apr-cli 0.30.0

CLI tool for APR model inspection, debugging, and operations
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

impl ChatSession {
        pub(super) fn new(path: &Path) -> Result<Self, CliError> {
            contract_pre_session_persistence!();
            println!("{}", "Loading model...".cyan());
            let start = Instant::now();

            // Read file bytes
            let mut file = File::open(path).map_err(|e| {
                CliError::ValidationFailed(format!("Failed to open model file: {e}"))
            })?;
            let mut model_bytes = Vec::new();
            file.read_to_end(&mut model_bytes).map_err(|e| {
                CliError::ValidationFailed(format!("Failed to read model file: {e}"))
            })?;

            // Detect format from magic bytes (Y14)
            let format = detect_format_from_bytes(&model_bytes);

            let elapsed = start.elapsed();
            let format_name = match format {
                ModelFormat::Apr => "APR",
                ModelFormat::Gguf => "GGUF",
                ModelFormat::SafeTensors => "SafeTensors",
                ModelFormat::Demo => "Demo",
            };
            println!(
                "{} {} format in {:.2}s ({:.1} MB)",
                "Loaded".green(),
                format_name,
                elapsed.as_secs_f32(),
                model_bytes.len() as f32 / 1_000_000.0
            );

            let (llama_tokenizer, qwen_tokenizer) = load_tokenizers(format, &model_bytes, path)?;

            if format == ModelFormat::SafeTensors {
                print_safetensors_config(path);
            }

            // Detect chat template from model architecture
            // GH-339: Warn on Raw fallback instead of silent degradation
            let model_name = detect_model_architecture(format, &model_bytes, path);
            let template_format = detect_format_from_name(&model_name);
            let chat_template = auto_detect_template(&model_name);

            if matches!(template_format, TemplateFormat::Raw) {
                eprintln!(
                    "{} Could not detect chat template for '{}', using raw format (no ChatML/Instruct wrapping)",
                    "Warning:".yellow(),
                    model_name.dimmed()
                );
            } else {
                println!(
                    "{} {} chat template",
                    "Detected".green(),
                    template_format_name(template_format).cyan()
                );
            }

            // GH-224: Eagerly initialize GPU models during "Loading model..." phase
            let model_path_buf = path.to_path_buf();

            let mut cached_gguf_mapped = None;
            #[cfg(feature = "cuda")]
            let mut cached_gguf_cuda = None;
            #[cfg(feature = "cuda")]
            let mut cuda_init_failed = false;

            if format == ModelFormat::Gguf {
                match realizar::gguf::MappedGGUFModel::from_path(&model_path_buf) {
                    Ok(mapped) => {
                        #[cfg(feature = "cuda")]
                        {
                            let (cuda, failed) = try_init_gguf_cuda(&mapped);
                            cached_gguf_cuda = cuda;
                            if failed { cuda_init_failed = true; }
                        }
                        cached_gguf_mapped = Some(mapped);
                    }
                    Err(e) => {
                        eprintln!("[GGUF mmap failed: {}, will mmap per message]", e);
                    }
                }
            }

            #[cfg(feature = "cuda")]
            let mut cached_apr_cuda = None;
            #[cfg(feature = "cuda")]
            if format == ModelFormat::Apr {
                let (cuda, failed) = try_init_apr_cuda(&model_bytes, path);
                cached_apr_cuda = cuda;
                if failed { cuda_init_failed = true; }
            }

            #[cfg(feature = "cuda")]
            let mut cached_safetensors_cuda = None;
            #[cfg(feature = "cuda")]
            if format == ModelFormat::SafeTensors {
                let (cuda, failed) = try_init_safetensors_cuda(&model_path_buf);
                cached_safetensors_cuda = cuda;
                if failed { cuda_init_failed = true; }
            }

            let session = Self {
                model_bytes,
                model_path: model_path_buf,
                format,
                history: Vec::new(),
                chat_template,
                template_format,
                llama_tokenizer,
                qwen_tokenizer,
                cached_gguf_mapped,
                #[cfg(feature = "cuda")]
                cached_gguf_cuda,
                #[cfg(feature = "cuda")]
                cached_apr_cuda,
                #[cfg(feature = "cuda")]
                cached_safetensors_cuda,
                #[cfg(feature = "cuda")]
                cuda_init_failed,
            };
            contract_post_session_persistence!(&());
            Ok(session)
        }
}