dynamo-renderer 10.4.5

Standalone OpenAI chat-template / prompt formatting (HF chat_template via minijinja).
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
// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

//! Python value spellings. Reference encoders and HF chat templates render through
//! Python, so the prompt text they produce uses Python's `json.dumps` and `str`.

use std::io;

use serde_json::ser::Formatter;

/// Recover serde_json numbers after MiniJinja's serializer has represented
/// arbitrary-precision Number structs as maps. Keep this at the JSON boundary;
/// canonical request values never pass through MiniJinja's bounded floats.
pub(crate) fn template_json<T: serde::Serialize>(
    value: T,
) -> serde_json::Result<serde_json::Value> {
    serde_json::from_value(serde_json::to_value(value)?)
}

/// MiniJinja doesn't understand serde_json's arbitrary-precision Number struct.
/// Construct its native numeric values explicitly, retaining the Python float
/// behavior of the template engine rather than exposing private Number fields.
pub(crate) fn template_value(value: &serde_json::Value) -> minijinja::Value {
    use serde_json::Value;
    match value {
        Value::Null => minijinja::Value::from(()),
        Value::Bool(value) => minijinja::Value::from(*value),
        Value::String(value) => minijinja::Value::from(value.clone()),
        Value::Number(value) => {
            if let Some(value) = value.as_i64() {
                minijinja::Value::from(value)
            } else if let Some(value) = value.as_u64() {
                minijinja::Value::from(value)
            } else {
                minijinja::Value::from(value.to_string().parse::<f64>().expect("valid JSON number"))
            }
        }
        Value::Array(values) => values.iter().map(template_value).collect(),
        Value::Object(values) => values
            .iter()
            .map(|(key, value)| (key.clone(), template_value(value)))
            .collect(),
    }
}

/// Python's `repr(float)`, which `json.dumps` and `str` use: the shortest round-trip
/// digits, fixed notation for exponents in [-4, 16) (always with a fraction, `100.0`),
/// otherwise scientific with a signed, two-digit exponent (`1e-06`, `1e+16`).
/// Non-finite values are `nan`, `inf`, and `-inf`.
pub(crate) fn python_float_repr(value: f64) -> String {
    // serde_json's ryu picks the same shortest digits as Python, halfway ties
    // included (`{:e}` breaks some differently); only the notation is respelled.
    let Some(shortest) = serde_json::Number::from_f64(value) else {
        let non_finite = if value.is_nan() {
            "nan"
        } else if value < 0.0 {
            "-inf"
        } else {
            "inf"
        };
        return non_finite.to_string();
    };
    python_number_repr(&shortest.to_string())
}

/// Preserve arbitrary-precision digits while applying Python's float notation.
fn python_number_repr(shortest: &str) -> String {
    if !shortest.contains(['.', 'e', 'E']) {
        return shortest.to_owned();
    }
    let (sign, unsigned) = match shortest.strip_prefix('-') {
        Some(unsigned) => ("-", unsigned),
        None => ("", shortest),
    };
    let (mantissa, exponent) = match unsigned.split_once(['e', 'E']) {
        Some((mantissa, exponent)) => match exponent.parse::<i64>() {
            Ok(exponent) => (mantissa, exponent),
            Err(_) => return shortest.to_owned(),
        },
        None => (unsigned, 0),
    };
    let (integer, fraction) = mantissa.split_once('.').unwrap_or((mantissa, ""));
    let all_digits = format!("{integer}{fraction}");
    let significant = all_digits.trim_start_matches('0');
    let digits = significant.trim_end_matches('0');
    if digits.is_empty() {
        return format!("{sign}0.0");
    }
    let leading_zeros = (all_digits.len() - significant.len()) as i64;
    let Some(exponent) = exponent.checked_add(integer.len() as i64 - 1 - leading_zeros) else {
        return shortest.to_owned();
    };
    if (-4..16).contains(&exponent) {
        let point = exponent + 1;
        let fixed = if point <= 0 {
            format!("0.{}{digits}", "0".repeat(point.unsigned_abs() as usize))
        } else if point as usize >= digits.len() {
            format!("{digits}{}.0", "0".repeat(point as usize - digits.len()))
        } else {
            format!(
                "{}.{}",
                &digits[..point as usize],
                &digits[point as usize..]
            )
        };
        format!("{sign}{fixed}")
    } else {
        let (head, tail) = digits.split_at(1);
        let mantissa = if tail.is_empty() {
            head.to_string()
        } else {
            format!("{head}.{tail}")
        };
        let exponent_sign = if exponent < 0 { '-' } else { '+' };
        format!(
            "{sign}{mantissa}e{exponent_sign}{:02}",
            exponent.unsigned_abs()
        )
    }
}

/// Writes a float as Python `json.dumps` does, instead of serde_json's `1e-6`/`0.00001`.
/// serde_json writes non-finite floats as `null` before reaching a formatter, so the
/// guard here is only defensive; Python would write `NaN`/`Infinity`, which request
/// JSON can't carry.
fn write_python_float<W: ?Sized + io::Write>(writer: &mut W, value: f64) -> io::Result<()> {
    if !value.is_finite() {
        return writer.write_all(b"null");
    }
    writer.write_all(python_float_repr(value).as_bytes())
}

/// Formatter matching Python `json.dumps` output for the given `separators`, `indent`
/// and `ensure_ascii`, with Python float spelling. Python's default separators are
/// `", "` and `": "`; serde_json's `CompactFormatter` writes `","`/`":"` instead, and
/// chat templates embed these strings directly into the prompt, so the separator
/// choice is model-visible.
pub(crate) struct PyJsonFormatter<'a> {
    item_sep: &'a str,
    key_sep: &'a str,
    /// Python `indent`: each item starts on a new line, indented by one copy of this
    /// per nesting level.
    indent: Option<&'a str>,
    level: usize,
    has_value: bool,
    ensure_ascii: bool,
}

impl<'a> PyJsonFormatter<'a> {
    pub(crate) fn new(
        item_sep: &'a str,
        key_sep: &'a str,
        indent: Option<&'a str>,
        ensure_ascii: bool,
    ) -> Self {
        Self {
            item_sep,
            key_sep,
            indent,
            level: 0,
            has_value: false,
            ensure_ascii,
        }
    }

    fn newline_indent<W: ?Sized + io::Write>(&self, writer: &mut W) -> io::Result<()> {
        if let Some(indent) = self.indent {
            writer.write_all(b"\n")?;
            for _ in 0..self.level {
                writer.write_all(indent.as_bytes())?;
            }
        }
        Ok(())
    }

    fn open<W: ?Sized + io::Write>(&mut self, writer: &mut W, bracket: &[u8]) -> io::Result<()> {
        self.level += 1;
        self.has_value = false;
        writer.write_all(bracket)
    }

    fn close<W: ?Sized + io::Write>(&mut self, writer: &mut W, bracket: &[u8]) -> io::Result<()> {
        self.level -= 1;
        if self.has_value {
            self.newline_indent(writer)?;
        }
        writer.write_all(bracket)
    }

    fn begin_item<W: ?Sized + io::Write>(&mut self, writer: &mut W, first: bool) -> io::Result<()> {
        if !first {
            writer.write_all(self.item_sep.as_bytes())?;
        }
        self.newline_indent(writer)
    }
}

/// Python's `json.dumps(x, ensure_ascii=False)`.
impl Default for PyJsonFormatter<'static> {
    fn default() -> Self {
        Self::new(", ", ": ", None, false)
    }
}

impl Formatter for PyJsonFormatter<'_> {
    fn write_f64<W: ?Sized + io::Write>(&mut self, writer: &mut W, value: f64) -> io::Result<()> {
        write_python_float(writer, value)
    }

    fn write_number_str<W: ?Sized + io::Write>(
        &mut self,
        writer: &mut W,
        value: &str,
    ) -> io::Result<()> {
        writer.write_all(python_number_repr(value).as_bytes())
    }

    fn begin_array<W: ?Sized + io::Write>(&mut self, writer: &mut W) -> io::Result<()> {
        self.open(writer, b"[")
    }

    fn end_array<W: ?Sized + io::Write>(&mut self, writer: &mut W) -> io::Result<()> {
        self.close(writer, b"]")
    }

    fn begin_array_value<W: ?Sized + io::Write>(
        &mut self,
        writer: &mut W,
        first: bool,
    ) -> io::Result<()> {
        self.begin_item(writer, first)
    }

    fn end_array_value<W: ?Sized + io::Write>(&mut self, _writer: &mut W) -> io::Result<()> {
        self.has_value = true;
        Ok(())
    }

    fn begin_object<W: ?Sized + io::Write>(&mut self, writer: &mut W) -> io::Result<()> {
        self.open(writer, b"{")
    }

    fn end_object<W: ?Sized + io::Write>(&mut self, writer: &mut W) -> io::Result<()> {
        self.close(writer, b"}")
    }

    fn begin_object_key<W: ?Sized + io::Write>(
        &mut self,
        writer: &mut W,
        first: bool,
    ) -> io::Result<()> {
        self.begin_item(writer, first)
    }

    fn begin_object_value<W: ?Sized + io::Write>(&mut self, writer: &mut W) -> io::Result<()> {
        writer.write_all(self.key_sep.as_bytes())
    }

    fn end_object_value<W: ?Sized + io::Write>(&mut self, _writer: &mut W) -> io::Result<()> {
        self.has_value = true;
        Ok(())
    }

    // serde_json already escaped control characters, `"` and `\`. Python's
    // `ensure_ascii` additionally escapes everything outside `' '..='~'` as
    // `\uXXXX`, using UTF-16 surrogate pairs above the BMP.
    fn write_string_fragment<W: ?Sized + io::Write>(
        &mut self,
        writer: &mut W,
        fragment: &str,
    ) -> io::Result<()> {
        if !self.ensure_ascii {
            return writer.write_all(fragment.as_bytes());
        }
        let mut start = 0;
        for (i, c) in fragment.char_indices() {
            if !(' '..='~').contains(&c) {
                writer.write_all(&fragment.as_bytes()[start..i])?;
                for unit in c.encode_utf16(&mut [0; 2]) {
                    write!(writer, "\\u{unit:04x}")?;
                }
                start = i + c.len_utf8();
            }
        }
        writer.write_all(&fragment.as_bytes()[start..])
    }
}

/// Python's `str.isspace` for one character: Rust's `char::is_whitespace` plus
/// U+001C..U+001F, which Python also counts as whitespace.
pub(crate) fn is_python_space(c: char) -> bool {
    c.is_whitespace() || ('\u{1c}'..='\u{1f}').contains(&c)
}

/// Python's `str.split()` with no separator: whitespace runs separate the words,
/// and after `maxsplit` splits the rest is one word, its trailing whitespace kept.
pub(crate) fn python_split_whitespace(s: &str, maxsplit: Option<usize>) -> Vec<&str> {
    let mut words = Vec::new();
    let mut rest = s;
    loop {
        rest = rest.trim_start_matches(is_python_space);
        if rest.is_empty() {
            return words;
        }
        if maxsplit == Some(words.len()) {
            words.push(rest);
            return words;
        }
        let end = rest.find(is_python_space).unwrap_or(rest.len());
        words.push(&rest[..end]);
        rest = &rest[end..];
    }
}

/// Python's `str.splitlines`: `\r\n` and each of `\n \r \v \f \x1c \x1d \x1e \x85
/// U+2028 U+2029` end a line, kept with `keepends`.
pub(crate) fn python_splitlines(s: &str, keepends: bool) -> Vec<&str> {
    let mut lines = Vec::new();
    let mut start = 0;
    let mut chars = s.char_indices().peekable();
    while let Some((i, c)) = chars.next() {
        if !matches!(
            c,
            '\n' | '\r'
                | '\u{b}'
                | '\u{c}'
                | '\u{1c}'
                | '\u{1d}'
                | '\u{1e}'
                | '\u{85}'
                | '\u{2028}'
                | '\u{2029}'
        ) {
            continue;
        }
        let mut end = i + c.len_utf8();
        if c == '\r' && chars.next_if(|&(_, next)| next == '\n').is_some() {
            end += 1;
        }
        lines.push(&s[start..if keepends { end } else { i }]);
        start = end;
    }
    if start < s.len() {
        lines.push(&s[start..]);
    }
    lines
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn template_boundaries_keep_numbers_numeric() {
        let json = serde_json::json!({"values": [1, 1e-7, 1e15, -0.0]});
        let template = template_value(&json);
        let values = template.get_attr("values").unwrap();
        for value in values.try_iter().unwrap() {
            assert!(value.is_number());
        }
        assert_eq!(template_json(template).unwrap(), json);
        let externally_serialized = minijinja::Value::from_serialize(&json);
        assert_eq!(template_json(externally_serialized).unwrap(), json);
    }

    #[test]
    fn number_spelling_preserves_exact_digits() {
        for (input, expected) in [
            ("1e-7", "1e-07"),
            ("1e15", "1000000000000000.0"),
            ("1E-6", "1e-06"),
            ("-0.0", "-0.0"),
            ("9007199254740993", "9007199254740993"),
            ("9007199254740993.25", "9007199254740993.25"),
            ("1e-400", "1e-400"),
            ("1e99999999999999999999", "1e99999999999999999999"),
        ] {
            let mut compact = Vec::new();
            PyJsonFormatter::default()
                .write_number_str(&mut compact, input)
                .unwrap();
            let mut indented = Vec::new();
            PyJsonFormatter::new(",", ": ", Some("  "), false)
                .write_number_str(&mut indented, input)
                .unwrap();
            assert_eq!(String::from_utf8(compact).unwrap(), expected);
            assert_eq!(String::from_utf8(indented).unwrap(), expected);
        }
    }

    #[test]
    fn floats_match_python_repr() {
        for (value, python) in [
            (0.000001, "1e-06"),
            (0.0001, "0.0001"),
            (0.00001234, "1.234e-05"),
            (1e16, "1e+16"),
            (1e15, "1000000000000000.0"),
            (1.5e-7, "1.5e-07"),
            (2.5, "2.5"),
            (100.0, "100.0"),
            (0.1, "0.1"),
            (-0.0, "-0.0"),
            (0.0, "0.0"),
            (-123.456, "-123.456"),
            (1.7976931348623157e308, "1.7976931348623157e+308"),
            (5e-324, "5e-324"),
            // Exactly halfway between two shortest candidates.
            (1e15 + 0.25, "1000000000000000.2"),
            (1e14 + 0.125, "100000000000000.12"),
            (f64::NAN, "nan"),
            (f64::INFINITY, "inf"),
            (f64::NEG_INFINITY, "-inf"),
        ] {
            assert_eq!(python_float_repr(value), python, "{value:?}");
        }
    }

    /// Expected values from CPython 3.12's `str.split()`, `str.splitlines()` and `str.isspace()`.
    #[test]
    fn whitespace_matches_python() {
        let all_breaks = "a\rb\r\nc\u{b}d\u{c}e\u{1c}f\u{1d}g\u{1e}h\u{85}i\u{2028}j\u{2029}k\nl\n";
        let split: &[(&str, Option<usize>, &[&str])] = &[
            ("\u{1c} answer \u{1f}", None, &["answer"]),
            ("  a  b \u{1c} c\u{1d}", None, &["a", "b", "c"]),
            ("  a  b \u{1c} c\u{1d}", Some(1), &["a", "b \u{1c} c\u{1d}"]),
            ("  lead and trail  ", Some(1), &["lead", "and trail  "]),
            ("  lead and trail  ", Some(0), &["lead and trail  "]),
            ("\u{85}\u{2028}x\u{2029}\u{3000}", None, &["x"]),
            ("\u{1c}\u{1d}\u{1e}\u{1f}", None, &[]),
            ("", None, &[]),
        ];
        for &(s, maxsplit, words) in split {
            assert_eq!(
                python_split_whitespace(s, maxsplit),
                words,
                "{s:?} {maxsplit:?}"
            );
        }
        let splitlines: &[(&str, &[&str], &[&str])] = &[
            (
                all_breaks,
                &["a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l"],
                &[
                    "a\r",
                    "b\r\n",
                    "c\u{b}",
                    "d\u{c}",
                    "e\u{1c}",
                    "f\u{1d}",
                    "g\u{1e}",
                    "h\u{85}",
                    "i\u{2028}",
                    "j\u{2029}",
                    "k\n",
                    "l\n",
                ],
            ),
            (
                "\u{1c} answer \u{1f}",
                &["", " answer \u{1f}"],
                &["\u{1c}", " answer \u{1f}"],
            ),
            ("plain text", &["plain text"], &["plain text"]),
            ("", &[], &[]),
        ];
        for &(s, lines, kept) in splitlines {
            assert_eq!(python_splitlines(s, false), lines, "{s:?}");
            assert_eq!(python_splitlines(s, true), kept, "{s:?}");
        }
        assert!("\u{1c}\u{1d}\u{1e}\u{1f}".chars().all(is_python_space));
        assert!(!"\u{200b}".chars().all(is_python_space));
    }
}