context window must still embed successfully -- the
/// admission guard must not reject an at-or-below-window prompt.
#[test]
fn embed_text_accepts_prompt_exactly_at_context_window() {
    let (mut cfg, weights, vision_weights) = tiny_vlm_fixture();
    cfg.max_position_embeddings = 1;
    let tokenizer = single_char_tokenizer();
    let model = VisionEmbeddingModel::new(weights, cfg, vision_weights, tokenizer);

    model
        .embed_text("a", PoolingStrategy::LastToken)
        .expect("a prompt exactly at the context window boundary must still embed");
}

#[test]
fn resolve_single_shard_rejects_multi_shard_index() {
    let tmp = tempfile::tempdir().expect("tempdir");
    let index_path = tmp.path().join("model.safetensors.index.json");
    std::fs::write(
        &index_path,
        r#"{"metadata":{},"weight_map":{"a":"shard1.safetensors","b":"shard2.safetensors"}}"#,
    )
    .expect("write index");

    let err = resolve_qwen35_single_decoder_safetensors(tmp.path())
        .expect_err("multi-shard must be rejected");
    let msg = err.to_string();
    assert!(msg.contains("sharded across 2 files"), "got: {msg}");
}

#[test]
fn resolve_single_shard_rejects_missing_checkpoint() {
    let tmp = tempfile::tempdir().expect("tempdir");
    let err = resolve_qwen35_single_decoder_safetensors(tmp.path())
        .expect_err("missing checkpoint must be rejected");
    assert!(matches!(err, InferenceError::ModelNotFound(_)));
}

#[test]
fn from_directory_without_checkpoint_reports_actionable_error() {
    let tmp = tempfile::tempdir().expect("tempdir");
    let config_json = include_str!(concat!(
        env!("CARGO_MANIFEST_DIR"),
        "/../inference/tests/fixtures/qwen35_0_8b_config.json"
    ));
    std::fs::write(tmp.path().join("config.json"), config_json).expect("write config.json");
    write_tiny_tokenizer_json(tmp.path());

    let Err(err) = VisionEmbeddingModel::from_directory(tmp.path()) else {
        panic!("a directory with no checkpoint must be rejected")
    };
    assert!(matches!(err, EmbedError::ModelInitialization(_)));
    let msg = err.to_string();
    assert!(
        msg.contains("model.safetensors") && msg.contains("model.safetensors.index.json"),
        "error must name the supported checkpoint layouts, got: {msg}"
    );
}

#[test]
fn from_directory_rejects_missing_tokenizer_before_checkpoint_materialization() {
    let tmp = tempfile::tempdir().expect("tempdir");
    write_tiny_vlm_checkpoint(tmp.path(), false);
    std::fs::remove_file(tmp.path().join("tokenizer.json")).expect("remove tokenizer fixture");
    std::fs::write(tmp.path().join("model.safetensors"), u64::MAX.to_le_bytes())
        .expect("replace checkpoint with a corrupt header");

    let Err(err) = VisionEmbeddingModel::from_directory(tmp.path()) else {
        panic!("a checkpoint without tokenizer.json must be rejected")
    };
    assert!(matches!(err, EmbedError::ModelInitialization(_)));
    let msg = err.to_string();
    assert!(msg.contains("tokenizer.json"), "got: {msg}");
    assert!(
        !msg.contains("vision weights") && !msg.contains("decoder weights"),
        "tokenizer admission must fail before tensor materialization, got: {msg}"
    );
}

#[test]
fn from_directory_rejects_quantized_checkpoint_before_tensor_loading() {
    let tmp = tempfile::tempdir().expect("tempdir");
    std::fs::write(tmp.path().join("quantize_index.json"), b"not valid json")
        .expect("write quantized checkpoint sentinel");

    let Err(err) = VisionEmbeddingModel::from_directory(tmp.path()) else {
        panic!("the f16 pooled decoder loader must reject quantized checkpoints")
    };
    assert!(matches!(err, EmbedError::ModelInitialization(_)));
    let msg = err.to_string();
    assert!(msg.contains("quantize_index.json"), "got: {msg}");
    assert!(msg.contains("not supported"), "got: {msg}");
    assert!(
        !msg.contains("config.json"),
        "the unsupported file-set must fail before unrelated component loading, got: {msg}"
    );
}

#[test]
fn from_directory_loads_single_model_safetensors_without_index() {
    let tmp = tempfile::tempdir().expect("tempdir");
    write_tiny_vlm_checkpoint(tmp.path(), false);

    let model = VisionEmbeddingModel::from_directory(tmp.path())
        .expect("single-file VLM checkpoint must load without a synthetic index");
    assert_eq!(model.dimensions(), 8);
}

/// Issue #1408, sibling-crate half: a `tokenizer.json` with no explicit
/// cap parses through `BpeTokenizer::from_tokenizer_json` at its
/// constructor default (4096, `DEFAULT_BPE_MAX_SEQ_LEN` in
/// `lattice_inference::tokenizer::bpe`). A checkpoint whose
/// `max_position_embeddings` sits above that default (here 5000, still
/// under the 8192 ceiling `capped_tokenizer` applies) must come out of
/// `from_directory` with its tokenizer cap raised to match -- otherwise
/// prompts between 4097 and 5000 tokens are silently truncated before
/// the forward pass ever sees them, even though the checkpoint can
/// process them. Reverting `VisionEmbeddingModel::new`'s call to
/// `capped_tokenizer` reddens this (cap stays at the tokenizer's own
/// default, 4096, strictly below `max_position_embeddings`, 5000).
#[test]
fn from_directory_raises_tokenizer_cap_to_context_window_above_tokenizer_default() {
    let tmp = tempfile::tempdir().expect("tempdir");
    write_tiny_vlm_checkpoint_with_max_position_embeddings(tmp.path(), false, 5000);

    let model = VisionEmbeddingModel::from_directory(tmp.path())
        .expect("VLM checkpoint with a wide context window must load");
    assert_eq!(
        model.tokenizer.max_seq_len(),
        5000,
        "tokenizer cap must be raised to max_position_embeddings, not left at the \
             tokenizer's own 4096 default"
    );
}

#[test]
fn single_file_and_one_shard_index_produce_identical_image_embeddings() {
    let single = tempfile::tempdir().expect("single tempdir");
    let indexed = tempfile::tempdir().expect("indexed tempdir");
    write_tiny_vlm_checkpoint(single.path(), false);
    write_tiny_vlm_checkpoint(indexed.path(), true);

    let single_model = VisionEmbeddingModel::from_directory(single.path())
        .expect("single-file VLM checkpoint loads");
    let indexed_model = VisionEmbeddingModel::from_directory(indexed.path())
        .expect("equivalent one-shard indexed VLM checkpoint loads");
    let image = make_test_png(4, 4, 17);
    let from_single = single_model
        .embed_image(&image, "a", PoolingStrategy::MeanVisualTokens)
        .expect("single-file image embedding succeeds");
    let from_index = indexed_model
        .embed_image(&image, "a", PoolingStrategy::MeanVisualTokens)
        .expect("indexed image embedding succeeds");

    assert_eq!(
        from_single, from_index,
        "equivalent single-file and indexed layouts must produce parity embeddings"
    );
}

#[test]
fn from_directory_rejects_index_map_that_contradicts_opened_shard_header() {
    let tmp = tempfile::tempdir().expect("tempdir");
    write_tiny_vlm_checkpoint(tmp.path(), true);
    std::fs::write(
        tmp.path().join("model.safetensors.index.json"),
        r#"{"weight_map":{"not.a.real.tensor":"model-00001-of-00001.safetensors"}}"#,
    )
    .expect("replace index with contradictory weight_map");

    let Err(err) = VisionEmbeddingModel::from_directory(tmp.path()) else {
        panic!("an authoritative index that omits the physical tensors must be rejected")
    };
    let msg = err.to_string();
    assert!(
        msg.contains("weight_map/header inventory mismatch"),
        "got: {msg}"
    );
}

#[cfg(unix)]
#[test]
fn from_directory_binds_visual_and_decoder_weights_across_path_replacement() {
    let tmp = tempfile::tempdir().expect("tempdir");
    write_tiny_vlm_checkpoint(tmp.path(), false);
    let checkpoint_path = tmp.path().join("model.safetensors");
    let replacement = tmp.path().join("replacement-checkpoint");
    write_f32_safetensors_with_offset(&replacement, &tiny_vlm_checkpoint_shapes(), 10.0);

    let model = with_after_visual_load_hook(
        move || {
            std::fs::rename(&replacement, &checkpoint_path)
                .expect("atomically replace checkpoint pathname with checkpoint B");
        },
        || {
            VisionEmbeddingModel::from_directory(tmp.path())
                .expect("constructor keeps both components on checkpoint A")
        },
    );

    assert_eq!(
        model.vision_weights.patch_embed_weight[0], 0.14,
        "visual weights must remain bound to checkpoint A"
    );
    let mut expected_embed = [0u16];
    f32_to_f16_slice(&[0.01], &mut expected_embed);
    assert_eq!(
        model.weights.embed_tokens[0], expected_embed[0],
        "decoder weights must remain bound to checkpoint A"
    );
}

#[test]
fn resolve_single_shard_prefers_existing_index_over_plain_file() {
    let tmp = tempfile::tempdir().expect("tempdir");
    std::fs::write(tmp.path().join("model.safetensors"), b"plain").expect("write convenience file");
    std::fs::write(
        tmp.path().join("model.safetensors.index.json"),
        r#"{"weight_map":{"tensor":"indexed.safetensors"}}"#,
    )
    .expect("write index");

    let resolved =
        resolve_qwen35_single_decoder_safetensors(tmp.path()).expect("single-shard index resolves");
    assert_eq!(resolved, tmp.path().join("indexed.safetensors"));
}

#[test]
fn resolve_single_shard_rejects_index_entry_escaping_model_directory() {
    let tmp = tempfile::tempdir().expect("tempdir");
    std::fs::write(
        tmp.path().join("model.safetensors.index.json"),
        r#"{"weight_map":{"tensor":"../outside.safetensors"}}"#,
    )
    .expect("write index");

    let err = resolve_qwen35_single_decoder_safetensors(tmp.path())
        .expect_err("an index entry must not escape the checkpoint directory");
    assert!(
        err.to_string().contains("escapes the model directory"),
        "got: {err}"
    );
}