apr-cli 0.4.15

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
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


// =============================================================================
// Fallback ChatSession without realizar (demo mode only)
// =============================================================================

/// Fallback ChatSession stub when inference feature is disabled.
/// Chat requires realizar for inference — aprender is training only.
#[cfg(not(feature = "inference"))]
struct ChatSession;

#[cfg(not(feature = "inference"))]
impl ChatSession {
    fn new(_path: &Path) -> Result<Self, CliError> {
        Err(CliError::ValidationFailed(
            "Chat requires the 'inference' feature (realizar). Rebuild with: \
             cargo install --path crates/apr-cli --features inference"
                .to_string(),
        ))
    }

    fn generate(&mut self, _user_input: &str, _config: &ChatConfig) -> String {
        unreachable!("ChatSession::new always returns Err without inference feature")
    }
}

// =============================================================================
// REPL implementation with inference feature
// =============================================================================

/// Read one line of user input. Returns `None` on EOF, `Some("")` for empty, `Some(text)` otherwise.
fn read_repl_line() -> Result<Option<String>, CliError> {
    print!("{}", "You: ".green().bold());
    io::stdout().flush()?;
    let mut input = String::new();
    if io::stdin().read_line(&mut input)? == 0 {
        println!();
        return Ok(None);
    }
    Ok(Some(input.trim().to_string()))
}

/// Generate a response, update history, and print (inference mode).
#[cfg(feature = "inference")]
fn generate_and_print(session: &mut ChatSession, input: &str, config: &ChatConfig) {
    let response = session.generate(input, config);
    session.add_to_history("user", input);
    session.add_to_history("assistant", &response);
    println!("{} {}", "Assistant:".blue().bold(), response);
    if config.inspect {
        print_inspection_info_inference(session);
    }
    println!();
}

/// Process a single REPL input line. Returns `true` to quit, `false` to continue.
#[cfg(feature = "inference")]
fn process_repl_input(
    input: &str,
    session: &mut ChatSession,
    config: &ChatConfig,
) -> Result<bool, CliError> {
    if input.starts_with('/') {
        match handle_command_inference(input, session)? {
            CommandResult::Continue => return Ok(false),
            CommandResult::Quit => return Ok(true),
        }
    }
    generate_and_print(session, input, config);
    Ok(false)
}

#[cfg(feature = "inference")]
fn run_repl(path: &Path, config: &ChatConfig) -> Result<(), CliError> {
    let mut session = ChatSession::new(path)?;

    while let Some(input) = read_repl_line()? {
        if input.is_empty() {
            continue;
        }
        if process_repl_input(&input, &mut session, config)? {
            break;
        }
    }

    println!("{}", "Goodbye!".cyan());
    Ok(())
}

#[cfg(feature = "inference")]
fn handle_command_inference(
    input: &str,
    session: &mut ChatSession,
) -> Result<CommandResult, CliError> {
    let parts: Vec<&str> = input.splitn(2, ' ').collect();
    let cmd = parts[0].to_lowercase();

    match cmd.as_str() {
        "/quit" | "/exit" | "/q" => {
            return Ok(CommandResult::Quit);
        }
        "/clear" => {
            session.clear_history();
            println!("{}", "Conversation cleared.".yellow());
        }
        "/system" => {
            if parts.len() > 1 {
                println!("{} {}", "System prompt set:".yellow(), parts[1]);
            } else {
                println!("{}", "Usage: /system <prompt>".yellow());
            }
        }
        "/help" | "/h" | "/?" => {
            println!();
            println!("{}", "Commands:".white().bold());
            println!("  /quit, /exit, /q   Exit the chat");
            println!("  /clear             Clear conversation history");
            println!("  /system <prompt>   Set system prompt");
            println!("  /help, /h, /?      Show this help");
            println!();
        }
        _ => {
            println!("{} {}", "Unknown command:".red(), cmd);
        }
    }

    Ok(CommandResult::Continue)
}

#[cfg(feature = "inference")]
fn print_inspection_info_inference(session: &ChatSession) {
    println!();
    println!("{}", "[DEBUG] Session info:".dimmed());
    println!("{}", format!("  Format: {:?}", session.format()).dimmed());
    println!(
        "{}",
        format!("  Template: {:?}", session.template_format()).dimmed()
    );
    println!(
        "{}",
        format!("  History: {} messages", session.history_len()).dimmed()
    );
}

// =============================================================================
// REPL implementation without inference feature (fallback)
// =============================================================================

/// Process a single REPL input line (fallback). Returns `true` to quit, `false` to continue.
#[cfg(not(feature = "inference"))]
fn process_repl_input_fallback(
    input: &str,
    session: &mut ChatSession,
    config: &ChatConfig,
) -> Result<bool, CliError> {
    let _ = (input, session, config);
    unreachable!("ChatSession::new always returns Err without inference feature")
}

#[cfg(not(feature = "inference"))]
fn run_repl(path: &Path, config: &ChatConfig) -> Result<(), CliError> {
    let mut session = ChatSession::new(path)?;

    while let Some(input) = read_repl_line()? {
        if input.is_empty() {
            continue;
        }
        if process_repl_input_fallback(&input, &mut session, config)? {
            break;
        }
    }

    println!("{}", "Goodbye!".cyan());
    Ok(())
}

/// Generate a response, update history, and print (fallback mode).
#[cfg(not(feature = "inference"))]
fn generate_and_print_fallback(_session: &mut ChatSession, _input: &str, _config: &ChatConfig) {
    unreachable!("ChatSession::new always returns Err without inference feature")
}

enum CommandResult {
    Continue,
    Quit,
}

#[cfg(not(feature = "inference"))]
fn handle_command(_input: &str, _history: &mut Vec<String>) -> Result<CommandResult, CliError> {
    unreachable!("ChatSession::new always returns Err without inference feature")
}

#[cfg(not(feature = "inference"))]
fn print_inspection_info(_session: &ChatSession) {
    unreachable!("ChatSession::new always returns Err without inference feature")
}

/// Display top-k token probabilities (spec E1)
#[cfg(not(feature = "inference"))]
fn print_top_k(logits: &[f32], tokenizer: &Qwen2BpeTokenizer, k: usize) {
    println!();
    println!("{}", "[TOP-K CANDIDATES]".cyan().bold());

    // Compute softmax
    let max_val = logits.iter().fold(f32::NEG_INFINITY, |a, &b| a.max(b));
    let exp_vals: Vec<f32> = logits.iter().map(|&v| (v - max_val).exp()).collect();
    let sum: f32 = exp_vals.iter().sum();
    let probs: Vec<f32> = exp_vals.iter().map(|&v| v / sum).collect();

    // Get top-k indices
    let mut indexed: Vec<(usize, f32)> = probs.iter().copied().enumerate().collect();
    indexed.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));

    for (i, (token_id, prob)) in indexed.iter().take(k).enumerate() {
        let token_str = tokenizer.decode(&[*token_id as u32]);
        let display = if token_str.trim().is_empty() {
            format!("<token_{}>", token_id)
        } else {
            format!("\"{}\"", token_str.escape_debug())
        };

        let bar_len = (prob * 50.0) as usize;
        let bar = "█".repeat(bar_len);

        println!(
            "  {}. {} {:>6.2}% {}",
            i + 1,
            display.yellow(),
            prob * 100.0,
            bar.green()
        );
    }
}

/// Format parameter count in human-readable form
#[cfg(not(feature = "inference"))]
fn format_params(count: usize) -> String {
    if count >= 1_000_000_000 {
        format!("{:.1}B", count as f64 / 1_000_000_000.0)
    } else if count >= 1_000_000 {
        format!("{:.1}M", count as f64 / 1_000_000.0)
    } else if count >= 1_000 {
        format!("{:.1}K", count as f64 / 1_000.0)
    } else {
        format!("{count}")
    }
}