chat_completions.rs 15 KB
Newer Older
1
// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
3
// SPDX-License-Identifier: Apache-2.0

4
5
use dynamo_runtime::protocols::annotated::AnnotationsProvider;
use serde::{Deserialize, Serialize};
6
use utoipa::ToSchema;
7
8
use validator::Validate;

9
use crate::engines::ValidateRequest;
10
use crate::preprocessor::media::MediaDecoder;
11
12

use super::{
13
    OpenAIOutputOptionsProvider, OpenAISamplingOptionsProvider, OpenAIStopConditionsProvider,
14
    common_ext::{CommonExt, CommonExtProvider},
15
16
    nvext::NvExt,
    nvext::NvExtProvider,
17
    tools, validate,
18
};
19

20
pub mod aggregator;
21
mod delta;
Ryan Olson's avatar
Ryan Olson committed
22
pub mod jail;
23

Paul Hendricks's avatar
Paul Hendricks committed
24
pub use aggregator::DeltaAggregator;
25
26
pub use delta::DeltaGenerator;

27
/// A request structure for creating a chat completion, extending OpenAI's
28
/// `CreateChatCompletionRequest` with [`NvExt`] extensions and common fields.
29
30
31
///
/// # Fields
/// - `inner`: The base OpenAI chat completion request, embedded using `serde(flatten)`.
32
33
34
/// - `common`: Common extension fields (ignore_eos, min_tokens) at root level, embedded using `serde(flatten)`.
/// - `nvext`: The optional NVIDIA extension field. See [`NvExt`] for more details.
///   Note: If ignore_eos is specified in both common and nvext, the common (root-level) value takes precedence.
35
#[derive(ToSchema, Serialize, Deserialize, Validate, Debug, Clone)]
36
pub struct NvCreateChatCompletionRequest {
Paul Hendricks's avatar
Paul Hendricks committed
37
    #[serde(flatten)]
38
    pub inner: dynamo_protocols::types::CreateChatCompletionRequest,
39

40
41
42
    #[serde(flatten, default)]
    pub common: CommonExt,

43
    #[serde(skip_serializing_if = "Option::is_none")]
44
    pub nvext: Option<NvExt>,
45
46

    /// Extra args to pass to the chat template rendering context
47
48
49
50
51
52
    /// Also accepts "chat_template_kwargs" as an alias for compatibility
    #[serde(
        default,
        skip_serializing_if = "Option::is_none",
        alias = "chat_template_kwargs"
    )]
53
    pub chat_template_args: Option<std::collections::HashMap<String, serde_json::Value>>,
54

55
56
57
58
59
60
    /// Runtime media decoding parameters.
    /// When provided, these override the MDC defaults
    /// Example: `{"video": {"num_frames": 16}}`
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub media_io_kwargs: Option<MediaDecoder>,

61
62
63
    /// Catch-all for unsupported fields - checked during validation
    #[serde(flatten, default, skip_serializing)]
    pub unsupported_fields: std::collections::HashMap<String, serde_json::Value>,
64
65
}

66
/// A response structure for unary chat completion responses, embedding OpenAI's
67
68
69
70
/// `CreateChatCompletionResponse` with optional NVIDIA extension metadata.
#[derive(Serialize, Deserialize, Debug, Clone, PartialEq)]
pub struct NvCreateChatCompletionResponse {
    #[serde(flatten)]
71
    pub inner: dynamo_protocols::types::CreateChatCompletionResponse,
72
73
74
    #[serde(skip_serializing_if = "Option::is_none")]
    pub nvext: Option<serde_json::Value>,
}
75

76
/// A response structure for streamed chat completions, embedding OpenAI's
77
78
79
80
/// `CreateChatCompletionStreamResponse` with optional NVIDIA extension metadata.
#[derive(Serialize, Deserialize, Debug, Clone, PartialEq)]
pub struct NvCreateChatCompletionStreamResponse {
    #[serde(flatten)]
81
    pub inner: dynamo_protocols::types::CreateChatCompletionStreamResponse,
82
83
84
    #[serde(skip_serializing_if = "Option::is_none")]
    pub nvext: Option<serde_json::Value>,
}
85

86
87
/// Implements `NvExtProvider` for `NvCreateChatCompletionRequest`,
/// providing access to NVIDIA-specific extensions.
88
impl NvExtProvider for NvCreateChatCompletionRequest {
89
    /// Returns a reference to the optional `NvExt` extension, if available.
90
91
92
93
    fn nvext(&self) -> Option<&NvExt> {
        self.nvext.as_ref()
    }

94
    /// Returns `None`, as raw prompt extraction is not implemented.
95
96
97
98
99
    fn raw_prompt(&self) -> Option<String> {
        None
    }
}

100
101
/// Implements `AnnotationsProvider` for `NvCreateChatCompletionRequest`,
/// enabling retrieval and management of request annotations.
102
impl AnnotationsProvider for NvCreateChatCompletionRequest {
103
    /// Retrieves the list of annotations from `NvExt`, if present.
Biswa Panda's avatar
Biswa Panda committed
104
105
106
107
108
109
    fn annotations(&self) -> Option<Vec<String>> {
        self.nvext
            .as_ref()
            .and_then(|nvext| nvext.annotations.clone())
    }

110
111
112
113
114
115
116
    /// Checks whether a specific annotation exists in the request.
    ///
    /// # Arguments
    /// * `annotation` - A string slice representing the annotation to check.
    ///
    /// # Returns
    /// `true` if the annotation exists, `false` otherwise.
Biswa Panda's avatar
Biswa Panda committed
117
118
119
120
121
122
123
124
    fn has_annotation(&self, annotation: &str) -> bool {
        self.nvext
            .as_ref()
            .and_then(|nvext| nvext.annotations.as_ref())
            .map(|annotations| annotations.contains(&annotation.to_string()))
            .unwrap_or(false)
    }
}
125

126
127
/// Implements `OpenAISamplingOptionsProvider` for `NvCreateChatCompletionRequest`,
/// exposing OpenAI's sampling parameters for chat completion.
128
impl OpenAISamplingOptionsProvider for NvCreateChatCompletionRequest {
129
    /// Retrieves the temperature parameter for sampling, if set.
130
    fn get_temperature(&self) -> Option<f32> {
Paul Hendricks's avatar
Paul Hendricks committed
131
        self.inner.temperature
132
133
    }

134
    /// Retrieves the top-p (nucleus sampling) parameter, if set.
135
    fn get_top_p(&self) -> Option<f32> {
Paul Hendricks's avatar
Paul Hendricks committed
136
        self.inner.top_p
137
138
    }

139
    /// Retrieves the frequency penalty parameter, if set.
140
    fn get_frequency_penalty(&self) -> Option<f32> {
Paul Hendricks's avatar
Paul Hendricks committed
141
        self.inner.frequency_penalty
142
143
    }

144
    /// Retrieves the presence penalty parameter, if set.
145
    fn get_presence_penalty(&self) -> Option<f32> {
Paul Hendricks's avatar
Paul Hendricks committed
146
        self.inner.presence_penalty
147
148
    }

149
    /// Returns a reference to the optional `NvExt` extension, if available.
150
151
152
    fn nvext(&self) -> Option<&NvExt> {
        self.nvext.as_ref()
    }
153
154
155
156
157
158
159
160
161
162
163
164
165
166
    /// Retrieves the seed value for random number generation, if set.
    fn get_seed(&self) -> Option<i64> {
        self.inner.seed
    }

    /// Retrieves the number of completions to generate for each prompt, if set.
    fn get_n(&self) -> Option<u8> {
        self.inner.n
    }

    /// Retrieves the best_of parameter, if set.
    fn get_best_of(&self) -> Option<u8> {
        None // Not supported in chat completions
    }
167
168
}

169
170
171
172
173
174
175
176
177
/// Implements `CommonExtProvider` for `NvCreateChatCompletionRequest`,
/// providing access to common extension fields.
impl CommonExtProvider for NvCreateChatCompletionRequest {
    /// Returns a reference to the CommonExt struct.
    fn common_ext(&self) -> Option<&CommonExt> {
        Some(&self.common)
    }

    /// Guided Decoding Options
178
179
180
181
182
    fn get_guided_json(&self) -> Option<serde_json::Value> {
        if let Some(value) = self.common.guided_json.clone() {
            return Some(value);
        }

183
184
185
186
187
188
189
190
191
192
193
194
195
        // 1) Tool-call guided decoding (highest precedence after explicit guided_json)
        if let (Some(tool_choice), Some(tools)) =
            (self.inner.tool_choice.as_ref(), self.inner.tools.as_deref())
        {
            match tools::get_json_schema_from_tools(Some(tool_choice), Some(tools)) {
                Ok(Some(schema)) => return Some(schema),
                Ok(None) => {}
                Err(err) => {
                    tracing::warn!(
                        error = %err,
                        "failed to derive guided_json from tool_choice"
                    );
                }
196
197
            }
        }
198
199
200

        // 2) OpenAI `response_format` (applies to assistant content, not tool calls)
        if let Some(response_format) = self.inner.response_format.as_ref() {
201
            use dynamo_protocols::types::ResponseFormat;
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
            match response_format {
                ResponseFormat::Text => {}
                ResponseFormat::JsonObject => {
                    // Minimal JSON Schema for "any JSON object"
                    return Some(serde_json::json!({
                        "type": "object"
                    }));
                }
                ResponseFormat::JsonSchema { json_schema } => {
                    // validate_response_format ensures schema is present when type=json_schema
                    if let Some(schema) = json_schema.schema.clone() {
                        return Some(schema);
                    }
                }
            }
        }

        None
220
221
222
    }

    fn get_guided_regex(&self) -> Option<String> {
223
        self.common.guided_regex.clone()
224
225
226
    }

    fn get_guided_grammar(&self) -> Option<String> {
227
        self.common.guided_grammar.clone()
228
229
230
    }

    fn get_guided_choice(&self) -> Option<Vec<String>> {
231
        self.common.guided_choice.clone()
232
233
234
    }

    fn get_guided_decoding_backend(&self) -> Option<String> {
235
        self.common.guided_decoding_backend.clone()
236
    }
237

238
    fn get_guided_whitespace_pattern(&self) -> Option<String> {
239
        self.common.guided_whitespace_pattern.clone()
240
241
    }

242
    fn get_top_k(&self) -> Option<i32> {
243
        self.common.top_k
244
245
    }

246
    fn get_min_p(&self) -> Option<f32> {
247
        self.common.min_p
248
249
    }

250
    fn get_repetition_penalty(&self) -> Option<f32> {
251
        self.common.repetition_penalty
252
    }
253
254
255
256

    fn get_include_stop_str_in_output(&self) -> Option<bool> {
        self.common.include_stop_str_in_output
    }
257
258
259
260

    fn get_skip_special_tokens(&self) -> Option<bool> {
        self.common.skip_special_tokens
    }
261
262
}

263
264
/// Implements `OpenAIStopConditionsProvider` for `NvCreateChatCompletionRequest`,
/// providing access to stop conditions that control chat completion behavior.
265
impl OpenAIStopConditionsProvider for NvCreateChatCompletionRequest {
266
    /// Retrieves the maximum number of tokens allowed in the response.
267
    #[allow(deprecated)]
Paul Hendricks's avatar
Paul Hendricks committed
268
    fn get_max_tokens(&self) -> Option<u32> {
269
        self.inner.max_completion_tokens.or(self.inner.max_tokens)
270
271
    }

272
    /// Retrieves the minimum number of tokens required in the response.
273
274
    /// Returns `min_tokens` Value
    /// `min_tokens` is not an OpenAI-supported parameter.
Paul Hendricks's avatar
Paul Hendricks committed
275
    fn get_min_tokens(&self) -> Option<u32> {
276
        self.common.min_tokens
277
278
    }

279
280
281
282
283
284
285
    /// Retrieves the stop conditions that terminate the chat completion response.
    ///
    /// Converts OpenAI's `Stop` enum to a `Vec<String>`, normalizing the representation.
    ///
    /// # Returns
    /// * `Some(Vec<String>)` if stop conditions are set.
    /// * `None` if no stop conditions are defined.
286
    fn get_stop(&self) -> Option<Vec<String>> {
287
        self.inner.stop.as_ref().map(|stop| match stop {
288
289
            dynamo_protocols::types::Stop::String(s) => vec![s.clone()],
            dynamo_protocols::types::Stop::StringArray(arr) => arr.clone(),
290
        })
291
292
    }

293
    /// Returns a reference to the optional `NvExt` extension, if available.
294
295
296
    fn nvext(&self) -> Option<&NvExt> {
        self.nvext.as_ref()
    }
297
298
299
300
301

    /// Get ignore_eos from CommonExt.
    fn get_common_ignore_eos(&self) -> Option<bool> {
        self.common.ignore_eos
    }
302

303
    /// Get the effective ignore_eos value from CommonExt.
304
    fn get_ignore_eos(&self) -> Option<bool> {
305
        self.common.ignore_eos
306
    }
307
}
308

Greg Clark's avatar
Greg Clark committed
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
impl OpenAIOutputOptionsProvider for NvCreateChatCompletionRequest {
    fn get_logprobs(&self) -> Option<u32> {
        match self.inner.logprobs {
            Some(true) => match self.inner.top_logprobs {
                Some(top_logprobs) => Some(top_logprobs as u32),
                None => Some(1_u32),
            },
            Some(false) => None,
            None => None,
        }
    }

    fn get_prompt_logprobs(&self) -> Option<u32> {
        None
    }

    fn get_skip_special_tokens(&self) -> Option<bool> {
326
        CommonExtProvider::get_skip_special_tokens(self)
Greg Clark's avatar
Greg Clark committed
327
328
329
330
331
332
333
    }

    fn get_formatted_prompt(&self) -> Option<bool> {
        None
    }
}

334
335
336
337
/// Implements `ValidateRequest` for `NvCreateChatCompletionRequest`,
/// allowing us to validate the data.
impl ValidateRequest for NvCreateChatCompletionRequest {
    fn validate(&self) -> Result<(), anyhow::Error> {
338
        validate::validate_no_unsupported_fields(&self.unsupported_fields)?;
339
340
341
342
        validate::validate_messages(&self.inner.messages)?;
        validate::validate_model(&self.inner.model)?;
        // none for store
        validate::validate_reasoning_effort(&self.inner.reasoning_effort)?;
343
        // none for metadata
344
345
346
347
348
349
350
351
352
353
354
        validate::validate_frequency_penalty(self.inner.frequency_penalty)?;
        validate::validate_logit_bias(&self.inner.logit_bias)?;
        // none for logprobs
        validate::validate_top_logprobs(self.inner.top_logprobs)?;
        // validate::validate_max_tokens(self.inner.max_tokens)?; // warning depricated field
        validate::validate_max_completion_tokens(self.inner.max_completion_tokens)?;
        validate::validate_n(self.inner.n)?;
        // none for modalities
        // none for prediction
        // none for audio
        validate::validate_presence_penalty(self.inner.presence_penalty)?;
355
        validate::validate_response_format(&self.inner.response_format)?;
356
357
358
359
360
361
362
363
364
365
366
367
368
        // none for seed
        validate::validate_service_tier(&self.inner.service_tier)?;
        validate::validate_stop(&self.inner.stop)?;
        // none for stream
        // none for stream_options
        validate::validate_temperature(self.inner.temperature)?;
        validate::validate_top_p(self.inner.top_p)?;
        validate::validate_tools(&self.inner.tools.as_deref())?;
        // none for tool_choice
        // none for parallel_tool_calls
        validate::validate_user(self.inner.user.as_deref())?;
        // none for function call
        // none for functions
369
370
        // Common Ext
        validate::validate_repetition_penalty(self.get_repetition_penalty())?;
371
372
        validate::validate_min_p(self.get_min_p())?;
        validate::validate_top_k(self.get_top_k())?;
373
374
        // Cross-field validation
        validate::validate_n_with_temperature(self.inner.n, self.inner.temperature)?;
375
376
377
378

        Ok(())
    }
}
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

#[cfg(test)]
mod tests {
    use super::*;
    use crate::protocols::common::OutputOptionsProvider;
    use serde_json::json;

    #[test]
    fn test_skip_special_tokens_none() {
        let json_str = json!({
            "model": "test-model",
            "messages": [
                {"role": "user", "content": "Hello"}
            ]
        });

        let request: NvCreateChatCompletionRequest =
            serde_json::from_value(json_str).expect("Failed to deserialize request");

        assert_eq!(request.common.skip_special_tokens, None);

        let output_options = request
            .extract_output_options()
            .expect("Failed to extract output options");

        assert_eq!(output_options.skip_special_tokens, None);
    }

    #[test]
    fn test_skip_special_tokens_propagates() {
        for skip_value in [true, false] {
            let json_str = json!({
                "model": "test-model",
                "messages": [
                    {"role": "user", "content": "Hello"}
                ],
                "skip_special_tokens": skip_value
            });

            let request: NvCreateChatCompletionRequest =
                serde_json::from_value(json_str).expect("Failed to deserialize request");

            let output_options = request
                .extract_output_options()
                .expect("Failed to extract output options");

            assert_eq!(output_options.skip_special_tokens, Some(skip_value));
        }
    }
}